text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'We show that preferential rewiring, which is supposed to mimick the behaviour of financial agents, changes a directed-network Ising ferromagnet with a single critical point into a model with robust critical behaviour. For the non-rewired random graph version, due to a constant number of out-links for each site, we write a simple mean-field-like equation describing the behaviour of magnetization; we argue that it is exact and support the claim with extensive Monte Carlo simulations. For the rewired version, this equation is obeyed only at low temperatures. At higher temperatures, rewiring leads to strong heterogeneities, which apparently invalidates mean-field arguments and induces large fluctuations and divergent susceptibility. Such behaviour is traced back to the formation of a relatively small core of agents which influence the entire system.'
author:
- Adam Lipowski
- Krzysztof Gontarek
- Dorota Lipowska
title: Robust criticality of Ising model on rewired directed networks
---
Introduction
============
To understand a complex behaviour of financial markets is one of the main objectives of econophysics. Fat-tail non-Gaussian fluctuations, volatility clustering or rapid decay of autocorrelations of returns characterize most of the financial markets, suggesting that these stylized facts [@stylized] have some more fundamental explanation. Searching for such an explanation, one can resort to the approach particularly suited for physicists, namely agent modeling [@agent]. In the spirit of statistical mechanics, one considers a collection of agents involved in interactions resembling functioning of financial markets. Since buying and selling are the most important activities of such agents, a number of models of financial markets bear some similarity to the two-state percolation [@cont] or Ising-like models [@ising].
An important agent’s interaction is mimicking some other agents’ behaviour, which suggests a similarity to ferromagnetic systems. However, ferromagnets typically exhibit rather small fluctuations, which is much different from the behaviour of finacial markets. Ferromagnets exhibit large fluctuations only at the critical point separating ferromagnetic and paramagnetic phases. To place the system at the critical point requires, however, a fine tuning of control parameters. On the other hand, financial markets seem to be more robust with strong fluctuations appearing without any tuning of parameters. Some models were proposed, where mimicking other agents’ behaviour is compensated with the tendency to be in the minority [@bornholdt] or where agents with more complex strategies were used [@lux]. They do reproduce some of the stylized facts but their considerable complexity hinders deeper understanding.
Apparently, the analogy with simple ferromagnetic systems is not sufficient to model financial markets and one should search for more suitable extensions. In our opinion, an important ingredient of models of financial markets should be the possibility to choose and sometimes also change neighbours that a given agent would like to mimick. The objective of the present paper is to implement such a rewiring mechanism and to show that it drastically affects the behaviour of the model. When the neighbours to be mimicked are selected at random and kept fixed, the model behaves as an ordinary ferromagnet with ferromagnetic and paramagnetic phases separated at a critical point. However, when agents might switch the neighbours and preferentially select those they consider as more influential, the system generically exhibits divergent fluctuations. Such behaviour indicates that preferential rewiring induces a robust criticality, which is a required feature of stock-market models [@tanaka]. We also examine the mechanism leading to the robust criticality.
Model without rewiring
======================
In our model we consider $N$ agents represented by spin-like variables $s_i=\pm 1, \ i=1,2\ldots, N$. At each time step $t$, each agents decides whether to buy ($s_i=1$) or sell ($s_i=-1$) an asset. To make the decision, an agent tries to mimick the behaviour of its neighbours and the model evolves according to the heat-bath dynamics: $$s_i(t+1)=\left\{ \begin{array}{ll} 1 & {\rm \ with \ probability\ } p=\frac{1}{1+\exp[-2h_i(t)/T]}\\
-1 & {\rm \ with \ probability\ } 1-p
\end{array}\right.
\label{heat-bath}$$ where $$h_i(t)=\sum_{j} s_j(t)
\label{field}$$ is the local field acting on a given agent $i$ and summation in Eq. (\[field\]) is over its neighbours. The control parameter $T$ is the analogue of the temperature in the magnetic Ising model and determines the level of fluctuations in the decision process.
The neighbourhood of a given agent is set randomly, namely, each agent has a fixed number of $z$ randomly selected neighbours, which it interacts with via the local field. The neighbouring relation is not necessarily symmetric: if agent $j$ enters the expression for the local field of agent $i$, it does not imply that agent $i$ enters the expression for the local field of agent $j$. In other words, agents are nodes of a directed random network and each node has $z$ out-links (arrows point at the nodes that contribute to the local field). The number of in-links of a given agent, which specifies how many agents it influences, is not fixed and it can vary among agents (of course, the average over all agents equals $z$). Equal numbers of out-links and unequal numbers of in-links constitute an important feature of our model, which we will refer to as the out-homogeneity.
Taking into account the spin variables, the above rules define actually an Ising ferromagnet on a directed random graph. Models of this kind were already analysed and shown to exibit an ordinary ferro-para phase transition belonging to the mean-field universality class [@directed-ising].
In the following, we present a more detailed analysis of our model for $z=4$. Due to the out-homogeneity, one can write a relatively simple equation, which governs the evolution of magnetization. Let $P_{i}(t)$ denote the probability that agent $i$ at time $t$ takes the value $s_i=1$. Assuming that $P_{i}(t)$ is spatially homogenous and does not depend on $i$, from the heat-bath rules we obtain that $$P(t+1)=\sum_{k=0}^{4} {4 \choose k} P^k(t)[1-P(t)]^{4-k}\frac{1}{1+\exp{[-4(k-2)/T]}} .
\label{mfa}$$ Of course, Eq. (\[mfa\]) can be easily rewritten in terms of magnetization ($m(t)=2P(t)-1$), which is common in Ising-model studies. In the steady-state limit ($t\rightarrow \infty$). Eq. (\[mfa\]) becomes a 4-th order polynomial equation, which can be easily solved numerically (and with some more effort even analytically). Moreover, the critical temperature $T_c$ can be found using the standard procedure of expanding the $t\rightarrow \infty$ limit of Eq. (\[mfa\]) in the vicinity of the critical point. Elementary calculations reveal that $T_c$ obeys $$2=\tanh{(4/T_c)}+2\tanh{(2/T_c)} .
\label{tc}$$ The solution of the equation above can be written as $$T_c=\frac{4}{\ln{(\frac{1+x}{1-x})}} ,
\label{tc1}$$ where $$x=\frac{1}{3}\Bigg[1-5\sqrt[3]{\frac{2}{11+3\sqrt{69}}}+\sqrt[3]{\frac{1}{2}(11+3\sqrt{69})} \Bigg] .
\label{x}$$ We thus obtain $T_c \approx 3.08982$, approximately.
The factorized form of the probabilities suggests that Eq. (\[mfa\]) is nothing more than the mean-field equation for our model and thus it is only approximate. This would certainly be the case for undirected graphs, where neighbours $j$ and $k$ of agent $i$ are strongly correlated (since $i$ contributes to the local fields of both $j$ and $k$). For undirected random graphs, some insight into the behaviour of the Ising model can obtained using a replica method [@replica] or some recurrence relations based on the similarity of random graphs to Cayley trees [@dorog]. On the other hand, in directed networks even though $j$ and $k$ are neighbours of $i$, they are not more correlated than any other two randomly selected nodes (Fig. \[ijk\]). Since the graph is sparse, we expect that in the limit $N\rightarrow\infty$, such correlations are negligible, and consequently, the factorization in Eq. (\[mfa\]) should be legitimate.
![In a directed random graph, neighbours $j$ and $k$ of node $i$ are not more correlated than any other two randomly selected nodes.[]{data-label="ijk"}](ijk.eps){width="3cm"}
Monte Carlo simulations of our model confirm the above analysis (Fig. \[magnet\]). Calculating the magnetization for $N=10^4$ and $z=4$, we find it in a very good agreement with $m=2P(t=\infty)-1$ obtained from the numerical solution of the steady-state limit of Eq. (\[mfa\]).
![Temperature dependence of the magnetization $m$ for the rewired and non-rewired models ($z=4$). The data are obtained from Monte Carlo simulations and are compared with the numerical solution of Eq. (\[mfa\]). Simulation and equilibration times were equal to $10^4$ Monte Carlo steps.[]{data-label="magnet"}](magnet.eps){width="\columnwidth"}
For $T=3$, we made much more extensive calculations (Fig. \[check\]). The linear extrapolation $N\rightarrow\infty$ based on simulations for $N\leq 3\cdot 10^6$ gives $m=0.34723(2)$, which is in perfect agreement with $m=0.347225\ldots$ obtained from the numerical solution of Eq. (\[mfa\]). In our opinion, such agreement strongly supports the claim that Eq. (\[mfa\]) is exact (at least in the limit $t\rightarrow\infty$).
![Magnetization $m$ as a function of the inverse of size calculated for the non-rewired model of size $N=10^4$, $3\cdot 10^4$, $10^5$, $3\cdot 10^5$, $10^6$, and $3\cdot 10^6$, with $z=4$ and $T=3$. Simulation and equilibration times were equal to $10^7$ and $10^4$ Monte Carlo steps, respectively. In the limit $N\rightarrow\infty$, a perfect agreement with the solution of Eq. (\[mfa\]) can be seen.[]{data-label="check"}](check.eps){width="\columnwidth"}
Model with rewiring
===================
The model analysed in the previous section behaves similarly to some other Ising-like models with ferromagnetic and paramagnetic phases separated at the critical point. Our primary motivation is to modify such ordinary ferromagnets so that they would resemble the behaviour of financial markets, at least to some extent. We are particularly interested in supplanting a fine-tuned critical point with a more generic critical behaviour, which would exist in some, possibly large, temperature range. So far our agents make the decision to buy or sell based on the observation of their $z$ neighbours, and the assignement of these neighbours is fixed during the entire evolution of the model. In the present section we modify this rule and allow to change the neighbours. The rewiring we use is preferential: each agent has its status equal to the number of in-links that are (currently) attached to it. The selection of a new neighbour takes place with probability proportional to its status [@liproulette]. A single step of the dynamics of our model is thus defined as follows:
- update spin variables $S_i$ ($i=1, 2,\ldots, N$) according to the heat-bath algorithm (\[heat-bath\]).
- rewire each agent selecting preferentially anew its $z$ out-links.
Since we keep the dynamics of spin variables basically unchanged, one might expect that Eq. (\[mfa\]) still describes the behaviour of our model. Monte Carlo simulations show that to some extent this is indeed the case (Fig. \[magnet\]) and a very good agreement with Eq. (\[mfa\]) can be seen over much of the temperature range. However, close to the critical point $T=T_c$, the rewired model shows much lower and perhaps zero magnetization. It would be desirable to understand the reasons why Eq. (\[mfa\]) is no longer obeyed at higher temperatures. Possible explanations include appearance of correlations (that we argued are negligible in the non-rewired case) or a breakdown of homogeneity (which is also one of the assumptions leading to Eq. (\[mfa\])). Some efforts to understand the origin of this behaviour will be made in the next section.
What is even more interesting, the magnetization in the rewired version shows large fluctuations also at temperatures much higher than $T_c$ (Fig. \[timer0r1\]).
![Time dependence of magnetization $m$ for the rewired and non-rewired models ($z=4$, $N=10^4$). Simulations were made for $T=4$, which for the non-rewired model is deep in the paramagnetic phase.[]{data-label="timer0r1"}](timer0r1.eps){width="\columnwidth"}
To measure these fluctuations more quantitatively, we calculated the susceptibility $\chi$ that up to the temperature factor is equal to the variance of magnetization $\chi = \frac{1}{N}[\langle (\sum_{i=1}^N S_i)^2 \rangle-\langle \sum_{i=1}^N S_i \rangle^2]$. Numerical values indicate that as a function of system size $N$ the susceptibility diverges as $\chi \sim N^\alpha$, where $\alpha\sim 0.67-0.91$ depends slightly on temperature (Fig. \[susc\]). Such behaviour is observed in a large temperature range ($3\leq T\leq 6$) for the system size $10^3\leq N\leq 3 \cdot 10^4$ . The divergence of susceptibility indicates that the model exhibits a robust critical behaviour. Together with data from Fig. \[magnet\], this suggests that the model with rewiring has two phases: low-temperature, which is ferromagnetic, and high-temperature, which is critical. It is difficult for us to locate precisely the transition point between these two phases. For longer simulations, it seems to shift slightly toward lower temperatures. Moreover, one cannot exclude that at sufficiently large temperature the critical phase will be replaced with the paramagnetic one (having much smaller fluctuations).
![Size dependence of susceptibility $\chi$ for the rewired model ($z=4$). The power-law fit $\chi\sim N^{\alpha}$ shows that $\alpha$ varies from 0.67 for $T=6$ up to 0.91 for $T=3$.[]{data-label="susc"}](susc.eps){width="\columnwidth"}
The critical behaviour in our model is also robust with respect to the frequency of rewiring. We made simulations with rewiring taking place, e.g., with probability 0.1 (i.e., with probability 0.9, the out-links of a given agent at a given step were left unchanged). Such modification slows down the dynamics but retains the power-law divergence of the susceptibility.
Dynamics of rewiring
====================
To get some understanding of our model, we looked at the structure of the network that emerges during the rewiring. Let us notice that rewiring is not affected by spin variables and thus might be considered as an independent process (but not vice versa—spin dynamics depends of course on the structure of the network). Some insight is already obtained from simulations of a small system (Fig. \[conf-full\]). One can notice that most agents have no in-links and thus they do not influence any other agent. There is only a small core of agents which are responsible for the decision formation of the other agents. Such structure appears also for larger systems (Fig. \[conf-core\]). One can notice a substantial heterogeneity of the resulting core as for the number of agents that a given agent is influencing.
![Network structure after simulations of $t=10^3$ steps for $N=30$ agents and $z=2$ [@graphviz]. Nodes with no in-links (open circles) represent agents which do not inluence decisions of any other agent (links that go out from them are drawn with dotted lines). Nodes with some in-links (filled red circles) represent the only agents that influence other agents (their out-links are drawn with solid lines).[]{data-label="conf-full"}](net-full.eps){width="6cm"}
![Network structure after simulations of $t=10^4$ steps for $N=10^3$ agents and $z=2$ [@graphviz]. Agents with no in-links are omitted. The size of a circle is proportional to the number of in-links.[]{data-label="conf-core"}](net2a.eps){width="6cm"}
A more detailed analysis shows, however, that the core size $L$ slowly diminishes in time (Fig. \[time-size-size\]). This is not surprising since once an agent looses all of its in-links, it cannot get them back because the probability to be selected in a rewiring process is proportional to the current number of in-links (which is 0 for such an agent). Although the process of diminishing of $L$ is irreversible, it is extremely slow for $z>1$. Only at $z=1$, this process is considerably faster and in a large time interval consistent with $t^{-1}$ decay. Such a slow decay for $z>1$ suggests that at long (but not infinitely long) time, the core is almost in a steady state and has a certain size. Numerical calculations show that for $z=2$ and 4 it increases with the system size approximately as $N^{1/2}$ (inset in Fig. \[time-size-size\]).
![Time dependence of the size of the core (log-log scale); calculations were made for $N=10^4$. Inset shows that for $z=2$ and 4 the ’steady-state’ size of the core increases approximately as $N^{1/2}$. To generate the network of fractional $z$ $(1<z<2)$, with probability $2-z$ we created one out-link and with probability $z-1$ we created two such links for each agent. Thus fractional $z$ has only an average sense.[]{data-label="time-size-size"}](time-size-size.eps){width="\columnwidth"}
Some insight into the stability of the core can be obtained from the analysis of the time $\tau$ needed for the system to condensate, i.e., for a given $z$ to reach the core size $z+1$ (which is the smallest core size that the system can reach). Numerical calculations show that for $z>1$, $\tau$ exhibits a fast, possibly exponential, increase with the system size (Fig. \[size-tau\]). And again, a slower increase ($\sim N$) is obtained only at $z=1$.
![The system size dependence of the time $\tau$ to condensate, ie., to reach the core size $z+1$. Let us notice that even for $z$ slightly larger than 1, $\tau$ shows a rapid increase.[]{data-label="size-tau"}](size-tau.eps){width="\columnwidth"}
Our results in Figs. \[time-size-size\]–\[size-tau\] show that rewiring for $z=1$ leads to a rather fast condensation while for $z>1$ the dynamics is basically trapped in a core of size $\sim N^{1/2}$. Even though the condensed state could be reached in principle, for large $N$ and $z>1$ it virtually never happens. The situation is reminiscent of some models with the so-called absorbing states: for some values of control parameters, the absorbing state of the dynamics is basically unavailable and the model remains in the active phase (the lifetime of which in that regime is also exponentially divergent with the system size) [@absorbing].
A network structure with a nearly stable core suggests an explanation of the generic divergence of susceptibility that we reported in the previous section. Indeed, since agents are influenced only by agents from the core, on average there are $N^{1/2}$ agents that are influenced by a single agent belonging to the core. Considering core agents as independent (and influencing $N^{1/2}$ other agents), we easily obtain that $\chi \sim N^{1/2}$. Numerical results (Fig. \[susc\]) suggest a faster increase at low temperatures (for $T=3$, we obtained $\chi\sim N^{0.91}$) that most likely result from some correlations between core agents. Another factor affecting our simple estimations of the divergence of $\chi$ might be some heterogeneity of the core (Fig. \[conf-core\]).
conclusions and remarks
=======================
In the present paper we have shown that preferential rewiring changes an Ising ferromagnet, which has a single critical point, into a model with robust critical behaviour. The rewiring mechanism that we used is supposed to mimick the behaviour of financial agents who try to follow their neighbours but at the same time have also some freedom to choose the ones to follow. We assume that the preference in the rewiring process is proportional to the number of in-links of a given agent. It is thus not a (more or less) objective measure of its performance but solely how the agent is perceived by the population of other agents. Similar recipes turned out to be successful in, e.g., some page rank algoritms used by search engines [@brin] or various recommendation systems [@amazon].
Our model is of some interest from the statistical-mechanics point of view. Due to out-homogeneity, we could write a simple mean-field-like equation (Eq. (\[mfa\])) that can be used to obtain the magnetization of the model. We argued, however, that for the present model in the non-rewired version this equation should be exact, and numerical simulations provide a very strong support for the claim. A very good agreement with this equation was obtained also for the rewired case, but only in the low-temperature regime. We made some attempts to explain why rewiring invalidates this equation at higher temperatures and at the same time leads to the divergence of suceptibility and criticality. In our opinion this is related with the formation of a relatively small subset of agents that retain some in-links and thus drive the entire system.
The change of the dynamical regime in the rewired version at $z=1$ is also of some interest. It is tempting to associate the change with some percolation transition that for random graphs is known to take place at $z=1$ [@random-graphs]. However, rewiring redistributes links in a highly nonrandom fashion and a possible relation with random graphs is by no means obvious.
We hope that our model might be useful also in the econophysics context. Relatively simple rules that generate a robust criticality might serve as a starting point for further modifications and analysis. For example, one might consider a model where an agent that no longer has any in-links still retains some (small) status $\epsilon$ and can be thus selected during the rewiring process. We expect that for small (possibly $N$-dependent) $\epsilon$ such a model would be similar to our ($\epsilon=0$) model, but certainly numerical simulations would be needed to support such a claim. One of the important stylized facts that apparently is missing in our model is volatility clustering. One might hope that some extensions where agents, for example, try to be in the minority (like in the so-called minority games [@minority]) or use more sophisticated strategies (like ’fundamentalist’, ’trend follower’ or ’noise trader’) will provide a more realistic description of financial markets and at the same time will retain simplicity of the model .
Acknowledgements: The research for this work was supported by NCN grant 2013/09/B/ST6/02277 (A.L.), NCN grant 2011/01/B/HS2/01293 (D.L.) and Ministry of Science and Higher Education grant N N202 488039 (K.G.).
A. Chakraborti, I. M. Toke, M. Patriarca, and F. Abergel, Quantitative Finance [**11**]{}, 991 (2011).
B. LeBaron, Quantitative Finance [**1**]{}, 254 (2001). E. Samanidou, E. Zschischang, D. Stauffer, and T. Lux, Reports on Progress in Physics [**70**]{}, 409 (2007).
R. Cont and J. P. Bouchaud, Macroeconomic Dynamics, [**4**]{}, 170 (2000).
T. Kaizoji, Physica A [**287**]{}, 493 (2000). W.-X. Zhou and D. Sornette, Europ. Phys. J. B [**55**]{}, 175 (2007).
S. Bornholdt, Int. J. Mod. Phys. C [**12**]{}, 667 (2001). T. Kaizoji, S. Bornholdt, and Y. Fujiwara, Physica A [**316**]{}, 441 (2002).
T. Lux and M. Marchesi, Int. J. Theor. Appl. Finance [**3**]{}, 675 (2000).
Some other rewiring mechanisms are known to lead to interesting complex networks: J. Ohkubo and T. Horiguchi, J. Phys. Soc. Jpn. [**74**]{}, 1334 (2005). J. Ohkubo, M. Yasuda, and K. Tanaka, Phys. Rev. E [**72**]{}, 065104 (2005).
F. W. S. Lima and M. A. Sumour, Physica A [**391**]{}, 948 (2012).
M. Leone, A. Vazquez, A. Vespignani, and R. Zecchina, Eur. Phys. J. B [**28**]{}, 191 (2002).
S. N. Dorogovtsev, A. V. Goltsev, and J. F. F. Mendes, Phys. Rev. E [**66**]{}, 016104 (2002).
Such selection mechanism is known as a roulette-wheel selection. Recently, an efficient O(1) algorithm of this process was proposed: A. Lipowski and D. Lipowska, Physica A [**391**]{}, 2193 (2012).
Network graphs were drawn using Graphviz. J. Ellson, E. R. Gansner, E. Koutsofios, S. C. North, and G. Woodhull, [*Graphviz and Dynagraph: Static and Dynamic Graph Drawing Tools*]{}. In [*Graph Drawing Software*]{}, edited by M. Junger and P. Mutzel (Berlin, Springer, 2003) pp. 127-148.
H. Hinrichsen, Adv. Phys. [**49**]{}, 815 (2000). G. Ódor, Rev. Mod. Phys. [**76**]{}, 663 (2004).
S. Brin and L. Page, Computer networks and ISDN systems [**30**]{}, 107 (1998).
G. Linden, B. Smith, and J. York, IEEE Internet Computing [**7**]{}, 76 (2003).
P. Erdös and A. Rényi, Publ. Math. Debrecen [**6**]{}, 290 (1959). M. E. J. Newman, [*Networks: An Introduction*]{} (Oxford, 2010).
W. B. Arthur, Amer. Econ. Rev. [**84**]{}, 406 (1994). D. Challet and Y. C. Zhang, Physica A [**246**]{}, 498 (1997). D. Challet, M. Marsili, and Y. C. Zhang, Physica A [**294**]{}, 514 (2001).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Manas Joglekar
- Rohan Puttagunta
- Chris Ré
bibliography:
- 'ajar-bib\_ARXIV.bib'
title: Aggregations over Generalized Hypertree Decompositions
---
We study a class of aggregate-join queries with multiple aggregation operators evaluated over annotated relations. We show that straightforward extensions of standard multiway join algorithms and generalized hypertree decompositions (GHDs) provide best-known runtime guarantees. In contrast, prior work uses bespoke algorithms and data structures and does not match these guarantees. Our extensions to the standard techniques are a pair of simple tests that (1) determine if two orderings of aggregation operators are equivalent and (2) determine if a GHD is compatible with a given ordering. These tests provide a means to find an optimal GHD that, when provided to standard join algorithms, will correctly answer a given aggregate-join query. The second class of our contributions is a pair of complete characterizations of (1) the set of orderings equivalent to a given ordering and (2) the set of GHDs compatible with some equivalent ordering. We show by example that previous approaches are incomplete. The key technical consequence of our characterizations is a decomposition of a compatible GHD into a set of (smaller) [*unconstrained*]{} GHDs, i.e. into a set of GHDs of sub-queries without aggregations. Since this decomposition is comprised of unconstrained GHDs, we are able to connect to the wide literature on GHDs for join query processing, thereby obtaining improved runtime bounds, MapReduce variants, and an efficient method to find approximately optimal GHDs.
Introduction
============
Generalized hypertree decompositions (GHDs), introduced by Gottlob et al. [@GHDIntro; @gottlob:ght] and further developed by Grohe and Marx [@FHTW], provide a means for performing early projection in join processing, which can result in dramatically faster runtimes. In this work, we extend GHDs to handle queries that include aggregations, which allows us to capture both SQL-aggregate processing and message passing problems. Motivated by our own database engine based on GHDs [@Duncecap1; @Duncecap2; @EH], we seek to more deeply understand the space of optimization for aggregate-join queries.
We build upon work by Green, Karvounarakis, and Tannen [@Semiring] on annotated relations to define our notion of aggregation. These annotations provide a general definition of aggregation, allowing us to represent a wide-ranging set of problems as aggregate-join queries. Our queries, which we call ${\textsc{Ajar}\xspace}$ (Aggregations and Joins over Annotated Relations) queries, contain semiring quantifiers that “sum over” or “marginalize out” values. We formally define ${\textsc{Ajar}\xspace}$ queries in Section \[sec:prelim\], but they are easy to illustrate by example:
\[ex:firstex\] Consider two relations with attributes $\{A,B\}$ and $\{B,C\}$ such that each tuple is annotated with some integer; we call these relations $\mathbb{Z}$-relations. Consider the query: $$\sum_C \sum_B R(A,B) \Join S(B,C)$$ Our output will then be a $\mathbb{Z}$-relation with attribute set $\{A\}$. Each value $a$ of attribute $A$ in $R$ is associated with a set $X_a$ of pairs $(b,z_R)$ composed of a value $b$ of attribute $B$ and an annotation $z_R$. Furthermore for each $b$ value in $X_a$, there is a set $X_b$ from relation $S$ of pairs $(c ,z_S)$ composed of a value $c$ of attribute $C$ and an annotation $z_S$. Given $X_a$ and each $X_b$ associated with a given value $a$, the annotation associated with $a$ in our output will simply be $$\sum_{b,z_1 \in X_a} \sum_{c, z_2 \in X_b} z_1 * z_2.$$
${\textsc{Ajar}\xspace}$ queries capture both classical SQL-style queries and newer data processing problems like probabilistic inference via message passing on graphical models [@Kask:2005:UTD:1090725.1090730]. In fact, Aji and McEliece proposed the “Marginalize a Product Function” (MPF) problem [@Aji], which is a special case of an ${\textsc{Ajar}\xspace}$ query, and showed how the problem and its solution capture a number of classic problems and algorithms, including fast Hadamard transforms, Viterbi’s algorithm, forward-backward algorithm, FFT, and probabilistic inference in Bayesian networks. These algorithm are fundamental to various fields; for example the forward-backward algorithm over conditional random fields forms the basis for state of the art solutions to named entity recognition, part of speech tagging, noun phrase segmentation, and other problems in NLP [@CRF]. We are motivated by the wide applicability of queries over annotated relations; annotated relations may provide a framework for combining classical query processing, linear algebra, and statistical inference in a single data processing system.
We consider a generalization of MPF with multiple aggregation operators. We represent an aggregate-join query as a join $Q$ and an aggregation ordering, which specifies both the ordering and the aggregation of each attribute. Our language directly follows from the work of Abo Khamis, Ngo, and Rudra [@FAQ], who investigated the “Functional Aggregate Query” (FAQ) problem. In addition to MPF, FAQ is a generalization of Chen and Dalmau’s QCQ problem [@QCQChenDalmau], in which the only aggregates are logical quantifiers (AND and OR).
The key technical challenge in both problems is characterizing the permissible aggregations orders to answer the query. Chen and Dalmau give a complete characterization of which variable orders are permissible for QCQ via a procedure. We first give a simple (complete) procedure for our more general class of queries with multiple aggregations, and then we provide a complete characterization of permissible orders.
- [*A Simple Test for Equivalence:* ]{} A query can be thought of as a body $Q$ and a string of attribute-operator pairs $\alpha$. Given a query $Q$ and two orders $\alpha$ and $\beta$, we provide a simple test to determine whether $\alpha$ and $\beta$ are equivalent (i.e., return the same output for any input database). The technical challenge is that different aggregation operators (e.g., $\sum$ and $\max$) cannot freely commute. We show that attribute-operator pairs can commute for only two reasons: $(1)$ their operators commute or $(2)$ their attributes are “independent” in the query, e.g., in the query $\min_{B} \max_{A} \sum_{C} R(A,B),S(B,C)$ the aggregations involving $A$ and $C$ can commute – even though $\max$ and $\sum$ do not commute as operators, the query body renders them [*independent*]{} given $B$. We show that these two conditions are [*complete*]{}, which leads to a simple test for equivalence (Algorithm \[algo:equivalance-test\]).
- [*A Simple Test for GHD and Order Compatibility:* ]{} We say a GHD is compatible with an ordering if we can run standard join algorithms on the GHD while performing aggregations in the order given by the ordering. We show that testing for compatibility amounts to verifying that for any two attributes $A$, $B$, if the topmost GHD node containing $A$ occurs above the topmost node containing $B$, then $A$ occurs before $B$ in the ordering.
This pair of results gives us a simple algorithm that achieves the best known runtime results. Given a query $(Q,\alpha)$, enumerate each order $\beta$ and each GHD $G$, checking if $\alpha$ is equivalent to $\beta$ and $G$ is compatible with $\beta$. If so, record the cost of solving the query using $G$, according to (say) fractional hypertreewidth. Solve the query using the lowest cost $(G,\beta)$ with a standard join algorithm [@FHTW].[^1]
The preceding simple algorithm runs in time exponential in the query size. But finding the optimal GHD even without aggregation is an $\mathsf{NP}$-hard problem, so the brute force optimizer has essentially optimal runtime. It is easy to implement, and a variant is in our prototype database [@EH; @Duncecap1].
The more interesting problem is to characterize the notions of equivalence, mirroring Chen and Dalmau. To that end, we give two new, complete characterizations:
- [*A Complete Characterization of Equivalent Orders:*]{} Given an order $\alpha$ and two attribute-operator pairs $x,y \in \alpha$, we describe a set of constraints of the form [*“in any order, $x$ must appear after $y$.”*]{} Our constraints are sound and complete, i.e., a string $\beta$ satisfies these constraints if and only if it is equivalent to $\alpha$. In contrast, previous approaches have an incomplete characterization, as shown in Example \[example:faq-incompleteness\] in the Appendix.
- [*A Complete Characterization of GHDs compatible with any Equivalent Order.*]{} Given an order $\alpha$ and a query hypergraph $Q$, we call a GHD ‘valid’ if it is compatible with any ordering equivalent to $\alpha$. We give a succinct characterization for all valid GHDs. We then describe a decomposition of the query $(Q, \alpha)$ into a series of *characteristic hypergraphs* (without attached aggregation orderings). GHDs for these hypergraphs can be combined into a valid GHD for the original query. We show that for any “node-monotone” [^2] width function, there is a GHD with optimal width $w$ that can be constructed with this decomposition. [^3] Treewidth, Fractional hypertreewidth, and Submodular width are all node-monotone.
Conceptually, we think the latter result is especially important for tying our work to existing GHD literature; the result reduces our problem to operating on standard GHDs. Pragmatically, we can apply existing GHD results to our characteristic hypergraphs and obtain the following results for free:
- Based on Grohe and Marx [@AGM], we are able to describe our runtime in terms of classical metrics like fractional hypertreewidth. In turn, we can use standard notions to upper bound the runtime like fractional hypertree width, Marx’s submodular width [@Marx:2010:THP:1806689.1806790], or Joglekar’s efficiently computable variant [@2015arXiv150801239J].
- Based on Afrati et al. [@GYM], who bound the communication costs of join processing in terms of a “width” parameter for GHDs, we can develop efficient MapReduce algorithms for solving [<span style="font-variant:small-caps;">Ajar</span>]{}queries.
- Based on Marx’s approximation [@Marx:2010:AFH:1721837.1721845] for GHDs, we can find approximately optimal GHDs for the popular fractional hypertreewidth measure in polynomial time.
We get the above results essentially for free from forging this connection to GHDs. We view this simple link as a strength of our approach.
Finally, we discuss an extension to handle “product aggregations” that allows us to aggregate away an attribute *before* we join the relations containing the attribute when the aggregation operator is the multiplication operator of the semiring. FAQ was the first to observe that this special case can improve certain types of logical queries. This opens up a new space of equivalent orderings and valid GHDs; mirroring the above results, we give a simple test and a complete characterization of the valid GHDs for queries that include this aggregation. As a result, we obtain similar improvements in runtime relative to previous work.
**Outline.** We discuss related work in Section \[sec:rel\]. In Section \[sec:prelim\], we introduce notation and algorithms that are relevant to our work before defining the ${\textsc{Ajar}\xspace}$ problem and discussing its solution, which involves running existing algorithms on a restricted class of GHDs. Section \[sec:equivalent-orderings\] provides a succinct characterization of all orderings that are equivalent to a given ordering. Section \[sec:decomposing\] discusses how to connect our work with recent research on GHDs, explaining how to construct valid optimal query plans and how to further improve and parallelize our results. In Section \[sec:univ-aggregation\], we discuss how to incorporate product aggregations.
Related Work {#sec:rel}
============
**Join Algorithms.** The Yannakakis algorithm, introduced in 1981, guarantees a runtime of $O({\textsc{IN}}+ {\textsc{OUT}})$ for $\alpha$-acyclic join queries [@Yannakakis81]. Modern multiway algorithms can process any join query and have worst-case optimal runtime. In particular, Atserias, Grohe, and Marx [@AGM] derived a tight bound on the worst-case size of a join query given the input size and structure. Ngo et al. [@NPRR] presented the first algorithm to achieve this runtime bound, i.e. the first worst-case optimal algorithm. Soon after, Veldhuizen presented Leapfrog Triejoin, a very simple worst-case optimal algorithm that had been implemented in LogicBlox’s commercial database system [@LFTJ]. Ngo et al. [@NRR] later presented the simplified and unified algorithm GenericJoin (GJ) that captured both of the previous worst-case optimal algorithms.
**GHDs.** First introduced by Gottlob, Leone, and Scarcello [@GHDIntro], hypertree decompositions and the associated hypertree width generalize the concept of tree decompositions [@treedecomp]. Conceptually, the decompositions capture a hypergraph’s cyclicity, allowing them to facilitate the selective use of GJ and Yannakakis in the standard hybrid join algorithm GHDJoin. There are deep connections between variable orderings and GHDs [@FAQ], which we leverage extensively. Grohe and Marx [@FHTW] introduced the idea of fractional hypertree width over GHDs, which bounds the runtime of GHDJoin by ${\widetilde{O}}(IN^w + OUT)$ (${\widetilde{O}}$ hides poly-logarithmic factors) for $w$ defined to be the minimum fractional hypertree width among all GHDs.
**Semirings and Aggregations.** Green, Karvounarakis, and Tannen developed the idea of annotations over a semiring [@Semiring]. Our notation for the annotations is superficially different from theirs, solely for notational convenience. We delve into more detail in Section \[sec:prelim\]. This also has been used as a mechanism to capture aggregation in probabilistic databases [@DBLP:journals/vldb/ReS09].
**MPF.** Aji and McEliece [@Aji] defined the “Marginalize a Product Function” (MPF) problem, which is equivalent to the the space of ${\textsc{Ajar}\xspace}$ queries with only one aggregation operator. They showed that MPF generalizes a wide variety of important algorithms and problems, which also implies that ${\textsc{Ajar}\xspace}$ queries are remarkably general. They also provided a message passing algorithm to solve MPF, which has since been refined [@Kask:2005:UTD:1090725.1090730]. We provide runtime guarantees that improve the current state of the art.
**Aggregate-Join Queries.** There is a standard modification to Yannakakis to handle aggregations [@Yannakakis81], but the classic analysis provides only a $O({\textsc{IN}}\cdot {\textsc{OUT}})$ bound. Bakibayev, Kocisky, Olteanu, and Zavodny study aggregation-join queries in factorized databases [@OZVLDB13], and later Olteanu and Zavodny connected factorized databases and GHDs/GHDJoin [@OZTODS15]. They develop the intuition that if output attributes are above non-output attributes, the $+{\textsc{OUT}}$ runtime is preserved; we use the same intuition to develop and analyze AggroGHDJoin, a variant to GHDJoin for aggregate-join queries.
Abo Khamis, Ngo, and Rudra present the “Functional Aggregate Query” (FAQ) problem [@FAQ], which is equivalent to ${\textsc{Ajar}\xspace}$. The FAQ/${\textsc{Ajar}\xspace}$ problems arose out of discussions between Ngo, Rudra, and Ré at PODS12 about how to extend the worst-case result to queries using aggregation and message passing via Green et al.’s semiring formulation. We originally worked jointly on the problem, but we developed substantially different approaches. As a result, we split our work. We argue the the ${\textsc{Ajar}\xspace}$ approach is simpler, as it yields the best known runtime results in only a few simple statements in Section \[sec:prelim\]. We also describe new complete characterizations as described above. Pragmatically, these completeness results allow us to connect to more easily to existing literature. We have already implemented the algorithm described here in the related database engine EmptyHeaded [@EH]. [^4] This engine has run motif finding, pagerank, and single-source shortest path queries dramatically faster than previous high-level approaches that take datalog-like queries as input.
A primary application of multiple aggregation operators is quantified conjunctive queries (QCQ) and the counting variant, which can be expressed as ${\textsc{Ajar}\xspace}$ queries over the semiring $(\vee, \wedge)$ with aggregations involving both operators. Here, we follow FAQ’s idea to formulate this as a query with product aggregation. Chen and Dalmau [@QCQChenDalmau] completely characterized the space of tractable QCQ by defining a notion of width that relies on variable orderings. Chen and Dalmau’s width definition includes a complete characterization of the permissible variable orderings for a QCQ instance. Their characterization is similar in spirit to the partial ordering we define in Section \[sec:equivalent-orderings\] that characterizes the space of valid GHDs for an ${\textsc{Ajar}\xspace}$ query. However, their results are focused on tractability rather than the optimal runtime exponents; our characterization extends theirs and has improved runtime bounds.
AJAR and A Simple Solution {#sec:prelim}
==========================
We start by describing some background material needed to define the [<span style="font-variant:small-caps;">Ajar</span>]{}problem. After that, we formally define the [<span style="font-variant:small-caps;">Ajar</span>]{}problem and our solution to it.
Background {#subsec:background-1}
----------
We use the classic hypergraph representation for database schema and queries [@ALICE]. A hypergraph ${\mathcal{H}}$ is a pair $({\mathcal{V}}, {\mathcal{E}})$, where ${\mathcal{V}}$ is a non-empty set of [*vertices*]{} and ${\mathcal{E}}\subseteq 2^{\mathcal{V}}$ is a set of [*hyperedges*]{}. Each $A \in {\mathcal{V}}$ is called an [*attribute*]{}. Each attribute has a corresponding [*domain*]{} ${\mathcal{D}}^A$.
- **Data** For each hyperedge $F \in {\mathcal{E}}$, there is a corresponding relation $R_F \subseteq \prod_{A \in F} {\mathcal{D}}^A$; we use the notation ${\mathcal{D}}^F$ to denote the domain of the tuples $\prod_{A
\in F} {\mathcal{D}}^A$.
- **Join Query** Given a set ${\mathcal{E}}$ and a relation $R_F$ for each $F \in {\mathcal{E}}$, let ${\mathcal{V}}= \cup_{F \in {\mathcal{E}}} F$. The join query is written $\Join_{F \in {\mathcal{E}}} R_F$ and is defined as $$\left\lbrace t \in {\mathcal{D}}^{{\mathcal{V}}} \mid \forall F \in {\mathcal{E}}: \pi_{F}(t) \in R_F \right\rbrace$$
We use $n$ to denote the number of attributes $|{\mathcal{V}}|$ and $m$ to denote the number of relations $|{\mathcal{E}}|$. ${\textsc{IN}}$ denotes the sum of sizes of input relations in a query, and ${\textsc{OUT}}$ denotes the output size.
A [*path*]{} from $A \in {\mathcal{V}}_{\mathcal{H}}$ to $B \in {\mathcal{V}}_{\mathcal{H}}$ in a hypergraph ${\mathcal{H}}$ is a sequence of attributes, starting with $A$ and ending with $B$, such that each consecutive pair of attributes in the sequence occur together in a hyperedge. The number of attributes in the sequence is the [*length*]{} of the path.
We now define a GHD of a hypergraph.
[^5] Given a hypergraph ${\mathcal{H}}= ({\mathcal{V}}_{\mathcal{H}}, {\mathcal{E}}_{\mathcal{H}})$, a *generalized hypertree decomposition* is a pair $({\mathcal{T}}, \chi)$ of a tree ${\mathcal{T}}= ({\mathcal{V}}_{\mathcal{T}}, {\mathcal{E}}_{\mathcal{T}})$ and function $\chi: {\mathcal{V}}_{\mathcal{T}}\to 2^{{\mathcal{V}}_{\mathcal{H}}}$ such that
- For each relation $F \in {\mathcal{E}}_{\mathcal{H}}$, there exists a tree node $t \in {\mathcal{V}}_{\mathcal{T}}$ that covers the edge, i.e. $F \subseteq \chi(t)$.
- For each attribute $A \in {\mathcal{V}}_{\mathcal{H}}$, the tree nodes containing $A$, i.e. $\{t \in {\mathcal{V}}_{\mathcal{T}}| A \in \chi(t)\}$, form a connected subtree.
The latter condition is called the “running intersection property”. The $\chi(t)$ sets are referred to as ‘bags’ of the GHD. GHDs are assumed to be ‘rooted’ trees, which imposes a top-down partial order on their nodes. Leveraging this order, for any GHD $({\mathcal{T}}, \chi)$ and attribute $A \in {\mathcal{V}}_{\mathcal{H}}$, we define $TOP_{\mathcal{T}}(A)$ to be the top-most node $v \in {\mathcal{V}}_{\mathcal{T}}$ such that $A \in \chi(v)$.
When each bag of a GHD consists of the attributes of a single relation, the GHD is also called a [*join tree*]{}. Joins over a join tree can be processed using Yannakakis’ algorithm [@Yannakakis81] (pseudo-code in Algorithm \[NoAggroY\]). The runtime of Yannakakis’ algorithm is $O({\textsc{IN}}+ {\textsc{OUT}})$.
GHDs can be interpreted as query plans for joins. Given a GHD, we first join the attributes in each bag using worst case optimal algorithms [@NPRR; @LFTJ] to get one intermediate relation per bag. The intermediate relations can then be joined using Yannakakis’ algorithm. This combined algorithm is called GHDJoin; Algorithm \[NoAggroGHD\] in Appendix \[sec:background\] gives the pseudo-code for GHDJoin.
**Input:** Join tree ${\mathcal{T}}= ({\mathcal{V}}, {\mathcal{E}})$, Relations $R_F$ for each $F \in {\mathcal{V}}$
$P \gets$ parent of $F$ $R_P \gets R_P \ltimes R_F$ $P \gets$ parent of $F$ $R_F \gets R_F \ltimes R_P$ $P \gets$ parent of $F$ $R_P \gets R_P \Join R_F$ $R_R$ for the root $R$
The runtime of GHDJoin can be expressed in terms of the [*fractional hypertree width*]{} of the GHD:
Given a hypergraph ${\mathcal{H}}= ({\mathcal{V}}_{\mathcal{H}}, {\mathcal{E}}_{\mathcal{H}})$ and a GHD $({\mathcal{T}}, \chi)$, the *fractional hypertree width*, denoted $fhw({\mathcal{T}}, {\mathcal{H}})$, is defined to be $\max_{t \in {\mathcal{T}}} \rho^*_t$ in which $\rho^*_t$ is the optimal value of the following linear program defined for each $t \in {\mathcal{V}}_{\mathcal{T}}$: $$\begin{aligned}
\textrm{Minimize } \sum_{F \in {\mathcal{E}}_{\mathcal{H}}} x_F \log_{{\textsc{IN}}}(|R_F|) \text{ such that } \\
\forall A \in \chi(t) : \sum_{F : A \in F} x_F \ge 1, \forall F \in {\mathcal{E}}_{\mathcal{H}}: x_F \ge 0\end{aligned}$$
The fractional hypertree width is just the AGM bound [@AGM] placed on the bags. Thus $IN^{fhw({\mathcal{T}}, {\mathcal{H}})}$ is an upper bound on the sizes of the intermediate relations of GHDJoin. GHDJoin runs in time ${\widetilde{O}}({\textsc{IN}}^{fhw({\mathcal{T}}, {\mathcal{H}})} + {\textsc{OUT}})$ for Join queries.
#### Annotated Relations {#annotated-relations .unnumbered}
To define a general notion of aggregations, we look to relations annotated with semirings [@Semiring].
A *commutative semiring* is a triple $(S, \oplus, \otimes)$ of a set $S$ and operators $\oplus: S \times S \to S$, $\otimes: S \times S \to S$ where there exist $0, 1 \in S$ such that for all $a,b,c \in S$ the following properties hold:
- Identity and Annihilation: $a \oplus 0 = a$, $a \otimes 1 = a$, $0 \otimes a = 0$
- Associativity: $(a \oplus b) \oplus c = a \oplus (b \oplus c)$, $(a \otimes b) \otimes c = a \otimes (b \otimes c)$
- Commutativity: $a \oplus b = b \oplus a$, $a \otimes b = b \otimes a$
- Distributivity: $a \otimes (b \oplus c) = (a \otimes b) \oplus (a \otimes c)$
Suppose we have some domain $\mathbb{K}$ and an operator set $O = \{\oplus^1, \oplus^2,\ldots \oplus^k, \otimes\}$ such that $0$ is the identity for each $\oplus^i \in O$ and $(\mathbb{K}, \oplus^i, \otimes)$ forms a commutative semiring for each $i$. We then define a relation with an annotation from $\mathbb{K}$ for each tuple.
\[annotated\] An *annotated relation* with annotations from $\mathbb{K}$, or a $\mathbb{K}$-relation, over attribute set $F$ is a set $\{(t_1, \lambda_1)$, $(t_2, \lambda_2)$, $\dots$, $(t_N, \lambda_N)\}$ such that for all $1 \le i \le N$, $t_i \in {\mathcal{D}}^F, \lambda_i \in \mathbb{K}$ and for all $1 \le j \le N : i \neq j \rightarrow t_i \neq t_j$.
Green et al. define a $\mathbb{K}$-relation to be a function $R_F: {\mathcal{D}}^F \to \mathbb{K}$ [@Semiring]. Our notion can be viewed as an explicit listing of this function’s support. Note that unlike an explicit listing of the function’s support, our table does allow tuples with $0$ annotations. However, under our definitions of the operators below, an annotation of $0$ is semantically equivalent to a tuple being absent (we discuss this further in Section \[sec:univ-aggregation\]). Note that we can have an annotated relation of the form $R_\emptyset$ of size $1$ containing the empty tuple with some annotation. We now define joins and aggregations over annotated relations.
#### Joins over Annotated Relations {#joins-over-annotated-relations .unnumbered}
Informally, a join over annotated relations is obtained as follows: (i) We perform a regular join on the non-annotated part of the relations. (ii) For each output tuple $t$ of the join, we set its annotation to the product of the annotations of the input tuples used to produce $t$. We define a join $\Join_{F \in {\mathcal{E}}} R_F$ as: $$\Join_{F \in {\mathcal{E}}} R_{F} = \{ (t,\lambda) : \lambda = \prod_{F \in {\mathcal{E}}} \lambda_{F} \text{ in which } (\pi_{F}(t),\lambda_{F}) \in R_{F} \}$$
#### Aggregations over Annotated Relations {#aggregations-over-annotated-relations .unnumbered}
An aggregation over an annotated relation $R_F$ is specified by a pair $(A,\oplus)$ where $A
\in F$, and $\oplus \in O$. The aggregation takes groups of tuples in $R_F$ that share values of all attributes other than $A$, and produces a single tuple corresponding to each group, whose annotation is the $\oplus$-aggregate of the annotations of the tuples in the group. Suppose that $R$ has schema $R(A,B)$ in which $A$ is a single attribute and $B$ is a set of attributes. Then, the result of aggregation $(A,\oplus)$ has only the attributes $B$ and $$\sum_{(A,\oplus)} R_{A,B} = \{ (t_{B},\lambda) : t_{B} \in \pi_{B} R \text{ and } \lambda = \sum^{\oplus}_{ (t,\lambda_t) \in R : \pi_B t = t_B } \lambda_t \}$$
One can define the meaning of aggregate queries in a straightforward way: first compute the join and then perform aggregations. Figure \[fig:operations-ex\] shows some examples of operators on relations. For the remainder of our work, we assume that all relations are $\mathbb{K}$-relations.
The AJAR problem
----------------
Given some global attribute set ${\mathcal{V}}$ and operator set $O$, we define an *aggregation ordering* to be a sequence $\alpha = \alpha_1, \alpha_2, \dots, \alpha_s$ such that for each $1 \le i \le s$, $\alpha_i = (a_i, \oplus_i)$ for some $a_i \in {\mathcal{V}}, \oplus_i \in O$ [^6]. In addition, attributes occur at most once, i.e., $a_j \neq a_k$ for each $1\le j < k \le s$.
Informally, the aggregation ordering is just a sequence of attribute-operator pairs such that each attribute in the sequence occurs at most once. Note that the aggregation ordering specifies the order and manner in which attributes are aggregated. The ordering does not need to contain every attribute; we use the term *output attributes* to denote the attributes not in the ordering.
$V(\alpha)$ represents the set of attributes that appear in $\alpha$, and $V(-\alpha)$ represents $V \backslash V(\alpha)$ (i.e. the output attributes). When $F \subseteq V(\alpha)$, we use $\alpha_F$ to represent a sequence $\beta$ that is equivalent to $\alpha$ restricted to the attributes in $F$, i.e. $V(\beta) = F$, and any $(A, \oplus), (B, \oplus') \in \alpha$ such that $A,B \in F$ must also appear in $\beta$ with their order preserved.
Given some hypergraph ${\mathcal{H}}= ({\mathcal{V}}, {\mathcal{E}})$ and an aggregation ordering $\alpha$, an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$ is a function over instances of ${\mathcal{H}}$ such that $$Q_{{\mathcal{H}}, \alpha} (\{R_F | F \in {\mathcal{E}}\}) = \Sigma_{\alpha_1} \cdots \Sigma_{\alpha_{|\alpha|}} \Join_{F \in {\mathcal{E}}} R_F .$$
For an [<span style="font-variant:small-caps;">Ajar</span>]{}query, we define ${\textsc{OUT}}$ to be the final output size, rather than the output size of the join. There are two technical challenges when it comes to solving an [<span style="font-variant:small-caps;">Ajar</span>]{}query:
- Multiple aggregation orders can give the same output over any database instance, and using some aggregation orders may give faster runtimes than others, e.g. some orders may allow early aggregation. Thus we need to identify which orders are equivalent to the given order and which order leads to the smallest runtime.
- ${\textsc{OUT}}$ for an [<span style="font-variant:small-caps;">Ajar</span>]{}query with $|\alpha| > 0$ is smaller than the output size of the join part of the query. Thus the standard GHDJoin runtime of ${\textsc{IN}}^{fhw} + {\textsc{OUT}}$ is harder to achieve for [<span style="font-variant:small-caps;">Ajar</span>]{}queries. Naively applying a variant of GHDJoin that performs aggregations (Appendix Algorithm \[Aggro\]) to [<span style="font-variant:small-caps;">Ajar</span>]{}leads to a higher runtime of ${\textsc{IN}}^{fhw}\cdot {\textsc{OUT}}$ (see Appendix \[sec:background\]). Thus we need to identify which GHDs can be used for efficient processing of [<span style="font-variant:small-caps;">Ajar</span>]{}queries.
We handle these technical challenges in turn.
Equivalent Orderings {#subsec:simple-solution}
--------------------
Distinct aggregation orders can be equivalent in that they produce the same output on every instance. For example, suppose $\alpha = ((A, +), (B, +))$ and $\beta = ((B,+), (A,+))$, where $A, B$ are two attributes in some ${\mathcal{H}}$. Then two ${\textsc{Ajar}\xspace}$ queries with orderings $\alpha$ and $\beta$ clearly produce the same output for any instance $I$ over ${\mathcal{H}}$. This is because we can obtain $\beta$ from $\alpha$ by switching the positions of two adjacent aggregations [*with the same aggregation operator*]{}. Similarly, if ${\mathcal{H}}$ consists only of relations $\{A,B\}, \{B,C\}$, then the orderings $\alpha = ((A, +),(C, \max))$ and $\beta = ((C, \max),(A, +))$ are equivalent, since you can independently aggregate the two attributes away before joining the two relations on $B$. We now formally define equivalent orderings.
Given a hypergraph ${\mathcal{H}}$, define the equivalence relation between orderings $\equiv_{\mathcal{H}}$ such that $\alpha \equiv_{\mathcal{H}}\beta$ if and only if $Q_{{\mathcal{H}}, \alpha} (I) = Q_{{\mathcal{H}}, \beta}(I)$ for all database instances $I$ over the schema ${\mathcal{H}}$.
We say that two operators $\oplus$, $\oplus'$ are distinct over a domain $\mathbb{K}$ (denoted by $\oplus \neq \oplus'$) if $\exists x, y \in \mathbb{K} : x \oplus y \neq x \oplus' y$. And $\oplus = \oplus'$ means that $\forall x,y \in \mathbb{K}$, $x \oplus y = x \oplus' y$. Of course, distinct operators do not, in general, commute.
We now state a theorem specifying two conditions under which aggregations can commute. We will later show these conditions to be complete.
\[commute\] Suppose we are given a relation $R_F$ such that $A,B \in F$ and two operators $\oplus', \oplus \in O$. Then $$\Sigma_{(A, \oplus)} \Sigma_{(B, \oplus')} R_F = \Sigma_{(B, \oplus')} \Sigma_{(A, \oplus)} R_F$$ if one of the following conditions hold:
- $\oplus = \oplus'$
- There exist relations $R_{F_1}$ and $R_{F_2}$ such that $A \notin F_1$, $B \notin F_2$, and $R_{F_1} \Join R_{F_2} = R_F$.
The first condition follows trivially from the commutativity of our operators. The second condition follows from the fact that we can “push down” aggregations. $$\begin{aligned}
\Sigma_{(A, \oplus)} \Sigma_{(B, \oplus')} R_{F_1} \Join R_{F_2} & = \left(\Sigma_{(B, \oplus')} R_{F_1} \right) \Join \left(\Sigma_{(A, \oplus)} R_{F_2} \right) \\
& = \Sigma_{(B, \oplus')} \Sigma_{(A, \oplus)} R_{F_1} \Join R_{F_2}\end{aligned}$$
**Input:** Query hypergraph ${\mathcal{H}}$, orderings $\alpha$, $\beta$.\
**Output:** True if $\alpha \equiv_{{\mathcal{H}}} \beta$, False otherwise.
True Remove $V(-\alpha)$ from ${\mathcal{H}}$, then divide ${\mathcal{H}}$ into connected components $C_1,\ldots C_m$. $\land_{i} \text{TestEquivalence}({\mathcal{H}}, \alpha_{C_i}, \beta_{C_i})$ Choose $j$ such that $\beta_j = \alpha_1$. Let $\beta_j = (b_j, \oplus'_j)$. False Let $\beta'$ be $\beta$ with $\beta_j$ removed. Let $\alpha'$ be $\alpha$ with $\alpha_1$ removed. $\text{TestEquivalence}({\mathcal{H}}, \alpha', \beta')$
These two conditions give us a simple procedure for testing when an ordering $\beta$ is equivalent to the given $\alpha$. Algorithm \[algo:equivalance-test\] gives the procedure’s pseudo-code. To avoid triviality, we assume $\alpha$ and $\beta$ have the same set of attributes and assign the same operator to the same attributes. First we return true if both $\alpha$ and $\beta$ are empty. Then we check if $\alpha$ can be shown to be equivalent to $\beta$ using the conditions from Theorem \[commute\]. This procedure is both sound and complete:
\[lemma:equivalence-test-sound-complete\] Algorithm \[algo:equivalance-test\] returns True iff $\alpha \equiv_{\mathcal{H}}\beta$.
We omit this lemma’s proof because it is very similar to and implied by the proofs required in Section \[sec:equivalent-orderings\].
To answer [<span style="font-variant:small-caps;">Ajar</span>]{}queries, we need one more component in addition to Algorithm \[algo:equivalance-test\]; namely AggroGHDJoin, a straightforward variant of GHDJoin that performs aggregations (Algorithm \[Aggro\] in Appendix \[sec:background\]). The first step of AggroGHDJoin is similar to that of GHDJoin, namely performing joins within each bag of the GHD to get intermediate relations. We need to do some extra work to ensure that each annotation is multiplied only once, since a relation may be joined in multiple bags. After that, instead of calling Yannakakis’ algorithm on the intermediate relations, AggroGHDJoin calls AggroYannakakis (Algorithm \[AggroY\] in Appendix \[sec:background\]), a well-known variant of Yannakakis that performs aggregations. AggroYannakakis initially performs semijoins like Yannakakis (lines $1$-$8$ in Algorithm \[NoAggroY\]). But in the bottom-up join phase (line $11$), AggroYannakakis aggregates out all attributes that have $F$ as their $TOP$ node, before joining $R_F$ with $R_P$.
Armed with Algorithm \[algo:equivalance-test\] and AggroGHDJoin, we have a simple way to answer an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$. We search through all orders, running Procedure $1$ to check for equivalence with $\alpha$. For each order $\beta$ such that $\beta \equiv_{\mathcal{H}}\alpha$, we search all through GHDs and check if they are compatible with $\beta$. A GHD ${\mathcal{T}}$ is defined to be [*compatible*]{} with an ordering $\beta$ if, for all attribute pairs $A,B$, $TOP_{\mathcal{T}}(A)$ being an ancestor of $TOP_{\mathcal{T}}(B)$ implies that either $A$ is an output variable or $A$ occurs before $B$ in $\beta$ (note this precludes $B$ from being an output variable). We can run AggroGHDJoin on any compatible GHD to answer the [<span style="font-variant:small-caps;">Ajar</span>]{}query. The runtime of AggroGHDJoin on a compatible GHD $({\mathcal{T}}, \chi)$ is given by ${\widetilde{O}}({\textsc{IN}}^{fhw({\mathcal{T}},{\mathcal{H}})} + {\textsc{OUT}})$. We choose the compatible GHD that has the smallest $fhw$, and use it to answer the query. The theorem below states our runtime:
\[thm:ajar-runtime\] Given a [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}},\alpha}$, let $w^*$ denote the smallest fhw for a GHD compatible with an ordering $\equiv_{\mathcal{H}}\alpha$; the runtime of our approach is ${\widetilde{O}}({\textsc{IN}}^{w^*} + {\textsc{OUT}})$.
#### Comparison to Prior Work {#comparison-to-prior-work .unnumbered}
Work by Olteanu and Zavodny [@OZVLDB13; @OZTODS15] focuses on a special case of [<span style="font-variant:small-caps;">Ajar</span>]{}queries, having a single aggregation operator. For these queries, they have a similar algorithm that iterates over GHDs to find the best compatible one. Their algorithm achieves the same runtime as ours, but cannot handle queries with more than one type of aggregation operator. The FAQ paper uses an algorithm called InsideOut to answer general [<span style="font-variant:small-caps;">Ajar</span>]{}queries. The running time of InsideOut equals ${\widetilde{O}}({\textsc{IN}}^{faqw})$ where faqw (FAQ-width) is a new notion of width defined by the FAQ authors [@FAQ Section 9.1]. Our algorithm has runtime that is no worse than InsideOut ($w^* \leq faqw, {\textsc{OUT}}\leq {\textsc{IN}}^{faqw}$), and can be much better when output attributes are present.
\[thm:ajar-vs-faq-runtime\] For any [<span style="font-variant:small-caps;">Ajar</span>]{}query, $w^* \leq faqw$ and ${\textsc{OUT}}\le O({\textsc{IN}}^{faqw})$.
This theorem is proved in Appendix \[subsec:faq-comparison-proof\]. Notice that the InsideOut runtime is not [*output-sensitive*]{}, i.e. it does not have a $+\text{ }{\textsc{OUT}}$ term. As a result the runtime can be very high when the output is small relative to the number of output attributes; this is demonstrated by Example \[ex:faqoutput\] in the appendix. FAQ does have a high-level discussion of approaches to make InsideOut output-sensitive [@FAQ Section 10.2]; indeed, simply using GHDJoin instead of their bespoke algorithm can achieve output-sensitive bounds, which we discuss in Appendix \[sec:app-related\].
#### Discussion {#discussion .unnumbered}
We presented a remarkably simple procedure for solving [<span style="font-variant:small-caps;">Ajar</span>]{}queries. The procedure involves a brute force search over different orderings and GHDs, but this is usually unavoidable as finding the best ordering and GHD is NP-Hard. Deciding if an ordering is equivalent to the given ordering is enabled by Algorithm \[algo:equivalance-test\], which takes time polynomial in the number of attributes. Determining if a GHD is compatible with an ordering is straightforward as well. Once the best GHD is found, we use well known, standard algorithms like AggroGHDJoin to answer the query efficiently. The resulting runtime exponents are smaller than those of previous work. The simplicity of the algorithm makes it easy to implement; we have already implemented a special case of a single additive operator $\oplus$ in our engine [@EH].
The equivalence/compatibility tests raise the technically interesting question of finding succinct characterizations of:
- All orderings equivalent to any given $\alpha$.
- All GHDs that are compatible with at least one of the equivalent orderings.
We answer the first question in Section \[sec:equivalent-orderings\] by providing a simple characterization of all equivalent orderings, and the second question in Section \[sec:decomposing\] by defining ‘valid’ GHDs and characterizing their structure in relation to unrestricted GHDs.
Characterizing Equivalent Orderings {#sec:equivalent-orderings}
===================================
We described a procedure for determining when two orderings are equivalent. The equivalence relation $\equiv_{\mathcal{H}}$ defines equivalence classes among the orderings, but these classes may be exponential in size; we find a more succinct characterization that lets us enumerate all equivalent orderings. Chen and Dalmau [@QCQChenDalmau] obtained a similar order-equivalence characterization for a special case of the [<span style="font-variant:small-caps;">Ajar</span>]{}problem, namely for aggregations “and” and “or”. The characterization was based on a procedure that generated all equivalent orderings. We improve on this result by providing a simple and succinct characterization of the equivalence class of an aggregation ordering with any number operators.
To that end, we develop an enumeration of the constraints that are sufficient and necessary for an ordering to be in the equivalence class of $\alpha$. The constraints are of the form “$A$ must always occur before $B$”:
Given an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$, define a constraint ${\textsf{PREC}}\subseteq {\mathcal{V}}\times {\mathcal{V}}$ such that $(A,B) \in {\textsf{PREC}}$ if and only if $A$ precedes $B$ in all orderings that are equivalent to $\alpha$.
We say ${\textsf{PREC}}(A,B)$ is true if and only if $(A,B) \in {\textsf{PREC}}$.
Trivially, the number of pairs in ${\textsf{PREC}}$ is less than $n^2$. We note that we can use ${\textsf{PREC}}$ to define a (strict) partial ordering on the attributes; the constraints are clearly antireflexive, antisymmetric, and transitive. We use $<_{{\mathcal{H}}, \alpha}$ to denote this partial order. Given an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$, $<_{{\mathcal{H}}, \alpha}$ is a partial order of attribute-operator pairs such that for any $(A, \oplus), (B, \oplus') \in \alpha$, $(A, \oplus) <_{{\mathcal{H}}, \alpha} (B, \oplus')$ if ${\textsf{PREC}}(A,B)$ (see Definition \[def:partial-order\] for the exact definition). The partial order $<_{{\mathcal{H}}, \alpha}$ is easier to use for proofs; we use the partial order to show the soundness and completeness of these constraints.
\[thm:dag-soundness-completeness\] Suppose we are given a hypergraph ${\mathcal{H}}= ({\mathcal{V}}, {\mathcal{E}})$ and aggregation orderings $\alpha, \beta$. Then $\alpha \equiv_{\mathcal{H}}\beta$ if and only if $\beta$ is a linear extension of $<_{{\mathcal{H}},\alpha}$.
We first describe a procedure to compute the precedence relation ${\textsf{PREC}}$. After that, we reason about its completeness.
#### Computing ${\textsf{PREC}}$ {#computing-textsfprec .unnumbered}
To assist in building ${\textsf{PREC}}$, we define a constraint of the form ‘$A$ and $B$ cannot commute’’:
Given an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$, define a constraint ${\textsf{DNC}}\subseteq {\mathcal{V}}\times {\mathcal{V}}$ such that $(A,B) \in {\textsf{DNC}}$ if and only if $A$ and $B$ are in the same order in any $\beta$ such that $\beta \equiv_{{\cal H}} \alpha$.
Once again, we say ${\textsf{DNC}}(A,B)$ is true if and only if $(A,B) \in {\textsf{DNC}}$. We prefer to work with ${\textsf{DNC}}$ because we have already discussed when aggregations can commute in Theorem \[commute\]; the conditions of that theorem specify when ${\textsf{DNC}}$ is $FALSE$. However, we can immediately derive a simple relationship between ${\textsf{PREC}}$ and ${\textsf{DNC}}$:
\[lemma:PREC&DNC\] Given an [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}}, \alpha}$, for any $A, B \in {\mathcal{V}}$, ${\textsf{PREC}}(A,B)$ iff ${\textsf{DNC}}(A,B)$ and $A$ precedes $B$ in $\alpha$.
We now develop conditions when ${\textsf{DNC}}$ is true. Recall that Theorem \[commute\] states that two aggregations can commute if $(1)$ they have the same operator or $(2)$ if they can be separated in the join query; the simplest structure that violates both of these conditions is an edge that contains two attributes with differing aggregating operators.
\[lemma:DNCbasecase\] Given an [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}}, \alpha}$, suppose $(A, \oplus)$, $(B, \oplus')$ $\in \alpha$. If $\oplus \neq \oplus'$ and there exists an edge $E \in {\mathcal{E}}$ such that $A, B \in E$, then ${\textsf{DNC}}(A,B)$.
Lemma \[lemma:DNCbasecase\] serves as a base case, but we want to extend the violation of Theorem \[commute\]’s conditions beyond single edges to paths. To do so, consider the following examples of how our commuting conditions interact with paths of length two.
Consider the query $$\sum_A \max_B \max_C R(A,B) \Join S(B,C) \text{ hence } \alpha = (A, B, C).$$ No two attributes can be separated, which implies ${\textsf{DNC}}(A,B)$ and ${\textsf{DNC}}(A,C)$. Lemma \[lemma:DNCbasecase\] gives us the former constraint, but not the latter one. This example indicates that it may be possible to extend a constraint ${\textsf{DNC}}(A,B)$ along an edge $\{B,C\}$. On the other hand, consider the query $$\max_B \sum_A \max_C R(A,B) \Join S(B,C) \text{ so } \alpha = (B, A, C).$$ Note that $A$ and $C$ can be separated, which implies that only ${\textsf{DNC}}(A,B)$ holds. Note that, as before, Lemma \[lemma:DNCbasecase\] gives us this constraint. This example suggests that we cannot extend every ${\textsf{DNC}}(A,B)$ constraint along an additional edge.
The key difference between the two examples is the [*relative order*]{} of $A$ and $B$ in $\alpha$, which suggests that we can only extend ${\textsf{DNC}}(A,B)$ along an edge if $A$ precedes $B$ in $\alpha$, i.e. if ${\textsf{PREC}}(A,B)$.
\[lemma:DNCextension\] Given an [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}}, \alpha}$, suppose $(A, \oplus)$, $(B, \oplus')$ $\in \alpha$. If $\oplus \neq \oplus'$ and $\exists C \in {\mathcal{V}}, E \in {\mathcal{E}}: {\textsf{PREC}}(A,C)$ and $B,C \in E$, then ${\textsf{DNC}}(A,B)$.
${\textsf{PREC}}$ is transitive, which implies:
\[lemma:DNCtransitive\] Given an [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}}, \alpha}$, suppose $(A, \oplus), (B, \oplus') \in \alpha$. If $\exists C: {\textsf{PREC}}(A,C) \text{ and } {\textsf{PREC}}(C,B)$, then ${\textsf{DNC}}(A,B)$.
The above transitivity condition interacts with the condition from Lemma \[lemma:DNCextension\] in interesting ways.
Consider the query with $\alpha = (A, B, C, D)$, $$\sum_A \max_B \max_C \sum_D R(A,B) \Join S(B,D) \Join T(C,D).$$ No attributes can be separated, which implies ${\textsf{DNC}}(A,B)$, ${\textsf{DNC}}(A,C)$, ${\textsf{DNC}}(B,D)$, and ${\textsf{DNC}}(C,D)$. Transitivity gives ${\textsf{DNC}}(A,D)$ as well. Now let us derive these constraints using Lemmas \[lemma:DNCbasecase\], \[lemma:DNCextension\], and \[lemma:DNCtransitive\]. Lemma \[lemma:DNCbasecase\] gives us ${\textsf{DNC}}(A,C)$, ${\textsf{DNC}}(B,D)$, and ${\textsf{DNC}}(C,D)$. Note that at this point, Lemma \[lemma:DNCextension\] gives us no more constraints. Only after the transitivity of Lemma \[lemma:DNCtransitive\] adds the constraint ${\textsf{DNC}}(A,D)$ can Lemma \[lemma:DNCextension\] add the constraint ${\textsf{DNC}}(A,B)$, completing the set of constraints.
It turns out that these three relatively simple lemmas are the sufficient and necessary constraints on the equivalence classes of orderings; no other conditions are necessary to complete the proofs the soundness and completeness of $<_{{\mathcal{H}}, \alpha}$.
We note that our current specifications of ${\textsf{PREC}}$ and ${\textsf{DNC}}$ are mutually recursive. The ${\textsf{PREC}}$ and ${\textsf{DNC}}$ sets build up in rounds; Lemma \[lemma:DNCbasecase\] provides their initial values, and Lemmas \[lemma:PREC&DNC\], \[lemma:DNCextension\], and \[lemma:DNCtransitive\] iteratively build up the sets further. We keep applying these lemmas until the sets reach a fixed point. This takes at most $2|\alpha|^2$ iterations, as we must add at least one additional attribute pair per iteration, and there can be only $|\alpha|^2$ pairs of attributes in each set. Thus the overall runtime of computing these constraints is polynomial in the number of attributes. We detail this process in Appendix \[sec:app-equiv\].
For convenience of notation, we make one modification to the definition of the partial order $<_{{\mathcal{H}}, \alpha}$. When $A$ is an output attribute and $B$ is not, we define $A <_{{\mathcal{H}}, \alpha} B$ to be true. So we can formally state the definition as:
\[def:partial-order\] Given a ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}},\alpha}$, we define $A <_{{\mathcal{H}},\alpha} B$ to be true if either (i) $A$ is an output attribute and $B$ is not, or (ii) ${\textsf{PREC}}(A,B)$ is true.
#### Soundness and Completeness of $<_{{\mathcal{H}},\alpha}$ {#soundness-and-completeness-of-_mathcalhalpha .unnumbered}
To give an intuition on how we prove the soundness and completeness of $<_{{\mathcal{H}}, \alpha}$, we now state two key lemmas (with proofs in Appendix \[sec:app-equiv\]) illustrating properties of $<_{{\mathcal{H}}, \alpha}$.
\[Pathing\] Suppose we are given a hypergraph ${\mathcal{H}}= ({\mathcal{V}}, {\mathcal{E}})$ and an aggregation ordering $\alpha$. Suppose $(A, \oplus), (B, \oplus') \in \alpha$ for differing operators $\oplus \neq \oplus'$. Then, for any path $P$ in ${\mathcal{H}}$ between $A$ and $B$, there must exist some attribute in the path $C \in P$ such that $C <_{{\mathcal{H}}, \alpha} A$ or $C <_{{\mathcal{H}}, \alpha} B$.
Lemma \[Pathing\] intuitively states that incomparable attributes with different operators must be separated in ${\mathcal{H}}$ by their common predecessors in $<_{{\mathcal{H}}, \alpha}$.
\[Pathing2\] Given a hypergraph ${\mathcal{H}}= ({\mathcal{V}}, {\mathcal{E}})$ and an aggregation ordering $\alpha$, suppose we have two attributes $A, B \in V(\alpha)$ such that $A <_{{\mathcal{H}}, \alpha} B$. Then there must exist a path $P$ from $A$ to $B$ such that for every $C \in P, C \neq A$ we have $A <_{{\mathcal{H}}, \alpha} C$.
Given these two lemmas, the proof of Theorem \[thm:dag-soundness-completeness\] is straightforward. Lemma \[Pathing\] implies that, given an attribute ordering $\beta$ that is a linear extension of $<_{{\mathcal{H}}, \alpha}$, each inversion of attribute-operator pairs must either have equal operators or have attributes that can be separated, allowing us to repeatedly use Theorem \[commute\] to transform $\beta$ into $\alpha$. Lemma \[Pathing2\] implies that, given an attribute ordering $\beta$ that is not a linear extension of $<_{{\mathcal{H}}, \alpha}$, we can construct a counterexample.
#### Discussion {#discussion-1 .unnumbered}
We obtained a sound and complete characterization of all orderings equivalent to any given ordering. This result extends the work of Chen and Dalmau [@QCQChenDalmau], who had characterized equivalent orderings for queries with logical “and” and “or” operators. Our characterization is simple, consisting of a partial order whose linear extensions are precisely the equivalent orderings. FAQ [@FAQ]’s method for identifying equivalent orderings is sound but not complete. That is, there exist equivalent orderings that the FAQ method does not identify as being equivalent (Appendix Example \[example:faq-incompleteness\]). In contrast, our characterization is guaranteed to cover all valid orderings. This completeness property lets us create a decomposition that is guaranteed to preserve all [*node-monotone*]{} widths (see Definition \[def:node-monotone\]). This in turn lets us get tighter guarantees on our runtime exponent, using the notion of submodular width (Section \[subsec:dbp-width\]).
Decomposing Valid GHDs {#sec:decomposing}
======================
We express our [<span style="font-variant:small-caps;">Ajar</span>]{}algorithm directly in terms of GHDs, rather than in terms of aggregation orderings. As such, our goal is the characterization of GHDs that are compatible with at least one equivalent ordering, i.e. the GHDs that can be used to answer an [<span style="font-variant:small-caps;">Ajar</span>]{}query. We call a GHD [*valid*]{} if it is compatible with at least one equivalent ordering. We first give a simple characterization of valid GHDs. Then we demonstrate a way to reduce the problem of finding a minimum-width valid GHD to multiple subproblems on unconstrained GHDs (Section \[subsec:decomposing\]). This decomposition of the problem gets us three things:
- We can speed up our brute force search for an optimal valid GHD. We can also find approximately optimal valid GHDs in polynomial time using Marx’s GHD approximation algorithm [@Marx:2010:AFH:1721837.1721845] (Section \[subsec:optimal-valid\]).
- We can apply existing MapReduce join algorithms that utilize GHDs [@GYM], obtaining efficient parallel algorithms for solving [<span style="font-variant:small-caps;">Ajar</span>]{}queries (Section \[subsec:GYM\]).
- We can apply improved join algorithms [@Marx:2010:THP:1806689.1806790; @2015arXiv150801239J] to further reduce our runtime exponent (Section \[subsec:dbp-width\]).
Valid and Decomposable GHDs {#subsec:decomposing}
---------------------------
We can easily characterize valid GHDs by combining the definition of compatible GHDs with Theorem \[thm:dag-soundness-completeness\].
\[thm:validcorrect\] For a [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}},\alpha}$, a GHD $({\mathcal{T}}, \chi)$ is *valid* if and only if for every pair of attributes $A,B$ such that $TOP_{\mathcal{T}}(A)$ is an ancestor of $TOP_{\mathcal{T}}(B)$, $B \not <_{{\mathcal{H}}, \alpha} A$.
Theorem \[thm:validcorrect\] gives us a criterion specifying which GHDs can act as query plans. We now consider the problem of finding a minimum width valid GHD for any ${\textsc{Ajar}\xspace}$ query. We call a GHD [*optimal*]{} if it has the minimum width possible for valid GHDs. We show how to reduce the problem of finding an optimal valid GHD into smaller problems of finding ordinary optimal GHDs. This unlocks a trove of powerful GHD results and makes them applicable to our problem.
Given an [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}}, \alpha}$, suppose we have a subset of the nodes $V \subseteq {\mathcal{V}}$. Define ${\mathcal{E}}_V$ to be $\{E \in {\mathcal{E}}| E \cap V \neq \emptyset\}$, i.e. the set of edges that intersect with $V$. As before, $\alpha_V$ denotes the aggregation ordering restricted to the nodes in $V$. Additionally define $V^O$ to be $\{ v \in V | \forall w \in V, w \not <_{{\mathcal{H}}, \alpha} v\}$, i.e. the nodes in $V$ that have no predecessors in $V$ according to the partial ordering $<_{{\mathcal{H}}, \alpha}$. Finally, note that $\alpha_{V\backslash V^O}$ is then $\alpha_V$ with all the nodes in $V^O$ removed (note that this makes the nodes in $V^O$ output attributes).
\[def:decomp\] Given an [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}}, \alpha}$, we say a GHD $({\mathcal{T}}, \chi)$ is *decomposable* if:
- There exists a rooted subtree ${\mathcal{T}}_0$ of ${\mathcal{T}}$ such that $\chi({\mathcal{T}}_0) = {\mathcal{V}}(-\alpha)$ (i.e. output attributes).
- For each connected component $C$ of ${\mathcal{H}}\backslash V_{-\alpha}$, there is exactly one subtree ${\mathcal{T}}_{C} \in {\mathcal{T}}\backslash {\mathcal{T}}_0$ such that ${\mathcal{T}}_C$ is a decomposable GHD of $Q_{(\cup_{E \in {\mathcal{E}}_C} E, {\mathcal{E}}_C), \alpha_{C \backslash C^O}}$.
We start by connecting this idea of decomposable GHDs to valid GHDs. We only give proof sketches here; see appendix \[sec:app-decomp\] for the full proofs.
\[thm:decompisvalid\] Every decomposable GHD is valid.
(Sketch) Suppose the ${\textsc{Ajar}\xspace}$ query is $Q_{{\mathcal{H}}, \alpha}$. We need to show for any $A,B$ such that $TOP_{\mathcal{T}}(A)$ is an ancestor of $TOP_{\mathcal{T}}(B)$, $A \not <_{{\mathcal{H}}, \alpha} B$. We use induction on $|\alpha|$. If $|\alpha|=0$, all GHDs are valid and decomposable. For $|\alpha| > 0$, ${\mathcal{T}}_0$ ensures that the output attributes are above non-output attributes. If $A$ and $B$ are non-output attributes and $TOP_{\mathcal{T}}(A)$ is an ancestor of $TOP_{\mathcal{T}}(B)$, then both are in some ${\mathcal{T}}_C$. By the inductive hypothesis, ${\mathcal{T}}_C$ is valid with respect to $Q_{(\cup_{E \in {\mathcal{E}}_C} E, {\mathcal{E}}_C), \alpha_{C \backslash C^O}}$. By inspecting the partial order created by this subgraph, we conclude that $A \not <_{{\mathcal{H}}, \alpha} B$ as desired.
Every valid GHD may not be decomposable. However, every valid GHD can be transformed into a corresponding decomposable GHD using some simple transformations. Each bag of the resulting decomposable GHD is a subset of one of the bags of the original GHD. Thus the fhw of the decompsable GHD is at most the fhw of the original valid GHD. In fact, we can make a more general claim, using a notion of node-monotone functions, defined next.
\[def:node-monotone\] Given a hypergraph ${\mathcal{H}}= ({\mathcal{V}}_{\mathcal{H}}, {\mathcal{E}}_{\mathcal{H}})$, we define a function to be [*node-monotone*]{} if it is a function $\gamma: 2^{{\mathcal{V}}_{\mathcal{H}}}
\rightarrow \mathbb{R}$ such that $\forall \text{ } A \subseteq B
\subseteq {\mathcal{V}}_{\mathcal{H}}: \gamma(A) \leq \gamma(B)$. Given any node-monotone function $\gamma$, we define the $\gamma$-width of a GHD $({\mathcal{T}}, \chi)$ over ${\mathcal{H}}$ as $\max_{v \in {\mathcal{V}}_{\mathcal{T}}} \gamma(\chi(v))$.
Many standard notions of widths can be expressed as $\gamma$-widths for a suitably chosen $\gamma$. Specifically:
\[prop:node-monotone-widths\] Suppose we are given a hypergraph ${\mathcal{H}}= ({\mathcal{V}}_{\mathcal{H}}, {\mathcal{E}}_{\mathcal{H}})$ and database instance $I$ on ${\mathcal{H}}$. Then for each of following notions of width: (i) Treewidth (ii) Generalized Hypertree Width (iii) Fractional Hypertree Width (iv) Submodular Width, there exists a node-monotone function $\gamma$ such that $\gamma$-width equals the given notion of width.
As a simple example, tree-width can be expressed as $\gamma$-width for $\gamma(A) = |A|-1$. We can now relate valid and decomposable GHDs with respect to their $\gamma$-widths.
\[thm:decompwidth\] For every valid GHD $({\mathcal{T}}, \chi)$, there exists a decomposable GHD $({\mathcal{T}}', \chi')$ such that for all node-monotone functions $\gamma$, the $\gamma$-width of $({\mathcal{T}}', \chi')$ is no larger than the $\gamma$-width of $({\mathcal{T}}, \chi)$.
(Sketch) Suppose the ${\textsc{Ajar}\xspace}$ query is $Q_{{\mathcal{H}}, \alpha}$. We transform the given GHD $({\mathcal{T}}, \chi)$ into $({\mathcal{T}}', \chi')$ such that for each $v' \in {\mathcal{V}}_{\mathcal{T}}'$, there exists a $v \in {\mathcal{V}}_{\mathcal{T}}$ such that $\chi'(v') \subseteq \chi(v)$. The result then follows from the node-monotonicity of $\gamma$ and the definition of $\gamma$-width. Any transformation of a GHD that ensures that all new bags are subsets of old bags, is called width-preserving.
We then transform the GHD $({\mathcal{T}}, \chi)$ to satisfy the following properties (using width-preserving transformations):
- Every $t \in {\mathcal{T}}$ is $TOP_{\mathcal{T}}(A)$ for exactly one attribute $A$.
- For any node $t \in {\mathcal{T}}$ and the subtree ${\mathcal{T}}_t$ rooted at $t$, the attributes $\{v \in {\mathcal{V}}| TOP_{\mathcal{T}}(v) \in {\mathcal{T}}_t\}$ form a connected subgraph of ${\mathcal{H}}$.
We can show, by induction, any valid GHD that satisfies these two properties is decomposable. Intuitively, the first transformation ensures the subtree ${\mathcal{T}}_0$ exists as desired. The second transformation ensures that each of the ${\mathcal{T}}_C$’s exists and satisfies the requisite properties.
This theorem lets us restrict our search to the smaller space of decomposable GHDs (instead of all valid GHDs) when looking for the optimal valid GHD. Moreover, the space of decomposable GHDs is simpler; it can be factored into smaller spaces of unconstrained GHDs, as we show next. We present the definition of *characteristic hypergraphs*, which are intuitively the set of hypergraphs that specify the factors, i.e. the unconstrained GHDs.
Our goal is two-fold: $(1)$ to be able to split a decomposable GHD into component GHDs of the characteristic hypergraphs and $(2)$ to be able to take arbitrary GHDs of the characteristic hypergraphs and connect them to create a decomposable GHD of the original [<span style="font-variant:small-caps;">Ajar</span>]{}problem. The definition of decomposable GHDs decomposes a GHD into a series of sub-trees ${\mathcal{T}}_0, \dots, {\mathcal{T}}_k$. The definition specifies that the subtrees ${\mathcal{T}}_1, \dots, {\mathcal{T}}_k$ must be decomposable GHDs of (smaller) [<span style="font-variant:small-caps;">Ajar</span>]{}problems. Additionally, it is simple to show ${\mathcal{T}}_0$ is a GHD of the hypergraph $(V(-\alpha), \{E \in {\mathcal{E}}| E \subseteq V(-\alpha)\})$. If we apply this decomposition recursively to the subtrees ${\mathcal{T}}_1, \dots, {\mathcal{T}}_k$, we can divide any decomposable GHD into a series of (unrestricted) GHDs of particular hypergraphs. This provides the basis of our definition of the characteristic hypergraphs; we define a hypergraph ${\mathcal{H}}_0$ that specifies the hypergraph corresponding to ${\mathcal{T}}_0$ and then recurse on the smaller [<span style="font-variant:small-caps;">Ajar</span>]{}queries specified in Appendix Definition \[def:decomp2\].
However, if we are given arbitrary GHDs of the hypergraphs as defined thus far, we may not be able to stitch them together while preserving the running intersection property of GHDs. To ensure this stitching is possible, we need the characteristic hypergraphs to contain additional edges that we can use to guarantee the running intersection property. Intuitively the edges we add will be the intersections of the adjacent subtrees in our decomposition; for example, for any connected component $C$ of ${\mathcal{H}}\backslash V(-\alpha)$, ${\mathcal{T}}_0$ and ${\mathcal{T}}_C$ are adjacent, and we will add the edge $\chi({\mathcal{T}}_0) \cap \chi({\mathcal{T}}_C)$ to the corresponding hypergraphs. We can use these ‘intersection edges’ to connect particular nodes in the adjacent subtrees.
Given an ${\textsc{Ajar}\xspace}$ problem $Q_{{\mathcal{H}}, \alpha}$, suppose $C_1, \dots, C_k$ are the connected components of ${\mathcal{H}}\setminus {\mathcal{V}}_{-\alpha}$. Define a function $H({\mathcal{H}}, \alpha)$ that maps ${\textsc{Ajar}\xspace}$ queries to a set of hypergraphs as follows:
- $C_i^+ = \bigcup_{E \in {\mathcal{E}}_{C_i}} E$ for all $1 \le i \le k$
- ${\mathcal{H}}_0 = ({\mathcal{V}}_{-\alpha}, \{F \in {\mathcal{E}}| F \subseteq {\mathcal{V}}_{-\alpha}\} \cup \{{\mathcal{V}}_{-\alpha} \cap C_i^+ | 1 \le i \le k\})$
- ${\mathcal{H}}_i^+ = (C_i^+, {\mathcal{E}}_{C_i} \cup \{{\mathcal{V}}_{-\alpha} \cap C_i^+\})$
- $H({\mathcal{H}}, \alpha) = \{{\mathcal{H}}_0\} \cup \bigcup_{1 \le i \le k} H({\mathcal{H}}_i^+, \alpha_{C_i \backslash C_i^O})$
The hypergraphs in the set $H({\mathcal{H}}, \alpha)$ are defined to be the *characteristic hypergraphs*.
Note that the definition of characteristic hypergraphs depends only on $({\mathcal{H}}, \alpha)$, and not on a specific GHD or the instance. Now we state a key result that lets us reduce the problem of searching for an optimal [*valid*]{} GHD over ${\mathcal{H}}$ to that of searching for (not necessarily valid) optimal GHDs over characteristic hypergraphs. Each decomposable GHD corresponds to a GHD over each characteristic hypergraph; conversely, a combination of GHDs for characteristic hypergraphs gives us a decomposable GHD for ${\mathcal{H}}$. Formally:
\[thm:decomp\] For an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$, suppose ${\mathcal{H}}_0, \dots, {\mathcal{H}}_k$ are the characteristic hypergraphs $H({\mathcal{H}}, \alpha)$. Then GHDs $G_0, G_1, \dots, G_k$ of ${\mathcal{H}}_0, \dots, {\mathcal{H}}_k$ can be connected to form a decomposable GHD $G$ for $Q_{{\mathcal{H}}, \alpha}$. Conversely, any decomposable GHD $G$ of $Q_{{\mathcal{H}}, \alpha}$ can be partitioned into GHDs $G_0, G_1, \dots, G_k$ of the characteristic hypergraphs ${\mathcal{H}}_0, \dots, {\mathcal{H}}_k$. Moreover, in both of these cases, $\gamma\text{-width}(G) = \max_{i} \gamma\text{-width}(G_i)$.
The proof is provided in the appendix, but it is a straightforward application of definitions.
\[cor:optimal-construct\] Given an optimal GHD for each characteristic hypergraph of an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$, we can construct an optimal valid GHD. The width of the optimal valid GHD equals the maximum optimal-GHD-width over its characteristic hypergraphs.
This reduces the problem of finding the optimal valid GHD to smaller problems of finding optimal GHDs. We first present the decomposition in the FAQ [@FAQ] paper. Then we present several applications of our decomposition, and compare them to their FAQ analogues.
#### FAQ’s Decomposition {#faqs-decomposition .unnumbered}
The FAQ paper uses a decomposition of the problem that is not width-preserving. They remove the set of output attributes $V(-\alpha)$ and decompose the rest of the hypergraph into smaller hypergraphs. They construct a regular Variable-Ordering/GHD for each hypergraph. Then they add all output attributes $V(-\alpha)$ into each bag of each of the GHDs, and then stitch the GHDs together. This [*output addition*]{} to the bags of the GHDs leads to a potentially $2\times$ increase in width compared to our method which stitches the GHDs together without changing their width. As a result, FAQ’s decomposition incurs higher runtime costs in each application of the decomposition, as we see in the next three subsections.
\[example:faq-2x\] Consider a query with output attribute $A$ $$\sum_{B,+}\sum_{C,+} (R(A,B) \Join S(B,C)).$$ The optimal valid GHD for this query has bags $\{A,B\}$ and $\{B,C\}$, and thus has fhw $1$. The faqw is also $1$. If we apply our decomposition, we get a GHD with bags $\{A\}$, $\{A,B\}$, $\{B,C\}$ which still has fhw $1$. FAQ’s decomposition on the reduced hypergraph (with output attribute $A$ removed) has one bag $\{B,C\}$. Adding $A$ to it gives a single bag $\{A,B,C\}$ resulting in a fhw of $2$. More generally, consider query $Q_n$ with $\alpha =$ $((B_1,+), (B_2,+),\ldots (B_n,+))$ and relations $T(A_1,B_1)$ and also $R_{i,j}(A_i,A_j)$, $S_{i,j}(B_i,B_j)$ for $i,j \in \{1,2,\ldots,n\}$. Our decomposition gives a GHD with bags $\{A_1,A_2,\ldots,A_n\}$, $\{A_1,B_1\}$, $\{B_1,B_2,\ldots,B_n\}$, which has fhw $n/2$. FAQ’s decomposition has a single bag and fhw equal to $n$.
Finding optimal valid GHDs {#subsec:optimal-valid}
--------------------------
Armed with Corollary \[cor:optimal-construct\], we simplify the brute force search algorithm for finding optimal valid GHDs.
Let $Q_{{\mathcal{H}}, \alpha}$ be an ${\textsc{Ajar}\xspace}$ query. The optimal width valid GHD for this query can be found in time ${\widetilde{O}}(|{\mathcal{H}}|2^{{\widetilde{O}}(\max_{{\mathcal{H}}' \in H({\mathcal{H}}, \alpha)}(|{\mathcal{H}}'|))})$.
This runtime for finding the optimal valid GHD can be exponentially better than the naive runtime:
\[example:star-ajar\] Consider the star query ${\mathcal{H}}= (\{A, B_1,\ldots B_n\}$, $\{ \{A, B_i\} \mid 1 \leq i \leq n \} )$, $\alpha = (B_1, +), (B_2, +), \dots, (B_n, +)$. $A$ is the only output attribute. Removing $A$ breaks the hypergraph into $n$ components, so there are $n + 1$ characteristic hypergraphs, each of size $\leq 2$. Finding the optimal valid GHD takes time ${\widetilde{O}}(n)$, whereas the standard algorithm takes time exponential in $n$.
We can also approximate the GHD [@Marx:2010:AFH:1721837.1721845]:
Let $Q$ be a join query with hypergraph ${\mathcal{H}}$ and fractional hypertree width $w$. Then we can find a GHD for $Q$ in time polynomial in $|{\mathcal{H}}|$, that has width $w' \leq w^3$.
We can replicate Marx’s result for valid GHDs.
Let $Q_{{\mathcal{H}}, \alpha}$ be an ${\textsc{Ajar}\xspace}$ query, such that its minimum width valid GHD has width $w$. Then we can find a valid GHD in time polynomial in $|{\mathcal{H}}|$ that has width $w' \leq w^3$.
FAQ [@FAQ]’s decomposition lets them apply Marx’s approximation as well. However, their decomposition is not width-preserving i.e. the width of their final GHD is higher than the width of the GHDs they construct for the hypergraphs in the decomposition. Thus their decomposition gives a weaker width guarantee of $faqw^3 + faqw$ [@FAQ Theorem 9.49]. The extra $+ faqw$ factor is due to output addition. Our guarantee, $w^3$, is strictly smaller ($w$ is the width of the optimal valid GHD) as $w \leq faqw$ by Theorem \[thm:ajar-vs-faq-runtime\].
Tighter Runtime Exponents {#subsec:dbp-width}
-------------------------
Marx [@Marx:2010:THP:1806689.1806790] introduced the notion of submodular width ($sw$) that is tighter than $fhw$, and showed that a join query can be answered in time ${\textsc{IN}}^{O(sw)}$. The $O$ in the exponent is because Marx’s algorithm requires expensive preprocessing that takes ${\textsc{IN}}^{2 \times sw}$ time. After the pre-processing, the join can be performed in time ${\textsc{IN}}^{sw}$. Despite the $O$ in the exponent, this algorithm can be very valuable because there are families of hypergraphs that have unbounded $fhw$ but bounded $sw$. We can apply Marx’s algorithm to the characteristic hypergraphs, potentially improving our runtime. Marx also showed that joins on a family of hypergraphs are fixed parameter tractable if any only if the submodular width of the hypergraph family is bounded [@Marx:2010:THP:1806689.1806790]. Moreover, adaptive width [@AdaptiveWidth] (applicable only when relations are expressed as truth tables) is unbounded for a hypergraph family if and only if submodular width is unbounded. Corollary \[cor:optimal-construct\] gets us an analogous tractability result for ${\textsc{Ajar}\xspace}$ queries.
\[thm:submodular-width\] We can answer an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$ in time $O({\textsc{IN}}^{O(\max_{{\mathcal{H}}' \in H({\mathcal{H}}, \alpha)}(sw({\mathcal{H}}')))} + {\textsc{OUT}})$.
Recent work [@2015arXiv150801239J] uses degree information to more tightly bound the output size of a query. The bound in the reference, called the [*DBP bound*]{}, has a tighter exponent than the AGM bound, while requiring only linear preprocessing to obtain. The authors also provide algorithms whose runtime matches the DBP bound. We can define DBP-width $dbpw({\mathcal{T}}, {\mathcal{H}})$ such that ${\textsc{IN}}^{dbpw({\mathcal{T}}, {\mathcal{H}})}$ is the maximum value of the DBP bound over all bags of GHD $T$. We then use the improved algorithm in place of GJ in AggroGHDJoin. This lets us get tighter results “for free”, reducing our runtime to ${\textsc{IN}}^{dbpw}$ instead of ${\textsc{IN}}^{fhw}$. Formally:
\[thm:dbp-width\] Given an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$ and a valid GHD for ${\mathcal{H}}$, we can answer the query in time $O({\textsc{IN}}^{dbpw({\mathcal{T}}, {\mathcal{H}})} + {\textsc{OUT}})$. Equivalently, we can answer the query in time $O({\textsc{IN}}^{\max_{{\mathcal{H}}' \in H({\mathcal{H}}, \alpha)} dbpw({\mathcal{T}}, {\mathcal{H}})} + {\textsc{OUT}})$.
As discussed before, FAQ has a non-width-preserving decomposition. We can combine FAQ’s decomposition with the DBP bound as we did above. Suppose we perform FAQ’s decomposition, and ${\textsc{IN}}^{faqw+}$ denotes the highest value of the DBP bound on each of their characteristic hypergraphs, and on the set of output attributes. Thus the DBP-width of each of their characteristic hypergraphs, and the outputs, is $faqw+$. However, when they perform output addition, the DBP-width of the resulting GHDs can go up to $2faqw+$. This happens when the DBP bound on both the outputs and one of the characteristic hypergraphs equals ${\textsc{IN}}^{faqw+}$. So if we apply the DBP result to FAQ’s decomposition, we get a runtime of ${\widetilde{O}}({\textsc{IN}}^{2faqw+} + {\textsc{OUT}})$. Thus their decomposition causes them to incur an extra factor of $2$ in the exponent. They similarly incur a factor of $2$ increase in exponent for the submodular width algorithm.
MapReduce and Parallel Processing {#subsec:GYM}
---------------------------------
The GYM algorithm [@GYM] uses GHDs to efficiently process joins in a MapReduce setting. GYM makes use of the GHD structure to parallelize different parts of the join. Given a GHD of depth $d$, and width $w$, with $n$ attributes, GYM can perform a join in a MapReduce setting in $O(d + \log(n))$ rounds at a communication cost of $M^{-1}({\textsc{IN}}^w + {\textsc{OUT}})^2$ where $M$ is the memory per processor on the MapReduce cluster. Combining this with the degree-based MapReduce algorithm [@2015arXiv150801239J] gives us the following result:
\[thm:ajar-mapreduce-n-rounds\] Given an optimal valid GHD $({\mathcal{T}}^*, \chi)$ of depth $d$, and DBP-width $dbpw$, we can answer an ${\textsc{Ajar}\xspace}$ query with Communication Cost equal to $O(M^{-1}({\textsc{IN}}^{dbpw({\mathcal{T}}, {\mathcal{H}})} + {\textsc{OUT}})^2)$ in $d + \log(n)$ MapReduce rounds, where $n$ is the number of attributes and $M$ is the available memory per processor.
A GHD can have depth up to $O(n)$, in which case the algorithm can take a very large number of MapReduce rounds ($O(n)$). To address this, the GYM paper uses the ‘Log-GTA’ algorithm to reduce the depth of any given GHD to $\log(n)$ while at most tripling its width. This lets it process joins in $\log(n)$ MapReduce rounds at a cost of $M^{-1}({\textsc{IN}}^{3w} + {\textsc{OUT}})^2$.
Log-GTA involves some shuffling of the attributes in the GHD bags, so naively applying it to a valid GHD could make the GHD invalid (see example \[example:Log-GTA-invalidation\] in the Appendix). But our decomposition lets us apply Log-GTA to the GHD of each characteristic hypergraph, and then stitch the short GHDs together. Our decomposition is recursive in nature; let $d'$ be the maximum recursive depth of the decomposition for a given $Q_{{\mathcal{H}},\alpha}$. Then the depth of the shortened GHD of each characteristic hypergraph is $O(\log(n))$, and so the depth of the valid GHD obtained by stitching them together is $O(d'\log(n))$. This gives us the result:
\[thm:ajar-mapreduce-log-rounds\] If $dbpw$ is the DBP width of a [<span style="font-variant:small-caps;">Ajar</span>]{}query, we can answer that query with Communication Cost equal to $O(M^{-1}({\textsc{IN}}^{3 \times dbpw({\mathcal{T}}, {\mathcal{H}})} + {\textsc{OUT}})^2)$ in $d'\log(n)$ MapReduce rounds, where $n$ is the number of attributes and $M$ is the available memory per processor.
$d'$ can vary from $O(1)$ to $O(n)$ depending on the query. The star query from example \[example:star-ajar\] has $d' = 2$, which lets us process it in $\log(n)$ MapReduce rounds. Any query that only has a single type of aggregation will have $d' = 2$ as well. On the other hand, a query with one relation having $n$ attributes, $1$ output attribute, and alternating $\sum$ and $\max$ aggregations, will have $d' = n$, and will be hard to parallelize.
Olteanu and Zavodny [@OZVLDB13; @OZTODS15] use valid GHDs to answer [<span style="font-variant:small-caps;">Ajar</span>]{}queries for the special case of a single type of aggregation. But they have no notion of a decomposition and attempting to shorten a valid GHD directly, without using a decomposition, may make it invalid. FAQ’s decomposition may be used to shorten GHDs similarly to ours, but leads to an increased width of $4faqw$ compared to our $3w$ (where $w \leq faqw$ is the width of our optimal valid GHD). This is again because of output addition, if the output attributes have a width of $faqw$, and the shortened GHDs of the characteristic hypergraphs have a width of $3faqw$, then the total width will be $4faqw$.
Product Aggregations {#sec:univ-aggregation}
====================
The primary application of queries with multiple aggregations is to establish bounds for the Quantified Conjunctive Query ($QCQ$) problem [@FAQ], and its counting variant, $\#QCQ$. We now introduce a new type of aggregation, called product aggregation, that lets us efficiently handle $QCQ$ queries. We define the [<span style="font-variant:small-caps;">Ajar</span>]{}problem for product aggregations, and then extend our algorithm from Section \[subsec:simple-solution\] to handle this new type of [<span style="font-variant:small-caps;">Ajar</span>]{}query. We then define a decomposition analogous to that in Section \[sec:decomposing\]. A more detailed version of this section with additional motivation, examples, and proofs can be found in Appendix \[sec:univ-aggregation-proofs\].
AJAR queries with product aggregates
------------------------------------
A [*product aggregation*]{} aggregates using the $\otimes$ operator. Throughout the paper, we assumed that an absent tuple effectively has an annotation of $0$. Taking this into account, we formally define the product aggregation. Let $B = F \backslash A$:
$\displaystyle \sum_{(A, \otimes)} R_{AB} = \{(t_B, \lambda): \forall t_A \in {\mathcal{D}}^A, t_B \circ t_A \in R_{AB} \text{ and } \lambda = \prod_{(t, \lambda_t) \in R_{AB}: \pi_B t = t_B} \lambda_t \}$
We can adjust the definition of aggregation orderings and ${\textsc{Ajar}\xspace}$ queries to allow this new type of aggregation. $QCQ$ queries can now be expressed as [<span style="font-variant:small-caps;">Ajar</span>]{}queries on the $(\{0,1\}, \max, \cdot)$ semiring. We assume for this section that $\otimes$ is idempotent, i.e. $a \otimes a = a$ for all $a$. We describe how to work with non-idempotent products in Appendix \[subsec:non-idempotent-product\].
Algorithms for product aggregates
---------------------------------
For aggregation orderings that have product aggregations, the rules for determining when two orderings are equivalent are somewhat different; product aggregations can be performed [*before*]{} a join. We illustrate this with an example:
In the semiring $(\{0,1\}, \max, \cdot)$, suppose we have two relations $R(A,B) = \{((0,0), x), ((0,1), y)\}$ and $S(B,C) = \{((0, 1), p), ((1,1),q)\}$. Consider the [<span style="font-variant:small-caps;">Ajar</span>]{}query $\sum_{(B, \cdot)} R(A,B) \Join S(B,C)$. If we compute the join, we will get two tuples with the annotations $x \cdot p$ and $y \cdot q$, and then aggregating over $B$ will produce a relation with the element $((0,1), x \cdot p \cdot y \cdot q)$. However, note that $x \cdot p \cdot y \cdot q = (x \cdot y) \cdot (p \cdot q)$, implying that $\sum_{(B, \cdot)} R(A,B) \Join S(B,C) = (\sum_{(B, \cdot)} R(A,B)) \Join (\sum_{(B, \cdot)} S(B,C))$.
Now we describe our algorithm for solving [<span style="font-variant:small-caps;">Ajar</span>]{}queries when product aggregations are present. Our algorithm follows the same lines as the algorithm from Section \[subsec:simple-solution\]. Recall that the algorithm consisted of searching for [*equivalent orderings*]{}, then searching for GHD [*compatible*]{} with an equivalent ordering, and running AggroGHDJoin on the GHD with the smallest fhw. For product aggregations, we need to modify our algorithm for testing equivalent orderings, and our definition of compatibility; we do these in turn.
#### Testing orderings for equivalence {#testing-orderings-for-equivalence .unnumbered}
We describe how we modify Algorithm \[algo:equivalance-test\] when product aggregations are present. Let ${\textsf{PA}}(\alpha)$ denote the set of product attributes in ordering $\alpha$. We make two changes to Algorithm \[algo:equivalance-test\]. (1) Instead of removing $V(-\alpha)$ and dividing ${\mathcal{H}}$ into components, we remove $V(-\alpha) \cup {\textsf{PA}}(\alpha)$ and then divide ${\mathcal{H}}$ into components $C_1, C_2,\ldots,C_m$. Then for each $C_i$ we define $C'_i = C_i \cup \bigcup_{e \in {\mathcal{E}}_{C_i}} ({\textsf{PA}}(\alpha) \cap e)$. [^7] That is $C'_i$ has the attributes of $C_i$ as well as the product attributes that are in the same hyperedge as some attribute in $C_i$. Then we recursively call the equivalence test on $(\alpha_{C'_i}, \beta_{C'_i})$ instead of on $(\alpha_{C_i},\beta_{C_i})$. (2) When we are checking for a $i < j$ such that $\odot'_i \neq \odot'_j$ and there is a path in $\{b_i,b_{i+1},\ldots,b_{|\alpha|}\}$, we instead check for a path in $$(\{b_i,b_{i+1},\ldots,b_{|\alpha|}\} \setminus {\textsf{PA}}(\alpha)) \cup \{b_i,b_j\}$$ That is, we look for a $b_i$ that has a different operator that $b_j$, and has a path to $b_j$ consisting only of $b_i$, $b_j$, and semiring attributes in $\{b_i,b_{i+1},\ldots,b_{|\alpha|}\}$. Appendix \[sec:univ-aggregation-proofs\] gives the pseudo-code for the modified algorithm (Algorithm \[algo:equivalance-test-prod\]) and proves that it is sound and complete.
\[lemma:AlgoP-sound-complete\] The above Algorithm returns True if and only if $\alpha \equiv_{\mathcal{H}}\beta$.
#### Compatible GHDs {#compatible-ghds .unnumbered}
Product aggregates not only change the set of equivalent orderings, but also the set of GHDs compatible with a given ordering. In fact, product aggregates allow us to break the rules of GHDs without causing incorrect behavior. We express this using a simple variant of GHDs, called [*aggregating generalized hypertree decompositions*]{} (AGHDs). Informally, AGHDs are GHDs that can violate the running intersection property for attributes that have a product aggregation. AGHDs are formally defined in Appendix \[sec:univ-aggregation-proofs\]. We determine compatibility for AGHDs as follows: An AGHD is compatible with an ordering $\beta$ if for every attribute pair $a$, $b$ such that one of the $TOP(a)$ nodes is an ancestor of a $TOP(b)$ node, $a$ precedes $b$ in $\beta$.
We can now modify our algorithm from Section \[subsec:simple-solution\] to detect equivalent orderings using Algorithm \[algo:equivalance-test-prod\], then search for compatible AGHDs, and run AggroGHDJoin over the compatible AGHD with the smallest fhw. Our runtime is given by the next theorem.
\[thm:ajar-product-runtime\] Given a [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}},\alpha}$ possibly involving idempotent product aggregates, let $w^*$ be the smallest fhw for an AGHD compatible with an ordering equivalent to $\alpha$. Then the runtime for our algorithm is ${\widetilde{O}}({\textsc{IN}}^{w^*} + {\textsc{OUT}})$.
#### Decomposing AGHDs {#decomposing-aghds .unnumbered}
We can apply the ideas from Section \[sec:decomposing\] to [<span style="font-variant:small-caps;">Ajar</span>]{}queries with product aggregates as well. We can define a notion of decomposable AGHDs for queries with product aggregates, and show the following results:
\[thm:decompaghdisvalid\] All decomposable AGHDs are compatible with an ordering $\beta$ such that $\beta \equiv_{\mathcal{H}}\alpha$.
\[thm:decompaghdwidth\] For every valid AGHD $({\mathcal{T}}, \chi)$, there exists a decomposable $({\mathcal{T}}', \chi')$ such that for all node-monotone functions $\gamma$, the $\gamma$-width of $({\mathcal{T}}', \chi')$ is no larger than the $\gamma$-width of $({\mathcal{T}}, \chi)$.
We can define characteristic hypergraphs similarly to how we did in Section \[sec:decomposing\] (see Appendix \[sec:univ-aggregation-proofs\] for a formal definition). We have the following result:
\[thm:decompaghd\] For an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$ involving product aggregates, suppose ${\mathcal{H}}_0, \dots, {\mathcal{H}}_k$ are the characteristic hypergraphs $H({\mathcal{H}}, \alpha)$. Then GHDs $G_0, G_1, \dots, G_k$ of ${\mathcal{H}}_0, \dots, {\mathcal{H}}_k$ can be connected to form a decomposable AGHD $G$ for $Q_{{\mathcal{H}}, \alpha}$. Conversely, any decomposable AGHD $G$ of $Q_{{\mathcal{H}}, \alpha}$ can be partitioned into GHDs $G_0, G_1, \dots, G_k$ of the characteristic hypergraphs ${\mathcal{H}}_0, \dots, {\mathcal{H}}_k$. Moreover, in both of these cases, $\gamma\text{-width}(G) = \max_{i} \gamma\text{-width}(G_i)$.
These theorems let us apply all the optimizations from Section \[subsec:optimal-valid\], \[subsec:dbp-width\], and \[subsec:GYM\] to [<span style="font-variant:small-caps;">Ajar</span>]{}queries with product aggregates.
#### Comparison to FAQ {#comparison-to-faq .unnumbered}
The runtime of InsideOut on a query involving idempotent product aggregations is given by ${\widetilde{O}}({\textsc{IN}}^{faqw})$, where the faqw depends on the ordering, and the presence of product aggregations. Our algorithm for handling product aggregations recovers the runtime of FAQ. Formally,
For any [<span style="font-variant:small-caps;">Ajar</span>]{}query involving idempotent product aggregations, ${\textsc{IN}}^{w^*} + {\textsc{OUT}}\leq 2 \cdot {\textsc{IN}}^{faqw}$.
The proof is in Appendix \[subsec:faq-comparison-proof\]. By applying ideas from the FAQ paper to our setting, we can also recover the FAQ runtime on $\#QCQ$ (Appendix \[subsec:recovering-hash-qcq\]). Algorithm \[algo:equivalance-test-prod\] for detecting equivalence of orderings is both sound and complete; in contrast, FAQ’s equivalence testing algorithm is sound but not complete. Moreover, we have a width-preserving decomposition for queries with product aggregates. This allows us to get tighter runtime exponents in terms of submodular and DBP-widths (Theorems \[thm:submodular-width\], \[thm:dbp-width\]) and efficient MapReduce Algorithms (Theorems \[thm:ajar-mapreduce-n-rounds\], \[thm:ajar-mapreduce-log-rounds\]).
Conclusion {#sec:conclusion}
==========
We investigate solutions to and the structure of ${\textsc{Ajar}\xspace}$ queries: aggregate-join queries with multiple aggregators over annotated relations. We start by providing a very simple algorithm based on a variant of the standard GHDJoin algorithm that generates query plans by relying on a simple test of equivalence between aggregation orderings. This naive approach is sufficient to recover and surpass the runtime of state-of-the-art solutions. We proceed to investigate more interesting technical questions regarding the structure of ${\textsc{Ajar}\xspace}$ queries. We first develop a partial ordering that fully characterizes equivalent orderings. We then develop a characterization of the corresponding valid GHDs, describing how they can be decomposed into ordinary, unrestricted GHDs. This reduction connects us to a trove of parallel work on GHDs. We finish by extending our work to handle product aggregations.
#### Acknowledgements {#acknowledgements .unnumbered}
We thank Atri Rudra for invaluable insights and feedback developing our approach. CR gratefully acknowledges the support of the Defense Advanced Research Projects Agency (DARPA) XDATA Program under No. FA8750-12-2-0335 and DEFT Program under No. FA8750-13-2-0039, DARPAs MEMEX program under No. FA8750-14-2-0240, the National Science Foundation (NSF) under CAREER Award No. IIS-1353606, Award No. CCF-1356918 and EarthCube Award under No. ACI-1343760, the Office of Naval Research (ONR) under awards No. N000141210041 and No. N000141310129, the Sloan Research Fellowship, the Moore Foundation Data Driven Investigator award, and gifts from American Family Insurance, Google, Lightspeed Ventures, and Toshiba.
Background {#sec:background}
==========
The AggroGHDJoin algorithm is a simple variant of some well-known join algorithms. We describe these algorithms next.
#### GenericJoin {#genericjoin .unnumbered}
We first describe the AGM bound on the join output size developed by Atserias, Grohe, and Marx [@AGM]. Given query hypergraph ${\mathcal{H}}_Q = ({\mathcal{V}}, {\mathcal{E}})$ and relations $\{R_F | F \in {\mathcal{E}}\}$, consider the following linear program: $$\begin{aligned}
\textrm{Minimize } &\sum_{F \in {\mathcal{E}}} x_F \log_{{\textsc{IN}}}(|R_F|) \\
\forall v \in {\mathcal{V}}: &\sum_{F : v \in F} x_F \ge 1 \\
\forall F \in {\mathcal{E}}: & x_F \ge 0\end{aligned}$$
Any feasible solution $\overrightarrow{x}$ is a *fractional edge cover*. Suppose $\rho^*$ is the optimal objective. Then the *AGM bound* on the worst-case output size of join $\Join_{F \in {\mathcal{E}}} R_F$ is given by ${\textsc{IN}}^{\rho^*} = \prod_{F \in {\mathcal{E}}} |R_F|^{x_F^*}$. We will use ${\textsc{IN}}^{AGM(Q)}$ to denote the AGM bound on a query $Q$. The GenericJoin (GJ) algorithm [@NRR] computes a join in time ${\widetilde{O}}({\textsc{IN}}^{AGM(Q)})$ for any join query. GJ will be used as a subroutine in a later algorithm, where $GJ({\mathcal{H}}, \{ R_F | F \in {\mathcal{E}}_{\mathcal{H}}\})$ denotes a call to GenericJoin with one input relation $R_F$ per hyperedge $F$ in hypergraph ${\mathcal{H}}$.
#### Yannakakis {#yannakakis .unnumbered}
Yannakakis’ algorithm [@Yannakakis81] operates on $\alpha$-acyclic queries. There are several different equivalent definitions of $\alpha$-acyclicity; we provide the definition that builds a tree out of the relations as it most naturally relates to generalized hypertree decompositions.
Given a hypergraph ${\mathcal{H}}= ({\mathcal{V}}_{\mathcal{H}}, {\mathcal{E}}_{\mathcal{H}})$, a *join tree* over ${\mathcal{H}}$ is a tree ${\mathcal{T}}= ({\mathcal{V}}_{\mathcal{T}}, {\mathcal{E}}_{\mathcal{T}})$ with ${\mathcal{V}}_{\mathcal{T}}= {\mathcal{E}}_{\mathcal{H}}$ such that for every attribute $A \in {\mathcal{V}}_{\mathcal{H}}$, the set $\{v \in {\mathcal{V}}_{\mathcal{T}}| A \in v\}$ forms a connected subtree in ${\mathcal{T}}$.
A hypergraph ${\mathcal{H}}$ is $\alpha$-acyclic if there exists a join tree over ${\mathcal{H}}$ [@ALICE; @Yannakakis81]. We can use the classic GYO algorithm to produce a join tree [@ALICE ch.6]. The Yannakakis algorithm takes a join tree as input. It’s pseudo-code is given in Section \[subsec:background-1\].
Algorithm \[NoAggroY\] runs in $O({\textsc{IN}}+ {\textsc{OUT}})$ where ${\textsc{IN}}$ and ${\textsc{OUT}}$ are the sizes of the input and output, respectively.
To leverage the speed of Yannakakis for cyclic queries, we look to GHDs [@FHTW; @gottlob:ght]. The intuition behind a GHD is to group the attributes into bags (as specified by the function $\chi$) such that we can build a *join tree* over these bags. This allows us to run $GJ$ within each bag and then Yannakakis on the join tree. The resulting algorithm is GHDJoinwhose pseudo-code is given in Algorithm \[NoAggroGHD\]. The runtime of GHDJoin is given by ${\widetilde{O}}({\textsc{IN}}^{fhw({\mathcal{T}},{\mathcal{H}})} + {\textsc{OUT}})$
**Input:** Query hypergraph ${\mathcal{H}}$, GHD $({\mathcal{T}}, \chi)$, Relations $R_F$ for each $F \in {\mathcal{E}}_{\mathcal{H}}$
$S_R \gets \emptyset$ ${\mathcal{H}}_t \gets (\chi(t), \{ \pi_{\chi(t)} F | F \in {\mathcal{E}}_{\mathcal{H}}\})$ $S_R \gets S_R \cup GJ({\mathcal{H}}_t, \{ \pi_{\chi(t)} R_F | F \in {\mathcal{E}}_{\mathcal{H}}\})$ $Yannakakis({\mathcal{T}}, S_R)$
Algorithm \[NoAggroGHD\] runs in ${\widetilde{O}}({\textsc{IN}}^{fhw({\mathcal{T}}, {\mathcal{H}})} + {\textsc{OUT}})$.
We can make some straightforward modifications to the above join algorithms to perform aggregations. The traditional Yannakakis and GHDJoin algorithms perform the join in a bottom up fashion, after a semijoin phase to ensure that there are no dangling tuples. The modified algorithms above handle aggregations using the same intuition as in traditional query plans: “push down” aggregations as far as possible. Since each attribute must occur in a connected subtree of the GHD, we can push its aggregation down to the root of this connected subtree, which is the $TOP$ node of the attribute. There is a standard modification to Yannakakis for project-join queries that projects away attributes at their $TOP$ node [@Yannakakis81]. Instead of projecting, we perform aggregation.
We provide the pseudo-code of AggroYannakakis, which is a simple variant of the well-known Yannakakis [@Yannakakis81] algorithm, in Algorithm \[AggroY\]. Algorithm \[Aggro\] gives the pseudo-code of AggroGHDJoin, which is a variant of GHDJoin that calls AggroYannakakis instead of Yannakakis. AggroGHDJoin also does some extra work to ensure we pass each annotation to $GJ$ only once. The $\pi^1$ operator in AggroGHDJoin denotes a projection that projects tuples while replacing the annotation by $1$, to ensure that the same annotation isn’t counted more than once.
**Input:** Join tree ${\mathcal{T}}= ({\mathcal{V}}, {\mathcal{E}})$, Aggregation order $\alpha$, Relations $R_F$ for each $F \in {\mathcal{V}}$
Semi-join reduction up $P \gets$ parent of $F$ $R_P \gets R_P \ltimes R_F$ Semi-join reduction down $P \gets$ parent of $F$ $R_F \gets R_F \ltimes R_P$ Aggregation $\beta \gets \alpha \cap \{a \in {\mathcal{V}}| TOP_{\mathcal{T}}(a) = F\}$ $R' \gets \Sigma_{\beta} R_F$ $P \gets$ parent of $F$ $R_P \gets R_P \Join R'$ Compute the join $R_R$ for the root $R$
**Input:** Query hypergraph ${\mathcal{H}}$, GHD $({\mathcal{T}}, \chi)$, Relations $R_F$ for each $F \in {\mathcal{E}}_{\mathcal{H}}$
$S_R \gets \emptyset$ ${\mathcal{H}}_t \gets (\chi(t), \{ \pi_{\chi(t)} F | F \in {\mathcal{E}}_{\mathcal{H}}\})$ $I \gets \{ R_F | F \subseteq \chi(t), \exists a \in F : TOP_{\mathcal{T}}(a) = t\} \cup \{ \pi_{\chi(t)}^1 R_F | F \not\subseteq \chi(t) \text{ or } \forall a \in F : TOP_{\mathcal{T}}(a) \neq t\}$ $S_R \gets S_R \cup GJ({\mathcal{H}}_t, I)$ AggroYannakakis$({\mathcal{T}}, S_R)$
In the classic analysis of Yannakakis, the runtime of the semi-join portion is bounded by $O({\textsc{IN}})$ and the bottom-up join is bounded by $O({\textsc{OUT}})$. In AggroYannakakis, the analysis of the semi-join portion is unchanged, but the aggregation reduces the size of the output, thereby making the ${\textsc{OUT}}$ bound harder to achieve. In particular, during the bottom-up join, we may compute an intermediate relation whose attributes are not a subset of the output attributes, meaning that its size may not bounded by ${\textsc{OUT}}$. These potentially large intermediate relations are the underlying cause for the traditional ${\textsc{IN}}\cdot {\textsc{OUT}}$ runtime.
However, using intuition discovered in reference [@OZVLDB13], if we require the output attributes to appear above non-output attributes, we can preserve the ${\textsc{IN}}+ {\textsc{OUT}}$ runtime.
\[thm:aggro-complexity\] Suppose we have a GHD such that for any output attribute $A$ and non-output attribute $B$, $TOP_{\mathcal{T}}(B)$ is not an ancestor of $TOP_{\mathcal{T}}(A)$. AggroGHDJoin runs in ${\widetilde{O}}({\textsc{IN}}^{fhw({\mathcal{T}}, {\mathcal{H}})} + {\textsc{OUT}})$ given this GHD.
$GJ$ on each bag still runs in ${\widetilde{O}}({\textsc{IN}}^{fhw({\mathcal{T}}, {\mathcal{H}})})$. We need to prove the Yannakakis portion runs in $O({\textsc{IN}}+{\textsc{OUT}})$ after running $GJ$.
The semijoin portion runs in $O(IN)$ as in the original Yannakakis algorithm. In the join phase, we have two types of joins. In the first type, $F \setminus \beta \subseteq P$. This implies the join output is a subset of $R_P$ (with different annotations). So the total runtime of this type of join is $O({\textsc{IN}})$. For the second type, $F \setminus \beta \subsetneq P$. This means some attribute in $(F \setminus \beta) \setminus P$ must be an output attribute, and all attributes in $P$ must be output attributes as well (as their $TOP$ value is an ancestor of $F$). So the result of our join must be a subset of the output table; the total runtime of this type of join is $O({\textsc{OUT}})$. Thus the total runtime of the algorithm is $O({\textsc{IN}}+{\textsc{OUT}})$.
Note that while AggroGHDJoin runs in ${\widetilde{O}}({\textsc{IN}}^{fhw({\mathcal{T}}, {\mathcal{H}})} + {\textsc{OUT}})$ time on the GHDs above, it may not necessarily produce the right output unless the GHD satisfies additional conditions, to ensure that aggregations can be done in the proper order. In particular, recall our definition a GHD $({\mathcal{T}}, \chi)$ is compatible with $\alpha$ if for all attribute pairs $A,B$, $TOP_{\mathcal{T}}(A)$ being an ancestor of $TOP_{\mathcal{T}}(B)$ implies that either $A$ is an output variable or $A$ occurs before $B$ in $\alpha$.
\[thm:aggro-compatible\] If a GHD $({\mathcal{T}}, \chi)$ is compatible with $\alpha$, then AggroGHDJoin given $({\mathcal{T}}, \chi)$ correctly computes $Q_{{\mathcal{H}}, \alpha}$.
We first show that AggroYannakakis works as expected. We note that the semi-join reduction does not change the output; it only quickens the process. We only consider the bottom-up join. For each node $t$ in the join tree, let $R(t)$ be the relation associated with that node before this loop (i.e. after the semi-join portion). Let $R'(t)$ be the final relation associated with node $t$ when we are processing node $t$ (i.e. after the bottom up join with $t$’s descendants is done, and after the aggregation in $t$). Let ${\mathcal{T}}_t$ be the subtree that includes $t$ and all of its descendants. Let $s(t)$ be the attributes aggregated at node $t$, i.e. $\alpha \cap \{a \in {\mathcal{V}}| TOP_{\mathcal{T}}(a) = t\}$, and let $s({\mathcal{T}}_t) = \cup_{t \in {\mathcal{T}}_t} s(t)$. For each non-leaf node $t$, let $c(t)$ be the set of $t$’s children.
For each node $t$, we claim $R'(t) = \sum_{\alpha_{s({\mathcal{T}}_t)}} \Join_{t' \in {\mathcal{T}}_t} R(t')$. Proof by induction on the tree. For each leaf $l$, $R'(l) = \sum_{\alpha_{s(l)}} R(l)$ by definition.
For a non-leaf node $t$, $$\begin{aligned}
R'(t) &= \sum_{\alpha_{s(t)}} R(t) \Join (\Join_{t_c \in c(t)} R'(t_c)) \\
&= \sum_{\alpha_{s(t)}} R(t) \Join \left(\Join_{t_c \in c(t)} \sum_{\alpha_{s({\mathcal{T}}_{t_c})}} \Join_{t' \in {\mathcal{T}}_{t_c}} R(t')\right)\\
&= \sum_{\alpha_{s({\mathcal{T}}_t)}} \Join_{t' \in {\mathcal{T}}_t} R(t')\end{aligned}$$ The second step is due to the inductive hypothesis. The final step is simply “pulling out" the aggregations from the sub-orderings one at a time; we can arbitrarily interleave the aggregation orders $\alpha_{s({\mathcal{T}}_{t_c})}$. We can simply interleave them to match $\alpha_{\cup_{t_c \in c(t)} s({\mathcal{T}}_t)}$. Since the original GHD is compatible with $\alpha$, we know the aggregations $\alpha_{s(t)}$ precede $\alpha_{s({\mathcal{T}}_{t_c})}$ in $\alpha$, implying that $\sum_{\alpha_{s({\mathcal{T}}_t)}} \sum_{\alpha_{\cup_{t_c \in c(t)} s({\mathcal{T}}_t)}} = \sum_{\alpha_{s({\mathcal{T}}_t)}}$. Our output is $R'(t_r)$ where $t_r$ is the root node, which is $\sum_{\alpha} \Join_{t \in {\mathcal{T}}} R(t)$ as desired.
Since AggroYannakakis works as expected, we simply need to ensure that the bags are computed appropriately. Note the GHD ensures for every relation $R_F$, there is a node $t$ such that $F \subseteq \chi(t)$. This means that no tuple is lost; computing AggroYannakakis on the bags will compute the correct tuples. To ensure it computes the correct annotations, we need to ensure every annotation appears in the bags at most once; our algorithm places the annotation of a relation $R_F$ in the top-most node that contains all of the attributes $R_F$.
[**Product Aggregations:**]{} When product aggregations are present in an [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}}, \alpha}$, we have a notion of product partition hypergraphs, AGHDs over product partition hypergraphs, and a corresponding notion of AGHDs compatible with an ordering. We now prove theorem \[thm:ajar-product-runtime\] that extends theorems \[thm:aggro-complexity\] and \[thm:aggro-compatible\] to the case where product aggregations are present.
A product partition partition $P = ({\mathcal{V}}_P, {\mathcal{E}}_P)$ essentially creates multiple renamed copies of each product attribute $a$ ($a_1, a_2,\ldots,a_{|P_a|}$), and assigns one of the renamed copies to each relation containing $a$. An AGHD is essentially a GHD over $P$. Given $P$, and $a \in {\mathcal{V}}_{\mathcal{H}}$, let $P(a)$ equal $\{a\}$ if $a$ is not a product attribute, and $\{a_1,\ldots,a_{|P_a|}\}$ otherwise. Given $a' \in {\mathcal{V}}_P$, let $P^{-1}(a')$ equal $a$ such that $a' \in P(a)$. Given an edge $F \in {\mathcal{E}}_P$, let $P^{-1}(F)$ denote the edge $\{P^{-1}(a') \mid a' \in F\}$. We define a modified ordering $\alpha^P$ over ${\mathcal{V}}_P$ that takes $\alpha$ and replaces each occurrence of $(a,\otimes)$ with $(a_1,\otimes)$,$(a_2,\otimes)$,$\ldots$,$(a_{|P_a|},\otimes)$ for each product attribute $a$. For any $F \in {\mathcal{E}}_P$, we define the relation $R_F$ to be same as the $R_{P^{-1}(F)}$ (but with the attribute name changed. This gives us the modified [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q^P_{P, \alpha^P} = \sum_{\alpha^P} \Join_{F \in {\mathcal{E}}_P} R_F$. Then we have,
\[lemma:product-renaming-one\] Suppose $R'(A,C_1)$ is a copy of $R(A,C)$ with $C$ renamed to $C_1$, and $S'(B,C_2)$ is a copy of $S(B,C)$ with $C$ renamed to $C_2$. Then $$\sum_{(C_1,\otimes)}\sum_{(C_2,\otimes)} R'(A,C_1) \Join S'(B,C_2) = \sum_{(C,\otimes)} R(A,C) \Join S(B,C)$$
Suppose the annotations for the $C$ values in $R$ are $n_1,n_2,\ldots,n_k$ and in $S$ are $m_1,m_2,\ldots,m_k$ (assume all annotations are present i.e. absent tuples have a zero-annotation). Then the RHS is $\otimes_{i=1}^{k}n_im_i$. The LHS will have $\otimes_{i=1}^{k}n_i \otimes_{j=1}^{k}m_j$. The RHS is equal to the LHS because of idempotence of $\otimes$. Note that if $\otimes$ wasn’t idempotent, the LHS would have the $m_j$ terms multiplied $k$ times while the RHS has them once.
\[lemma:product-renaming-ajar\] For each database instance $I$, $Q_{{\mathcal{H}},\alpha}(I) = Q^P_{P,\alpha^P}(I)$.
This lemma can be proved by repeated application of Lemma \[lemma:product-renaming-one\].
Now we can easily prove Theorem \[thm:ajar-product-runtime\]. Suppose we have a AGHD $D = ({\mathcal{T}}, \chi, P)$ which is compatible with an ordering $\alpha$. Then the GHD $({\mathcal{T}}, \chi)$ over hypergraph $P$, is compatible with $\alpha^P$. Running AggroGHDJoin over this GHD, with ordering $\alpha^P$ correctly computes $Q^P_{P,\alpha^P}(I)$, due to theorem \[thm:aggro-compatible\]. And by Lemma \[lemma:product-renaming-ajar\], this also equals $Q_{{\mathcal{H}},\alpha}(I)$, which is the output we want. Also, since the AGHD is compatible with $\alpha$, the GHD must satisfy the condition of Theorem \[thm:aggro-complexity\], and hence AggroGHDJoin runs on it in time ${\widetilde{O}}({\textsc{IN}}^{fhw} + {\textsc{OUT}})$.
Comparison with Related Work {#sec:app-related}
============================
Section \[sec:prelim\] {#subsec:faq-comparison-proof}
----------------------
In Section \[sec:prelim\], we define a simple approach to solving ${\textsc{Ajar}\xspace}$ queries, and we claim in Theorem \[thm:ajar-vs-faq-runtime\] that our runtime guarantee of ${\widetilde{O}}(IN^{w^*} + {\textsc{OUT}}) \le {\widetilde{O}}(IN^{faqw})$. We note that the $faqw$ exponent is actually the optimum value of $faqw(\sigma)$ over the equivalent orderings $\sigma$ they consider (we discuss the space of orderings they consider in the next subsection). Our approach will recognize $\sigma$ as being equivalent, and will search for the best compatible GHD for $\sigma$. We will show that there exists a compatible AGHD $({\mathcal{T}}, \chi, P)$ for every equivalent ordering $\sigma$ such that $fhw({\mathcal{T}}, {\mathcal{H}}) = faqw(\sigma)$ (as Example \[ex:faqoutput\] shows, the compatible AGHD ${\mathcal{T}}$ we construct may not be the optimal compatible GHD).
We start by briefly summarizing FAQ’s algorithm, with the pseudo-code (written in the notation of this paper) given in Algorithm \[algo:InsideOut\]. Let $\sigma$ be the ordering used for aggregation. Let $n$ denote the total number of attributes $|{\mathcal{V}}_{\mathcal{H}}|$ and $f$ denote the number of output attributes (thus $|\sigma| = n - f$). For notational convenience, we will be using $\sigma[i]$ to denote both the attribute and the operator that make up the $i^{th}$ operator-attribute pair in the ordering.
**Input:** Hypergraph ${\mathcal{H}}= ({\mathcal{V}}_{\mathcal{H}}, {\mathcal{E}}_{\mathcal{H}})$, Aggregation ordering $\sigma$, Relations $R_F$ for each $F \in {\mathcal{E}}_{\mathcal{H}}$
$E_n \gets \{R_F \mid F \in {\mathcal{E}}_{\mathcal{H}}\}$ $\delta(k) \gets \{R_F \in E_k \mid \sigma[k-f] \in F\}$ $U_k \gets \Join_{R \in \delta(k)} R$ $E_{k-1} \gets (E_k \setminus \delta(k)) \cup \{\sum_{\sigma[k-f]} U_k\}$ $E_{k-1} \gets (E_k \setminus \delta(k)) \cup \{\sum_{\sigma[k-f]} R \mid R \in \delta(k) \}$ $\Join_{R \in E_f} R$
FAQ relies on a worst-case optimal algorithm to compute each of the joins, implying that in the ${\widetilde{O}}(IN^{faqw})$ runtime guarantee, faqw is defined as the maximum AGM bound placed on each of the computed joins. Define $p^*_H : 2^{{\mathcal{V}}_{\mathcal{H}}} \to \mathcal{R}$ to be a function that maps a subset of the attributes to the AGM bound on the subset (i.e. the optimal value of the canonical linear program). Then $faqw = \max(\max_k p^*_H(U_k), p^*_H(V(-\sigma)))$ [@FAQ].
We will build up the compatible AGHD $({\mathcal{T}}, \chi, P)$ in rounds corresponding to each of the $k$ values of $InsideOut$. We first describe how to construct $({\mathcal{T}}, \chi)$, and later describe how to obtain $P$. At the start of round corresponding to a particular $k$, we will have a forest of AGHDs, each of which will have a root mapped (by $\chi$) to the attribute sets of $E_k$, and at the end of each round, the forest’s roots will be mapped to the relations of $E_{k-1}$.
For an attribute set $F$, let $t(F)$ represent the node such that $\chi(t(F)) = F$. We start by creating the $|{\mathcal{E}}_{\mathcal{H}}|$ nodes $\{t(F) | F \in E_n\}$, which are simply nodes mapped to the input relations. Then for each $k$ from $n$ to $f+1$, let $T$ represent the set of nodes $\{t(F) | F \in \delta(n)\}$; these are the nodes that will be processed (i.e. the nodes for whom we will create parents). If $\sigma[k-f]$ is not a product aggregation, we create a node $t(U_k)$ and set $parent(t) = t(U_k)$ for all $t \in T$. We then create a node $t(U_k \backslash \{\sigma[k-f]\})$ and set it to be $parent(t(U_k))$. Note that this process has transformed the set of the forest’s roots by removing $T$ and adding $t(U_k \backslash \{\sigma[k-f]\})$, mirroring the transformation between $E_k$ and $E_{k-1}$. If $\sigma[k-f]$ is a product aggregation, then for each $F \in \delta(n)$, we create a node $t(F \backslash \{\sigma[k-f]\})$ and set it to be $parent(t(F))$; in this case as well the set of the forest’s roots match $E_{k-1}$.
At the end of this process, we will have a forest of AGHDs whose roots map to the relations in $E_f$. To conclude our construction, we simply construct the node $t(V(-\sigma))$ and set it to be $parent(t(F))$ for all $F \in E_f$. If there are no product aggregations, then $({\mathcal{T}}, \chi)$ forms a GHD.
$({\mathcal{T}}, \chi)$ satisfies the running intersection property for all non-product attributes, but a product attribute $a$ can be present in multiple disconnected parts of ${\mathcal{T}}$. We now describe a product partition $P$ such that $({\mathcal{T}},\chi,P)$ forms an AGHD for the ordering $\sigma$. Let $P_a$ denote the number of distinct connected components of ${\mathcal{T}}$ in which $a$ is present. Then we create $P_a$ copies of $a$ ($a_1$, $a_2$,$\ldots$,$a_{|P_a|}$), and assign a copy to each component in some order. For each $F \in {\mathcal{E}}_{\mathcal{H}}$ that contains $a$, if the component that $t(F)$ belongs to is assigned $a_i$, then $P$ assigns $a_i$ to $F$. Then $({\mathcal{T}}, \chi, P)$ is an AGHD for $\sigma$.
The AGHD $({\mathcal{T}}, \chi, P)$ as described is trivially compatible with $\sigma$ since we construct $parent(TOP_{\mathcal{T}}(\sigma[k-f]))$ explicitly in round $k$; this ensures that $TOP_{\mathcal{T}}(\sigma[i])$ cannot be an ancestor of $TOP_{\mathcal{T}}(\sigma[j])$ if $i > j$.
Define $p^*_{\mathcal{H}}$ to be a function that maps a set of attributes to the AGM bound on the set (the optimal value of the canonical linear program). The AGHD $({\mathcal{T}}, \chi, P)$ constructed as described satisfies $$fhw({\mathcal{T}}, {\mathcal{H}}) = \max(p^*_{\mathcal{H}}(V(-\alpha)), \max_{k}{p^*_{\mathcal{H}}(U_k)}) = faqw(\sigma).$$
The nodes in our tree that do not map to $V(-\alpha)$ or the $U_k$ either map to an input relation or to a relation created by aggregating an attribute from a single child node. In the former case, $p^*_{\mathcal{H}}$ would evaluate to $1$, so we can ignore them in our maximum. In the latter case, the attributes are a strict subset of its child’s attributes, implying we can ignore them too. As such, the fractional hypertree width is simply the maximum fractional cover over $V(-\alpha)$ and the $U_k$. This shows the first part of the equality.
The second part of the equality is the definition of $faqw$ [@FAQ].
Theorem \[thm:ajar-vs-faq-runtime\], as well as its analogue for product aggregations, follow as a simple corollary. We now show an example where the runtime of InsideOut is much worse than the runtime of our Algorithm, primarily due to the fact that it is not output-sensitive.
\[ex:faqoutput\] Let $n$ be an even number, and consider an [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}}, \alpha}$ where ${\mathcal{H}}= (\{A_i \mid 1 \leq i \leq n\}, \{\{A_i, A_{i+1}\} \mid 1 \leq i \leq n-1 \} \cup \{\{A_n,A_1\}\})$, and $\alpha$ is empty (i.e. the query is just a join). Also let each attribute take values $1, 2, 3,\ldots 2\times\lfloor \sqrt{N} \rfloor$. Suppose each relation $\{A_i,A_{i+1}\}$ for $1 \leq i \leq n-1$ connects values of the same parity, while relation $\{A_n,A_1\}$ connectes values of opposite parities. Thus each relation has size $N$, and ${\textsc{IN}}= O(N)$ ($n$ is a constant), and the join output is empty. There is a GHD with bags $\{A_1,A_2,A_3\}, \{A_1,A_3,A_4\}, \ldots \{A_1, A_{n-1},A_n\}$ that is compatible with the empty ordering. The fhw of this GHD is $2$, so we have $w^* = 2$. Thus the runtime of our algorithm will be ${\widetilde{O}}({\textsc{IN}}^2)$. InsideOut will compute an intermediate output consisting of the join of $n-1$ of the relations, which has size $N^{(n-1)/2}$ , so InsideOut’s runtime will be at least ${\widetilde{O}}({\textsc{IN}}^{(n-1)/2})$.
FAQ does discuss, at a very high-level and without proofs, changes to InsideOut that will allow their runtime to be output-sensitive [@FAQ Section 10.2]. Their most general and useful change involves building a GHD for the output variables and running a message passing algorithm between the bags, which exactly describes GHDJoin. Implementing this change would make InsideOut completely equivalent to AggroGHDJoin. We note that the FAQ paper frames these changes as decisions in how to *represent* the output, whereas we present the optimization in an algorithmic context, independent of any other storage optimizations.
Section \[sec:equivalent-orderings\]
------------------------------------
In Section \[sec:equivalent-orderings\], we define a partial order $<_{{\mathcal{H}}, \alpha}$ that exactly characterizes the constraints an aggregation ordering must satisfy to be equivalent to a given ordering $\alpha$. Our partial ordering is complete, which is a result that FAQ cannot match. Much like our approach, FAQ actually defines their own partial ordering, which we denote $<_{FAQ}$, and their work only considers orderings that are linear extensions of $<_{FAQ}$. However, we will show an example where $<_{FAQ}$ has unnecessary constraints:
\[example:faq-incompleteness\] Consider the ${\textsc{Ajar}\xspace}$ query given by $\sum_A \max_B \sum_C R(A,B) S(A,C)$. By our characterization, $A <_{{\mathcal{H}}, \alpha} B$ is the only constraint, giving rise to $3$ different valid orderings. The FAQ characterization, however, has two constraints: $A <_{FAQ} B$ and $C <_{FAQ} B$, which only allows for $2$ different valid orderings. Note that FAQ constraints *preclude the original ordering* ABC.
Section \[sec:decomposing\]
---------------------------
In Section \[sec:decomposing\], we define a decomposition that relates the width of a valid GHD to the widths of a series of ordinary GHDs. Variable orderings (as used by FAQ) are not as readily suited as GHDs are for decompositions. FAQ does derive their own version of a decomposition, but the difficulties that arise when using variable orderings are exemplified in the way FAQ switches between GHDs and variable orderings in their proofs [@FAQ]. In addition, the FAQ decomposition is demonstrably weaker than ours; their decomposition incurs some overhead costs when combining the sub-orderings to build the overall ordering, precluding a result like Corollary \[cor:optimal-construct\] that provides the groundwork for the variety of extensions we provide. To exemplify the gap in the two decompositions, we inspect a specific ${\textsc{Ajar}\xspace}$ query:
\[example:faq-decomposition\] Consider the query $\sum_B \sum_C \sum_D R(A,B) S(B,C) T(C,D), U(D,A)$. Suppose $|A| = \sqrt{N}$, $|B| = 2 = |D|$, $|C| = N$, and all of the pairwise relations are constructed as complete cross products of the attributes’ values. Our decomposition will result in the chain GHD $A-ABD-BCD$, while the FAQ decomposition will result in the GHD $A-ABC-ACD$. The runtimes of both FAQ and GHDJoin using the former GHD is ${\widetilde{O}}(N)$, whereas the runtimes using the latter GHD are ${\widetilde{O}}(N^{3/2})$. As such, the FAQ decomposition will perform asymptotically worse than our decomposition.
More generally, consider a query $Q_n$ with relations $R_i(A_i,B)$, $S_i(B,C_i)$, $T_i(C_i,D)$, $U_i(D,A_i)$ for $1 \leq i \leq n$. Like before, all $|A_i|$’s are $\sqrt{N}$, all $|C_i|$’s are $N$, and $|B| = |D| = 2$. And the aggregation ordering only has the $+$ operator, on $B$, $D$ and all $C_i$’s. Our decomposition gives the chain $A_1\ldots A_n$ $-$ $A_1\ldots A_nBD$ $-$ $C_1\ldots C_n BD$. This results in a runtime of ${\widetilde{O}}(N^n)$. FAQ’s decomposition gives $A_1\ldots A_nBC_1\ldots C_n$ $-$ $A_1\ldots A_n D C_1\ldots C_n$. This decomposition, and its corresponding ordering, give a runtime of ${\widetilde{O}}(N^{3n/2})$. Thus the difference between runtime exponents caused by FAQ’s decomposition and our decomposition can be arbitrarily high.
Section \[subsec:GYM\]
----------------------
\[example:Log-GTA-invalidation\] Suppose we a [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}}, \alpha}$ with $${\mathcal{H}}= (\{A,B,C,D,E,F\}, \{\{A,B\}, \{B,C\}, \{B,D,E\}, \{D,F\}\})$$ and $\alpha = ((D, \sum), (E, \sum), (F, \sum))$. We start with the width-$1$ valid GHD $({\mathcal{T}}, \chi)$ with ${\mathcal{V}}-{\mathcal{T}}= \{v_1,v_2,v_3,v_4\}$ and $${\mathcal{E}}_{\mathcal{T}}= \{v_1,v_2\}, \{v_2,v_3\}, \{v_3,v_4\}$$ such that $v_1$ is the root, and $\chi(v_1) = \{A,B\}$, $\chi(v_2) = \{B,C\}$, $\chi(v_3) = \{B,D,E\}$, $\chi(v_4) = \{D,F\}$.
Applying Log-GTA gives us a shorter GHD $({\mathcal{T}}', \chi')$ with ${\mathcal{V}}_{\mathcal{T}}' = \{u, v_1, v_2, v_3, v_4\}$, $${\mathcal{E}}_{\mathcal{T}}' = \{\{v_1,u\}, \{u,v_2\}, \{u,v_3\}, \{u,v_4\} \}$$ with $v_1$ as the root. $\chi'(u) = \{B,D\}$ and $\chi'(v_i) = \chi(v_i)$ for all $i$. Now $TOP(D) = u$ which is an ancestor of $TOP(C) = v_2$, despite $C$ being an output attribute and $D$ not being an output attribute. This means GHD $({\mathcal{T}}', \chi')$ is invalid, showing that applying Log-GTA to a valid GHD may make it invalid.
As the above example shows, we cannot directly apply Log-GTA to a valid GHD to get a shorter valid GHD.
Characterizing Equivalent Orderings: Proofs {#sec:app-equiv}
===========================================
We now formally present our partial order $<_{{\mathcal{H}},\alpha}$ that characterizes the interaction of the two forms of commuting. As we said in Section \[sec:equivalent-orderings\], we have two relations ${\textsf{PREC}}$ and ${\textsf{DNC}}$, that are mutually recursive. We initialize the constraints to a base case and iteratively update them till we reach a fixed point. We now formalize this. We use binary operator $<^i_{{\mathcal{H}},\alpha}$ to denote the constraint ${\textsf{PREC}}$ after $i$ iterations, and operator $\sim^i_{{\mathcal{H}},\alpha}$ to denote ${\textsf{DNC}}$ after $i$ iterations, with one difference; both operators behave slightly differently for output attributes. To readily incorporate output attributes into the constraints, we define an augmented aggregation ordering below:
For any aggregation ordering $\alpha$, let $F$ be the set of output variables. Then define $\alpha^O = \alpha^O_1, \alpha^O_2, \dots, \alpha^O_n$ to be a sequence such that $\alpha^O_i = (F_i, \texttt{NULL})$ for $1 \le i \le |F|$ and $\alpha^O_i = \alpha_{i+|F|}$ for $|F|+1 \le i \le n$.
Note that $n$ is defined to be the number of attributes in the query. Now we can formally define $<^i_{{\mathcal{H}},\alpha}$ and $\sim^i_{{\mathcal{H}},\alpha}$. Both of these binary operators operator over attribute-operator pairs, but since each attribute occurs at most once in an ordering, we can equivalently think of them as operating over attributes. We use these two interchangeably e.g. $A <_{{\mathcal{H}}, \alpha} B$ denotes the same thing as $(A, \oplus) <_{{\mathcal{H}}, \alpha} (B, \oplus')$.
For a given query $Q_{{\mathcal{H}}, \alpha}$ with ${\mathcal{H}}= ({\mathcal{V}}, {\mathcal{E}})$, we define relations $\sim_{{\mathcal{H}}, \alpha}^i$ and partial orders $<_{{\mathcal{H}}, \alpha}^i$ over attribute-operator pairs in $\alpha^O$. For any $A,B \in {\mathcal{V}}$, suppose $(A, \oplus), (B, \oplus') \in \alpha^O$. Then, for $i = 0$, $(A, \oplus) \sim_{{\mathcal{H}}, \alpha}^0 (B, \oplus')$ if and only if one of the following is true:
- $\oplus \neq \oplus'$ and $\exists E \in {\mathcal{E}}: A,B \in E.$ $(0.1)$
- $\oplus \neq \oplus'$ and either $\oplus = \texttt{NULL}$ or $\oplus' = \texttt{NULL}$ $(0.2)$
For $i > 0$, $(A, \oplus) \sim^i_{{\mathcal{H}}, \alpha} (B, \oplus')$ if and only if $(A, \oplus) \not\sim^j_{{\mathcal{H}}, \alpha} (B, \oplus')$ for all $j < i$ and one of the following is true:
- $\oplus \neq \oplus'$ and $\exists E \in {\mathcal{E}}, (C, \oplus'') \in \alpha^O: B,C \in E, (A, \oplus) <_{{\mathcal{H}}, \alpha}^{i-1} (C, \oplus'')$ $(i.1)$
- $\exists (C, \oplus'') \in \alpha^O$ and $j,k < i: (A, \oplus) <_{{\mathcal{H}}, \alpha}^j (C, \oplus'') <_{{\mathcal{H}}, \alpha}^k (B, \oplus')$ $(i.2)$
For any $i \ge 0$, $(A, \oplus) <_{{\mathcal{H}}, \alpha}^i (B, \oplus')$ if and only if $(A, \oplus) \sim_{{\mathcal{H}}, \alpha}^i (B, \oplus')$ and $(A, \oplus)$ precedes $(B, \oplus')$ in $\alpha^O$.
Finally, $(A, \oplus) \sim_{{\mathcal{H}}, \alpha} (B, \oplus')$ if and only if $(A, \oplus) \sim_{{\mathcal{H}}, \alpha}^i (B, \oplus')$ for some $i \ge 0$. Similarly, $(A, \oplus) <_{{\mathcal{H}}, \alpha} (B, \oplus')$ if and only if $(A, \oplus) <_{{\mathcal{H}}, \alpha}^i (B, \oplus')$ for some $i \ge 0$.
The core of our definition is the four labeled conditions for $\sim$. The condition $0.1$ represents the simplest structure that violates both conditions of Theorem \[commute\]; it represents our base case. Condition $0.2$ simply ensures the output attributes precede non-output attributes. Our condition $i.1$ extends the structure from $0.1$ beyond single relations. If $A < C$ and $C$ appears in a relation with $B$, we can guarantee that $A$ and $B$ cannot be separated in the way the second condition of Theorem \[commute\] requires, and if $\oplus \neq \oplus'$, the first condition is violated as well. Condition $i.2$ simply ensures that transitivity interacts properly with condition $i.1$.
We now prove the two lemmas stated in Section \[sec:equivalent-orderings\], followed by proving soundness and completeness of $<_{{\mathcal{H}},\alpha}$.
Suppose we are given a hypergraph ${\mathcal{H}}= ({\mathcal{V}}, {\mathcal{E}})$ and an aggregation ordering $\alpha$. Fix two arbitrary attributes $A,B \in {\mathcal{V}}$ such that $(A, \oplus), (B, \oplus') \in \alpha^O$ for differing operators $\oplus \neq \oplus'$. Then, for any path $P$ in ${\mathcal{H}}$ between $A$ and $B$, there must exist some attribute in the path $C \in P$ such that $C <_{{\mathcal{H}}, \alpha} A$ or $C <_{{\mathcal{H}}, \alpha} B$.
We use induction on the length of path $P$.\
[*Base Case:*]{} Let $|P| = 2$. This implies that there exists some edge $E \in {\mathcal{E}}$ such that $A,B \in E$. Thus $A \sim_{{\mathcal{H}}, \alpha}^0 B$. Then, by definition, either $A <_{{\mathcal{H}}, \alpha} B$ or $B <_{{\mathcal{H}}, \alpha} A$ depending on which attribute appears first in $\alpha$.
[*Induction:*]{} Suppose $|P| = N > 2$ and assume the lemma is true for paths of length $< N$. We call this assumption the [*outer inductive hypothesis*]{}, for reasons that will become apparent later. Path $P$ can be rewritten as $P = AP'B$ where $P'$ is a path of length at least $1$. Let $C$ be the node in $P'$ that appears earliest in $\alpha^O$; this implies that there exists no attribute in our path $D \in P'$ such that $D <_{{\mathcal{H}}, \alpha} C$. Define an operator $\oplus''$ such that $(C, \oplus'') \in \alpha^O$. Since $\oplus \neq \oplus'$, either $\oplus \neq \oplus''$ or $\oplus' \neq \oplus''$. Without loss of generality, assume that $\oplus \neq \oplus''$.
Consider the subpath of $P$ from $A$ to $C$. It is shorter than $N$ and connects two attributes with different operators. We apply our inductive hypothesis to get that there exists some $D \in P$ such that either $D <_{{\mathcal{H}}, \alpha} A$ or $D <_{{\mathcal{H}}, \alpha} C$. In the first case, we have found an attribute that satisfies our conditions and we are done. In the second case, we know that $D \notin P'$ by our definition of $C$. Thus $D$ must be $A$; we have that $A <_{{\mathcal{H}}, \alpha} C$.
Consider the subpath of $P$ from $C$ to $B$; let $X_i$ denote the $i^{th}$ node in this path for $0 \le i \le k$, where $X_0 = C$ and $X_k = B$. We claim that for all $i < k$, $A <_{{\mathcal{H}}, \alpha} X_i$. We argue this inductively; for our base case, we are given that $A <_{{\mathcal{H}}, \alpha} C = X_0$. Now let $i \ge 1$, and assume $A <_{{\mathcal{H}}, \alpha} X_{j}$ for $j < i$. Call this the [*inner inductive hypothesis*]{}.
Note that we have $A<_{{\mathcal{H}}, \alpha} C$ and that $C$ must precede $X_i$ by definition. Thus $A$ precedes $X_i$ in $\alpha^O$. All that remains is showing that $A \sim_{{\mathcal{H}}, \alpha} X_i$. Define $\oplus^i$ such that $(X_i, \oplus^i) \in \alpha^O$. Since we assumed earlier that $\oplus \neq \oplus''$, we know that either $\oplus^i \neq \oplus$ or $\oplus^i \neq \oplus''$.
- $\oplus^i \neq \oplus$\
By our (inner) inductive hypothesis, we know that $A <_{{\mathcal{H}}, \alpha} X_{i-1}$. We also know that there must exist some edge $E \in {\mathcal{E}}$ such that $X_{i-1}, X_i \in E$. Thus by condition $i.1$, $A \sim_{{\mathcal{H}}, \alpha} X_i$.
- $\oplus^i \neq \oplus''$\
By our (outer) inductive hypothesis, we know that for some $0 \le j \le i$, $X_j <_{{\mathcal{H}}, \alpha} C$ or $X_j <_{{\mathcal{H}}, \alpha} X_i$. By our definition of $C$, the first case is impossible. And by our (inner) inductive hypothesis, we have that $A <_{{\mathcal{H}}, \alpha} X_j$. We thus have that $A <_{{\mathcal{H}}, \alpha} X_j <_{{\mathcal{H}}, \alpha} X_i$, which implies that $A \sim_{{\mathcal{H}}, \alpha} X_i$ by condition $i.2$.
This gives us that $A <_{{\mathcal{H}}, \alpha} X_{k-1}$. Since there exists an edge $E \in {\mathcal{E}}$ such that $X_{k-1}, B \in E$, condition $i.1$ tells us that $A \sim_{{\mathcal{H}}, \alpha} B$. As before, this implies that either $A <_{{\mathcal{H}}, \alpha} B$ or $B <_{{\mathcal{H}}, \alpha} A$.
Given a hypergraph ${\mathcal{H}}= ({\mathcal{V}}, {\mathcal{E}})$ and an aggregation ordering $\alpha$, suppose we have two attributes $A, B \in V(\alpha)$ such that $A <_{{\mathcal{H}}, \alpha} B$. Then there must exist a path $P$ from $A$ to $B$ such that for every $C \in P, C \neq A$ we have $A <_{{\mathcal{H}}, \alpha} C$.
Define $\oplus$ and $\oplus'$ such that $(A, \oplus), (B, \oplus') \in \alpha$. In addition define $i$ such that $A <_{{\mathcal{H}}, \alpha}^i B$, which implies that $A$ precedes $B$ in $\alpha$ and that $A \sim_{{\mathcal{H}}, \alpha}^i B$. Our proof is by induction on $i$. For our basecase, if $A \sim_{{\mathcal{H}}, \alpha}^0 B$, we know that $\exists E \in {\mathcal{E}}: A, B \in E$. Thus the path $P = AB$ satisfies our conditions.
For $i > 0$, we have the following cases:
- $\oplus \neq \oplus'$ and $\exists E \in {\mathcal{E}}, (C, \oplus'') \in \alpha^O: B,C \in E, A <_{{\mathcal{H}}, \alpha}^{i-1} C$\
By our inductive hypothesis, there must exist a path $P'$ from $A$ to $C$ such that for all $D \in P', D \neq A$ we have $A <_{{\mathcal{H}}, \alpha} D$. Then the path $P = P'B$ satisfies our conditions.
- $\oplus \neq \oplus'$ and $\exists E \in {\mathcal{E}}, (C, \oplus'') \in \alpha^O: A,C \in E, B<_{{\mathcal{H}}, \alpha}^{i-1} C$\
By our inductive hypothesis, there must exist a path $P'$ from $B$ to $C$ such that for all $D \in P', D \neq B$ we have $B <_{{\mathcal{H}}, \alpha} D$, which also implies that $A <_{{\mathcal{H}}, \alpha} D$ by $i.2$. Let $\overline{P'}$ be the reverse of $P'$. Then the path $P = A\overline{P'}$ satisfies our condition.
- $\exists C \in {\mathcal{V}}$ and $j,k < i: A <_{{\mathcal{H}}, \alpha}^j C <_{{\mathcal{H}}, \alpha}^k B$\
By our inductive hypothesis, there must exist two paths $P'$ and $P''$. $P'$ is a path from $A$ to $C$ such that for all $D \in P', D \neq A$ we have $A <_{{\mathcal{H}}, \alpha} D$. Similarly, $P''$ is a path from $C$ to $B$ such that for all $D \in P'', D \neq C$ we have $C <_{{\mathcal{H}}, \alpha} D$, which implies $A <_{{\mathcal{H}}, \alpha} D$. Thus the path $P = P'P''$ satisfies our conditions.
Suppose we are given a hypgergraph ${\mathcal{H}}= ({\mathcal{V}}, {\mathcal{E}})$ and aggregation orderings $\alpha, \beta$. Then $\alpha \equiv_{\mathcal{H}}\beta$ if and only if $\beta$ is a linear extension of $<_{{\mathcal{H}},\alpha}$.
[**Soundness:**]{}
We use induction on the number of inversion in $\beta$ with respect to the ordering $\alpha$. [*Base Case:*]{} $0$ inversions. Then $\beta$ is identical to $\alpha$ and $\alpha \equiv_{\mathcal{H}}\beta$.
[*Induction:*]{} Suppose $\beta$ has $N > 0$ inversions, and assume the lemma is true for orderings with $<N$ inversions. There must be some $\beta_i$ and $\beta_{i+1}$ that are inverted with respect to $\alpha$. Consider the ordering $\beta'$ derived by swapping $\beta_i$ and $\beta_{i+1}$. It has $N-1$ inversions with respect to $\alpha$ and is clearly a linear extension of $<_{{\mathcal{H}}, \alpha}$. Thus, by the inductive hypothesis, $\alpha \equiv_{\mathcal{H}}\beta'$.
We now show that $\beta \equiv_{\mathcal{H}}\beta'$. Suppose $\beta_i = (A, \oplus)$ and $\beta_{i+1} = (B, \oplus')$. We have two cases to consider.
- $\oplus = \oplus'$\
By Theorem \[commute\], we can swap $\beta_i$ and $\beta_{i+1}$ without affecting the output. This implies that $\beta \equiv_{\mathcal{H}}\beta'$.
- $\oplus \neq \oplus'$\
By Lemma \[Pathing\] and since we know $A$ and $B$ are incomparable under $<_{{\mathcal{H}}, \alpha}$, any path between $A$ and $B$ must go through some attribute $C$ such that $C<_{{\mathcal{H}}, \alpha} A$ or $C<_{{\mathcal{H}}, \alpha} B$. Since $\beta$ is a valid linear extension of $<_{{\mathcal{H}},\alpha}$, these attributes $C$ appear earlier than index $i$ in $\beta$. This implies that $A$ and $B$ are in separate connected components in $\pi_{V(\beta_i, \beta_{i+1}, \dots, \beta_{|\beta|})} {\mathcal{H}}$, which implies that we can swap $\beta_i$ and $\beta_{i+1}$ without affecting the output by Theorem \[commute\]. This implies that $\beta \equiv_{\mathcal{H}}\beta'$.
[**Completeness:**]{}
We prove the contrapositive: we assume that we are given aggregation orderings $\alpha, \beta$ such that $\beta$ is not a linear extension of $<_{{\mathcal{H}}, \alpha}$, and we will show that $\alpha \not \equiv_{{\mathcal{H}}} \beta$. We will do so by constructing an instance $\hat{I}$ such that $Q_{{\mathcal{H}}, \alpha}(\hat{I}) \neq Q_{{\mathcal{H}}, \beta}(\hat{I})$.
We assume without loss of generality that ${\mathcal{V}}= V(\alpha)$, i.e. that there are no output attributes. We will provide an example where $\beta$ and $\alpha$ must differ in the single annotation that comprises the output. If there are output attributes, we can augment our example by putting $1$s in all the output attributes; our output will be composed of a single tuple composed of all $1$s with the same annotation as in our example below.
Consider the set of all valid linear extensions of $<_{{\mathcal{H}}, \alpha}$. Suppose the maximum length prefix identical to the prefix of $\beta$ is of size $k$. Among all linear extensions with maximum length identical prefixes, suppose the minimum possible index for $\beta_{k+1}$ is $k'$. Consider a linear extension $\alpha'$ such that $\alpha'_i = \beta_i$ for $i \le k$ and $\alpha'_{k'} = \beta_{k+1}$. By the soundness part of our proof, $\alpha' \equiv_{\mathcal{H}}\alpha$; to show that $\alpha \not\equiv_{\mathcal{H}}\beta$ we can simply show that $\alpha' \not\equiv_{\mathcal{H}}\beta$.
Suppose $\alpha'_{k'} = (A, \oplus) = \beta_{k+1}$ and $\alpha'_{k'-1} = (B, \oplus')$. We know that $B <_{{\mathcal{H}}, \alpha} A$ since $k'$ is the minimum possible index for $\beta_{k+1}$ in any linear extension of $<_{{\mathcal{H}}, \alpha}$. Also, since $B$ and $A$ are adjacent in $\alpha'$, we know that there cannot exist any $C$ such that $B <_{{\mathcal{H}}, \alpha} C <_{{\mathcal{H}}, \alpha} A$. These two facts combine to imply $\oplus \neq \oplus'$. Then, by Lemma \[Pathing2\], there exists a path $P$ from $A$ to $B$ such that every attribute in our path $C \in P$ other than $A$ and $B$ must appear after index $k'$ in $\alpha'$.
Since $\oplus \neq \oplus'$, there must exist $x,y \in \textbf{D}$ such that $x \oplus y \neq x \oplus' y$. Define a relation $\widehat{R_{\mathcal{V}}}$ with two tuples. The first tuple will contain a $1$ for each attribute and an annotation $x$. The second tuple will contain a $2$ for each attribute in $P$ (including $A$ and $B$) and a $1$ for every other attribute. The second tuple will be annotated with $y$. Note that among the attributes in $P$, $(A, \oplus)$ is the outermost aggregation in $\beta$ and $(B, \oplus')$ is the outermost aggregation in $\alpha$. This implies that $$\Sigma_{\alpha'_1} \Sigma_{\alpha'_2} \cdots \Sigma_{\alpha'_{n'}} \widehat{R_{\mathcal{V}}} = x \oplus' y
\hspace{15mm}
\Sigma_{\beta_1} \Sigma_{\beta_2} \cdots \Sigma_{\beta_{|\beta|}}
\widehat{R_{\mathcal{V}}} = x \oplus y$$
Let $C$ be the attribute in $P$ right before $B$; by definition there must exist an edge $E \in {\mathcal{E}}$ such that $B,C \in E$. Consider the following instance over the schema ${\mathcal{H}}$: $$\hat{I} = \{ \pi_{E}\widehat{R_{\mathcal{V}}} \} \cup \{ \pi_F^1 \widehat{R_{\mathcal{V}}} | F \in {\mathcal{E}}, F \neq E\} .$$
By definition, $\Join_{R_F \in \hat{I}} R_F = \widehat{R_{\mathcal{V}}}$. Since we know that $x \oplus y \neq x \oplus' y$, we have that $$\Sigma_{\alpha'_1} \Sigma_{\alpha'_2} \cdots \Sigma_{\alpha'_{n'}} \Join_{R_F \in \hat{I}} R_F
\neq
\Sigma_{\beta_1} \Sigma_{\beta_2} \cdots \Sigma_{\beta_{|\beta|}} \Join_{R_F \in \hat{I}} R_F$$
We thus have that $Q_{{\mathcal{H}}, \alpha'}(\hat{I}) \neq Q_{{\mathcal{H}}, \beta}(\hat{I})$, which implies that $\alpha' \not\equiv_{\mathcal{H}}\beta$.
Decomposing Valid GHDs: Proofs {#sec:app-decomp}
==============================
We start by stating and proving a useful lemma about the aggregation orderings seen in the sub-trees of a decomposable GHD.
\[lemma:sub-orders\] Given a hypergraph ${\mathcal{H}}= ({\mathcal{V}}, {\mathcal{E}})$ and an aggregation ordering $\alpha$, suppose $C$ is a connected component of ${\mathcal{H}}\backslash V(-\alpha)$. Define ${\mathcal{H}}_C = (\bigcup_{E \in {\mathcal{E}}_C} E, {\mathcal{E}}_C)$[^8]. For any $A \in C, B \in {\mathcal{V}}$, if $A <_{{\mathcal{H}}, \alpha} B$ then $A <_{{\mathcal{H}}_C, \alpha_{C \backslash C^O}} B$. Similarly, if $A <_{{\mathcal{H}}_C, \alpha_C \backslash C^O} B$, either $A \in C^O$ or $A <_{{\mathcal{H}}, \alpha} B$.
First we show that for any $A \in C$, if $A <_{{\mathcal{H}}, \alpha} B$ then $B \in C \backslash C^O$. If $B \in C^O$, then by definition, $A \not <_{{\mathcal{H}}, \alpha} B$. If $B \notin C$, then every path between $A$ and $C$ must go through attributes in $V(-\alpha)$. Thus, by the contrapositive of Lemma \[Pathing2\], $A \not <_{{\mathcal{H}}, \alpha} B$. This implies that $A <_{{\mathcal{H}}, \alpha} B$ only for $B \in C \backslash C^O$.
Note that for any $A \in C^O$, since $A$ is an output attribute in $\alpha_{C \backslash C^O}$, $A <_{{\mathcal{H}}_C, \alpha_{C \backslash C^O}} B$ for all $B \in C \backslash C^O$. This proves our lemma for $A \in C^O$.
For $A \in C \backslash C^O$, we prove the lemma by showing that for any $i \ge 0$, $\{B | A <_{{\mathcal{H}}, \alpha}^i B\} = \{B | A <^i_{{\mathcal{H}}_C, \alpha_{C \backslash C^O}} B\}$. Note that our earlier result shows that both of these sets are subsets of $C \backslash C^O$, so we know that for any $i$, any $B$ in either set appears in both aggregation orderings.
Proof by induction on $i$. We first consider the base case: $i = 0$. We note that since $A$ is not an output attribute condition $(0.2)$ is irrelevant. Since ${\mathcal{H}}_C$ contains all edges involving attributes in $C$ and $\alpha_{C \backslash^i C^O}$ preserves the ordering and the operators of elements of $\alpha$, condition $(0.1)$ applies to the same set of attributes in both ${\textsc{Ajar}\xspace}$ queries $Q_{{\mathcal{H}}, \alpha}$ and $Q_{{\mathcal{H}}_C, \alpha_{C \backslash C^O}}$. Thus $\{B | A <_{{\mathcal{H}}, \alpha}^0 B\} = \{B | A <_{{\mathcal{H}}_C, \alpha_{C \backslash C^O}}^0 B\}$.
For $i > 0$, the inductive hypothesis supposes $\{B | A <_{{\mathcal{H}}, \alpha}^j B\} = \{B | A <_{{\mathcal{H}}_C, \alpha_{C \backslash C^O}}^j B\}$ for all $j < i$. Again since ${\mathcal{H}}_C$ contains all edges involving attributes in $C$ and $\alpha_{C \backslash^i C^O}$ preserves the ordering and the operators of elements of $\alpha$, the inductive hypothesis trivially implies conditions $(i.1)$ and $(i.2)$ apply to the same set of attributes.
Every decomposable GHD is valid.
Suppose the ${\textsc{Ajar}\xspace}$ query is $Q_{{\mathcal{H}}, \alpha}$. We need to show for any $A,B$ such that $TOP_{\mathcal{T}}(A)$ is an ancestor of $TOP_{\mathcal{T}}(B)$, $A \not <_{{\mathcal{H}}, \alpha} B$. Proof by induction on $|\alpha|$. If $|\alpha|=0$, all GHDs are valid and decomposable. For $|\alpha > 0|$, we note ${\mathcal{T}}_0$ ensure the output attributes are above non-output attributes.
If $A$ and $B$ are non-output attributes and $TOP_{\mathcal{T}}(A)$ is an ancestor of $TOP_{\mathcal{T}}(B)$, then both are in some ${\mathcal{T}}_C$. By the inductive hypothesis, ${\mathcal{T}}_C$ is valid with respect to $Q_{(\cup_{E \in {\mathcal{E}}_C} E, {\mathcal{E}}_C), \alpha_{C \backslash C^O}}$. By Lemma \[lemma:sub-orders\], this implies $A \not <_{{\mathcal{H}}, \alpha} B$.
For every valid GHD $({\mathcal{T}}, \chi)$, there exists a decomposable GHD $({\mathcal{T}}', \chi')$ such that for all node-monotone functions $\gamma$, the $\gamma$-width of $({\mathcal{T}}', \chi')$ is no larger than the $\gamma$-width of $({\mathcal{T}}, \chi)$.
In the proof sketch provided in Section \[sec:decomposing\], we claim to have width-preserving transformations of a GHD that can enforce two additional properties, which we now present and name:
- *TOP-unique*: every node $t \in {\mathcal{T}}$ is $TOP_{\mathcal{T}}(A)$ for exactly one attribute $A$\
- *subtree-connected*: for any node $t \in {\mathcal{T}}$ and the subtree ${\mathcal{T}}_t$ rooted at $t$, the attributes $\{v \in {\mathcal{V}}| TOP_{\mathcal{T}}(v) \in {\mathcal{T}}_t\}$ form a connected subgraph of ${\mathcal{H}}$
We first have two lemmas proving the transformations required to enforce these properties are width-preserving.
\[lemma:decomp-prop1\] Given a valid GHD $({\mathcal{T}}, \chi)$ with $\gamma$-width $w$, we can transform it to be TOP-unique while ensuring $\gamma$-width $\le w$.
Define a function $TOP^{-1}_{\mathcal{T}}: {\mathcal{T}}\to 2^{\mathcal{V}}$ from nodes to sets of attributes such that $TOP^{-1}_{\mathcal{T}}(t) = \{A | TOP_{\mathcal{T}}(A) = t\}$.
First we eliminate nodes $t \in {\mathcal{T}}$ such that $|TOP^{-1}_{\mathcal{T}}(t)| = 0$. We note, by definition, $\chi(t) \subseteq \chi(parent(t))$. This implies that we can simply remove $t$, connecting all of its children to $parent(t)$ without violating any properties of the valid GHD. And the width is trivially preserved.
Now suppose for some node $t \in {\mathcal{T}}$, $|TOP^{-1}_{\mathcal{T}}(t)| = k > 1$. Let $A_1$ be the attribute in $TOP^{-1}_{\mathcal{T}}(t)$ that is earliest in the aggregation ordering. Let $X = \chi(t) \cap \chi(parent(t))$. Then create a new node $t'$ such that $\chi(t') = \{A_1\} \cup X$ and add it to ${\mathcal{T}}$ between $t$ and $parent(t)$. All of properties of the valid GHD must still hold, and since the new node contains a subset of the attributes in $t$, the width must be preserved. Note that after adding this node, $|TOP^{-1}_{\mathcal{T}}(t)| = k-1$; we can repeat this process until the set is of size $1$.
\[lemma:decomp-prop2\] Given a valid GHD $({\mathcal{T}}, \chi)$ with $\gamma$-width $w$ that is TOP-unique, we can transform it to be subtree-connected while preserving TOP-unique and $\gamma$-width $\le w$.
For any node $t \in {\mathcal{T}}$, define ${\mathcal{V}}_t = \{v \in {\mathcal{V}}| TOP_{\mathcal{T}}(v) \in {\mathcal{T}}_t\}$.
We proceed with a proof by (bottom-up) induction on the tree ${\mathcal{T}}$. As our base case, we consider the leaves of ${\mathcal{T}}$. Since $l$ does not have any children, ${\mathcal{V}}_l$ must contains exactly one attribute, which is trivially a connected subgraph of ${\mathcal{H}}$.
Now we consider the subtree ${\mathcal{T}}_t$ rooted at some internal node $t$. Let $A$ be the attribute such that $TOP_{\mathcal{T}}(A) = t$. Let $c_1, c_2, \dots, c_k$ be the children of $t$. By the inductive hypothesis, the subtrees rooted at these children satisfy all of the desired properties. We note that, by definiton, $\chi(t) \backslash A \subseteq \chi(parent(t))$. For any child $c_i$ such that ${\mathcal{V}}_{c_i}$ and $A$ are not connected in ${\mathcal{H}}$, we can remove $A$ from $\chi({\mathcal{T}}_{c_i})$ and set $parent(c_i)$ to be $parent(t)$. By doing so for all such children $c_i$, we ensure that ${\mathcal{V}}_t$ is a connected subgraph of ${\mathcal{H}}$. Since $A$ is not connected to ${\mathcal{V}}_{c_i}$, this transformation does not violate the properties of GHDs. Since we are not creating any new ancestral relationships between nodes, the transformation does not violate the properties of valid GHDs. Finally, the $\gamma$-width $\le w$ and TOP-unique properties are preserved trivially.
We have thus established that we can transform any valid GHD to additionally satisfy TOP-unique and subtree-connected while preserving width. We now show that any valid GHD satisfying the two additional properties is decomposable. Combined with the two lemmas above, this will complete the proof of Theorem \[thm:decompwidth\]. Before we dive into the proof, we prove two helpful lemmas.
\[lemma:inbetween\] Given an AJAR+ query $Q_{{\mathcal{H}}, \alpha}$, if $A <_{{\mathcal{H}}, \alpha} B$ for $A,B$ with identical operators, there must exist some $C$ with a different operator such that $A <_{{\mathcal{H}}, \alpha} C <_{{\mathcal{H}}, \alpha} B$.
$A <_{{\mathcal{H}}, \alpha}^i B$ for some fixed $i$. If $A,B$ have identical operators, the only way $A <_{{\mathcal{H}}, \alpha}^i B$ is via rule $(i.2)$, which requires some $C$ and $j,k < i$ such that $A <_{{\mathcal{H}}, \alpha}^j C <_{{\mathcal{H}}, \alpha}^k B$. If this $C$ has the same operator as $A$ and $B$, we can repeatedly apply this rule until we find some attribute between $A$ and $B$ with a different operator (since both of the rules for $i=0$ only apply to attributes with differing operators).
\[lemma:ancestor\] Given an AJAR+ query $Q_{{\mathcal{H}}, \alpha}$ and valid GHD $({\mathcal{T}}< \chi)$. Suppose $A<_{{\mathcal{H}}, \alpha} B$, $A$ is not an output attribute, and $TOP_{\mathcal{T}}(A)$ is a top node only for $A$. Then, $TOP_{\mathcal{T}}(A)$ must be an ancestor of $TOP_{\mathcal{T}}(B)$ in any valid GHD.
Lemma \[Pathing2\] implies that there exists a path from $A$ to $B$ such that for every $C$ in the path such that $C \neq A$, $A <_{{\mathcal{H}}, \alpha} C$. Let $C_0, C_1, \dots, C_k$ represent the path, where $C_0 = A$ and $C_k = B$. We claim that $TOP_{\mathcal{T}}(A)$ is an ancestor of $TOP_{\mathcal{T}}(C_i)$ for all $1 \le i \le k$. Proof by induction on $i$. Our base case is for $i=1$. By the definition of a path, $A$ and $C_1$ must appear together in some hyperedge, implying that they appear together in some bag of ${\mathcal{T}}$. Both $TOP_{\mathcal{T}}(C_1)$ and $TOP_{\mathcal{T}}(A)$ must either be equal to or an ancestor of this bag. Since $TOP_{\mathcal{T}}(C_1)$ cannot be equal to or an ancestor of $TOP_{\mathcal{T}}(A)$, $TOP_{\mathcal{T}}(A)$ is an ancestor of $TOP_{\mathcal{T}}(C_1)$.
For $i>1$, we note since $C_i$ and $C_{i-1}$ appear in an edge together, by the same logic as above, $TOP_{\mathcal{T}}(C_i)$ and $TOP_{\mathcal{T}}(C_{i-1})$ must both be equal to or an ancestor of some node $t \in {\mathcal{T}}$. By the inductive hypothesis, $TOP_{\mathcal{T}}(A)$ is an ancestor of $TOP_{\mathcal{T}}(C_{i-1}$, implying that $TOP_{\mathcal{T}}(A)$ is an ancestor of $t$. Since $TOP_{\mathcal{T}}(C_i)$ cannot equal or be an ancestor of $TOP_{\mathcal{T}}(A)$, $TOP_{\mathcal{T}}(A)$ must be an ancestor of $TOP_{\mathcal{T}}(C_i)$.
Any valid GHD $({\mathcal{T}}, \chi)$ that is TOP-unique and subtree-connected must be decomposable.
We actually prove a slightly stronger statement. Define the property TOP-semiunique as follows: every non-root node $t \in {\mathcal{T}}$ is the $TOP_{\mathcal{T}}$ node for exactly one attribute and the root node is either the $TOP_{\mathcal{T}}$ for exactly one attribute or more than one output attribute (and zero non-output attributes). Note that the TOP-unique property directly implies the TOP-semiunique property. We will show that if $({\mathcal{T}}, \chi)$ is a valid, TOP-semiunique, and subtree-connected GHD for the ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$, it must be decomposable.
Proof by induction on $|\alpha|$. If $|\alpha = 0|$, then every GHD is decomposable.
Suppose $|\alpha > 0|$. Consider the set of nodes that are $TOP_{\mathcal{T}}$ nodes for output attributes, i.e. $\{t \in {\mathcal{T}}| \exists A \in V(-\alpha): TOP_{\mathcal{T}}(A) = t\}$. Since no non-output attribute can have a top node above an output attribute’s top node, the TOP-semiunique property guarantees that this set of nodes forms a rooted subtree ${\mathcal{T}}_0$ of ${\mathcal{T}}$ such that $\chi({\mathcal{T}}_0) = V(-\alpha)$.
Consider the subtrees in ${\mathcal{T}}\backslash {\mathcal{T}}_0$. Call them ${\mathcal{T}}_1, {\mathcal{T}}_2, \dots, {\mathcal{T}}_k$. For any ${\mathcal{T}}_i$, let ${\mathcal{V}}_i$ be the attributes that have $TOP_{\mathcal{T}}$ nodes in ${\mathcal{T}}_i$, i.e. ${\mathcal{V}}_i = \{A \in {\mathcal{V}}| TOP_{\mathcal{T}}(A) \in {\mathcal{T}}_i\}$. None of these ${\mathcal{V}}_i$ can contain any output attributes, and connected-subtree guarantees that each of the ${\mathcal{V}}_i$ are connected. Thus, the ${\mathcal{V}}_i$ must be the connected components of ${\mathcal{H}}\backslash V(-\alpha)$. So for each connected component $C$ of ${\mathcal{H}}\backslash V(-\alpha)$, the corresponding subtree ${\mathcal{T}}_C$ is the subtree ${\mathcal{T}}_i$ such that ${\mathcal{V}}_i = C$. Since for any $A \in C$, $TOP_{\mathcal{T}}(A) \in {\mathcal{T}}_C$, the attributes in $C$ only appear in ${\mathcal{T}}_C$. Note that for every edge $E \in {\mathcal{E}}$, there exists a node $t \in {\mathcal{T}}$ such that $E \subseteq \chi(t)$. This implies that for every edge $E \in {\mathcal{E}}_C$, there exists a node $t \in {\mathcal{T}}_C$ such that $E \subseteq \chi(t)$. As such, we can conclude that each ${\mathcal{T}}_C$ is a GHD for the hypergraph $(\bigcup_{E \in {\mathcal{E}}_C} E, {\mathcal{E}}_C)$.
Define ${\mathcal{V}}_C = \bigcup_{E \in {\mathcal{E}}_C} E$. To complete this proof, we now need to show that each ${\mathcal{T}}_C$ is a decomposable GHD for the ${\textsc{Ajar}\xspace}$ query $Q_{({\mathcal{V}}_C, {\mathcal{E}}_C), \alpha_{C \backslash C^O}}$. By the inductive hypothesis, if ${\mathcal{T}}_C$ is valid, TOP-semiunique and subtree-connected, it must be decomposable. Note that since ${\mathcal{T}}$ is TOP-semiunique and subtree-connected, ${\mathcal{T}}_C$ must also be TOP-semiunique and subtree-connected. We have also established that ${\mathcal{T}}_C$ is a GHD for $({\mathcal{V}}_C, {\mathcal{E}}_C)$. Thus to finish this proof, we only need to show that for any $A, B \in {\mathcal{V}}_C$ such that $TOP_{{\mathcal{T}}_C}(B)$ is an ancestor of $TOP_{{\mathcal{T}}_C}(A)$, $A \not <_{({\mathcal{V}}_C, {\mathcal{E}}_C), \alpha_{C \backslash C^O}} B$.
For ease of notation, in the rest of this proof we will use $<_C$ to represent $<_{({\mathcal{V}}_C, {\mathcal{E}}_C), \alpha_{C \backslash C^O}}$. We show the contrapositive: if $A <_C B$, $TOP_{{\mathcal{T}}_C}(B)$ is not an ancestor of $TOP_{{\mathcal{T}}_C}(A)$. We consider a few cases. If $A \in {\mathcal{V}}_C \backslash C$, $A$ must be in $V(-\alpha)$, implying $TOP_{{\mathcal{T}}_C}(A)$ is the root of ${\mathcal{T}}_C$. For any $A \in C$, note that $TOP_{{\mathcal{T}}_C}(A) = TOP_{\mathcal{T}}(A)$. By Lemma \[lemma:sub-orders\], for any $A \in C$, if $A <_C B$ then either $A <_{{\mathcal{H}}, \alpha} B$ or $A \in C^O$. In the former case, the fact that ${\mathcal{T}}$ is valid ensures $TOP_{\mathcal{T}}(B)$ is not an ancestor of $TOP_{\mathcal{T}}(A)$. For the latter case, assume for contradiction that there exist $A, B$ such that $TOP_{\mathcal{T}}(B)$ is an ancestor of $TOP_{\mathcal{T}}(A)$, $A <_C B$, and $A \in C^O$.
We first claim that, without loss of generality, we can suppose that $A$ and $B$ have different operators. To do so, we show that if $A$ and $B$ have the same operator, there must exist a $B'$ with a different operator such that $TOP_{\mathcal{T}}(B')$ is an ancestor of $TOP_{\mathcal{T}}(A)$ and $A <_C B'$. By the definition of $C^O$, there must exist some $A' \in C^O$ such that $A' <_{{\mathcal{H}}, \alpha} B$. Since $A', A \in C$, there must exist a path exclusively in $C$ that connects the two. And since $A', A \in C^O$, no attribute along the path precedes either $A$ or $A'$ in $<_{{\mathcal{H}}, \alpha}$. The contrapositive of Lemma \[Pathing\] implies that $A'$ and $A$ must have the same operator, which implies that $A'$ and $B$ have the same operator. Lemmas \[lemma:inbetween\] and \[lemma:ancestor\] imply that there exists some $B'$ with a different operator such that $A' <_{{\mathcal{H}}, \alpha} B' <_{{\mathcal{H}}, \alpha} B$ and $TOP_{\mathcal{T}}(B')$ is an ancestor of $TOP_{\mathcal{T}}(B)$. The former result implies $B' \in C \backslash C^O$ and $A <_C B'$. The latter result implies $TOP_{\mathcal{T}}(B')$ is an ancestor of $TOP_{\mathcal{T}}(A)$.
We now suppose $A$ and $B$ have different operators without loss of generality. Since $A \in C^O$, any $O \in {\mathcal{V}}$ such that $O <_{{\mathcal{H}}, \alpha} A$ must be an output attribute, thereby implying $O <_{{\mathcal{H}}, \alpha} B$ as well. This fact, combined with Lemma \[Pathing\], implies every path between $A$ and $B$ must contain some $D$ such that $D <_{{\mathcal{H}}, \alpha} B$. Since ${\mathcal{T}}$ is valid and TOP-semiunique, the $TOP_{\mathcal{T}}(D)$ for each of these $D$ cannot be in the subtree rooted at $TOP_{\mathcal{T}}(B)$. This implies that $A$ and $B$ are disconnected in the subtree rooted at $TOP_{\mathcal{T}}(B)$, contradicting the subtree-connected property.
For an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$, suppose ${\mathcal{H}}_0, \dots, {\mathcal{H}}_k$ are the characteristic hypergraphs $H({\mathcal{H}}, \alpha)$. Then GHDs $G_0, G_1, \dots, G_k$ of ${\mathcal{H}}_0, \dots, {\mathcal{H}}_k$ can be connected to form a decomposable GHD $G$ for $Q_{{\mathcal{H}}, \alpha}$. Conversely, any decomposable GHD $G$ of $Q_{{\mathcal{H}}, \alpha}$ can be partitioned into GHDs $G_0, G_1, \dots, G_k$ of the characteristic hypergraphs ${\mathcal{H}}_0, \dots, {\mathcal{H}}_k$. Moreover, in both of these cases, $\gamma\text{-width}(G) = \max_{i} \gamma\text{-width}(G_i)$.
Proof by induction on $|\alpha|$. Our base case is $|\alpha| = 0$. In this case, the only characteristic hypergraph is the input hypergraph, that is $H({\mathcal{H}}, \alpha) = {\mathcal{H}}$. The theorem is then trivially true.
Suppose $|\alpha > 0|$. Any decomposable GHD $G$ for $Q_{{\mathcal{H}}, \alpha}$ must be decomposable into subtrees ${\mathcal{T}}_0, \dots, {\mathcal{T}}_l$ such that $\chi({\mathcal{T}}_0) = V(-\alpha)$ and ${\mathcal{T}}_i$ is a decomposable GHD for $Q_{(\cup_{E \in {\mathcal{E}}_{C_i}} E, {\mathcal{E}}_{C_i}), \alpha_{C_i \backslash C_i^O}}$ where $C_i$ is the $i^{th}$ connected component of $H \backslash V(-\alpha)$. Define ${\mathcal{V}}_{C_i}$ to be $\cup_{E \in {\mathcal{E}}_{C_i}} E$. To preserve the running intersection property of a GHD, the root of ${\mathcal{T}}_i$ and its parent (in ${\mathcal{T}}_0$) must contain the attributes ${\mathcal{V}}_{C_i} \cap V(-\alpha)$. This implies each of the ${\mathcal{T}}_i$ are decomposable GHDs of $Q_{{\mathcal{H}}_i^+, \alpha_{C_i \backslash C_i^O}}$, where ${\mathcal{H}}_i^+$ is the hypergraph defined in the definition of the characteristic hypergraphs. By the inductive hypothesis, the ${\mathcal{T}}_i$ (for $i \ge 1$) can be broken down into GHDs $G_1, \dots G_k$ of the characteristic hypergraphs ${\mathcal{H}}_1, \dots, {\mathcal{H}}_k$. In addition, ${\mathcal{T}}_0$ must also have nodes that contain the edge $E \in {\mathcal{E}}$ such that $E \subseteq V(-\alpha)$, implying it is the GHD $G_0$ of the characteristic hypergraph ${\mathcal{H}}_0$.
In the other direction, by the inductive hypothesis, the GHDs $G_1, \dots G_k$ can be stitched together to form ${\mathcal{T}}_1, \dots {\mathcal{T}}_l$ such that each ${\mathcal{T}}_i$ is the decomposable GHD for the [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}}_i^+, \alpha_{C_i \backslash C_i^O}}$. Note that, by definition, for each $i$, ${\mathcal{T}}_i$ and $G_0$ must both have a node containing the attributes ${\mathcal{V}}_{C_i} \cap V(-\alpha)$; let $t_i$ and $g_i$ denote the appropriate node in ${\mathcal{T}}_i$ and $G_0$, respectively. We can re-root ${\mathcal{T}}_i$ at $t_i$ without violating any conditions since it amounts to re-rooting the top-most GHD of its decomposition; re-rooting ${\mathcal{T}}_i$ at $t_i$ can only change the ancestor relationship between $TOP_{\mathcal{T}}$ nodes of output attributes. Once we re-root the ${\mathcal{T}}_i$ appropriately, we can simply set $parent(t_i)$ to be $g_i$ to generate a decomposable GHD for the [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}}, \alpha}$.
Product Aggregations (Detailed version) {#sec:univ-aggregation-proofs}
=======================================
The primary application of queries with multiple aggregations is to establish bounds for the Quantified Conjunctive Query ($QCQ$) problem [@FAQ]. A $QCQ$ query consists of an arbitrary conjunctive query preceded by a series of (existential and/or universal) quantifiers, and a solution must report the satisfying assignments to the non-quantified variables. A $\#QCQ$ query is similar to a $QCQ$ query, but instead of reporting satisfying assignments, we report the number of satisfying assignments.
We now introduce a new type of aggregation, called product aggregation, that lets us efficiently handle $QCQ$ queries. We define the [<span style="font-variant:small-caps;">Ajar</span>]{}problem for product aggregations, and then extend our algorithm from Section \[subsec:simple-solution\] to handle this new type of [<span style="font-variant:small-caps;">Ajar</span>]{}query.
[<span style="font-variant:small-caps;">Ajar</span>]{}queries with product aggregates
-------------------------------------------------------------------------------------
In order to recover $QCQ$ as an [<span style="font-variant:small-caps;">Ajar</span>]{}query, we need [*product aggregations*]{} i.e. aggregations that use the $\otimes$ operator. Throughout the paper, we have assumed that an absent tuple effectively has an annotation of $0$. To maintain this for product aggregations, we need to define product aggregation so that it returns $0$ if any tuple is absent. In particular, we redefine $\sum_{(A, \otimes)}R_F$ to include a projected tuple $t_{F \backslash A}$ in the output only if $(t_{F\backslash A} \circ t_A)$ exists in $R_F$ for *every* possible value $t_{A} \in {\mathcal{D}}^A$. More formally, let $B = F \backslash A$:
$\displaystyle \sum_{(A, \otimes)} R_{AB} = \{(t_B, \lambda): \forall t_A \in {\mathcal{D}}^A, t_B \circ t_A \in R_{AB} \text{ and } \lambda = \prod_{(t, \lambda_t) \in R_{AB}: \pi_B t = t_B} \lambda_t \}$
Note that this adjusted definition implies an annotation of $0$ is once again fully equivalent to absence. We can adjust the definition of aggregation orderings (and ${\textsc{Ajar}\xspace}$ queries) to possibly include this new type of aggregation. We can construct valid GHDs for such aggregations as before, and run AggroGHDJoin to solve them.
Consider the semiring $(\{0,1\}, \max, \cdot)$. Note that in this domain $\max$ is equivalent to a disjunction (and the logical existential quantifier) and $\prod$ is equivalent to a conjunction (and the logical universal quantifier). Thus the space of [<span style="font-variant:small-caps;">Ajar</span>]{}queries that use these two aggregators recover all $QCQ$ queries.
An aggregation using $\otimes$ is called a [*product aggregation*]{}, and an attribute that is aggregated using a product aggregation is called a [*product attribute*]{}. Aggregations that are not product aggregations are called [*semiring aggregations*]{}, while attributes that are neither output attributes nor product attributes are called [*semiring attributes*]{}.
[**Idempotence Assumption:**]{} Using the product aggregation as defined raises one issue. Our semiring aggregates satisfy the distributive property, which is integral in our ability to push-down aggregations and for our results about commuting aggregations (Theorem \[commute\]). In general, product aggregations do not distribute: $(a \otimes b) \otimes (a \otimes c) = (a \otimes a) \otimes (b \otimes c) \neq a \otimes (b \otimes c)$. However if we require our product aggregations to be idempotent, that is that $a \otimes a = a$ for any element $a$, our product aggregations will distribute. And for $QCQ$, the domain is restricted to $\{0,1\}$, in which product aggregations are idempotent. So in this section, we will study idempotent product aggregations; we will generalize to non-idempotent aggregations in Appendix \[subsec:non-idempotent-product\].
Solving [<span style="font-variant:small-caps;">Ajar</span>]{}queries with product aggregates
---------------------------------------------------------------------------------------------
For aggregation orderings that have product aggregations, the rules for determining when two orderings are equivalent are somewhat different. We now discuss how we can optimize this new type of aggregation further; product aggregations are fundamentally different from ordinary aggregation because we can do the aggregation *before* the join, as seen in the following example:
In the semiring $(\{0,1\}, \max, \cdot)$, suppose we have two relations $R(A,B) = \{((0,0), x), ((0,1), y)\}$ and $S(B,C) = \{((0, 1), p), ((1,1),q)\}$. Consider the [<span style="font-variant:small-caps;">Ajar</span>]{}query $\sum_{(B, \cdot)} R(A,B) \Join S(B,C)$. If compute the join, we will get two tuples with the annotations $x \cdot p$ and $y \cdot q$, and then aggregating over $B$ will produce a relation with the element $((0,1), x \cdot p \cdot y \cdot q)$. However, note that $x \cdot p \cdot y \cdot q = (x \cdot y) \cdot (p \cdot q)$, implying that $\sum_{(B, \cdot)} R(A,B) \Join S(B,C) = (\sum_{(B, \cdot)} R(A,B)) \Join (\sum_{(B, \cdot)} S(B,C))$.
Now we describe our algorithm for solving [<span style="font-variant:small-caps;">Ajar</span>]{}queries when product aggregations are present. Our algorithm follows the same lines as the algorithm from Section \[subsec:simple-solution\]. Recall that the algorithm consisted of searching for [*equivalent orderings*]{}, then searching for GHD [*compatible*]{} with an equivalent ordering, and running AggroGHDJoin on the GHD with the smallest fhw. For product aggregations, we need to modify our algorithm for testing equivalent orderings, and our definition of compatibility; we do these in turn.
#### Testing orderings for equivalence {#testing-orderings-for-equivalence-1 .unnumbered}
Algorithm \[algo:equivalance-test-prod\] gives the pseudo-code for our equivalence test for orderings containing product aggregates.
**Input:** Query hypergraph ${\mathcal{H}}$, orderings $\alpha$, $\beta$.\
**Output:** True if $\alpha \equiv_{{\mathcal{H}}} \beta$, False otherwise.
True Remove $V(-\alpha)$ from ${\mathcal{H}}$, then divide ${\mathcal{H}}$ into connected components $C_1,\ldots C_m$. $\land_{i} \text{TestEquivalence}({\mathcal{H}}, \alpha_{C_i}, \beta_{C_i})$ Choose $j$ such that $\beta_j = \alpha_1$. Let $\beta_j = (b_j, \oplus'_j)$. False Let $\beta'$ be $\beta$ with $\beta_j$ removed. Let $\alpha'$ be $\alpha$ with $\alpha_1$ removed. $\text{TestEquivalence}({\mathcal{H}}, \alpha', \beta')$
We have the lemma analogous to Lemma \[lemma:equivalence-test-sound-complete\].
Algorithm \[algo:equivalance-test-prod\] returns True if and only if $\alpha \equiv_{\mathcal{H}}\beta$.
[**Soundness:**]{} Suppose Algorithm \[algo:equivalance-test-prod\] returns true; we will show $\alpha \equiv_{\mathcal{H}}\beta$. We induct on the length $|\alpha|$. For our base case, when $|\alpha|=0$, we return true when $|\beta|=0$. In this case, the two (empty) orderings are trivially equivalent.
Suppose $|\alpha| > 0$. We have two cases: when ${\mathcal{H}}\backslash (V(-\alpha) \cup {\textsf{PA}}(\alpha))$ has one component and when it has multiple components. We first consider the multiple components case. Let the components be $C_1, \dots, C_m$. Then we define $C'_1, \dots, C'_m$ as in the algorithm i.e. For $1 < i < m$, let ${\mathcal{E}}_i$ be $\{E \in {\mathcal{E}}| E \cap C_i \neq \emptyset\}$, the elements of ${\mathcal{E}}$ that intersect with $C_i$. Then $C'_i = C_i \cup \bigcup_{E \in {\mathcal{E}}_i} E \cap {\textsf{PA}}(\alpha)$. We define ${\mathcal{E}}_0$ to be ${\mathcal{E}}\backslash (\bigcup_{1 \le i m} {\mathcal{E}}_i)$ (these are relations with only output attributes or product aggregations). Accordingly let $C'_0$ be the product aggregations that appear in ${\mathcal{E}}_0$. We can then express the following identities: $$\begin{aligned}
\Join_{F \in {\mathcal{E}}} R_F &= \Join_{0 \le i \le m} \Join_{F \in {\mathcal{E}}_i} R_F\\
\sum_{\alpha} \Join_{F \in {\mathcal{E}}} R_F &= \Join_{0 \le i \le m} \sum_{\alpha_{C'_i}} \Join_{F \in {\mathcal{E}}_i} R_F\end{aligned}$$ The RHS may have a product aggregation $(a,\otimes)$ happening in multiple components, but it happens exactly once per relation containing $a$. We note this identity holds for $\beta$ as well. This identity implies that $\alpha \equiv_{\mathcal{H}}\beta$ if $\alpha_{C'_i} \equiv_{\mathcal{H}}\beta_{C'_i}$ for all $i$. We note that for $i=0$, all of the aggregations contain the same operator, so any ordering is equivalent. For $i>0$, we note that we return true only if all of the recursive calls return true, implying $\alpha_{C'_i} \equiv_{\mathcal{H}}\beta_{C'_i}$ by the inductive hypothesis.
When ${\mathcal{H}}\backslash (V(-\alpha) \cup {\textsf{PA}}(\alpha))$ has one component, we choose $j$ such that $\beta_j = \alpha_1$ and define $\beta'$ to be $\beta$ with $\beta_j$ removed. Note $\alpha'$ is defined to be $\alpha$ with $\alpha_1 = \beta_j$ removed. To show $\beta \equiv_{\mathcal{H}}\alpha$, we need to show $\alpha' \equiv_{\mathcal{H}}\beta'$ and $\beta \equiv_{\mathcal{H}}\beta_j \beta'$. Since we return true only when our recursive call on $\alpha'$ and $\beta'$ returns true, the former equivalence holds by the inductive hypothesis.
To show $\beta \equiv_{\mathcal{H}}\beta_j \beta'$, we ensure $\beta_j$ and $\beta_i$ can commute for all $i<j$. More specifically, we ensure that if $\beta_j$ can be moved to index $i+1$, it can be moved to index $i$. For any $\beta_i$ with the same operator, $\beta_i$ and $\beta_j$ trivially commute. If $\beta_i$ has a different operator, we know there is no path between their attributes $b_i$ and $b_j$ among the nodes $$(\{b_i,b_{i+1},\ldots,b_{|\alpha|}\} \setminus {\textsf{PA}}(\alpha)) \cup \{b_i,b_j\}.$$ Let $V$ be this set of attributes. Define $V_1 \subset V$ to be the set of nodes connected to $b_i$ in the hypergraph restricted to $V$ (we know $b_j \notin V_1$). Let ${\mathcal{E}}_1$ be the set of edges that contain some attribute in $V_1$, i.e. $\{ E \in {\mathcal{E}}| E \cap V_1 \neq \emptyset\}$. We note that the attributes of $V \backslash V_1$ do not appear in the edges of ${\mathcal{E}}_1$. Let ${\mathcal{E}}_2 = {\mathcal{E}}\backslash {\mathcal{E}}_1$; the attributes of $V \backslash V_1$ all appear in ${\mathcal{E}}_2$. We can then express the following identities: $$\begin{aligned}
\Join_{F \in {\mathcal{E}}} R_F = &(\Join_{F \in {\mathcal{E}}_1} R_F) \Join (\Join_{F \in {\mathcal{E}}_2} R_F)\\
\sum_{\beta_{V \cup {\textsf{PA}}(\alpha)}} \Join_{F \in {\mathcal{E}}} R_F = &\left(\sum_{\beta_{V_1 \cup {\textsf{PA}}(\alpha)}} \Join_{F \in {\mathcal{E}}_1} R_F \right) \Join \\
&\left(\sum_{\beta_{V_2 \cup {\textsf{PA}}(\alpha)}} \Join_{F \in {\mathcal{E}}_2} R_F\right)\end{aligned}$$ We note, by definition, that $\beta_i$ and $\beta_j$ must be pushed down into different aggregations in the previous expression. This implies that we can commute $\beta_i$ and $\beta_j$ when they are adjacent, completing the soundness proof.
[**Completeness:**]{} We prove that if Algorithm \[algo:equivalance-test-prod\] returns false, then there must exist a database instance $I$ such that $Q_{{\mathcal{H}}, \alpha}(I) \neq Q_{{\mathcal{H}}, \beta}(I)$.
If Algorithm \[algo:equivalance-test-prod\] returns false, there must be a component $C'$, $\alpha' = \alpha_{C'}$, $\beta' = \beta_{C'}$, such that $\beta_j = \alpha_1$, and there exists a $i < j$ such that $\beta_i = (b_i, \odot'_i)$, $\beta_j = (b_j, \odot'_j)$, $\odot'_i \neq \odot'_j$ and there is a path from $b_i$ to $b_j$ that consists of only $b_i$, $b_j$, and semiring attributes in $\{b_i, b_{i+1},\ldots,b_{|\alpha'|}\}$. We now define our instance $I$ that gives different outputs on these orderings.
If neither $\odot'_i$ nor $\odot'_j$ are product operators, then choose $x$, $y$ such that $x \odot'_i y \neq x \odot'_j y$. If one of them is a product operator while the other is not, choose $x = y = 1$. Now we define the attribute domains. Let $B$ be the set of attributes in the path from $b_i$ to $b_j$ consisting of $b_i$, $b_j$ and semiring attributes in $\{b_i, b_{i+1},\ldots,b_{|\alpha'|}\}$. For every $b \in B$, we set ${\mathcal{D}}^b = \{0,1\}$. For every $b' \notin B$, we set its ${\mathcal{D}}^{b'}$ to $\{0\}$. In every relation that has at least one attribute from $B$, it has two tuples. One tuple has value $0$ for all attributes in $B$, the other has value $1$ for all attributes in $B$. The values of the other attributes are of course always $0$. One of the relations containing a attribute from $B$ has annotation $x$ for the tuple with $0$s and annotation $y$ for the tuple with $1$s. All other annotations are $1$.
Clearly, each aggregation for an attribute $b' \notin B$ is a no-op, since the domain size $|{\mathcal{D}}^{b'}| = 1$. Moreover, all aggregations other than $\beta_i$, $\beta_j$ in $\beta$ and $\alpha$ are also no-ops, because they are non-product aggregations (from the way we chose $B$) and there is a unique value of the attribute for each tuple it maps to after aggregation.
Thus if both $\beta_i$ and $\beta_j$ are non-product aggregations themselves, then we have $Q_{{\mathcal{H}}, \alpha}(I) = x \odot'_j y$, $Q_{{\mathcal{H}}, \beta}(I) = x \odot'_i y$ which are unequal due to how we chose $x$ and $y$. If one of them, say $\beta_j$ is a product aggregation, then $Q_{{\mathcal{H}}, \alpha}(I) = 1$ while $Q_{{\mathcal{H}}, \beta}(I) = 0$ (and vice versa if $\beta_i$ is a product aggregation). This is because in $\beta$, when we do the product aggregation $\beta_j$, there is only one value of $b_j$ per corresponding output value, so the product annotation is $0$ (and finally the $\beta_i$ aggregation adds two $0$’s to get $0$). On the other hand, for $\alpha$, $\beta_j = \alpha_1$ happens when $b_j$ has two values $0$, $1$ corresponding to a single output tuple, so their annotations are multiplied to get $x \otimes y = 1$. This shows that Algorithm \[algo:equivalance-test-prod\] is complete.
#### Compatible GHDs {#compatible-ghds-1 .unnumbered}
Product aggregations not only change the set of equivalent orderings, but also the set of GHDs compatible with a given ordering. In fact, product aggregations allow us to break the rules of GHDs without causing incorrect behavior. In particular, we can have a product attribute $P$ appear in completely disparate parts of the GHD. Thus before defining compatibility for GHDs, we define the notion of [*product partitions*]{}.
Given a hypergraph ${\mathcal{H}}= ({\mathcal{V}}, {\mathcal{E}})$ and aggregation ordering $\alpha$, let $S = \{a \in {\mathcal{V}}| (a, \otimes) \in \alpha\}$ be the set of attributes with product aggregations. A *product partition* is a set $\{P_a | a \in S\}$ where $P_a$ is a partition of $\{F \in {\mathcal{E}}| a \in F\}$ (the relations that contain $a$).
We will duplicate each attribute $a$ for each partition of $P_a$ and have the partition specify which edges contain each instance of $a$.
Suppose we are given a hypergraph ${\mathcal{H}}= ({\mathcal{V}}, {\mathcal{E}})$, aggregation ordering $\alpha$, and product partition $P$. The *product partition hypergraph* ${\mathcal{H}}_P$ is the pair $({\mathcal{V}}_P, {\mathcal{E}}_P)$ such that
- $S = \{a \in {\mathcal{V}}| P_a \in P\}$
- ${\mathcal{V}}_P = \left(\bigcup_{a \in S} \{a_1, a_2, \dots, a_{|P_a|}\} \right) \cup {\mathcal{V}}\backslash S$
- $p: {\mathcal{V}}\times {\mathcal{E}}\to {\mathcal{V}}_P \text{ where } p(a, F) = a \text{ if } a \notin S \text{ otherwise }$\
$a_i \text{ where } F \text{ is in } i^{th} \text{ partition of } P_a$
- ${\mathcal{E}}_P = \bigcup_{F \in {\mathcal{E}}} \{ p(a, F) | a \in F \}$
Given a hypergraph ${\mathcal{H}}$ and aggregation ordering $\alpha$, an *aggregating generalized hypertree decomposition* (AGHD) is a triple $({\mathcal{T}}, \chi, P)$ such that $({\mathcal{T}}, \chi)$ is a GHD of the product partition hypergraph ${\mathcal{H}}_P$.
For any attribute $a$ in the [<span style="font-variant:small-caps;">Ajar</span>]{}query, $TOP_{\mathcal{T}}(a)$ for an AGHD $({\mathcal{T}}, \chi, P)$ can be defined as the set $\{TOP_{\mathcal{T}}(a_1)$, $TOP_{\mathcal{T}}(a_2)$,$\ldots$,$TOP_{\mathcal{T}}(a_{|P_a|})\}$. Now we can define the notion of compatibility of an AGHD, with an ordering.
A AGHD $({\mathcal{T}}, \chi, P)$ for an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$ is compatible with an ordering $\beta \equiv_{\mathcal{H}}\alpha$ if for each attribute pair $a$, $b$ for which there exists $v_1 \in TOP_{\mathcal{T}}(a)$, $v_2 \in TOP_{\mathcal{T}}(b)$ such that $v_1$ is an ancestor of $v_2$, $a$ must occur before $b$ in the ordering $\beta$.
#### Solving [<span style="font-variant:small-caps;">Ajar</span>]{}queries with product aggregates {#solving-ajarqueries-with-product-aggregates-1 .unnumbered}
In our proofs and discussions for the remainder of this section, we will treat the set of $TOP_{\mathcal{T}}$ as a single element for convenience, implicitly placing an existential quantifier before the statement. For example, when we say $TOP_{\mathcal{T}}(A)$ is an ancestor of $TOP_{\mathcal{T}}(B)$, we mean $\exists t_A \in TOP_{\mathcal{T}}(A), t_B \in TOP_{\mathcal{T}}(B)$ such that $t_A$ is an ancestor of $t_B$. We also often omit the partition $P$ when referring to an AGHD $G = ({\mathcal{T}}, \chi, P)$; the partition $P$ can be uniquely defined by $({\mathcal{T}},\chi)$, so we will always assume it is defined appropriately.
We can now modify our algorithm from Section \[subsec:simple-solution\] to detect equivalent orderings using Algorithm \[algo:equivalance-test-prod\], then search for compatible AGHDs, and run AggroGHDJoin over the compatible AGHD with the smallest fhw. Our runtime is given by the next theorem. Note that any AGHD of the original hypergraph is also a GHD of some product partition hypergraph.
Given a [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}},\alpha}$ possibly involving idempotent product aggregates, let $w^*$ be the smallest fhw for an AGHD compatible with an ordering equivalent to $\alpha$. Then the runtime for our algorithm is ${\widetilde{O}}({\textsc{IN}}^{w^*} + {\textsc{OUT}})$.
The theorem is proved in Appendix \[sec:background\].
#### Decomposing AGHDs {#decomposing-aghds-1 .unnumbered}
We can apply the ideas from Section \[sec:decomposing\] to [<span style="font-variant:small-caps;">Ajar</span>]{}queries with product aggregates as well. In this section we will assume without loss of generality that for any relation $R_F$, the last aggregation in $\alpha_F$ is not a product aggregation. Suppose this assumption is violated, i.e. there exists some relation $R_F$ such that the last aggregation in $\alpha_F$ is the product aggregation $(A_P, \otimes)$. We can then immediately perform this aggregation, transforming the relation to $R_{F \backslash \{A_P\}}$ and removing the product aggregation. This assumption ensures that every relation appears in one of the subtrees in the decomposition defined below. We now define some terms.
Given an [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}}, \alpha}$, suppose we have a subset of the nodes $V \in {\mathcal{V}}$. Define ${\mathcal{E}}_V$ to be $\{E \in {\mathcal{E}}| E \cap V \neq \emptyset\}$, i.e. the set of edges that intersect with $V$. Additionally, define $\alpha_{-[i]}$ to be $\alpha$ with the first $i$ elements removed. We will be looking at the connected components of ${\mathcal{H}}\backslash (V_{-\alpha} \cup {\textsf{PA}}(\alpha))$. For any connected component $C$, let $C^+ = C \cup \{ v \in {\textsf{PA}}(\alpha) | \exists E \in {\mathcal{E}}_C : v \in E \}$. Additionally, given an ordering $\alpha$, we define $\alpha^O$ based on a conditional: if $\alpha_1$ is a product aggregation, let $\alpha^O$ be just $\alpha_1$; if $\alpha_1$ is not a product aggregations, let $\alpha^O$ be the set of attributes that can be commuted to the beginning of the ordering. To be more precise for this second case, given an attribute $A$ that appears in $\alpha_j$ with operator $\odot$, $A \in \alpha^O$ if for all $\alpha_i = (B, \odot')$ such that $i<j$ either $\odot' = \odot$ or $A$ and $B$ are not connected among the nodes $(\alpha_{-[i-1]} \backslash {\textsf{PA}}(\alpha_{-[i-1]}) \cup \{A, B\}$.
\[def:decomp2\] Given an [<span style="font-variant:small-caps;">Ajar</span>]{}query $Q_{{\mathcal{H}}, \alpha}$, we say an AGHD $({\mathcal{T}}, \chi, P)$ is *decomposable* if:
- There exists a rooted subtree ${\mathcal{T}}_0$ of ${\mathcal{T}}$ such that $\chi({\mathcal{T}}_0) = {\mathcal{V}}(-\alpha)$ (i.e. output attributes).
- For each connected component $C$ of ${\mathcal{H}}\backslash (V_{-\alpha} \cup {\textsf{PA}}(\alpha))$, there is exactly one subtree ${\mathcal{T}}_{C} \in {\mathcal{T}}\backslash {\mathcal{T}}_0$ such that ${\mathcal{T}}_C$ is a decomposable AGHD of $Q_{(\cup_{E \in {\mathcal{E}}_C} E, {\mathcal{E}}_C), \alpha_{C^+ \backslash \alpha_{C^+}^O}}$.
Then we have theorems analogous to theorems \[thm:decompisvalid\], \[thm:decompwidth\], and \[thm:decomp\].
All decomposable AGHDs are compatible with an ordering $\beta$ such that $\beta \equiv_{\mathcal{H}}\alpha$.
Suppose we are given an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$ and a decomposable AGHD $G$ for this query. We show a stronger statement: all decomposable AGHDs are compatible with an ordering $\beta$ such that $\beta \equiv_{\mathcal{H}}\alpha$ and ${\textsf{PA}}(\beta) = {\textsf{PA}}(\alpha)$ (i.e. the order of the product attributes does not change). Proof by induction on $|\alpha|$. When $|\alpha| = 0$, all GHDs are decomposable and all GHDs are compatible with $\alpha$.
Suppose $|\alpha| > 0$. By definition, there is a subtree ${\mathcal{T}}_0$ of $G$ such that $\chi({\mathcal{T}}_0) = V(-\alpha)$. And for each connected component $C$ of ${\mathcal{H}}\backslash (V(-\alpha) \cup {\textsf{PA}}(\alpha))$, we have a subtree ${\mathcal{T}}_C$ that is a decomposable GHD for the query $Q_{(\cup_{E \in {\mathcal{E}}_C} E, {\mathcal{E}}_C), \alpha_{C^+ \backslash \alpha_{C^+}^O}}$. We will use ${\mathcal{V}}_C$ to denote $\cup_{E \in {\mathcal{E}}_C} E$ and ${\mathcal{H}}_C$ to denote $({\mathcal{V}}_C, {\mathcal{E}}_C)$. Similarly, we will use $\alpha^C$ to represent $\alpha_{C^+ \backslash \alpha_{C^+}^O}$. By the inductive hypothesis, each of these subtrees ${\mathcal{T}}_C$ is compatible with some ordering $\beta^C$ such that $\beta^C \equiv_{{\mathcal{H}}_C} \alpha^C$ and ${\textsf{PA}}(\beta^C) = {\textsf{PA}}(\alpha^C)$. Note that $\beta^C \equiv_{{\mathcal{H}}_C} \alpha^C$ trivially implies $\beta^C \equiv_{\mathcal{H}}\alpha^C$.
For each $C$ we will construct a $\beta^C+$ such ${\mathcal{T}}_C$ is compatible with $\beta^C+$, $\beta^C+ \equiv_{\mathcal{H}}\alpha_{C^+}$, and ${\textsf{PA}}(\beta^C+) = {\textsf{PA}}(\alpha_{C^+})$. Since $\alpha^C = \alpha_{C^+ \backslash alpha_{C^+}^O}$, this requires adding the elements of $\alpha_{C^+}^O$ to $\beta^C$. Define $\beta^O$ to be some ordering of the elements compatible with $G$ (i.e. for any $A,B \in V(\alpha_{C^+}^O)$ if $TOP_{{\mathcal{T}}_C}(A)$ is an ancestor of $TOP_{{\mathcal{T}}_C}(B)$, $A$ precedes $B$ in $\beta^O$). We claim the ordering $\beta^C+ = \beta^O \circ \beta^C$ satisfies our three conditions.
The first condition is that ${\mathcal{T}}_C$ is compatible with this $\beta^C+$. This is trivially true because we constructed the ordering by adding output attributes to the start of $\beta^C$, with which ${\mathcal{T}}_C$ is already compatible, in an order that is guaranteed to be compatible.
The second condition is that $\beta^C+ \equiv_{\mathcal{H}}\alpha_{C^+}$. By the definition of $\alpha_{C^+}^O$, $\alpha_{C^+} \equiv_{\mathcal{H}}\alpha_{C^+}^O \circ \alpha^C$. We know $\beta^C \equiv_{{\mathcal{H}}} \alpha^C$ by the inductive hypothesis. And we claim $\beta^O \equiv_{\mathcal{H}}\alpha_{C^+}^O$, which implies $\beta^C+ \equiv_{\mathcal{H}}\alpha_{C^+}$ by definition. We show this claim by showing that the operators of $\alpha_{C^+}^O$ are uniform, implying that its elements can be reordered freely. In particular, consider the first element $(A_1, \odot_1)$ of $\alpha_{C^+}$. Since $C^+$ is a connected component, there must exist a path between $A_1$ and every other node among the nodes $C^+$. Thus, for any $(B, \odot') \in \alpha_{C^+}$ such that $\odot' \neq \odot_1$, $A_1$ will violate the path condition for commuting and ensure $B \notin V(\alpha_{C^+}^O)$.
The third condition is that ${\textsf{PA}}(\beta^C+) = {\textsf{PA}}(\alpha_{C^+})$. By the inductive hypothesis, ${\textsf{PA}}(\beta^C) = {\textsf{PA}}(\alpha^C)$. We simply need to show ${\textsf{PA}}(\beta_{C^O}) = {\textsf{PA}}(\alpha_{C^+}^O)$. There are two cases to consider, from the definition of $\alpha_{C^+}^O$. In the first case, both $\beta_{C^O})$ and $\alpha_{C^+}^O$ contains only one (product) aggregation. In the second case, the two orderings have no product aggregations. In either case, ${\textsf{PA}}(\beta_{C^O}) = {\textsf{PA}}(\alpha_{C^+}^O)$ trivially.
We now need to combine the $\beta^C+$ for each $C$ to construct the desired ordering $\beta$ as desired. We construct $\beta$ by repeating the two following steps algorithm until every $\beta^C$ is empty: $(1)$ remove the non-product output prefixes of $\beta^C+$ and append them to $\beta$ (interleaved arbitrarily) and $(2)$ remove the earliest remaining product aggregation of ${\textsf{PA}}(\alpha)$ from the start of the appropriate $\beta^C+$ and append it to $\beta$. Note that this procedure ensures $\beta_{C^+} = \beta^C+$ for each $C$, which implies $\beta_{C^+} \equiv_{\mathcal{H}}\alpha_{C^+}$ and (by the soundness of Algorithm \[algo:equivalance-test-prod\]) $\beta \equiv_{\mathcal{H}}\alpha$. Also note that the procedure preserves the ordering of the product aggregates, so ${\textsf{PA}}(\beta) = {\textsf{PA}}(\alpha)$. Finally, the given AGHD $G$ must be compatible with $\beta$. The construction of $G$ ensure the top nodes of output attributes are all above the top nodes of non-output attributes, and the top nodes of non-output attributes are in the subtrees ${\mathcal{T}}_C$, which means the fact that $\beta_{C^+} = \beta^C+$ ensures these top nodes are ordered in a compatible manner.
For every valid AGHD $({\mathcal{T}}, \chi)$, there exists a decomposable $({\mathcal{T}}', \chi')$ such that for all node-monotone functions $\gamma$, the $\gamma$-width of $({\mathcal{T}}', \chi')$ is no larger than the $\gamma$-width of $({\mathcal{T}}, \chi)$.
We first modify the definition of subtree-connected from Appendix \[sec:app-decomp\]:
- *subtree-connected*: for any node $t \in {\mathcal{T}}$ and the subtree ${\mathcal{T}}_t$ rooted at $t$, consider the set the attributes $V_t = \{v \in {\mathcal{V}}| TOP_{\mathcal{T}}(v) \in {\mathcal{T}}_t\}$; we require for any two attributes $A,B \in V_t$, there exists a path from $A$ to $B$ in the set $(V_t \backslash {\textsf{PA}}(\alpha)) \cup \{A,B\}$.
This same transformation described Lemma \[lemma:decomp-prop2\] can be used for this adjusted definition. Note that this transformation ensures that any node that is $TOP_{\mathcal{T}}$ for a product aggregation has only one child. Also note that the described transformation might change the partition function $P$ of the AGHD, but it does not change the compatible order.
Suppose the given [<span style="font-variant:small-caps;">Ajar</span>]{}problem is $Q_{{\mathcal{H}}, \alpha}$. Since $({\mathcal{T}}, \chi)$ is valid, there must exist an ordering $\beta$ such that $({\mathcal{T}}, \chi)$ is compatible with $\beta$ and $alpha \equiv_{\mathcal{H}}\beta$. The width-preserving transformations of Appendix \[sec:app-decomp\] preserve the compatibility with an ordering. So we can apply them to get a TOP-unique and subtree-connected AGHD $({\mathcal{T}}', \chi')$ that is compatible with $\beta$ and has $\gamma$-width no larger than that of $({\mathcal{T}}, \chi)$. We claim that this AGHD is decomposable.
As in Appendix \[sec:app-decomp\], we prove that any valid, TOP-semiunique, and subtree-connected GHD for an is decomposable. Proof by induction on $|\alpha|$. If $|\alpha = 0|$, then every GHD is decomposable.
Suppose $|\alpha| > 0$. Consider the set of nodes that are $TOP_{\mathcal{T}}$ nodes for output attributes, i.e. $\{t \in {\mathcal{T}}| \exists A \in V(-\alpha): TOP_{\mathcal{T}}(A) = t\}$. Since $({\mathcal{T}}', \chi')$ is compatible with $\beta$, no non-output attributes can have a top node above an output attributes top node. Thus, the TOP-semiunique property guarantees that this set of nodes forms a rooted subtree ${\mathcal{T}}_0$ of ${\mathcal{T}}$ such that $\chi({\mathcal{T}}_0) = V(-\alpha)$.
Consider the subtrees in ${\mathcal{T}}\backslash {\mathcal{T}}_0$. Call them ${\mathcal{T}}_1, {\mathcal{T}}_2, \dots, {\mathcal{T}}_k$. For any ${\mathcal{T}}_i$, let ${\mathcal{V}}_i$ be the attributes that have $TOP_{\mathcal{T}}$ nodes in ${\mathcal{T}}_i$, i.e. ${\mathcal{V}}_i = \{A \in {\mathcal{V}}| TOP_{\mathcal{T}}(A) \in {\mathcal{T}}_i\}$. None of these ${\mathcal{V}}_i$ can contain any output attributes, and connected-subtree guarantees that each of the ${\mathcal{V}}_i$ are connected. Thus, the ${\mathcal{V}}_i$ must be the $C_i^+$ as defined earlier. So for each connected component $C$ of ${\mathcal{H}}\backslash (V(-\alpha) \cup {\textsf{PA}}(\alpha))$, the corresponding subtree ${\mathcal{T}}_C$ is the subtree ${\mathcal{T}}_i$ such that ${\mathcal{V}}_i = C^+$. Since for any $A \in C$, $TOP_{\mathcal{T}}(A) \in {\mathcal{T}}_C$, the attributes in $C$ only appear in ${\mathcal{T}}_C$. Note that for every edge $E \in {\mathcal{E}}$, there exists a node $t \in {\mathcal{T}}$ such that $E \subseteq \chi(t)$. This implies that for every edge $E \in {\mathcal{E}}_C$, there exists a node $t \in {\mathcal{T}}_C$ such that $E \subseteq \chi(t)$. As such, we can conclude that each ${\mathcal{T}}_C$ is a GHD for the hypergraph $(\bigcup_{E \in {\mathcal{E}}_C} E, {\mathcal{E}}_C)$.
Define ${\mathcal{V}}_C = \bigcup_{E \in {\mathcal{E}}_C} E$. To complete this proof, we now need to show that each ${\mathcal{T}}_C$ is a decomposable GHD for the ${\textsc{Ajar}\xspace}$ query $Q_{({\mathcal{V}}_C, {\mathcal{E}}_C), \alpha_{C^+ \backslash \alpha_{C^+}^O}}$. By the inductive hypothesis, if ${\mathcal{T}}_C$ is valid, TOP-semiunique and subtree-connected, it must be decomposable. Note that since ${\mathcal{T}}$ is TOP-semiunique and subtree-connected, ${\mathcal{T}}_C$ must also be TOP-semiunique and subtree-connected. We have also established that ${\mathcal{T}}_C$ is a GHD for $({\mathcal{V}}_C, {\mathcal{E}}_C)$. Thus to finish this proof, we only need to show that there exista an ordering $\beta'$ such that $\beta' \equiv_{({\mathcal{V}}_C, {\mathcal{E}}_C)} \alpha_{C^+ \backslash \alpha_{C^+}^O}$ and ${\mathcal{T}}_C$ is compatible with $\beta'$.
We know ${\mathcal{T}}$ is compatible with $\beta$ and $\beta \equiv \alpha$. We set $\beta' = \beta_{C^+ \backslash \beta_{C^+}^O}$; this implies that $\beta' \equiv_{\mathcal{H}}\alpha_{C^+ \backslash \alpha_{C^+}^O}$ since left hand and right hand sides are simply sub-orderings of $beta$ and $\alpha$, respectively. Furthermore, this implies $\beta' \equiv_{({\mathcal{V}}_C, {\mathcal{E}}_C)} \alpha_{C^+ \backslash \alpha_{C^+}^O}$, as $({\mathcal{V}}_C, {\mathcal{E}}_C)$ is simply ${\mathcal{H}}$ with some output attributes (of $\beta'$) removed.
We now need to show that ${\mathcal{T}}_C$ is compatible with $\beta'$. In other words, we need to show for any two attributes $A,B \in {\mathcal{V}}_C$, if $TOP_{{\mathcal{T}}_C}(A)$ is an ancestor of $TOP_{{\mathcal{T}}_C}(B)$, either $A$ is an output attribute or $A$ precedes $B$ in $\beta'$. We show the contrapositive: if $A$ is not an output attribute and $A$ does not precede $B$ in $\beta'$, then $TOP_{{\mathcal{T}}_C}(A)$ is not an ancestor of $TOP_{{\mathcal{T}}_C}(B)$. There are a couple of cases to consider. If $B \in {\mathcal{V}}_C \backslash C^+$, $B$ must be in $V(-\alpha)$, implying $TOP_{{\mathcal{T}}_C}(B)$ is the root of ${\mathcal{T}}_C$. We note that for attributes in $C^+$, $TOP_{{\mathcal{T}}_C}$ and $TOP_{\mathcal{T}}$ are equivalent, so we use them interchangeably. If $B \in C^+ \backslash \beta_{C^+}^O$, then we know $B$ must precede $A$ in $\beta'$, which implies $B$ precedes $A$ in $\beta$. The fact that ${\mathcal{T}}$ is compatible with $B$ implies $TOP_{\mathcal{T}}(A)$ is not an ancestor of $TOP_{\mathcal{T}}(B)$. The final case to consider is $B \in \beta_{C^+}^O$.
Even in this case, we have two cases to consider, based on the two definitions of $\beta_{C^+}^O$. If $B$ has a product aggregation, then $B$ must be the first element of $\beta_{C^+}$. This implies $B$ precedes $A$ in $\beta$, guaranteeing that $TOP_{\mathcal{T}}(A)$ is not an ancestor of $TOP_{\mathcal{T}}(B)$. The other case is a bit more involved.
Assume for contradiction that there exist $A,B$ such that $TOP_{\mathcal{T}}(A)$ is an ancestor of $TOP_{\mathcal{T}}(B)$, $B \in \beta_{C^+}^O$, and $A \in \beta'$. We first claim that, without loss of generality, we can suppose that $A$ and $B$ have different operators. To do so, we show that if $A$ and $B$ have the same operator, there must exist a $A' \in \beta'$ with a different operator such that $TOP_{\mathcal{T}}(A')$ is an ancestor $TOP_{\mathcal{T}}(B)$. The fact that $A \notin \beta_{C^+}^O$ implies there is an attribute $A'$ with a different operator such that there exists a path between $A'$ and $A$ composed of attributes that appear after $A'$ in $\beta_{C^+}$. We claim $TOP_{\mathcal{T}}(A')$ is an ancestor of $TOP_{\mathcal{T}}(A)$, which implies $TOP_{\mathcal{T}}(A')$ is an ancestor of $TOP_{\mathcal{T}}(B)$. Suppose the path between $A'$ and $A$ is $X_0, X_1, X_2, \dots, X_k$ where $A' = X_0$ and $A = X_k$; we will show $TOP_{\mathcal{T}}(A')$ is an ancestor of $TOP_{\mathcal{T}}(A)$ by showing $TOP_{\mathcal{T}}(A')$ is an ancestor of all $X_i$ for $i \ge 1$. Proof by induction on $i$. For $i=1$, $A'$ and $X_1$ share an edge, implying they appear in $\chi(t)$ together for some tree node $t$. By definition, $TOP_{\mathcal{T}}(A')$ and $TOP_{\mathcal{T}}(X_1)$ are both ancestors of $t$. Since ${\mathcal{T}}$ is TOP-semiunique (so $TOP_{\mathcal{T}}(A') \neq TOP_{\mathcal{T}}(X_1)$) and ${\mathcal{T}}$ is compatible with $\beta$ (so $TOP_{\mathcal{T}}(X_1)$ cannot be an ancestor of $TOP_{\mathcal{T}}(A')$, this means that $TOP_{\mathcal{T}}(A')$ is an ancestor of $TOP_{\mathcal{T}}(X_1)$. For $i > 1$, we know that $X_{i-1}$ and $X_i$ share an edges, implying they appear together in $\chi(t)$ for some tree node $t$. $TOP_{\mathcal{T}}(X_i)$ and $TOP_{\mathcal{T}}(X_{i-1})$ must both ancestors $t$. Note that the inductive hypothesis gives us that $TOP_{\mathcal{T}}(A')$ is an ancestor of $TOP_{\mathcal{T}}(X_{i-1})$, implying it is an ancestor of $t$. By the same logic as before, this implies that $TOP_{\mathcal{T}}(A)$ is an ancestor of $TOP_{\mathcal{T}}(X_i)$. We thus have that $TOP_{\mathcal{T}}(A')$ is an ancestor of $TOP_{\mathcal{T}}(A)$.
We now suppose, without loss of generality, that $A$ and $B$ have different operators. Since $TOP_{\mathcal{T}}(A)$ is an ancestor of $TOP_{\mathcal{T}}(B)$, we know $A$ comes before $B$ in the compatible ordering $\beta$. However, the fact that $B \in \beta_{C^+}^O$ implies that every path between $B$ and $A$ includes an attribute $X$ that is either an output attribute or comes before $B$ in $\beta$. Either way, none of these $X$ is in the subtree rooted at $TOP_{\mathcal{T}}(A)$, implying that $A$ and $B$ are disconnected in the subtree rooted at $TOP_{\mathcal{T}}(A)$. This contradicts the subtree-connected property.
Given an ${\textsc{Ajar}\xspace}$ problem $Q_{{\mathcal{H}}, \alpha}$, suppose $C_1, \dots, C_k$ are the connected components of ${\mathcal{H}}\setminus ({\mathcal{V}}_{-\alpha} \cup {\textsf{PA}}(\alpha))$. Define a function $H({\mathcal{H}}, \alpha)$ that maps ${\textsc{Ajar}\xspace}$ queries to a set of hypergraphs as follows:
- $C_i^{++} = \bigcup_{E \in {\mathcal{E}}_C} E$ for all $1 \le i \le k$
- ${\mathcal{H}}_0 = ({\mathcal{V}}_{-\alpha}, \{F \in {\mathcal{E}}| F \subseteq {\mathcal{V}}_{-\alpha}\} \cup \{{\mathcal{V}}_{-\alpha} \cap C_i^{++} | 1 \le i \le k\})$
- ${\mathcal{H}}_i^+ = (C_i^{++}, {\mathcal{E}}_C \cup \{{\mathcal{V}}_{-\alpha} \cap C_i^+\})$
- $H({\mathcal{H}}, \alpha) = \{{\mathcal{H}}_0\} \cup \bigcup_{1 \le i \le k} H({\mathcal{H}}_i^+, \alpha_{C_i^+ \backslash \alpha_{C_i^+}^O})$
The hypergraphs in the set $H({\mathcal{H}}, \alpha)$ are defined to be the *characteristic hypergraphs*.
For an ${\textsc{Ajar}\xspace}$ query $Q_{{\mathcal{H}}, \alpha}$ involving product aggregates, suppose ${\mathcal{H}}_0, \dots, {\mathcal{H}}_k$ are the characteristic hypergraphs $H({\mathcal{H}}, \alpha)$. Then AGHDs $G_0, G_1, \dots, G_k$ of ${\mathcal{H}}_0, \dots, {\mathcal{H}}_k$ can be connected to form a decomposable AGHD $G$ for $Q_{{\mathcal{H}}, \alpha}$. Conversely, any decomposable AGHD $G$ of $Q_{{\mathcal{H}}, \alpha}$ can be partitioned into AGHDs $G_0, G_1, \dots, G_k$ of the characteristic hypergraphs ${\mathcal{H}}_0, \dots, {\mathcal{H}}_k$. Moreover, in both of these cases, $\gamma\text{-width}(G) = \max_{i} \gamma\text{-width}(G_i)$.
The proof is the exact same as the proof of Theorem \[thm:decomp\] provided in Appendix \[sec:app-decomp\].
This lets us apply all the optimizations from Section \[subsec:optimal-valid\], \[subsec:dbp-width\], and \[subsec:GYM\] to [<span style="font-variant:small-caps;">Ajar</span>]{}queries with product aggregates.
#### Comparison to FAQ {#comparison-to-faq-1 .unnumbered}
The runtime of InsideOut on a query involving idempotent product aggregations is given by ${\widetilde{O}}({\textsc{IN}}^{faqw})$, where the faqw depends on the ordering, and the presence of product aggregations. Our algorithm for handling product aggregations recovers the runtime of FAQ. Formally,
For any [<span style="font-variant:small-caps;">Ajar</span>]{}query involving idempotent product aggregations, ${\textsc{IN}}^{w^*} + {\textsc{OUT}}\leq 2 \cdot {\textsc{IN}}^{faqw}$.
The proof is in Appendix \[subsec:faq-comparison-proof\]. By applying ideas from the FAQ paper to our setting, we can also recover the FAQ runtime on $\#QCQ$ (Appendix \[subsec:recovering-hash-qcq\]). Our algorithm for detecting when two orderings involving product aggregates are equivalent (Algorithm \[algo:equivalance-test-prod\]) is both sound and complete; in contrast, FAQ’s equivalence testing algorithm is sound but not complete. Moreover, we have a width-preserving decomposition for queries with product aggregates. This allows us to apply all the optimizations from Section \[sec:decomposing\], giving us tighter runtimes in terms of submodular and DBP-widths (Theorems \[thm:submodular-width\], \[thm:dbp-width\]) and efficient MapReduce Algorithms (Theorems \[thm:ajar-mapreduce-n-rounds\], \[thm:ajar-mapreduce-log-rounds\]). As shown before, FAQ gives a worse runtime exponent in each of these cases.
Recovering \#QCQ {#subsec:recovering-hash-qcq}
----------------
We discussed idempotent product aggregations and how they can help [<span style="font-variant:small-caps;">Ajar</span>]{}generalize $QCQ$ in Section \[sec:univ-aggregation\]. There is a variant called $\#QCQ$ in which solutions are expected to output the number of solutions to a given $QCQ$ (instead of the solutions themselves). At first this seems like a fairly straightforward extension to $QCQ$. If we use ${\textsc{Ajar}\xspace}$ to solve a given $QCQ$, the output is a relation that lists the satisfying assignments, where each tuple’s annotation is $1$; to count the number of tuples, we simply need to prefix the $QCQ$ query with aggregations using the operator $+$.
An issue arises because these new aggregations need to occur in in the domain $\textbf{Z}_+$ (the non-negative integers) instead of $\{0,1\}$. Though $(\textbf{Z}_+, \max, \cdot)$ is still a semi-ring, the product aggregations are no longer idempotent in the given domain; we discuss how to handle non-idempotent aggregation in Appendix \[subsec:non-idempotent-product\], but the added complexity (and runtime) required to deal with non-idempotent aggregations seems unnecessary in our case. Even though multiplication is not idempotent over the larger domain, we can guarantee that it is idempotent whenever a product aggregation occurs; the annotations do not leave the $\{0,1\}$ domain until the $+$ aggregations, which must occur after the product aggregations.
To handle this extra structure, we introduce the concept of specifying restricted domains in [<span style="font-variant:small-caps;">Ajar</span>]{}queries. To recover $\#QCQ$, we translate the approach of FAQ [@FAQ Section 9.5], which is the minimal application of the restricted domain concept to [<span style="font-variant:small-caps;">Ajar</span>]{}queries.
Given a domain $\mathbb{K}$ and operator set $O$, we define a *restriction* to subsets of the domain $\mathbb{K}_r \subset \mathbb{K}$ and operator set $O_r \subseteq O$ such that $\{0, 1\} \subseteq \mathbb{K}_r, \otimes \in O_r$ and for any $a,b \in \mathbb{K}_r$ and $\odot \in O_r$, $a \odot b \in \mathbb{K}_r$.
In the context of $\#QCQ$, $\mathbb{K} = \textbf{Z}_+$ and $O = \{+, \max, \otimes\}$. The restriction is $\mathbb{K}_r = \{0,1\}$ and $O_r = \{\max, \otimes\}$.
Note that if we ensure that the specified operators are closed in the restricted domain, the semiring properties will all hold in the restricted domain. We then define an aggregation ordering that incorporates these restrictions - we will define an index $l$ divides the unrestricted and restricted portions of the ordering.
Given an attribute set ${\mathcal{V}}$, domain $\mathbb{K}$, operator set $O$, and restriction $\mathbb{K}_r$ and $O_r$, an *restriction-compatible* aggregation ordering is an aggregation ordering $\alpha$ and index $l$ such that $1 \le l \le |\alpha|$ and for each $k \ge l$, $\alpha_k = (A, \odot)$ for $A \in {\mathcal{V}}$ and $\odot \in O_r$.
Any single operator $\odot$ that appears both before and after the division index $l$ will be treated as different operators (this issue does not come up in the context of $\#QCQ$). We can then define an [<span style="font-variant:small-caps;">Ajar</span>]{}query to use a restriction-compatible ordering, and any instance of the query must have $\mathbb{K}_r$-relations. Under this definition, we can treat the product aggregations as idempotent, allowing us to use the work in Section \[sec:univ-aggregation\] to recover $\#QCQ$.
This set-up is essentially a translation of FAQ’s results to our language/notation. Using the exact same construction described in the previous Appendix section, we can now recover FAQ’s runtime on $\#QCQ$ as well. We note that we could extend this idea of restricting domains even further by relying on our GHDs. In particular, we can have every single element of the aggregation ordering specify its own domain, and a valid GHD would have to ensure that for any $A,B$ such that $TOP_{\mathcal{T}}(A)$ is an ancestor of $TOP_{\mathcal{T}}(B)$, the semiring domain corresponding to $A$ is a superset of the semiring domain corresponding to $B$.
Non-Idempotent Product Aggregations {#subsec:non-idempotent-product}
-----------------------------------
Our AggroYannakakis algorithm actually implicitly assumes that any product aggregation that arises consists of an *idempotent* operator.
Given a set $S$, an operator $\oplus$ is *idempotent* if and only if for any element $a \in S$, $a \oplus a = a$.
This is a reasonable assumption, as the problems that we’ve discovered using product aggregation all tend to have idempotent products. The key difference between an idempotent and non-idempotent operator is the distributive property; $(a \otimes b) \otimes (a \otimes c) = a \otimes (b \otimes c)$ only if $\otimes$ is idempotent. Note that the non-idempotent case would require an $a^2$. So, to be complete, we can support non-idempotent operators by raising the annotations of every other relation to a power. In particular, if we have a non-idempotent aggregator for an attribute $A$, we should raise the annotations for the relations in every other node in our tree to the $|{\mathcal{D}}^A|$ power when we aggregate the attribute $A$ away.
Extension: Computing Transitive Closure
=======================================
A standard extension to the basic relational algebra is the transitive closure or Kleene star operator. In this section, we explore how our framework for solving ${\textsc{Ajar}\xspace}$ queries can be applied to computing transitive closures. First we define the operator using the language of ${\textsc{Ajar}\xspace}$. Given a relation $R$ with two attributes, consider the query $$Q_{k} = \sum_{A_2} \cdots \sum_{A_k} \Join_{1 \le i \le k} R(A_i, A_{i+1})$$ where each of $R(A_i, A_{i+1})$ are identical copies of $R$ with the attributes named as specified. Note that our output $Q_k$ is going to be a two-attribute relation. Suppose there exists some $k^*$ such that $Q_k$ is identical for all $k \ge k^*$. We can then define the transitive closure of a relation $R$, denoted $R^*$, to be $Q_{k^*}$.
This classic operator has natural applications in the context of graphs. If our relation $R$ is a list of (directed) edges (without meaningful annotations), computing $R^*$ is equivalent to computing the connected components of our graph. If we add annotations over the semiring $(\mathbb{Z} \cup \{\infty\}, \min, +)$ where each edge is annotated with a weight, then computing $R^*$ is equivalent to computing all pairs shortest paths [@FunSemirings]. Note that we can guarantee $R^*$ exists as long $(i)$ our graph contains no negative weight cycles and $(ii)$ our relation contains self-edges with weight $0$. We will discuss computing $R^*$ in the context of graphs, applying it to the all pairs shortest path problem. Let $E$ be the number of edges and $V$ the number of nodes in the graph; we will derive the complexity of computing all pairs shortest paths in terms of $E$ and $V$.
A naive algorithm for finding $R^*$ is to compute $Q_{1}, Q_{2}, Q_{4}, \dots$ until we find two consecutive results that are identical. This approach requires answering $O(\log k^*)$ ${\textsc{Ajar}\xspace}$ queries. In the context of all pairs shortest path, we know $k^* \le V$, which means that the number of queries to answer is $O(\log V)$. We start by analyzing the computation required to answer a query of the form $Q_{2^n}$.
We define the GHD to use for $Q_{2^n}$ recursively. Our base case, when $n=1$, is to have a single bag containing all three attributes $A_1, A_2, A_3$. For $n>1$, the root of our GHD will contain the attributes $A_1, A_{2^{n-1}+1}, A_{2^n+1}$. It will have two children: on the left, it will have the GHD corresponding to $Q_{R^{2^{n-1}}}$, and on the right it will have an identical GHD over the attributes $A_{2^{n-1}+1}, A_{2^{n-1}+2}, \dots, A_{2^n+1}$ instead of the attributes $A_1, A_2, \dots, A_{2^{n-1}+1}$. Note that each bag of our constructed GHD has $3$ attributes, but they may not appear in any relation together. Additionally, note that the depth of our GHD is simply $n$.
If we naively apply the AGM bound to derive the fractional hypertree width, we get a width of $E^3$. However, if, for each attribute $A_i$, we (virtually) create a relation $S(A_i)$ of size $V$, our fractional hypertree width becomes $V^3$. Alternatively, we can also use DBP-width to derive the $V^3$ bound without introducing these relations.
Applying the results of GYM [@GYM] gives us that we can answer $Q_{2^n}$ in $O(n)$ MapReduce rounds with $O(V^3)$ communication cost. Given that we need to answer $O(\log V)$ of these queries and that $n \le O(\log V)$ for each of these queries, we have a $O(\log^2 V)$ round MapReduce algorithm with ${\widetilde{O}}(V^3)$ total communication cost for all pairs shortest paths, which is within poly-log factors of standard algorithms for this problem.
In addition, if we allow a $O(k^* \log k^*)$ round MapReduce algorithm, we can reduce the total communication cost to ${\widetilde{O}}(EV)$ by using a chain GHD. In particular, for a query $Q_k$, the GHD will be a chain of $k$ bags such that the $i^{th}$ bag in our chain consists of $A_i$, $A_{i+1}$ and $A_{k+1}$. This construction ensures that two of the three attributes in each bag appear in a relation together, reducing the width to $EV$.
We note that we derived this MapReduce bound with our generic algorithms, without any specialization for this particular problem. We can also derive a serial algorithm for the problem with the same bound, but it requires a small optimization. By construction, our (original, non-chain) GHD has the property that every subtree whose root is at a particular level is completely identical. This means that AggroGHDJoin does not need to visit each bag; it simply needs to visit one bag per level, and then assign the result to the other bags on the level. With this optimization, our algorithm computes all pairs shortest paths in ${\widetilde{O}}(V^3)$, again within poly-log factors of specialized graph algorithms.
[^1]: Two technical notes: (1) methods like submodular width [@Marx:2010:THP:1806689.1806790] or Joglekar and Ré [@2015arXiv150801239J] require that we first partition the instances and then run the above algorithm; (2) FAQ [@FAQ] is not output sensitive (it does not use GHDs), and so it handles output attributes less efficiently than the above algorithm, as seen in Example \[ex:faqoutput\].
[^2]: Informally, a map is [*node monotone*]{} if adding more [*nodes*]{} to a graph does not reduce the measure, but additional edges may reduce the measure, see Definition \[def:node-monotone\].
[^3]: In contrast, FAQ’s decomposition strategy may miss the optimal GHD. Appendix Example \[example:faq-2x\] shows a case in which using the FAQ decomposition gives a width $2n$ while AJAR obtains width $n$ for $n\geq 1$. We also exhibit a family of queries and instances on which FAQ runs in time $\Omega(N^{3n/2})$ while AJAR runs in time $O(N^{n})$ for $n\geq 1$, see Appendix Example \[example:faq-decomposition\].
[^4]: We have been told that LogicBlox has implemented a similar algorithm recently, but their approach is not public. We shared our implementation with them several months ago.
[^5]: Traditionally GHDs are defined as a triple $({\mathcal{T}}, \chi, \lambda)$ where the function $\lambda: {\mathcal{V}}_{\mathcal{T}}\to 2^{{\mathcal{E}}_{\mathcal{H}}}$ assigns relations to each bag. Here we omit this function and implicitly assign *every* relation to each bag (so $\lambda(t) = {\mathcal{E}}_{\mathcal{H}}$ for all $t \in {\mathcal{V}}_{\mathcal{T}}$). Though this makes a difference for certain notions of width, it leaves the fractional hypertree width unchanged, as adding more relations to the linear program will never make the objective value worse.
[^6]: Note that, by this definition, the operators in aggregation ordering can be the product aggregation $\otimes$. However, product aggregations require different definitions, see Section \[sec:univ-aggregation\].
[^7]: Recall for any $V \subseteq {\mathcal{V}}$, ${\mathcal{E}}_V$ is defined to be $\{E \in {\mathcal{E}}| E \cap V \neq \emptyset\}$.
[^8]: Recall ${\mathcal{E}}_C = \{E \in {\mathcal{E}}| E \cap C \neq \emptyset\}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Recent experiments have shown that many species of microorganisms leave a solid surface at a fixed angle determined by steric interactions and near-field hydrodynamics. This angle is completely independent of the incoming angle. For several collisions in a closed body this determines a unique type of billiard system, an aspecular billiard in which the outgoing angle is fixed for all collisions. We analyze such a system using numerical simulation of this billiard for varying tables and outgoing angles, and also utilize the theory of one-dimensional maps and wavefront dynamics. When applicable we cite results from and compare our system to similar billiard systems in the literature. We focus on examples from three broad classes: the ellipse, the Bunimovich billiards, and the Sinai billiards. The effect of a noisy outgoing angle is also discussed.'
author:
- 'Madison S. Krieger'
bibliography:
- 'newrefs.bib'
title: Microorganism billiards in closed plane curves
---
[^1]
\[sec:microint\]Introduction
============================
Chaotic billiards [@chernmakar; @wojt1986; @donnay1991; @tabachbook] continue to be an extremely active area of research and its objects of study are of interest in varied disciplines of mathematics. Billiard systems are also useful tools and models in physics [@chatterjee96; @bhpz2004; @ulambilliard]. Current extensions to well-studied billiard problems include modifications of the table geometry, the shape of the inter-collision trajectories, and the rule for generating a new trajectory upon contact with the table boundary. In this direction, recent attention has been paid to aspecular reflection laws, especially in dissipative billiard systems commonly referred to as pinball billiards [@ams2009; @mps2009; @ams2012; @mddgp2012] and slap maps [@mddg2015; @mddgp2014; @mddg2014], as well as to aspecular reflection laws arising from other physical effects [@adh2008].
Recent experiments have revealed that microorganisms also play mathematical billiards [@kdpg2012]. This was a surprising result in the field of microorganism locomotion and fluid dynamics; the role of boundaries has been known for some time to have a strong effect on microorganism trajectories, even at large distances — results on mammalian spermatozoa [@rothschild1963; @fm1995; @smithblake2009; @sgbk2009] as well as on bacteria and self-propelled particles [@lpcp2010; @mbss2014; @lauga06; @ElgetiGompper2009; @GiaccheIshikawaYamaguchi2010; @hdf1992; @btbl2008; @ShumGaffneySmith2010] gave rise to a standard theory on the reorientation of microorganism trajectories in the presence of solid boundaries. The theory [@LaugaPowers2009; @sl2012] can be summarized succinctly: far-field hydrodynamics cause microorganisms to reorient themselves either parallel or perpendicular/antiperpendicular to a solid boundary, depending on the sign of their hydrodynamic dipole. This is the leading-order term of the Green’s function for the appropriate hydrodynamic equation describing the global flow caused by the shape-deforming microorganism. The two possibilities are commonly referred to in the literature as being either of “pusher” type, describing microorganisms which experience a drag force on the head (or foremost body part in the direction of motion) and propel from the rear, or of “puller” type, which propel from the head and experience the largest drag on the rear. The theory predicts that pushers parallelize with solid boundaries; some common microorganisms of this class include spermatozoa and *E. coli*, while pullers orient along the normal line to the solid boundary; some model organisms from this class include many algal cells such as *Chlamydomonas reinhardtii*, which locomotes by performing a breaststroke-type motion with two fore flagellar appendages.
While this theory is accurate for many microorganisms (especially prokaryotes), it seems that scattering from solid boundaries is completely different for the microorganisms studied in the experiment (mostly eukaryotes, whose flagellar structure is much more complex). For these organisms, the scattering events are defined by near-field hydrodynamics and steric interactions; the swimmer collides with the solid boundary and is reoriented a certain amount depending on its morphology, which includes both the body geometry (species) and the flagellar/ciliar expression (which depends on temperature, possible mutations, and so forth). During this process, memory of the incoming angle is lost; after the collision, the outgoing angles are governed by a distribution with well-defined peaks. This is the motivation for the current work, which considers a model of such collision events as an aspecular reflection law for two-dimensional closed billiard tables with a single moving particle (the microorganism). This extends previous work [@wlst2016] on a similar billiard system in some two-dimensional polygonal tables. The main results of that work which differ from classical polygonal billiards were of a hyperbolic nature; for different tables and outgoing angles, regions of large negative Lyapunov exponent with periodic attractors were seen, and some very small regions of positive Lyapunov exponent correlated with ergodic components (unlike non-hyperbolic ergodic orbits in polygonal tables with specular reflection). Because the aspecular law prohibits us from using several quantitative tools, especially those involving invariant measures, we focus primarily on numerical simulations. When configuration spaces are plotted, our code for generating these graphs is based on previous work [@lanselcode]. Our work may also be applicable to the design of robots which re-orient themselves based on tactile rather than visual data. These two cases of aspecular billiards have developed in parallel; for instance, the motion of aspecularly-reflecting robots [@el2013] is analogous in certain respects to microorganism billiards, while rectification of ensembles of such locomoters by gears have been considered in both in the microorganism case [@adr2009; @saga2010; @ladriscmaf2010] and in the robotic case [@lz2013].
The organization of the paper is as follows: in Sec. \[sec:theory\], we define the microorganism billiard and assess the quantities of interest. The aspecular nature of the reflection law prohibits the use of several analytical tools known in the literature, so these quantities are primarily determined via numerical simulation, though we do provide qualitative results from the theory when possible. We then provide three examples from billiards classes of interest: the ellipse (Sec. \[sec:ellipse\]), the Bunimovich stadium (Sec. \[sec:stadium\]), and the Sinai billiard in a square unit cell (Sec \[sec:sinai\]). We summarize the results in Sec. \[sec:summary\].
\[sec:theory\]Microorganism billiards
=====================================
![(Color online) Geometry of microorganism billiards. The microorganism collides with the curved boundary of the table at an angle $\theta_i$ and departs at a fixed angle $\theta_o$. All angles are measured with respect to the normal at the point of collision. []{data-label="setupfig"}](setupfig2){width="30.00000%"}
We consider two-dimensional billiard tables populated by a single particle which moves with constant (unit) speed and whose free-flight path is always a straight line. The only departure from a classical billiard is our reflection law, which is aspecular: rather than the specular reflection law $\theta_o=\theta_i$, where alphabetical subscripts $\{o,i\}$ refer to “out” and “in”, respectively, our reflection law is $\theta_o=C$, where $C$ is a constant or a random variable with a prescribed distribution. Because the outgoing angle is fixed, the microorganism billiard is effectively a one-dimensional map [@demelostrein]. However, unlike the case of bouncing robots [@el2013], we choose our reflection law to preserve orientation of the trajectory with respect to the tangent line immediately before and after a collision; that is, the microorganism always crosses the normal[^2]. The outgoing angle is therefore one of two possibilities, because crossing the normal can change the orientation of the following links from clockwise to counter-clockwise depending on the incoming angle and the table curvature. Equivalently, the map is only completely determined at a boundary point $q$ if the pre-image of that point is also given, as these two pieces of data determine the subsequent orientation. The one-dimensional map is therefore only a map of the true collision space when this orientation is preserved for all collisions. We point out the ramifications of these orientation reversals and when the dynamics is completely or only partially encapsulated in the appropriate one-dimensional map $F(q)$, where $q$ is the arclength parameter for the table boundary. The derivative of this map is given [@ams2009; @mps2009] by $$F'(q_0)=-\frac{t K_0 + \cos \theta_{o,0}}{\cos \theta_{i,1}}, \label{onedimmap}$$ where the numeric subscripts $\{0,1\}$ indicate that the map $F:q_0 \rightarrow q_1$ takes a position on the table boundary $q_0$ with boundary curvature $K_0$ and (fixed) outgoing angle $\theta_{o,0}$ with respect to the normal to the table boundary at $q_0$ and maps it to a new boundary position $q_1$. When a straight line is drawn in the table between $q_0$ and $q_1$, representing the actual billiard trajectory, it has length $t$ and intersects the table boundary at an angle $\theta_{i,1}$ with respect to the normal to the boundary at $q_1$. To avoid a cumbersome amount of indices, we write $\eta \equiv \cos \theta_{i,1}$, and henceforth the numeric subscripts of $0$ on all other quantities will not be written.
The unusual reflection law has strong implications for the toolbox which is available to analyze microorganism billiards. Perhaps the most important is that these systems do not preserve area in phase space, and indeed have no natural invariant measures. As a consequence, we do not speak at all to possible ergodic and statistical properties of microorganism billiards, since it is not clear how to derive results in these directions in the absence of such a measure. In addition, the involutive or time-reversal property of specular billiards is not preserved in microorganism billiards, and thus we must be extremely cautious in invoking pre-images or continued fractions reaching backwards in time from the current point in a microorganism’s orbit. This also restricts the extent to which we can truly utilize one-dimensional dynamics; we discuss this on a case-by-case basis in the text.
Unlike previous work on microorganism billiards [@wlst2016], we do not consider the role of “skids”, or events in which the billiard particle collides with the wall at an arclength position $q$ and departs at an arclength position $q \pm s$, where $s$ is the length of the “skid”; one of the conclusions of that work was that skids had negligible impact on the dynamics. In our system, it would be possible to “skid” from regions of low curvature to regions of high curvature in some tables; however, recent experiments [@szs2015] suggest that surface roughness reduces the length of the skid dramatically, so we consider our billiard tables to have roughness on a small scale. This assumption may affect the outgoing angle of a true microorganism; we assume that re-orientations due to the roughness of the table are incorporated as noise into the distribution of the outgoing angle. The complete problem is schematized in Fig. \[setupfig\].
We will focus our efforts on quantities that we believe will be of interest to both the dynamics and the physics communities. We compare the orbit structures in various tables to the classical billiard problem to understand the potential behavior of a microorganism as well as to extract visual information about the topological dynamics, such as rotation numbers of periodic orbits. We use this information to construct a phase diagram of the different behaviors as a function of table geometry and outgoing angle, and also use this information to design useful microfluidic sorting and trapping mechanisms that might find application in experiment. Motivated by results [@wlst2016] suggesting that the hyperbolic dynamics of a microorganism billiard differs dramatically from a classical billiard, we place special emphasis on Lyapunov exponents. These will be calculated using standard methods [@oseledets] from specular billiards simulation and theory rather than the one-dimensional map, due to the possibility of orientation-reversals occuring multiple times over many collisions.
\[sec:ellipse\]Ellipse
======================
\[subsec:ellone\]One-dimensional dynamics
-----------------------------------------
![(Color online) Diagram of the regions of negative and positive Schwarzian derivative for the one-dimensional map. The solid line indicates the curve given by Eq. (\[critellipse\]), above which the numerator of Eq. (\[onedimmap\]) cannot change signs. The red dots indicate the lowest value of $E$ for a given $\theta_o$ where a globally piecewise-negative Schwarzian derivative is seen in the numerical simulations. []{data-label="schwarz"}](SchwarzLine2){width="50.00000%"}
![(Color online) Plots of the billiard map $F(q)$ (blue), its deriviative $F'(q)$ (orange), and the quantity $\frac{1}{\sqrt{F'(q)}}$ (green), which is convex when the Schwarzian derivative is negative. The plot is shown on the range $q \in [0, \pi]$ and is extended to the full table by symmetry. $Sing(F)$ is denoted with red circles. The plots are shown in an ellipse with $E=0.86$ and $\theta_o=0.327$ (top), $\theta_o=1.09$ (bottom). These values have been chosen such that the top panel displays a pair $(E,\theta_o)$ lying above the curve Eq. (\[critellipse\]), while the lower panel displays a pair below this transition point. This illustrates the global transition from a one-dimensional billiard map with piecewise negative Schwarzian derivative to one with piecewise positive Schwarzian derivative. In crossing the curve shown in Fig. \[schwarz\], all critical points vanish, changing the behavior of the Schwarzian derivative. []{data-label="ellipsetriptych"}](1ddip){width="50.00000%"}
We begin with an analysis of the one-dimensional map, Eq. (\[onedimmap\]), specialized to the case of the ellipse. The map is only piecewise continuous; we denote the singular set, comprised of critical points and points of discontinuity, by $Sing(F)$. An example of such a set is shown in Fig. \[ellipsetriptych\]. We are interested to know the regime in which the Schwarzian derivative of this map is negative. This simplifies the dynamics dramatically, for negativity of the Schwarzian is preserved under forward iterations of the map and much is known in the literature [@demelostrein; @bpp2016] about hyperbolicity and the structure of attractors for such maps. The reader can easily check that the Schwarzian derivative $S(F(q))=\frac{F'''(q)}{F'(q)}-\frac{3}{2}\left(\frac{F''(q)}{F'(q)}\right)^2$ is negative if the quantity $\frac{1}{\sqrt{|F'(q)|}}$ is convex; thus, this behavior can be deduced by analysis of Eq. (\[onedimmap\]). After lengthy algebra it is possible to demonstrate that the only way for $F'(q)$ to preserve one sign for all values of $q$ is if $\mathrm{min}(t K) > \cos \theta_o$. In terms of the eccentricity [^3] $E = \sqrt{1-b^2/a^2}$, where $a$ is the length of the major axis and $b$ is the length of the minor axis, the critical ellipse separating the two sign behaviors of $F'(q)$ has eccentricity $$E_c = \sqrt{1-\frac{\cos \theta_o}{2}}. \label{critellipse}$$ By sampling a few values of $E$ and $\theta_o$ above this line, it is easy to see that the quantity $1/\sqrt{|F'(q)|}$ is always convex away from the singular points, where the Schwarzian derivative diverges; thus all microorganism billiards above this curve have negative Schwarzian derivative $\forall q \notin Sing(F)$.
To confirm this, we integrated Eq. (\[onedimmap\]) numerically to obtain $F(q)$. Note that due to the symmetry of the ellipse, we can describe orientation-preserving links using iterations of $F(q)$, and can include orientation reversals by using reflections across the axes. A generic orbit is therefore described by a word on two generators of the type $FFFRFRFFR...$, where $F$ is the billiard map and $R$ is a suitable reflection. We measured the Schwarzian derivative of $F$ and took note of the first value of $E$ for a fixed $\theta_o$ when the Schwarzian derivative became globally negative. The values collapse well on to the curve predicted by Eq. (\[critellipse\]), see Fig. \[schwarz\]. Below this curve, the Schwarzian derivative is positive $\forall q \notin Sing(F)$ — surprisingly, the Schwarzian has only one sign $\forall q \notin Sing(F)$ for every ellipse and outgoing angle away from the critical line. Note that the pre-image of any point under $F$ is unique in the ellipse, so that the inverse is well-defined; this means that the curve Eq. (\[critellipse\]) also defines the sign of the Schwarzian derivative for the inverse map, which is the opposite as the sign for the forward map. However, for the actual microorganism billiard, information about the inverse map is only useful if there are no changes in orientation. To provide an illustration of how Eq. (\[critellipse\]) organizes the (piecewise) global sign of the Schwarzian derivative, Fig. \[ellipsetriptych\] shows the maps $F(q)$, $F'(q)$, and $\frac{1}{\sqrt{|F'(q)|}}$ for one ellipse and three different values of $\theta_o$ which lie on different sides of the critical curve. The disappearance of several points from $Sing(F)$ is responsible for the global change in sign.
![(Color online) The phase diagram for microorganism billiards in the ellipse. The solid black line indicates the division between periodic orbits and rotator-like orbits whose limit set is dense. This line is identical to the curve Eq. (\[critellipse\]) only for $\pi/4 < \theta_o < \pi/2$; the continuation of that curve is shown by a dashed red line. Below the solid black line, the dynamics are orientation-preserving; the dynamics also preserve orientation in a small region above this curve for $\theta_0 > \pi/4$, and this region is demarcated by another dashed red line. The period of periodic orbits increases with $E$ and $\theta_o$ until the line $\theta_o=\pi/4$ is approached, and the pattern begins anew for $\theta_o > \pi/4$. When $\pi/\theta_o \in \mathbb{Z}$, there can be exchange-of-interval type orbits which are topologically transitive on a finite set of intervals. The insets give examples in the configuration space for the two phases. []{data-label="ellphase"}](ellphasediag){width="50.00000%"}
Much about the attractor structure for maps with piecewise-negative Schwarzian derivatives has been described in a recent work [@bpp2016]. In this regime we expect strong attractors in microorganism elliptical billiards without orientation reversals, with all values of $q$ belonging to at least one basin of attraction, and attractors primarily of periodic or change-of-interval type. We confirm this expectation and discuss attractors and hyperbolic dynamics in the next section.
\[subsec:ellhyp\] Billiard simulation and hyperbolic dynamics
-------------------------------------------------------------
We choose to determine the values of $(E,\theta_o)$ where the one-dimensional dynamics requires reflection maps $R$ to complete words representing orbits via simulation. We simulate the full microorganism billiards follwing a standard method: we embed the billiard table in the Euclidean plane and calculate the itinerary $\{ q_i \}$ of a trajectory by finding the intersection point of the straight line departing from initial position $q_0$ with angle $\theta_o$ via Newton-Raphson type methods and continuing in this manner to find $q_i$ $\forall i<N$, where $N$ is the maximum number of collisions considered for the orbit with the given initial position. To get an overall picture of the orbit structure, we simulated the billiard on a $30 \times 30$ grid evenly distributed on table eccentricity $0<E<1$ and outgoing angle $0.1<\theta_o<\pi/2-0.01$, with $N=1000$ and 5-10 different initial positions for each billiard table. The orbit behavior is organized in a phase diagram shown in Fig. \[ellphase\]. Roughly half of the area of the parameter space describes dynamics which preserve orientation, so that the billiard is governed exactly by Sec. \[subsec:ellone\]. However, the region of negative Schwarzian derivative only preserves orientation for $\theta_o > \pi/4$ and $E \lesssim 0.96$.
For piecewise negative Schwarzian maps, the maximum number of attractors [@bpp2016] can be as large as $2^{2 \ \mathrm{card}(Sing(F))}$, which for a generic elliptical microorganism billiard can be very large. Interestingly, our simulations never revealed more than one attractor, with a basin of attraction including the entire boundary. The set of periodic attractors has full measure in the parameter space $(E,\theta_o)$ — however, there are change-of-interval type orbits when $\pi/\theta_o \in \mathbb{Z}$. This is true as well in the orientation-reversing region with negative Schwarzian, and also in a small region of positive Schwarzian derivative for the one-dimensional map with $\theta_o < \pi/4$. Excluding this region, the rest of the orientation-preserving maps with positive Schwarzian derivative have identical structure (apart from $\pi/\theta_o \in \mathbb{Z}$), which we call “rotator-like” orbits due to their orbits staying exclusively outside the foci, in analogy with the classical elliptical billiard. These billiards have a caustic which is itself somewhat elliptical in appearance, but rotated with respect to the table boundary. Note that since these tables preserve orientation, their dynamics are equivalently defined by the one-dimensional map and therefore the inverse, which has everywhere negative Schwarzian, also satisfies the conditions given [@bpp2016] to have a finite set of attractors with global basin between them of periodic or change-of-interval type. Yet here we again find only single attractor in the reverse dynamics; all the rotator-like orbits with $\pi/\theta_o \notin \mathbb{Z}$ have periodic orbits as limit sets of the inverse map.
![(Color online) A surface plot of the Lyapunov exponent for microorganism billiards in the ellipse averaged over three initial conditions. Deep in the rotator-like phase, the Lyapunov exponents are all zero. Nearing the curve Eq. (\[critellipse\]), the exponent becomes negative; despite the map having positive Schwarzian, the orbits in this regime are attracted to periodic orbits as shown in Fig. \[ellphase\]. After crossing the curve, the Schwarzian derivative is negative on every component, leading to negative Lyapunov exponent — the strongest attractors are 4- and 6-periodic orbits which do not preserve orientation. []{data-label="lyell"}](LYEL){width="50.00000%"}
We also calculate the Lyapunov exponent, using a method identical to that used in previous work on microorganism billiards [@wlst2016]. Due to the presence of global attractors for high $E$, we expect the Lyapunov exponent to be negative. This is indeed true, with the results shown in Fig. \[lyell\]. The curve of the most significance for the Lyapunov exponent seems to be the solid black line in Fig. \[ellphase\], which includes some regions of both positive and negative Schwarzian derivative, but only displays periodic orbits. These are the strongest attractors, with 4- and 6-periodic orbits being exceptionally strong. Beneath this curve, the Lyapunov exponents vanish, except for a small region of rotator-like orbits which seem to have slightly non-zero Lyapunov exponents; this could be due to numerical error, or could signal the gradual formation of periodic orbits.
Lastly, we considered the role of a noisy outgoing angle. Noise-mediated transitions between basins of attraction [@ms1995] seem not to be possible because each table has only one global attractor. To check this fact, we simulated several more initial conditions for each of our parameter points with weak noise, to see if orbits transitioned between multiple attractors. One attractor was seen, with the orbit slightly “smeared” by weak noise. If multiple attractors coexisted we did not detect them; since the attractors seem to have global basins, the role of noise in these tables is trivial.
\[sec:stadium\]Bunimovich systems
=================================
\[subsec:stadtheory\]Wavefront dynamics
---------------------------------------
In this section we consider a class of chaotic billiards where the mechanism for hyperbolicity and positive Lyapunov exponents is that of *defocusing*. This mechanism was first discovered by Bunimovich [@bunim1974a], and so systems employing the mechanism are often referred to as Bunimovich systems. The most generic requirement for the mechanism to exist in a billiard table is the presence of *absolutely focusing arcs* [@bunim1992]. These systems therefore include an immensely large number of billiard tables; since our results are primarily of a computational nature, we focus on the stadium [@bunim1979] when speaking of hyperbolic dynamics, and then focus on an example of the applicability of billiards theory to practical problems of microbiology by presenting a mushroom-shaped billiard as a trap for microorganisms.
The stadium is constructed using two half-circles of radius $r$ connected by straight lines of length $L$. The family of stadia is therefore defined by a nondimensional parameter $L/r$ [@Note2]. In the limit $L/r \rightarrow 0$, we recover microorganism billiards in a circle. For nonzero $L/r$, it is expected that a billiard trajectory will change its orientation with respect to the table during its orbit; as a consequence, we neglect the one-dimensional map given in Eq. . As described in Sec. \[sec:ellipse\], it would be possible to describe the dynamics using words on the one-dimensional map $F$ and a map $R$ which takes advantage of the four-fold symmetry to effect reversals of orientation in the actual billiard, but we prefer to use the language of wavefronts to provide supporting intuition for our simulations. The wavefront in question is the plane curve which is locally normal to each trajectory in a fan of initially-nearby trajectories, so changes in curvature of a flat wavefront indicate that initially parallel trajectories are either converging or diverging after a collision, see Fig. \[scatter\].
![(Color online) Illustration of ray dynamics in the Bunimovich stadium (left) and Sinai unit cell billiard (right, see Sec. \[sec:sinai\]), for $\theta_o=\pi/3$ (top) and $\theta_o = \pi/20$ (bottom). Rays which would contact the table along the normal have been slightly perturbed to avoid this ill-defined event. The wavefronts in question are locally normal to the ray packets.[]{data-label="scatter"}](scatterers){width="50.00000%"}
We follow standard conventions [@chernmakar], according to which the wavefront curvature $\mathcal{B}$ is defined to be zero for a flat wave, positive for a dispersing/expanding wave, and negative for a focusing/contracting wave; at a focusing point of the wave, $\mathcal{B}=\infty$. We denote the curvature of a front immediately before a collision as $\mathcal{B}^-$ and immediately after as $\mathcal{B}^+$. The two are related by an aspecular variant of the mirror equation [@mps2009]: $$\mathcal{B}^+=\frac{K}{\cos \theta_o}. \label{mirroreqn}$$ Note that the specular mirror equation reads $\mathcal{B}^+ = \mathcal{B}^- + \frac{2 K}{\cos \theta_o}$; in particular, our outgoing curvature does not depend at all on the incoming curvature. During the subsequent free-flight, the curvature evolves according to $$\mathcal{B}_t = \frac{1}{t+1/\mathcal{B}^+}. \label{freeflight}$$ These two expressions can be combined in a continued fraction formula to give the curvature of the front evolving at any time during the billiard dynamics. These dynamics will be very different from the specular limit due to the absence of $\mathcal{B}^-$ in Eq. (\[mirroreqn\]); in particular, a front “forgets” its past during any collision and adopts a new curvature which is sensitive to $\theta_o$. The aforementioned defocusing mechanism is based on an analysis of Eq. (\[freeflight\]) which reveals that a focusing wave front actually expands between two collisions taking place at a distance $\tau$ apart so long as $$\tau \mathcal{B}^+_0<-2, \label{defocus}$$ where $\mathcal{B}^+_0$ is the curvature emanating from the first of the two collisions.
The general theory of how hyperbolicity and chaos arise in Bunimovich billiards reveals which type of collision events lead to positive Lyapunov exponents, so we will analyze some basic situations using Eq. (\[mirroreqn\]) and Eq. (\[freeflight\]) to determine when Eq. (\[defocus\]) is satisfied. We first divide the types of collisions into three possible classes: **i)**, the microorganism leaves a point on a circular arc and contacts another point on the same circular arc, **ii)** the microorganism leaves a point on one flat component and contacts a point on another flat component, **iii)** the microorganism leaves a point on one circular arc and contacts a point on one flat component, or **iv)** the microorganism leaves one circular arc and contacts another. In the specular theory, (i),(ii) and (iii) are often called “nonessential collisions”, as one can show that the curvature of wavefronts remains unchanged by these collisions [@chernmakar]. In microorganism billiards, however, (ii) and (iii) immediately “reset” the curvature of the wavefront to zero, violating (\[defocus\]). If the radius of a circular cap is $r$, then an event of type (i) or (iv) is only defocusing if $$-\frac{\tau}{r \cos \theta_o} <-2,$$ so generically speaking type (iv) events are more likely to lead to defocusing. This leads to an interesting competition between the microorganism outgoing angle and the geometry of the table. For the stadium, having line sections much longer than the radius of the caps will increase the likelihood of defocusing type (iv) effects, but these are also increasingly often “reset” by a type (ii) or type (iii) event. We therefore expect shorter line segments to lead to more chaotic dynamics. However, decreasing the length too much will lead to focusing events as $\tau$ decreases; this is especially deleterious if $\theta_o > \pi/4$. In addition, motivated both by our results in Sec. \[sec:ellipse\] on microorganism billiards in the ellipse and the presence of marginally unstable periodic orbits (or MUPO’s) in the specular Bunimovich tables, we expect that the hyperbolic dynamics will be complicated by the presence of attractive periodic orbits — therefore the Lyapunov exponent is expected to be highly sensitive to table geometry, outgoing angle, and initial conditions. We do not attempt to demonstrate uniform hyperbolicity for these tables or for the tables in Sec. \[sec:sinai\], as it is obvious in the case of the stadium that uniform hyperbolicity does not exist for the vast majority of geometries; we discuss uniform hyperbolicity for Sinai tables very briefly, as it is obvious when it will be preserved or destroyed by the aspecular reflection law.
![(Color online) A sampling of configuration spaces for microorganism billiards in stadia.[]{data-label="stadcaps"}](stadcaps){width="50.00000%"}
\[subsec:stadsim\]Billiard simulation and hyperbolic dynamics
-------------------------------------------------------------
![(Color online) Contour plot of the Lyapunov exponent in the stadium averaged over three initial conditions chosen from the uniform distribution for various geometries and outgoing angles. They are generically much higher (approximately 5-10 times higher) than in the specular case [@dp1995] due to the removal of nonessential collisons. However, much unlike the specular case, pockets with negative Lyapunov exponent can be found, especially for $\theta_o > \pi/4$ — these correspond to weakly attracting periodic orbits (marginally unstable periodic orbits, or MUPO’s) comprised only of nonessential collisions. The introduction of noise destabilizes these orbits, resulting in a Lyapunov exponent landscape which is higher but also much more homogeneous.[]{data-label="lystad"}](Lystad){width="50.00000%"}
We studied the orbit types and hyperbolic dynamics of microorganism billiards in the Bunimovich stadium by simulating [^4] the dynamics in a $30 \times 30$ grid distributed evenly on the table geometry $L/r \in [0,2]$ and the outgoing angle $\theta_o \in [0.1,\pi/2-0.1]$ for several initial conditions. Again, the bounds on $\theta_o$ have been selected to avoid ill-defined events such as orbits tracing the boundary of the table ($\theta_o=\pi/2$) or the slap map ($\theta_o=0$), which for the stadium simply converges to the two-periodic attractors representing only-flat and only-cap collisions. Unlike the relatively simple orbit structure of microorganism ellipse billiards (Sec. \[sec:ellipse\]) or the generically-chaotic orbit structure of Sinai billiards (Sec. \[sec:sinai\]), the stadium showed an incredibly rich array of orbits. We estimate that there are between 15 and 25 distinct orbit types, including periodic orbits with a range of periods, some resembling “cuts” of the specular stadium orbit — by this we mean any coherent segment of links which could occur in a specular stadium trajectory that is compatible with our reflection law, repeated periodically. These latter structures are easily recognized by plotting the phase space trajectory for several initial conditions and comparing with the phase space of the specular case. We show some representative examples of configuration spaces in Fig. \[stadcaps\]. Other orbits were less easily classified, representing an interaction between the weakly-attracting periodic orbits and the defocusing trajectories. This competition is reflected in Fig. \[lystad\], where we plot the Lyapunov exponent averaged over three initial conditions chosen uniformly from Lebesgue measure on the table domain and initial orientation. For tables nearer the boundary of the figure (representing more extreme geometry/outgoing angle combinations), the Lyapunov exponent is lower than the specular one or even negative; for more generic situations, the Lyapunov exponent can be as much as an order of magnitude higher than the specular case due to the increase in defocusing links (equivalently, the reduction of nonessential collisions). We refer to the attractive periodic orbits as “weakly attracting” because for even very low thresholds of noise, such as the uniform distribution $\theta_o/|\theta_o| \in [0.9 , 1.1]$, the Lyapunov exponent landscape homogenizes at a high positive value and no periodic structures can be found in the orbits. This is a strong contrast to the results of Sec. \[sec:ellipse\], where noise simply tended to “smear” the globally attracting periodic orbits.
\[subsec:stadtrapfil\] Design for a microorganism trap
------------------------------------------------------
![(Color online) Microorganism billiards in the double-mushroom; the top lobe has a radius which is $10\%$ larger than that of the bottom lobe. The width of the stem is $20\%$ the radius of the top lobe and the length of the stem is $50\%$ the radius of the top lobe. Here, $\theta_o = \pi/6 + \theta_n$, where $\theta_n$ is uniformly distributed on $(-\pi/12,\pi/12)$. The microorganism begins in the bottom lobe and migrates to the top lobe after a few hundred collisions, where it remains thereafter — this calculation is continued for $N=3000$ collisions. The initial condition is shown as a red line. If the noise is Gaussian distributed rather than uniformly distributed, the “trap” is not permanent; an event from the tail of the distribution will eventually propel the microorganism back into the bottom lobe.[]{data-label="shroomtrap"}](mushroomtrap){width="30.00000%"}
Because flat walls and absolutely focusing arcs can be combined in an infinite number of ways, the Bunimovich class of billiards gives the table designer a large amount of control over the properties of the phase space. This is an extremely useful tool in the context of designing a trap for one or more microorganisms, which has been the subject of much recent work [@bjmvdvsetal2; @smbl2015; @jkp2016]. A prime example of a useful billiard table for constructing traps is a mushroom-shaped table [@bunim2001], which has a phase space which is divided into an ordered/integrable component comprised of periodic orbits trapped inside the spherical cap, and a chaotic component with orbits that include collisions with the flat walls of the stem and the spherical cap [^5]. The boundary between the two regions is usually of a fractal nature, and also ‘sticky’ [@amk2005; @dettmann2016], meaning that orbits originating in the chaotic region can spend large amounts of time at the border with the integrable region due to the existence of marginally unstable periodic orbits, or MUPO’s. Motivated by this phase space structure, as well as previous results on the role of noise in trapping events [@ae2010], we designed a passive, tunable trap for microorganisms based on the mushroom shape. Ours is an asymmetric double-mushroom, pictured in Fig. \[shroomtrap\]. The mechanism of the trapping depends on both the asymmetry of the table and noise in the outgoing angle of the microorganism. In the limit of zero noise, the behavior of the microorganism is entirely determined by the radii of the two lobes, the outgoing angle and the initial condition: the latter two set a fixed angular momentum $\mathcal{L}$ (defined as the distance of closest approach to the focus of the arc). If $\mathcal{L}$ is larger than the radius of the lobe in which the microorganism originates, the orbit will be periodic and contained in that lobe; if $\mathcal{L}$ is smaller, the microorganism will eventually migrate to the other lobe, where the process is repeated. This description is only approximate, ignoring several fine details [@dettmann2016], but serves to illustrate the idea that lobe asymmetry can filter microorganisms from an escapable lobe to an inescapable lobe.
The inclusion of noise and its distribution are important to the design of this trap — not only do they simplify the picture by removing MUPO’s, they allow for rather precise tunability in first-return properties when desirable, or suggest lobe radii for different purposes. In our example, we consider the noise to be uniformly distibuted; this leads to a rather simple design for a trap, for uniform noise in the outgoing angle leads to a bounded distribution of angular momenta; one need simply choose one lobe to have a radius outside of this bound to trap a microorganism for all time. This is the situation pictured in Fig. \[shroomtrap\]. If the noise is instead Gaussian distributed, events from the tail of the distribution will cause the microorganism to return after many collisions. We leave the precise details of first-return times and their dependency on $\theta_o$ and table geometry to future work. This division of phase space is a property of a robust class of billiard tables, and we believe that it can be very useful in the design of microfluidic devices for various purposes involving mixing, filtering, and trapping of microorganisms based only on table geometry and outgoing angle.
\[sec:sinai\]Sinai Systems
==========================
\[subsec:sinaitheory\]Unfoldings and wavefront dynamics
-------------------------------------------------------
Our last system of interest is the chaotic billiard system discovered by Sinai [@sinai1963; @sinai1970]. The model has deep connections to the fundamentals of statistical mechanics [@chernmakar]. The tables are in some sense “opposite” to the Bunimovich tables, in that they are comprised only of dispersing walls (those having positive curvature $K$). This leads to a wealth of tables inhabiting different types of spaces, of which we focus on only one example. We briefly discuss the tools of wavefronts and unfoldings before presenting the results of our simulations, comparing the Lyapunov exponents obtained to other works in the literature [@dp1995; @dahl1997].
![(Color online) Two possible unfoldings for the microorganism Sinai billiard. The unfolding A naively applies the aspecular reflection law on all boundaries, and may represent a microorganism responding to external stimuli or ‘tumbling’ type events. The unfolding B has no such events; to preserve the direction of free-flight in the lattice, specular reflections are applied on the walls of the unit cell. Numbered labels of cells indicate the order of re-folding to obtain the trajectory in the unit cell.[]{data-label="unfoldings"}](unfoldings){width="50.00000%"}
{width="\textwidth"}
![(Color online) Lyapunov exponents for the B-type unfolding for various geometries and outgoing angles, averaged over three initial conditions. Equipping the walls of the primitive cell with a specular reflection law leads to higher Lyapunov exponents (more scattering events) than the A-type unfolding. []{data-label="lysinb"}](LySinaiB){width="50.00000%"}
Our analysis begins with Eq. (\[mirroreqn\]). Combining this with Eq. (\[freeflight\]) reveals that wavefront curvature decreases hyperbolically in free-flight, asymptoting to zero, and jumps to a fixed positive number at collisions with dispersing walls or to zero at collisions with a flat wall, tracing an erratic saw-toothed curve in time. As an example, we consider single dispersing disk of radius $R$ centered in a square cell of length $w$ [@Note2]. By creating copies of this unit cell and reflecting collisions from the flat walls across the unit cell, we arrive at an *unfolding* of the unit cell in which the billiard particle now navigates a square lattice of scatterers of radius $R$ and lattice spacing $w$. This geometry therefore captures the dynamics of two billiard tables at once. If we naively play microorganism billiards, however, the unfolded problem illustrates a rather strange feature, that the microorganism at some points in free-flight suddenly “decides” to reorient itself in response to the imaginary line representing the wall of the unit cell. Because of its deterministic character this is unlike a run-and-tumble event; we call this billiard the “drunken swimmer”, or Unfolding A (see Fig. \[unfoldings\]). As a representative of a relevant physical system, the unit cell dynamics of Unfolding A are likely more realistic than the lattice dynamics. We therefore study a second pair of unit cell-lattice systems, one which has aspecular reflections on scatters and specular reflections on flat walls, which we call Unfolding B. This latter system is more realistic for microorganisms navigating lattices of scatterers, which has been a subject of previous inquiry [@wrnr2008; @gkca2007]. We further note that this billiard system is of physical interest due to its previous use to calculate rheological quantities [@bs1996], which in our case are expected to be highly dependent on the outgoing angle — we leave such simulation work to future endeavors.
It is worth noting immediately that Unfolding B has wave dynamics very similar to the specular Sinai tables; it has a curvature which stays bounded away from zero, $\mathcal{B}_{min} = \frac{1}{\tau_{max}+1/\mathcal{R}}$, where $\mathcal{R} = \frac{K}{\cos \theta_o}$, and of course a maximum $\mathcal{B}_{\max}=\mathcal{R}$. Exponential growth of the length of the wavefront and uniform hyperbolicity follow generally, where the exponent in the growth of wavefront length is sensitive to $\theta_o$. Unfolding A is not uniformly hyperbolic due to the “resetting” of the curvature. Moreover, the existence of strongly attracting periodic orbits in the square microorganism billiard [@wlst2016] implies the existence of attractors which *avoid* the scatterer for certain values of $R/w$ and certain initial conditions; depending on the geometry and the strength of the attractor, this can lead to entirely polygonal dynamics, and in the lattice problem these correspond to “drunken” swimmers which are able to successfully navigate the lattice of scatterers without collisions. Such events are expected to dramatically decrease the Lyapunov exponent measured in these tables. Because our systems are already highly chaotic, we do not consider noise; it will have a strong effect on the periodic attractors in Unfolding A, and we expect the full implications of noise to be represented by results in other work [@shsl2009].
\[subsec:sinaisim\]Billiard simulation and hyperbolic dynamics
--------------------------------------------------------------
As in the previous sections, we simulated the microorganism billiard dynamics on a $30 \times 30$ grid, here distributed evenly on $R/w \in [0.025, 0.5]$ and $\theta_o \in [0.1, \pi/2-0.1]$, where $R$ is the radius of the scatterer and $w$ the width of the cell; the upper bound of our parameter $R/w$ therefore corresponds to the limit of an absolutely finite horizon, where the particle is trapped in one quadrant of the unit cell or between four abutting scatterers in the unfolded lattice. We do not show any orbits; in both unfoldings, the trajectories from several initial conditions in both configuration and phase spaces are visually identical to those in the specular case, and seem to be statistically identical, though we cannot say so precisely without well-defined invariant measures. Note that with our reflection law links between two parallel flat surfaces happen to reflect specularly, so both Unfolding A and Unfolding B alternate between specular and aspecular reflections, though Unfolding A will have more aspecular events.
We plot the corresponding Lyapunov exponents for Unfolding A and Unfolding B in Fig. \[sinaitriptych\] and Fig. \[lysinb\], respectively. In both cases the asymptotic dependence on the geometry $R/w$ is identical to that known in the specular case [@dp1995; @dahl1997], but the actual value of the exponent is either larger or smaller than in the specular case depending on outgoing angle. For Unfolding A there is the influence of strongly attracting periodic orbits discussed in previous work [@wlst2016] which do not contact the scatterer. These seem to be rather pervasive; we plot the Lyapunov exponent averaged over an increasing number of initial conditions and still find regions of negative Lyapunov exponent. Unlike similar events in Sec. \[sec:stadium\], these are not so easily removed by noise; the range of noise necessary to make all Lyapunov exponents positive, when drawn from a uniform distribution, is proportional to $\frac{w^2 \theta_o}{R^2}$. Unfolding B behaves much more like the specular Sinai billiard, and except for a few instances its Lyapunov exponent landscape is on the same order of magnitude as the specular case.
\[sec:summary\]Summary
======================
Microorganism billiards, in which the outgoing angle is a constant or random variable, provide an example of a physically relevant non-conservative billiard system. The system is both simple and complicated — in many table geometries and for many outgoing angles $\theta_o$, the billiard dynamics will reduce to a single one-dimensional map. We have seen examples where this map has negative Schwarzian derivative everywhere, a condition for which much theoretical work in one-dimensional maps has been accomplished. On the other hand, there is no simple invariant measure, so that many of the standard tools and theorems are not applicable, complicating the cases when the dynamics are words on two generators, the one-dimensional map and rotations which account for orientation reversals due to the microorganisms preserving their forward (tangential) momentum. The system is of interest not only as a novel billiard system, but also because several interest results from billiard theory can be made useful in the study of swimming microorganisms. Bunimovich tables and more generally tables with divided phase space can be utilized to design robust classes of traps, providing essentially infinite tunability to first-passage times and methods of combining and filtering different strains of microorganisms. Sinai tables provide insight into scattering by a lattice of obstacles, and will form a fundamental template for studying and optimizing scattering by differently shaped obstacles in future work. We hope that our work will be of some value to both the study of billiards and the physics of swimming microorganisms.
[^1]: Special thanks to Yuliy Baryshnikov, Sergei Tabachnikov, Vadim Zharnitsky, Maxim Arnold and Saverio Spagnolie for helpful comments and guidance.
[^2]: It is true that in the aforementioned experiments this condition is not always satisfied, however such events seem to be somewhat rare and confined to particular species, so we ignore them in this initial work
[^3]: \[note1\] For all tables under consideration we assume the speed of the particle scales with the table size, so that the table geometry can be specified by a single nondimensional number
[^4]: The systems of interest are hyperbolic, so we assume that the shadowing lemma applies to ensure the applicability of our simulations.
[^5]: Note that depending on definitions the mushroom may not qualify as a true Bunimovich billiard; the system is usually equipped with a condition that the completion of any circular arc must lie entirely within the table boundary
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Stochastic reaction-diffusion models are now a popular tool for studying physical systems in which both the explicit diffusion of molecules and noise in the chemical reaction process play important roles. The Smoluchowski diffusion-limited reaction model (SDLR) is one of several that have been used to study biological systems. Exact realizations of the underlying stochastic process described by the SDLR model can be generated by the recently proposed First-Passage Kinetic Monte Carlo (FPKMC) method. This exactness relies on sampling analytical solutions to one and two-body diffusion equations in simplified protective domains.
In this work we extend the FPKMC to allow for drift arising from fixed, background potentials. As the corresponding Fokker-Planck equations that describe the motion of each molecule can no longer be solved analytically, we develop a hybrid method that discretizes the protective domains. The discretization is chosen so that the drift-diffusion of each molecule within its protective domain is approximated by a continuous-time random walk on a lattice. New lattices are defined dynamically as the protective domains are updated, the numerical convergence and accuracy of our method in for both smooth and discontinuous potentials.
author:
- 'Ava J. Mauro [^1]'
- 'Jon Karl Sigurdsson [^2]'
- 'Justin Shrake [^3]'
- 'Paul J. Atzberger [^4]'
- 'Samuel A. Isaacson [^5]'
title: 'A First-Passage Kinetic Monte Carlo Method for Reaction-Drift-Diffusion Processes'
---
Introduction
============
A fundamental challenge in cell biology is to understand how to predict and control the dynamics of cellular processes [@AlbertsMOLECELLBIO]. Stochasticity in the quantities and movements of molecules can have significant effects on the outcomes of cellular processes, particularly given the low copy numbers of many signaling and regulatory proteins and mRNAs present in a cell. For such species, the actual number and locations of molecules can provide a more accurate and useful description than the local concentration. The method we present in this paper allows for the explicit simulation of the stochastically varying numbers and locations of molecular species undergoing chemical reactions and drift-diffusion.
Experimental studies and mathematical models have shown that stochasticity in the chemical reaction process plays a role in many cellular processes, for example gene expression [@CollinsNatureEukNoise; @ElowitzScienceNoiseExper; @RaserScience2004], cell-fate decision making [@ElowitzNatureCompentence06; @RajComptence07], and signaling pathways in development [@AriasNoiseDevelop06]. Mathematical models of such processes frequently treat an individual cell as a single well-mixed volume or a small number of well-mixed compartments, such as the nucleus and the cytoplasm. However, the heterogeneous spatial distribution of chemical species, as well as interactions with internal membranes and organelles, often have significant effects on cellular processes. For example, after gene regulatory proteins enter the nucleus through nuclear pores, the time required to find specific DNA binding sites can be significantly influenced by the spatial structure of DNA within the nucleus [@IsaacsonPNAS2011]. Similarly, the spatial distribution of components of cellular signaling processes can play a decisive role in the successful propagation of signals from the cell membrane to the nucleus [@MunozGarcia:2009hd; @Kholodenko:2010jv].
A number of stochastic reaction-diffusion models have been introduced to understand the combined influence of noise due to the chemical reaction process and spatial diffusion [@AndrewsMinCDE04; @ElfIEESys04; @WoldeEgfrdPNAS2010; @IsaacsonPNAS2011]. These mathematical models resolve the explicit spatial movement of proteins and mRNAs within cells. There are three such mathematical models that have been commonly used: the spatially-continuous Smoluchowski diffusion-limited reaction model (SDLR) [@SmoluchowskiDiffLimRx; @KeizerJPhysChem82]; what we call the Doi model [@TeramotoDoiModel1967; @DoiSecondQuantA; @DoiSecondQuantB]; and the lattice-based reaction-diffusion master equation model (RDME) [@GardinerRXDIFFME; @McQuarrieJAppProb].
These models often treat the movement of molecules as purely diffusive; however, drift can also play a significant role in the dynamics of cellular processes. Examples of sources of such drift include active transport, variations in chemical potential, material heterogeneities in the cytoplasm and nucleoplasm, and local interactions with cellular structures. The incorporation of drift has played a key role in developing models for molecular-motor based active transport [@PeskinOsterMotor1995; @Atzberger2006], movement of proteins on DNA [@MirnyDNASlide04], and protein movement subject to the influence of volume exclusion by chromatin [@IsaacsonPNAS2011]. It has been shown how to extend the RDME to incorporate drift due to potentials [@IsaacsonPNAS2011]. In the present work we consider a generalization of the SDLR model that allows for drift due to fixed potentials. For creating realizations of the stochastic process described by this model, we present a new numerical method, combining elements of the First-Passage Kinetic Monte Carlo (FPKMC) and the lattice methods of [@ElstonPeskinJTB2003; @IsaacsonPNAS2011]. We will refer to our method as Dynamic Lattice FPKMC or DL-FPKMC.
The SDLR and RDME Models
------------------------
In this subsection we compare the SDLR and RDME models to provide motivation for our choice of the SDLR model. In the standard SDLR approach [@SmoluchowskiDiffLimRx], the positions of molecules are modeled as point particles undergoing Brownian motion. The state of a chemical system is given by the collection of stochastic processes for the positions of each molecule of each chemical species at a given time. Bimolecular, or second-order, reactions between two molecules are modeled as occurring either instantaneously, or , when the molecules’ separation reaches a specified reaction radius [@KeizerJPhysChem82]. Bimolecular reactants are not allowed to approach closer than their reaction radius. Unimolecular, or first-order, reactions involving a single molecule represent internal processes, such as decay or splitting, and are assumed to occur with specified probabilities per unit time. (The Doi model is similar, but allows bimolecular reactants to approach arbitrarily close to each other. Bimolecular reactions then occur with a fixed probability per unit time when the separation between two molecules is *less than* the reaction-radius [@TeramotoDoiModel1967; @DoiSecondQuantA; @DoiSecondQuantB].) The SDLR model can be extended to incorporate drift due to potentials. The underlying reaction process remains unchanged, but molecules move by a drift-diffusion process instead of pure Brownian motion.
In contrast to the spatially-continuous SDLR model, space in the RDME is partitioned by a mesh into a collection of voxels. The diffusion of molecules is approximated by a continuous-time random walk on the mesh, with bimolecular reactions occurring with a fixed probability per unit time for molecules within the same voxel. Unimolecular reactions are modeled in the same manner as in the SDLR model. The state of a chemical system is then described by the collection of stochastic processes for the number of each chemical species within each voxel at a given time. Molecules are assumed to be well-mixed within each voxel (i.e. uniformly distributed). The RDME can be considered an extension of the chemical master equation (CME), a standard non-spatial model for stochastic chemical kinetics [@GardinerRXDIFFME; @McQuarrieJAppProb; @GardinerHANDBOOKSTOCH; @VanKampenSTOCHPROCESSINPHYS; @RamaswamyNatCommun2012]. One advantage of the RDME is the ability to construct a hierarchy of more macroscopic approximations for which efficient numerical solution methods have been developed [@Rossinelli2008136].
When the RDME is interpreted as an independent physical model [@ErbanChapman2009; @RamaswamyJChemPhys2011] there is a nonzero, lower bound on the lattice spacing that arises from the assumption that the diffusive mixing timescale within a voxel is faster than the timescale for a (well-mixed) bimolecular reaction to occur [@IsaacsonSJSC2006; @ElfPNASRates2010]. There is also a simultaneous upper bound on the lattice spacing to ensure the random walk approximation of molecular diffusion is accurate. Only when these bounds are both satisfied is the RDME considered “physically valid”. However, in the absence of nonlinear reactions, the solution to the RDME should *converge* to that of the SDLR model. This can be seen by considering the particle-tracking representation of the RDME derived in [@IsaacsonRDMENote]. It is also standard to choose the diffusive hopping rates in the RDME so that, in the absence of any reactions, the solution to the RDME recovers the Brownian motion of point particles as the lattice spacing is taken to zero [@IsaacsonSJSC2006; @LotstedtFERDME2009]. For these reasons, in applications the RDME is often considered an approximation to the more microscopic SDLR model [@RamaswamyJChemPhys2011; @ElfPNASRates2010; @Hellander:2012jk].
For systems that include bimolecular reactions, interpreting the RDME as an approximation to the SDLR model can be problematic. It has been proven that in the continuum limit that the RDME lattice spacing is taken to zero bimolecular reactions are lost in two or more dimensions [@IsaacsonRDMELims; @IsaacsonRDMELimsII; @Hellander:2012jk]. As such, the time required for two molecules to react becomes infinite. The loss of bimolecular reactions is consistent with the physical lower bound on the lattice spacing, and demonstrates that the RDME can only be interpreted as an approximation to the SDLR model for lattice spacings that are neither too large nor too small. The error of this approximation cannot be made arbitrarily small. In [@IsaacsonRDMELimsII] the two molecule $\textrm{A} + \textrm{B}
\to \varnothing$ reaction was studied in $\R^3$ in both the RDME and SDLR model. It was shown that for certain biologically relevant parameter values the reaction time distribution in the RDME could *at best* approximate the reaction time distribution in the SDLR model to within 5-10% percent for an optimal choice of lattice spacing.
Several recent efforts have derived renormalized bimolecular reaction rates for use in the RDME that are designed to accurately capture *one specific statistic* of the SDLR model over a range of sufficiently large lattice spacings [@ElfPNASRates2010; @Hellander:2012jk]. For example, [@ElfPNASRates2010] matches the mean equilibration time for the $\textrm{A} + \textrm{B} \leftrightarrows \textrm{C}$ reaction in a system with one molecule and one molecule. More recently, a modified convergent RDME (CRDME) that approximates the Doi model was proposed in [@IsaacsonCRDME]. Still, as of yet there is no RDME-like approximation of the (microscopic) SDLR model in which the approximation error can be made arbitrarily small.
There are several additional challenges to using RDME-like lattice models to study cellular processes in realistic geometries. Foremost, approximate the domain geometry by either unstructured meshes [@LotstedtFERDME2009] or Cartesian grid embedded boundary methods [@IsaacsonSJSC2006]. In the former it can be difficult to construct meshes for which a continuous-time random walk approximation of diffusion is well-defined at all mesh voxels when in three dimensions [@LotstedtFERDME2009]. The latter tends to lose accuracy in voxels cut by the boundary [@IsaacsonSJSC2006].
For the preceding reasons, in this work we have chosen to focus on developing a convergent numerical method for directly approximating the SDLR model. In particular, our approach does not constrain molecules to remain on a fixed lattice as in the RDME. While our method is presented in one dimension to illustrate that it is converging to the SDLR model, we expect that the method should be well-suited for handling complex geometries in two and three dimensions by
Exact numerical realizations of the stochastic processes described by the RDME can be generated by the well-known Stochastic Simulation Algorithm (SSA) method [@GillespieJPCHEM1977], which is a Kinetic Monte Carlo (KMC) method [@KalosKMC75]. More recently, the First-Passage Kinetic Monte Carlo (FPKMC) method was developed to generate exact realizations of the stochastic processes described by the SDLR model [@KalosDSMC06; @OppelstrupPRE2009; @DonevJCP2010; @WoldeEgfrdPNAS2010].
An exact FPKMC method has also been introduced incorporating spatially and temporally varying transition or annihilation rates for single particles [@SchwarzRieger2013], which could be used to simulate transitions from diffusive to ballistic modes in models of intracellular transport [@LoverdoNatPhys2008]. It should be noted that there are a number of alternative numerical methods, that have also been proposed for approximating the stochastic processes described by the SDLR or Doi models [@AndrewsBrayPhysBio2004; @ErbanChapman2011; @HellanderGFRRD2011]. Each of these methods is approximate, and ultimately based on a discretization in time.
While it is conceptually simple to modify the SDLR model to include drift due to potentials, numerical realizations of the underlying stochastic processes described by the model. In this work we propose a dynamic lattice version of the FPKMC method to allow for the inclusion of drift due to a potential. While our Dynamic Lattice FPKMC method (DL-FPKMC) no longer generates exact realizations of the SDLR model, the error introduced in the method is controlled. The method can be extended in a straightforward manner to also include spatially varying diffusion coefficients.
A DL-FPKMC Method for the SDLR Model with Drift
-----------------------------------------------
The central idea of the original FPKMC method is to enclose one or two molecules within a ‘protective domain’ that isolates them from all other molecules of the system. A significant change in the state of the system occurs only when a molecule leaves its protective domain or a reaction occurs. In the cases discussed below, first-passage time distributions for such events can be determined from information about the underlying physical system. The first-passage time distributions can then be sampled to determine the time and type of state change that occurs next. This event-driven approach provides an especially efficient simulation algorithm by allowing each update of the algorithm to span the time interval between significant state changes (as opposed to proceeding over many small fixed time steps during which the change in relevant state variables is minor).
In the rather special case of pure Brownian motion in simple domains (spheres or rectangular regions), the first-passage time distributions for a molecule to leave a protective domain [@KalosDSMC06; @OppelstrupPRE2009; @DonevJCP2010; @WoldeEgfrdPNAS2010] or for two molecules to reach a threshold radius for reaction [@WoldeEgfrdPNAS2010] can be computed analytically by solving the diffusion equation. The use of these expressions allows for the generation of exact realizations of the stochastic process described by the SDLR model with the FPKMC. However, for many situations in cell biology, pure Brownian motion does not provide the most realistic description of the movement of molecules as a consequence of active transport, chemical gradients, interactions with cellular structures, etc. In such cases, significant drift terms are inherent to the particle dynamics and can be modeled as arising from a fixed potential field. The DL-FPKMC method we develop extends the FPKMC to allow for such drift. Analytical expressions for the first-passage time distributions from protective domains are no longer possible with the addition of spatially varying drift. In DL-FPKMC, we therefore approximate the drift-diffusion process each molecule undergoes within a protective domain by a continuous-time random walk on a discretized mesh. The “hopping rates” for these walks are finite-difference discretization of the Fokker-Planck equation [@ElstonPeskinJTB2003]. For this reason, our method can be interpreted as a dynamic-lattice master equation model. Unlike the standard RDME, it has the benefit of converging to the SDLR model as the lattice spacing is reduced.
We present results demonstrating both the convergence and accuracy of the DL-FPKMC method in one dimension as the mesh spacing in the discretization is decreased. In particular, we apply our algorithm to the bimolecular reaction $A + B \to \varnothing$ where the molecules of species $A$ and $B$ undergo drift-diffusion subject to various types of potential functions (smooth, discontinuous, and constant). Our results indicate that the method is approximately second-order accurate for smooth potentials and approximately first-order accurate for discontinuous potentials. In this paper, we focus on convergence of the DL-FPKMC in one dimension because of the availability of exact analytical solutions and high-accuracy numerical solutions with which to assess the error. For a one-dimensional domain containing only one molecule of species $A$ and one molecule of species $B$, the SDLR model for the reaction system $A + B \to \varnothing$ can be described by a single two-dimensional PDE for the probability density the particles have not reacted and are located at specified positions. This PDE can be solved numerically to high accuracy with general potential fields using finite difference discretizations, and can solved analytically when the potential is constant. Having these numerical and analytic solutions allows us to check both the accuracy and the convergence of the DL-FPKMC simulation results. If the domain containing the two molecules were instead two- or three-dimensional, the corresponding PDE would be four- or six-dimensional, respectively, and would therefore be challenging to solve to *high accuracy* by standard PDE discretization techniques. For the two-molecule system in one dimension, we provide detailed numerical results demonstrating the convergence of DL-FPKMC to the SDLR model. For systems with more than two molecules, where high-accuracy solutions to the equations for the probability density of being in a given state are not available, we show more qualitative convergence results.
The FPKMC method was originally presented as an efficient way to simulate reaction-diffusion systems at low particle densities “without all the hops” by using larger “superhops” [@KalosDSMC06]. While DL-FPKMC uses more hops than FPKMC due to the random walk approximation of molecular motion, we demonstrate that DL-FPKMC maintains efficiency at low particle densities by requiring far fewer hops than fixed lattice methods with comparable resolution. By discretizing each individual protective region, DL-FPKMC allows fine meshes to be used in localized regions when needed for accuracy considerations. Examples where fine meshes may be necessary include resolving bimolecular reactions, boundary conditions, or rapidly varying potential fields. For protective domains in which such features are not present, coarser meshes can be used. In this way the DL-FPKMC offers an alternative to the types of global adaptive mesh methods that have been proposed for RDME-based models [@Bayati201113].
The paper is organized as follows. Section \[S:overview\] presents our approach for incorporating drift into the SDLR model. In Section \[S:FPKMCSect\] we give an overview of the implementation and steps of the FPKMC or DL-FPKMC algorithm to generate realizations of the stochastic process described by the SDLR model. The general discussion in Sections \[S:overview\] and \[S:FPKMCSect\] assumes the molecules move in $\R^{n}$. The specific implementation we develop in Section \[S:oneDimMethod\] restricts the molecules to intervals in $\R$, but our technique can be extended to higher dimensions through use of the Walk on Rectangles method [@LejayExactWOS2006]. Section \[S:oneDimMethod\] presents our numerical method for using a dynamic lattice to incorporate drift into the FPKMC algorithm, and in Section \[S:methConv\] we demonstrate the convergence and accuracy of this DL-FPKMC method. In Section \[S:RunningTime\] we provide a running time analysis of DL-FPKMC, in which we demonstrate $O(N)$ scaling with the number of molecules in the system and compare DL-FPKMC to a fixed lattice method. Section \[S:Applications\] applications . In \[S:ComparisonPotentials\] we compare the effects of drift due to several potentials on reaction time and location statistics. We conclude in \[S:DNAproteinApplication\] by investigating a simplified model of a coupled protein-polymer fiber system, in which two drift-diffusion along a polymer, and may also unbind from the polymer and diffuse in three dimensions. We study the interaction between
Incorporating Drift into the SDLR Model {#S:overview}
=======================================
The SDLR reaction-diffusion model can be described by a system of partial integro-differential equations (PIDEs) for the probability densities of having a given number of molecules of each chemical species at a specified set of positions, similar to the stochastic reaction-diffusion PIDE models in [@IsaacsonRDMENote] and [@DoiSecondQuantA; @DoiSecondQuantB]. Alternatively, one can consider the collection of stochastic processes for the numbers of molecules and positions of each molecule of each chemical species in the system. Due to the high-dimensionality of the system of PIDEs for the probability densities, numerical methods for solving the SDLR model, including FPKMC methods, are typically based on Monte Carlo approaches that approximate the underlying stochastic processes. In the remainder of the paper, we develop a Dynamic Lattice First-Passage Kinetic Monte Carlo method (DL-FPKMC) for generating realizations of stochastic processes. The realizations are approximate, but the error is small and goes to zero as the lattice spacing is decreased.
In the modified SDLR model that includes drift, molecules are modeled as points or hard spheres undergoing drift-diffusion processes. In a system with $K$ chemical species, we label the $k^{\textrm{th}}$ chemical species by $S^k$, $k = 1,\dots,K$. We denote by the stochastic process for the number of molecules of species $S^k$ at time $t {{\color{black}{\ge 0}}}$. The position vector of the $l^{\textrm{th}}$ molecule of species $S^k$ at time $t {{\color{black}{\ge 0}}}$ is given by the vector stochastic process $\vQ_{l}^{k}(t) \in \R^{n}$, $l = 1, \dots, M^k(t)$.
In the absence of any possible chemical reactions we assume the $l^{\textrm{th}}$ molecule of species $S^k$ with position $\vQ^k_l(t)$ undergoes diffusion with diffusion coefficient $D^k$ and experiences drift due to a potential $V^k \big(\vQ^k_l(t)\big)$. In this case, $\vQ^k_l(t)$ satisfies the stochastic differential equation (SDE) $$\label{eq:driftDiffSDEs}
d \vQ^k_l(t) = \frac{-D^k}{k_{\textrm{B}}{T}} \, \nabla V^k \big(\vQ^k_l(t)\big) \, dt + \sqrt{2 D^k} \, d \vec{W}^k_l(t),$$ where $k_{\textrm{B}}$ is Boltzmann’s constant, $T$ is absolute temperature, $\nabla$ denotes the gradient in the coordinates of $\vQ^k_l(t)$, and ${{\mathbf{W}}}^k_l(t)$ denotes the standard $n$-dimensional Wiener process which describes Brownian motion. Our method for generating realizations of this stochastic process will be presented in Section \[S:oneDimMethod\]. The form of drift due to a potential in Eq. is useful for modeling environmental interactions such as volume exclusion, attractive DNA forces felt by regulatory proteins, and effective potentials felt by molecular motors. It does not allow the possibility of interactions between the diffusing molecules of the chemical system. To incorporate such potentials into our method would be feasible for short-range pair interactions, but for simplicity we assume no potential interactions between molecules.
When bimolecular and unimolecular reactions are added into the system, the molecules continue to move by the drift-diffusion process in Eq. under the additional constraint that any pair of bimolecular reactants are not allowed to approach closer than their corresponding reaction radius. Unimolecular reactions representing internal processes are modeled as occurring with exponentially distributed times based on a specified reaction-rate constant.
We will assume that reaction products are placed at the locations specified in Table \[tab:RxnProductLocations\]. In the cases of two reaction products, for either unimolecular or bimolecular reactions, the angular orientation of the product separation vector about the center of mass is chosen randomly,
When using the partial absorption Robin boundary condition-based bimolecular reaction mechanism, the separation distance can be chosen to be exactly the reaction radius. In the case that two products react by the pure absorption Dirichlet boundary condition-based mechanism, such a choice would lead to an immediate re-association reaction. This issue is generally handled through the introduction of an unbinding radius [@AndrewsBrayPhysBio2004]. In the remainder, we restrict our focus to irreversible bimolecular reactions and assume reactions occur immediately when the separation between two reactants equals a specified reaction radius (a pure-absorption reaction).
[| p[1.9cm]{} ![width 2pt]{} p[4.3cm]{} | p[8cm]{} |]{} & *One Reaction Product* & *Two Reaction Products*\
*Unimolecular Reaction* & The product is placed at the same location as the reactant. & The products are placed a specified distance apart, with their center of mass at the location of the reactant.\
*Bimolecular Reaction* & The product is placed at the center of mass of the two reactants. & The products are placed a specified distance apart, with their center of mass at the same location as the center of mass of the reactants.\
First-Passage Kinetic Monte Carlo Methods {#S:FPKMCSect}
=========================================
Overview of FPKMC Approaches
----------------------------
First-Passage Kinetic Monte Carlo methods have been developed to generate exact realizations of the SDLR model in the absence of drift [@KalosDSMC06; @OppelstrupPRE2009; @DonevJCP2010; @WoldeEgfrdPNAS2010; @SchwarzRieger2013]. These novel algorithms are based on the Walk on Spheres method for solving exit time problems in complicated geometries [@MullerWOSAlgo56]. They rely on being able to derive exact analytical solutions of the diffusion equation in spheres and rectangular solids. In these FPKMC algorithms, a spherical or rectangular region called a ‘protective domain’ is drawn around every molecule in the system, with the collection of protective domains chosen to be disjoint. The first-passage time for each molecule, meaning the time when the molecule will first hit the boundary of its protective domain, can be sampled exactly using the corresponding analytical solution to the diffusion equation. The molecule that exits its protective domain first is updated to its exit position, and a new protective domain is defined. When two reactants are sufficiently close that a sphere can be drawn that contains only them, the corresponding two-body solution to the SDLR PIDEs can be used to exactly sample a candidate time and location for their reaction [@WoldeEgfrdPNAS2010].
A central feature of each of these prior methods is the assumption that the molecular species undergo purely diffusive stochastic dynamics within their protective domains. When drift due to a potential is present in addition to diffusion, the probability densities for the locations of one or two molecules within their protective domains are no longer described by the diffusion equation, but rather by a Fokker-Planck equation. Let $\Omega \subset \R^{n}$ denote the overall domain with boundary $\partial \Omega$, and let $U \subset \Omega$ label a protective domain with boundary $\partial U$. We are interested in the time a molecule first leaves $U$, leading to a zero Dirichlet boundary condition on $\partial U$. Let $V(\vx)$ denote the strength of the potential at $\vx \in \Omega$. The probability density, $\rho(\vx,t)$, for a single molecule to be at location $\vx$ within its protective domain $U$ at time $t > 0$ evolves according to the equations $$\begin{aligned}
\label{eq:FPE_Rn_singlePD}
\frac{\partial \rho( \vx ,t)}{\partial t} = D \nabla \cdot
\left( \rho( \vx ,t) \frac{ \nabla V(\vx )} {k_{\textrm{B}} T} + \nabla \rho(\vx ,t) \right),
& \qquad \text{on } U, \\
\rho (\vx ,t)= 0, & \qquad \text{on } \partial U \backslash (\partial U \cap \partial \Omega), \notag \\
\rho(\vx,0) = \delta(\vx - \vx_0), \notag &\end{aligned}$$ where $\vx_0$ is the initial position of the molecule within the protective domain. If $\partial U$ intersects $\partial \Omega$, the boundary conditions on $\partial U \cap \partial \Omega$ will agree with the boundary conditions on $\partial \Omega$.
For two bimolecular reactants within one protective domain $U
{{\color{black}{\subset \Omega \subset \R^{n}}}}$, the joint probability density $\rho( \vx, \vy ,t)$ for one molecule to be at location $\vx {{\color{black}{\in U}}}$ and the other molecule to be location $\vy {{\color{black}{\in U}}}$ at time $t > 0$ $V_{1}(\vx)$ and $V_2(\vy)$ the potential fields that impart drift to the $\vx$ molecule and the $\vy$ molecule respectively, $$\begin{aligned}
\label{eq:FPE_Rn_pairPD}
\frac{\partial \rho( \vx, \vy ,t)}{\partial t} = {{\color{black}{ \nabla \cdot \mathcal{D} }}}
\left( \rho( \vx, \vy ,t) \frac{ \nabla {{\color{black}{V(\vx , \vy)}}} } {k_{\textrm{B}} T} + \nabla \rho(\vx, \vy ,t) \right),
& \qquad \text{on } {{\color{black}{W}}}, \\
\rho(\vx,\vy,t) = 0, & \qquad \text{on } {{\color{black}{ \partial W_\textrm{rxn} }}}, \notag \\
\rho (\vx, \vy, t)= 0, & \qquad \text{on } {{\color{black}{\partial W
\backslash (\partial W_\textrm{rxn} \cup \partial W_\textrm{outer} )}}}, \notag \\
\rho(\vx, \vy, 0) = \delta(\vx - \vx_0, \vy - \vy_0), \notag \end{aligned}$$ where $\vx_0$ and $\vy_0$ are the initial positions of the molecules within the protective domain, and the gradient and divergence operators are in the $(\vx,\vy)$ coordinates. the single particle case, if $\partial U$ intersects $\partial \Omega$, the boundary conditions on will agree with the boundary conditions on $\partial \Omega$. The boundary condition $\rho(\vx,\vy,t) = 0$ when $||\vx-\vy||
= r_{\textrm{R}}$ models the pure-absorption reaction mechanism. This reactive boundary condition could be modified to use a Robin partial-absorption mechanism if desired. In the remainder we assume all molecules experience the same potential field, so that $V_1=V_2$. In the case that the potential $V$ is a constant function, Eqs. and reduce to diffusion equations.
Our method for using discretizations of these Fokker-Planck equations to sample times and locations of first-passage and reaction events is described in Section \[S:oneDimMethod\].
![\[fig:fpkmcSchematic\] Schematic of the FPKMC or DL-FPKMC algorithm as described in Subsection \[StepsOfAlgorithm\]. Here, $\tau_i$ denotes the next event time for the molecule or pair of molecules in the $i^{th}$ protective domain, and $\tau_{\textrm{next}} = \tau_5$ is the time of a global event.](figure1){width="15cm"}
Main Steps of the FPKMC or DL-FPKMC Algorithm {#StepsOfAlgorithm}
---------------------------------------------
In this subsection we describe the role of protective domains and the processing of events in our implementation of the FPKMC or DL-FPKMC algorithm. We then list the main steps of the algorithm. Our implementation is based on the FPKMC algorithm developed in [@KalosDSMC06; @OppelstrupPRE2009; @DonevJCP2010], with some modifications. The content of this section applies to both the FPKMC and DL-FPKMC algorithms. The difference between the two methods is the approach for sampling event times within a protective domain. In FPKMC, which can only be used in the case of constant $V(\vx)$, event times are sampled from *exact* solutions of the diffusion equation. In DL-FPKMC, which allows for arbitrary $V(\vx)$, event times are *approximated* by generating sample paths of continuous-time random walks on meshes within protective domains (see Section \[S:oneDimMethod\]).
###
To apply the algorithm, every molecule in the system is placed in a protective domain. In one dimension the protective domains are intervals and in higher dimensions the protective domains are usually rectangular or spherical regions. In general, the boundaries of protective domains are absorbing. The boundary of a protective domain can contain a portion of the boundary of the overall spatial domain, in which case the protective domain boundary conditions will depend on the overall domain boundary conditions. We allow protective domains to contain either one or two molecules. Protective domains containing only one molecule are referred to as ‘single protective domains’, and those containing two molecules are referred to as ‘pair protective domains’. Molecules in separate protective domains behave independently. Each molecule undergoes drift-diffusion within its protective domain, and may undergo unimolecular reactions. Two molecules in the same protective domain may additionally participate in bimolecular reactions. To maintain independence when bimolecular reactants are in different protective domains, we require at least one reaction radius. For non-reacting molecular species we allow for overlap to prevent the size of protective domains from going to zero.
###
Each event that may occur will have a type, time, and location. The two major event types are first passage from a protective domain and reaction. First passage from a protective domain occurs when a molecule first reaches an absorbing boundary of its protective domain. In DL-FPKMC with general $V(\vx)$ (resp. FPKMC with constant $V(\vx)$), times and locations for first-passage events from single protective domains are sampled from probability densities determined from approximate (resp. exact) solutions of Eq. . Similarly, for pair protective domains, solutions of Eq. are used to sample times and locations for first-passage events or for bimolecular reactants to first reach a separation of one reaction radius. The time for a unimolecular reaction to occur is sampled from an exponential distribution with a specified reaction rate, and a corresponding reaction location is sampled from a “no-passage” probability density for the molecule involved (see below).
To facilitate the discussion of these events, we use specific names for three times. The ‘global time’ will refer to when the most recent event has occurred, irrespective of its particular type or which molecules were involved. An ‘individual time’ and a ‘next event time’ will be associated with each particular molecule. ‘Individual time’ will refer to when the molecule was last updated, and ‘next event time’ will refer to the sampled time at which the molecule might next undergo an event. Individual times are less than or equal to the global time, and next event times are greater than the global time.
Usually, the individual time and location of a molecule will only be updated when the molecule undergoes a major event (first-passage or reaction). In this case, the time and location of the molecule will be updated to the time and location of the event. However, a molecule can also be updated to any specified time prior to its next event time, by sampling a new position for the molecule within the protective domain from the conditional probability density for the molecule to be at a position within the domain, at the specified time, and not yet have undergone a first-passage or reaction event. This procedure is called a ‘no-passage’ update.
### {#S:algoSteps}
The algorithm is carried out according to the following steps:
1. Protective domains are defined around each molecule or pair of molecules, as shown in Figure \[fig:fpkmcSchematic\]b.
2. The next individual event for each molecule or pair of molecules is determined by sampling an event type, time, and location. In Figure \[fig:fpkmcSchematic\]c, each next event time is labeled by a $\tau_i$.
3. To determine global events, the individual events are sorted in a priority queue ordered from the shortest event time to the longest event time. For example, $\tau_5$ denotes the shortest event time in Figure \[fig:fpkmcSchematic\]c.
4. The next global event is determined from the priority queue using the next individual event with the shortest time. The global time and the individual time(s) of the participating molecule(s) are updated to the event time. In the case of a first-passage event for a molecule to leave its protective domain, the molecule’s location is updated to the sampled first-passage location, as shown in Figure \[fig:fpkmcSchematic\]d. If this molecule is in a pair protective domain with another molecule, the other molecule is no-passage updated to the new global time. In the case of a reaction event, the reaction products are placed at or about the reaction location, as specified in Table \[tab:RxnProductLocations\].
5. Molecules in protective domains that are close to or overlap the newly updated molecules are no-passage updated to the new global time.
6. New protective domains are constructed only for those molecules that have undergone an update to reach the current global time, as shown in Figure \[fig:fpkmcSchematic\]f. New events are sampled for these updated molecules, and the event times are sorted into the priority queue. All other molecules and events remain unchanged.
7. Steps 4 through 6 are then repeated.
Note that Step 5 is used to keep the sizes of the protective domains from becoming too small [@DonevJCP2010], in which case the *effective* time steps used in the FPKMC or DL-FPKMC methods could become very short.
We remark that information about the state of any molecule in the system is available for any particular time in the simulation. For instance, if one would like to sample the locations of all molecules at a specified time, this can obtained by taking the state of the system at the largest global time before or equal to the specified time and then no-passage updating each molecule to the specified time.
Protective Domain Changes during One Simulation
-----------------------------------------------
![\[fig:SampleRun1\] One simulation of the reaction $A + B \to \varnothing $, with one molecule each of $A$ and $B$ present initially and $V(x)=0$. Both panels show the same run of the simulation. In the left panel, the vertical axis is the number of times the simulation cycled through Steps $4$ to $6$ of the algorithm. In the right panel, the vertical axis is the time of the most recent event.](figure2)
During simulations, updates are made to the protective domains sequentially as events occur changing the state of the system. To illustrate this process, we consider the simulation in one dimension of the reaction $A + B \to \varnothing$ starting with one molecule each of $A$ and $B$ and using our DL-FPKMC algorithm. One simulation is shown in Figure \[fig:SampleRun1\].
In the left panel of Figure \[fig:SampleRun1\], the vertical axis is the number of times that the simulation cycled through Steps $4$ to $6$ of the algorithm; we call this number $N_{\textrm{update}}$. In the right panel, the vertical axis is the time of the most recent event. At $N_{\textrm{update}}=1$ in this particular run of the simulation, molecule $A$ is first-passage updated to the right endpoint of its initial protective domain. This location is close to the left endpoint of molecule $B$’s protective domain, so molecule $B$ is no-passage updated and new protective domains are defined around each molecule. From $N_{\textrm{update}}=1$ to $N_{\textrm{update}}=3$, molecule $B$ is first-passage updated but does not come close to the protective domain of molecule $A$, so molecule $A$ is not updated. At $N_{\textrm{update}}=6$, the distance between molecules $A$ and $B$ is less than a specified pair threshold, so they are placed in a pair protective domain. At $N_{\textrm{update}}=7$, the distance between the molecules reaches the reaction radius and the reaction occurs.
{#S:oneDimMethod}
To introduce our methods, we consider the case where the simulation domain is a one-dimensional interval . We assume that bimolecular reactions occur instantaneously when the reactants’ separation reaches the reaction radius. We allow the simulation domain to have reflecting, absorbing, or periodic boundaries. Reflecting boundaries are modeled using zero-flux boundary conditions, and absorbing boundaries are modeled using zero Dirichlet boundary conditions. Protective domains are proper subintervals of the overall domain.
{#S:WPE_LatticeDiscretization}
-
pure diffusion, the probability distributions for first-passage times, first-passage locations, and no-passage locations can all be determined from analytic solutions of the diffusion equation [@KalosDSMC06; @OppelstrupPRE2009; @DonevJCP2010; @WoldeEgfrdPNAS2010]. In contrast, once drift is considered, such analytic approaches are no longer possible in general. Instead one must consider probability densities that satisfy Fokker-Planck equations such as Eqs. and , in which we restrict to the case where the drift arises from a spatially varying potential energy function $V(x)$.
in DL-FPKMC, we introduce approximations by treating the movement of each molecule within its protective domain as a discrete-space continuous-time Markov chain, more specifically a continuous-time random walk on discrete mesh points. Jump rates between neighboring mesh points are obtained using the Wang–Peskin–Elston [@ElstonPeskinJTB2003] (WPE) spatial discretization of the one-dimensional Fokker-Planck equation $$\label{eq:FPE}
\frac{\partial \rho(x,t)}{\partial t} = D \frac{\partial }{\partial x}
\left(\rho(x,t) \frac{d V(x)}{d x} +
\frac{\partial \rho(x,t) }{\partial x} \right).$$ We remark that the factor $(k_B T)^{-1}$ is absorbed into the potential function $V$. The WPE discretization weights then determine the jump rates (i.e. probabilities per unit time) for molecules to hop from one mesh point to another. In particular, the jump rate for a molecule to hop from the mesh point $x_i$ to a neighboring mesh point $x_j$, in the case of a uniform mesh of width $h$, is given by
$$\label{eq:JumpRate}
a_{ij} = \left\{
\begin{array}{cl}
\frac{D }{h^2} \frac{V(x_j) - V(x_i)} {\exp[V(x_j) - V(x_i) ] -1} & \text{for } V(x_i)\ne V(x_j) \\
\frac{D }{h^2} & \text{otherwise. } \end{array} \right.$$
Let $p_i(t)$ be the probability that a molecule is located at mesh point $x_i$ at time $t$. Then the time evolution of $p_i(t)$ is described by the master equation $$\label{eq:MasterEqn}
\frac{d \, p_i(t)}{d t} =
a_{i-1,i} \, p_{i-1}(t) - (a_{i,i-1} + a_{i,i+1}) \, p_i(t) + a_{i+1,i} \, p_{i+1}(t).$$ If $x_{i \pm 1}$ is an absorbing boundary, then $p_{i \pm 1}(t)=0$ in Eq. . We shall extend Eqs. and for non-uniform discretizations and in Appendix \[S:DerivationNonUnifJumpRates\].
The discretization given by Eq. has the following properties:
- Converges at second-order for smooth potentials, and can handle discontinuous potentials [@ElstonPeskinJTB2003].
- Satisfies a discrete version of detailed balance (zero net flux at equilibrium), which helps reduce artificial drift due to numerical discretization errors [@ElstonPeskinJTB2003].
- Is consistent with the standard second-order-accurate discretization of the Laplacian operator, in that $a_{ij}$ converges to $ D / h^2$ as $V(x_j) - V(x_i)$ approaches zero.
- *Can be extended to higher dimensions. The jump rates in each coordinate are then given by Eq. .*
- Can incorporate a spatially dependent diffusion coefficient $D(x)$. For example, in the case that $D(x)$ is continuous, the constant $D$ in Eq. can be replaced by $[D(x_i) + D(x_j)]/2$ [@ElstonPeskinJTB2003].
{#S:SamplePaths}
To make use of the discretization, a mesh is defined within each protective domain so that every molecule is located at a mesh point. Exact sample paths of the molecules’ random walks are generated using the event-driven Stochastic Simulation Algorithm In this method, the times of the hops are sampled from exponential distributions. There is no fixed time step. By varying the mesh width, the resolution of this process can be adjusted depending on the desired trade-off between computational efficiency and accuracy. Our specific approach for choosing the mesh width and the locations of mesh points is described in more detail in Subsection \[S:ChoosingMeshWidth\].
Any protective domain endpoints on the interior of the overall domain are absorbing, as are endpoints that coincide with an absorbing boundary of the overall domain. If one endpoint of a protective domain is located at a reflecting boundary of the overall domain that endpoint is made reflecting. each protective domain either has absorbing Dirichlet boundaries ($\rho = 0$) at both endpoints, or an absorbing boundary at one endpoint and a reflecting boundary ($ \rho
\frac{d V}{d x} + \frac{\partial \rho}{ \partial x} = 0$) at the other.
For a newly constructed protective domain containing a single molecule, we determine the molecule’s next event time by an exact random-walk path for the molecule to hop on the mesh points until it reaches an absorbing boundary of the protective domain. The time that molecule reaches an absorbing endpoint is the first-passage time, and the endpoint that the molecule reaches is the first-passage location. For pair protective domains with two molecules, we perform random walks for each molecule until either: (i) one molecule reaches an absorbing boundary of the protective domain; or (ii) the distance between the two molecules is equal to the reaction radius $r_\textrm{R}$. The mesh width for pair protective domains is always chosen to exactly divide $r_\textrm{R}$, so that the reaction occurs when the two molecules are exactly one reaction radius apart. A no-passage location at any specified time before the next event time can be obtained by finding the last time in the sample path less than or equal to the specified time and taking the location of the molecule at that time.
{#S:NonUnif}
Non-uniform mesh cells are used when needed to conform to a boundary or to move molecules onto a uniform mesh where the mesh width exactly divides the reaction radius, as will be described in Subsection \[S:ChoosingMeshWidth\]. Let $x_0$ be the initial location of a molecule on a non-uniform mesh, with $x_1$ and $x_2$ denoting the locations of the neighboring mesh points in either direction. Note that we may have either $x_1 < x_0 < x_2$ or $x_2 < x_0 < x_1$ (see Fig. \[fig:NonUnifSubLattice\], top row). Let $h_1 = |x_0 - x_1|$ and $h_2 = |x_0 - x_2|$. The jump rates from $x_0$ to $x_j$ for $ j = 1,2$ are given by
$$\label{eq:NonUnifJumpRate}
a_{0j} = \left\{
\begin{array}{cl}
\frac{2 D}{h_j (h_1+h_2) } \frac{V(x_j) - V(x_0) }{\exp[V(x_j) - V(x_0) ] -1} & \text{for } V(x_i)\ne V(x_j) \\
\frac{2 D}{h_j (h_1+h_2) } & \text{otherwise. } \end{array} \right.$$
The non-uniform discretization in Eq. is derived in Appendix \[S:DerivationNonUnifJumpRates\] by modifying the WPE discretization of the Fokker-Planck equation [@ElstonPeskinJTB2003]. In the case of constant $V(x)$, the non-uniform discretization reduces to the non-uniform spatial discretization of the Laplacian at a Dirichlet boundary given by equation (20) of [@FedkiwJCP2002]. For solving the Poisson equation with Dirichlet boundary conditions, using a uniform interior mesh and non-uniform mesh cells at the boundaries, this discretization is second-order accurate [@FedkiwJCP2002]. To our knowledge, Eq. gives a new discretization of the Fokker-Planck equation for non-uniform meshes.
Choosing the Mesh within Protective Domains {#S:ChoosingMeshWidth}
-------------------------------------------
Single-molecule protective domains with absorbing boundaries are chosen to be symmetric about the location of the molecule. A maximum mesh width for all single protective domains, $h_s^{\textrm{max}}$, is specified . Then, for each individual protective domain, the mesh width $h_s$ is chosen to be the largest value less than or equal to $h_s^{\textrm{max}}$ that exactly divides is constructed so that the molecule and both endpoints of the protective domain lie exactly on mesh points. Having the endpoints lie on mesh points allows enforcement of the absorbing Dirichlet boundary conditions at the endpoints the jumps rates in Eq. .
![\[fig:NonUnifBoundary\] Non-uniform mesh cell at an absorbing Dirichlet boundary of a pair protective domain (left panel), or a reflecting boundary of a pair or single protective domain (right panel). Arrows are shown only where the jump rates differ from those given by Eq. for uniform mesh cells. the jump rates $a_{01}$ and $a_{02}$ are given by the non-uniform rates in Eq. with $h_1 = h_p$ and $h_2 = h_{\textrm{abs}}$. At a reflecting boundary, the distance from the boundary to the nearest mesh point, $x_0$. $h_{\textrm{ref}}/2 \le h_p$ or $h_s$. In this case, from $x_0$ toward the boundary due to the reflecting boundary condition. The jump rate $a_{01}$, going away from the reflecting boundary, is given by Eq. with $h_1 = h_p \textrm{ or } h_s$, and $h_2 = h_{\textrm{ref}}$. ](figure4){width="16cm"}
For single-molecule protective domains with one absorbing endpoint and one reflecting endpoint, the mesh width $h_s$ is chosen to be the largest value less than or equal to $h_s^{\textrm{max}}$ that exactly divides the distance from the molecule to the absorbing endpoint. A mesh is defined so that the molecule and the absorbing endpoint lie exactly on mesh points. The mesh is uniform with the exception of one non-uniform cell used immediately adjacent to the reflecting boundary, as shown in Figure \[fig:NonUnifBoundary\] (right panel).
In pair protective domains, the mesh width $h_p$ is a specified value chosen to exactly divide $r_\textrm{R}$. Each time that two molecules are placed in a new pair protective domain, the initial distance between the molecules will not necessarily be divisible by $h_p$. Rather than perturbing the molecules, non-uniform mesh cells are used to move one of the molecules, as shown in Figure \[fig:NonUnifSubLattice\], so that a uniform mesh of width $h_p$ can be defined with both molecules lying exactly on mesh points. Since this uniform mesh is chosen based on the locations of the two molecules, the endpoints of the protective domain may not conform with the mesh. In this case, one non-uniform mesh cell is used at each endpoint, which may be absorbing or reflecting (see Figure \[fig:NonUnifBoundary\]).
Since $h_p$ is always chosen to exactly divide $r_{\textrm{R}}$, it necessarily follows that $h_p \le r_{\textrm{R}}$. For single molecule protective domains we allow $h_s^{\textrm{max}}$ to be larger than $r_{\textrm{R}}$. In the convergence studies , we set $h_s^{\textrm{max}} = k h_p$ where $k \ge1$, and hold the ratio of $h_s^{\textrm{max}}$ to $h_p$ constant as both are reduced to study convergence. As discussed , the actual mesh widths $h_s$ used in single protective domains are almost always strictly less than $h_s^{\textrm{max}}$, and non-uniform mesh widths are used in both single and pair protective domains. For these reasons, we keep track of the mean of the mesh widths that are actually used in each simulation. In calculating this mean, each mesh width is weighted by the number of times that it is actually used in a sample path. Then, for all simulations performed with fixed $h_s^{\textrm{max}}$ and $h_p$, we calculate an overall mean mesh width by taking the arithmetic mean of the means for each simulation.
Convergence of DL-FPKMC in One Dimension {#S:methConv}
========================================
for the annihilation reaction $A + B \to \varnothing$, where the molecular species $A$ and $B$ undergo drift-diffusion subject to various potentials on the interval $[0,1]$. Our results demonstrate both the convergence and accuracy of our method as the mesh widths, $h_s^{\textrm{max}}$ and $h_p$, in the discretization are decreased. We denote by $M^A(t)$ and $M^B(t)$ the number of molecules of $A$ and $B$, respectively, at time $t$. In the first set of convergence studies only two molecules are simulated, $M^A(0) = M^B(0) = 1$. Each simulation runs until the two molecules have reacted. A large number ($10^7$) of simulations are performed in order to minimize the statistical error, so that the error due to the spatial discretization and the rate of convergence can be studied. In the next set of convergence studies multiple molecules each of $A$ and $B$ are simulated, $M^A(0) = M^B(0) = 10$ or $M^A(0) = M^B(0) = 50$, and each simulation runs until all the molecules have reacted. We will denote the $i^{\textrm{th}}$ molecule of species $A$ by $A_i$, and the location of $A_i$ at time $t$ by $Q_i^A(t)$. $B_j$ and $Q_j^B(t)$ are defined analogously. In the case $M^A(0) = M^B(0) = 1$, we will drop the subscripts $i$ and $j$.
### {#section-4}
The convergence studies are performed using three different potential functions: (i) zero potential, $V_{\textrm{zero}}(x)=0$ (which results in pure diffusion); (ii) a cosine potential with two energy wells, $V_{\cos}(x)= \cos(4 \pi x)$; and (iii) a step potential with one step, $$V_{\textrm{step}}(x) = \left\{
\begin{array}{rl} 2 & \text{if } x < \frac{1}{2}\\
0 & \text{if } x \ge \frac{1}{2} \ . \end{array} \right.$$ The step potential is used to demonstrate that our DL-FPKMC algorithm with the WPE discretization of the Fokker-Planck equation can handle discontinuous potentials. Note that adding a constant to any potential would not change the results, since the Fokker-Planck equation depends on the derivative of the potential but not the potential itself. In particular, any constant potential would produce the same results as $V(x)=0$.
In all the convergence studies, the length $L$ of the overall domain is $1$ unit, the boundaries of the overall domain are reflecting, and the diffusion coefficient $D$ is $1$ unit$^2/$sec for both $A$ and $B$. The values used for the reaction radius $r_\textrm{R}$ will be specified in each subsection. We will use the notation $\mathcal{U}(a,b)$ for the uniform random distribution on the interval $(a,b)$. The initial locations $Q_i^A(0)$ and $Q_j^B(0)$ are drawn from $\, \mathcal{U}(a,b)$, where $(a,b) \subseteqq (0,L)$ will be specified in each subsection. If $| Q_{i*}^A(0) - Q_{j*}^B(0)| \le
r_{\textrm{R}}$ for some $i^*$ and $j^*$, then $A_{i*}$ and $B_{j*}$ will react immediately, at $t=0$. For $ t > 0 $, a reaction occurs if $| Q_{i*}^A(t) - Q_{j*}^B(t)| = r_{\textrm{R}}$.
### {#section-5}
In general, to describe a stochastic reaction-drift-diffusion system of many molecules by the probability density of having a given number of molecules at specified positions, a large coupled system of partial integro-differential equations is required [@IsaacsonRDMENote]. In the special case of only two molecules, $M^A(0) = M^B(0) = 1$, with both molecules having the same diffusion coefficient $D$, the reaction-drift-diffusion system $A + B
\to \varnothing$ in 1D can be described by a single 2D PDE: a Fokker-Planck equation (or a diffusion equation when $V$ is constant). Let $\rho(x,y,t)$ denote the probability density for finding molecule $A$ at location $x$ and molecule $B$ at location $y$ at time $t> 0$. $\rho(x,y,t)$ evolves according to the equations $$\begin{aligned}
\label{eq:FPEbvp}
\frac{\partial \rho(x,y,t)}{\partial t} = D \nabla \cdot \left( \rho(x,y,t)
\left( \frac{d V(x)}{d x} + \frac{d V(y) }{d y} \right) + \nabla \rho(x,y,t) \right),
& \qquad \textrm{on } {{\color{black}{ \Omega}}}, \notag \\
\rho(x,y,t) = 0, & \qquad \textrm{on }
{{\color{black}{ \partial \Omega \cap \{ |x-y| = r_{\textrm{R}} \} }}}, \\
{{\color{black}{ \rho (x,y,t) \frac{\partial}{\partial \veta} (V(x)+V(y)) + }}}
\frac{\partial \rho (x,y,t)}{\partial \veta } = 0, & \qquad \textrm{on }
{{\color{black}{ \partial \Omega \backslash \{ |x-y| = r_{\textrm{R}} \} }}} ,
\notag\end{aligned}$$ where $\veta = \veta(x,y)$ denotes the outward pointing normal at the point $(x,y)$, The 2D domain for Eq. is illustrated in Appendix \[S:AnalytlcAndNumericalPDEsolns\], Figure \[fig:PDEdomains\]. When the initial locations of the two molecules in the DL-FPKMC simulations are drawn from $\,
\mathcal{U}(0,L)$, the corresponding initial condition for the 2D Fokker-Planck or diffusion equation is a constant, $\rho(x,y,0) =
1/L^2$. Note, in the following we define $\rho$ on $0 \leq x, y \leq
L$ by also defining $\rho(x,y,t) = 0$ for $\abs{x-y} < r_{\textrm{R}}$ and $t > 0$. The 2D Fokker-Planck and diffusion equations can both be solved numerically by finite difference methods, and the diffusion equation can be solved analytically using an eigenfunction expansion. These numerical and analytic solutions are discussed in Appendix \[S:AnalytlcAndNumericalPDEsolns\], and provide a baseline against which we compare the results of two-molecule DL-FPKMC simulations in Subsection \[S:2ParticleConvResults\].
Let $T$ denote the random variable for the time at which the two molecules react. Using the solution $\rho(x,y,t)$ of Eq. , we can calculate the survival probability, $$\begin{aligned}
\label{eq:SurvProb}
S(t) & = \Pr \brac{ T > t } = \int_0^L \int_0^L \rho(x,y,t) \, dx \, dy ,\end{aligned}$$ and the mean reaction time, $$\label{eq:MeanRxnTime}
\avg{T} = -\int_0^\infty t \, S'(t) \, dt = \int_0^\infty S(t) \, dt .$$ Note that $1-S(t)$ is the reaction time distribution function
### {#section-6}
In what follows, we will use the term ‘statistical error’ to refer to the difference between the empirical value of a statistic (e.g., mean reaction time) estimated from the DL-FPKMC simulations and the upper, or lower, bound of the $99\%$ confidence interval for the statistic. By ‘discretization error,’ we will mean the difference between the empirical value from the DL-FPKMC simulations and the actual value. In the two-molecule case, actual values are known exactly from the analytic solution $\rho(x,y,t)$ when $V$ is constant, and are estimated from the numerical PDE solver described in Appendix \[S:AnalytlcAndNumericalPDEsolns\] when $V$ is not constant.
Since we perform a large number of simulations ($10^7$) at each mesh width in the two-molecule case, the statistical error is quite small, generally between $0.04\%$ and $0.19\%$ for statistics. Although we perform fewer simulations ($4 \times 10^4$) when using multiple molecules each of species $\textrm{A}$ and $\textrm{B}$, the statistical error is still reasonably small, generally between $0.4\%$ and $1\%$. Our results show that as the mesh width is decreased, the discretization error rapidly decreases to below the statistical error. This demonstrates that the DL-FPKMC algorithm converges and accurately resolves the underlying reaction-drift-diffusion processes.
Two-Molecule Convergence Studies {#S:2ParticleConvResults}
--------------------------------
In this subsection we consider the reaction $A + B \to \varnothing$ for a system with $M^A(0) = M^B(0) = 1$. Here, the initial locations $Q^A(0)$ and $Q^B(0)$ are drawn randomly from $\, \mathcal{U}(0,1)$; the reaction radius $r_{\textrm{R}}$ is $0.02$ units; and the pair threshold $r_{\textrm{pair}}$ is equal to $2 r_{\textrm{R}}$ (i.e. the molecules are placed in a pair protective domain when $| Q^A(t) -
Q^B(t) | \le r_{\textrm{pair}}$). We study the convergence of as the mesh widths $h_s^{\textrm{max}}$ and $h_p$ are reduced.
For $V=0$, the errors in the DL-FPKMC simulation results are calculated relative to the exact analytic solution, as determined from the eigenfunction expansion in Appendix \[S:AnalytlcAndNumericalPDEsolns\]. For $V_{\cos}$ and $V_{\textrm{step}}$, the errors are relative to the numerical solution from the PDE solver described in Appendix \[S:AnalytlcAndNumericalPDEsolns\]. .
As the mesh widths in the DL-FPKMC simulations are decreased, for all three potentials. In each of the panel results for $V=0$, the panel for $V_{\cos}$, and the panel for $V_{\textrm{step}}$.
### {#section-7}
Let $\E_{\textrm{emp}}[T]$ be the empirical mean reaction time calculated from the DL-FPKMC simulations. Define $\E_{\textrm{upp}}[T]$ of the $99\%$ confidence interval for the empirical mean reaction time. We denote by $\E_{\textrm{act}}[T]$ the exact analytic mean reaction time in the case that $V=0$, or the mean reaction time determined from the numerical PDE solution in the $V \ne 0$ cases . We calculate the relative error by $$\label{eq:RelErrMeanRxnTm}
\frac{ | \E_{\textrm{act}}[T] - \E_{\textrm{emp}}[T] | }{ \E_{\textrm{act}}[T] }
\pm \frac{ {{\color{black}{ | \E_{\textrm{bd}}[T] - \E_{\textrm{emp}}[T] | }}}}{ \E_{\textrm{act}}[T] }
{{\color{black}{ \quad \text{ where } \E_{\textrm{bd}}[T] = \E_{\textrm{upp}}[T] \text{ or } \E_{\textrm{low}}[T]. }}}$$
![\[fig:MeanRxnTimes\] Convergence of mean reaction time, $\avg{T}$, for the two-molecule $ A + B \to \varnothing $ reaction, as the mesh width is decreased. Note the expanded scales of the vertical axes. Each DL-FPKMC data point shows $\E_{\textrm{emp}}[T]$ from $10^7$ simulations, $99\%$ confidence intervals. are shown here in the $V=0$ case to demonstrate the . The mean reaction time calculated from the numerical PDE solution was resolved to an absolute error tolerance of $10^{-5}$ for $V_{\cos}$ and $10^{-4}$ for $V_{\textrm{step}}$. ](figure5)
![\[fig:RelErrMeanRxnTimes\] Relative errors in the empirical mean reaction time, $\E_{\textrm{emp}}[T]$, for the two-molecule $A + B \to
\varnothing$ reaction. Note that the vertical axes have different scales in each panel. Each data point is based on $10^7$ simulations with fixed $h_s^{\textrm{max}}$ and $h_p$. The error bars are determined by Eq. using the $99\%$ confidence intervals for the empirical mean reaction times. ](figure6)
Figure \[fig:MeanRxnTimes\] $\E_{\textrm{emp}}[T]$ plotted against $h_s^{\textrm{max}}$ as the mesh widths are varied, \[fig:RelErrMeanRxnTimes\] the relative errors in $\E_{\textrm{emp}}[T]$ plotted against the mean mesh width. The insets in Figure \[fig:MeanRxnTimes\] show the respective potentials. the discretization errors decrease to less than the corresponding statistical errors. Note that the statistical errors are very small since $10^7$ simulations were performed at each mesh width.
We estimate the rate of convergence to be approximately second-order for $V_{\cos}$ and approximately first-order for $V_{\textrm{step}}$. This is consistent with the convergence rates of the WPE discretization of the Fokker-Planck equation for smooth and discontinuous potentials. In the $V=0$ case, it is difficult to draw a conclusion about the rate of convergence since the discretization errors are small relative to the statistical errors; however, for the same reason, we can conclude that very accurate in this case.
### {#section-8}
![\[fig:SurvProb\] ](figure7)
For each mesh width, the empirical survival probability and the associated $99\%$ confidence bounds are calculated using the MATLAB function ‘`ecdf`’. $S_{\textrm{act}}(t)$ the analytic numerical survival probability .
The discrete $L^1$, $L^2$, and $L^\infty$ norms of a function $u(t)$ are given by The relative error for $S_{\textrm{emp}}(t)$ we report in each norm is then given by $$\label{eq:RelErrSurvProb}
\frac{ || S_{\textrm{act}}(t) - S_{\textrm{emp}}(t) || }{|| S_{\textrm{act}}(t) ||},$$ where the norms are evaluated on the interval $t\in [0, {{\color{black}{\
S^{-1} ( 10^{-6} ) }}} ] $ $\Delta{t_i} = t_{i+1} -
t_{i}$. The time points, $t_i$, used in evaluating the norms correspond to those at which the numerical PDE solutions were calculated (see Appendix \[S:AnalytlcAndNumericalPDEsolns\] for more information). In the $V=0$ case, the analytic expression for the survival probability was evaluated at those same $t_i$’s. In all cases, the empirical DL-FPKMC survival probabilities were linearly interpolated to obtain values at every $t_i$.
![\[fig:RelErrSurvProb\] Errors in the empirical survival probability $S_{\textrm{emp}}(t)$, for the two-molecule $ A + B \to
\varnothing $ reaction: $h_s^{\textrm{max}}
= 8 h_p$ . Each empirical survival probability function is based on $10^7$ simulations. ](figure8)
### {#section-9}
### {#section-10}
Results of Multiple-Molecule Convergence Studies {#S:MultParticleConvResults}
------------------------------------------------
In the convergence studies for multiple molecules undergoing the reaction $A + B \to \varnothing$, we start the simulations with either $20$ molecules ($M^A(0) = M^B(0) = 10$) or $100$ molecules ($M^A(0) = M^B(0) = 50$). In the $20$-molecule simulations: $r_{\textrm{R}}$ = $0.02$ units; $r_{\textrm{pair}} =
2r_{\textrm{R}}$; $Q_{i}^A(0) \sim \mathcal{U}(0.1,0.4)$ and $Q_{j}^B(0) \sim \mathcal{U}(0.6,0.9)$ for $ 1 \le i, \, j \le 10$. In the $100$-molecule simulations: $r_{\textrm{R}}$ = $0.001$ units; $r_{\textrm{pair}} = 4r_{\textrm{R}}$; $Q_{i}^A(0)$ and $Q_{j}^B(0)
\sim \mathcal{U}(0,1)$ for $ 1 \le i, \, j \le 50$. Two molecules $A_{i*}$ and $B_{j*}$ are placed in a pair protective domain if they are closer to each other than to any other molecules of the opposite type, i.e. if $$| Q_{i*}^A(t) - Q_{j*}^B(t) | = \min_i | Q_{i}^A(t) - Q_{j*}^B(t) | = \min_j | Q_{i*}^A(t) - Q_{j}^B(t) |,$$ and if the distance between them satisfies $$\begin{aligned}
\label{eq:PairConditions}
| Q_{i*}^A(t) - Q_{j*}^B(t) | &\le r_{\textrm{pair}} \notag \\
\text{ and}& \notag \\
| Q_{i*}^A(t) - Q_{j*}^B(t) | &\le r_{\textrm{R}} + \min( \min_{i \ne i*}( | Q_{i*}^A(t) - Q_{i}^A(t) | , \ \min_{j \ne j*} | Q_{j*}^B(t) - Q_{j}^B(t) | ).\end{aligned}$$ The last condition in Eq. was added to prevent the length of the protective domain of a molecule, say $A_i$, from approaching zero when $A_i$ is close to another molecule of the same type, say $A_{i*}$, where $A_{i*}$ would otherwise have been placed in a pair with $B_{j*}$.
![\[fig:MeanTimeLastRxn\] Mean time for all molecules to react via the reaction $A + B \to \varnothing$. Error bars show 99% confidence intervals, based on $4 \times 10^4$ simulations per data point. *Left panel*: $r_{\textrm{R}} = 0.02$ units. *Right panel*: $r_{\textrm{R}} = 0.001$ units. ](figure11){width="15cm"}
![\[fig:NmoleculeDistributionsVcos\] Mean number of molecules of $A$ remaining at time $t$, $\E_{\textrm{emp}}[M^A(t)]$ (left panel), and probability that at least one molecule of $A$ remains at time $t$, $\Pr \brac{ M^A(t) \ge 1}$ (center and right panels). $V= \cos(4 \pi x)$. Each graph is based on $4 \times 10^4$ simulations. $\E_{\textrm{emp}}[M^A(t)]$ is not plotted in the case $M^A(0) = M^B(0) = 50$, because the results for the different mesh widths are essentially indistinguishable until $\E_{\textrm{emp}}[M^A(t)] \lesssim 4$. *Left and center*: $r_{\textrm{R}} = 0.02$ units, $h_s^{max} = 8 h_p$. *Right*: $r_{\textrm{R}} = 0.001$ units, $h_s^{max} = 200 h_p$. ](figure12)
Each simulation runs until all of the molecules have reacted. Let $T_n$ denote the random variable for the time at which the $n^{\textrm{th}}$ reaction occurs. Since we are working with the irreversible reaction $A + B \to \varnothing$, we have that $M^A(t) =
M^A(0) - n$ for $t \in [T_n, T_{n+1})$, and similarly for $M^B(t)$. Figure \[fig:MeanTimeLastRxn\] shows the empirical mean time for all molecules to react; this is $\E_{\textrm{emp}}[T_{10}]$ when $M^A(0) =
M^B(0) = 10$ and $\E_{\textrm{emp}}[T_{50}]$ when $M^A(0) = M^B(0) =
50$. With these values of $M^A(0)$ and $M^B(0)$, the SDLR model described in Section \[S:overview\] will correspond to a large coupled system of partial integro-differential equations for the probability densities of having a specified number of molecules at specified locations. It is no longer feasible to solve these equations numerically to obtain high-accuracy solutions for assessing the empirical convergence of our DL-FPKMC method. As such, we now estimate the accuracy of reaction time statistics by comparing DL-FPKMC results at coarser mesh widths to results obtained with the finest mesh width.
When $V(x)= \cos(4 \pi x)$ (Fig. \[fig:MeanTimeLastRxn\], top panels), we see convergence as the mesh width is decreased. The percent difference between $\E_{\textrm{emp}}[T_{10}]$ for the coarsest mesh width compared to the finest mesh width is approximately $5.4\%$; this percent difference for $\E_{\textrm{emp}}[T_{50}]$ is approximately $7\%$. These differences are comparable in size to the explicitly calculated discretization error of approximately $6.4\%$ at the coarsest mesh width in the two-molecule case (previous subsection).
When $V=0$ (Fig. \[fig:MeanTimeLastRxn\], bottom panels), the confidence intervals for all mesh widths overlap, indicating that the discretization error is less than the statistical error even for the coarsest mesh width. The statistical errors when $V=0$ are between $0.49\%$ and $0.81\%$. If the unknown discretization error here is comparable in size to the known discretization error in the two-molecule case, which was approximately $0.87\%$ at the coarsest mesh width, that would provide an explanation for why can not observe convergence when $V=0$.
$\E_{\textrm{emp}}[M^A(t)]$ is the mean number of $A$ molecules remaining at time $t$, and $\Pr \brac{ M^A(t) \ge 1}$ is the probability that at least one molecule of $A$ remains at time $t$. Figure \[fig:NmoleculeDistributionsVcos\] shows convergence of $\E_{\textrm{emp}}[M^A(t)]$ and $\Pr \brac{ M^A(t) \ge 1} $ as the mesh width is decreased, in the case $V(x)= \cos(4 \pi x)$. Results are not plotted for the $V=0$ case, because the confidence bounds for different mesh widths overlap; this indicates that the results are resolved to within the statistical error, even for coarser mesh widths.
Running Time Analysis {#S:RunningTime}
=====================
In this section we demonstrate that the running time of the DL-FPKMC algorithm when simulating the $A + B \to \varnothing$ reaction scales linearly with the number of molecules in the system. We also compare the computational performance of DL-FPKMC to a second method in which all molecules hop on a fixed, global, uniform lattice. Both methods were implemented in MATLAB, and an attempt was made to take advantage of reasonable and standard optimizations. That said, we make no claim that our implementation of either method provides optimal computational performance. All DL-FPKMC and fixed lattice simulations were performed in MATLAB on a Sun Fire X4600 M2 x64 server. The server was configured with four AMD Opteron Model 8220 processors (2.8 GHz dual-core) and 16 GB of RAM.
In both methods, bimolecular reactions occur when the distance between two reactants is exactly equal to $r_\textrm{R}$. To enforce this condition in the fixed lattice method, the lattice spacing $h$ will be chosen to equal $r_\textrm{R}$. We set $V(x)=0$, so the spatial hopping rates in the fixed lattice method are simply $D/h^2$. The results of the convergence studies in Section \[S:methConv\] indicate that in DL-FPKMC taking $h_p
\approx r_\textrm{R}$ and $h_s^{\textrm{max}} \approx L/50$ is sufficient to resolve the reaction and diffusion processes to within statistical error for $n \le 10^7$ simulations. For biologically relevant parameter values, e.g. $L = 10 \mu$m and $r_\textrm{R} =
1$nm, this will allow $h_s^{\textrm{max}}$ in DL-FPKMC to be a couple orders of magnitude larger than $h$ in the fixed lattice method without compromising accuracy.[^6]
To our knowledge, the only reaction system that has previously been simulated *in one dimension* using any FPKMC-type method is the $A + A \to \varnothing$ [@KalosDSMC06; @OppelstrupPRE2009] reaction, in which any two molecules annihilate as soon as they collide. In contrast, for the reaction system $A + B \to
\varnothing$, molecules of the same type do not react. As mentioned in Subsection \[StepsOfAlgorithm\], we allow the protective domains of non-reacting molecules to overlap to prevent the size of protective domains from going to zero when two non-reacting molecules approach each other. We expect that this issue could also be addressed by allowing protective domains to contain more than two molecules. In order to simulate the same underlying process with both DL-FPKMC and the fixed lattice method, non-reacting molecules in the fixed lattice simulations are allowed to occupy the same lattice site and cross each other.
{#section-11}
In comparing the DL-FPKMC and fixed lattice methods, the following parameters values are used: $L = 10 \mu \text{m}$, $r_\textrm{R} =
1\text{nm}$, $D = 10 \mu \text{m}^2/ \text{sec}$, and $V(x) = 0$. The overall simulation domain, the interval $(0, L)$, has reflecting boundaries. The initial number of molecules of $A$ and $B$ are equal, $M^A(0) = M^B(0)$, and each simulation runs until all molecules have reacted. The initial locations of molecules are uniformly distributed over the interval $(0, L)$. In both DL-FPKMC and the fixed lattice method, if the initial distance between a molecule of $A$ and a molecule of $B$ is less than or equal to $r_\textrm {R}$, then they react immediately. All later reactions occur when the distance between an $A$ molecule and a $B$ molecule equals $r_\textrm{R}$. In the fixed lattice simulations, $h = r_\textrm{R} = 1\text{nm}$. In DL-FPKMC, $h_s^\textrm{max} = L/50 = 200 \text{nm}$, $h_p =
r_\textrm{R} = 1\text{nm}$, and $r_\textrm{pair} = 50 \text{nm}$. To check that using a coarser value for $h_s^\textrm{max}$ than $h_p$ is still sufficient to obtain accurate results, we ran $10^6$ simulations with $M^A(0) = M^B(0) =1$. In this case the exact mean reaction time, $\E_{\textrm{act}}[T]$, is known analytically, see Eq. . Using the preceding parameters, the resulting $\E_{\textrm{emp}}[T]$ from the DL-FPKMC simulations agrees with $\E_{\textrm{act}}[T]$ to within statistical error, which is approximately $0.34\%$.
{#section-12}
![\[fig:RunningTime\] Comparison of DL-FPKMC and fixed lattice methods for the reaction $A + B \to \varnothing $ as the number of molecules present initially, $N = M^A(0) + M^B(0)$, is increased. $M^A(0) = M^B(0)$, $Q_{i}^A(0)$ and $Q_{j}^B(0)
\sim \mathcal{U}(0,L)$ for $ 1 \le i, \, j \le M^A(0)$, $L = 10 \mu \text{m}$, $r_\textrm{R} = 1\text{nm}$, $D = 10 \mu \text{m}^2/ \text{sec}$, and $V(x) = 0$. Each DL-FPKMC data point is based on $10^3$ simulations, and each fixed lattice data point is based on $10^2$ simulations. Error bars indicate $99 \%$ confidence intervals. ](figure13)
Figure \[fig:RunningTime\] compares the results and computational performance of DL-FPKMC to those of the fixed lattice method as the total number of molecules in the system, $N = M^A(0) + M^B(0)$, is varied. As shown in the left panel, the resulting values of $\E_{\textrm{emp}}[T_{M^A(0)}]$, the mean time for all molecules to react within the simulations, from the two methods agree to within statistical error. The running time (center panel) for DL-FPKMC is two to three orders of magnitude faster than the running time for the fixed lattice. DL-FPKMC also requires fewer hops than the fixed lattice by two to three orders of magnitude (right panel). The slope of the line for DL-FPKMC on the log-log plot of running time (center panel) indicates that the method is approximately $O(N)$, i.e., the running time scales linearly with the number of molecules, $N = M^A(0)
+ M^B(0)$. Although the fixed lattice method has better scaling with $N$, it would not appear to become more efficient than DL-FPKMC until $N$ is substantially larger than $10^5$. At that large a number of molecules, it is common to transition to more macroscopic stochastic reaction-diffusion models.
As the number of molecules in the system is increased, the proportion of running time spent on different steps of the DL-FPKMC algorithm changes. For the parameter values used in this section, the changes in proportions of running time as $N$ is increased from $2$ to $2048$ are as follows (approximated using the MATLAB ‘`profile`’ function):
- Generating sample paths: decreases from $67 \%$ to $38 \%$,
- Identifying neighboring molecules and defining new protective domains for updated molecules based on locations of neighbors: increases from $ 17\%$ to $ 52 \%$,
- Determining which molecules (possibly none) to no-passage update after each first-passage update: decreases from $ 4\%$ to $ 3\%$,
- Sorting the event queue: decreases from $7 \%$ to $3 \%$.
These percentages could vary substantially for different parameter values and mesh widths, but we expect the overall trends would hold.
As was described in Subsection \[S:SamplePaths\], the DL-FPKMC algorithm generates a sample path for each molecule within its protective domain until a first-passage or reaction event occurs. Whenever a molecule is no-passage updated before its next event time, part of its sample path goes unused. The mean number of hops per simulation shown in Figure \[fig:RunningTime\] (right panel) is based on the total number of hops in all sample paths that are generated, not just the hops that are actually used. Particularly long paths tend to be generated for protective domains touching an overall domain boundary, since molecules can only exit through one endpoint of such protective domains. If we cap the length of single-molecule protective domains that are near the boundaries, then the total number of hops decreases, but the number of times that the protective domains are updated increases and the overall running time also increases. Some guidelines for and difficulties in optimizing the partitioning of space among protective domains in FPKMC are discussed in [@DonevJCP2010], but we have not yet attempted to address this difficult optimization problem.
{#section-13}
It is expected that the implementation of DL-FPKMC in higher dimensions can be done in such a way as to maintain the $O(N)$ scaling for the following reasons:
- FPKMC in two and three dimensions can be implemented to have $O(N)$ scaling [@DonevJCP2010]. One might expect identifying neighboring molecules to be more costly in higher dimensions,
- The only algorithmic difference between DL-FPKMC and FPKMC is that DL-FPKMC generates sample paths within protective domains using continuous-time random walks, whereas FPKMC samples from analytic solutions of the diffusion equation;
- Since the fraction of running time spent generating sample paths in DL-FPKMC decreases as $N$ increases, the scaling of overall running time as $N$ increases is expected to depend mainly on other steps of the algorithm. These steps can all be implemented in the same way in DL-FPKMC as in FPKMC.
Applications {#S:Applications}
============
Comparison of Potentials {#S:ComparisonPotentials}
------------------------
To demonstrate the contrasting effects that can be produced by different drifts, we consider the reaction $A + B \to \varnothing$ where the molecules diffuse within various potential energy landscapes. We consider the following three cases: (i) zero potential, $V(x)=0$; (ii) a one-well potential, $V(x)= \cos(2 \pi x)$; and (iii) a two-well potential, $V(x)= \cos(4 \pi x)$. We use a domain of length $L = 1$ unit with reflecting boundaries, and diffusion coefficient $D = 1$ unit$^2/$sec for both molecular species $A$ and $B$.
![\[fig:RxnLocDensity\] Reaction locations from $ A + B \to
\varnothing $ DL-FPKMC simulations with $M^A(0) = M^B(0) = 1$, and the Gibbs-Boltzmann distribution for each potential. The reaction location is $( Q^A(t) + Q^B(t) )/2$, where $t$ is the time of the reaction. Each graph of reaction locations is based on $10^7$ simulations. $r_{\textrm{R}}= 0.02$. $h^{\textrm{max}}_s = h_p = r_{\textrm{R}}/8$. $Q^A(0)$ and $Q^B(0) \sim \mathcal{U}(0,1)$. The plotted densities were determined by binning the reaction locations into $100$ bins. ](figure14)
In the absence of reactions, the equilibrium probability density for a molecule to be at location $x {{\color{black}{ \in (0, L) }}}$ is given by the . We compare the Gibbs-Boltzmann distributions for each of the three potentials to the reaction locations from the $A + B \to \varnothing$ DL-FPKMC simulations in the particular case of two molecules, $M^A(0) = M^B(0)
= 1$. The results are shown in Figure \[fig:RxnLocDensity\]. The potentials serve to spatially “confine” molecules, in the sense that molecules are most likely to be found in locations where the potential energy is lowest. Consequently, the reactions are most likely to occur in such low energy locations.
![\[fig:RxnTimeCDFs\] *Left and center panels*: Empirical survival probability $S_{\textrm{emp}}(t)$ for the two-molecule $A + B \to \varnothing$ reaction. Both panels show the same survival probabilities, however the axes have different scales. The dashed lines show $99\%$ confidence bounds based on $10^7$ simulations. $r_{\textrm{R}} = 0.02$. $h^{\textrm{max}}_s = h_p =
r_{\textrm{R}}/8$. *Right panel*: Empirical mean number of molecules of $A$ remaining at time $t$, for the reaction $ A + B \to
\varnothing$ with $M^A(0) = M^B(0) = 10$. The dashed lines show $99\%$ confidence bounds based on $4 \times 10^4$ simulations. $r_{\textrm{R}} = 0.02$. $h_p = r_{\textrm{R}}/16 $. $h^{\textrm{max}}_s = 8 h_p$.](figure15){width="16cm"}
As would be expected, the mean reaction time for the two-molecule $A +
B \to \varnothing$ reaction is faster with the one-well potential ($\sim 0.03481$ sec) than with no potential ($\sim 0.06483$ sec), while slower with the two-well potential ($\sim 0.09887$ sec). Figure \[fig:RxnTimeCDFs\] (left and center panels) compares the survival probabilities $S_{\textrm{\textrm{emp}}}(t)$ for the three different potentials. The semi-log graphs of $S_{\textrm{\textrm{emp}}}(t)$ for all three potentials appear linear except at short times, indicating that the reaction time distributions could be described by exponential distributions for larger times (Fig. \[fig:RxnTimeCDFs\], left panel). For $V=0$, this is as to be expected since the survival probability distribution is known analytically and is given by the eigenfunction expansion in Appendix \[S:AnalytlcAndNumericalPDEsolns\]. At short times (Fig. \[fig:RxnTimeCDFs\], center panel), the graphs of $S_{\textrm{\textrm{emp}}}(t)$ for the different potentials are not linear and behave differently from each other. Initially, reactions occur more quickly with the two-well potential than with either the one-well potential or no potential. However, after $S_{\textrm{emp}}(t)$ has decreased to about $50\%$, reactions occur more slowly with the two-well potential than with the other potentials. As would be expected, if the initial locations of the two molecules are in the same well of a potential, then they tend to react more quickly; whereas, if the two molecules start in different energy wells, then the time until they react tends to be longer.
Figure \[fig:RxnTimeCDFs\] (right panel) shows $\E_{\textrm{emp}}[M^A(t)]$, the mean number of molecules of $A$ remaining at time $t$, when $M^A(0) = M^B(0) = 10$, $Q_{i}^A(0) \sim \mathcal{U}(0.1,0.4)$ and $Q_{j}^B(0) \sim \mathcal{U}(0.6,0.9)$. In this case, $\E_{\textrm{emp}}[M^A(t)]$ could be described by exponential distributions for all three potentials. In the case of $M^A(0) = M^B(0) = 50$ with $Q^A_i(0)$ and $Q^B_j(0) \sim \mathcal{U}(0,1)$, $\E_{\textrm{emp}}[M^A(t)]$ is not plotted but behaves very similarly to $S_{\textrm{emp}}(t)$ in the two-molecule case with $Q^A(0)$ and $Q^B(0) \sim \mathcal{U}(0,1)$ (Fig. \[fig:RxnTimeCDFs\], left and center panels).
Role of Biopolymer Geometry in Protein Diffusive Search {#S:DNAproteinApplication}
-------------------------------------------------------
![ Biopolymer diffusion-excursion search and protein-protein interactions. The geometry of an individual biopolymer or network is expected to be influential in the search kinetics of proteins which diffuse in one dimension along the biopolymers, but may also detach to undergo excursions in three dimensions.[]{data-label="fig_ProteinDNASchematic"}](figure16){width="2.8in"}
Many proteins diffuse effectively in one dimension by adhering loosely to biopolymer filaments, such as actin, microtubules, or DNA [@AlbertsMOLECELLBIO]. In the case of DNA regulatory proteins, single molecule experiments and theoretical work indicate that observed rapid kinetic rates are achieved by a search process involving a combination of one-dimensional diffusion of the protein sliding along the DNA in conjunction with three-dimensional diffusive excursions [@Wang2006; @Riggs1970; @Shimamoto1999; @Gorman2008; @MirnyDNASlide04; @MirnySlideGlob09; @Hu2006; @Berg1981]. The geometry of the individual biopolymer or network of biopolymers is expected to play an influential role in such search kinetics. We introduce a model to investigate this phenomena in the diffusion of proteins that interact with biopolymers having basic knotted closed-loop configurations, see Figure \[fig\_ProteinDNASchematic\]. In our model we also allow for position-dependent forces acting on the proteins, such as might occur from heterogeneities in DNA sequence or protein binding sites. We perform simulations of two proteins undergoing a search process until they encounter each other on a biopolymer. The process involves one-dimensional drift-diffusion along the polymer in combination with absorption/desorption events associated with excursions from the polymer. For example, this process could model the formation of a regulatory complex at a non-specific DNA binding site. Movement along the polymer is simulated using the drift-diffusion DL-FPKMC method introduced in the preceding sections. The effects of the biopolymer geometry are taken into account through the absorption/desorption statistics computed by solving a diffusion equation in three dimensions for each specified biopolymer configuration. We remark that the model and methods we introduce are readily extended to more complex geometries, polymer networks, and filament bundles.
### Model of the Biopolymer Drift-Diffusion Process and Three-Dimensional Excursions {#S:ModelPolymerExcursions}
The biopolymer is represented geometrically as a one-dimensional filament embedded within a three-dimensional space $\Omega$. The heterogeneity along the biopolymer strand, for example changes in DNA sequence, density of bound proteins, or other chemical factors, are taken into account through the potential in the drift-diffusion process. To account for diffusive excursions in three dimensions, we introduce desorption and absorption events for the protein from the biopolymer. Let $\lambda_{\textrm{off}}$ be the first-order rate at which a protein desorbs from the biopolymer. We assume the protein must undergo a conformational change before regaining affinity to bind to the biopolymer (otherwise it would instantly re-absorb at the desorption position). This conformational change is considered as a first-order reaction with the rate $\mu_a$. To model this process, let $p(s_2, t_2 | s_1, t_1)$ denote the probability density for a protein that desorbs at location $s_1$ at time $t_1$ to re-absorb at location $s_2$ at time $t_2$.
This probability density can be calculated by considering the two stages of the protein desorption and re-absorption to the biopolymer. In the first stage, the protein desorbs at location $s_1$ along the biopolymer and performs three-dimensional diffusion until the occurrence of the conformational change at a time $\tau_a$ after desorption. At this time the probability density the protein is at the location ${{\mathbf{X}}}$ is given by $p_s ({{\mathbf{X}}} \, | s_1, \tau_a) =
\frac{1}{(2\pi Dt)^{3/2}} \exp\left\lbrack -\frac{({{\mathbf{X}}} -
{{\mathbf{Z}}}(s_1))^2}{4D\tau_a}\right\rbrack$, where ${{\mathbf{Z}}}(s_1)$ denotes the position on the biopolymer from which the protein desorbed and $D$ the three-dimensional diffusion coefficient of the protein. In the second stage, the protein has affinity for the biopolymer and diffusively re-absorbs to the biopolymer at the location $s_2$ at time $t_2$. We denote this probability density by $p_a(s_2, t_2 | {{\mathbf{X}}},
t_1 + \tau_a)$. This gives the density $p(s_2, t_2 | s_1, t_1) =$ $\int_\Omega \int_0^{t_2-t_1} p_a(s_2,
t_2 | {{\mathbf{x}}}, t_1 + t) \, p_s ({{\mathbf{x}}} | s_1, t) \, \mu_a e^{-\mu_a t}
\, dt \, d{{\mathbf{x}}}$,
For simplicity we shall assume the biopolymer geometric conformation is fixed, resulting in a desorption and absorption process that is stationary. This gives $p_a(s_2, t_2 | s_1, t_1 + \tau_a) = p_a(s_2,
t_2 - t_1 - \tau_a | s_1, 0)$. In practice, we tabulate $p_a(s_2, t
\, | \, {{\mathbf{X}}}, 0)$ at select locations, ${{\mathbf{X}}}$, as a one-time off-line calculation using a numerical diffusion equation solver. This provides a very efficient method to simulate the excursions. To obtain $\tau_a$, we first generate a random exponential time at which the affinity conformational change of the protein occurs. We then sample the protein location, ${{\mathbf{X}}} = {{\mathbf{x}}}$, using the density $p_s({{\mathbf{X}}} \, | s_1, \tau_a)$. From our pre-constructed table, we can then sample $p_a(s_2, t_2 - t_1 - \tau_a \, | \, {{\mathbf{x}}} )$ to find a time of re-absorption, $t_2$, and the protein association location, $s_2$. The protein will eventually be re-absorbed, because the three-dimensional spatial domain has reflecting boundaries and re-absorption to the polymer is modeled using a sink term which has support on a set of positive measure (see Subsection \[S:PolymerSink\]).
The modeled search process of the two proteins proceeds by drift-diffusion along the biopolymer until the proteins encounter each other . The drift-diffusion along the biopolymer is handled by our one-dimensional DL-FPKMC algorithm with periodic boundary conditions. Upon a desorption event from location $s_1$, the protein is repositioned on the biopolymer at location $s_2$ at the re-attachment time $t_2$ according to the density $p(s_2, t_2 | s_1, t_1)$. This repositioning process models the three-dimensional diffusive excursion of the protein until re-absorbing to biopolymer. We remark that reflecting boundary conditions in our DL-FPKMC method could be used to model obstructions on the biopolymer. Either potential or absorbing Dirichlet boundary conditions could be used to model irreversible binding sites on the polymer.
### Numerical Methods for Excursion-Time Probability Distribution and Reattachment Locations {#S:PolymerSink}
We numerically solve the three-dimensional diffusion equation with diffusivity $D$ for the position of a detached molecule at a given time. Denote by $\rho(\vx,t)$ the corresponding probability density that solves the diffusion equation. Reattachment sites on a biopolymer are modeled using sink terms in the equation. We let $s$ label the parameterization variable along the biopolymer, and $s_0$ the detachment position. From the probability mass absorbed at the locations of the sinks, we can obtain the probability densities $p(s,t|s_0,0)$, where $p(s, t | s_0, 0)$ is as defined in the previous subsection. In the diffusion equation we use a sink term of the form $$g(\vy,t) = - \paren{ \int \lambda(s)q(\vy-\vx(s)) \, ds } \rho (\vy,t),$$ where $\lambda$ is the intensity of the sink and $q$ is an averaging kernel function. The rate of the reattachment flux is given by $$\frac{dQ}{dt}(s,t) = \int\frac{\lambda(s)q(\vy-\vx(s))}{\int \lambda(s')q(\vy-\vx(s'))ds'}g(\vy,t)\, d\vy
= -\int\lambda(s)q(\vy-\vx(s))\rho(\vy,t) \, d\vy.$$ In the discretized form this becomes $$g(\vy_m,t) = -\sum_k \lambda_k q(\vy_m-\vx(s_k))\Delta s_k \rho(\vy_m,t)$$ and the rate of the flux becomes $$\frac{dQ_k(t)}{dt} = - \sum_m \lambda_k q(\vy_m-\vx_k)\Delta{s}_k \rho (\vy_m,t) \Delta{\vy}_m.$$
The particular choice that we make for the kernel $q$ is the four-point Peskin-$\delta$ kernel [@Peskin2002]. For $\vx =
(x,y,z)$ this is given by $q(\vx) =
\phi_p(x/\Delta{x})\phi_p(y/\Delta{x})\phi_p(z/\Delta{x}),$ where $$\begin{aligned}
\phi_p(u) = \frac{1}{16}\left\{
\begin{array}{lrl}
3-2u + \sqrt{1+4u-4u^2} & 0 \leq &u \leq 1,\\
5-2u - \sqrt{-7+12u-4u^2} & 1 \leq &u \leq 2,\\
0 & 2 \leq &u,
\end{array} \right. \nonumber
\label{eq_PeskinDelta}\end{aligned}$$ with $\phi_p(-u) = \phi_p(u)$. This particular function is chosen to reduce numerical error induced by the off-lattice shifts of the polymer adsorption locations relative to the underlying discretization lattice used in the three-dimensional diffusion solver. We take the absorption parameter $\lambda$ to have two stages. Namely, a “non-sticky" state $\lambda = 0$ and a “sticky" state $\lambda(s) = \lambda_k =
\lambda_\textrm{absorb}$. The latter is taken to be uniform along the biopolymer. As discussed in Subsection \[S:ModelPolymerExcursions\], the protein becomes sticky after undergoing a conformational change, the time for which is modeled as an exponentially distributed random variable.
![ Reattachment probability distributions for the different topologies. (a–c) Probability density of absorbing at a given location ${{\mathbf{Z}}}(s)$ when the protein desorbed at location ${{\mathbf{Z}}}(0.5)$. $P_{s_1}(s) = \int_0^\infty p(s,t|s_1 = 0.5,0)\, dt$. (d–f) Conditional probability densities for a protein that desorbed at location ${{\mathbf{Z}}}(0.5)$ to absorb at a point $s_2$. For select locations these are indicated in the first row by a colored dot with the same color convention holding for the plots in the second row. $P_{s_1,s_2}(t) = p(s_2, t\, | \, s_1=0.5, 0)$. (g–i) Biopolymer geometries from left to right: a circular loop, trefoil knot, and figure-eight knot.[]{data-label="fig_DNAAttachProb"}](figure17){width="3.5in"}
We consider three distinct geometric configurations for the biopolymer, each with the same arc-length but a different topology. Biopolymers represented by a circular unknotted loop, a trefoil knot, and a figure-eight knot are studied. We use parameterizations given by $X(s) = c*(x(s), y(s), z(s))$ for $0\leq s \leq 1$, where $c$ is chosen so that the arc-length of each of the knots is $L_{\textrm{biopolymer}} = 1000$nm. The $x$, $y$, and $z$ parameterizations used for the three polymer configurations are: unknotted loop, $x(s) = \cos(2\pi s)$, $y(s) = \sin(2\pi s)$, $z(s) =
0$; trefoil knot, $x(s) = (2+\cos(6\pi s))\cos(4 \pi s)$, $y(s) =
(2+\cos(6 \pi s))\sin(4\pi s)$, $z(s) = \sin(6\pi s)$; and figure-eight knot, $x(s) = (2+\cos(4\pi s))\cos(6 \pi s)$, $y(s) =
(2+\cos(4 \pi s))\sin(6\pi s)$, $z(s) = \sin(8\pi s)$. For each configuration we use the diffusion equation solver to tabulate the joint reattachment time and location distribution (see Fig. \[fig\_DNAAttachProb\]).
### Simulation Results: Diffusion-Excursion Search with Different Biopolymer Geometries
3D Parameters Value 1D Parameters Value
--------------------------- ----------------------------------------------- -- -------------------------- -------------------------------------------
3D diffusion coefficient $ 2.183823 \, \mu \text{m}^2 \text{sec}^{-1}$ 1D diffusion coefficient $0.01 \, \mu \text{m}^2 \text{sec}^{-1}$
$ \mu_a$ $5.1728\times 10^7 \mbox{ns}^{-1}$ $\lambda_{\textrm{off}}$ $0.02$ to $200 \, \text{sec}^{-1}$
$\Delta{x}$ $12.5 \mbox{nm}$ $r_{\textrm{R}}$ $20 \,$nm
$\Delta{t}$ $4292.9 \mbox{ns}$ $h_s^{\textrm{max}}$ $10 \,$nm
$N_{\textrm{biopolymer}}$ $100$ $h_p$ $5 \,$nm
$N_x,N_y,N_z$ $40$ $r_{\textrm{pair}}$ $ 4 \, r_{\textrm{R}} = 80 \, \text{nm} $
: \[tab:biopolyParams\] Diffusion-excursion search parameters.
![\[fig:VaryingOffRateV0\] Reaction location densities for $ A + B \to \varnothing $ with $M^A(0) = M^B(0) = 1$ and $V(s)=0$. The reaction location is the midpoint between the locations of $A$ and $B$ at the time of the reaction. Each graph is based on $10^6$ simulations. The plotted densities were determined by binning the reaction locations into $50$ bins. Note that no potentials are imposed in this figure. The non-uniformity in the reaction location density is a result of the polymer geometry. Mean reaction time and mean time that both molecules are on the biopolymer. Each data point is based on $10^6$ simulations. Error bars indicate $99\%$ confidence intervals. ](figure18)
![\[fig:vAcos6px\] Reaction location densities for $ A + B \to \varnothing $, with $M^A(0) = M^B(0) = 1$, $V(s) = \alpha \cos{(6 \pi s)}$, . Each graph is based on $10^6$ simulations. The plotted densities were determined by binning the reaction locations into $50$ bins. Mean reaction times. Each data point is based on $10^6$ simulations. Error bars indicate $99\%$ confidence intervals. ](figure19)
We assume one molecule each of protein species $A$ and $B$ are initially present on the biopolymer. The initial positions are drawn from a uniform distribution over the length of the biopolymer. The two molecules undergo drift-diffusion along the biopolymer and diffusion in the three-dimensional space by the method developed in the preceding subsections. Detachment times are sampled from an exponential distribution with rate $\lambda_{\textrm{off}}$. A reaction between the two molecules can only occur when both are on the biopolymer.
For the parameters we take three-dimensional diffusion coefficient $D = k_B{T}/\gamma$, where $\gamma = 6 \pi \eta a$ is the Stokes drag, , $T = 298.15$ is room temperature, $k_B$ is Boltzmann constant, and the dynamic fluid viscosity, $\eta$, is chosen to be 10 times the viscosity of water. The parameter for the exponential distribution that represents when the protein becomes sticky is denoted by $\mu_a =
6D/ {{\color{black}{ r^2}}}$, . For convenience, the absorption strength, $\lambda_\textrm{absorb} =
0.005 \textrm{ns}^{-1}$, is chosen to be larger than the timescale required for numerical stability in the diffusion equation solver. The other parameters are specified in Table \[tab:biopolyParams\].
A natural question is how diffusion in a potential energy landscape may enhance or counteract the effects of biopolymer geometry on the reaction locations. From Subsection \[S:ComparisonPotentials\], we expect the density of reaction locations to decrease in areas where the potential is large and to increase in areas where the potential is small. To test this idea, we investigated the trefoil knot conformation with potentials of the form $V(s) = \alpha \cos{(6 \pi
s)}$ for several values of $\alpha \in \brac{-6,6}$. The resulting reaction location densities as $\alpha$ is varied are shown in Figure \[fig:vAcos6px\] (left panel). As expected, the potential enhances the effects of the trefoil knot geometry when $\alpha<0$, and counteracts the effects when $\alpha>0$. When $\alpha = 1.5$, the effects of the potential and the trefoil geometry essentially cancel each other out. For comparison, we also ran simulations with $\alpha = 1.5$ for each of the other two biopolymer geometries (Fig. \[fig:vAcos6px\], center panel).
When the potential $V(s) = \alpha \cos{(6 \pi s)}$ is used, the mean reaction time is fastest when the amplitude $| \alpha |$ is large (see Fig. \[fig:vAcos6px\], right panel). As discussed in Subsection \[S:ComparisonPotentials\], molecules that are in the same energy well of a potential tend to react more quickly. However, if the proteins can only move on the biopolymer, potentials with large amplitudes as used here would represent large energy barriers, greatly slowing the time for the two molecules to find each other. For example, with $\alpha = -3$ and $\lambda_{\textrm{off}} = 0$, the mean reaction time is approximately $80 \pm 9$ seconds, based on $10^3$ simulations (compare to Fig. \[fig:vAcos6px\], right panel). The three-dimensional diffusion excursion provides an alternative path for proteins to circumvent the energy barriers present on the biopolymer. This could be an important mechanism in protein-protein interactions associated with biopolymers. As shown in Figure \[fig:vAcos6px\] (right panel), for the circle and figure-eight conformations, the maximum mean reaction time occurs when $\alpha = 0$, and the mean reaction times with positive and negative $\alpha$ of the same magnitude are similar. In contrast, for the trefoil conformation, the maximum mean reaction time occurs when $\alpha = 1.5$, which is the same value of $\alpha$ that resulted in the most uniform distribution of reaction locations (as was shown in Fig. \[fig:vAcos6px\], left panel). This demonstrates that the effect of the potential is not independent of the polymer geometry.
Overall, these results demonstrate the importance of incorporating drift-diffusion and other spatial heterogeneities, such as biopolymer geometry, when investigating biological chemical processes. They illustrate the power of the presented DL-FPKMC method in capturing such effects. Our approach and methods can be readily extended to more complex chemical kinetics, biopolymer geometries, or filament networks.
Conclusion
==========
We have presented a new Dynamic Lattice First-Passage Kinetic Monte Carlo method (DL-FPKMC) that is capable of incorporating the roles played by drift and spatial heterogeneities into the stochastic dynamics of chemically reacting molecular species. We have shown that our numerical method is convergent for smooth potentials with approximately second-order accuracy and for discontinuous potentials with approximately first-order accuracy. Unlike the standard lattice RDME, our method retains bimolecular reactions as the lattice spacings within protective domains are taken to zero, and converges to the underlying SDLR model. In higher dimensions we expect the use of Walk on Rectangles [@LejayExactWOS2006] techniques will facilitate the incorporation of complex geometries into our method, while still allowing the use of basic Cartesian-grid meshes within each protective domain (hence avoiding the need for unstructured or embedded boundary meshes as commonly used for RDME models.)
We have further demonstrated how our method can be utilized in practice. In particular, the many examples we investigated demonstrate that drift can have a significant effect on reaction locations, reaction time distributions, and number of reactions that occur in a system. To demonstrate how our approach might be used for more complicated systems, we investigated the process of diffusion-excursion of proteins interacting with a biopolymer, taking into account geometric effects due to polymer shape. Such processes are thought to occur in the interactions between regulatory proteins and DNA. We considered a simplified model where the biopolymer has basic knotted configurations. Our results demonstrate that the combined effects of the biopolymer-protein interactions, biopolymer geometry, and drift-diffusion play a significant role in the chemical kinetics. The results also illustrate that capturing important features of biological systems may require models that account for drift-diffusion effects and spatial heterogeneities. We expect the DL-FPKMC methods we have introduced to be useful in performing more realistic simulations of the chemical kinetics of biological systems.
Acknowledgments
===============
A.J.M. was supported by NSF DMS-0602204 EMSW21-RTG and Boston University. S.A.I. and A.J.M. were supported by NSF DMS-0920886 and NSF DMS-1255408. P.J.A. and J.K.S. were supported by NSF CAREER - 0956210. J.S. was supported by the CCS Summer Undergraduate Research Fellowship (SURF) program at UCSB and from P.J.A.’s startup funds (UCSB Math).
{#S:ProtectiveDomainsAppendix}
This appendix describes our method for constructing protective domains (PDs). It should be noted that this is just one of many possible approaches, and it is still an open question how best to optimize the partitioning of space among PDs. For the reader interested in this question we refer and defer to [@DonevJCP2010]. Regardless of the method used for constructing the PDs, our lattice approach of Section \[S:oneDimMethod\] can be used to propagate molecules within their respective PDs.
The PDs should be defined in a such a way that the distance between any two bimolecular reactants in separate PDs will remain strictly greater than their reaction radius, $r_\textrm{R}$, for as long as they remain in their respective PDs. This is necessary to ensure that the movements of molecules in separate PDs are independent. We allow the PDs of non-reacting molecules to overlap.
The following is the approach for defining PDs that resulted in the scaling demonstrated in Section \[S:RunningTime\]. Steps 3 through 5 are specific to one dimension, but could be extended to higher dimensions by following a similar procedure in each coordinate. When constructing PDs for all molecules initially, or when updating the PDs of more than one molecule, we begin with the molecule closest to the left endpoint of the overall domain and then proceed to the right. An alternative approach would be to begin with a seed molecule and proceed outward.
1. For each molecule in need of a protective domain, identify the nearest potential reaction partners. In one dimension, this can by done by keeping a list of all molecules ordered by location. In higher dimensions, the near-neighbor list (NNL) method described in [@DonevJCP2010] can be used.
2. Determine which molecules will be placed in pair PDs. Two potential reaction partners are placed in a pair if:
- the two molecules are closer to each other than to any other potential reaction partners, and
- the distance between the two molecules is less than a pair threshold, $r_\textrm{pair}$, which is a parameter chosen by the user.
For the reaction system $ A + B \to \varnothing $, we also enforce the condition in Eq. when determining if two molecules will be placed in a pair. No other PDs are allowed to overlap with a pair PD.
3. For each molecule, identify it’s “limiting neighbor," which we define in the following way:
- For each molecule that will be in a pair PD, the limiting neighbor is the next nearest molecule of any type outside the pair.
- For a molecule that will be in a single PD, the limiting neighbor is the nearest molecule that is either a potential reaction partner or in a pair with another molecule.
- Let $d_\textrm{nbr}$ be the distance from a molecule to it’s limiting neighbor. If a PD for the limiting neighbor has already been defined, let $d_\textrm{nbrPD}$ be the distance from the molecule in question to the nearest endpoint of it’s limiting neighbor’s PD.
- Note, a molecule is not necessarily the limiting neighbor of it’s limiting neighbor.
4. Determine the size of each pair PD and then the size of each single PD as follows, ignoring overall domain boundaries until Step 5:
- Pair PDs will be symmetric about the midpoint of the two molecules’ locations. Single PDs will be symmetric about the location of the molecule.
- Let $r_\textrm{PD}$ for a pair PD be the distance from either molecule to the nearest endpoint of the PD. For a single PD, $r_\textrm{PD}$ will denote the distance from the molecule to either endpoint of the PD.
- Define Condition 1 to be that the limiting neighbor is a potential reaction partner, and Condition 2 to be that a PD for the limiting neighbor has not yet been defined.
For a single PD, calculate $r_\textrm{PD}$ by
$$\label{rPD}
r_\textrm{PD} = \left\{
\begin{array}{cl}
(d_\textrm{nbr} - r_\textrm{R}) / 2
& \text{ if Conditions 1 and 2 hold} \\
d_\textrm{nbrPD} - r_\textrm{R} & \text{ if only Condition 1 holds} \\
d_\textrm{nbr}/ 2
& \text{ if only Condition 2 holds} \\
d_\textrm{nbrPD} & \text{ otherwise.} \end{array} \right.$$
For a pair PD, calculate the quantity in Eq. for each molecule and then set $r_\textrm{PD}$ for the pair to be the minimum of the two quantities.
- We recommend capping the size of pair PDs, so that two molecules will not remain in a pair if they have moved sufficiently far away from each other. For example, $r_\textrm{PD}$ for a pair PD could be set to the minimum of the value calculated above and half the initial distance between the two molecules in the pair.[^7]
5. If a PD as defined in Step 4 extends beyond an endpoint of the overall domain, truncate the PD so that one endpoint of the PD will coincide with the overall domain endpoint. Such PDs will no longer be symmetric. An alternative approach would be to treat the overall domain boundary as a neighbor in Steps 1 through 4, and only allow a molecule’s PD to touch the domain boundary if the distance to the boundary is less than $r_\textrm{pair}$.
Derivation of Non-Uniform Jump Rates {#S:DerivationNonUnifJumpRates}
====================================
To derive the non-uniform rates in Eq. , we use the fluxes from the WPE discretization [@ElstonPeskinJTB2003]. As illustrated in Figure \[fig:NonUnifSubLattice\] (top row, left), $x_1 < x_0 <
x_2$ are the locations of mesh points with non-uniform spacing $h_j
= |x_0 - x_j|$. The jump rates $a_{0j}$ from $x_0$ to $x_j$, $j=1,2$, are derived in this appendix. The solution $\rho(x,t)$ of the Fokker-Planck equation , gives the probability density of being at location $x$ at time $t$. Let $\rho^{\textrm{eq}}(x)$ denote the equilibrium value of $\rho(x,t)$. Define $p_i(t)$ to be the probability of being at the mesh point $x_i$ at time $t$ in the discrete master equation model. We consider the point $x_1$ to represent the interval $(x_1-\frac{h_1}{2}, x_1+\frac{h_1}{2})$ in the sense that $$\label{eq:p1}
p_1(t) \approx \int_{ x_1-\frac{h_1}{2} }^{ x_1+\frac{h_1}{2} }
\rho(x,t) \, dx \approx \rho(x_1, t) h_1.$$ Similarly, $x_0$ represents $(x_0-\frac{h_1}{2}, x_0+\frac{h_2}{2})$ and $x_2$ represents $(x_2-\frac{h_2}{2}, x_2+\frac{h_2}{2})$, so $$\label{eq:p0andp2}
p_0(t) \approx \rho(x_0, t) \frac{h_1+h_2}{2} \qquad \text{and} \qquad
p_2(t) \approx \rho(x_2, t) h_2.$$
Let $J(x,t)$ denote the flux $$J(x,t) = - D \left(\rho(x,t) \frac{\partial V(x)}{\partial x} +
\frac{\partial \rho(x,t) }{\partial x} \right).$$ For convenience, we define
$$A_{ik}(t) = \left\{
\begin{array}{cl}
\frac{V(x_k) - V(x_i)} {\exp[V(x_k) - V(x_i) ] -1}
& \text{for } V(x_k)\ne V(x_i) \\
1 & \text{otherwise. } \end{array} \right.$$
Based on the WPE discretization, we approximate the unidirectional outward flux from $x_0$ to $x_j$ by $$J_{0j}(t) = \frac{D }{h_j} A_{0j}(t) \rho(x_0,t),$$ and the unidirectional inward flux from $x_j$ to $x_0$ by $$J_{j0}(t) = \frac{D }{h_j} A_{j0}(t) \rho(x_j,t).$$ Then, the net flux from $x_0$ to $x_j$ is $J_{0j}(t) - J_{j0}(t)$. In the case that $x_j$, for $j=1$ or $2$, lies on an absorbing Dirichlet boundary, $\rho(x_j, t) = 0$, and so we would have $J_{j0} = 0$ throughout the following calculation. The Fokker-Planck PDE at $x_0$ is approximated by $$\begin{aligned}
\frac{\partial \rho(x_0,t)}{\partial t}
&= - \frac{\partial }{\partial x} J(x_0,t)
\approx \frac{2}{h_1+h_2} \int_{x_0-\frac{h_1}{2} }^{x_0+\frac{h_2}{2} }
- \frac{\partial }{\partial x} J(x,t) \, dx \notag \\
&= \frac{2}{h_1+h_2} \left( - J(x_0+\frac{h_2}{2},t) +J(x_0-\frac{h_1}{2},t) \right) \label{eq:fluxDiff} \\
& \approx \frac{2}{h_1+h_2} \left( (J_{20} - J_{02}) - (J_{01} - J_{10}) \right) \notag \\
& = \frac{2}{h_1+h_2}
\left( \frac{D}{h_2} (A_{20} \, \rho(x_2,t) - A_{02} \, \rho(x_0,t)) -
\frac{D}{h_1} (A_{01} \, \rho(x_0,t) - A_{10} \, \rho(x_1,t)) \right). \label{eq:WPEDensity}
\end{aligned}$$ When $x_0$, $x_1$, and $x_2$ are interior points, Taylor series expansion shows the spatial truncation error to be $O(h_1-h_2)$ at $x_0$. For $h_1=h_2=h$ the discretization is second-order accurate and we recover the WPE discretization. Note, as described in Section \[S:oneDimMethod\], we only use a non-uniform mesh for mesh cells bordering a boundary, or to move molecules onto a uniform mesh in a newly formed pair protective domain. When one of $x_1$ or $x_2$ corresponds to a (non-uniform) Dirichlet boundary point, and all interior mesh cells are uniform, the spatial discretization in Eq. in the case that $V(x) =0$ is known to be second-order accurate for the Poisson equation, see Ref. [@FedkiwJCP2002]. For a reflecting boundary at $x_0-\frac{h_1}{2}$ or $x_0+\frac{h_2}{2}$, the corresponding flux term in Eq. is zero (since we assume reflecting boundaries are at the edges of mesh cells). The corresponding terms involving $A_{0 j}$ and $A_{j 0}$ then drop out of Eq. .
Multiplying through by $\frac{h_1+h_2}{2}$ and making the substitutions in Eqs. and yields the master equation $$\frac{d \, p_0(t)}{d t} =
\left(\frac{D}{h_2^2} A_{20} \, p_2(t) - \frac{2D}{h_2(h_1+h_2)} A_{02} \, p_0(t) \right) -
\left( \frac{2D}{h_1(h_1+h_2)} A_{01} \, p_0(t) - \frac{D}{h_1^2} A_{10} \, p_1(t) \right).$$ Thus, we obtain the non-uniform jump rates $$a_{0j} = \frac{2 D}{h_j (h_1+h_2) } A_{0j}
\qquad j = 1,2.$$
Whenever $x_j$ is *not* a boundary point, the jump rate $a_{j0}$ in the opposite direction agrees with the uniform rate in Eq. with $h=h_j$. Hence, the system still satisfies a discrete detailed balanced condition at equilibrium (similar to the uniform WPE discretization): $$\begin{aligned}
a_{0j}\, \rho^{\textrm{eq}}(x_0) \frac{h_1+h_2}{2} &= a_{j0} \rho^{\textrm{eq}}(x_j) h_j ,\end{aligned}$$ since $\rho^{\textrm{eq}}(x)$ is given by the Gibbs-Boltzmann distribution $ \rho^{\textrm{eq}}(x) \propto \exp[-V(x)] $ and $$\frac{\rho^{\textrm{eq}}(x_j)}{\rho^{\textrm{eq}}(x_0)} = \frac{\exp[-V(x_j)] }{\exp[-V(x_0)] } = \frac{1}{ \exp [V(x_j) -V(x_0)] } .$$
Analytic and Numerical Solutions for the Two-Molecule Annihilation Reaction $A + B \to \varnothing$ {#S:AnalytlcAndNumericalPDEsolns}
===================================================================================================
In this Appendix we discuss the numerical solution of the Fokker-Planck equation and the analytic solution of the diffusion equation to which we compared the DL-FPKMC simulation results in Section \[S:2ParticleConvResults\]. The 1D reaction-drift-diffusion system of two molecules undergoing the reaction $ A + B \to
\varnothing$ can be described by Eq. on the 2D domain in the left panel of Figure \[fig:PDEdomains\]. The solutions of the Fokker-Planck equation on the two disjoint triangular components of this domain are independent of each other. By symmetry, solving the Fokker-Planck equation on the domain in the left panel of Figure \[fig:PDEdomains\] can be reduced to solving the same equation on the single triangular domain in the center panel. We have written a PDE solver to solve the Fokker-Planck equation on the triangular domain. The PDE solver uses the rates in Eq. from the WPE discretization [@ElstonPeskinJTB2003] of the Fokker-Planck equation. This discretization is second-order accurate for smooth potentials and first-order accurate for discontinuous potentials. The mesh in the PDE solver is uniform.
![\[fig:PDEdomains\] *Left*: 2D domain equivalent to 1D simulation domain of length $L$ in which the two molecules are located. The zero Dirichlet boundaries on the diagonals correspond to the reaction occurring between the two molecules when they are one reaction radius apart. The zero-flux boundaries on the outer edges correspond to the reflecting endpoints of the 1D simulation domain. *Center*: PDE solver domain. *Right*: Eigenfunction expansion domain. ](figure20)
The solution to the diffusion equation on the triangular domain in Figure \[fig:PDEdomains\] (center) can be recovered by solving on a square domain formed by adjoining four copies of the triangular domain at their reflecting (zero Neumann) edges, as shown in the right panel of Figure \[fig:PDEdomains\]. On the square domain, an eigenfunction expansion for the solution can be determined analytically. This provides a check for the DL-FPKMC simulations in the $V=0$ case, as well as a check for the PDE solver in that case. The solution $\rho(x,y,t) $ to the diffusion equation with constant initial condition $\rho_0$ and diffusion coefficient $D$ on a square domain with sides of length $l$ is given by [@HandbookPDE2002]: $$\begin{aligned}
\rho(x,y,t) = \frac{16 \rho_0}{\pi^2} & \left[ \sum_{n=0}^\infty \frac{1}{2n+1}
\sin \left( \frac{(2n+1) \pi x}{l} \right)
\exp \left( \frac{-\pi^2 (2n+1)^2 D t}{l^2} \right) \right]
\nonumber \\
\times & \left[ \sum_{m=0}^\infty \frac{1}{2m+1}
\sin \left( \frac{(2m+1) \pi y}{l} \right)
\exp \left( \frac{-\pi^2 (2m+1)^2 D t}{l^2} \right) \right] .\end{aligned}$$ The survival probability $S(t)$, the probability that the two particles have not yet reacted by time $t$, is $$\begin{aligned}
\label{eq:AnalyticSurvProb}
S(t) &= \int_0^l \int_0^l \rho(x,y,t) \, dx \, dy
= \frac{64 \, \rho_0 \, l^2}{\pi^4} \left[ \sum_{n=0}^\infty
\left( \frac{1}{2n+1} \right)^2
\exp \left( \frac{-\pi^2 (2n+1)^2 D t}{l^2} \right) \right] ^2 .\end{aligned}$$ The mean reaction time is $$\begin{aligned}
\label{eq:AnalyticMeanRxnTime}
\avg{T} = -\int_0^\infty t \, S'(t) \, dt &= \int_0^\infty S(t) \, dt
= \frac{64 \,\rho_0 \, l^4}{D \, \pi^6} \sum_{n=0}^\infty
\sum_{m=0}^\infty
\left( \frac{1}{2n+1} \right)^2
\left( \frac{1}{2m+1} \right)^2
\frac{1}{(2n+1)^2 + (2m+1)^2} \notag \\
&{{\color{black}{
= \frac{16 \,\rho_0 \, l^4}{D \, \pi^5} \sum_{n=0}^\infty
\left( \frac{1}{2n+1} \right)^4
\left( \frac{\pi}{2} - \frac{\tanh(\frac{\pi}{2} (2n+1))}{(2n+1)} \right) }}}\end{aligned}$$ Evaluating $\avg{T}$ at the parameter values corresponding to the two-molecule DL-FPKMC simulations gives a mean reaction time of 0.064831881311 seconds.
For each of the potential functions, we ran the PDE solver using the Crank–Nicolson method in time, with spatial step sizes ranging from $\Delta
x=r_{\textrm{R}} = 0.02$ to $\Delta x=r_{\textrm{R}}/16$ and time steps $\Delta t= \Delta x /16$. For $V=0$, we can check the results of the PDE solver against the analytic solution; in this case, the numerical mean reaction times determined using the Crank–Nicolson method converge at approximately second-order ($m \approx 2.0025$) to the analytic mean reaction time. For $V \ne 0$, there is no analytic solution to which the numerical solutions can be compared; however, the decrease in the pairwise differences between the mean reaction times determined from the numerical PDE solutions at successive step sizes indicates convergence ($m \approx 1.9968$ for $V_{\textrm{cos}}$ and $m \approx 0.9844$ for $V_{\textrm{step}}$). In the DL-FPKMC convergence results in Section \[S:2ParticleConvResults\], the empirical mean reaction times are compared to the analytic mean reaction time in the case $V=0$, and to the numerical mean reaction times determined using the Crank–Nicolson PDE solver with the finest spatial step size, $\Delta x=r_{\textrm{R}}/16$, in the cases of $V_{\textrm{cos}}$ and $V_{\textrm{step}}$.
The survival probabilities calculated using the Crank–Nicolson method show small numerical oscillations during the first few time steps. This is due to the incompatibility of the initial condition with the zero Dirichlet boundary condition. In order to numerically resolve the survival probabilities more accurately at short times, we re-ran the PDE solver using the Twizell–Gumel–Arigu (TGA) method [@TwizellL0Meth] from $t=0$ to $t=0.07$ seconds. The TGA method is a second order, $L_0$ stable time discretization. We also used a finer time step, $\Delta t = (\Delta x)^2$ where $\Delta x=r_{\textrm{R}}/16$, to further improve the accuracy at short times when the survival probabilities change most rapidly. We then determined the numerical survival probabilities using the results from the TGA method for $t=0$ to $t=0.07$ seconds and using the results from Crank–Nicolson method for $t>0.07$ seconds. At $t=0.07$, the absolute difference in survival probabilities between the two methods is less than $10^{-8}$ for $V_{\textrm{cos}}$ and less than $10^{-7}$ for $V_{\textrm{step}}$.
To check that using the TGA method with a finer time step improved the accuracy of the numerical survival probabilities over the Crank–Nicolson method, we compared the numerical survival probabilities for $V=0$ to the analytic survival probability on the interval from $t=0$ to $t=0.07$. Table \[tab:CN\_TGA\] shows the absolute errors between the numerical survival probability calculated using each method and the analytic survival probability.
------------------ ----------------------------------------------------------- ----------------------------------------------------------
Crank–Nicolson TGA
$\Delta x = r_{\textrm{R}}/1$6 , $\Delta t = \Delta x/16$ $\Delta x=r_{\textrm{R}}/16$ , $\Delta t = (\Delta x)^2$
$L^1$ Error 4.1117e-6 4.7909e-8
$L^2$ Error 8.2513e-5 1.6171e-6
$L^\infty$ Error 5.8634e-3 2.2463e-4
------------------ ----------------------------------------------------------- ----------------------------------------------------------
: \[tab:CN\_TGA\] Improved errors in numerical survival probabilities.
[^1]: Department of Mathematics and Statistics, Boston University, 111 Cummington St., Boston, MA 02215 ([email protected])
[^2]: Department of Mathematics, University of California, Santa Barbara
[^3]: Department of Mathematics, University of California, Santa Barbara
[^4]: 6712 South Hall, Department of Mathematics, University of California, Santa Barbara, CA 93106 ([email protected])
[^5]: Department of Mathematics and Statistics, Boston University, 111 Cummington St., Boston, MA 02215 ([email protected])
[^6]: Since $V(x)=0$, in the fixed lattice method each possible spatial hop is equally likely. This allows optimization of the underlying Stochastic Simulation Algorithm [@GillespieJPCHEM1977; @KalosKMC75]. Incorporating drift into the fixed lattice method would give spatially-varying hopping rates, so additional computational cost might be required to sample which molecule hops and in which direction.
[^7]:
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate the form of the Coulomb potential of a point charge in a noncommutative geometry, using a state of minimal dispersion. We find the deviation of the potential at large distances from the point, distinguishing between coordinate distance and measured distance. Defining the “effective” value of an operator as its expectation value in a minimum dispersion state centered at a point, we find the effective potential to be finite at the origin, the effective charge density to be Gaussian and the effective total electrostatic energy to be finite. However, the true total electrostatic energy operator is shown to still be infinite.'
address: |
Dept. of Physics\
University of Illinois at Chicago\
Chicago, Illinois 60607\
[email protected]
author:
- 'A. Lewis Licht'
title: '**Coulomb Potential of a Point Mass in Theta Noncommutative Geometry**'
---
Introduction {#S:intro}
============
We consider the form of noncommutative geometry where the operators that measure position, the “physical” position operators, satisfy $$\label{E:noncomtheta}
\left[ {x^i ,\;x^j } \right] = \theta ^{ij}$$
Where $\theta^{i,j}$ is a constant antisymmetric tensor, and where the indexes i, j = 1, 2, 3. The components $\theta^{0,i}$ are assumed to vanish. The physical momentum operators are assumed to have the conventional commutation relations [@douglas] $$\label{E:comconven}
\begin{array}{c}
\left[ {p^i ,\;p^j } \right] = 0 \\
\left[ {x^i ,\;p^j } \right] = i\delta ^{ij} \\
\end{array}$$
We investigate for such a geometry the problem of measuring distance from a point. We also investigate the operational definition of the value at a point of such quantities as the Coulomb potential, the charge distribution and the electric field.
Let $\left| {\Psi _0 } \right\rangle $ denote a point-like state at the origin. This could be a wave packet spherically symmetric about the origin in commutative theory, or as spherically symmetric a state as possible in NC theory. The spherical symmetry implies $$\label{E:symm}
\left\langle {\Psi _0 \left| {{\bf{x}}\left| {\Psi _0 } \right.} \right.} \right\rangle = 0$$ We translate the state $\left| {\Psi _0 } \right\rangle $ to a point a distance ${\bf{a}}$ from the origin, obtaining the state $$\left| {\Psi _0 \left( {\mathbf{a}} \right)} \right\rangle = \exp \left( { - i{\mathbf{a}} \cdot {\mathbf{p}}} \right)\left| {\Psi _0 } \right\rangle$$
The vector ${\bf{a}}$ is a coordinate displacement. Its length $a$ is not actually the displacement distance that would be measured. The measured distance is $$\begin{gathered}
\left\langle r \right\rangle = \left\langle {\Psi _0 \left( {\mathbf{a}} \right)} \right|\sqrt {{\mathbf{x}} \cdot {\mathbf{x}}} \left| {\left| {\Psi _0 \left( {\mathbf{a}} \right)} \right.} \right\rangle \\
= \left\langle {\Psi _0 } \right|\sqrt {{\mathbf{x}} \cdot {\mathbf{x}} + 2{\mathbf{x}} \cdot {\mathbf{a}} + a^2 } \left| {\Psi _0 } \right\rangle \\
= a + \left\langle {\Psi _0 \left| {{\mathbf{x}} \cdot {\mathbf{\hat n}} + \frac{1}
{{2a}}\left( {{\mathbf{x}}^2 - \left( {{\mathbf{x}} \cdot {\mathbf{\hat n}}} \right)^2 } \right)} \right|\Psi _0 } \right\rangle + O\left( {\frac{1}
{{a^2 }}} \right) \\
\end{gathered}$$
Where ${\mathbf{\hat n}}$ is the unit vector along ${\mathbf{a}}$. Using Eq. (\[E:symm\]), this becomes $$\label{E:dist}
\left\langle r \right\rangle = a + \frac{1}
{{2a}}\left\langle {\Psi _0 \left| {\left( {{\mathbf{x}}^2 - \left( {{\mathbf{x}} \cdot {\mathbf{\hat n}}} \right)^2 } \right)} \right|\Psi _0 } \right\rangle + O\left( {\frac{1}
{{a^2 }}} \right)$$ The Coulomb potential that would be measured at the origin due to a point particle at the coordinate position ${\bf{a}}$ is
$$\label{E:coulpota}
\begin{gathered}
\left\langle {\frac{1}
{r}} \right\rangle = \left\langle {\Psi _0 \left( {\mathbf{a}} \right)} \right|\frac{1}
{{\sqrt {{\mathbf{x}}^2 } }}\left| {\Psi _0 \left( {\mathbf{a}} \right)} \right\rangle \\
= \left\langle {\Psi _0 } \right|\frac{1}
{{\sqrt {{\mathbf{x}}^2 + 2{\mathbf{x}} \cdot {\mathbf{a}} + a^2 } }}\left| {\Psi _0 } \right\rangle \\
= \frac{1}
{a} - \left\langle {\Psi _0 \left| {\frac{1}
{{2a^3 }}\left( {{\mathbf{x}}^2 - 3\left( {{\mathbf{x}} \cdot {\mathbf{\hat n}}} \right)^2 } \right)} \right|\Psi _0 } \right\rangle + O\left( {\frac{1}
{{a^4 }}} \right) \\
\end{gathered}$$
This is however expressed in terms of coordinate distance ${\bf{a}}$. Eq. (\[E:dist\]) may be used to express it in terms of the observed distance $\left\langle r \right\rangle $ $$\label{E:obsvcoupot}
\left\langle {\frac{1}
{r}} \right\rangle = = \frac{1}
{{\left\langle r \right\rangle }} + \frac{1}
{{\left\langle r \right\rangle ^3 }}\left\langle {\Psi _0 \left| {\left( {{\mathbf{x}} \cdot {\mathbf{\hat n}}} \right)^2 } \right|\Psi _0 } \right\rangle + O\left( {\frac{1}
{{\left\langle r \right\rangle ^4 }}} \right)$$
For any operator A, the expectation value in the translated state $$\begin{gathered}
\left\langle {\Psi _0 \left( {\mathbf{a}} \right)} \right|A\left| {\Psi _0 \left( {\mathbf{a}} \right)} \right\rangle = \left\langle {\Psi _0 } \right|\exp \left( { + i{\mathbf{a}} \cdot {\mathbf{p}}} \right)A\exp \left( { - i{\mathbf{a}} \cdot {\mathbf{p}}} \right)\left| {\Psi _0 } \right\rangle \\
= \left\langle {\Psi _0 } \right|A\left( {\mathbf{a}} \right)\left| {\Psi _0 } \right\rangle \\
\end{gathered}$$ can also be interpreted as measuring the operator A at the “coordinate point” ${\bf{a}}$ in the compact state $\Psi _0$. We interprete this as the “effective” value of the operator A at ${\bf{a}}$ in the state $\Psi _0$.
We examine measured distance and Coulomb potential for a commutative geometry in Section (2) and for $\theta$ NC geometry in Section (3). In Section (4) it is shown that in $\theta$ NC geometry the effective Coulomb potential, as seen from a compact state, is finite at the origin. The effective charge density is derived in Section (5), and the effective elective field in Section (6). The total “effective” electrostatic energy is calculated and found to be finite. However, in Section (7) we consider the true electric field operator and show that the integral over all space of the corresponding electrostatic energy density can be defined in $\theta$ NC geometry and is infinite.
Commutative Geometry {#S:comm}
====================
We consider here an ordinary free particle in a Gaussian wasve packet of coordinate dispersion $\sigma $. The wave function in x space is $$\Psi \left( {\mathbf{x}} \right) = \frac{1}
{{\left( {2\pi \sigma } \right)^{\tfrac{3}
{4}} }}\exp \left( { - \frac{{{\mathbf{x}}^2 }}
{{4\sigma }}} \right)$$
For this particle, $$\begin{gathered}
\left\langle {{\mathbf{x}}^2 } \right\rangle = 3\sigma \\
\left\langle {\left( {{\mathbf{x}} \cdot {\mathbf{\hat n}}} \right)^2 } \right\rangle = \sigma \\
\end{gathered}$$ Then the measured distance is $$\left\langle r \right\rangle = a + \frac{\sigma }
{a} + O\left( {\frac{1}
{{a^2 }}} \right)$$ and the Coulomb potential is $$\left\langle {\frac{1}
{r}} \right\rangle = \frac{1}
{{\left\langle r \right\rangle }} + \frac{\sigma }
{{\left\langle r \right\rangle ^3 }} + O\left( {\frac{1}
{{\left\langle r \right\rangle ^4 }}} \right)$$
Of course, for this system $\sigma$ can be made arbitrarily small, and the Coulomb potential arbitrarily close to 1/r .
Theta Noncommutativity {#S:thetnc}
======================
The operators $$\xi ^i = x^i + \frac{{\theta ^{ik} }}
{2}p^k$$ have the nice property that $$\begin{gathered}
\left[ {\xi ^i ,\;\xi ^j } \right] = 0 \\
\left[ {\xi ^i ,\;p^j } \right] = i\delta ^{ij} \\
\end{gathered}$$ The $\xi ^i $ can be considered as hypothetical coordinate operators, where the $x^i $ are the operators that correspond to actual physical measurements of position.
The anti-symmetry of the $\theta ^{ij} $matrix allows us to introduce a vector ${\mathbf{\theta}}$ such that $$\begin{gathered}
\theta ^{ij} = \varepsilon _{ijk} \theta ^k \\
\theta ^k = \frac{1}
{2}\varepsilon _{ijk} \theta ^{ij} \\
\end{gathered}$$ Then
$${\mathbf{x}} = {\mathbf{\xi}} + \tfrac{1}
{2}{\mathbf{\theta}} \times {\mathbf{p}}$$
and the distance squared operator becomes
$${\mathbf{x}}^2 = {\mathbf{\xi}} ^2 - {\mathbf{\theta}} \cdot
{\mathbf{\xi}} \times {\mathbf{p}} + \frac{{{\mathbf{\theta}} ^2 }}
{4}\left( {{\mathbf{p}}^2 - \left( {{\mathbf{p}} \cdot {\mathbf{\theta}} } \right)^2 } \right)$$
We find this operator’s spectrum by introducing creation and annihilation operators:
$$\begin{gathered}
u_\alpha = \frac{1}
{{\sqrt \theta }}\xi _\alpha + i\frac{{\sqrt \theta }}
{2}p_\alpha \\
\zeta = \xi _3 \\
\end{gathered}$$
where $\alpha = 1,2$. Then
$$\label{E:xsquar}
{\mathbf{x}}^2 = \zeta ^2 + i\theta \left( {u_1^\dag u_2 - u_2^\dag u_1 } \right) + \theta \left( {u_1^\dag u_1 + u_2^\dag u_2 + 1} \right)$$
With the substitution
$$\left( {\begin{array}{*{20}c}
{u_1 } \\
{u_2 } \\
\end{array} } \right) = \frac{1}
{{\sqrt 2 }}\left( {\begin{array}{*{20}c}
1 & 1 \\
i & { - i} \\
\end{array} } \right)\left( {\begin{array}{*{20}c}
a \\
b \\
\end{array} } \right)$$
Where a and b are annihilation operators, Eq. (\[E:xsquar\]) becomes
$${\mathbf{x}}^2 = \zeta ^2 + \theta \left( {2b^\dag b + 1} \right)$$
The angular momentum operator ${\mathbf{\theta}} \cdot
{\mathbf{\xi}} \times {\mathbf{p}}$ corresponding to rotations about the 3-axis is now $${\mathbf{\theta}} \cdot
{\mathbf{\xi}} \times {\mathbf{p}} = a^\dag a - b^\dag b$$
The general basis state can be written as
$$\left| {\Psi _{nm} \left( \zeta \right)} \right\rangle = \psi \left( \zeta \right)\left| {nm} \right\rangle$$
where $\psi \left( \zeta \right)$ is an arbitrary wavefunction, and $\left| {nm} \right\rangle$ is a harmonic oscillator basis state, with
$$\begin{gathered}
\left| {nm} \right\rangle = \frac{{a^{\dag n} b^{\dag m} }}
{{\sqrt {n!m!} }}\left| {00} \right\rangle \\
a\left| {00} \right\rangle = 0 \\
b\left| {00} \right\rangle = 0 \\
\end{gathered}$$
To get the most compact state, with maximum symmetry, so that Eq. (\[E:symm\])is satisfied, we take $\psi \left( \zeta \right)$ to be a gaussian with dispersion $\sigma$, and n = m = 0. Then our standard state has the wavefunction
$$\left\langle {{\mathbf{\xi}} _ \bot ,\;\zeta \left| {\Psi _{00} } \right.} \right\rangle = \frac{1}
{{\left( {2\pi } \right)^{\frac{3}
{4}} \sigma ^{\frac{1}
{4}} \left( {\frac{\theta }
{4}} \right)^{\frac{1}
{2}} }}\exp \left[ { - \frac{{{\mathbf{\xi}} _ \bot ^2 }}
{\theta } - \frac{{\zeta ^2 }}
{{4\sigma }}} \right]$$
Then with ${\mathbf{\hat n}} = \left( {\alpha ,\;\phi } \right)$, we get
$$\begin{gathered}
\left\langle {{\mathbf{x}}^2 } \right\rangle = \sigma + \theta \\
\left\langle {\left( {{\mathbf{x}} \cdot {\mathbf{\hat n}}} \right)^2 } \right\rangle = \sigma \cos ^2 \alpha + \frac{\theta }
{2}\sin ^2 \alpha \\
\end{gathered}$$
and
$$\left\langle {\frac{1}
{r}} \right\rangle = \frac{1}
{{\left\langle r \right\rangle }} + \frac{1}
{{\left\langle r \right\rangle ^3 }}\left( {\sigma \cos ^2 \alpha + \frac{\theta }
{2}\sin ^2 \alpha } \right) + O\left( {\frac{1}
{{\left\langle r \right\rangle ^4 }}} \right)$$
For this system, the deviation from the 1/r Coulomb potential depends on the direction along which the potential is measured. In terms of the coordinate distance a, this is
$$\label{E:acoul}
\left\langle {\frac{1}
{r}} \right\rangle = \frac{1}
{a} + \frac{1}
{{2a^3 }}\left( {\sigma \left( {3\cos ^2 \alpha - 1} \right) + \frac{\theta }
{2}\left( {3\sin ^2 \alpha - 2} \right)} \right) + O\left( {\frac{1}
{{a^4 }}} \right)$$
Effective Potential at Origin {#S:coordpos}
=============================
We distinguish the vector operator that measures position, ${\mathbf{x}}$, from the auxiliary vector operator ${\mathbf{\xi}} =
\left( {{\mathbf{\xi}} _ \bot ,{\kern 1pt} \,\zeta } \right)$. Eigenvalues of the latter we will consider as a kind of “coordinate” position, analogous to the coordinate position of general relativity. Expectation values of the operator ${\mathbf{x}}$ give the “physical” position.
It is of interest ot ask what an operator, in particular the Coulomb potential, looks like in terms of coordinate position. A qualitative treatment of this problem has been given by Colatto et al. [@colatto] To determine this, we consider the states
$$\left| {\Psi \left( {\sigma ,\;\sigma _ \bot ,\;{\mathbf{a}}}
\right)} \right\rangle = \int {d^3 \xi } \left|{\mathbf{ \xi}} \right\rangle \frac{1}
{{\left( {2\pi } \right)^{\frac{3}
{4}} \sigma ^{\frac{1}
{4}} \sigma _ \bot ^{\frac{1}
{2}} }}\exp \left[ { - \frac{{\left( {\zeta - a^3 } \right)^2 }}
{{4\sigma }} - \frac{{\left( {{\mathbf{\xi}} _ \bot - {\mathbf{a}}_ \bot } \right)^2 }}
{{4\sigma _ \bot }}} \right]$$
These states give us a kind of window at the point ${\mathbf{a}}$, of widths $\sqrt \sigma ,\;\sqrt {\sigma _ \bot } $ in the z, r directions, to determine the “effective” value of any operator A at that point:
$$\begin{gathered}
\left\langle {A\left( {\mathbf{a}} \right)} \right\rangle = \left\langle {\Psi \left( {\sigma ,\;\sigma _ \bot ,\;{\mathbf{a}}} \right)} \right|A\left| {\Psi \left( {\sigma \;,\sigma _ \bot ,\;{\mathbf{a}}} \right)} \right\rangle \\
= \left\langle {\mathbf{a}} \right|A\left| {\mathbf{a}} \right\rangle \\
\end{gathered}$$
To look at the Coulomb potential we use the Fourier transform expression
$$\frac{1}
{r} = \frac{1}
{{2\pi ^2 }}\int {d^3 q} \frac{1}
{{q^2 }}\exp \left[ {i{\mathbf{q}} \cdot {\mathbf{x}}} \right]$$
We note that
$$\begin{gathered}
\exp \left[ {i{\mathbf{q}} \cdot {\mathbf{x}}} \right] = \exp
\left[ {i{\mathbf{q}} \cdot \left( {{\mathbf{\xi}} + \tfrac{1}
{2}{\mathbf{\theta}} \times {\mathbf{p}}} \right)} \right] \\
= \exp \left[ {i{\mathbf{q}} \cdot {\mathbf{\xi}} } \right]\exp \left[ {i{\mathbf{q}} \cdot \tfrac{1}
{2}{\mathbf{\theta}} \times {\mathbf{p}}} \right] \\
\end{gathered}$$
since
$$\left[ {{\mathbf{q}} \cdot {\mathbf{\xi}} ,\;{\mathbf{q}} \cdot
{\mathbf{\theta}} \times {\mathbf{p}}} \right] = i{\mathbf{q}} \cdot {\mathbf{\theta}} \times {\mathbf{q}} = 0$$
also
$$\exp \left[ {i{\mathbf{q}} \cdot \tfrac{1}
{2}{\mathbf{\theta}} \times {\mathbf{p}}} \right]\left| {\mathbf{\xi}} \right\rangle = \left| {{\mathbf{\xi}} + \tfrac{1}
{2}{\mathbf{\theta}} \times {\mathbf{q}}} \right\rangle$$
Then
$$\begin{aligned}
\left\langle {\mathbf{a}} \right|\frac{1}
{r}\left| {\mathbf{a}} \right\rangle = & \frac{1}
{{2\pi ^2 \left( {2\pi } \right)^{\frac{3}
{2}} }}\int {\frac{{d^3 \xi 'd^3 \xi }}
{{\sigma _ \bot \sqrt \sigma }}\frac{{d^3 q}}
{{q^2 }}\exp \left[ {i{\mathbf{q}} \cdot {\mathbf{\xi}} ' - \frac{{\left( {\zeta ' - a^3 } \right)^2 + \left( {\zeta - a^3 } \right)^2 }}
{{4\sigma }} - \frac{{\left( {{\mathbf{\xi}} '_ \bot - {\mathbf{a}}_ \bot } \right)^2 + \left( {{\mathbf{\xi}} _ \bot - {\mathbf{a}}_ \bot } \right)^2 }}
{{4\sigma _ \bot }}} \right]} \\
& \times \left\langle {\mathbf{\xi}} '\vert \, {\mathbf{\xi}} + \frac{1}
{2}{\mathbf{\theta}} \times {\mathbf{q}} \right\rangle \\
\end{aligned}$$
which becomes
$$\label{E:potina}
\left\langle {\mathbf{a}} \right|\frac{1}
{r}\left| {\mathbf{a}} \right\rangle = \frac{1}
{{2\pi ^2 }}\int {\frac{{d^3 q}}
{{q^2 }}\exp \left[ {i{\mathbf{q}} \cdot {\mathbf{a}} - \frac{{\sigma \left( {q^3 } \right)^2 }}
{2} - \frac{{\sigma _p {\mathbf{q}}_ \bot ^2 }}
{2}} \right]}$$
where
$$\label{E:sigmap}
\sigma _p = \sigma _ \bot + \frac{{\theta ^2 }}
{{16\sigma _ \bot }}$$
At very large $\left| a \right|$, we can write this to first order in the dispersions,
$$\label{E:largea}
\begin{gathered}
\left\langle {\mathbf{a}} \right|\frac{1}
{r}\left| {\mathbf{a}} \right\rangle = \left[ {1 + \frac{\sigma }
{2}\left( {\frac{\partial }
{{\partial a^3 }}} \right)^2 + \frac{{\sigma _p }}
{2}\nabla _ \bot ^2 } \right]\frac{1}
{a} \\
= \frac{1}
{a} + \frac{1}
{{a^3 }}\left[ {\sigma \left( {3\cos ^2 \alpha - 1} \right) + \sigma _p \left( {3\sin ^2 \alpha - 2} \right)} \right] \\
\end{gathered}$$
The dispersion of Eq. (\[E:sigmap\]) is minimized by $\sigma _ \bot = \frac{\theta }
{4}$ , which makes $\sigma _p = \frac{\theta }
{2}$ and Eq. (\[E:largea\]) becomes identical with Eq. (\[E:acoul\]).
Unlike the Coulomb potential in commuting geometry, the effective potential in $\theta$ NC is finite at a = 0, where we have
$$\begin{gathered}
\left\langle 0 \right|\frac{1}
{r}\left| 0 \right\rangle = \frac{1}
{{2\pi ^2 }}\int {\frac{{d^3 q}}
{{q^2 }}\exp \left[ { - \frac{{\sigma \left( {q^3 } \right)^2 }}
{2} - \frac{{\sigma _p {\mathbf{q}}_ \bot ^2 }}
{2}} \right]} \\
= \frac{1}
{{2\pi }}\int_{ - 1}^{ + 1} {dc} \frac{1}
{{\sqrt {\sigma _p - \left( {\sigma _p - \sigma } \right)c^2 } }} \\
\end{gathered}$$
If $\sigma > \sigma _p $, this becomes
$$\left\langle 0 \right|\frac{1}
{r}\left| 0 \right\rangle = \frac{1}
{{\pi \sqrt {\sigma - \sigma _p } }}\sinh ^{ - 1} \left( {\sqrt {\frac{{\sigma - \sigma _p }}
{{\sigma _p }}} } \right)$$
At $\sigma = \sigma _p $ this becomes $\frac{1}
{{\pi \sqrt {\sigma _p } }}$ , and for $\sigma < \sigma _p $ we get
$$\left\langle 0 \right|\frac{1}
{r}\left| 0 \right\rangle = \frac{1}
{{\pi \sqrt {\sigma _p - \sigma } }}\arcsin \left( {\sqrt {\frac{{\sigma _p - \sigma }}
{{\sigma _p }}} } \right)$$
which is finite at $\sigma = 0$ and equal to $\frac{1}
{{2\sqrt {\sigma _p } }}$ .
The Charge Density {#S:chgden}
==================
The charge density, expressed as
$$\begin{gathered}
\rho = - \frac{1}
{{4\pi }}\nabla ^2 \frac{1}
{r} \\
= \frac{1}
{{4\pi }}\left[ {{\mathbf{p}} \cdot ,\;\left[ {{\mathbf{p}},\;\frac{1}
{r}} \right]} \right] \\
\end{gathered}$$
has expectation value in the coordinate states given by
$$\begin{gathered}
\rho \left( {\mathbf{a}} \right) = \left\langle {\mathbf{a}} \right|\rho \left| {\mathbf{a}} \right\rangle \\
= - \frac{1}
{{4\pi }}\nabla ^2 \left\langle {\mathbf{a}} \right|\frac{1}
{r}\left| {\mathbf{a}} \right\rangle \\
\end{gathered}$$
Using Eq. (\[E:potina\]) this becomes a Gaussian distribution:
$$\label{E:rho}
\rho \left( {\mathbf{a}} \right) = \frac{1}
{{\left( {2\pi } \right)^{\frac{3}
{2}} \sqrt \sigma \sigma _p }}\exp \left( { - \frac{{\left( {a^3 } \right)^2 }}
{{2\sigma }} - \frac{{{\mathbf{a}}_ \bot ^2 }}
{{2\sigma _p }}} \right)$$
The “Effective” Electric Field {#S:elecfld}
==============================
With an effective potential at the point ${\mathbf{a}}$ defined as
$$\phi \left( {\mathbf{a}} \right) = \left\langle {\mathbf{a}} \right|\frac{1}
{r}\left| {\mathbf{a}} \right\rangle$$
we can define an effective field at ${\mathbf{a}}$ as
$${\mathbf{E}}\left( {\mathbf{a}} \right) = - \nabla _a \phi \left( {\mathbf{a}} \right)$$
The “effective” energy contained in this field may be of some interest. Defining it as
$$U_E = \frac{1}
{{8\pi }}\int {d^3 a{\mathbf{E}}^2 \left( {\mathbf{a}} \right)}$$
we find
$$\begin{gathered}
U_E = \frac{1}
{{8\pi }}\int {d^3 a\left( {\nabla \phi } \right)} ^2 \\
= - \frac{1}
{{8\pi }}\int {d^3 a\phi \nabla ^2 \phi } \\
= \frac{1}
{2}\int {d^3 a\phi \rho } \\
\end{gathered}$$
From Eqs. (\[E:potina\]) and (\[E:rho\]) we get, for $\sigma < \sigma _p $ ,
$$U_E = \frac{1}
{{2\sqrt 2 \pi \sqrt {\sigma _p - \sigma } }}\arcsin \left( {\sqrt {\frac{{\sigma _p - \sigma }}
{{\sigma _p }}} } \right)$$
At the extreme limit, when $\sigma = 0$ and $\sigma _p = \frac{\theta }
{2}$, this becomes $U_E = \frac{1}
{{4\sqrt \theta }}$.
The True Electric Field {#S:trufld}
=======================
Although the integral over all space of the square of the expectation value of the electric field operator is finite, this is not true of the operator $
{\mathbf{E}}^2$ itself. Let $A\left({\mathbf{x}} \right)$ be any operator function of ${\mathbf{ x}}$. It can be written as
$$A\left( {\mathbf{x}} \right) = \int {\frac{{d^3 q}}
{{\left( {2\pi } \right)^4 }}\tilde A\left( {\mathbf{q}} \right)\exp
\left[ {i{\mathbf{q}} \cdot \left( {{\mathbf{\xi}} + \frac{1}
{2}{\mathbf{\theta}} \times {\mathbf{p}}} \right)} \right]}$$
where
$$\tilde A\left( {\mathbf{q}} \right) = \int {d^3 a\,A\left( {{\mathbf{x}} + {\mathbf{a}}} \right)\exp \left( { - i{\mathbf{q}} \cdot {\mathbf{a}}} \right)\exp \left[ { - i{\mathbf{q}} \cdot {\mathbf{x}}} \right]}$$
in particular, the operator’s space integral can be found from
$$\tilde A\left( 0 \right) = \int {d^3 a\,A\left( {{\mathbf{x}} + {\mathbf{a}}} \right)}$$
The operator $E^2$ can be written as
$${\mathbf{E}}^2 \left( {\mathbf{x}} \right) = \frac{1}
{{4\pi ^4 }}\int {\frac{{d^3 q'd^3 q}}
{{q'^2 \left( {{\mathbf{q}} + {\mathbf{q'}}} \right)^2 }}{\mathbf{q'}} \cdot } \left( {{\mathbf{q}} + {\mathbf{q'}}} \right)\exp \left[ { - \frac{i}
{2}{\mathbf{q'}} \cdot \theta \times {\mathbf{p}}} \right]\exp \left( {i{\mathbf{q}} \cdot {\mathbf{x}}} \right)\exp \left[ { + \frac{i}
{2}{\mathbf{q'}} \cdot \theta \times {\mathbf{p}}} \right]$$
from which it follows that
$$\int {d^3 a\,} {\mathbf{E}}^2 \left( {{\mathbf{x}} + {\mathbf{a}}} \right) = \frac{1}
{{4\pi ^4 }}\int {\frac{{d^3 q}}
{{q^2 }}}$$
and is infinite.
Conclusions
===========
In $\theta$ NC geometry the physical operators corresponding to the measurement of position, the $x^{i}$ , are non-commuting, but can be exoressed in terms of the momentum operators $p^{i}$, and commuting “coordinate” operators $\xi^{i}$. States of minimum dispersion in $x^{i}$ can be expressed as Gaussian eigenfunctions of the $\xi^{i}$. These states are the closest one can get in $\theta$ NC to a point particle.
The antisymmetric tensor $\theta^{ij}$ defines a vector $\theta^{k}$. The physical distance to a point charge, for fixed coordinate distance, depends on the direction relative to ${\bf{\theta}}$ along which the distance is measured. The departures of the measured Coulomb potential from 1/r are also direction dependent.
In commutative geometry, the expectation value of a field operator such as the potential, the charge distribution, or the electric field, in a Gaussian state located at a point, becomes equal to the value of the field at that point when the state’s dispersion goes to zero. One might therefore consider in $\theta$ NC the expectation of the value of the field in a minimum dispersion state centered at some point to be the “effective” equivalent of the value of the field at that point. We find the effective value of the potential at zero position to be finite. Also the integral over all space of the”effective” electrostatic energy density of a point charge is finite, as discussed by Colatto et al. [@colatto] However, the “effective” value involves just one matrix element. We show that the integral over all space of the energy density operator is still infinite, just as in commutative geometry.
Acknowledgements
================
I would like to thank O. W. Greenberg for many helpful discussions, and also the University of Maryland Physics Dept. for their very kind hospitality.
[9]{} Michael R. Douglas and Nikita A. Nekrasov, *Noncommutative Field Theory*, Rev. Mod. Phys. ${\bf{73}}$, 977-1029 (2001), hep-th/0106048 C. P. Colatto et al., *Non Commutative Geometry Induced by Spin Effects*, hep-th/0512266
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study swept-out Monge-Ampère measures of plurisubharmonic functions and boundary values related to these measures.'
address:
- |
Department of Mathematics and Mathematical Statistics\
Ume[å]{} University\
SE-901 87 Ume[å]{}\
Sweden
- |
Department of Mathematics and Mathematical Statistics\
Ume[å]{} University\
SE-901 87 Ume[å]{}\
Sweden
author:
- Urban Cegrell
- Berit Kemppe
title: 'Monge-Ampère boundary measures'
---
Introduction
============
The purpose of this paper is to study certain boundary measures related to plurisubharmonic functions on hyperconvex domains. These measures are obtained as swept-out Monge-Ampère measures and generalize the boundary measures studied by Demailly in [@Dem1], see Section \[sec\_constr\]. A number of properties of the measures, such as density, support and convergence, are given in Section \[sec\_facts\]. The idea is then to use these measures to define and study boundary values of plurisubharmonic functions on the given domain. This is done in Section \[sec\_bdryval\], where we also describe some situations where this coincides with other notions of boundary values. Finally in Section \[sec\_more\] we study more general boundary measures on a more restricted class of hyperconvex domains. Here we start with a measure on the boundary and find a sequence of Monge-Ampère measures approximating the given measure.
It is a great pleasure for us to thank Phm Hoàng Hip for many fruitful comments.
Preliminaries {#sec_prel}
=============
We first recall some definitions needed in this paper. Let $\Omega$ be a domain in $\mathbb C^n$, $n\ge 2$. Denote by $PSH(\Omega)$ the plurisubharmonic functions on $\Omega$ and by $PSH^-(\Omega)$ the subclass of nonpositive functions. A set $\Omega\subset\mathbb C^n$ is said to be a hyperconvex domain if it is open, connected and if there exists a function $\varphi\in
PSH^-(\Omega)$ such that $\{z\in\Omega: \varphi(z)<-c\}\subset\subset\Omega$, $\forall\, c>0$. If $\Omega$ is a bounded hyperconvex domain, then it can be shown that the exhaustion function $\varphi$ can be chosen in $C^\infty(\Omega)\cap
C(\bar\Omega)$ and such that $\int_\Omega (dd^c \varphi)^n<+\infty$ (see [@Ceg5]). This implies for example that the classes defined below are nontrivial. Unless otherwise stated, $\Omega$ will throughout this paper denote a bounded hyperconvex domain in $\mathbb C^n$. Also, by a measure we mean a positive regular Borel measure.
Let ${\mathcal{E}}_0(\Omega)$, ${\mathcal{F}}(\Omega)$, ${\mathcal{E}}(\Omega)$ and ${\mathcal{F}}^{a}(\Omega)$ be the subclasses of $PSH^{-}(\Omega)$ defined as in [@Ceg1] and [@Ceg2], namely as follows:
- ${\mathcal{E}}_0(\Omega)$ is the set of functions $u\in PSH(\Omega)\cap L^\infty(\Omega)$ such that $\int_\Omega (dd^c u)^n < +\infty$ and $\lim_{z\to\xi} u(z)=0$, $\forall\,\xi\in{{\partial\Omega}}$
- ${\mathcal{F}}(\Omega)$ is the set of functions $u\in PSH(\Omega)$ such that there is a sequence $\{u_j\}$ in ${\mathcal{E}}_0(\Omega)$ with the properties that $u_j\searrow u$ and $\sup_j \int_\Omega (dd^c u_j)^n < +\infty$
- ${\mathcal{E}}(\Omega)$ is the set of functions $u\in PSH(\Omega)$ such that for each $\omega\subset\subset\Omega$ there is function $u_\omega\in{\mathcal{F}}(\Omega)$ with the properties that $u_\omega\ge u$ on $\Omega$ and $u_\omega=u$ on $\omega$
- ${\mathcal{F}}^a(\Omega)$ is the set of functions $u\in{\mathcal{F}}(\Omega)$ such that $\int_E(dd^c u)^n = 0$ for each pluripolar set $E\subset\Omega$
For the convenience of the reader, we state some of the results, concerning these classes, that we use most frequently in this paper. If nothing else is mentioned, proofs can be found in [@Ceg2].
First, observe that $PSH^-(\Omega)\cap L^\infty_{\text{loc}}(\Omega)$ is contained in ${\mathcal{E}}(\Omega)$ and that ${\mathcal{E}}_0(\Omega) \subset{\mathcal{F}}^{a}(\Omega)\subset{\mathcal{F}}(\Omega)
\subset{\mathcal{E}}(\Omega)$. The following lemma explains why the functions in ${\mathcal{E}}_0(\Omega)$ sometimes are called *test functions*.
\[l\_testfcns\] If $\varphi\in C_0^\infty(\Omega)$, then there are $\phi_1,\phi_2\in{\mathcal{E}}_0(\Omega)\cap C(\bar\Omega)$ such that $\varphi=\varphi_1-\varphi_2.$
If $u_1,\ldots,u_n\in{\mathcal{E}}(\Omega)$, then $dd^c u_1\wedge\ldots\wedge dd^c u_n$ is defined as the limit measure obtained by combining the following two theorems.
\[th\_appr\_d\] Suppose that $u\in PSH^-(\Omega)$. Then there is a sequence $\{u_j\}\subset{\mathcal{E}}_0(\Omega)\cap C(\bar\Omega)$ such that $u_j\searrow u$ on $\Omega$ and $\text{supp}\,(dd^c u_j)^n\subset\subset\Omega$ for each $j$.
\[th\_conv\_d\] For $k=1,\ldots,n$, let $u_k\in{\mathcal{E}}(\Omega)$ and $\{g_{kj}\}_{j=1}^\infty\subset{\mathcal{E}}_0(\Omega)$ be such that $g_{kj}\searrow u_k$ as $j\to\infty$. Then $dd^c g_{1j}\wedge\ldots\wedge dd^c g_{nj}$ is weak\*-convergent and the limit measure is independent on the sequences $\{g_{kj}\}$.
A function $u\in{\mathcal{E}}(\Omega)$ is a maximal plurisubharmonic function if and only if $(dd^cu)^n=0$ (see [@Blocki] and [@Ceg1_2]). If $u\in{\mathcal{F}}(\Omega)$ and $(dd^cu)^n=0$, then $u=0$ (see Theorem 5.15 in [@Ceg2]). Theorem \[th\_conv\_d\] can be generalized as follows, see e.g. Lemma 3.2 in [@Ceg00].
\[l\_gen\_conv\] For $k=1,\ldots,n$, let $u_k\in{\mathcal{E}}(\Omega)$ and $\{g_{kj}\}_{j=1}^\infty\subset{\mathcal{E}}(\Omega)$ be such that $g_{kj}\ge u_k$ and $g_{kj}$ tends weakly to $u_k$ as $j\to\infty$. If $h\in PSH^-(\Omega)\cap L^\infty(\Omega)$, then $h\,dd^c g_{1j}\wedge\ldots\wedge dd^c g_{nj}$ tends weak\* to $h\,dd^c u_1\wedge\ldots\wedge dd^c u_n$. Moreover, if $u_k\in{\mathcal{F}}(\Omega)$ then $\lim_{j\to\infty}\int_\Omega
h\,dd^c g_{1j}\wedge\ldots\wedge dd^c g_{nj}=
\lim_{j\to\infty}\int_\Omega
h\,dd^c u_1\wedge\ldots\wedge dd^c u_n$.
The next lemma contains some useful basic properties of the classes we use.
\[lemma\_propert\] Let $\mathcal{K}\in\{{\mathcal{E}}_0,{\mathcal{F}}^{a},{\mathcal{F}},{\mathcal{E}}\}$, then the following holds.
(i) If $u,v\in\mathcal{K}(\Omega)$ and $\alpha,\beta \ge 0$, then $\alpha u + \beta v \in\mathcal{K}(\Omega)$.
(ii) If $u\in\mathcal{K}(\Omega)$ and $v\in PSH^-(\Omega)$, then $\text{max}\,\{u,v\}\in\mathcal{K}(\Omega)$. In particular, if $u\in\mathcal{K}(\Omega)$, $v\in PSH^-(\Omega)$ and $v\ge u$, then $v\in\mathcal{K}(\Omega)$.
Note that functions in ${\mathcal{F}}(\Omega)$ have finite total Monge-Ampère mass. Also, they have in some sense boundary values zero, which can be seen e.g. in the following formula for partial integration.
\[th\_pi\] Let $v,u_1,\ldots,u_n\in{\mathcal{F}}(\Omega)$. Then $$\int_\Omega v\,dd^c u_1\wedge dd^c u_2\wedge\ldots
\wedge dd^c u_n =
\int_\Omega u_1\,dd^c v\wedge dd^c u_2\wedge\ldots
\wedge dd^c u_n.$$
Since bounded function cannot put Monge-Ampère mass on pluripolar sets (see e.g. ), we have that ${\mathcal{F}}(\Omega)\cap L^\infty(\Omega)\subset{\mathcal{F}}^{a}(\Omega)$. Moreover, Theorem 5.5 and Theorem 5.8 in [@Ceg2] gives:
\[l\_pp\] If $u_1,\ldots,u_{n-1}\in{\mathcal{F}}(\Omega)$ and $v\in{\mathcal{F}}^{a}(\Omega)$ or $v\in PSH^-(\Omega)\cap L^\infty(\Omega)$, then $dd^c u_1\wedge\ldots \wedge dd^c u_{n-1}\wedge dd^c v$ vanishes on pluripolar sets.
We conclude this section with some notation needed in this paper. Let $\Omega$ and $u\in{\mathcal{E}}(\Omega)$ be given and choose a fundamental sequence $\{\Omega_j\}$ of strictly pseudoconvex domains, i.e.$\Omega_j\subset\subset\Omega_{j+1}\subset\subset\Omega$ and $\cup_{j=1}^\infty\Omega_j=\Omega$. For each $j$ define $$\label{uj_def}
u^j = \sup\,\{\varphi\in PSH(\Omega):
\varphi|_{\Omega\setminus\Omega_j}\le
u|_{\Omega\setminus\Omega_j}\}.$$ Note that since $\Omega_j$ has $C^2$ boundary, it follows that $u^j=(u^j)^*$, the smallest upper semicontinuous majorant of $u^j$, so $u^j$ is plurisubharmonic. Moreover $u\le u^j \le u^{j+1}\le 0$, so each $u^j\in{\mathcal{E}}(\Omega)$ and the same holds for $\tilde u = (\lim_{j\to\infty} u^j)^{*}$. It follows that $\tilde u$ is the smallest maximal plurisubharmonic majorant of $u$ and that $\tilde u$ is independent of the chosen sequence $\{\Omega_j\}$. In [@Ceg00] the following classes were defined: $$\begin{split}
\mathcal{N}(\Omega) & =
\{u\in{\mathcal{E}}(\Omega): \tilde u = 0\}\\
{\mathcal{M}}(\Omega) & =
\{u\in {\mathcal{E}}(\Omega): (dd^c u)^n = 0\}
\end{split}$$ Thus ${\mathcal{M}}(\Omega)$ is the class of maximal plurisubharmonic functions in ${\mathcal{E}}(\Omega)$. Note that $\mathcal{N}(\Omega)$ contains ${\mathcal{F}}(\Omega)$, since if $u\in{\mathcal{F}}(\Omega)$, then $\tilde u$ is a maximal function in ${\mathcal{F}}(\Omega)$ so $\tilde u=0$. It also follows that if $u\in{\mathcal{F}}(\Omega)$, then $u^j\nearrow
0$ outside a pluripolar subset of $\Omega$ (see [@Lel2] or ).
Finally, we say that $u\in{\mathcal{E}}(\Omega)$ has boundary values $\tilde u$ if there is a function $\psi\in\mathcal{N}(\Omega)$ such that $\tilde u\ge u \ge \tilde u + \psi$. Given $H\in{\mathcal{M}}(\Omega)$ we define $${\mathcal{F}}(\Omega,H) =
\{u\in PSH(\Omega): H\ge u \ge H+\psi,\,\psi\in {\mathcal{F}}(\Omega)\},$$ which is a subclass of ${\mathcal{E}}(\Omega)$. It follows that if $u\in{\mathcal{F}}(\Omega,H)$ then $\tilde u = H$. Also, ${\mathcal{F}}(\Omega,0)={\mathcal{F}}(\Omega)$.
Construction of the boundary measures $\mu_u$ {#sec_constr}
=============================================
In this section we show that every function in ${\mathcal{F}}(\Omega)$ gives rise to a measure on the boundary of $\Omega.$ Let $u\in{\mathcal{F}}(\Omega)$ be given, choose a fundamental sequence $\{\Omega_j\}$ of strictly pseudoconvex domains and let $u^j$ be defined by (\[uj\_def\]). Then $u\le u^j \le u^{j+1}\le 0$, so each $u^j\in{\mathcal{F}}(\Omega)$. Moreover, Stokes’ theorem implies that $\int_\Omega(dd^c u^j)^n = \int_\Omega(dd^c u)^n<+\infty$, and by maximality $(dd^c u^j)^n$ is concentrated on $\Omega\setminus\Omega_j$.
\[th1\] Suppose that $u\in{\mathcal{F}}(\Omega)$. Then $\{(dd^cu^j)^n\}$ is a weak\*-convergent sequence, which defines a positive measure $\mu_u$ on ${{\partial\Omega}}$. Also $\lim_{j\to\infty}\int_\Omega\varphi\,(dd^cu^j)^n$ exists for all $\varphi\in PSH(\Omega)\cap L^{\infty}(\Omega)$.
Choose W to be a strictly pseudoconvex set containing the closure of $\Omega$. First assume that $\varphi\in PSH(\Omega)\cap L^\infty(\Omega)$ and $\varphi\le 0$, then $$\label{ineq_1}
-\infty<\int_\Omega \varphi\,(dd^c u)^n
\le \int_\Omega \varphi\,(dd^c u^j)^n
\le \int_\Omega \varphi\,(dd^c u^{j+1})^n
\le \sup_\Omega \varphi\int_\Omega (dd^c u)^n.$$ To see this, approximate $\varphi$ with functions in ${\mathcal{E}}_0(\Omega)$ and use partial integration in ${\mathcal{F}}(\Omega)$ (see Section \[sec\_prel\]). Since all Monge-Ampère measures involved have the same total mass, it follows that (\[ineq\_1\]) holds for all $\varphi\in PSH(\Omega)\cap L^\infty(\Omega)$. Thus $\{\int_\Omega \varphi\,(dd^c u^j)^n\}$ is a bounded monotone sequence, so $\lim_{j\to\infty}\int_\Omega \varphi\,(dd^c u^j)^n$ exists for all $\varphi\in PSH(\Omega)\cap L^\infty(\Omega)$. In particular the limit exists for $\varphi\in C_0^\infty(W)$ (see Lemma \[l\_testfcns\]). Since each $(dd^c u^j)^n$ is a positive distribution on $C_0^\infty (W)$, it follows from standard distribution theory that the convergence in fact holds for all $\varphi\in C_0(W)$. Also the limit distribution itself is positive and thus defines a positive regular Borel measure $\mu_u$ on $W$, which by the construction is concentrated on ${{\partial\Omega}}$.
In this manner we may, to each $u\in{\mathcal{F}}(\Omega)$, associate a positive measure $\mu_u$, and it follows for example that $$\label{limit_1}
\int_{{\partial\Omega}}\varphi\, d\mu_u =
\lim_{j\to\infty}\int_\Omega \varphi\,(dd^c u^j)^n$$ holds for all $\varphi\in C_0(W)$, in particular for $\varphi\in C(\bar\Omega)$. We also have that $\int_{{\partial\Omega}}d\mu = \int_\Omega (dd^c u)^n$, which implies that $\mu_u=0$ if and only if $u=0$ (since $u\in{\mathcal{F}}(\Omega)$). Note that $\mu_u$ does not depend on the chosen sequence $\{\Omega_j\}$. Note also that by applying (\[ineq\_1\]) to $\varphi$ and $-\varphi$ we get that $$\label{eq_1}
\int_\Omega \varphi\,(dd^c u^j)^n=
\int_\Omega \varphi\,(dd^c u)^n,\quad
\forall\,\varphi\in PH(\Omega)\cap L^\infty(\Omega),$$ where $PH(\Omega)$ denotes the pluriharmonic functions on $\Omega$.
In [@Dem1] Demailly defines a set of Monge-Ampère boundary measures in the following setting. Let $X$ be a Stein manifold of dimension $n$ and $\Omega\subset\subset X$ an open hyperconvex subset. Assume that $\phi:\Omega\to [-\infty,0)$ is a continuous plurisubharmonic exhaustion function such that $\int_\Omega (dd^c \phi)^n < +\infty$. For each $r<0$ define: $$\begin{aligned}
B(r)&=&\{z\in\Omega : \phi(z) < r\}\\
S(r)&=&\{z\in\Omega : \phi(z) = r\}\\
\phi_r(z)&=&\max\,\{\phi(z),r\}\end{aligned}$$ It is then shown that $$(dd^c \phi_r)^n =
\chi_{\Omega\setminus B(r)}\cdot (dd^c \phi)^n +
\mu_{\phi,r}$$ where $\mu_{\phi,r}$ is a positive measure concentrated on $S(r)$. Furthermore, when $r\to 0$ then $\mu_{\phi,r}$ converges in a weak sense to a positive measure $\tilde{\mu}_\phi$ concentrated on ${{\partial\Omega}}$. (More explicitly it is shown that $\lim_{r\to 0}\int h\,d\mu_{\phi,r}$ exists $\forall\, h\in C^2(X,\mathbb{R})$.)
Now consider the case when $X=\mathbb{C}^n$, then the function $\phi$ is in ${\mathcal{F}}(\Omega)$ so we can define $\mu_\phi$ according to Theorem \[th1\]. Choose a sequence $\{r_j\}$ such that $r_j\nearrow 0$ and let $\Omega_j = B(r_j)$. Then $ \phi_{r_j}=\max\,\{\phi,r_j\}$ is equal to the function $\phi^j$ defined as in (\[uj\_def\]). Note that $\Omega_j$ is not necessarily strictly pseudoconvex in this setting, only hyperconvex. However, this is enough in the proof of Theorem \[th1\], since we only use the smoothness of ${{\partial\Omega}}_j$ to ensure that the function $\phi^j$ is plurisubharmonic. Hence $$(dd^c \phi_{r_j})^n =
\chi_{\Omega\setminus B(r_j)}\cdot (dd^c \phi)^n +
\mu_{\phi,r_j},$$ where the left hand side converges to the boundary measure $\mu_\phi$ and the right hand side to $0+\tilde{\mu}_\phi$ (since $\int_\Omega (dd^c \phi)^n < +\infty$). This shows that $\mu_\phi = \tilde{\mu}_\phi$, so in particular Demailly’s boundary measures form a subset of those defined in Theorem \[th1\], when $X=\mathbb{C}^n$.
Also, note that if $u\in{\mathcal{E}}_0(\Omega)\cap C(\bar\Omega)$ then $u$ satisfies the conditions in Demailly’s definition, so for boundary measures corresponding to such functions we may use Demailly’s results.
The following theorem, where $u^j$ is defined by (\[uj\_def\]), generalizes a formula considered by Demailly in [@Dem1].
\[th\_formula\] Assume that $u\in {\mathcal{F}}(\Omega)$, $h\in {\mathcal{E}}(\Omega)$, $\int_\Omega h\,(dd^c u)^n > -\infty $ and that $dd^ch\wedge (dd^c u)^{n-1}$ vanishes on pluripolar sets. Then $$\lim_{j\to\infty} \int_\Omega h\,(dd^c u^j)^n =
\int_\Omega h\,(dd^c u)^n - \int_\Omega u\,dd^c h\wedge (dd^c u)^{n-1}.$$
Note that the conditions in this theorem are satisfied if for example $u\in{\mathcal{F}}(\Omega)$ and $h\in PSH^{-}(\Omega)\cap L^\infty(\Omega)$ (see Lemma \[l\_pp\]). Actually, it is enough that $h\in PSH(\Omega)\cap L^\infty(\Omega)$, since $\int_\Omega(dd^c u^j)^n = \int_\Omega(dd^c u)^n$.
First we claim the following.
(i) \[c1\] $\displaystyle
\int_\Omega u\,dd^c h\wedge(dd^c u)^{n-1} > -\infty$
(ii) \[c2\] $\displaystyle\lim_{j\to\infty}
\int_\Omega u^j\,dd^c h\wedge(dd^c u)^{n-1}=0$
(iii) \[c3\] $\displaystyle
\int_\Omega h\,(dd^c u^j)^{n-p+1}\wedge (dd^c u)^{p-1}\ge
\int_\Omega h\,(dd^c u^j)^{n-p}\wedge (dd^c u)^{p}\ge
\int_\Omega h\,(dd^c u)^{n}$, $1\le p\le n-1$
(iv) \[c4\] $\displaystyle
\int_\Omega
h\,dd^c(u^j-u)\wedge(dd^c u^j)^{n-p}\wedge(dd^c u)^{p-1}=$
$\displaystyle =\int_\Omega u^j\,dd^c h\wedge dd^c(u^j-u)
\wedge(dd^c u^j)^{n-p-1}\wedge(dd^c u)^{p-1}=$
$\displaystyle =\int_\Omega
(u^j-u)\,dd^c h\wedge(dd^c u^j)^{n-p}\wedge(dd^c u)^{p-1}\ge0$, $1\le p\le n$
For the proof of (\[c1\]), choose a sequence $\{h_k\}$ in ${\mathcal{E}}_0(\Omega)$ decreasing to $h$ on $\Omega$. Then $dd^c h_k\wedge(dd^c u)^{n-1}$ converges weak\* to $dd^c h\wedge(dd^c u)^{n-1}$ (Lemma \[l\_gen\_conv\]). Combining this with the fact that $u$ is upper semicontinuous it follows that $$\begin{split}
\int_\Omega (-u)\,dd^c h\wedge(dd^c u)^{n-1}&\le
\limsup_{k\to\infty}
\int_\Omega (-u)\,dd^c h_k\wedge(dd^c u)^{n-1}=\\
& =\limsup_{k\to\infty}\int_\Omega(-h_k)\,(dd^c u)^{n}
= \int_\Omega (-h)\,(dd^c u)^{n}<+\infty
\end{split}$$ (where we have used partial integration in ${\mathcal{F}}(\Omega)$). Since $u^j\nearrow 0$ outside a pluripolar set (see Section \[sec\_prel\]) and since $dd^ch\wedge (dd^c u)^{n-1}$ puts no mass there, (\[c1\]) implies (\[c2\]) by dominated convergence. To see (\[c3\]), use the same technique as in Theorem \[th1\]. Finally (\[c4\]) follows from partial integration, using the fact that $h$ is locally in ${\mathcal{F}}(\Omega)$ and that $u^j-u$ is compactly supported in $\Omega$. This proves the claim.
Now using (\[c4\]) we have that $$\begin{split}
\int_\Omega u\,dd^c h\wedge(dd^c u)^{n-1} &=
\int_\Omega (u-u^j)\,dd^c h\wedge(dd^c u)^{n-1} +
\int_\Omega u^j\,dd^c h\wedge(dd^c u)^{n-1}=\\
& = \int_\Omega h\,dd^c(u-u^j)\wedge(dd^c u)^{n-1} +
\int_\Omega u^j\,dd^c h\wedge(dd^c u)^{n-1},
\end{split}$$ so we can write $$\begin{gathered}
\int_\Omega h\,(dd^c u^j)^n - \int_\Omega h\,(dd^c u)^n +
\int_\Omega u\,dd^c h\wedge (dd^c u)^{n-1} =\\
= \int_\Omega h\,(dd^c u^j)^n -
\int_\Omega h\,dd^c u^j\wedge(dd^c u)^{n-1} +
\int_\Omega u^j\,dd^c h\wedge(dd^c u)^{n-1}
\end{gathered}$$ where the last integral tends to $0$ according to (\[c2\]). Moreover $$\begin{gathered}
\int_\Omega h\,(dd^c u^j)^n -
\int_\Omega h\,dd^c u^j\wedge(dd^c u)^{n-1} =\\
=\sum_{p=1}^{n-1}
\left(\int_\Omega h\,(dd^c u^j)^{n-p+1}\wedge(dd^c u)^{p-1}-
\int_\Omega h\,(dd^c u^j)^{n-p}\wedge(dd^c u)^{p}\right) =
\sum_{p=1}^{n-1} a_p
\end{gathered}$$ where each $a_p\ge 0$ by (\[c3\]). Using (\[c4\]) we have that $$\begin{split}
a_p & = \int_\Omega
h\,dd^c(u^j-u)\wedge(dd^c u^j)^{n-p}\wedge(dd^c u)^{p-1}=\\
& = \int_\Omega u^j\,dd^c h\wedge dd^c(u^j-u)
\wedge(dd^c u^j)^{n-p-1}\wedge(dd^c u)^{p-1}\le\\
& \le -\int_\Omega
u^j\,dd^c h\wedge(dd^c u^j)^{n-p-1}\wedge(dd^c u)^p.
\end{split}$$ Now, the second expression in (\[c4\]) implies that $\int_\Omega u^j\,dd^c h\wedge(dd^c u^j)^{n-k}\wedge(dd^c u)^{k-1}$ is decreasing in $k$, so it follows that $0\le a_p\le -\int_\Omega u^j\,dd^c h\wedge(dd^c u)^{n-1}$. Hence (\[c2\]) implies that each term $a_p\to 0$ as $j\to\infty$ and the theorem is proved.
\[rem\_pi\] Combining the preceeding theorem with (\[limit\_1\]), we have the following formula. Given $u\in{\mathcal{F}}(\Omega)$, $$\label{pi_formula}
\int_\Omega h\,(dd^c u)^n=
\int_\Omega u\,dd^c h\wedge (dd^c u)^{n-1} +
\int_{{\partial\Omega}}h\,d\mu_u,
\quad\forall\,h\in PSH(\Omega)\cap C(\bar\Omega).$$ In Section \[sec\_facts\] (Corollary \[cor\_supp\_S\]) we will show that there is a set $S\subset{{\partial\Omega}}$ such that $\text{supp}\,\mu_u=S$ for each $u\in{\mathcal{F}}(\Omega)$, $u\ne 0$. Hence (\[pi\_formula\]) gives a partial integration formula for $h\in PSH^-(\Omega)\cap C(\bar\Omega)$ such that $h|_S=0$. From Theorem \[th\_equal\] in Section \[sec\_bdryval\] it follows that if $u\in{\mathcal{F}}^{a}(\Omega)$, then (\[pi\_formula\]) is valid for $h\in PSH(W)\cap L^\infty(W)$, where $W$ is some neighbourhood of $\Omega$. We also get a Jensen-type inequality; given $u\in{\mathcal{F}}(\Omega)$, $$\label{J_ineq}
\int_\Omega h\,(dd^c u)^n \le
\int_{{\partial\Omega}}h\,d\mu_u,
\quad\forall\,h\in PSH(\Omega)\cap C(\bar\Omega).$$ If $h\in PSH(W)$ for some neighbourhood $W$ of $\Omega$, then using convolution we may find functions $h_k\in PSH(W')\cap C(W')$, where $\bar\Omega\subset W'\subset\subset W$, such that $h_k\searrow h$ on $W'$. Therefore (\[J\_ineq\]) holds true if $h\in PSH(W)$ and $u\in{\mathcal{F}}(\Omega)$.
Some properties of the boundary measures $\mu_u$ {#sec_facts}
================================================
In this section we investigate some properties of the boundary measures $\mu_u$ defined in Section \[sec\_constr\]. Recall that a hyperconvex domain $\Omega$ is called B-regular if each continuous function on ${{\partial\Omega}}$ can be extended continuously to a plurisubharmonic function on $\Omega$ (see [@Sib]).
\[th3\] Let $\mu$ be a finite positive measure on $\partial\Omega$, where $\Omega$ is a bounded B-regular domain. Then $\mu$ is in the weak\* closure of $\{\mu_u : u\in {\mathcal{F}}(\Omega)\}$.
For simplicity, assume that $\mu(\partial\Omega)=1$. Choose a sequence of measures $$\mu_k=\sum_{j=1}^{N_k}a_j^k\delta_{z_j^k},
\text{ where }
{\{z_j^k\}}_{j=1}^{N_k}\subset\Omega
\text{ and }
\sum_{j=1}^{N_k}a_j^k=1$$ such that $$\label{th3_eq1}
\lim_{k\to\infty}\int_\Omega h\,d\mu_k =
\int_{{\partial\Omega}}h\,d\mu,
\quad\forall\,h\in C(\bar\Omega).$$ Let e.g. $a_j^k=\mu(A_j^k)$ and $z_j^k\in A_j^k\cap\Omega$, where ${\{A_j^k\}}_{j=1}^{N_k}$ is a partition of $\bar\Omega$ such that $\text{diam}({A_j^k})\le\frac{1}{2^k}$, and use the fact that $h$ is uniformly continuous on $\bar\Omega$. For each $k$, consider $g_k(z)$, the multipole pluricomplex Green’s function for $\Omega$ with poles at $\{z_j^k\}$ with weights $\{(a_j^k)^{1/n}\}$ (see [@Lel] and [@Pol]). Then $g_k\in {\mathcal{F}}(\Omega)$ and $(dd^c g_k)^n = \mu_k$. Form $\tilde\mu_k = \lim_{i\to\infty}(dd^c(g_k)^i)^n$ as in section \[sec\_constr\]. Then for each $k$ $$\label{th3_eq2}
\int_{{\partial\Omega}}d\tilde\mu_k = \int_\Omega(dd^c g_k)^n =
\int_\Omega d\mu_k = 1 = \int_{{\partial\Omega}}d\mu$$ and from (\[limit\_1\]) and (\[ineq\_1\]) it follows that $$\label{th3_eq3}
\int_{{\partial\Omega}}\varphi\,d\tilde\mu_k =
\lim_{i\to\infty}\int_\Omega \varphi\,(dd^c (g_k)^i)^n \ge
\int_\Omega \varphi\,(dd^c g_k)^n =
\int_\Omega \varphi\,d\mu_k$$ for $\varphi\in PSH(\Omega)\cap C(\bar\Omega)$. Let $\{\tilde\mu_{k_m}\}$ be any weak\*-convergent subsequence of $\{\tilde\mu_k\}$. (Such a subsequence exists since the measures $\{\tilde\mu_k\}$ have uniformly bounded total mass.) Now let $t\in C({{\partial\Omega}})$, $t\le 0$ be given. Since $\Omega $ is B-regular there is a $\varphi\in PSH(\Omega)\cap C(\bar\Omega)$ with $\varphi = t$ on ${{\partial\Omega}}$. Hence, by (\[th3\_eq1\]) and (\[th3\_eq3\]), $$\int_{{\partial\Omega}}t\,d\mu = \lim_{m\to\infty}\int_\Omega \varphi\,d\mu_{k_m}
\le \lim_{m\to\infty}\int_{{\partial\Omega}}\varphi\,d\tilde\mu_{k_m} =
\lim_{m\to\infty}\int_{{\partial\Omega}}t \,d\tilde\mu_{k_m}.$$ This shows that $\mu\ge\lim_{m\to\infty}\tilde\mu_{k_m}$. It then follows from (\[th3\_eq2\]) that they have the same total mass, so $\mu=\lim_{m\to\infty}\tilde\mu_{k_m}$ and the theorem is proved. Note that since the argument is valid for any weak\*-convergent subsequence, it follows that $\{\tilde\mu_k\}$ itself tends weak\* to $\mu$.
Later in this section, we will show that not every positive measure on ${{\partial\Omega}}$ is in $\{\mu_u: u\in{\mathcal{F}}(\Omega)\}$, see for example Proposition \[prop\_support\]. Moreover, the assumption of B-regularity cannot be removed in Theorem \[th3\], see for example Corollary \[cor\_supp\_S\] and Example \[ex\_bidisc\]. Before we can prove this, we need the following convergence property.
\[appr\_prop\] Suppose that $u\in{\mathcal{F}}(\Omega)$ and that $\{u_k\}$ is a decreasing sequence in ${\mathcal{F}}(\Omega)$ such that $u_k\searrow u$ on $\Omega$. Then $\mu_{u_k}$ converges weak\* to $\mu_u$.
Let $h\in{\mathcal{E}}_0(\Omega')\cap C(\bar\Omega')$ where $\Omega'\supset\bar\Omega$. Then (\[pi\_formula\]) gives that $$\int_{{\partial\Omega}}h\,d\mu_u =
\int_\Omega h\,(dd^c u)^n -
\int_\Omega u\,dd^c h\wedge(dd^c u)^{n-1}$$ and that for each $k$ $$\int_{{\partial\Omega}}h\,d\mu_{u_k} =
\int_\Omega h\,(dd^c u_k)^n -
\int_\Omega u_k\,dd^c h\wedge(dd^c u_k)^{n-1}.$$ From Lemma \[l\_gen\_conv\] it follows that $\lim_{k\to\infty} \int_\Omega h\,(dd^c u_k)^n =
\int_\Omega h\,(dd^c u)^n$. Moreover, $\lim_{k\to\infty}\int_\Omega u_k\,dd^c h\wedge(dd^c u_k)^{n-1}
=\int_\Omega u\,dd^c h\wedge(dd^c u)^{n-1}$ by the following calculations. Since $u\le u_k$ for each $k$, Lemma 3.3 in [@ACCP] implies that $$\int_\Omega u\,dd^c h\wedge(dd^c u)^{n-1}\le
\int_\Omega u\,dd^c h\wedge(dd^c u_k)^{n-1}\le
\int_\Omega u_k\,dd^c h\wedge(dd^c u_k)^{n-1}$$ for each $k$. Hence, for fixed $k_0$, $$\begin{split}
\int_\Omega u\,dd^c h\wedge(dd^c u)^{n-1}
& \le \liminf_{k\to\infty}
\int_\Omega u_k\,dd^c h\wedge(dd^c u_k)^{n-1}\\
& \le\limsup_{k\to\infty}
\int_\Omega u_k\,dd^c h\wedge(dd^c u_k)^{n-1}\\
& \le\limsup_{k\to\infty}
\int_\Omega u_{k_0}\,dd^c h\wedge(dd^c u_k)^{n-1}\\
& \le \int_\Omega u_{k_0}\,dd^c h\wedge(dd^c u)^{n-1},
\end{split}$$ where the last inequality follows since $dd^c h\wedge(dd^c u_k)^{n-1}$ is weak\*-convergent to $dd^c h\wedge(dd^c u)^{n-1}$ (Lemma \[l\_gen\_conv\]) and $u_{k_0}$ is upper semicontinuous. Now, the claim follows if we let $k_0\to\infty$.
Thus $$\label{limit_h}
\lim_{k\to\infty}\int_{{\partial\Omega}}h\,d\mu_{u_k} =
\int_{{\partial\Omega}}h\,d\mu_{u}$$ holds true for $h\in{\mathcal{E}}_0(\Omega')\cap C(\bar\Omega')$ and therefore for $h\in C_0^\infty(\Omega')$. By standard distribution theory it follows that (\[limit\_h\]) holds for $h\in C_0(\Omega')$ and hence for $h\in C({{\partial\Omega}})$.
Recall from Section \[sec\_constr\] that for functions in ${\mathcal{E}}_0(\Omega)\cap C(\bar\Omega)$ we can apply the results of Demailly in [@Dem1]. We make use of this fact in the proof of the following proposition.
\[ineq\_prop\] If $u$ and $v$ are functions in ${\mathcal{F}}(\Omega)$ such that $u\le v$, then $\mu_u\ge\mu_v$.
Take $\{u_k\},\{w_k\}\subset
{\mathcal{E}}_0(\Omega)\cap C(\bar\Omega)$ such that $u_k\searrow u$ and $w_k\searrow v$. Let $v_k=\max\,\{u_k,w_k\}$. Then $v_k\in
{\mathcal{E}}_0(\Omega)\cap C(\bar\Omega)$, $v_k\searrow v$ and $u_k\le v_k$. By Theorem 3.4 in [@Dem1] $\mu_{u_k}\ge\mu_{v_k}$ for each $k$. Using Proposition \[appr\_prop\] it follows that $\mu_u\ge\mu_v$.
When $\Omega$ is B-regular there is a slightly more direct proof of Proposition \[ineq\_prop\], not using Demailly’s results. If in that case $f\in C({{\partial\Omega}})$, $f\le 0$ is given, it may be extended to a function in $PSH^-(\Omega)\cap C(\bar\Omega)$. Since $u\le v$ we have that $u^j\le v^j$ for each $j$, which (see the proof of Theorem \[th1\]) implies that $\int_\Omega f\,(dd^c u^j)^n\le \int_\Omega f\,(dd^c v^j)^n$ for each $j$. From (\[limit\_1\]) it follows that $\int_{{\partial\Omega}}f\,d\mu_u\le\int_{{\partial\Omega}}f\,d\mu_v$, so we have, by the regularity of $\mu_u$ and $\mu_v$, that $\mu_u\ge\mu_v$.
\[cor\_max\] Suppose that $u\in {\mathcal{F}}(\Omega)$, then $\mu_u = \mu_{\max\,\{u,-1\}}$.
Let $v = \max\,\{u,-1\}$, then $\mu_u\ge \mu_v$ by Proposition \[ineq\_prop\]. Take $\{u_k\}\subset{\mathcal{E}}_0(\Omega)$ such that $u_k\searrow u$ and let $v_k = \max\,\{u_k,-1\}$. Then $v_k\in{\mathcal{E}}_0(\Omega)$, $v_k\searrow v$ and $v_k = u_k$ on $\Omega\setminus\{u_k<-1\}$ (note that $\{u_k<-1\}\subset\subset\Omega$). Using Theorem 5.1 in [@Ceg2] and Stokes theorem, it follows that $$\begin{gathered}
\int_{{\partial\Omega}}d\mu_u =
\int_\Omega (dd^c u)^n =
\lim_{k\to\infty}\int_\Omega (dd^c u_k)^n =\\
=\lim_{k\to\infty}\int_\Omega (dd^c v_k)^n =
\int_\Omega (dd^c v)^n
= \int_{{\partial\Omega}}d\mu_v,
\end{gathered}$$ so $\mu_u = \mu_v$.
We will now use this corollary to show that each $\mu_u$ vanishes on pluripolar sets. We start with two technical lemmas.
\[tec\_lemma\] Suppose that $u\in{\mathcal{F}}(\Omega)$ and that $\varphi$ is in $PSH(\Omega)\cap L^\infty(\Omega)$ and upper semicontinuous on some neighbourhood of $\bar\Omega$. Then $$\lim_{j\to\infty}\int_\Omega \varphi\,(dd^c u^j)^n \le
\int_{{\partial\Omega}}\varphi\,d\mu_u.$$
Choose $\Omega'$ and $\Omega''$ such that $\varphi$ is upper semicontinuous on $\Omega'$ and $\Omega\subset\subset\Omega''\subset\subset\Omega'$. Then there is a decreasing sequence $\{\varphi_k\}$ of continuous functions on $\Omega''$ that are bounded above and that converge to $\varphi$ on $\Omega''$. Using equality (\[limit\_1\]) we have that $$\lim_{j\to\infty}\int_\Omega \varphi\,(dd^c u^j)^n\le
\lim_{j\to\infty}\int_\Omega \varphi_k\,(dd^c u^j)^n=
\int_{{\partial\Omega}}\varphi_k\,d\mu_u$$ for each $k$. Hence the lemma follows by letting $k\to\infty$.
\[lemma\_support\] Let $E\subset{{\partial\Omega}}$ be a pluripolar set and $u\in{\mathcal{F}}(\Omega)$. Suppose that there is a function $g\in PSH(\Omega')$, where $\Omega'\supset\Omega$, such that $E\subset\mathcal{S}_g=\{z: g(z)=-\infty\}$ and $(dd^c u)^n$ is concentrated on $\Omega\setminus\mathcal{S}_g$. Then $\mu_u(E)=0$.
By subtracting a suitable constant we may assume that $g\le 0$ on $\bar\Omega$. For each positive integer $k$, define $h_k=\max\,\{\frac{1}{k}\cdot g, -1\}$. Then from ($\ref{ineq_1}$) and Lemma \[tec\_lemma\] it follows that $$-\infty<\int_\Omega h_k\,(dd^c u)^n \le
\lim_{j\to\infty} \int_\Omega h_k\,(dd^c u^j)^n\le
\int_{{\partial\Omega}}h_k\,d\mu_u\le
\int_E h_k\,d\mu_u = -\mu_u(E),$$ since $h_k\le 0$ on $\bar\Omega$ and $h_k=-1$ on $E$. Moreover, $h_k (z)\nearrow 0$ for all $z\in\Omega\setminus\mathcal{S}_g$, as $k\to\infty$, so $\lim_{k\to\infty}\int_\Omega h_k\,(dd^c u)^n = 0$. Hence $\mu_u(E)=0$.
\[prop\_support\] If $u\in{\mathcal{F}}(\Omega)$, then $\mu_u$ vanishes on pluripolar subsets of ${{\partial\Omega}}$.
If $u\in{\mathcal{F}}(\Omega)$ then $v = \max\,\{u,-1\}\in{\mathcal{F}}^{a}(\Omega)$ and from Corollary \[cor\_max\] we know that $\mu_u=\mu_v$. Now, for functions in ${\mathcal{F}}^{a}(\Omega)$ the conditions in Lemma \[lemma\_support\] are satisfied for each pluripolar set $E\subset{{\partial\Omega}}$, so the proposition follows.
The next proposition enables us to say more about the support of the $\mu_u$-measures.
\[prop\_compare\] Assume that $u,v\in{\mathcal{E}}_0(\Omega)$ are strictly negative functions such that $\text{supp}\,(dd^c u)^n\subset\subset\Omega$ and $\text{supp}\,(dd^c v)^n\subset\subset\Omega$. Then there are constants $a,b > 0$ such that $$a\mu_u\le\mu_v\le b\mu_u.$$ In particular, $\text{supp}\,\mu_u = \text{supp}\,\mu_v$.
\[lemma1\] Assume that $u\in{\mathcal{F}}^{a}(\Omega)$, $u\ne0$, $v\in\mathcal E(\Omega)$ and that $u \geq v$ on $\text{supp}\,(dd^cu)^n$. Then $u \geq v$ on $\Omega$.
Assume that $u(z_0)<v(z_0)$ for some $z_0\in\Omega$. Let $\psi\in{\mathcal{E}}_0(\Omega)\cap C^{\infty}(\Omega)$ be a strictly plurisubharmonic exhaustion function and let $s>0$ be such that $u(z_0)<s\psi(z_0) + v(z_0).$ Corollary 3.6 in [@Ceg00] gives, with $A= \{u(z)<s\psi(z) + v(z)\}$, $$\int_{A}(dd^c(s\psi + v))^n \leq \int_{A}(dd^cu)^n = 0.$$ Hence $s^n\int_{A}(dd^c\psi)^n =0$ which implies that $A$ has Lebesgue measure $0$. Since the functions involved are plurisubharmonic, this means that $A=\emptyset$. This is a contradiction and the lemma is proved.
Let $K=\text{supp}\,(dd^c u)^n$. Since $K$ is compact, and since $u$ and $v$ are bounded upper semicontinuous functions, $\alpha>0$ may be chosen such that $\alpha v\le u$ on $K$. It then follows from Lemma \[lemma1\] that $\alpha v\le u$ holds on all of $\Omega$. Similarly, there is $\beta>0$ such that $\beta u \le v$ on $\Omega$. Then Proposition \[ineq\_prop\] implies that $\mu_{\alpha^{-1}u}\le \mu_v\le\mu_{\beta u}$. Hence, if we let $a=\alpha^{-n}$ and $b=\beta^n$, the proposition follows.
\[cor\_supp\_S\] There is a set $S\subset{{\partial\Omega}}$ such that $\text{supp}\,\mu_u=S$ for each $u\in{\mathcal{F}}(\Omega)$, $u\ne 0$.
Choose a function $v_0\in{\mathcal{E}}_0(\Omega)$ with $\text{supp}\,(dd^c v_0)^n\subset\subset\Omega$, and let $S=\text{supp}\,\mu_{v_0}$. Let $u$ be an arbitrary function in ${\mathcal{F}}(\Omega)$. Choose a sequence $\{u_j\}\subset{\mathcal{E}}_0(\Omega)$ such that $u_j\searrow u$ and $\text{supp}\,(dd^c u_j)^n\subset\subset\Omega$. Then Proposition \[prop\_compare\] implies that $\text{supp}\,\mu_{u_j}=S$ for each $j$. Moreover, $\mu_{u_1}\le\mu_{u_2}\le\cdots\le\mu_u$ and $\mu_{u_j}$ tends weak\* to $\mu_u$, by Proposition \[ineq\_prop\] and Proposition \[appr\_prop\]. Hence $\text{supp}\,\mu_u= S$.
Note that if $\mu$ is in the the weak\* closure of $\{\mu_u : u\in {\mathcal{F}}(\Omega)\}$, then $\text{supp}\,\mu\subset S$. Hence if $\Omega$ is B-regular, then the support set $S$ has to be all of ${{\partial\Omega}}$, because of Theorem \[th3\].
On the other hand, if $\Omega=\omega_1\times\omega_2\subset\mathbb{C}^n=\mathbb{C}^{n_1+n_2}$, where $\omega_1\subset\mathbb{C}^{n_1}$ and $\omega_2\subset\mathbb{C}^{n_2}$ are bounded hyperconvex domains, then $S\subset\partial\omega_1\times\partial\omega_2$. To see this, consider the function $u(z,w)=\max\,\{g_1(z),g_2(w)\}$ where $g_k$ is the pluricomplex Green’s function for $\omega_k$ with pole at some point in $\omega_k$. Note that $g_k$ is continuous outside the pole and tends to zero at the boundary of $\omega_k$. Then $u\in{\mathcal{F}}(\Omega)$ and $\text{supp}\,(dd^c u)^n\subset\{(z,w)\in\Omega: g_1(z)=g_2(w)\}$. Choose a sequence $\{{\varepsilon}_j\}$ such that ${\varepsilon}_j\searrow 0$. Then $\Omega_j=\{(z,w)\in\Omega:u(z,w)<-{\varepsilon}_j\}$ defines a fundamental sequence of $\Omega$ and $u^j:=\sup\,\{\varphi\in PSH(\Omega):
\varphi|_{\Omega\setminus\Omega_j}\le u|_{\Omega\setminus\Omega_j}\}=
\max\,\{u,-{\varepsilon}_j\}$. It follows that $\text{supp}\,(dd^c u^j)^n\subset
\{(z,w)\in\Omega: g_1(z)=g_2(z)\ge-{\varepsilon}_j\}$, which implies that $\text{supp}\,\mu_u\subset\partial\omega_1\times\partial\omega_2$. Hence the claim follows from Corollary \[cor\_supp\_S\].
Using a similar argument, the following example shows that when $\Omega=\mathbb{D}\times\mathbb{D}\subset\mathbb{C}^2$, then we have equality, $S=\partial\mathbb{D}\times\partial\mathbb{D}$.
\[ex\_bidisc\] Let $\Omega$ be the unit bidisc $\mathbb{D}\times\mathbb{D}$ in $\mathbb{C}^2$. Then $\text{supp}\,\mu_u$ is equal to the distinguished boundary $\partial\mathbb{D}\times\partial\mathbb{D}$ for each $u\in{\mathcal{F}}(\Omega)$, $u\ne 0$. This follows from Corollary \[cor\_supp\_S\], if we for example consider the pluricomplex Green’s function $g$ for $\Omega$ with pole at the origin. We then have that $g(z,w)=m\cdot\max\,\{\log|z|,\log|w|\}$, where the constant $m>0$ is chosen such that $\int_\Omega(dd^c g)^n = 1$. This is a function in ${\mathcal{F}}(\Omega)$, and we can compute $\mu_g$ explicitly. For $j=1,2,\ldots$, let $\Omega_j=\{(z,w):|z|<r_j, |w|<r_j\}$ where $r_j=1-\frac{1}{j}$. Then $g^j:= \sup\,\{\varphi\in PSH(\Omega):
\varphi|_{\Omega\setminus\Omega_j}\le
g|_{\Omega\setminus\Omega_j}\}=
m\cdot\max\,\{\log|z|,\log|w|, \log(r_j)\}$, from which it follows that $(dd^c g^j)^2 =
m^2\cdot dd^c(\max\,\{\log|z|,\log(r_j)\})\wedge
dd^c(\max\,\{\log|w|,\log(r_j)\})$. Since $\int_\Omega(dd^c g^j)^2=1$ for each $j$ (see Section \[sec\_constr\]), we can conclude that $(dd^c g^j)^2=\sigma_j\times\sigma_j$, where $\sigma_j$ is the normalized Lebesgue measure on the the circle $\partial \mathbb{D}(0,r_j)$. This implies that $\mu_g=\sigma\times\sigma$, where $\sigma$ is the normalized Lesbegue measure on the unit circle.
\[rem\_supp\] Recall from Remark \[rem\_pi\] at the end of Section \[sec\_constr\] that Corollary \[cor\_supp\_S\] and (\[pi\_formula\]) together give the partial integration formula $$\label{pi_impl}
h|_S=0\ \Rightarrow\
\int_\Omega h\,(dd^c u)^n=
\int_\Omega u\,dd^c h\wedge (dd^c u)^{n-1}.$$ The implication (\[pi\_impl\]) holds true for $h\in PSH(\Omega)\cap C(\bar\Omega)$ if $u\in{\mathcal{F}}(\Omega)$, and for $h\in PSH(W)\cap L^\infty(W)$, $W\supset\bar\Omega$, if $u\in{\mathcal{F}}^{a}(\Omega)$ (using Theorem \[th\_equal\] of Section \[sec\_bdryval\]). Here $S$ is the support set defined in Corollary \[cor\_supp\_S\].
Furthermore, (\[J\_ineq\]) implies that $$\label{S_ineq}
\sup_\Omega\,h \le \sup_S\,h,
\quad\forall\,h\in PSH^-(\Omega)\cap C(\bar\Omega).$$ To see this, let $h\in PSH^-(\Omega)\cap C(\bar\Omega)$ be given. For $z\in\Omega$ fixed, let $g_z$ be the pluricomplex Green’s function for $\Omega$ with pole at $z$. Then $(dd^c g_z)^n=\delta_z$ and we have that $h(z) = \int_\Omega h\,(dd^c g_z)^n\le
\int_{{\partial\Omega}}h\,d\mu_{g_z}\le \sup_S\,h$. By the same argument, (\[S\_ineq\]) holds true if $h$ is an upper bounded function in $PSH(W)$, where $W\supset\bar\Omega$.
\[rem\_Henkin\] Another property of the measures $\mu_u$ is that they are so called *Henkin measures* (a kind of measure introduced by Henkin in [@Hen]). This means that $$\lim_{k\to\infty}\int_{{\partial\Omega}}f_k\,d\mu_u=0$$ for each uniformly bounded sequence $\{f_k\}$ in $A(\Omega)$ such that $\lim_{k\to\infty} f_k(z)=0$ for all $z\in\Omega$. Here $A(\Omega)$ denotes the functions that are holomorphic on $\Omega$ and continuous on $\bar\Omega$. To see that this holds, take such a sequence $\{f_k\}$ and let $\{\varphi_k\}=\{\text{Re}\,f_k\}$. From (\[limit\_1\]) and (\[eq\_1\]) it follows that $$\lim_{k\to\infty}\int_{{\partial\Omega}}\varphi_k\,d\mu_u =
\lim_{k\to\infty}\left(\lim_{j\to\infty} \int_\Omega\varphi_k\,(dd^c u^j)^n\right) =
\lim_{k\to\infty}\int_\Omega\varphi_k\,(dd^c u)^n = 0$$ for each $u\in{\mathcal{F}}(\Omega)$, since $\varphi_k$ is uniformly bounded and $\int_\Omega (dd^c u)^n<\infty$. Since the same holds for $\{\psi_k\}=\{\text{Im}\,f_k\}$, it follows that $\lim_{k\to\infty}\int_{{\partial\Omega}}f_k\,d\mu_u=0$.
This property can be used to show the following fact about the support of the measures $\mu_u$. Suppose that $u\in{\mathcal{F}}(\Omega)$ and that $K\subset{{\partial\Omega}}$ is a peak set for $A(\Omega)$. Let $f\in A(\Omega)$ be a peak function for $K$ and define $f_k(z) = (f(z))^k$, for $z\in\bar\Omega$ and $k=1,2,\ldots$. Then $\{f_k\}$ satisfies the assumptions above, so $\lim_{k\to\infty} \int_{{\partial\Omega}}f_k\,d\mu_u = 0$. But we also have that $\lim_{k\to\infty} \int_{{\partial\Omega}}f_k\,d\mu_u = \mu_u(K)$. Hence $\mu_u(K)=0$ for each peak set $K$ and each $u\in{\mathcal{F}}(\Omega)$.
Boundary values {#sec_bdryval}
===============
In this section we define and study boundary values of plurisubharmonic funtions, with respect to the measures $\mu_u$.
\[bdry\_lemma\] Assume that $u\in{\mathcal{F}}(\Omega)$ and $g\in PSH(\Omega)\cap L^\infty(\Omega)$. Then $\{g\,(dd^c u^j)^n\}$ is weak\*-convergent.
By the same argument as in Theorem \[th1\] it is enough to prove that the limit $\lim_{j\to\infty}
\int_\Omega \varphi g\,(dd^c u^j)^n$ exists for all $\varphi\in PSH^-(\Omega)\cap L^\infty(\Omega)$. Given such a function $\varphi$, take $M,N\ge 0$ such that $\varphi+M\ge0$ and $g+N\ge 0$. Then $(\varphi+M)^2$, $(g+N)^2$, $(\varphi+M+g+N)^2\in
PSH(\Omega)\cap L^\infty(\Omega)$, so if $\psi$ is any of these then $\lim_{j\to\infty} \int_\Omega \psi\,(dd^c u^j)^n$ exists by Theorem \[th1\]. Expanding $((\varphi+M)+(g+N))^2$, it follows that the limit exists for $\psi=(\varphi+M)(g+N)$ and then finally for $\psi=\varphi g$ (using Theorem \[th1\] again).
Using this lemma, together with standard measure theory, we can make the following definition.
\[b\_val\_def\] For $u\in{\mathcal{F}}(\Omega)$ and $g\in PSH(\Omega)\cap L^\infty(\Omega)$, let $g^u$ be the function in $L^\infty({{\partial\Omega}},\mu_u)$ such that $\lim_{j\to\infty} g\,(dd^c u^j)^n = g^u\,d\mu_u$.
We may consider $g^u$ as the boundary values of $g$ with respect to $\mu_u$. Note that, at least formally, $g^u$ depends on both $g$ and $u$. However, the following theorems describe some situations when this definition agrees with other notions of boundary values.
\[th\_equal\] Assume that $u\in{\mathcal{F}}^{a}(\Omega)$ and $g\in PSH(W)\cap L^\infty(W)$ where $W$ is a bounded domain containing $\bar\Omega$. Then $g^u=g|_{{\partial\Omega}}$ a.e. $(\mu_u)$.
Note that if $M$ is a constant then $(g-M)^{u}=g^{u}-M$, so we may assume that $g\le 0$. Let $t\in C(\bar\Omega)$, $t\ge 0$ be given. Then it follows, in the same way as in the proof of Lemma \[tec\_lemma\], that $$\int_{{\partial\Omega}}tg^u\,d\mu_u =
\lim_{j\to\infty}\int_\Omega tg\,(dd^c u^j)^n \le
\int_{{\partial\Omega}}tg\,d\mu_u.$$ Thus $g^u\le g$ a.e. $(\mu_u)$, so it remains to prove that $\int_{{\partial\Omega}}g^u\,d\mu_u = \int_{{\partial\Omega}}g\,d\mu_u$. Choose $K$ such that $\Omega\subset\subset K \subset\subset W$. Given ${\varepsilon}> 0$ there is an open set $U_{\varepsilon}\subset W$ and a function $g_{\varepsilon}\in C_0(W)$ such that $\inf_W g\le g_{\varepsilon}\le 0$, the relative capacity $cap\,(U_{\varepsilon}, W) < {\varepsilon}$ and $K\setminus U_{\varepsilon}\subset\{z\in W: g(z) = g_{\varepsilon}(z)\}$ (for definition and properties of relative capacity, see ). It follows that $$\begin{split}
\int_{{\partial\Omega}}g^u\,d\mu_u & =
\lim_{j\to\infty}\int_\Omega g\,(dd^c u^j)^n =
\\
& = \lim_{j\to\infty}\int_{\Omega\cap U_{\varepsilon}} g\,(dd^c u^j)^n +
\lim_{j\to\infty}\int_{\Omega\setminus U_{\varepsilon}}
g_{\varepsilon}\,(dd^c u^j)^n\ge\\
& \ge \lim_{j\to\infty}\int_{\Omega\cap U_{\varepsilon}} g\,(dd^c u^j)^n +
\int_{{\partial\Omega}}g_{\varepsilon}\,d\mu_u =\\
& = \lim_{j\to\infty}\int_{\Omega\cap U_{\varepsilon}} g\,(dd^c u^j)^n +
\int_{{{\partial\Omega}}\cap U_{\varepsilon}} g_{\varepsilon}\,d\mu_u +
\int_{{{\partial\Omega}}\setminus U_{\varepsilon}} g\,d\mu_u\ge\\
& \ge \lim_{j\to\infty}\int_{\Omega\cap U_{\varepsilon}} g\,(dd^c u^j)^n +
\int_{{{\partial\Omega}}\cap U_{\varepsilon}} g_{\varepsilon}\,d\mu_u +
\int_{{{\partial\Omega}}} g\,d\mu_u.
\end{split}$$ Let $h_{\varepsilon}=\sup\,\{\psi\in PSH^-(W): \psi|_{U_{\varepsilon}}\le-1\}$, we then have that $$\begin{split}
0&\ge \int_{{\partial\Omega}}g^u\,d\mu_u - \int_{{{\partial\Omega}}} g\,d\mu_u\ge\\
& \le
\lim_{j\to\infty}\int_{\Omega\cap U_{\varepsilon}} g\,(dd^c u^j)^n +
\int_{{{\partial\Omega}}\cap U_{\varepsilon}} g_{\varepsilon}\,d\mu_u\ge\\
& \ge\left(\inf_W g\right)
\left(\lim_{j\to\infty}\int_{\Omega\cap U_{\varepsilon}}(dd^c u^j)^n
+ \int_{{{\partial\Omega}}\cap U_{\varepsilon}}d\mu_u\right)=\\
& =\left(-\inf_W g\right)
\left(\lim_{j\to\infty}\int_{\Omega\cap U_{\varepsilon}}h_{\varepsilon}\,(dd^c u^j)^n
+ \int_{{{\partial\Omega}}\cap U_{\varepsilon}}h_{\varepsilon}\,d\mu_u\right)\ge\\
& \ge\left(-\inf_W g\right)
\left(\lim_{j\to\infty}\int_{\Omega}h_{\varepsilon}\,(dd^c u^j)^n
+ \int_{{{\partial\Omega}}}h_{\varepsilon}\,d\mu_u\right)\ge\\
&\ge 2\left(-\inf_W g\right)
\int_\Omega h_{\varepsilon}\,(dd^c u)^n,
\end{split}$$ where we have used (\[ineq\_1\]) and Lemma \[tec\_lemma\] in the last inequality. From Lemma 1.9 in [@CegKol], using that $u\in{\mathcal{F}}^{a}(\Omega)$ and that $cap\,(U_{\varepsilon}, W) < {\varepsilon}$, it follows that this last integral tends to zero as ${\varepsilon}\searrow 0$, which completes the proof.
The following theorem may be compared with the definitions in Section \[sec\_prel\].
\[th2\] Suppose that $H\in {\mathcal{M}}(\Omega)\cap L^\infty(\Omega)$. Then, for every $u\in{\mathcal{F}}^{a}(\Omega)$ and every $g\in
{\mathcal{F}}(\Omega,H)$ such that $\int_\Omega g\,(dd^cu)^n > -\infty$, $g\,(dd^cu^j)^n$ is weak\*-convergent to $H^{u}\,d\mu_u$.
By the same argument as in Theorem \[th1\], it is enough to prove that $$\lim_{j\to\infty} \int_\Omega tg\,(dd^c u^j)^n =
\lim_{j\to\infty} \int_\Omega t H (dd^c u^j)^n,
\quad\forall\,t\in PSH^-(\Omega)\cap L^\infty(\Omega).$$ Since $g\in {\mathcal{F}}(\Omega,H)$ there is a $\psi\in{\mathcal{F}}(\Omega)$ such that $\psi + H \le g\le H$. We may assume that $\psi\ge g$ (otherwise, look at $\psi_0=\max\,\{\psi,g\}$). We may also (after dividing by suitable constants) assume that $-1\le t\le 0$ and $-1\le H\le 0$. Now, $$\int_\Omega tg\,(dd^c u^j)^n =
\int_\Omega t(g-H)\,(dd^c u^j)^n +
\int_\Omega t H (dd^c u^j)^n$$ where $0\le \int_\Omega t(g-H)\,(dd^c u^j)^n =
\int_\Omega (-t)(H-g)\,(dd^c u^j)^n \le
\int_\Omega (-t)(-\psi)\,(dd^c u^j)^n \le
\int_\Omega (-\psi)\,(dd^c u^j)^n$. Using partial integration in ${\mathcal{F}}(\Omega)$ we have the following $$\begin{split}
\int_\Omega (-\psi)\,(dd^c u^j)^n & =
\int_\Omega (-u^j)\,dd^c\psi\wedge(dd^c u^j)^{n-1} \le
\int_\Omega (-u)\,dd^c\psi\wedge(dd^c u^j)^{n-1}=\\
& = \int_\Omega (-u^j)\,dd^c\psi\wedge dd^c u
\wedge(dd^c u^j)^{n-2} \le \ldots\le\\
&\le\int_\Omega (-u^j)\,dd^c\psi\wedge(dd^c u)^{n-1} =
I_j\le
\int_\Omega (-u)\,dd^c\psi\wedge(dd^c u)^{n-1}=\\
& =\int_\Omega (-\psi)\,(dd^c u)^n \le
\int_\Omega (-g)\,(dd^c u)^n <+\infty.
\end{split}$$ Since $u^j$ increases to zero outside a pluripolar set and $dd^c\psi\wedge(dd^c u)^{n-1}$ vanishes on pluripolar sets (see Section \[sec\_prel\], Lemma \[l\_pp\]), it follows that $I_j\searrow 0$ when $j\to+\infty$. This proves the theorem.
\[rem\_th2\] If $g\in L^\infty(\Omega)$ then $\int_\Omega g\,(dd^cu)^n > -\infty$ for every $u\in{\mathcal{F}}(\Omega)$. Furthermore, $\psi\ge g$ implies that $\psi$ is bounded as well, so $dd^c\psi\wedge(dd^c u)^{n-1}$ vanishes on pluripolar sets for every $u\in{\mathcal{F}}(\Omega)$ (Lemma \[l\_pp\]). Thus for bounded functions $g$ in ${\mathcal{F}}(\Omega,H)$, the conclusion $g^{u}\,d\mu_u = H^{u}\,d\mu_u$ holds for every $u\in{\mathcal{F}}(\Omega)$.
Suppose that we have a bounded plurisubharmonic function on $\Omega$ and want to approximate it with plurisubharmonic functions that are continuous on $\bar\Omega$. The following theorem gives a condition for when this implies weak\*-convergence on the boundary.
Assume that $u\in {\mathcal{F}}(\Omega)$ and $\mu_u = \lim_{j\to\infty}(dd^cu^{j})^{n}$. Let $\{\varphi_j\}$ be a sequence in $PSH(\Omega)\cap
C(\bar\Omega)$ such that $0\leq \varphi_j\leq 1$. If $\varphi_j$ tends to $\varphi\in PSH(\Omega)\cap L^\infty(\Omega)$ in the sense of distributions, then $\varphi_j\,d\mu_u $ tends weak\* to $\varphi^{u}\, d\mu_u$ if and only if $\lim_{j\to\infty}\int \varphi_j\,d\mu_u = \int\varphi^{u}\,d\mu_u$.
By Corollary \[cor\_max\] we may assume that $u\in{\mathcal{F}}^{a}(\Omega)$. The condition in the theorem is obviously necessary, we prove it is also sufficient. First, note that for $\{\psi_k\}\subset PSH(\Omega)\cap
C(\bar\Omega)$, $\psi_k\ge 0$, the following holds. For $k$ fixed, $(\sup_{l\ge k}\psi_l)^*\in PSH(\Omega)\cap L^\infty(\Omega)$, therefore $(\sup_{l\ge k}\psi_l)^*\,(dd^c u^j)^n$ is weak\*-convergent (as $j\to\infty$) by Lemma \[bdry\_lemma\]. Furthermore, since $(\sup_{l\ge k}\psi_l)=(\sup_{l\ge k}\psi_l)^*$ outside a pluripolar set and $u^j\in{\mathcal{F}}^{a}(\Omega)$ (since $u\in{\mathcal{F}}^{a}(\Omega)$), the star may be removed. We claim that $$\label{part_i}
\lim_{j\to\infty}\: (\sup_{l\ge k}\psi_l)\,(dd^c u^j)^n=
(\sup_{l\ge k}\psi_l)\,d\mu_u.$$ Given $f\in C(\bar\Omega)$, $f\ge0$ it follows from (\[limit\_1\]) that for each $m$ $$\lim_{j\to\infty}\int_\Omega f(\sup_{l\ge k}\psi_l)\,(dd^c u^j)^n \ge
\lim_{j\to\infty}\int_\Omega f(\sup_{m\ge l\ge k}\psi_l)\,(dd^c u^j)^n=
\int_{{\partial\Omega}}f(\sup_{m\ge l\ge k}\psi_l)\,d\mu_u,$$ where the last integral tends to $\int_{{\partial\Omega}}f(\sup_{l\ge k}\psi_l)\,d\mu_u$ as $m\to\infty$. It follows that $\lim_{j\to\infty}\: (\sup_{l\ge k}\psi_l)\,(dd^c u^j)^n\ge
(\sup_{l\ge k}\psi_l)\,d\mu_u$. On the other hand, by (\[ineq\_1\]) and (\[limit\_1\]) $$\int_\Omega (\sup_{m\ge l\ge k}\psi_l)\,(dd^c u^j)^n \le
\int_{{\partial\Omega}}(\sup_{m\ge l\ge k}\psi_l)\,d\mu_u$$ for each $m$ and $j$. So by letting $m\to\infty$ we have that $\int_\Omega (\sup_{l\ge k}\psi_l)\,(dd^c u^j)^n \le
\int_{{\partial\Omega}}(\sup_{ l\ge k}\psi_l)\,d\mu_u$, which proves the claim.
Now, let $\{\varphi_{j_m}\,d\mu_u\}$ be any weak\*-convergent subsequence of $\{\varphi_j\,d\mu_u\}$. (Such a sequence exists by the same reasoning as in the proof of Theorem \[th3\].) Then, by standard measure theory, the limit measure is equal to $\varphi_0\,d\mu_u$ for some $\varphi_0\in L^\infty(\mu)$. We will show that $\varphi_0=\varphi^u$ a.e. $(\mu)$. It then follows that the original sequence itself converges to $\varphi^{u}\,d\mu_u$, and the proof will be complete.
From $L^2$-theory it follows that we may choose $\psi_k=\frac{1}{M_k}\sum_{l=1}^{M_k}\varphi_{j_{m_l}}$ such that $\psi_k\to\varphi_0$ in $L^2(\mu)$ and then a subsequence converging to $\varphi_0$ a.e. $(\mu)$, for simplicity call it $\{\psi_k\}$. Since by assumption the original sequence $\{\varphi_j\}$ tends to $\varphi$ in the sense of distributions, the same holds for $\{\psi_k\}$. Now, for $f\in C(\bar\Omega)$, $f\ge 0$, using the definition of $\varphi^u$, (\[part\_i\]) and monotone convergence, $$\begin{split}
\int_{{\partial\Omega}}f\varphi^u\,d\mu_u & =
\lim_{j\to\infty}\int_\Omega f\varphi\,(dd^c u^j)^n =
\text{(Lemma 1.4 in \cite{CegKol})} =\\
& = \lim_{j\to\infty}\lim_{k\to\infty} \int_\Omega f\psi_k\,(dd^c u^j)^n \le
\lim_{k\to\infty}\lim_{j\to\infty} \int_\Omega f(\sup_{l\ge k}\psi_l)\,
(dd^c u^j)^n =\\
& = \lim_{k\to\infty}\int_{{\partial\Omega}}f(\sup_{l\ge k}\psi_l)\,d\mu_u =
\int_{{\partial\Omega}}f (\limsup_{k\to\infty}\psi_k)\,d\mu_u.
\end{split}$$ From this it follows that $\varphi^u\le\limsup_{k\to\infty}\psi_k$ a.e. $(\mu)$, which implies that $\varphi^u\le\varphi_0$ a.e. $(\mu)$. Furthermore, $\int_{{\partial\Omega}}\varphi_0\,d\mu_u =
\lim_{m\to\infty} \int_{{\partial\Omega}}\varphi_{j_m}\,d\mu_u =
\int_{{\partial\Omega}}\varphi^u\,d\mu_u$, by assumption, so $\varphi^u=\varphi_0$ a.e. $(\mu)$. Hence the theorem is proved.
More boundary measures {#sec_more}
======================
Let $\nu$ be a positive measure on $\Omega$ with finite total mass. Then there is a positive measure $\mu\ne 0$ which is supported by ${{\partial\Omega}}$, vanishes on pluripolar sets and such that$$\label{mes_ineq}
\int_\Omega\varphi\,d\nu \leq \int_{{\partial\Omega}}\varphi\,d\mu,
\quad\forall\,\varphi\in PSH^-(\bar\Omega),$$ where $PSH^-(\bar\Omega)
=\{\varphi: \varphi\in PSH^-(\Omega'),\, \Omega'\supset\bar\Omega\}$. To see this, let $P_\nu$ denote the pluricomplex potential of $\nu$ relative to $\Omega$, i.e. $P_\nu(z)=\int_\Omega g(z,w)\,d\nu(w)$, where $g(z,w)$ is the pluricomplex Green’s function for $\Omega$ with pole at $w$. Then Theorem 1.1 in [@Ceg4] says that $P_\nu\in{\mathcal{F}}(\Omega)$ and that $$\int_\Omega-\varphi\,(dd^c P_\nu)^n\le
\left(\nu(\Omega)\right)^{n-1}\int_\Omega -\varphi\,d\nu,
\quad\forall\,\varphi\in PSH^-(\Omega).$$ Moreover, $\int_\Omega\varphi\,(dd^c P_\nu)^n\le
\int_{{\partial\Omega}}\varphi\,d\mu_{P_\nu}$ for each $\varphi\in PSH^-(\bar\Omega)$, by Remark \[rem\_pi\] at the end of Section \[sec\_constr\]. Hence, the claim follows if we take $\mu = \left(\nu(\Omega)\right)^{-n+1}\mu_{P_\nu}$.
Conversely, if a positive measure $\mu$ on ${{\partial\Omega}}$ is such that (\[mes\_ineq\]) holds for some finite measure $\nu$ on $\Omega$, we would like to find an approximation procedure, similar to the one in Section \[sec\_constr\]. A motivation is that we are interested in boundary values of plurisubharmonic functions with respect to $\mu.$
We will study the case when $\nu$ vanishes on all pluripolar subsets of $\Omega$ and $\Omega$ belongs to a more restricted class of hyperconvex domains:
1. \[cond\_a\] $\Omega$ and $\{\Omega_k\}$ are hyperconvex domains with $\Omega\subset\subset\Omega_{k+1}\subset\subset\Omega_k$, such that for each $t\in{\mathcal{F}}(\Omega)$ there is a sequence $\{t_k\}$, where $t_k\in{\mathcal{F}}(\Omega_k)$ and $t_k\nearrow t$ a.e. on $\Omega$.
2. \[cond\_b\] $\Omega$ is not thin at any of its boundary points, so that $\limsup_{\Omega\ni z\to\xi}v(z)=v(\xi)$ for each $\xi\in{{\partial\Omega}}$ if $v\in PSH^-(\bar\Omega)$.
Conditions for the approximation property in (\[sec\_more\]\[cond\_a\]) to hold true have been studied in for example [@Benel] and [@CegHed]. Examples of domains satisfying (\[sec\_more\]\[cond\_a\]) and (\[sec\_more\]\[cond\_b\]) are polydiscs and strictly pseudoconvex domains. Note that if $t$ is bounded, we may assume that each $t_k$ is bounded.
\[th\_more\] Let $\Omega$ be a domain satisfying (\[sec\_more\]\[cond\_a\]) and (\[sec\_more\]\[cond\_b\]). Assume that $\mu$ is a positive measure on $\partial\Omega$, vanishing on pluripolar sets. Then there is a sequence $\{w_k\}$ in ${\mathcal{F}}^{a}(\bar\Omega) =
\{u: u\in {\mathcal{F}}^{a}(\Omega'),\, \Omega'\supset\bar\Omega\}$ such that $\text{supp}\,(dd^c w_k)^n\subset\subset\Omega$, $\int_\Omega (dd^c w_k)^n\le \int_{{\partial\Omega}}d\mu$, and $(dd^cw_k)^n$ tends weak\* to $\mu$ as $ k\to \infty$.
Furthermore, if there is a finite positive measure $\nu$ on $\Omega$, vanishing on pluripolar sets, such that (\[mes\_ineq\]) holds, then $\lim_{k\to\infty}\int_\Omega t\,(dd^cw_k)^n = 0$ for each $t\in{\mathcal{F}}(\Omega)\cap L^\infty(\Omega)$. Hence $t\,(dd^cw_k)^n$ tends weak\* to $0$ for each $t\in{\mathcal{F}}(\Omega)\cap L^\infty(\Omega)$.
If we compare this theorem with the results in the previous sections we have the following. In the setting of Section \[sec\_constr\] we know that if $u\in{\mathcal{F}}^{a}(\Omega)$ and $\varphi \in PSH(W)\cap L^\infty(W)$, $W\supset\bar\Omega$, then $\int_\Omega\varphi\,(dd^cu)^n\leq \int_\Omega\varphi\,(dd^cu^j)^n$ which increases to $\int_{{\partial\Omega}}\varphi\, d\mu_u$ (see Theorem \[th\_equal\]). In particular it follows that when $\mu=\mu_u$ for some $u\in{\mathcal{F}}^{a}(\Omega)$, then (\[mes\_ineq\]) is satisfied if we take $\nu=(dd^c u)^n$. We also have that $\int_\Omega (dd^c u^j)^n=\int_{{\partial\Omega}}d\mu_u$ and $\lim_{j\to\infty} t\,(dd^c u^j)^n=0$ for each $t\in{\mathcal{F}}(\Omega)\cap L^\infty(\Omega)$ (see Remark \[rem\_th2\]). Hence, the approximation procedure in Theorem \[th\_more\] is similar to the one in the previous sections, and it applies to a larger class of boundary measures, see also Example \[ex\_more\].
\[lemma2\] Let ${\{\mu^{j}_k\}}_{j,k}$ be a sequence of positive measures on $\bar\Omega$ with uniformly bounded mass. Suppose that, for each fixed $k$, $\mu^{j}_k$ tends weak\* to $\mu$ as $j\to\infty$. Then there is a subsequence ${\{\mu^{j_k}_k\}}_k$ such that $\mu^{j_k}_k$ tends weak\* to $\mu$ as $k\to \infty$.
Let $\{t_l\}$ be a dense sequence in $C(\bar\Omega)$. For each $k$ we choose $j_k$ such that $$\left|\int_{\bar\Omega} t_l\,d\mu -
\int_{\bar\Omega} t_l\,d\mu^{j_k}_k\right| <
\frac{1}{k},\quad 1\leq l \leq k.$$ It follows that $\mu^{j_k}_k$ tends weak\* to $\mu$ as $k\to \infty$, since $\{t_l\}$ is dense and the measures have uniformly bounded total mass.
For each $k$, the measure $\mu$ can be regarded as a finite measure on $\Omega_k$ which vanishes on pluripolar sets. Hence there is $u_k\in{\mathcal{F}}^{a}(\Omega_k)$ such that $(dd^cu_k)^n = \mu$, see Lemma 5.14 in [@Ceg2]. Choose a fundamental sequence $\{\omega_j\}$ of $\Omega$, i.e. $\omega_j\subset\subset \omega_{j+1}\subset\subset\Omega$ and $\cup_{j=1}^\infty\omega_j=\Omega$. For each $k$ and $j$, define $u_k^j = \sup\,\{\varphi\in PSH^-(\Omega_k):
\varphi|_{\omega_j}\leq u_k|_{\omega_j}\}$. Then $u_k^j\in{\mathcal{F}}^{a}(\Omega_k)$ (note that $(u_k^j)^*=u_k^j$ since $\omega_j$ is open, so $u_k^j$ is plurisubharmonic) and we have the following:
(i) \[stat\_1\] $\text{supp}\,(dd^cu_k^j)^n \subset \partial\omega_j$, $u_k^j \geq u_k$ on $\Omega_k$, $\int_{\Omega_k}(dd^cu_k^j)^n \leq
\int_{\Omega_k}(dd^cu_k)^n =\int_{{\partial\Omega}}d\mu$.
(ii) \[stat\_2\] If $j_1\leq j_2$ then $u_k^{j_1} \geq u_k^{j_2}$ on $\Omega_k$.
(iii) \[stat\_3\] $\lim_{j \to \infty}u_k^j = u_k$ on $\Omega_k$.
The first two statements are obvious. For the proof of the third, let $v_k=\lim_{j}u_k^j$. Then $v_k\in{\mathcal{F}}(\Omega_k)$, $v_k\ge u_k$ on $\Omega_k$ and $v_k=u_k$ on $\Omega$. Thus $v_k(\xi)=u_k(\xi)$ for $\xi\in{{\partial\Omega}}$, using the assumption (\[sec\_more\]\[cond\_b\]), so $v_k\le u_k$ on $\Omega_k$ by Lemma \[lemma1\] and the statement follows. Now, (\[stat\_2\]) and (\[stat\_3\]) imply that $(dd^c u_k^j)^n$ tends weak\* to $(dd^c u_k)^n=\mu$ as $j\to\infty$, for each fixed $k$. Hence, by (\[stat\_1\]) we can use Lemma \[lemma2\] to pick $\{j_k\}$ such that $(dd^cu_k^{j_k})^n$ tends weak\* to $\mu$ as $k\to\infty$. This completes the first part of the theorem, if we let $w_k = u_k^{j_k}$.
It remains to prove that $\lim_{k\to \infty}\int
t\,(dd^cu_k^{j_k})^n = 0$ for all $t\in{\mathcal{F}}(\Omega)\cap L^\infty(\Omega)$, assuming that (\[mes\_ineq\]) holds. Given $t\in{\mathcal{F}}(\Omega)\cap L^\infty(\Omega)$ there is by (\[sec\_more\]\[cond\_a\]) a sequence $\{t_k\}$ with $t_k\in{\mathcal{F}}(\Omega_k)\cap L^\infty(\Omega_k)$ such that $t_k$ increases a.e. to $t$ on $\Omega$. Now, $$\int_{\Omega_k} t\,(dd^cu_k^{j_k})^n\geq
\int_{\Omega_k} t_k\,(dd^cu_k^{j_k})^n \geq
\int_{\Omega_k} t_k\,(dd^cu_k)^n =
\int_{{\partial\Omega}}t_k\, d\mu \geq
\int_\Omega t_k\,d\nu>-\infty$$ so it follows that $$\liminf_{k\to \infty}
\int_{\Omega_k} t\,(dd^cu_k^{j_k})^n
\geq \int_\Omega t\,d\nu.$$ Define $t^{i}= \sup\,\{\varphi\in PSH(\Omega):
\varphi|_{\Omega\setminus\omega_i}\le
t|_{\Omega\setminus\omega_i}\}$. Then $t^{i}\in{\mathcal{F}}(\Omega)\cap L^\infty(\Omega)$ and $t^{i}=t$ on $\Omega\setminus\omega_i$, so $$\liminf_{k\to \infty}\int_{\Omega_k} t\,(dd^cu_k^{j_k})^n =
\liminf_{k\to \infty}\int_{\Omega_k} t^{i}\,(dd^cu_k^{j_k})^n\geq
\int_\Omega t^{i}\,d\nu,$$ by the above calculations. Now, the left hand side is independent of $i$, while the right hand side tends to $0$ when $i$ tends to $\infty$, since $\nu$ vanishes on pluripolar sets. This completes the proof.
The reason not to keep $k$ fixed in the proof above, is to be able to prove the second part of the theorem. Also, one can prove that $\lim_{k\to\infty} u_k^j=0$ a.e. on $\Omega$, for each fixed $j$.
Suppose that $v\in PSH^-(\Omega)$ satisfies $\tilde v \ge v \ge \tilde v + \psi$ for some $\psi\in{\mathcal{F}}(\Omega)\cap L^\infty(\Omega)$ and that $\tilde v\in C(\bar\Omega)$. (Thus, $v$ is a function in ${\mathcal{F}}(\Omega, \tilde v)$ with some additional properties, see Section \[sec\_prel\].) Then the preceeding theorem implies that $$\label{limit_6}
\lim_{k\to\infty} v\,(dd^c u_k^{j_k})^n = \tilde v\,d\mu,$$ where the limit is in weak\* sense. To see this, take $f\in C(\bar\Omega)$, $f\ge 0$. Then by the theorem we have that $\lim_{k\to\infty}\int_\Omega f\tilde v\,(dd^c u_k^{j_k})^n =
\int_{{\partial\Omega}}f\tilde v\,d\mu$ and that $0\ge\int_\Omega f\psi\,(dd^c u_k^{j_k})^n\ge
{\max f}\cdot\int_\Omega\psi\,(dd^c u_k^{j_k})^n$, where the last integral tends to $0$ as $k\to\infty$. Hence the inequality $f\tilde v \ge fv \ge f\tilde v + f\psi$ implies that $\lim_{k\to\infty}\int_\Omega fv\,(dd^c u_k^{j_k})^n =
\int_{{\partial\Omega}}f\tilde v\,d\mu$, and (\[limit\_6\]) follows.
Furthermore, if we assume that $\int_\Omega \varphi\,d\nu > -\infty$ for all $\varphi\in{\mathcal{F}}(\bar\Omega)$, then (\[limit\_6\]) holds for all $v\in{\mathcal{F}}(\Omega,\tilde v)$ where $\tilde v\in C(\bar\Omega)$. This is due to the fact that the boundedness of $t$ in the second part of Theorem \[th\_more\] is used only to ensure that $\int_\Omega t_k\,d\nu>-\infty$ (because if $t$ is bounded then $t_k$ is bounded). Hence the assumption that $t\in{\mathcal{F}}(\Omega)\cap L^\infty(\Omega)$ can be replaced by the assumption that $t\in{\mathcal{F}}(\Omega)$ and $\int_\Omega \varphi\,d\nu > -\infty$ for all $\varphi\in{\mathcal{F}}(\bar\Omega)$.
\[ex\_more\] Let $\Omega$ be the unit bidisc $\mathbb{D}\times\mathbb{D}$ in $\mathbb{C}^2$. Let $\mu$ and $\nu$ be defined by $$\mu=\sigma_1\times dV_\frac{1}{2}\ \ \text{and}\ \
\nu=\sigma_\frac{1}{2}\times dV_\frac{1}{2},$$ where $\sigma_r$ denotes the normalized Lebesgue measure on the circle $\partial\mathbb{D}(0,r)$ and $dV_\frac{1}{2}$ the normalized Lebesgue measure on the disc $\mathbb{D}(0,\frac{1}{2})$. Then $\mu$ and $\nu$ satisfies (\[mes\_ineq\]), so Theorem \[th\_more\] tells us that we can approximate $\mu$ from the inside of $\Omega$ by our procedure. Moreover, by Example \[ex\_bidisc\] we see that $\mu$ is not in the weak\* closure of $\{\mu_u: u\in{\mathcal{F}}(\Omega)\}$. Hence, we do reach more measures by the method in this section than we could before.
[99999]{}
hag, P., Cegrell, U., Czyż, R. and Phm, H. H., *Monge-Ampère measures on pluripolar sets,* manuscript Ume[å]{} University 2008.
Bedford, E. and Taylor, B. A., *A new capacity for plurisubharmonic functions,* Acta Math. 149:1-2 (1982), 1–40.
Benelkourchi, S., *A note on the approximation of plurisubharmonic functions,* C. R. Math. Acad. Sci. Paris 342:9 (2006) 647–650.
B[ł]{}ocki, Z., *On the definition of the complex Monge-Ampère operator in $\mathbb{C}^2$,* Math. Ann. 328:3 (2004), 415–423.
Cegrell, U., *Pluricomplex energy,* Acta Math. 180:2 (1998), 187–217.
Cegrell, U., *Maximal plurisubharmonic functions,* manuscript Ume[å]{} University 2002, revised 2006.
Cegrell, U., *The general definition of the complex Monge-Ampère operator,* Ann. Inst. Fourier (Grenoble) 54:1 (2004), 159–179.
Cegrell, U., *Potentials with respect to the pluricomplex Green function,* Research Reports 3, 2006, Mid Sweden University.
Cegrell, U., *A general Dirichlet problem for the Complex Monge-Ampère operator,* To appear in Ann. Polon. Math. Cegrell, U., *Approximation of plurisubharmonic functions in hyperconvex domains,* To appear in “Complex Analysis and Digital Geometry, Proceedings of the Kiselmanfest, Acta Universitatis Upsaliensis”.
Cegrell, U. and Hed, L., *Subextension and approximation of negative plurisubharmonic functions,* To appear in Michigan Math. J.
Cegrell, U. and Ko[ł]{}odziej, S., *The equation of complex Monge-Ampère type and stability of solutions,* Math. Ann. 334:4 (2006), 713–729.
Demailly, J.-P., *Mesures de Monge-Ampère et mesures pluriharmoniques,* Math. Z. 194:4 (1987), 519–564.
Henkin, G. M., *The Banach space of analytic functions on the ball and the bicylinder are not isomorphic,* Funct. Anal. Appl. 2:4 (1968), 334–341.
Lelong, P., *Fonctions entières de type exponentiel dans $\mathbb{C}^n$,* Ann. Inst. Fourier (Grenoble) 16:2 (1966), 269–318.
Lelong, P., *Fonction de Green pluricomplexe et lemmes de Schwarz dans les espaces de Banach,* J. Math. Pures Appl. 68:3 (1989), 319–347.
Poletsky, E. A., *Approximation of plurisubharmonic functions by multipole Green functions,* Trans. Amer. Math. Soc. 355:4 (2003), 1579–1591.
Sibony, N., *Une classe de domaines pseudoconvexes,* Duke Math. J. 55:2 (1987), 299–319.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The non-uniform demand network coding problem is posed as a single-source and multiple-sink network transmission problem where the sinks may have heterogeneous demands. In contrast with multicast problems, non-uniform demand problems are concerned with the amounts of data received by each sink, rather than the specifics of the received data. In this work, we enumerate non-uniform network demand scenarios under which network coding solutions can be found in polynomial time. This is accomplished by relating the demand problem with the graph coloring problem, and then applying results from the strong perfect graph theorem to identify coloring problems which can be solved in polynomial time. This characterization of efficiently-solvable non-uniform demand problems is an important step in understanding such problems, as it allows us to better understand situations under which the NP-complete problem might be tractable.'
author:
- 'Joseph C. Koo and John T. Gill, III[^1]'
bibliography:
- 'IEEEabrv.bib'
- 'references.bib'
title: '**Low-Complexity Non-Uniform Demand Multicast Network Coding Problems** '
---
Introduction {#sec-intro}
============
Network coding has been shown to enable higher transmission rates across communication networks, when compared against routing. This is because network coding allows data flows toward different sinks to share the same links, and—through appropriate coding of symbols—have the sinks still be able to decode out these disparate flows. In the butterfly network example first proposed by Ahlswede *et al.* [@ahlswede:network_info_flow] (see Figure \[fig:butterfly\_network\]), if the input data at node $w$ are coded together, it is possible to multicast two streams of information $b_1$ and $b_2$ from the source $s$ to both sinks $t_1$ and $t_2$ within a single time period. The benefits of allowing coding at nodes are evident; under routing, multiple time periods would be required to send both streams to both sinks. The authors show that in any network with a single source and multiple sinks, the information rate can achieve the minimum (over all sinks) of the maximum flow to the sink nodes. In subsequent work, Li *et al.* [@li:linear_net_coding] prove that linear network codes are sufficient for multicast, and Jaggi *et al.* [@jaggi:polytime_algs_multicast_code_constr] give a polynomial-time algorithm for constructing such linear codes.
Following the quick successes of characterizing and developing algorithms for multicast network coding problems, there has been much work concerning the construction of network codes for more general scenarios—although this has proven to be much more difficult. Koetter and Médard [@koetter:algebraic_approach_net_coding] give an algebraic characterization for achievable linear network codes, but prove that checking for the existence of such codes requires running time which is not polynomially bounded. Then, Rasala Lehman and Lehman [@rasalalehman:complex_class_net_info_flow] prove that for most network coding scenarios, finding linear network codes to satisfy arbitrary source and demand requirements is NP-hard. Of relevance to the current work is the problem of constructing network codes to send data from a single source to multiple sinks with arbitrary demands (potentially with different demands by different sinks).
In this work, we study networks where the single source may send data to multiple sinks at unequal rates. The motivation for this can be seen in the extended butterfly network of Figure \[fig:extended\_butterfly\_network\]. Here, the traditional butterfly network is augmented with an additional path between the source $s$ and sink $t_2$. Within a single time period, at most two streams can be transmitted to sink $t_1$, but it is possible to transmit more than two streams to sink $t_2$. If sink $t_2$ is constrained to only receive two streams, then available capacity is wasted.
The problem of sending unequal-rate data from a single source to multiple sinks has two flavors: the multiple multicast connections problem [@lun:net_coding_cost_criterion] and the non-uniform demand problem [@cassuto:non_uniform_demands]. In the multiple multicast connections problem, the sinks are allowed to receive data at different rates, but the subset of information demanded by each particular sink—while arbitrary—is identified in advance. On the other hand, in the non-uniform demand problem, the amount of information a particular sink must be able to receive is specified in advance, but it does not matter which specific pieces of information are received. This is a scenario where the source has a large set of messages which it wishes to send to the sinks, and each particular sink wishes to receive a subset of the source’s messages; however, the requirement at each sink is only that the messages it receives is a subset of a particular size rather than a requirement of receiving some specific subset of messages. This problem can be understood as a relaxed version of the multiple multicast connections network coding problem, since if it is known that a sink is unable to receive a subset of size $n$, then any specific demand for a subset of size $n$ can automatically be rejected as impossible. Conversely, if it is known that a sink is able to receive some subset of size $n$, then it may be possible to find a network coding solution with a specific demand for that sink which is of size $n$. The non-uniform demand problem makes it possible to determine the maximum possible data rates that can be received by the sinks.
The non-uniform demand problem was originally investigated by Cassuto and Bruck [@cassuto:non_uniform_demands]. For demand scenarios where all sinks require the same high rate except for two sinks demanding some lower rate, the authors prove that it is possible to satisfy these demands in all cases—and linear codes are sufficient. They also give limited conditions for the achievability of non-uniform demands when there are more than two lower demanded rate sinks along with any number of higher rate equal-rate sinks. In our work, we more specifically consider the case of non-uniform demands where each \[possibly heterogeneous\] sink demands a data rate of its own maximum-flow (i.e., maximum point-to-point rate from the source). From this seemingly more restrictive setting, however, we are able to describe a larger class of networks for which the non-uniform demand problem is solvable (and solvable in polynomial time), thus enabling non-uniform demand network coding to be more widely applicable.
Although the class of network demand scenarios for which we give polynomial-time solutions is not exhaustive, it may be difficult to enumerate the conditions more generally. This is because the non-uniform demand network coding problem is NP-hard [@cassuto:non_uniform_demands]. We give an alternate proof of the NP-hardness of the non-uniform demand network coding problem (see Appendix), which proves this result for slightly different demand scenarios than those addressed in [@cassuto:non_uniform_demands].
Related Work {#subsec-relatedwork}
------------
A technique which we use is that of transmitting data along paths, or through flows. This approach has been widely used in the network coding literature, and has enabled many significant results. In Jaggi *et al.* [@jaggi:polytime_algs_multicast_code_constr], the polynomial-time algorithms for multicast problems rely on the concept of sending data down \[perhaps overlapping\] paths. In [@fragouli:info_flow_decomp], Fragouli and Soljanin give a decomposition of networks into flows, in order to model data transmission in a network more simply. Using this decomposition and a graph coloring formulation, alphabet size bounds for any network code are then proven. Although the flow-based and path-based approaches are similar in many ways, the two techniques differ in that the flow-based approach creates a new flow every time a piece of data is transformed by coding, whereas the path-based approach keeps track of each piece of data as it is sent individually down a path, even if any transformations get applied to the data. We shall use a path-based approach.
We briefly mention some results regarding the multiple multicast connections problem, since achievable solutions for such problems are also achievable for the non-uniform demand problem with the same demanded rates. (Of course, the reverse is not always true.) Many of these results consider the case of two sinks. In [@ngai:multisource_net_coding_two_sinks], after enumeration of all possible scenarios, the authors conclude that in the case of two sinks with differing rates, linear coding is sufficient. The same conclusion is made in [@ramamoorthy:single_source_two_terminal_network_net_coding], although the authors use a different approach which considers a path-based enumeration. A characterization of the achievable data rate region using network coding is given for the two sink case. For more than two sinks, conditions under which solutions exist for the multiple connections problem have not been enumerated.
Separately from the multiple multicast connections problem, the non-uniform demand problem itself has also been the subject of study. In [@cassuto:non_uniform_demands], Cassuto and Bruck introduce the problem and give some results concerning the achievability of the individual max-flow rates to each sink. In our work, we address similar guarantees but for a wider class of network demands. The authors in [@cassuto:non_uniform_demands] also prove that the non-uniform demand problem is NP-hard, using a reduction from a 3-CNF problem, although the demand problems for which their result holds have network coding solutions that do not fully utilize the available data rates (in our terminology, the solutions are not saturating). We supplement their proof by considering whether or not networks in which the network coding solutions use all possible paths are still difficult to solve. We take a similar approach by considering contamination amongst data; however, we do not allow intermediate decoding as they do. The non-uniform demand problem is also studied in [@chekuri:achiev_rates_non_uniform_demands], which considers the case where demands are allowed to be relaxed in the solution. For general networks, the authors give bounds on the fraction of max-flow rate which is achievable, and show networks for which the bounds are tight. We take a different approach and instead characterize specific network demand scenarios for which the max-flow rate can be achieved.
Outline of Paper {#outline}
----------------
In this paper, we will investigate the case of non-uniform demand network coding in which each sink receives data at its individual point-to-point (i.e., max-flow min-cut) capacity rate. In Section \[sec-notationdef\], we discuss useful notation. In Section \[sec-nonunifdemandassign\], we define our approach to the problem, and analyze some of its characteristics. Following that, in Section \[sec-streamassignalg\], we give an algorithm for determining if a non-uniform demand solution exists, and discuss some of its performance issues. Using this algorithm, we characterize in Section \[sec-effnonunifdemand\] a class of networks for which the non-uniform demand network coding problem can be solved in polynomial time. Of course, not all non-uniform demand network coding problems can be solved efficiently; in the Appendix, we give an alternate proof of the NP-completeness of the non-uniform demand problem which accounts for the demand scenarios we are considering.
Notation and Definitions {#sec-notationdef}
========================
We will consider a directed acyclic network graph $G = (V,E)$. Each sink will be indexed as $j \in \{1,2,\ldots,t\}$, where $t$ is the total number of sink nodes. (Recall that there is only a single source node $s$.) Because the graph is acyclic, there exists a partial ordering of the nodes starting from the source $s$. A partial ordering of the edges can be constructed based on the ordering of the nodes from which the edges originate; for edges $e = (v,w)$ and $e' =
(v',w')$, we denote $e \preceq e'$ if and only if $v \preceq v'$ in the partial ordering of nodes.
For a particular sink $j$, we define $\mathcal{P}_j$ as the set of paths associated with sink $j$. These are unit-capacity edge-disjoint paths from the source $s$ to sink $j$ and can be determined from maximum-flow algorithms such as the Ford-Fulkerson augmenting path algorithm [@ford:maxflow]. Paths $p \in \mathcal{P}_j$ are given as the elements of $\mathcal{P}_j = \{p_{j,1}, p_{j,2}, \ldots,
p_{j,n_j}\}$, where $n_j = |\mathcal{P}_j|$ is the number of paths to sink $j$. Call the set of all paths $\mathcal{P} = \bigcup_{j=1}^{t}
\mathcal{P}_j$. In contrast to much of the literature on network coding for multicast, we will consider the *maximum* of the max-flows (instead of the minimum of the max-flows) and denote this quantity $n$, so $n = \max_j |\mathcal{P}_j| = \max_j n_j$.
In order to keep track of data that overlaps onto paths with other sinks, we introduce the concepts of contamination and contaminating set. We say that *contamination* from path $p_{jk}$ onto path $p_{j'k'}$ occurs when data transmitted on $p_{jk}$ gets combined into the data which is supposed to be transmitted on $p_{j'k'}$. This can occur, for example, when data on paths $p_{jk}$ and $p_{j'k'}$ are coded together in order to be transmitted across an edge where the two paths overlap. Then the *contamination set* of $p_{jk}$ is the set of all paths which experience contamination due to data from $p_{jk}$. If we call $\mathcal{D}_{jk}(e)$ as the set of paths which are contaminated by path $p_{jk}$ downstream of edge $e$, then $\mathcal{D}_{jk}(e)$ can be defined recursively as follows: $$\mathcal{D}_{jk}(e) = \bigcup_{p_{j'k'}} \left\{ p_{j'k'} \cup
\left( \bigcup_{e' \succ e} \mathcal{D}_{j'k'}(e')\right) \right\}
\mbox{,}$$ where the union over $p_{j'k'}$ is over all paths $p_{j'k'}$ which overlap path $p_{jk}$ at edge $e$ (and $j' \ne j$). Then $\mathcal{D}_{jk} = \bigcup_{e \in E} \mathcal{D}_{jk}(e)$ gives the contamination set of $p_{jk}$. This definition accounts for a data stream on a path to contaminate onto paths that it does not explicitly overlap, due to contamination being spread from path to overlapping path.
We also wish to keep track of the particular data streams sent to each of the sinks. A stream is defined as the identifier of the data that is being transmitted down a particular path—as opposed to the identifier of the path itself. We identify a particular stream with the index $i \in \{1,2,\ldots,n\}$, which means that a path is only allowed to transmit a stream from the set $\{1,2,\ldots,n\}$. Each stream represents one information unit, of which only $n$ unique information units are allowed to be transmitted. This restriction is not prohibitive, since $n$ is the maximum max-flow. Each sink receives a subset of the same $n$ streams, so different sinks will likely receive many of the same streams. The number of distinct streams a particular sink receives is its data rate, since each unit-capacity edge-disjoint path can transmit at most only a single data stream.
We also define decodable and saturating solutions.
A *decodable solution* to a network coding problem is one in which every sink is able to decode all of the information which is intended to be sent to it. In the example of streams assigned to paths, a decodable solution is one in which every sink can recover all of the streams which are assigned on paths to that sink.
A *saturating solution* is an assignment $f:\mathcal{P} \to
\{1,2,\ldots,n\}$ from paths $p_{jk} \in \mathcal{P}$ to streams $\{1,2,\ldots,n\}$, such that for each $j \in \{1,2,\ldots,t\}$, $$f(p_{jk}) \ne \emptyset, \ \forall k \in \{1,2,\ldots,n_j\}$$ and $$f(p_{jk}) \ne f(p_{jk'}), \ \forall k \ne k'
\mbox{.}$$
That is, a saturating stream assignment is a stream assignment in which all paths to every sink are assigned some data stream; no path is left unassigned. Moreover, any streams assigned to different paths to the same sink must be distinct. Otherwise, if two paths carried the same stream, one of the paths is redundant and does not carry additional information. Thus, a saturating stream assignment is one in which each sink $j$ achieves its maximum possible data rate of $n_j$.
We briefly mention the concept of intermediate decoding. Specifically, for the network codes we are considering, we do not allow intermediates nodes (i.e., nodes which are neither source nor sink) to decode data and retransmit only a part of the data on its outgoing links. In other words, intermediate nodes are not allowed to remove any contamination which it might receive on its incoming links, even if it possesses enough information to decode out the contamination. Although this condition may prevent certain network codes from being considered, it is still general enough that except for certain cases, we should be able to find the appropriate network coding solution if it exists.
The *non-uniform demand problem* is the following solvability problem (adapted from [@cassuto:non_uniform_demands]): Given a directed acyclic network graph $G = (V,E)$ (where each edge has capacity $1$), source $s$, sinks $j \in \{1,2,\ldots,t\}$, and demand function $d:\{1,2,\ldots,t\} \to \{1,2,\ldots,n\}$ (where $d(j)$ is the demanded rate of sink $j$), is there a network coding solution such that for all $j$, sink $j$ receives information at a rate $d(j)$?
The Non-Uniform Demand Stream Assignment Problem {#sec-nonunifdemandassign}
================================================
The goal is to determine whether or not an assignment of data streams to paths can give a decodable network solution.
The *non-uniform demand stream assignment problem* is the following: Given a network graph $G = (V,E)$ and a decomposition into paths, is there an assignment of streams to paths, such that no intermediate decoding occurs, and the solution is both saturating and decodable?
We establish necessary and sufficient conditions for a network to have a saturating and decodable solution.
\[thm:sat\_decode\_cond\] Given a set of paths between the source and the sinks, and if no intermediate decoding is allowed, there exists a saturating and decodable solution if and only if all streams which contaminate onto paths to a particular sink have also been assigned to some other path to the same sink.
Necessity follows from the fact that if the solution is decodable, then each sink can separate out all streams and all contamination sent to it. For a particular sink, each path carries an assigned stream mixed with the contamination from along that path. Because no intermediate decoding is allowed, all contamination arrives at the sink, but arrives mixed in with the assigned streams on the respective paths. If there are $n_j$ paths to the sink $j$ and the solution is saturating, then there are $n_j$ unique streams assigned on paths to the sink. Now, if a contaminant is not also assigned to some other path to that sink, then that means that there is data from at least $n_j + 1$ streams on inputs to the sink (the assigned $n_j$ streams of data plus at least one more data stream from the contamination). However, because there are only $n_j$ paths from the source to that sink, where each path supports a data rate of $1$, that means that no more than $n_j$ unique data streams can be received by the sink (or else the max-flow condition would be violated). Thus, any situation where more than $n_j$ data streams (perhaps mixed) can be seen by the sink is a situation where fewer than $n_j$ data streams can be decoded successfully by the sink, and the solution is either not decodable or not saturating.
For each sink, it is straightforward to show sufficiency of assigning the contamination onto a path to that sink as the primary stream on another of its paths, in order to guarantee decodability. If no intermediate decoding is allowed, all contamination arrives at the sink. If no other path has been assigned the same data stream, then there is no way to determine (or even have partial knowledge) of the data due to the contamination in order to either utilize or remove it. Thus, there must be an assigned stream on some other path to that sink which provides this information.
The concept of saturation is important, so that it is possible to state (using max-flow theorems) that contamination without a corresponding assigned stream can not be removed, as there will not be enough flow to support this additional data. Saturation is also useful because it enables us to determine whether or not the maximum data rate actually utilizes all paths, without repetitive data streams. In fact, if data is repeated (i.e., multiple paths to the same sink are assigned the same stream), then one of the multiple paths could be shut off with no harm to the data rate toward that sink, and possibly even increasing data rates across the entire network due to less contamination onto paths to other sinks.
An Algorithm for Assigning Streams {#sec-streamassignalg}
==================================
We now give an algorithm for solving the non-uniform demand stream assignment problem. Again, we restrict ourselves to the case where intermediate decoding does not occur within the network; that is, even though nodes within the network may encode information, there is no preliminary decoding (and removal of streams) except at the sink nodes which are receiving information. Equivalently, the output of any node can only be more contaminated (but not less contaminated) than any of the inputs to that node. Admittedly, precluding the removal of streams within the network does limit the solution space. However, since the goal is to maximize the data rate to each sink, solutions which do remove streams must be able to compensate for the loss of data rate due to the removal of streams with some other benefit (e.g., less contamination further down the network).
Our goal is to be able to assign streams to paths—with guarantees that the contamination will be decodable—while maximizing the number of streams transmitted. We give a method which guarantees saturation and decodability, by utilizing a polynomial transformation to graph coloring. We first give the transformation of the network graph into a coloring graph in Algorithm \[alg:color\_graph\_transform\], and then give the algorithm for finding the saturating and decodable solution in Algorithm \[alg:graph\_color\_alg\].
Original network graph $G = (V,E)$, decomposed into edge-disjoint paths $\mathcal{P}_j = \{p_{jk} \, | \,
k=1,\ldots,n_j\}$ for each sink $j$ \[it:vhat\] Create vertices $v_{jk} \in \hat{V}$, for $k=1,\ldots,n_j$, which are associated with paths $p_{jk} \in \mathcal{P}_j$. For each $j$, introduce $n-n_j$ additional vertices $w_{jk} \in \hat{V}$. ($v_{jk}$ are *regular vertices* and $w_{jk}$ are *fictitious vertices*.) Let $$\hat{V}_j = \{v_{jk} \, | \, 1 \leq k \leq n_j\} \cup \{w_{jk} \, | \,
1 \leq k \leq n-n_j\}
\mbox{.}$$ Call $\hat{V}_j$ the *sink subgraph* associated with sink $j$. Then $\hat{V} = \bigcup_{j=1}^{t} \hat{V}_j$. \[it:ehat\_complete\] For each $j$, connect all vertices $v_{jk}$ and vertices $w_{jk}$ together into a clique. Specifically, for each $j$, let $$\begin{aligned}
\hat{E}_j^{\mathrm{complete}}
& = & \{(v_{jk}, v_{jk'}) \, | \, 1 \leq k < k' \leq n_j\} \cup \\
&& \{(w_{jk}, w_{jk'}) \, | \, 1 \leq k < k' \leq n-n_j\} \cup \\
&& \{(v_{jk}, w_{jk'}) \, | \, 1 \leq k \leq n_j, 1 \leq k' \leq
n-n_j\}
\mbox{.}\end{aligned}$$ \[it:ehat\_overlaps\] For each vertex $v_{jk}$, connect $v_{jk}$ to vertices $w_{j'k'}$ for all $k'=1,\ldots,n-n_{j'}$, if path $p_{jk}$ contaminates onto some path to sink $j' \ne j$. For each $j$, call $$\begin{aligned}
\lefteqn{\hat{E}_j^{\mathrm{overlaps}} = } \\
&& \bigcup_{k=1}^{n_j} \{(v_{jk}, w_{j'k'}) \, | \, 1 \leq k' \leq
n-n_{j'} \ \mathrm{if} \ \exists \tilde{k} \ \mathrm{s.t.} \ p_{j'
\tilde{k}} \in \mathcal{D}_{jk}\}
\mbox{.}\end{aligned}$$ That is, if path $p_{jk}$ to sink $j$ contaminates onto some path $p_{j' \tilde{k}}$ to sink $j'$, then node $v_{jk}$ must connect to all $n-n_{j'}$ fictitious vertices associated with sink $j'$. \[it:ehat\] Let $\hat{E} = \bigcup_{j=1}^{t} (\hat{E}_j^{\mathrm{complete}} \cup
\hat{E}_j^{\mathrm{overlaps}})$. coloring graph $\hat{G} = (\hat{V}, \hat{E})$
Directed acyclic network graph $G = (V,E)$, source $s$, and sinks $j \in \{1,2,\ldots,t\}$ \[it:edge\_disjoint\] For each sink $j$, find a set of edge-disjoint paths from $s$ to $j$. Call the set of these paths $\mathcal{P}_j = \{p_{jk} \, | \,
k=1,\ldots,n_j\}$, where there are $n_j$ such paths. Let $n = \max_j
n_j$. \[it:create\_graph\] Using Algorithm \[alg:color\_graph\_transform\], construct coloring graph $\hat{G}$ from $\mathcal{P} = \bigcup_{j=1}^{t} \mathcal{P}_j$. \[it:color\] Color $\hat{G}$ using exactly $n$ colors. Let $c_{jk}$ be the color of $v_{jk}$ in $\hat{G}$. \[it:assign\_streams\] For each path $p_{jk} \in \mathcal{P}$, assign stream $c_{jk}$ to that path. (Each path $p_{jk} \in \mathcal{P}$ in the network graph $G$ is assigned the stream given by the color of its associated vertex $v_{jk} \in \hat{V}$ in the coloring graph $\hat{G}$.) \[it:netcode\] In the network graph, at each node where the inputs to the node are different streams, send as the output of the node a combination of the data from the input streams (e.g., linear combination, or some other combining method)—taking care that no input streams are nullified in the node output.
The coloring graph $\hat{G}$ can be interpreted with respect to the solution found by Algorithm \[alg:graph\_color\_alg\]. The vertices $v_{jk}$, $k=1,\ldots,n_j$, correspond to paths in the original network, and so are called regular vertices. The additional vertices $w_{jk}$, $k=1,\ldots,n-n_j$, correspond to fictitious paths, indicating the streams which are *not* assigned to paths leading to sink $j$—hence the name fictitious vertices. The edges in $\hat{E}_j^\mathrm{complete}$ form a complete subgraph among all $n_j$ regular vertices associated with sink $j$, guaranteeing saturation. (In fact, the entire induced subgraph of $\hat{V}_j$ is a clique.) For $\hat{E}_j^\mathrm{overlaps}$, the edges $(v_{jk},
w_{j'k'})$ connect the vertices from sink $j$ to the vertices of sink $j'$ if there is some overlap on paths toward these two sinks; these edges force a relationship between the streams assigned on paths to one sink and streams assigned on paths to other sinks, providing decodability.
Because $n = \max_j n_j$, there must exist at least one clique of size $n$ (associated with the induced subgraph of $\hat{V}_{j^{\star}}$, where $j^{\star} = \arg\max_j n_j$). Thus, $\hat{G}$ can not be colored with fewer than $n$ colors, so step \[it:color\] of Algorithm \[alg:graph\_color\_alg\] is equivalent to coloring $\hat{G}$ with *at most* $n$ colors. If the minimum coloring solution of $\hat{G}$ requires more than $n$ colors, then the following lemma tells us that decodability has been violated.
\[lem:coloring\_violation\] In the equivalent coloring graph constructed from Algorithm \[alg:color\_graph\_transform\], if the chromatic number $\chi(\hat{G}) > n$, then the original network is not decodable.
If $\chi(\hat{G}) > n$ and every vertex is a member of at least one induced clique of size $n$, then there must exist two cliques of size $n$ such that there is at least one edge connecting these two cliques. Moreover, because the coloring graph must be colored with more than $n$ colors, some pair $(j,j')$ of connected cliques (each clique associated with a different sink) must satisfy the following condition: If clique $j$ does not have color $c$ within it, then its fictitious vertices must be connected to a regular vertex in clique $j'$ which has color $c$. In this case, sink $j$ can not decode color $c$ even though some path to $j$ has contamination $c$ from a path to $j'$, and so decodability is violated.
\[thm:sat\_decod\_cond\] Algorithm \[alg:graph\_color\_alg\] succeeds in coloring the equivalent coloring graph with exactly $n$ colors if and only if the original network graph has a decodable and saturating solution with no intermediate decoding.
It is clear that $\chi(\hat{G}) = n$ is necessary for the solution to be decodable and saturating. From Lemma \[lem:coloring\_violation\], we know that if the network is decodable, then the equivalent coloring graph must have $\chi(\hat{G}) \leq n$. Now consider sink $j^{\star}$, where $j^{\star} = \arg\max_j n_j$. If the network is saturating, then sink $j^{\star}$ must be able to receive $n =
n_{j^{\star}}$ distinct streams. That is, the clique associated with sink $j^{\star}$ must be colored with at least $n$ colors. This gives us $\chi(\hat{G}) \geq n$. Thus, $\chi(\hat{G}) = n$.
Next we prove sufficiency of $\chi(\hat{G}) = n$ for a decodable and saturating solution. For the coloring to be valid, if path $p_{jk}$ contaminates onto *any* path $p_{j' \tilde{k}}$ to sink $j'$, then by construction of $\hat{E}^\mathrm{overlaps}$, none of the fictitious vertices $w_{j'k'}$ associated with sink $j'$ may have the same color as vertex $v_{jk}$ (call this color $c_{jk}$). Because $\chi(\hat{G}) = n$ and each sink subgraph is an induced clique of size $n$, some regular vertex $v_{j'k'}$ to sink $j'$ must be colored $c_{jk}$. Equivalently, contamination due to path $p_{jk}$ onto path $p_{j' \tilde{k}}$ has been assigned to some path $p_{j'k'}$ to sink $j'$ (and this is true for all possible contaminations), so by Theorem \[thm:sat\_decode\_cond\] the solution is decodable. Moreover, if $\chi(\hat{G}) = n$, then by construction of $\hat{E}^\mathrm{complete}$, all paths have assigned streams, and the assigned streams are distinct for different paths to the same sink. Thus, the solution is saturating.
Theorem \[thm:sat\_decod\_cond\] gives necessary and sufficient conditions for a saturating and decodable solution to exist. Moreover, the stream assignment algorithm tells us how to allocate streams in order to construct this solution.
\[ex:extended\_butterfly\_network\] Figure \[fig:extended\_butterfly\_example\] shows the result of Algorithm \[alg:graph\_color\_alg\] on the extended butterfly network.
Shortcomings {#subsec-alg_shortcomings}
------------
Assuming the correct set of edge-disjoint paths are chosen in the first step of the algorithm, then if the solution exists it will be found. However, we do not address the proper selection of edge-disjoint paths, even though there may be multiple path decompositions, where some decompositions lead to sub-optimal assignments. For example, it is possible to construct a counterexample network with a given path decomposition, where switching a single edge for one path greatly increases allowed throughput.
The algorithm determines—for a given set of paths—whether or not $n$ streams can be assigned, but to determine if some $\bar{n} >
n$ streams can be assigned, additional fictitious vertices need to be introduced. For each sink, an additional $\bar{n} - n$ fictitious vertices must be introduced in order to get sink subgraphs of size $\bar{n}$. The relationship between these larger graphs and the original coloring graph is unknown, and it is possible that no matter how large $\bar{n}$ is chosen, it will still be impossible to find a saturating and decodable assignment with $\bar{n}$ streams. If one wishes to determine how many additional available streams will guarantee saturation and decodability, $\bar{n}$ could be increased without bound while searching for a possible stream assignment.
In fact, it is possible to have networks where there does not exist a saturating solution which is also decodable, no matter how large the available stream set is. This can occur when there is too much overlap but not enough available paths to remove the contamination. For example, consider a two sink case, where $n_1 = 1$ and $n_2 = 2$. Suppose both paths of sink $2$ overlap the path to sink $1$ at some link\[s\]. No matter what the stream assignment (or how large the space of possible streams) for sink $1$, it will never be able to decode out both contaminants if saturation for both sinks is required.
Thus, our algorithm only works for the restrictive case where no intermediate decoding is allowed, yet all paths to sinks must be saturating. Either loosening the saturation or the no intermediate decoding restrictions would be beneficial, but at the moment, the algorithm relies on both conditions.
Another issue to keep in mind is that because coloring is an NP-complete problem, there are no known polynomial time algorithms which will perform the coloring step (unless $P = NP$). Additionally, there are no good approximation algorithms known for the graph coloring problem (see [@karger:approx_graph_coloring_sdp] for algorithms which can color $n$-colorable graphs with number of colors logarithmic in the number of vertices of the graph, but with no guarantees based on the actual chromatic number $n$). Even if there were good approximation algorithms, an approximation algorithm might not be enough to answer the question of whether or not a saturating solution exists (i.e., whether or not $\chi(\hat{G}) = n$) since we require finding the chromatic number exactly. One might conjecture that because the coloring graph $\hat{G}$ is carefully constructed, it might have some special structure which would allow for a polynomial-time coloring algorithm. In the next section, we give some structural properties of the coloring graph which can lead to a polynomial-time coloring, but we also show a counterexample network where this particular structural analysis is not sufficient to prove polynomial-time solvability.
Efficiently-Solvable Non-Uniform Demand Problems {#sec-effnonunifdemand}
================================================
The coloring step in the stream assignment problem is problematic, as the graph coloring problem is NP-complete and so in general no known polynomial-time algorithm can solve the problem. However, if we restrict our class of demand problems to only those for which the corresponding coloring graph is polynomial-time solvable, then such demand problems will also be polynomial-time solvable. Specifically, we consider the class of graphs known as Berge graphs, which are graphs characterized by the absences of both odd holes (induced cycles of odd length at least 5) and odd antiholes (complements of odd holes). Then by the strong perfect graph theorem [@chudnovsky:strong_perfect_graph_theorem], a Berge graph is also a perfect graph, so the chromatic number of a Berge graph is equal to the size of its maximum clique.
This fact is useful in finding solutions to the non-uniform demand problem because in our formulation, the maximum clique is easily found, so if the coloring graph is Berge, then the chromatic number is also readily determined. The main result of this section is that we can find the maximum clique for the coloring graph of Algorithm \[alg:graph\_color\_alg\] in polynomial time and hence also its chromatic number if the coloring graph is Berge. We first prove some preliminary results.
\[lem:no\_clique\_three\_sink\_cliques\] Any induced clique consisting of vertices from different sink subgraphs can only consist of vertices from at most two sink subgraphs. That is, it is impossible to induce a complete subgraph consisting of at least one vertex from each of $\hat{V}_j$, $\hat{V}_{j'}$, and $\hat{V}_{j''}$.
For vertices belonging to different sink subgraphs, i.e., $v \in
\hat{V}_j$ and $v' \in \hat{V}_{j'}$ where $j \ne j'$, either $v$ is regular and $v'$ is fictitious, or $v$ is fictitious and $v'$ is regular. Regular vertices are not connected to regular vertices, nor are fictitious vertices connected to fictitious vertices—unless they belong to the same sink subgraph. Any complete subgraph consisting of at least one vertex from each of $\hat{V}_j$, $\hat{V}_{j'}$, and $\hat{V}_{j''}$ must contain at least two vertices of the same type from different sink subgraph (e.g., two regular vertices and one fictitious vertex, where each vertex is from a different sink subgraph). However, such a scenario can not exist, as that implies that two vertices of the same type but from different sink subgraphs are connected.
The preceding lemma tells us that in order to find the maximum clique in the coloring graph, all we need to do is search for induced cliques pairwise between sink subgraphs. We can select sink subgraphs two at a time and determine the largest induced complete subgraph consisting only of vertices from these two sink subgraphs. This procedure requires solving $\binom{t}{2}$ subproblems, where each subproblem can be performed in time which is polynomial in $n$.
\[lem:max\_clique\_pairwise\_polytime\] For a pair of sink subgraphs $\hat{V}_j$ and $\hat{V}_{j'}$, finding the maximum induced complete subgraph of $\hat{V}_j \cup \hat{V}_{j'}$ takes time polynomial in the sink subgraph size $n$.
If sinks $j$ and $j'$ do not have any overlapping paths, then because $\hat{V}_j$ and $\hat{V}_{j'}$ are disjoint, the maximum induced complete subgraph of $\hat{V}_j \cup \hat{V}_{j'}$ is $\hat{V}_j$ (or $\hat{V}_{j'}$), which has size $n$. Disjointness of $\hat{V}_j$ and $\hat{V}_{j'}$ can be checked by considering each vertex $v_{jk} \in \hat{V}_j$ and seeing if it has an edge to any vertex in $\hat{V}_{j'}$. This requires $n$ steps.
If a path $p_{jk}$ to sink $j$ contaminates onto some path to sink $j'$, then the regular vertex $v_{jk}$ is connected to all of the fictitious vertices of $\hat{V}_{j'}$. Thus, the largest induced complete subgraph consisting of both regular vertices from $\hat{V}_j$ and fictitious vertices from $\hat{V}_{j'}$ has size $m_{j,j'} +
(n-n_{j'})$, where $m_{j,j'}$ is the number of regular vertices of $\hat{V}_j$ which are connected to the fictitious vertices of $\hat{V}_{j'}$. (Recall that $n-n_{j'}$ is the number of fictitious vertices of $\hat{V}_{j'}$.) Computing $m_{j,j'}$ takes $O(n)$ time, as it merely requires counting up the number of regular vertices of $\hat{V}_j$ that are connected to the fictitious vertices of $\hat{V}_{j'}$. Equivalently, $m_{j,j'}$ can be computed by counting the number of paths to sink $j$ which contaminate onto some path to sink $j'$. Of course, the largest induced complete subgraph of $\hat{V}_j \cup \hat{V}_{j'}$ may instead consist of regular vertices from $\hat{V}_{j'}$ and fictitious vertices from $\hat{V}_j$; by a similar argument, finding such a subgraph also takes polynomial time. Then we can find the maximum induced complete subgraph of $\hat{V}_j \cup \hat{V}_{j'}$, and the size of this induced subgraph is $n + \max(0, m_{j,j'} - n_{j'}, m_{j',j} - n_j)$.
From this, it can be readily shown that finding the maximum clique in the coloring graph is polynomial-time. We can then conclude that it is possible to determine the existence of a saturating and decodable solution (again, disregarding intermediate decoding) in polynomial time.
\[thm:berge\_polytime\] For a particular non-uniform demand scenario, if the associated coloring graph is a Berge graph, then it is a polynomial-time operation to determine whether or not there exists a saturating and decodable solution which does not require intermediate decoding. Moreover, if the solution exists, it can be found using the non-uniform demand stream assignment algorithm (Algorithm \[alg:graph\_color\_alg\]).
From Lemma \[lem:max\_clique\_pairwise\_polytime\], we know that finding the maximum induced clique between two sink subgraphs is a polynomial time operation. Thus, finding the maximum clique of the coloring graph takes polynomial time, as it consists of solving $\binom{t}{2} = \frac{t(t-1)}{2}$ such subproblems. Because this coloring graph is a Berge graph, then its chromatic number can be found in polynomial time, since the chromatic number is equal to the maximum clique size. From Theorem \[thm:sat\_decod\_cond\], we can then determine if the original network graph has a saturating and decodable solution. Not only that, but if the coloring requires no more than $n$ colors, then the coloring found from the stream assignment algorithm immediately gives the non-uniform demand solution.
The interpretation of Theorem \[thm:berge\_polytime\] is that for coloring graphs which are Berge graphs, if we find that the maximum clique has size $n$, then we can conclude that the coloring graph can be colored with $n$ colors, and so the original non-uniform demand network coding problem has a saturating and decodable solution. If, however, the maximum clique has size greater than $n$, then we can also conclude that no saturating and decodable solution exists—at least no solution which does not require intermediate decoding. This result is particularly promising, as we can then quickly enumerate a sufficient condition under which a saturating and decodable solution can be found in polynomial time—specifically, if the associated coloring graph is Berge.
One might ask if the step of determining whether or not a graph is Berge might be a difficult task, as any difficulties in doing so would outweigh any benefits gained by solving the demand problem efficiently. However, a polynomial-time algorithm for recognizing Berge graphs does exist [@chudnovsky:recog_berge]. Consequently, if it is determined that the coloring graph is a Berge graph, then the stream assignment algorithm can be used to find the saturating and decodable solution to the non-uniform demand problem in polynomial time. Or, if it is determined that the coloring graph is not a Berge graph, then some other, perhaps superpolynomial time, algorithm will be needed to perform the coloring step.
However, non-uniform demand scenarios which lead to non-Berge coloring graphs do exist. We give an example.
\[ex:non\_berge\_coloring\_graph\] Consider the network given in Figure \[fig:non\_berge\_example\]. Its corresponding coloring graph has an odd hole of length $5$, induced by the vertices $v_{1,1}$, $w_{1,1}$, $v_{2,1}$, $v_{2,2}$, and $w_{3,1}$, so it is not Berge. However, a valid coloring of size $n=3$ does exist.
\
Conclusion {#sec-conclusion}
==========
In this paper, we have considered the non-uniform demand network coding problem, where the sinks are allowed to receive data at unequal rates. We give an algorithm for finding network coding solutions which satisfy the decodability and saturation properties. Additionally, we show that for certain types of networks, i.e., those which can be transformed into equivalent Berge graphs, our algorithm can find the solution in polynomial time. Moreover, it will be difficult to do much better for the general case, as the non-uniform demand problem is NP-hard, even when demands are restricted to only those which are saturating.
Our results can be interpreted relative to the case of equal-rate multicast, where network coding solutions can always be found in polynomial time [@jaggi:polytime_algs_multicast_code_constr]. Our algorithm relies on introducing fictitious vertices in the coloring graph, which corresponds to introducing fictitious paths in the original network. These fictitious paths do not overlap anywhere with any other paths, but only serve to bring the total number of paths to each sink up to $n$. We can perform multicast on the expanded network consisting of the original network plus the network induced by the fictitious paths; in this expanded network, each sink is guaranteed to receive $n$ streams of information. The main contribution of our work is that using our algorithm, we can directly specify that the subset of paths in the expanded network corresponding to paths in the original network is assigned mutually-decodable streams, without needing any of the information transmitted on the fictitious paths. From the perspective of the equal-rate multicast problem on the expanded graph, the interpretation of our algorithm is that it provides a partitioning of information between data transmitted on the original network paths and data transmitted on the fictitious paths.
There are a few issues which require further study. We have not considered the optimal selection of paths; our assumption is that the set of paths we use are the ones which give the demand solution if it exists. The optimal selection of paths is a challenging problem in itself, as it requires knowing what types of demand solutions may arise from the particular choice of paths. When implementing this algorithm, heuristics—such as minimizing the number of overlapping links on paths to different sinks—will most likely be sufficient. We also mention that although our conditions guarantee that a network code will exist if our algorithm finds a solution, the actual construction of the network code is not detailed. In the case of no intermediate decoding, linear codes will be sufficient. However, if intermediate decoding were to be allowed, then we must be more careful, as it has been shown that sometimes nonlinear codes are required to solve certain other network coding problems [@dougherty:insuff_linear_coding].
Our approach considers network coding scenarios which are scalar, where the same code is employed during every time period. Although this allows for a wide variety of codes and is also practically implementable, there are certain network coding problems where vector solutions (i.e., solutions where the network code may be different at each time period) exist, but scalar solutions do not [@medard:on_coding_non_multicast]. One avenue of inquiry would be the adaptation of our algorithms to find vector solutions in the cases where scalar solutions do not exist; this should be possible by augmenting our network graphs to also include a time dimension. However, characterizing the set of networks with polynomial-time-solvable vector solutions (but no scalar solutions) will require more work.
\[app-npcomp\]
We give an alternate proof of NP-completeness of the non-uniform demand network coding problem, via a polynomial reduction from a general graph coloring problem. Unlike [@cassuto:non_uniform_demands], in which the network demand problems shown to be NP-hard do not have fully saturated demands, our proof considers sink demands in which saturation must occur. The coloring problem which we consider is the following: Given an undirected graph $\hat{G} = (\hat{V}, \hat{E})$, is there a coloring using $n$ \[or fewer\] colors? We first give the reduction, followed by a proof that the reduction leads to an equivalent problem.
Undirected graph $\hat{G} = (\hat{V},\hat{E})$ to be colored \[it:constr\_paths\] For each edge $e_j = (v_j,w_j) \in \hat{E}$, construct a sink $j$ in the network graph consisting of two paths $p_{j,1}$ and $p_{j,2}$. If two edges $e_j = (v,w_j) \in \hat{E}$ and $e_i = (v,w_i) \in \hat{E}$ share a vertex $v$, then force the paths $p_{j,1}$ and $p_{i,1}$ to overlap at some link. Call the overlapping link in the network graph by $v$. (If the shared vertex is $w$ such that $e_j = (v_j,w) \in
\hat{E}$ and $e_i = (v_i,w) \in \hat{E}$, then force paths $p_{j,2}$ and $p_{i,2}$ to overlap at some link $w$.) Thus, vertices in the coloring graph determine the intersections of paths in the network graph—where the link of intersection occurs according to the vertex in the coloring graph. \[it:same\_paths\] Introduce another $|V|$ sinks, with only a single path to each sink. Label these sinks $1,2,\ldots,|V|$. For a particular sink $v$, call the single path $p_v$, and make path $p_v$ intersect with all other paths which cross through link $v$ in the network graph. \[it:force\_n\_colors\] Introduce one additional sink, with $n$ paths. These $n$ paths do not intersect any paths defined in prior steps. Solve the non-uniform demand stream assignment problem on the resulting network graph.
Step \[it:constr\_paths\] of the above algorithm sets up most of the network graph. Overlaps between paths reflect the fact that a vertex can not be colored two different colors. The addition of $|V|$ sinks in step \[it:same\_paths\] forces the stream assignment algorithm to assign the same color to all paths crossing through the same link $v$; otherwise, in the sinks with two paths, it may be possible that the corresponding stream will be assigned to the path in the pair which does not overlap at the considered link. Furthermore, the single sink with $n$ paths in step \[it:force\_n\_colors\] guarantees that at least $n$ different streams will be assigned.
\[lem:nudsa\_equiv\_coloring\] Performing non-uniform demand stream assignment on the network digraph $G$ formed from Algorithm \[alg:reduc\_from\_graph\_color\] is equivalent to coloring the original undirected graph $\hat{G}$.
First we show that if there exists a coloring solution for $\hat{G}$ using $n$ colors, then there will also be a non-uniform demand stream assignment on the constructed network graph with $n$ streams. To do so, start with a coloring solution. For a particular vertex $v$ in the coloring graph, assign the stream corresponding to the color of vertex $v$ to all paths which intersect at the associated link $v$ in the network graph. Because no path has more than one link which has overlap, then there is no ambiguity about the stream which is assigned to that path. Because graph coloring guarantees that the vertices connected by an edge will be colored different colors, each sink from step \[it:constr\_paths\] will receive two paths that have different stream assignments. Thus, the solution is saturating. Not only that, in the network graph, any intersecting paths only intersect at one link, so contamination is mitigated by assigning the same stream to all paths which intersect at the same edge. Thus, the solution to the graph coloring with $n$ colors gives a non-uniform demand stream assignment for the constructed network graph using exactly $n$ distinct streams. This stream assignment is both saturating and decodable.
Conversely, assume that there exists a saturating and decodable stream assignment to the non-uniform demand stream assignment problem on the constructed network graph, which uses exactly $n$ streams. Then this solution can be used to determine a graph coloring of the original graph, with $n$ colors. To prove this, first consider the sinks which have single paths. From these sinks, suppose path $p_v$ to sink $v$ (associated with vertex $v$ in the coloring graph) is assigned stream $c$. By decodability, any other paths which intersect path $p_v$ must be carrying stream $c$. That is, for a path $p'_j$ which intersects $p_v$, then its pair path $p''_j$ (i.e., toward the same sink) is not the path of the pair which is carrying stream $c$. Otherwise, sink $v$ would need to decode out the stream on path $p'_j$ (which would be some $c' \ne c$), but sink $v$ can not, since it is only assigned to receive stream $c$ from the single path $p_v$. Thus, all paths associated with the same vertex $v$ in the original coloring graph must have the same stream assignment; this is the color assigned to vertex $v$. Now, consider the sinks with paired paths. Because the non-uniform demand stream assignment solution is saturating, that means that the two paths are assigned different streams. This is equivalent to the requirement that vertices which are connected by an edge in the original coloring graph be assigned different colors. Thus, the stream assignment on $G$ using $n$ streams gives a coloring on $\hat{G}$ using $n$ colors.
From the preceding construction, it is straightforward to determine the complexity of the stream assignment problem.
\[thm:nudsa\_np\_complete\] The non-uniform demand stream assignment problem with saturated demands is NP-complete.
First we show that the problem is in NP, by showing that it takes polynomial time to check if a given solution to the non-uniform demand stream assignment problem is feasible. For every overlap link in the network graph $G$, check if the sinks receiving the paths crossing through the link have additional streams assigned to them which are the same as all the contaminations from that link. This takes at most $O(n|E|)$ time, where $|E|$ is the number of links in the corresponding network graph (and $|E|$ can be as small as $4|\hat{E}|
+ 3|\hat{V}| + n$).
Next, we show that the transformation from the coloring graph to the related network graph given in Algorithm \[alg:reduc\_from\_graph\_color\] is polynomial. The network graph has $|\hat{E}| + |\hat{V}| + 1$ sinks (more precisely, it has $2|\hat{E}| + |\hat{V}| + n$ paths), so the reduction is polynomial. Because any instance of graph coloring can be polynomially reduced to a non-uniform demand stream assignment problem, and solutions can be checked in polynomial time, the non-uniform demand stream assignment problem is NP-complete.
From the above arguments, we also conclude that non-uniform demand stream assignment can be solved in polynomial time when $n = 2$ (i.e., when the maximum data rate to any sink is upper bounded by $2$). This is because graph coloring is polynomial time (i.e., by searching for bipartiteness in the graph) when only $2$ colors are allowed [@garey:np_completeness].
Acknowledgements {#sec-ack .unnumbered}
================
The authors would like to thank Amin Saberi for insightful discussions, and William Wu and Caleb Lo for helpful comments which improved the readability of this paper. Part of this work was completed while J. Koo was supported by a National Defense Science and Engineering Graduate Fellowship.
[^1]: J. Koo and J. Gill are both with the Department of Electrical Engineering, Stanford University, Stanford, CA 94305, USA. E-mail: [{jckoo, gill}@stanford.edu]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
#### Subject
This work is dedicated to the questions of the contemporary medical image visualization, the architecture design of the cloud server systems and the using of methods for the .DICOM data representation for the distributed smart healthcare industry systems.
#### Purpose
In modern medicine and biology, in the pace of research along with the objective need for a constant increase, there is a sharp necessity of the three-dimensional data representation [@Tractica]: requiring high-performance methods of the three-dimensional visualization, processing, decomposition, reconstruction and analysis. In many ways, the reason for the sharp growth is associated with the development of the computer technologies (from quantity into quality): the quantitative characteristics of the devices and computers used, the emergence of three-dimensional technologies for processing, visualization and research of data, and, on this basis, the creation of new three-dimensional methods for the human-machine interaction.
#### Research methodology
This paper proposes a method for visualization using direct volumetric rendering based on the cubic spline interpolation on a ray emission technology. Modifications of these algorithms, based on a block decomposition, are proposed and still investigated. The server architecture is proposed as a cloud hypervisors server system for the grid processing of the data required.
#### Research results
The well-known method of block decomposition of the medical data was studied: the idea of block searching was implemented, auxiliary structures reducing the volume range were used to skip empty spaces (empty space leaping). The cloud architecture for the server processing was developed as well as the user interface was proposed.
#### Scope of application
The work aims to investigate the possible contemporary interactive cloud based solutions in the fields of the applied medicine for the smart Healthcare as the data visualization open-source free system distributed under the MIT license.
#### Conclusions
A comparative study of a number of the well-known implementations of the Ray Casting algorithms was studied. A new method of numerical calculus is proposed for calculating the volume – the method of spheres, as well as a proposal for paralleling the algorithm on graphic accelerators in a linearly homogeneous computing environment using the block decomposition methods. For the artifacts control – algorithm of the cubic interpolation was used. The cloud server architecture was proposed. The work is done as a part of the PhD thesis of the author for non-profit/non-commercial, educational/research only reasons under the MIT License.
address:
- '[email protected]'
- National University of Science and Technology MISiS
author:
- Kondybayeva Almagul Baurzhanovna
bibliography:
- 'mybibfile.bib'
title: 'Interactive distributed cloud-based web-server systems for the smart healthcare industry'
---
`Data Science for Smart Healthcare`,healthcare ,internet of things ,computer science ,interactive visualization ,distributed ,web-server ,cloud ,parallel ,systems ,scientific research ,cloud technologies ,communications ,information ,cloud healthcare systems ,web-services ,web-technologies ,information systems ,experimental research ,DICOM ,CT ,computed tomography ,medical imaging ,volume render 68-00,97U99
Introduction
============
#### Introduction
One of the modern highly informative methods in medical diagnostics is the *computed tomography* technology (CT). *X-ray computed tomography* or CT is an imaging process that reproduces cross-sectional image representing the X-ray attenuation properties of the body. Three-dimensional reconstruction of CT image slices allows to visualize in three-dimensional space the localization of blood vessels, pathologies and other features, making this technique useful as an interactive visualization tool for the Healthcare purposes. In present days, the duration of the whole body computed tomography (with a slice layer thickness of less than $ \leq$ 1 mm) is about $ \sim$ 10-15 seconds, and the result of the study is from several hundred up to several thousand images [@Calhoun]. Actually modern *Multislice CT* (MSCT) is a method for volumetric studies of the entire human body (due to the the resulting axial transverse tomography constitute a three-dimensional data array) allowing to perform any reconstruction of images, including multi-plane reformation, volumetric (if necessary, stereo) visualization, virtual endoscopy [@Ding:2697996].
#### Aims of the study
The object of research is spatial data science in the fields of the medicine and proper healthcare industry in computed tomography (CT) of existing types in .DICOM format. The subject of the study are methods and algorithms of the high-quality visualization of medical spatial data, especially medical image tomography data, as well as methods and algorithms for morphological representation of surfaces and features on GPU/CPU systems including server side cloud system architecture with the following steps:
(a) to create the sustainable model of the homogeneous cloud based server architecture,
(b) to create the reliable model of the web-server,
(c) to create the representation for the volume data render tasks (including mobile operating systems).
*The objects of study* are the models of synthesis systems’ images for the three-dimensional models of human-machine communication in real time in the analysis of medical and biological spatial data, as well as the development and the study of models and architectures for the GRID computing and the cloud server hypervisor implementations. This goal requires the following tasks:
- to study the existing methods of the three-dimensional visualization in applied medicine, and science in general, an analysis of approaches for improving the quality and productivity of the 3D renderings,
- to develop existing methods for their complementary application in conditions of budgetary GPUs and the creation of 3D visualization technology for mass medical applications, including:
(a) modification of the method of the block decomposition of gigavoxel (more than $\sim$ $10 ^ {9}$ voxels) data for visualization on the GPU, preserving the possibilities and quality of visualization for not decomposed data (the possibility of using cubic visual interpolation, lighting, casting shadows, skipping blank areas, different integration conditions along the beam, ...),
(b) to develop methods to improve the performance of 3D visualization and suppression in the various kinds of render defects (artifacts) rational in GPU conditions,
(c) to develop a method for quantifying visualization quality for achieving the required quality and comparing the real (in compliance with a given quality) the effectiveness of the proposed methods.
- to study and develop an online prototype of the software package that implements the proposed methods on modern parallel hardware GPU architectures, and experimentally investigate the effectiveness of methods.
#### Render example
The Figure 1 represents the 3D volume render data visualization made with the cubic spline interpolation used in the study which is considered to be used as a basic algorithm for the volume render process in this system.
Methodology
===========
There are two main ways to visualize scalar field isosurfaces:
(a) to restore isosurface geometry (usually in polygonal form). In this case, the well-known marching cubes method is usually used. For surfaces reconstructed by this method, there is an algorithm for their effective compression. The main advantage of this approach is low hardware requirements: in contrast to the method of emitting rays, the process of rendering a polygon surface does not impose high requirements on the GPU, in addition, there is no need to store the initial field data for visualizing the surface. Another advantage of the approach is access to polygons, for example, for analysis of surface morphology. About the disadvantages, the duration of restoration of isosurface geometry can be noted [@Flohr],
(b) visualization using the method of emitting rays (the Ray Casting method) does not imply the restoration and preservation of the surface mesh in the RAM (the Random Access Memory). The method is to emit a beam for each pixel in the image in order to find collisions of the beam with the surface and determining the illumination of the surface at a given point, and thus determining the color of the pixel. Obviously, generating an image of an isosurface by this method is a much more resource-intensive process than rendering a polygonal surface, so the program to find collisions are usually performed on the GPU. In this case, the advantage is the lack of the need to restore polygons of the surface, allowing interactive adjustment of the value for the isosurface. Often, on modern video cards, the method of emitting rays is superior to the usual rendering of the polygonal mesh, both in quality and in performance. However, the need to store the original data array and high hardware requirements limit the use of this method [@Hadwiger].
The ray tracing method and its modification, called the method of the emission of the rays (Ray Casting – see Figure 2) allows to achieve the best quality and informative volume rendering. In this method, the color of each pixel of the desired image is calculated and the corresponding beam then is generated; which is a point in the space (for example, the position of the observer) and the direction of the ray. Moving in this direction with a certain step, the beam “accumulates” the color of the pixel. An important advantage of the ray emission method is that the algorithm is easy paralleling on the graphics processing unit (GPU) since each pixel of the desired image is processed independently of the rest. Medical image size *512\*512\*512* voxels can easily fit in GPU memory and the most efficient storage for such data is a three-dimensional texture; since the GPU provides texture access caching and automatic trilinear interpolation of data when sampling at an arbitrary point in space [@Engel]. In this work the cubic interpolation is being used. The following method is proposed in this study: modification of the method of block decomposition of data in the Ray Casting algorithm – different optimal sequence bypass blocks using volume optimized auxiliary structures of the method skipping empty areas (empty space leaping), providing the ability to build local lighting and shadows combined with tables of the previously integrated rendering. The task of three-dimensional visualization of scalar and vector fields in medical and scientific imaging in general: visualization of scalar fields defined on regular and irregular grids.
#### Voxel graphics
The method is based on the display process of three-dimensional textures in space. Naturally, texture elements (voxels) are usually translucent. As a rule, scalar fields are visualized in this way. The method of direct volumetric rendering, which is largely devoted to this work, is intended to visualize voxels as a translucent medium [@Hernell].
#### Isosurfaces
Isosurfaces are intended for visualization of three-dimensional scalar fields. Sometimes several isosurfaces are displayed at once; their color may depend on the field value [@Donatelli:2704061]. Three-dimensional texture is an acceptable repository for volumetric data rendering performed on graphic video cards using OpenGL library extensions. However, there are restrictions on the size of the texture: the maximum size is 5123 voxels [@Calhoun] [@Ding:2697996]. Using a block view of the data circumvents this limitation. For programming on the GPU, the shader language GLSL was chosen, because at the moment, the performance of implementations on CUDA and, especially, OpenCL is often inferior to shader implementations [@Dobashi:2710749]. The main limitation of shaders is the lack of access to shared memory on the GPU [@Calhoun]. This access allows you to group the rays into packets (slabs). Using Slab-based rendering allows decomposition in the image space, grouping the rays into packets with a shared shared memory, which cannot be accessed from shaders [@MargretAnouncia:2697932]. When decomposing data, different data blocks are written into different three-dimensional textures. Textures are the same size, with the exception of those that capture only part of the source data. Using blocks of different sizes and covering only the visible part of the data can give an additional performance boost and save GPU memory. The texture map method also reduces the size of the used GPU memory [@Calhoun]. In this implementation, there is a limitation to excluding completely empty blocks and fitting the bounding box for each block to the visible voxels. To avoid artifacts at the junction, adjacent blocks must overlap at least one voxel thick. This will be enough if trilinear interpolation is used in the rendering algorithm when sampling values from the data. In this implementation, the blocks overlap by a thickness of three voxels is used. It is used due to the reasons that, firstly, for the Fong local lighting model [@Donchin], to calculate the gradient, for which it is necessary to make additional samples from neighboring voxels, and secondly, to use cubic interpolation during the sampling instead of trilinear, including calculating the gradient, which extends the radius of the sample by one more voxel. There are also approaches for stitching data blocks having different spatial resolutions.
Figure 3 represents decomposing data into overlapping blocks. Here are the blocks overlap by a thickness of two voxels. Blocks marked in red and yellow have regular size (8 voxels in this case). The remaining blocks capture the remains data and are reduced in size if one wants to visualize three-dimensional discrete data array (tomogram) by volume rendering method, then to each possible data value one must set certain optical properties. Transfer function *T(x)* (transfer function), or palette, in the simplest case, matches any color value and transparency (usually in practice, the value is stored in memory opacity, and *0 - full transparency, 1 - maximum opacity)*. *Postclassification* - the operation scheme of the basis algorithm on render in which coloring (classification) of a point in space occurs after interpolating the value selected at the current point of the beam, i.e. color is determined as the interpolated data value. *Preclassification* - the operation scheme of the volume rendering in which the color of a space point is defined as color interpolation nearby voxels that are classified (as received color) according to their values, i.e. coloring occurs before interpolation [@Calhoun]. If one visualizes the array as a lot of multi-colored translucent cubes, i.e. if interpolation between data cells (voxels) is not used, then the result of the 3D rendering will contain crude artifacts. As for *“smoothness”* of visualization it is necessary to use interpolation between the nodes of the source (data) grid; and with it two different approaches may arise. In the most widespread cases the postclassification method is being used [@Calhoun] [@Tarini].
The optical properties of a point in space is first calculating the interpolating value of *V* data at a point (usually using trilinear interpolation [@Calhoun]), the value of the transfer function is considered as point *V*, i.e. *T(V)* even if the interpolated value of *V* does not occur at all in the histogram of the source data . On the contrary, in preclassification process, all voxels are “painted” at first, but beyond optical properties, the arbitrary point in space is considered as the result of the interpolation between the optical properties (color with transparency) of voxels, i.e. classification occurs before interpolation, not after [@Hernell] [@Hadwiger] [@Tarini] [@Sattler]. In practice, a model with postclassification is more often used, because it gives:
(a) significantly improving visualization quality: significantly less noticeable stepping data comparing to the preclassification with tricubic interpolation for rendering (preclassification gives the same artifacts),
(b) better performance or better resource intensity: on practice for preclassification method each voxel and its optical properties must be either stored in the memory, or calculate these voxels’ properties during rendering.
In the first case instead of a 12-bit array with the source data in the GPU one shall need to load a 32-bit array of the same dimensions, which will store the color and voxel transparency instead of the original data values. And when changing the transfer function it is needed to re-calculate and load the entire array, instead of loading a new transfer function. In the second case one has to make samples of eight voxels (in the case of trilinear interpolation) calculate their colors and transparency and then find the interpolated value of color and transparency for the point. Whereas in postclassification one makes one sampling from a point using trilinear filtering of textures, which practically gives hardware for free trilinear interpolation [@Shanmugam] [@Stewart] [@Besset:768962].
In this work the preference is given to the tricubic spline interpolation using postclassification algorithm due to the results from the comparison presented on the Figure 4.
#### Analysis
Analysis of the problems of implementing the method of block decomposition under conditions of maintaining high quality 3D visualization.
#### Pros
:
(a) the ability to load large amounts of data. In addition, when divided into blocks of 643 voxels, one can save the blocks that do not contain useful information. For example, for CT scans, one can usually drop $ \sim$ 40 $ \%$ of the blocks, due to the reason that these blocks contain only air. The size of the visualized data is limited only by the capacity of the video card (in the case of a consumer video card with a memory of only 1 GB can visualize an array of data up to *512\*512\*2048* bytes in size).
(b) decomposition significantly improves visualization performance, due to the reason that during the rendering of a single small block, the selection comes from a small texture, which is much faster than selecting from a large texture. Thus, blocks are displayed in order from the observer, some of the blocks can be closed by previously drawn blocks. The early beam completion strategy is also applicable to block data rendering. In addition to saving GPU memory, it also saves space that you have to trace with rays. As in the case of saving memory, saving space depends on the choice of block size and, as a rule, the smaller the size, the greater the saving. In the block presentation of data, the strategy of skipping empty areas is applicable with greater efficiency, since the data is divided into small arrays, each of which is easier to fit the bounding box than to the data array as a whole. Blocking provides additional opportunities for paralleling rendering to multiple GPUs, which is especially important for the client-server architecture of the visualizer.
(c) despite the need to mix the rendering results of various blocks in a specific order, each individual block can be rendered independently of the others on any GPU. Computing and data can be distributed across different GPUs. There are works on distributed visualization by the method of volume rendering on cluster systems, which is especially important for scientific visualization, when as a result of numerical modeling huge data arrays are obtained, which cannot be transferred to one local machine.
#### Cons
:
(a) the main drawback of the block representation is the complexity of the rendering algorithm: when rendering a block, nothing is known about data from neighboring blocks (except for the overlap layer with neighboring blocks); if for each block to provide such access there is a need to access to the remaining blocks that process is able to negate the performance gain. Thus, for example, the implementation of casting shadows and various nonlocal lighting techniques [@Hernell] [@Hadwiger] [@Sattler] [@Shanmugam] [@Stewart] [@Tarini] [@Behrens] [@Zhukov] [@Desgranges] [@Hernell:01] [@Fairchild], including techniques requiring the generation of secondary rays [@Wyman] [@Rezk-Salama] [@Weiskopf] [@Magnor], becomes more complicated [@Kniss] [@Jensen].
(b) It is worth to highlight the complication of the multi-volume rendering algorithm, in which it is necessary to perform joint rendering of two or more spatial data arrays overlapping in space, each of which has a block representation,
(c) block overlapping means that the voxels on the floors will be duplicated, so if the partition is too small (with block sizes less than 323), the GPU memory savings will be ineffective,
(d) if the block is too small, the total time spent on “switching” between the blocks noticeably increases; after rendering the next block, it is necessary to copy the rendering results from the texture (which was performed on rendering) into the texture from which the reading will be carried out during the rendering of the following blocks. Even when copying only the area of the screen where the block was rendered, the performance already drops significantly when the regular block size is 323 voxels [@Gavrilov].
The work on slab-based rendering [@Mensmann] is very interesting as a study demonstrating not only a new method of slicing blocks along a beam to reduce misses for a ray packet in shared memory, yet also demonstrating inconstancy of gain and its limited amplitude ($ \sim 30 \%$); also shows the difficulty of supplementing the data between the blocks due to the above mentioned pros and cons.
#### The method of Spheres
The method of spheres for the analysis of the morphology of complex biological objects in *SVR* values: *Surface-to-Volume Ratio*, *SVR* or *Area-to-Volume Ratio*. It is one of the most important characteristics of all biological objects from the cell to the animal as a whole. This value characterizes the intensity of exchange between the biological object as a whole with the environment and has a characteristic dependence on the radius *(R)* of the object as *$ \dfrac{1}{R}$*. With the same measure, you can approach the characterization of the local properties of the object. Therefore, it is very important to have a quantitative local characteristic morphology of the studied object in *SVR* values ($ SVR = \dfrac{S}{V} $) in order to discover, understand and explore the functions performed by its parts and organs. This is especially relevant to a challenge with brain cells such as astrocytes, due to the extraordinary complexity of their shape and relatively poor knowledge[@Donchin]. However, to date has not formed sustainable methodology for calculating local *SVR*s. In addition, there is a problem. Settlements in an acceptable time, as computational complexity of the analysis task local *SVR* is proportional to the square of the number of vertices or triangles $ \sim (O = n^{2})$, and astrocytes reconstructed from electron micrographs (a microscope with a resolution of units of nanometers) have in its composition hundreds of thousands $ n \sim (10^{5}) $ triangles[@Gavrilov].
#### Surface-to Volume approach
Let’s define a scalar field *SVR* *(X)* in 3D space as following: let *X* is an arbitrary point in space, *G* is the surface of the object under study (in our case it is a closed polygonal surface). Then construct a sphere *$\Omega$* of radius *R* with center at point *X*. Let the intersection of *$\Omega$* and *G* be nonzero, then let *S* be the area the surface of an object lying inside the sphere *$\Omega$* (denote this surface by *S*), and *V* as it is the intersection volume of the domains *$ {V_{G}}$* and *$V_{\Omega}$* bounded by *G* and *$\Omega$* (denote this is the locus of points as *$\Theta$*. Then the value of the field *SVR* *(X)* is calculated as (1) and presented on the Figure 3: $$SVR(X) =\frac{S}{V}$$ In the case *$ V = 0$*, i.e. if *$ {\Omega \cap G = 0}$*, then the field at the point *X* is not defined.
As a result, one defines the method of spheres as the method that consists in calculating the field *SVR (X)* at points *X* belonging to the boundary of the investigated part of the object. A significant influence on the field values in this method is a choice the radius of the sphere. To calculate *SVR (X)*, in accordance with the formulated by definition (1), it is necessary to calculate the area *S* and the volume of only the part *G* located inside the sphere $ \Omega$. Area *S* is the sum of the areas of the triangles located entirely inside $ \Omega$, and from the cut-off part of the areas of triangles intersected by a sphere $ \Omega$. The area of the part of the triangle inside $ \Omega$ (see Figure 5 – b) is calculated by the Monte Carlo method [@Sobol]. In a similar way, the volume *G* inside $ \Omega$. However, there are difficulties that should be considered.
#### Cloud architecture
Cloud data storage - an online storage model in which data is stored on numerous servers distributed on the network and provided for use by customers, mainly by a third party. In contrast to the model for storing data on its own dedicated servers purchased or leased specifically for such purposes, the number or any internal structure of the servers is generally not visible to the client. Data is stored and processed in the so-called cloud, which represents, from the user’s viewpoint, as one large virtual server. Physically, such servers can be located remotely from each other geographically, up to the location on different continents.
#### Types of Cloud Computing
The concept of a cloud computing is often associated with such service-providing (i.e. everything as a service) technologies, such as:
(1) “Infrastructure as a Service” (“Infrastructure as a Service” or “IaaS”),
(2) “Platform as a Service” (“Platform as a Service”, “PaaS”),
(3) “Software as a Service” (“Software as a Service” or “SaaS”).
The solution in the study is developing as an open-source solution based on a SaaS model.
SaaS is an application deployment model that involves delivering an application to the end user as an on demand service. Access to such an application is carried out through the network, and most often through an Internet browser. In this case, the main advantage of the SaaS model for the user is the lack of costs associated with installing, updating and maintaining the use of the equipment and software running on it. In this SaaS model presented on the Figure 6 includes the following points:
(1) the application is studied to be adapted for the remote use,
(2) multiple clients can use one application,
(3) the application can be upgraded by developers smoothly and transparently to users.
In fact, SaaS software can be seen as a more convenient and cost-effective alternative to internal information systems. The development of SaaS logic is the concept of WaaS (Workplace as a Service - workplace as a service). That is, the client receives at his disposal a virtual workstation fully equipped with all the necessary software. Here on the Figure 6 the approximate cloud architecture is presented.
Results
=======
There is an arising significant complication of the visualization algorithms coming from the block decomposition of the data. In practice, it’s necessity inevitably arises not only for visualizing large arrays but also to speed up rendering, especially on graphics cards with low performance. Regrading this, the method of the block decomposition of data for volume visualization based on the ray emission algorithm (Ray Casting method) is still being studied. Based on the study, a method for quantifying *DVR* artifacts in a method for ray emission (Ray Casting method) caused by an insufficiently short beam pitch in the form similar to the peaks of the signal-to-noise ratio (PSNR - Peak Signal-to-Noise Ratio) is still being developed. According to the results from the Ray Casting processing in volume render: using the *PSNR* ratio brings the noise level to a logarithmic scale in dB, which values from 30 to 40 dB corresponding to the acceptable image synthesis quality. The causes of artifacts arising from the trilinear interpolation are investigated further as much as the cubic interpolation (and either their approaches to the quantitative assessment involving the assessment of structural similarity of two *SSIM* images). A new method using the postclassification for eliminating volumetric errors (artifacts) – visualizations characterized by using pre-integrated rendering in a virtual data sampling method when integrated along the beam, which is optimal for a class of visualization cases where tricubic interpolation and local lighting is proposed, still investigated. A study was conducted on the 3D visualization methods based on the Ray Casting (RC) algorithm. Although no RC algorithm was found in the course of the experiments that was optimal in terms of quality and performance under any visualization conditions, the evaluation method showed the necessity of optimization: due to the reason that the not optimized *UDVR* approach is inferior to other approaches in any visualization conditions, despite its high performance.
Applying the approaches to volume render of medical image data studied in this work, it was managed to design the 3D visualization software prototype and achieve interactive and high-quality volumetric visualization of medical image of about 2 GB in size (512x512x5382 voxels) on the localhost machines. Also, the design for the cloud based grid machines was developed. The solution mostly implements the hypervisor technologies for the grid server computing. A method for implementing a fully functional system based on GRID cloud computing servers is proposed, examples of client-side are shown on the Figures 7, 8, 9.
Discussion
==========
Today there are variety of the approaches developed in the works of the following scientists: Klaus Engel, Bernhard Kainz, Daniel Ruijters, Stefan Guthe, Johanna Beyer, Vincent Vidal, Markus Hadwiger, Daniel Weiskopf, Thomas Ertl, Wolfgang Strasser, Byeonghun Lee, Jihye Yun, Jinwook Seo , Yeong-Gil Shin, Bohyoung Kim, Byonghyo Shim, and others, allowing to process the real-time volumetric visualization using GPU computing. The world market offers several tomography software systems that provide fusion and three-dimensional visualization of tomograms. These commercial systems use the most productive versions of the commercial 3D visualizer systems [@Tractica]. The same can be said about the growth of information flow and the need to build productive methods for processing it in the high-energy physics. The most critical situation, due to the huge amount of data, is observed in three-dimensional processing of data from particle accelerator detectors with a resolution of a few picoseconds ($ \sim10^{-12}$ seconds) [@Cheung:1495074]. For various mathematical models, it is necessary to visualize the obtained data in such a way that certain field properties are revealed. For example, for the result of numerical simulation of unsteady fluid flows, these properties can be revealed:
(1) dynamics of the velocity field,
(2) the formation and decay of vortices,
(3) vortex flows and shock waves.
Of particular difficulty is the task of constructing animations for vector velocity fields (both steady and unsteady flows). Texture methods for solving this problem allow to obtain high-quality results in the visualization of two-dimensional flows. The pinnacle of their development are methods based on the construction of a Motion Map for stationary flows. However, when trying to apply them to the three-dimensional flows, a number of problems arise, associated primarily with the high density of texture data, and also, importantly, with large computational costs. The bottleneck of the well-known three-dimensional texture methods is the construction of interactive animation with high quality animated paintings. Due to the high computational complexity, it is possible to build only a certain animation sequence for the purpose of its subsequent visualization. Despite significant progress in solving the problems mentioned above, there are a number of unsolved problems:
(1) high-quality three-dimensional visualization of medical images today is “tied” to the tomograph due to the high demands on the productivity of the workstation, yet it is not available to the ordinary clinician and, especially, the medical students. Thus, a transition to the online software for the mass accessibility technology without the loss of visualization quality is necessary,
(2) the volume of the medical image available for the 3D reconstruction on the GPU is limited by the size of the GPU memory (today for the mass office video cards that are for sale the volume is about $ \sim$ 1-2 GB), while the constant growth of data requires the removal of restrictions on their volume and the construction of decomposition algorithms for parallel block data processing while preserving all the capabilities and quality of visualization on the side of user services,
(3) despite the increase in productivity and quality of 3D visualization, there is no practice of quantitatively assessing the quality of visualization; as for the gaming technologies, for example, the so-called light distribution calculation method is used [@Magro]
(4) There are several open and commercial programs for semi-automatic three-dimensional geometric reconstruction of cells, but there are no methods and programs for automating detailed morphological analysis of cells, while the computational complexity of such analysis in SVR (Surface-to-Volume Ratio) values is proportional to the square of the number of vertices of triangles $O^{2}$ with characteristic values of approximately $\textsl{10*5}$ triangles and deviation variance of the vortex data given under the volume render [@MargretAnouncia:2697932].
#### Software rights
All rights on the software are under the MIT License.
{#section .unnumbered}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We provide a unified framework for modeling LIBOR rates using general semimartingales as driving processes and generic functional forms to describe the evolution of the dynamics. We derive sufficient conditions for the model to be arbitrage-free which are easily verifiable, and for the LIBOR rates to be true martingales under the respective forward measures. We discuss when the conditions are also necessary and comment on further desirable properties such as those leading to analytical tractability and positivity of rates. This framework allows to consider several popular models in the literature, such as LIBOR market models driven by Brownian motion or jump processes, the Lévy forward price model as well as the affine LIBOR model, under one umbrella. Moreover, we derive structural results about LIBOR models and show, in particular, that only models where the forward price is an exponentially affine function of the driving process preserve their structure under different forward measures.'
address:
- 'Center for Mathematics, Technical University of Munich, Parkring 11, 85748 Garching b. München, Germany'
- 'Laboratoire de Probabilit[é]{}s et Modèles Al[é]{}atoires, Universit[é]{} Paris Diderot, 75205 Paris Cedex 13, France'
- 'Institute of Mathematics, TU Berlin, Straße des 17. Juni 136, 10623 Berlin, Germany'
author:
- Kathrin Glau
- Zorana Grbac
- Antonis Papapantoleon
bibliography:
- 'references.bib'
title: A unified view of LIBOR models
---
\[section\] \[theorem\][Lemma]{} \[theorem\][Proposition]{} \[theorem\][Corollary]{} \[theorem\][Claim]{} \[theorem\][Definition]{}
\[theorem\][Remark]{} \[theorem\][Note]{} \[theorem\][Example]{} \[theorem\][Assumption]{}
\#1\#2\#3
[^1]
[^1]: Financial support from the PROCOPE project “Financial markets in transition: mathematical models and challenges” and the Europlace Institute of Finance project “Post-crisis models for interest rate markets” is gratefully acknowledged.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'S. Chakraborty,'
- 'R. S. Hansen,'
- 'I. Izaguirre,'
- 'and G. G. Raffelt'
title: 'Self-induced neutrino flavor conversion without flavor mixing'
---
=1
Introduction {#sec:introduction}
============
Most of the energy liberated in stellar core collapse or in neutron-star mergers appears in the form of neutrinos and antineutrinos of all flavors, but with fluxes and spectra that differ strongly between $\nu_e$, $\bar\nu_e$ and the other species, collectively referred to as $\nu_x$. The subsequent flavor evolution of these neutrinos influences energy deposition beyond the decoupling region, neutrino-driven nucleosynthesis, and detection opportunities of the neutrino signal from the next nearby supernova or the diffuse supernova neutrino flux from all past core-collapse events [@Janka:2012wk; @Scholberg:2012id; @Mirizzi:2015eza]. However, a true understanding of flavor evolution in dense environments has remained elusive because of many complications engendered by the nonlinear nature of collective flavor oscillations [@Mirizzi:2015eza; @Chakraborty:2016yeg]. We study a new item on this list which has eluded most workers in this field with the notable exception of Ray Sawyer [@Sawyer:2005jk; @Sawyer:2015dsa], i.e., the surprising insight that collective flavor conversion need not depend on neutrino mixing parameters.
Collective neutrino flavor oscillations manifest themselves in the form of two generic phenomena. One is the effect of synchronisation: different modes of the neutrino mean field oscillate in lockstep even though they have different vacuum oscillation frequencies $\omega=\Delta m^2/2E$ [@Kostelecky:1994dt; @Pantaleone:1998xi; @Pastor:2001iu; @Raffelt:2010za; @Akhmedov:2016gzx]. The other is the phenomenon of self-induced flavor conversion, corresponding to collective run-away modes [@Kostelecky:1993dm; @Samuel:1995ri; @Duan:2005cp; @Hannestad:2006nj; @Duan:2007mv; @Banerjee:2011fj; @Raffelt:2011yb]. Surprisingly, the growth rate in the linear regime and the overall evolution need not depend on $\Delta
m^2/2E$ and therefore the effect can occur even for unmixed neutrinos if given an appropriate seed to grow from [@Sawyer:2005jk; @Sawyer:2015dsa].
Collective “flavor conversion” actually does not represent any change of flavor in the overall ensemble, but a reshuffling among different modes. In the simplest case, a gas of $\nu_e$ and $\bar\nu_e$ can convert to $\nu_\mu$ and $\bar\nu_\mu$ without change of lepton number or flavor-lepton number. Such pair processes certainly occur in the form of non-forward scattering with a rate proportional to ${G_{\rm F}}^2$, but can also occur on the refractive level with a rate proportional to ${G_{\rm F}}$. For most cases studied in the literature, the conversion rate was actually found to be of order $\Delta
m^2/2E$ instead, i.e., driven by the frequency $\omega$. Another possible driving frequency is the neutrino-neutrino interaction energy $\mu=\sqrt{2}
G_{\rm F}n_{\nu}$. The very definition of a “dense” neutrino gas is precisely that $\mu\gg\omega$. However, this dominant scale cancels when the neutrino and antineutrino angle distributions are too similar. On the other hand, with sufficiently different angle distributions the conversion rate can be driven by $\mu\gg\omega$, corresponding to much faster conversion. Moreover, these fast conversions can exist even without any vacuum frequency $\omega$ and thus in the absence of neutrino masses.
In general, therefore, self-induced flavor conversion—in the sense of flavor reshuffling among modes—can occur without flavor mixing, provided there exist fluctuations in flavor space to seed the unstable modes. One may speculate that even quantum fluctuations of the mean-field quantities could suffice as seeds. However, in practice ordinary neutrino oscillations driven by their masses and mixing parameters exist, so disturbances in flavor space to seed self-induced flavor conversion always exist even on the mean-field level.
The main purpose of our paper is to present a few simple examples which illustrate these general points and which are even more basic than those presented by Sawyer. We begin in section \[sec:beam\] with the simplest possible case, two colliding beams of neutrinos and antineutrinos, which shows fast flavor conversion if we allow for inhomogeneities. In section \[sec:homogeneous\] we also construct a homogeneous example, consisting of four modes in the form of two beams intersecting with a nonvanishing angle. We finally turn in section \[sec:two-bulb\] to the example of a spherical source which emits neutrinos and antineutrinos with different zenith-angle distributions in analogy to the schematic supernova model proposed by Sawyer [@Sawyer:2015dsa]. We conclude with a discussion and outlook in section \[sec:conclusion\].
Colliding beams {#sec:beam}
===============
The current-current structure of the low-energy neutrino-neutrino interaction implies that we need at least two different propagation directions to obtain any effects at all. Therefore, the simplest possible example is an initially homogeneous gas of neutrinos and antineutrinos, allowing only for two opposing directions of motion, i.e., a system that is one-dimensional in momentum space and that we can view as two colliding beams (figure \[fig:beam-setup\]). This type of simple model was recently used by several groups to study the impact of spontaneously breaking various symmetries [@Raffelt:2013isa; @Hansen:2014paa; @Chakraborty:2015tfa; @Duan:2014gfa; @Abbar:2015mca; @Mangano:2014zda; @Mirizzi:2015fva].
Linearized equations of motion
------------------------------
On the refractive level, the interacting neutrino system is best represented in terms of the mean-field $\varrho_i$ for every momentum mode $i$. The diagonal components of this matrix in flavor space are phase-space densities of the different flavor states, whereas the off-diagonal elements represent flavor correlations. Antineutrinos are represented by negative energies and we use the “flavor isospin convention,” where the $\varrho$ matrices for antineutrinos correspond to negative phase-space densities. The advantage is that we do not need to distinguish explicitly between neutrino and antineutrino modes. The $\varrho$ matrices of $N$ modes evolve according to [@Sigl:1992fn] $$\label{eq:EOM1}
{{\rm i}}{\left(}\partial_t+{\bf v}_i\cdot{\bm \nabla}{\right)}\,\varrho_i
={\left[}{\sf H}_i,\varrho_i{\right]}\,,
\quad\hbox{where}\quad
{\sf H}_i=\frac{{\sf M}^2}{2E_i}
+\mu\sum_{j=1}^N{\left(}1-{\bf v}_i\cdot{\bf v}_j{\right)}\varrho_j\,,$$ where ${\sf M}^2$ is the matrix of neutrino mass-squares. We assume neutrinos to be ultra-relativistic so that the velocities ${\bf v}_i$ are unit vectors giving the directions of the individual modes. The neutrino-neutrino interaction energy is $\mu=\sqrt{2}{G_{\rm F}}n_\nu$ with the effective neutrino density $n_\nu={{\textstyle\frac{1}{2}}}(n_{\nu_e}+ n_{\bar\nu_e}-n_{\nu_x}-n_{\bar\nu_x})$. We always consider two-flavor oscillations between $\nu_e$ and another flavor $\nu_x$ which is a suitable combination of $\nu_\mu$ and $\nu_\tau$. These conventions follow reference [@Chakraborty:2014lsa] and are chosen such that a fixed $\mu$ corresponds to a fixed density of $\nu_e$ plus $\bar\nu_e$, even when we modify, for example, their relative abundance. In much of the previous literature, instead either the number of $\nu_e$ or of $\bar\nu_e$ was held fixed, but we here prefer a more symmetric definition.
![Initially homogeneous ensemble of four neutrino modes (“colliding beams” of neutrinos and antineutrinos). The system is taken to be infinite in all directions. The normalized $\nu$ flux is $1+a$, the $\bar\nu$ flux $1-a$ with the asymmetry parameter $a$ in the range $-1\leq a\leq+1$. The left-right asymmetry is parametrized by $b$ such that the upper beam in this figure has normalized strength $1+b$, the lower beam $1-b$ with $-1\leq b\leq+1$. The relation of parameters $r$, $\bar{r}$, $l$ and $\bar{l}$ to parameters $a$ and $b$ can be found in equation (\[Eq:param\]).[]{data-label="fig:beam-setup"}](fig1.pdf){width="40.00000%"}
In order to identify unstable modes, we consider a linearized version of equation (\[eq:EOM1\]). We first note that ${\rm Tr}\,\varrho_i$ is conserved by flavor conversion and, if the system was initially homogeneous, it is not modified by the transport term in equation (\[eq:EOM1\]). It is convenient to define traceless normalized $\varrho$ matrices in the form $$\varrho_i-\frac{1}{2}\,{\rm Tr}\,\varrho_i
=\frac{g_i}{2}\begin{pmatrix}s_i&S_i\\ S_i^*&-s_i\end{pmatrix}\,,$$ where $s_i$ is a real and $S_i$ a complex number with $s_i^2+|S_i|^2=1$. Moreover, if neutrinos are initially prepared in $\nu_e$ or $\bar\nu_e$ eigenstates (our usual example), then initially $s_i=+1$. The “spectrum” $g_i$ gives the actual density of neutrinos in mode $i$ and is positive for an initial $\nu_e$ and negative for an initial $\bar\nu_e$, corresponding to our flavor-isospin convention. Our definition of the effective neutrino density $n_\nu$ corresponds to the normalization $\sum_{i=1}^N|g_i|=2$. To linear order, $s_i=1$ remains constant, whereas the off-diagonal elements evolve according to $$\label{eq:EOM2}
{{\rm i}}{\left(}\partial_t+{\bf v}_i\cdot{\bm \nabla}{\right)}\,S_i=
\left[\omega_i+\mu\sum_{j=1}^N {\left(}1-{\bf v}_i\cdot{\bf v}_j{\right)}g_j\right]S_i
-\mu\sum_{j=1}^N{\left(}1-{\bf v}_i\cdot{\bf v}_j{\right)}g_j S_j\,.$$ We have assumed a very small vacuum mixing angle and use $\omega_i=\Delta
m^2/2E_i$ with $\Delta m^2$ positive and the convention that $\omega_i$ is positive for neutrinos and negative for antineutrinos.
As a next step, we transform this linear equation of the space-time coordinates $(t,{\bf r})$ into Fourier space $(\Omega,{\bf k})$ and we write $S_i(t,{\bf r})=Q_i(\Omega,{\bf k})\,e^{-{{\rm i}}(\Omega t-{\bf k}\cdot{\bf r})}$. The linearized equations of motion in Fourier space are $$\label{eq:EOM3}
\Omega\,Q_i=
\left[\omega_i+{\bf v}_i\cdot{\bf k} +\mu\sum_{j=1}^N
{\left(}1-{\bf v}_i\cdot{\bf v}_j{\right)}g_j\right]Q_i
-\mu\sum_{j=1}^N{\left(}1-{\bf v}_i\cdot{\bf v}_j{\right)}g_j Q_j\,.$$ We are looking for exponentially growing solutions, i.e., eigenvalues $\Omega$ with a nonvanishing imaginary part.
We finally turn to a system which is one-dimensional in momentum space (a beam) so that ${\bf v}_i\to v_i$ and ${\bf k}\to k$. The current-current factors $(1-{\bf v}_i\cdot{\bf v}_j)$ are 0 for parallel-moving modes or 2 for opposite moving ones. We consider four modes as in figure \[fig:beam-setup\] and use the vacuum oscillation frequency $+\omega$ for neutrinos and $-\omega$ for antineutrinos. We denote the amplitudes $Q_i$ for the different modes with $R$ for right-moving neutrinos, $\bar R$ for right-moving antineutrinos, and $L$ and $\bar L$ for the left movers. Likewise, we denote the mode occupations $g_i$ with $r$, $\bar r$, $l$ and $\bar l$, respectively. Equation (\[eq:EOM3\]) becomes $$\label{eq:EOM4}
\Omega\,\begin{pmatrix}R\\ \bar L\\ L\\ \bar R\end{pmatrix}=\left[
\begin{pmatrix}\omega+k&0&0&0\\0&-\omega-k&0&0\\0&0&\omega-k&0\\0&0&0&-\omega+k\end{pmatrix}
+2\mu
\begin{pmatrix}
l+\bar l&-\bar l&-l&0\\
-r&r+\bar r&0&-\bar r\\
-r&0&r+\bar r&-\bar r\\
0&-\bar l&-l&l+\bar l\\
\end{pmatrix}
\right]\begin{pmatrix}R\\ \bar L\\ L\\ \bar R\end{pmatrix}\,.$$ This is the most general one-dimensional four-mode case and the starting point for the discussion in the rest of this section.
Two modes
---------
The possible existence of unstable modes for $\omega=0$ is most easily understood in a yet simpler case consisting only of right-moving neutrinos and left-moving antineutrinos, i.e., $r=1+a$, $\bar l=-(1-a)$, and ${\bar
r}=l=0$. The parameter $-1\leq a\leq1$ codifies the neutrino-antineutrino asymmetry of our system. Equation (\[eq:EOM4\]), reduced to the two occupied modes, becomes $$\label{eq:EOM5}
\Omega\,\begin{pmatrix}R\\ \bar L\end{pmatrix}
=\left[\begin{pmatrix}\omega+k&0\\0&-\omega-k\end{pmatrix}+2\mu
\begin{pmatrix}-1+a,&1-a\\-1-a,&1+a\end{pmatrix}
\right]\begin{pmatrix}R\\ \bar L\end{pmatrix}\,.$$ Without further calculation we can see that the role of the vacuum oscillation frequency $\omega$ is here played by $\omega+k$. If we consider vanishing neutrino masses ($\omega=\Delta m^2/2E=0$) and a spatial Fourier mode $k>0$, the role usually played by $\omega$ will be taken over by $k$. The reason for this behavior is that we have constructed our system such that neutrinos (vacuum frequency $+\omega$) move right so that the spatial Fourier term ${\bf
v}\cdot{\bf k}$ enters as $+k$, and the other way round for the beam of left-moving antineutrinos.
For completeness we provide the explicit eigenvalues for this two-mode case. Using the notation $\tilde\omega=\omega+k$ we find $$\Omega=2a\mu\pm\sqrt{(2a\mu)^2+\tilde\omega(\tilde\omega-4\mu)}\,.$$ For $\tilde\omega=0$ the eigenvalues are purely real. The eigenfrequencies have an imaginary part for $$1-\sqrt{1-a^2}<\frac{\tilde\omega}{2\mu}<1+\sqrt{1-a^2}\,.$$ Notice that in this system, the solutions for positive or negative $k$ are different. Because $\mu$ is defined to be positive, we have unstable solutions only for $\tilde\omega>0$, i.e., for $k>-\omega$. The imaginary part has its maximum for $\mu=\tilde\omega/(2a^2)$ and is $${\rm Im}\,\Omega\vert_{\rm max}
=\tilde\omega\,\sqrt{\frac{1}{a^2}-1}\,.$$ Therefore, in the homogeneous case ($k=0$), the growth rate is indeed proportional to the vacuum oscillation frequency times a numerical factor which depends on the neutrino-antineutrino asymmetry. Conversely, for vanishing $\omega$, the maximum growth rate is proportional to the spatial Fourier wave number $k$.
We can turn this discussion around and ask which Fourier modes $k$ are unstable for a fixed $\mu$ value. The maximum growth rate occurs for $\tilde\omega=2\mu$. The fastest-growing $k$ mode grows with the rate $${\rm Im}\,\Omega\vert_{\rm max}
=2\mu\,\sqrt{1-a^2}\,.$$ This rate is indeed “fast” in the sense that it is proportional to $\sqrt{2}{G_{\rm F}}n_\nu$. Of course, we assume that $-1<a<1$ is not fine-tuned to be very close to $\pm1$ which would correspond to having only neutrinos or only antineutrinos in the system.
Four modes
----------
The results of the previous section came about because the system was constructed with maximal left-right asymmetry: neutrinos were only right-moving, antineutrinos left-moving. On the other hand, previous one-beam examples [@Raffelt:2013isa; @Hansen:2014paa; @Chakraborty:2015tfa; @Duan:2014gfa; @Abbar:2015mca; @Mangano:2014zda; @Mirizzi:2015fva] had been constructed to be left-right symmetric, although spontaneous symmetry breaking of the solution was allowed. In all previous cases, the colliding beams were stable against self-induced flavor conversion in the $\omega=0$ limit even for nonvanishing $k$. Therefore, the system must be prepared with some amount of left-right asymmetry to be unstable for vanishing $\omega$.
To study this condition, we now turn to a four-mode system with left- and right-moving neutrinos and antineutrinos. As before, we use the parameter $a$ to denote the neutrino-antineutrino asymmetry, and in addition the parameter $-1\leq b \leq 1$ to denote the left-right asymmetry. Specifically, we use the beam occupations
\[Eq:param\] $$\begin{aligned}
r &=& +{{\textstyle\frac{1}{2}}}(1+a)(1+b)\,,\\
\bar l&=& -{{\textstyle\frac{1}{2}}}(1-a)(1+b)\,,\\
l &=& +{{\textstyle\frac{1}{2}}}(1+a)(1-b)\,,\\
\bar r&=& -{{\textstyle\frac{1}{2}}}(1-a)(1-b)\,.\end{aligned}$$
The two-mode example of the previous section corresponds to $b=1$. The neutrino-neutrino interaction matrix in equation (\[eq:EOM4\]) becomes $$\mu
\begin{pmatrix}
2\,(a-b)&(1-a)(1+b)&~-(1+a)(1-b)&0\\
-(1+a)(1+b)&2\,(a+b)&~0&(1-a)(1-b)\\
-(1+a)(1+b)&0&~2\,(a+b)&(1-a)(1-b)\\
0&(1-a)(1+b)&~-(1+a)(1-b)&2\,(a-b)\\
\end{pmatrix}\,.$$ The eigenvalue equation is now quartic and the explicit solutions provide little direct insight.
However, in several special cases there are simple solutions. In the previous literature, one always used a system which was set up in a left-right symmetric configuration, meaning $b=0$. Considering the homogeneous case ($k=0$), one finds the eigenvalues $$\Omega=a\mu\pm\sqrt{(a\mu)^2+\omega(\omega-2\mu)}
\quad\hbox{and}\quad
\Omega=3a\mu\pm\sqrt{(a\mu)^2+\omega(\omega+2\mu)}\,.$$ The first solution corresponds to the usual “flavor pendulum” for inverted neutrino mass ordering, the second to the left-right symmetry breaking solution for normal ordering as discussed previously [@Raffelt:2013isa]. Notice that changing the mass ordering corresponds to ${\bf B}\to-{\bf B}$ in equation (\[eq:EOM1\]) and thus to $\omega\to-\omega$ in these expressions.
The corresponding inhomogeneous case ($|k|>0$) was studied in references [@Duan:2014gfa; @Chakraborty:2015tfa]. The system is stable for $\omega=0$ and the growth rate is proportional to $\omega$. Explicit results can be derived in the $k\to\infty$ limit [@Chakraborty:2015tfa].
Turning now to the left-right asymmetric case, one can actually find explicit solutions for $\omega=0$, $$\Omega=2a\mu\pm(k-2b\mu)
\quad\hbox{and}\quad
\Omega=2a\mu\pm\sqrt{(2a\mu)^2+k(k-4b\mu)}\,.$$ All solutions are real in the homogeneous case ($k=0$) and for any $k$ in the left-right symmetric system ($b=0$). In the general case, the second eigenvalue has an imaginary part if the expression under the square root is negative, which occurs for $$b-\sqrt{b^2-a^2}
<\frac{k}{2\mu}<
b+\sqrt{b^2-a^2}\,.$$ An instability exists only for $a^2<b^2$, i.e., the left-right asymmetry must exceed the neutrino-antineutrino asymmetry. For fixed $\mu$, the Fourier mode with the largest growth rate is $k=2b\mu$, growing with a rate $${\rm Im}\,\Omega\vert_{\rm max}
=2\mu\,\sqrt{b^2-a^2}\,.$$ Again, this growth is of order a typical neutrino-neutrino interaction energy and thus “fast”.
Intersecting beams {#sec:homogeneous}
==================
The crucial ingredient to obtain fast flavor conversion appears to be a sufficient difference between the direction distribution of neutrinos and antineutrinos. However, if the momentum distribution is one dimensional we need spatial inhomogeneities. As a next step we construct the simplest homogeneous system (${\bf k}=0$) that shows fast flavor conversion. We consider four modes with directions which intersect at a relative angle $\theta$ as shown in figure \[fig:twobeam\]. We continue to denote the modes as “left- and right-moving” in an obvious sense. The mode occupations are taken to be symmetric between left and right, but we include a neutrino-antineutrino asymmetry $a$, i.e., the mode occupations are taken to be
$$\begin{aligned}
r=l &=& +{{\textstyle\frac{1}{2}}}(1+a)\,,\\
\bar r=\bar l&=& -{{\textstyle\frac{1}{2}}}(1-a)\,.\end{aligned}$$
As before, we use the normalization $|r|+|l|+|\bar r|+|\bar l|=2$. The current-current factors $(1-{\bf v}_i\cdot {\bf v}_j)$ are equal to 2 for opposite moving modes, and $1\pm\cos\theta$ for the other pairs in obvious ways.
![Homogeneous ensemble of four neutrino modes (two beams with relative angle $\theta$.) The system is taken to be infinite in all directions.[]{data-label="fig:twobeam"}](fig2.pdf){width="40.00000%"}
The symmetries of this setup suggest to combine the neutrino and antineutrino amplitudes in a symmetric and antisymmetric form, $A_{\pm}={{\textstyle\frac{1}{2}}}(L\pm R)$ and $\bar{A}_\pm={{\textstyle\frac{1}{2}}}(\bar L\pm \bar R)$. Indeed, the linearized equations of motion decouple and we find with $c=\cos\theta$ $$\begin{aligned}
\kern-1em\Omega\,\begin{pmatrix}A_+\\ \bar A_+\end{pmatrix}&=&
\left[\begin{pmatrix}\omega&0\\ 0&-\omega\end{pmatrix}
+\frac{\mu\,(3-c)}{2}
\begin{pmatrix}
-1+a,&1-a,\\ -1-a&1+a
\end{pmatrix}
\right]\begin{pmatrix}A_+\\ \bar A_+\end{pmatrix},
\\[1.5ex]
\kern-1em\Omega\,\begin{pmatrix}A_-\\ \bar A_-\end{pmatrix}&=&
\left[\begin{pmatrix}\omega&0\\ 0&-\omega\end{pmatrix}
+\frac{a\mu\,(5+c)}{2}
+\frac{\mu}{2}
\begin{pmatrix}
-(1-3c)&-(1+c)(1-a)\\ (1+c)(1+a)&1-3c
\end{pmatrix}
\right]\begin{pmatrix}A_-\\ \bar A_-\end{pmatrix}.\end{aligned}$$ The first equation again corresponds to the usual flavor pendulum. Indeed, for $c=-1$ we return to the situation of a completely left-right asymmetric system of all neutrinos moving right and all antineutrinos moving left and we reproduce equation (\[eq:EOM5\]). As observed earlier, we then need a nontrivial spatial variation to obtain fast flavor conversion.
The second case with left-right symmetry breaking, on the other hand, provides nontrivial eigenvalues even for $\omega=0$ which are found to be $$\label{eq:twobeam-eigenvalue}
\Omega=\frac{a\mu(5+c)}{2}\pm\frac{\mu}{2}\sqrt{(1+c)^2a^2-8c(1-c)}\,.$$ In figure \[fig:twobeam-growthrate\] we show the imaginary part as a contour plot in the $\cos\theta$–$a$ plane. A fast growth rate occurs only for $\cos\theta>0$ and it is symmetric between positive and negative $a$. The absolute maximum obtains for $a=0$ and $\cos\theta={{\textstyle\frac{1}{2}}}$ and is found to be ${\rm Im}\Omega|_{\rm max}=\mu/\sqrt{2}$.
![Growth rate in units of $\mu$ for neutrino-antineutrino beams intersecting at an angle $\theta$ and a neutrino-antineutrino asymmetry $-1<a<1$. The mode occupations are taken to be symmetric between left and right. The analytic expression is given in equation (\[eq:twobeam-eigenvalue\]). A fast growth rate occurs for $\cos\theta>0$ and it is symmetric between positive and negative $a$.[]{data-label="fig:twobeam-growthrate"}](fig3.pdf){width="70.00000%"}
We have performed a numerical solution of the full nonlinear equations for typical parameters $\cos\theta$ and $a$ in the unstable regime. We find the usual behavior of an inverted oscillator. Given a small perturbation, there is a long phase of exponential growth of the transverse component, followed by a deep dip of the flavor content of type $\nu_e\bar\nu_e\to\nu_x\bar\nu_x$ and back to $\nu_e\bar\nu_e$ and so on, similar to the usual flavor pendulum.
Two-bulb supernova model {#sec:two-bulb}
========================
Setting up the model
--------------------
As a final example we consider the model proposed by Sawyer [@Sawyer:2015dsa] which is motivated by typical supernova emission characteristics. Neutrinos are taken to emerge from a spherical surface, the “neutrino bulb,” with a blackbody-like angular characteristic, i.e., isotropically into the outer half space [@Duan:2006an]. In this case, a distant observer measures a zenith-angle distribution which is uniform in the variable $\sin^2\theta$ up to a maximum which is determined by the angular size of the neutrino bulb at the observation point. The species $\nu_e$ and $\bar\nu_e$ decouple in different regions. Therefore, as a simple approximation one can assume that they are emitted from neutrino surfaces of different radii, which we call a two-bulb emission model. This setup leads to $\nu_e$ and $\bar\nu_e$ zenith-angle distributions of the type illustrated in figure \[fig:twobulb-spectrum\]. In a supernova, one usually expects the $\nu_e$ flux to exceed that of $\bar\nu_e$. However, the recently discovered LESA phenomenon (lepton-emission self-sustained asymmetry) implies that the relative fluxes show a strong hemispheric asymmetry [@Tamborra:2014aua]. Moreover, in neutron-star mergers, very different distributions may occur which also depend strongly on direction [@Dasgupta:2008cu; @Malkus:2014iqa].
![Zenith-angle distribution for neutrinos (blue) and antineutrinos (orange) implied by the two-bulb supernova emission model.[]{data-label="fig:twobulb-spectrum"}](fig4.pdf){width="45.00000%"}
The main point of this supernova-motivated setup is the neutrino velocity distribution in the transverse direction. One can formulate this problem in terms of velocities in the transverse plane [@Chakraborty:2015tfa] and it is then very similar to the colliding-beam examples of the previous sections, with different velocity distributions for $\nu_e$ and $\bar\nu_e$. Therefore, this case is conceptually quite similar to our previous ones.
We consider a stationary two-flavor neutrino flux and assume stationarity of the solution, i.e., we study the flavor evolution as a function of radius. We ignore small-scale effects in the transverse direction, i.e., the solution is constrained to depend on radius alone. The neutrino radiation field at some observation point beyond the emitting surface is described by the azimuth angle $\varphi$ and the zenith-angle variable $u\propto\sin^2\theta$. The range of occupied zenith angles is normalized to some chosen reference radius, so the $u$-range does not depend on the test radius where we perform the stability analysis. The emission spectrum $g(\omega,u)$ has the same meaning as $g_i$ in our earlier sections, except that we use the continuous labels $\omega$ and $u$. We assume axial symmetry of emission so that $g(\omega,u)$ does not depend on $\varphi$.
Eigenvalue equation
-------------------
We use the eigenvalue equation in the form developed in reference [@Raffelt:2013rqa] for the case of axially symmetric neutrino emission. As input for the eigenvalue equation we need the integrals $$I_n=\mu\int d\omega\,du\,\frac{u^n\,g(\omega,u)}{\omega+u\,\bar\lambda-\Omega}$$ for $n=0$, 1 and 2. We have dropped a possible $\varphi$ dependence because we assume axially symmetric emission. The parameter $\bar\lambda=\lambda+\epsilon\mu$ describes the effective multi-angle matter effect where $\lambda=\sqrt{2}{G_{\rm F}}n_e$ and $$\epsilon=\int d\omega\,du\,g(\omega,u)\,.$$ In contrast to reference [@Raffelt:2013rqa] we here normalize the spectrum in the same way as in the previous sections, i.e., $\int
du\,d\omega\,{\rm sign}(\omega)\,g(\omega,u)=2$ which also influences the meaning of $\mu$. The only physically relevant quantity is the product $\mu\,g(u,\omega)$ and it is somewhat arbitrary how to normalize the two quantities separately. The main point is to define a quantity $\mu$ which has the meaning of a typical neutrino-neutrino interaction energy. The eigenvalues $\Omega$ are found as solutions of one of the equations $$\label{eq:egnval}
(I_1-1)^2-I_0I_2=0
\qquad\hbox{and}\qquad
I_1+1=0\,.$$ The first equation corresponds to those solutions which respect axial symmetry, whereas the second corresponds to spontaneous axial symmetry breaking.
We look for instabilities in the limit $\omega=0$. In this case the contributions to $g(\omega,u)$ from emitted $\nu_x$ and $\bar\nu_x$ drop out exactly if their emission characteristics are the same. Notice, however, that the $\nu_x$ distribution enters indirectly through the definition of the effective neutrino density $n_\nu$ and the definition of $\mu$ and $\epsilon$. However, for simplicity we assume that no $\nu_x$ or $\bar\nu_x$ are emitted. We denote the $\omega$-integrated zenith-angle distributions as for neutrinos (positive $\omega$) and for antineutrinos (negative $\omega$). In this notation we have $$\int du\,\Bigl[h_{\nu_e}(u)+h_{\bar\nu_e}(u)\Bigr]=2
\qquad\hbox{and}\qquad
\int du\,\Bigl[h_{\nu_e}(u)-h_{\bar\nu_e}(u)\Bigr]=\epsilon\,.$$ After performing the trivial $\omega$ integration, the above integrals are $$I_n=\int du\,\frac{u^n}{u\,(\epsilon+m)-w}\,
\Bigl[h_{\nu_e}(u)-h_{\bar\nu_e}(u)\Bigr]\,,$$ where $w=\Omega/\mu$ is the normalized eigenvalue and $m=\lambda/\mu$ represents the matter effect.
The two-bulb model of neutrino emission implies the top-hat $u$ distributions shown in figure \[fig:twobulb-spectrum\]. We express the occupied $u$-ranges in terms of a width parameter $-1<b<+1$ in the form $u_{\nu_e}=1+b$ and $u_{\bar\nu_e}=1-b$. In the supernova context, the $\nu_e$ interact more strongly, thus decouple at a larger distance, and hence correspond to $b>0$. Moreover, we describe the normalized neutrino densities as $n_{\nu_e}=1+a$ and $n_{\bar\nu_e}=1-a$ in terms of an “asymmetry parameter” $-1<a<+1$. In the supernova context, deleptonization implies an excess of $\nu_e$ over $\bar\nu_e$ so that $a>0$. In other words, the traditional supernova-motivated situation corresponds to the first quadrant $a,b>0$ in the parameter space of our model. In terms of these parameters, the zenith-angle distributions are $$h(u)=\frac{1\pm a}{1\pm b}\times\begin{cases}1&\hbox{for $0\leq u \leq 1\pm b$}\,,\\
0&\hbox{otherwise}\,,\end{cases}$$ where the upper sign refers to $\nu_e$, the lower sign to $\bar\nu_e$. So finally the integrals are $$I_n=\frac{1+a}{1+b}\int_0^{1+b} du\,\frac{u^n}{u\,(2a+m)-w}
-\frac{1-a}{1-b}\int_0^{1-b} du\,\frac{u^n}{u\,(2a+m)-w}\,,$$ where we have used $\epsilon=2a$. These integrals can be performed analytically without problem, but the eigenvalues can be found only numerically.
One special case is $b=0$ where the zenith-angle distributions for $\nu_e$ and $\bar\nu_e$ are the same, yet their number density is different ($a\not=0$). In this case the integrals are $$I_n=2a\int_0^{1} du\,\frac{u^n}{u\,(2a+m)-w}\,.$$ Numerically it appears that in this case the eigenvalues are always real, i.e., fast flavor conversion indeed requires the top-hat distributions to have different widths. However, we have not tried to prove this point mathematically.
Solution without matter effect
------------------------------
![Growth rate in units of $\mu$ for the axial-symmetry breaking solution of the two-bulb supernova model without matter. The normalized $\nu_e$ density is $1+a$, for $\bar\nu_e$ it is $1-a$. The $\nu_e$ zenith-angle distribution is nonzero on the range $0<u<1+b$ and on $0<u<1-b$ for $\bar\nu_e$. The results show no instability in the SN motivated parameters ($a>0$ and $b>0$), i.e., the first quadrant. []{data-label="fig:twobulb-contour"}](fig5.pdf){width="70.00000%"}
As a first nontrivial case we ignore matter ($m=0$) and find that the first case in equation (\[eq:egnval\]), the axially symmetric solution, does not show any instabilities in a numerical scan over the space $-1<a<1$ and $-1<b<1$. The second equation (axial symmetry breaking) provides solutions and hence allows fast flavor conversion. We show the imaginary part of $w$, i.e., the growth rate in units of $\mu$ as a contour plot in figure \[fig:twobulb-contour\]. The first and third quadrants are stable, i.e., when $a$ and $b$ have the same sign. These results suggest that fast flavor conversion requires that the species $\nu_e$ or $\bar\nu_e$ with the broader zenith-angle distribution must have a smaller flux. This particular conclusion appears to be opposite of what Sawyer has found in his recent study [@Sawyer:2015dsa].
Of course, in accretion disks arising from neutron-star mergers or black hole-neutron star mergers, the flux is dominated by $\bar\nu_e$, not $\nu_e$, so that $a<0$. Also LESA can be another interesting scenario spanning parameters other than the traditional supernova-motivated case. Therefore, the main point is the possible existence of fast flavor conversion if nontrivial zenith-angle distributions are used.
Including matter
----------------
These above results change drastically in the presence of matter. A substantial matter effect is expected when $\lambda$ is at least of order $\mu$, so as a specific example we use $m=\lambda/\mu=1$ and show the growth rates in figure \[fig:twobulb-matter\]. We find fast growth rates for both the axially symmetric and the axial-symmetry breaking cases. While the latter (bottom panel) is simply a modified version of the matter-free case, we now find run-away solutions even in the axially symmetric case (upper panel). In particular, there are unstable solutions for supernova-motivated parameters, where the $\nu_e$ distribution is the broader one ($b>0$) and there are more $\nu_e$ than $\bar\nu_e$ ($a>0$), i.e., the first quadrant of our parameter space.
![Growth rate in units of $\mu$ in analogy to figure \[fig:twobulb-contour\], but now with matter $m=\lambda/\mu=1$. [*Top:*]{} Axially symmetric solution. [*Bottom:*]{} Axial symmetry spontaneously broken.[]{data-label="fig:twobulb-matter"}](fig6a.pdf "fig:"){width="70.00000%"} ![Growth rate in units of $\mu$ in analogy to figure \[fig:twobulb-contour\], but now with matter $m=\lambda/\mu=1$. [*Top:*]{} Axially symmetric solution. [*Bottom:*]{} Axial symmetry spontaneously broken.[]{data-label="fig:twobulb-matter"}](fig6b.pdf "fig:"){width="70.00000%"}
If we had used instead a background of antimatter ($m<0$), the unstable range would lie in the other half where $b<0$. For $b=0$, when the two zenith-angle distributions are the same, no fast instability seems to occur as remarked earlier.
If the matter effect is very large ($\lambda\gg\mu$, corresponding to $m\gg1$), the axially symmetric solution disappears, so it exists only for some range of matter density. For example, during the supernova accretion phase, this instability would be suppressed in analogy to the “slow” instabilities [@EstebanPretel:2008ni; @Sarikas:2011am; @Sarikas:2012vb; @Chakraborty:2011nf; @Chakraborty:2011gd]. Of course, we have here only considered the ${\bf k}=0$ case as well as stationarity of the solution. Therefore, what all of this means in practice remains to be understood.
Previous studies
----------------
Flavor-dependent angle distributions were previously investigated by Mirizzi and Serpico [@Mirizzi:2011tu; @Mirizzi:2012wp]. These authors have not reported fast flavor conversion in any of their cases. They have used forward-peaked distributions of the form $(1-u)^{\beta/2}$ where $\beta=0$ provides a top-hat distribution on the interval $0< u<1$ and for $\beta>0$ a distribution which is more concentrated for smaller $u$-values. However, these authors assumed equal $\beta$ for both $\nu_e$ and $\bar\nu_e$ and a different one for $\nu_x$ which was the same for $\bar\nu_x$, i.e., they only studied angle differences between the $x$-flavor and the $e$-species. As we remarked earlier, if $\nu_x$ and $\bar\nu_x$ have the same distribution, they drop out of the equation in our limit of $\omega=0$. In other words, the distributions used in references [@Mirizzi:2011tu; @Mirizzi:2012wp] indeed do not spawn fast flavor conversion.
As a cross check we have also considered angle distributions of this form, but taking different $\beta$ for $\nu_e$ and $\bar\nu_e$ as well as different abundances. We find fast instabilities which qualitatively agree with our earlier cases of top-hat distributions. Therefore, fast flavor conversion is not an artifact of the top-hat distribution.
A stability analysis was also performed by Saviano et al.[@Saviano:2012yh], using realistic energy and zenith-angle distributions taken from a few specific numerical supernova simulations. The growth rates reported in their figure 2 are always of order the vacuum oscillation frequency and thus not fast. The used numerical angle distributions are shown in the lower panels of their figure 1. The $\nu_e$ and $\bar\nu_e$ (dotted and solid curves) look visually very similar except for the overall normalization which represent the different fluxes. The matter effect was taken into account, but not the possibility of axial symmetry breaking. However, the fast conversion will remain absent when the ordinary matter effect is significantly larger than the effect of background neutrinos. Indeed, in the examples of reference [@Saviano:2012yh] the matter effect $\lambda$ is almost an order of magnitude larger than $\mu$. Also, probably in these specific models, the angle distributions were not different enough to spawn fast flavor conversion. In a later study [@Chakraborty:2014nma], these authors analyzed the same models for the axial symmetry breaking case. However, again the matter effects were large and the angle distributions were similar. The lepton asymmetry was supernova inspired, i.e., the first quadrant in figure \[fig:twobulb-contour\] and bottom panel of figure \[fig:twobulb-matter\]. Unsurprisingly, fast flavor conversion did not show up in this case either.
In summary, while a number of previous studies have considered nontrivial zenith-angle distributions, the chosen examples could not have found fast flavor conversion.
Conclusions {#sec:conclusion}
===========
We have studied a few simple examples of interacting neutrino systems which show the phenomenon of “fast flavor conversion,” i.e., they have unstable modes in flavor space which grow with rates of order the neutrino-neutrino interaction energy $\mu=\sqrt{2} G_{\rm F}n_{\nu}$ instead of the much smaller vacuum oscillation frequency $\omega=\Delta^2m/2E$. In these cases, self-induced flavor conversion in the sense of flavor shuffling among modes does not depend on $\Delta m^2$ or the vacuum mixing angle except for providing disturbances as seeds for the run-away modes. In other words, the main conceptual point is that self-induced flavor conversion does not depend on flavor mixing. In the supernova context, neutrino flavor evolution on the refractive level would have had to be considered even if flavor mixing among neutrinos did not exist.
Notice that to lowest order, neutrino-neutrino interactions are of neutral-current type and thus flavor blind. We ignore radiative corrections which introduce a flavor dependence in neutrino-neutrino refraction [@Mirizzi:2009td]. In this approximation, the overall flavor content of the ensemble remains conserved by the action of neutrino-neutrino refraction, i.e., self-induced flavor conversion corresponds to flavor reshuffling among modes which however can lead to flavor decoherence if neighboring modes become effectively uncorrelated.
The principle of fast flavor conversion was discovered ten years ago by Ray Sawyer [@Sawyer:2005jk] in a three-flavor setup of a small number of modes. He speculated that supernova neutrinos might flavor-equilibrate over very short distances, meters or even centimeters, in their decoupling region. With hindsight it is difficult to understand why the conceptual and practical points raised in this paper were completely lost on the community.
Fast flavor conversion by definition does not depend on the vacuum oscillation frequencies and thus not on neutrino energy. The energy spectrum plays no role, fast flavor conversion is driven by nontrivial angle distributions. In several of our examples, the spontaneous breaking of initial symmetries was also important. However, the crucial condition is that the initial angle distribution must not be too symmetric or too simple, although we cannot provide a general mathematical condition.
In the context of astrophysical applications in supernovae or neutron-star mergers, the main question is if neutrinos emerging from the decoupling region maintain spectral fluxes which strongly depend on species or if self-induced flavor conversion and its interplay with matter effects and vacuum oscillations leads to quick flavor decoherence. The effects of spatial [@Duan:2014gfa; @Abbar:2015mca; @Mangano:2014zda; @Mirizzi:2015hwa] and temporal [@Abbar:2015fwa; @Dasgupta:2015iia; @Chakraborty:2016yeg] symmetry breaking as well as the possibility of fast flavor conversion [@Sawyer:2005jk; @Sawyer:2015dsa] have been taken as evidence for quick decoherence. Still, the breaking of spatial homogeneity may be suppressed by the multi-angle matter effect [@Chakraborty:2015tfa], and the breaking of stationarity depends on a narrow resonance condition.
Actually, our stability studies as well as numerical solutions of the full equations in the free-streaming limit may not be appropriate to capture the realistic evolution at or near the neutrino decoupling region of a compact object. In this region, the description of the neutrino mean field in terms of a freely outward streaming neutrino flux is not appropriate: neutrinos flow in all directions, but with different intensity. Even at larger distances, the re-scattered neutrino flux plays an important role [@Cherry:2012zw; @Cherry:2013mv].
Therefore, the toy examples studied here and in the recent literature leave crucial questions open and do not yet provide clear-cut conclusions concerning realistic flavor evolution in core-collapse or neutron-star merger events. Neutrino flavor evolution in dense media remains a challenging subject where different pieces of the jigsaw puzzle keep showing up, but do not yet form a complete picture.
Acknowledgments {#acknowledgments .unnumbered}
===============
We acknowledge partial support by the Deutsche Forschungsgemeinschaft through Grant No. EXC 153 (Excellence Cluster “Universe”) and by the European Union through the Initial Training Network “Invisibles,” Grant No. PITN-GA-2011-289442.
[10]{}
H.-T. Janka, Explosion mechanisms of core-collapse supernovae, Ann. Rev. Nucl. Part. Sci. [**62**]{} (2012) 407 \[arXiv:1206.2503\]. K. Scholberg, Supernova neutrino detection Ann. Rev. Nucl. Part. Sci. [**62**]{} (2012) 81 \[arXiv:1205.6003\]. A. Mirizzi, I. Tamborra, H.-T. Janka, N. Saviano, K. Scholberg, R. Bollig, L. Hüdepohl and S. Chakraborty, Supernova neutrinos: Production, oscillations and detection, \[arXiv:1508.00785\]. S. Chakraborty, R. Hansen, I. Izaguirre and G. Raffelt, Collective neutrino flavor conversion: Recent developments, Nucl. Phys. B [****]{} (2016) \[arXiv:1602.02766\].
R. F. Sawyer, Speed-up of neutrino transformations in a supernova environment, Phys. Rev. D [**72**]{} (2005) 045003 \[hep-ph/0503013\]. R. F. Sawyer, Neutrino cloud instabilities just above the neutrino sphere of a supernova, arXiv:1509.03323. V. A. Kostelecký and S. Samuel, Selfmaintained coherent oscillations in dense neutrino gases, Phys. Rev. D [**52**]{} (1995) 621 \[hep-ph/9506262\]. J. T. Pantaleone, Stability of incoherence in an isotropic gas of oscillating neutrinos, Phys. Rev. D [**58**]{} (1998) 073002. S. Pastor, G. G. Raffelt and D. V. Semikoz, Physics of synchronized neutrino oscillations caused by selfinteractions, Phys. Rev. D [**65**]{} (2002) 053011 \[hep-ph/0109035\]. G. G. Raffelt and I. Tamborra, Synchronization versus decoherence of neutrino oscillations at intermediate densities, Phys. Rev. D [**82**]{} (2010) 125004 \[arXiv:1006.0002\]. E. Akhmedov and A. Mirizzi, Another look at synchronized neutrino oscillations, arXiv:1601.07842. V. A. Kostelecký and S. Samuel, Neutrino oscillations in the early universe with an inverted neutrino mass hierarchy, Phys. Lett. B [**318**]{} (1993) 127. S. Samuel, Bimodal coherence in dense selfinteracting neutrino gases, Phys. Rev. D [**53**]{} (1996) 5382 \[hep-ph/9604341\]. H. Duan, G. M. Fuller and Y.-Z. Qian, Collective neutrino flavor transformation in supernovae, Phys. Rev. D [**74**]{} (2006) 123004 \[astro-ph/0511275\]. S. Hannestad, G. G. Raffelt, G. Sigl and Y. Y. Y. Wong, Self-induced conversion in dense neutrino gases: Pendulum in flavour space, Phys. Rev. D [**74**]{} (2006) 105010; Erratum [*ibid.*]{} [**76**]{} (2007) 029901 \[astro-ph/0608695\]. H. Duan, G. M. Fuller, J. Carlson and Y.-Z. Qian, Analysis of collective neutrino flavor transformation in supernovae, Phys. Rev. D [**75**]{} (2007) 125005 \[astro-ph/0703776\]. A. Banerjee, A. Dighe and G. Raffelt, Linearized flavor-stability analysis of dense neutrino streams, Phys. Rev. D [**84**]{} (2011) 053013 \[arXiv:1107.2308\]. G. G. Raffelt, N-mode coherence in collective neutrino oscillations, Phys. Rev. D [**83**]{} (2011) 105022 \[arXiv:1103.2891\]. G. Raffelt and D. de Sousa Seixas, Neutrino flavor pendulum in both mass hierarchies, Phys. Rev. D [**88**]{} (2013) 045031 \[arXiv:1307.7625\]. R. S. Hansen and S. Hannestad, Chaotic flavor evolution in an interacting neutrino gas, Phys. Rev. D [**90**]{} (2014) 025009 \[arXiv:1404.3833\]. S. Chakraborty, R. S. Hansen, I. Izaguirre and G. Raffelt, Self-induced flavor conversion of supernova neutrinos on small scales, JCAP, in press (2016) \[arXiv:1507.07569\]. H. Duan and S. Shalgar, Flavor instabilities in the neutrino line model, Phys. Lett. B [**747**]{} (2015) 139 \[arXiv:1412.7097\]. S. Abbar, H. Duan and S. Shalgar, Flavor instabilities in the multiangle neutrino line model, Phys. Rev. D [**92**]{} (2015) 065019 \[arXiv:1507.08992\]. G. Mangano, A. Mirizzi and N. Saviano, Damping the neutrino flavor pendulum by breaking homogeneity, Phys. Rev. D [**89**]{} (2014) 073017 \[arXiv:1403.1892\]. A. Mirizzi, G. Mangano and N. Saviano, Self-induced flavor instabilities of a dense neutrino stream in a two-dimensional model, Phys. Rev. D [**92**]{} (2015) 021702 \[arXiv:1503.03485\]. G. Sigl and G. Raffelt, General kinetic description of relativistic mixed neutrinos, Nucl. Phys. B [**406**]{} (1993) 423. S. Chakraborty, G. Raffelt, H.-T. Janka and B. Müller, Supernova deleptonization asymmetry: Impact on self-induced flavor conversion, Phys. Rev. D [**92**]{} (2015) 105002 \[arXiv:1412.0670\]. G. Raffelt, S. Sarikas and D. de Sousa Seixas, Axial symmetry breaking in self-induced flavor conversion of supernova neutrino fluxes, Phys. Rev. Lett. [**111**]{} (2013) 091101; Erratum [*ibid.*]{} [**113**]{} (2014) 239903 \[arXiv:1305.7140\]. A. Mirizzi and P. D. Serpico, Instability in the dense supernova neutrino gas with flavor-dependent angular distributions, Phys. Rev. Lett. [**108**]{} (2012) 231102 \[arXiv:1110.0022\]. A. Mirizzi and P. D. Serpico, Flavor stability analysis of dense supernova neutrinos with flavor-dependent angular distributions,” Phys. Rev. D [**86**]{} (2012) 085010 \[arXiv:1208.0157\]. N. Saviano, S. Chakraborty, T. Fischer and A. Mirizzi, Stability analysis of collective neutrino oscillations in the supernova accretion phase with realistic energy and angle distributions, Phys. Rev. D [**85**]{} (2012) 113002 \[arXiv:1203.1484\]. S. Chakraborty, A. Mirizzi, N. Saviano and D. de Sousa Seixas, Suppression of the multi-azimuthal-angle instability in dense neutrino gas during supernova accretion phase, Phys. Rev. D [**89**]{} (2014) 093001 \[arXiv:1402.1767\]. A. Mirizzi, S. Pozzorini, G. G. Raffelt and P. D. Serpico, Flavour-dependent radiative correction to neutrino-neutrino refraction, JHEP [**0910**]{} (2009) 020 \[arXiv:0907.3674\]. A. Mirizzi, Breaking the symmetries in self-induced flavor conversions of neutrino beams from a ring, Phys. Rev. D [**92**]{} (2015) 105020 \[arXiv:1506.06805\]. S. Abbar and H. Duan, Neutrino flavor instabilities in a time-dependent supernova model, Phys. Lett. B [**751**]{} (2015) 43 \[arXiv:1509.01538\]. B. Dasgupta and A. Mirizzi, Temporal instability enables neutrino flavor conversions deep inside supernovae Phys. Rev. D [**92**]{} (2015) 125030 \[arXiv:1509.03171\]. H. Duan, G. M. Fuller, J. Carlson and Y.-Z. Qian, Simulation of coherent non-linear neutrino flavor transformation in the supernova environment: Correlated neutrino trajectories, Phys. Rev. D [**74**]{} (2006) 105014 \[astro-ph/0606616\]. I. Tamborra, F. Hanke, H.-T. Janka, B. Müller, G. G. Raffelt and A. Marek, Self-sustained asymmetry of lepton-number emission: A new phenomenon during the supernova shock-accretion phase in three dimensions, Astrophys. J. [**792**]{} (2014) 96 \[arXiv:1402.5418\]. B. Dasgupta, A. Dighe, A. Mirizzi and G. G. Raffelt, Collective neutrino oscillations in non-spherical geometry, Phys. Rev. D [**78**]{} (2008) 033014 \[arXiv:0805.3300\]. A. Malkus, A. Friedland and G. C. McLaughlin, Matter-neutrino resonance above merging compact objects, arXiv:1403.5797. A. Esteban-Pretel, A. Mirizzi, S. Pastor, R. Tomàs, G. G. Raffelt, P. D. Serpico and G. Sigl, Role of dense matter in collective supernova neutrino transformations, Phys. Rev. D [**78**]{} (2008) 085012 \[arXiv:0807.0659\]. S. Sarikas, G. G. Raffelt, L. Hüdepohl and H.-T. Janka, Suppression of self-induced flavor conversion in the supernova accretion phase, Phys. Rev. Lett. [**108**]{} (2012) 061101 \[arXiv:1109.3601\]. S. Sarikas, I. Tamborra, G. Raffelt, L. Hüdepohl and H.-T. Janka, Supernova neutrino halo and the suppression of self-induced flavor conversion, Phys. Rev. D [**85**]{} (2012) 113007 \[arXiv:1204.0971\]. S. Chakraborty, T. Fischer, A. Mirizzi, N. Saviano and R. Tomàs, No collective neutrino flavor conversions during the supernova accretion phase, Phys. Rev. Lett. [**107**]{} (2011) 151101 \[arXiv:1104.4031\]. S. Chakraborty, T. Fischer, A. Mirizzi, N. Saviano and R. Tomàs, Analysis of matter suppression in collective neutrino oscillations during the supernova accretion phase, Phys. Rev. D [**84**]{} (2011) 025002 \[arXiv:1105.1130\]. J. F. Cherry, J. Carlson, A. Friedland, G. M. Fuller and A. Vlasenko, Neutrino scattering and flavor transformation in supernovae, Phys. Rev. Lett. [**108**]{} (2012) 261104 \[arXiv:1203.1607\]. J. F. Cherry, J. Carlson, A. Friedland, G. M. Fuller and A. Vlasenko, Halo modification of a supernova neutronization neutrino burst, Phys. Rev. D [**87**]{} (2013) 085037 \[arXiv:1302.1159\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Given a sequence of resistance forms that converges with respect to the Gromov-Hausdorff-vague topology and satisfies a uniform volume doubling condition, we show the convergence of corresponding Brownian motions and local times. As a corollary of this, we obtain the convergence of time-changed processes. Examples of our main results include scaling limits of Liouville Brownian motion, the Bouchaud trap model and the random conductance model on trees and self-similar fractals. For the latter two models, we show that under some assumptions the limiting process is a FIN diffusion on the relevant space.'
author:
- 'D. A. Croydon, B. M. Hambly and [T. Kumagai]{}'
date: 'July 25, 2016'
title: |
Time-changes of stochastic processes\
associated with resistance forms
---
[**AMS 2010 Mathematics Subject Classification**]{}: Primary 60J35, 60J55; Secondary 28A80, 60J10, 60J45, 60K37.
[**Keywords and phrases**]{}: Bouchaud trap model, FIN diffusion, fractal, Gromov-Hausdorff convergence, Liouville Brownian motion, local time, random conductance model, resistance form, time-change.
Introduction
============
In recent years, interest in time-changes of stochastic processes according to irregular measures has arisen from various sources. Fundamental examples of such time-changed processes include the so-called *Fontes-Isopi-Newman (FIN) diffusion* [@FIN], the introduction of which was motivated by the study of the localisation and aging properties of physical spin systems, and the two-dimensional *Liouville Brownian motion* [@Beres; @GRV], which is the diffusion naturally associated with planar Liouville quantum gravity. More precisely, the FIN diffusion is the time-change of one-dimensional Brownian motion by the positive continuous additive functional with Revuz measure given by $$\label{onedfinmeasure}
\nu(dx)=\sum_{i}v_i\delta_{x_i}(dx),$$ where $(v_i ,x_i)_{i\in\mathbb{N}}$ is the Poisson point process with intensity $\alpha v^{-1-\alpha}dvdx$, and $\delta_{x_i}$ is the probability measure placing all its mass at $x_i$. Similarly, the two-dimensional Liouville Brownian motion is the time-change of two-dimensional Brownian motion by the positive continuous additive functional with Revuz measure given by $$\label{2dlmeasure}
\nu(dx)=e^{\kappa\gamma(x)-\frac{\kappa^2}{2}\mathbf{E}(\gamma(x)^2)}dx$$ for some $\kappa\in(0,2)$, where $\gamma$ is the massive Gaussian free field; actually the latter description is only formal since the Gaussian free field can not be defined as a function in two dimensions. In both cases, connections have been made with discrete models; the FIN diffusion is known to be the scaling limit of the one-dimensional Bouchaud trap model [@BC; @FIN] and the constant speed random walk amongst heavy-tailed random conductances in one-dimension [@CernyEJP], and the two-dimensional Liouville Brownian motion is conjectured to be the scaling limit of simple random walks on random planar maps [@GRV], see also [@DS]. The goal here is to provide a general framework for studying such processes and their discrete approximations in the case when the underlying stochastic process is strongly recurrent, in the sense that it can be described by a resistance form, as introduced by Kigami (see [@Kig] for background). In particular, this includes the case of Brownian motion on tree-like spaces and low-dimensional self-similar fractals.
To present our main results, let us start by introducing the types of object under consideration (for further details, see Section \[prelimsec\]). Let $\mathbb{F}$ be the collection of quadruples of the form $(F,R,\mu,\rho)$, where: $F$ is a non-empty set; $R$ is a resistance metric on $F$ such that $(F,R)$ is complete, separable and locally compact, and moreover closed balls in $(F,R)$ are compact; $\mu$ is a locally finite Borel regular measure of full support on $(F,R)$; and $\rho$ is a marked point in $F$. Note that the resistance metric is associated with a resistance form $(\mathcal{E},\mathcal{F})$ (see Definition \[resformdef\] below), and we will further assume that for elements of $\mathbb{F}$ this form is regular in the sense of Definition \[regulardef\]. In particular, this ensures the existence of a related regular Dirichlet form $(\mathcal{E},\mathcal{D})$ on $L^2(F,\mu)$, which we suppose is recurrent, and also a Hunt process $((X_t)_{t\geq 0},\:P_x,\: x\in F)$ that can be checked to admit jointly measurable local times $(L_t(x))_{x\in F,t\geq 0}$. The process $X$ represents our underlying stochastic process (i.e. it plays the role that Brownian motion does in the construction of the FIN diffusion and Liouville Brownian motion), and the existence of local times means that when it comes to defining the time-change additive functional, it will be possible to do this explicitly.
Towards establishing a scaling limit for discrete processes, we will assume that we have a sequence $(F_n,R_n,\mu_n,\rho_n)_{n\geq 1}$ in $\mathbb{F}$ that converges with respect to the Gromov-Hausdorff-vague topology (see Section \[ghpsec\]) to an element $(F,R,\mu,\rho)\in\mathbb{F}$. Our initial aim is to show that it is then the case that the associated Hunt processes $X^n$ and their local times $L^n$ converge to $X$ and $L$, respectively. To do this we assume some regularity for the measures in the sequence – this requirement is formalised in Assumption \[a1\], which depends on the following volume growth property. In the statement of the latter, we denote by $B_n(x,r)$ the open ball in $(F_n,R_n)$ centred at $x$ and of radius $r$, and also $r_0(n):= \inf_{x,y\in F_n,\:x\neq y}R_n(x,y)$ and $r_\infty(n):=\sup_{x,y\in F_n}R_n(x,y)$. We note that this control on the volume yields an equicontinuity property for the local times.
[A sequence $(F_n,R_n,\mu_n,\rho_n)_{n\geq 1}$ in $\mathbb{F}$ is said to satisfy *uniform volume growth with volume doubling (UVD)* if there exist constants $c_1,c_2,c_3\in(0,\infty)$ such that $$c_1v(r)\leq \mu_n\left(B_{n}(x,r)\right)\leq c_2v(r),\qquad\forall x\in F_n,\:r\in[r_0(n),r_\infty(n)+1]$$ for every $n\geq 1$, where $v:(0,\infty)\rightarrow(0,\infty)$ is non-decreasing function with $v(2r)\leq c_3v(r)$ for every $r\in\mathbb{R}_+$.]{}
[\[a1\] The sequence $(F_n,R_n,\mu_n,\rho_n)_{n\geq 1}$ in $\mathbb{F}$ satisfies UVD, and also $$\label{ghpconv}
\left(F_n,R_n,\mu_n,\rho_n\right)\rightarrow \left(F,R,\mu,\rho\right),$$ in the Gromov-Hausdorff-vague topology, where $(F,R,\mu,\rho)\in\mathbb{F}$.]{}
It is now possible to state our first main result. We write $D(\mathbb{R}_+,M)$ for the space of cadlag processes on $M$, equipped with the usual Skorohod $J_1$ topology. The definition of equicontinuity of the local times $L^n$, $n\geq 1$, should be interpreted as the conclusion of Lemma \[ltcont\].
[\[main1\] Suppose Assumption \[a1\] holds. It is then possible to isometrically embed $(F_n,R_n)$, $n\geq 1$, and $(F,R)$ into a common metric space $(M,d_M)$ in such a way that if $X^n$ is started from $\rho_n$, $X$ is started from $\rho$, then $$\left(X^n_t\right)_{t\geq 0}\rightarrow \left(X_t\right)_{t\geq 0}$$ in distribution in $D(\mathbb{R}_+,M)$. Moreover, the local times of $L^n$ are equicontinuous, and if the finite collections $(x_i^n)_{i=1}^k$ in $F_n$, $n\geq 1$, are such that $d_M(x_i^n,x_i)\rightarrow 0$ for some $(x_i)_{i=1}^k$ in $F$, then it simultaneously holds that $$\label{ltconv1}
\left(L^n_t\left(x_i^n\right)\right)_{i=1,\dots, k, t\geq 0}\rightarrow \left(L_t\left(x_i\right)\right)_{i=1,\dots, k, t\geq 0},$$ in distribution in $C(\mathbb{R}_+,\mathbb{R}^k)$.]{}
From the above result, we further deduce the convergence of time-changed processes. The following assumption adds the time-change measure to the framework.
[\[a2\] Assumption \[a1\] holds with (\[ghpconv\]) replaced by $$\left(F_n,R_n,\mu_n,\nu_n,\rho_n\right)\rightarrow \left(F,R,\mu,\nu,\rho\right),$$ in the (extended) Gromov-Hausdorff-vague topology (see Section \[ghpsec\]), where $\nu_n$ is a locally finite Borel regular measure on $F_n$, and $\nu$ is a locally finite Borel regular measure on $(F,R)$ with $\nu(F)>0$.]{}
The time-change additive functional that we consider is the following: $$\label{atdef}
A_t:=\int_FL_t(x)\nu(dx).$$ In particular, let $\tau(t):=\inf\{s>0:\:A_s>t\}$ be the right-continuous inverse of $A$, and define a process $X^\nu$ by setting $$\label{xnudef}
X^\nu_{t}:=X_{\tau(t)}.$$ As described in Section \[rfsec\], this is the trace of $X$ on the support of $\nu$ (with respect to the measure $\nu$), and its Dirichlet form is given by the corresponding Dirichlet form trace. We define $A^n$, $\tau^n$, and $X^{n,\nu_n}$ similarly. The space $L^1_{\rm loc}(\mathbb{R}_+,M)$ is the space of cadlag functions $\mathbb{R}_+\rightarrow M$ such that $\int_0^Td_M(\rho,f(t))dt<\infty$ for all $T\geq0$, equipped with the topology induced by supposing $f_n\rightarrow f$ if and only if $\int_0^Td_M(f_n(t),f(t))dt\rightarrow 0$ for any $T\geq 0$.
[\[maincor\] (a) Suppose Assumption \[a2\] holds, and that $\nu$ has full support. Then it is possible to isometrically embed $(F_n,R_n)$, $n\geq 1$, and $(F,R)$ into a common metric space $(M,d_M)$ in such a way that $$\label{xnnu}
X^{n,\nu_n}\rightarrow X^\nu$$ in distribution in $D(\mathbb{R}_+,M)$, where we assume that $X^n$ is started from $\rho_n$, and $X$ is started from $\rho$.\
(b) Suppose Assumption \[a2\] holds, and that $X$ is continuous. Then holds in distribution in $L^1_{\rm loc}(\mathbb{R}_+,M)$.]{}
The above results are proved in Section \[copsec\], following the introduction of preliminary material in Section \[prelimsec\]. In the remainder of the article, we demonstrate the application of Theorem \[main1\] and Corollary \[maincor\] to a number of natural examples. Firstly, we investigate the Liouville Brownian motion associated with a resistance form, showing in Proposition \[lbmconv\] that Assumption \[a1\] implies the convergence of the corresponding Liouville Brownian motions. This allows us to deduce the convergence of Liouville Brownian motions on a variety of trees and fractals, which we discuss in Example \[lbmexamples\]. We note that Liouville Brownian motion associated with a resistance form is a toy model and we discuss it merely as a simple example of our methods. The more interesting and challenging problem of analysing this process in two dimensions is not possible within our framework. Next, in Section \[bouchsec\], we proceed similarly for the Bouchaud trap model, describing the limiting process as the FIN process associated with a resistance form in Proposition \[btmresult\], and giving an application in Example \[SG55ex\]. Related to this, in Section \[rcmsec\], we study the heavy-tailed random conductance model on trees and a class of self-similar fractals, discussing a FIN limit for the so-called constant speed random walk in Propositions \[rcmtreeresult\], \[ssfrcmresult\] and Examples \[exa6-5\], \[exa6-18\]. Heat kernel estimates for the limiting FIN processes will be presented in a forthcoming paper [@CroHamKum].
Of the applications outlined in the previous paragraph, one that is particularly illustrative of the contribution of this article is the random conductance model on the (pre-)Sierpiński gasket graphs. More precisely, the random conductance model on a locally finite, connected graph $G=(V,E)$ is obtained by first randomly selecting edge-indexed conductances $(\omega_e)_{e\in E}$, and then, conditional on these, defining a continuous time Markov chain that jumps along edges with probabilities proportional to the conductances. For the latter process, there are two time scales commonly considered in the literature: firstly, for the *variable speed random walk (VSRW)*, the jump rate along edge $e$ is given by $\omega_e$, so that the holding time at a vertex $x$ has mean $(\sum_{e:\:x\in e}\omega_e)^{-1}$; secondly, for the *constant speed random walk (CSRW)*, holding times are assumed to have unit mean. From this description, it is clear that the CSRW is a time-change of the VSRW according to the measure placing mass $\sum_{e:\:x\in e}\omega_e$ on vertex $x$. Here, we will only ever consider conductances that are uniformly bounded below, but this still gives a rich enough model for there to exist a difference in the trapping behaviour experienced by the VSRW and CSRW. Indeed, in the one-dimensional case (i.e. when $G$ is $\mathbb{Z}$ equipped with edges between nearest neighbours) when conductances are i.i.d., it is easily checked that the VSRW has as its scaling limit Brownian motion (by adapting the argument of [@CernyEJP Appendix A] to the VSRW, for example); although the VSRW will cross edges of large conductance many times before escaping, it does so quickly, so that homogenisation still occurs. In the case of random conductances also uniformly bounded from above, the analogous result was proved in [@kk] for the VSRW on the fractal graphs shown in Figure \[sg\], with limit being Brownian motion on the Sierpiński gasket. In Section \[rcmfractalsec\], we extend this result significantly to show the same is true whenever the conductance distribution has at most polynomial decay at infinity. Specifically, writing $X^{n,\omega}$ for the VSRW on the $n$th level graph and $X$ for Brownian motion on the Sierpiński gasket, we prove that, under the annealed law (averaging over both process and environment), $$\label{sgvsrw}
\left(X^{n,\omega}_{5^nt}\right)_{t\geq 0}\rightarrow \left(X_t\right)_{t\geq0};$$ the time scaling here is the same as for the VSRW on the unweighted graph. For the CSRW, on the other hand, the many crossings of edges of large conductance lead to more significant trapping, which remains in the limit. In particular, if the conductance distribution satisfies $\mathbf{P}(\omega_e>u)\sim u^{-\alpha}$ for some $\alpha\in(0,1)$, then, as noted above, in the one-dimensional case the CSRW has a FIN diffusion limit [@CernyEJP]. Applying our time-change results, we are able to show that the corresponding result holds for the Sierpiński gasket graphs. Namely, writing $X^{n,\omega,\nu}$ for the CSRW on the $n$th level graph, we establish that there exists a constant $c$ such that, again under the annealed law, $$\label{sgcsrw}
\left(X^{n,\omega,\nu}_{c3^{n/\alpha}(5/3)^nt}\right)_{t\geq 0}\rightarrow\left(X^{\nu}_t\right)_{t\geq 0},$$ where the limit is now $\alpha$-FIN diffusion on the Sierpiński gasket, which is time-change of the Brownian motion on the limiting gasket by a Poisson random measure defined similarly to (\[onedfinmeasure\]), but with Lebesgue measure in the intensity replaced by the appropriate Hausdorff measure. (Note that, in the case that $\mathbf{E}\omega_e<\infty$, our techniques also yield convergence of CSRW to the Brownian motion, see Remark \[finitemoments\].) Full details for the preceding discussion are provided in Section \[rcmsec\]. At the start of the latter section, we also give an expanded heuristic explanation for the appearance of the FIN diffusion as a limit of the CSRW amongst heavy-tailed conductances. We remark that the specific conclusion of this interpretation is dependent on the point recurrence of the processes involved; by contrast, for the random conductance model on $\mathbb{Z}^d$ for $d\geq 2$, the same trapping behaviour gives rise in the limit to the so-called *fractional kinetics process*, for which the time-change and spatial motion are uncorrelated [@BarCern; @CernyEJP].
Finally, we note there are many other applications to which the notion of time-change is relevant, so that the techniques of this article might be useful. Although we do not consider it here, one such example is the *diffusion on branching Brownian motion*, as recently constructed in [@AndHar]. Moreover, whilst the examples of time-changes described above are based on measures that are constant in time, our main results will also be convenient for describing time-changes based on space-time measures, i.e. via additive functionals of the form $A_t:=\int_{F\times\mathbb{R}_+}\mathbf{1}_{\{s\leq L_t(x)\}}\nu(dxds)$. In particular, Theorem \[main1\] would be well-suited to extending the study of the scaling limits of randomly trapped random walks, as introduced in [@BCCR], from the one-dimensional setting to trees and fractals.
Preliminaries {#prelimsec}
=============
Resistance forms and associated processes {#rfsec}
-----------------------------------------
In this section, we define precisely the objects of study and outline some of their relevant properties; primarily this involves a recap of results from [@FOT] and [@Kig]. We start by recalling the definition of a resistance form and its associated resistance metric.
\[resformdef\] Let $F$ be a non-empty set. A pair $(\mathcal{E},\mathcal{F})$ is called a *resistance form* on $F$ if it satisfies the following five conditions.
RF1
: $\mathcal{F}$ is a linear subspace of the collection of functions $\{f:F\rightarrow\mathbb{R}\}$ containing constants, and $\mathcal{E}$ is a non-negative symmetric quadratic form on $\mathcal{F}$ such that $\mathcal{E}(f,f)=0$ if and only if $f$ is constant on $F$.
RF2
: Let $\sim$ be the equivalence relation on $\mathcal{F}$ defined by saying $f\sim g$ if and only if $f-g$ is constant on $F$. Then $(\mathcal{F}/\sim,\mathcal{E})$ is a Hilbert space.
RF3
: If $x\neq y$, then there exists a $f\in \mathcal{F}$ such that $f(x)\neq f(y)$.
RF4
: For any $x,y\in F$, $$\label{resdef}
R(x,y):=\sup\left\{\frac{\left|f(x)-f(y)\right|^2}{\mathcal{E}(f,f)}:\:f\in\mathcal{F},\:\mathcal{E}(f,f)>0\right\}<\infty.$$
RF5
: If $\bar{f}:=(f \wedge 1)\vee 0$, then $\bar{f}\in\mathcal{F}$ and $\mathcal{E}(\bar{f},\bar{f})\leq\mathcal{E}({f},{f})$ for any $f\in\mathcal{F}$.
We note that can be rewritten as $$R(x,y)=\left(\inf\left\{\mathcal{E}(f,f):\:
f\in \mathcal{F},\: f(x)=1,\: f(y)=0\right\}\right)^{-1},$$ which is the effective resistance between $x$ and $y$. The function $R:F\times F\rightarrow \mathbb{R}$ is actually a metric on $F$ (see [@Kig Proposition 3.3]); we call this the *resistance metric* associated with $(\mathcal{E},\mathcal{F})$. Henceforth, we will assume that we have a non-empty set $F$ equipped with a resistance form $(\mathcal{E},\mathcal{F})$ such that $(F,R)$ is complete, separable and locally compact. Defining the open ball centred at $x$ and of radius $r$ with respect to the resistance metric by $B_R(x,r):=\left\{y\in F:\:R(x,y)<r\right\}$, and denoting its closure by $\bar{B}_R(x,r)$, we will also assume that $\bar{B}_R(x,r)$ is compact for any $x\in F$ and $r>0$. Furthermore, we will restrict our attention to resistance forms that are regular, as per the following definition.
[\[regulardef\] Let $C_0(F)$ be the collection of compactly supported, continuous (with respect to $R$) functions on $F$, and $\|\cdot\|_F$ be the supremum norm for functions on $F$. A resistance form $(\mathcal{E},\mathcal{F})$ on $F$ is called *regular* if and only if $\mathcal{F}\cap C_0(F)$ is dense in $C_0(F)$ with respect to $\|\cdot\|_F$.]{}
We next introduce related Dirichlet forms and stochastic processes. First, suppose $\mu$ is a Borel regular measure on $(F,R)$ such that $0<\mu(B_R(x,r))<\infty$ for all $x\in F$ and $r>0$. Moreover, write $\mathcal{D}$ to be the closure of $\mathcal{F}\cap C_0(F)$ with respect to the inner product $\mathcal{E}_1$ on $\mathcal{F}\cap L^2(F,\mu)$ given by $$\label{e1def}
\mathcal{E}_1(f,g):=\mathcal{E}(f,g)+\int_Ffgd\mu.$$ Under the assumption that $(\mathcal{E},\mathcal{F})$ is regular, we then have the following. See [@FOT] for the definition of a regular Dirichlet form.
[The quadratic form $(\mathcal{E},\mathcal{D})$ is a regular Dirichlet form on $L^2(F,\mu)$.]{}
Given a regular Dirichlet form, standard theory then gives us the existence of an associated Hunt process $((X_t)_{t\geq 0},\:P_x,\: x\in F)$ (e.g. [@FOT Theorem 7.2.1]). Note that such a process is, in general, only specified uniquely for starting points outside a set of zero capacity. However, in this setting every point has strictly positive capacity (see [@Kig Theorem 9.9]), and so the process is defined uniquely everywhere. Moreover, since we are assuming closed balls are compact, we have from [@Kig Theorem 10.4] that $X$ admits a jointly continuous transition density $(p_t(x,y))_{x,y\in F,t>0}$. We note that the Dirichlet form for Brownian motion on ${\mathbb R}^d$ is a resistance form only when $d=1$. However, resistance forms are a rich class that contains various Dirichlet forms for diffusions on fractals, see [@kig1].
Key to this study will be the existence of local times for $X$. As a first step to introducing these, note that the strict positivity of the capacity of points remarked upon above implies that all points are regular (see [@FukuChen Theorems 1.3.14 and 3.1.10, and Lemma A.2.18], for example). Thus $X$ admits local times everywhere (see [@BG (V.3.13)]). In the following lemma, by studying the potential density of $X$, we check that these local times can be defined in a jointly measurable way and satisfy an occupation density formula.
[(a) Define the (one-)potential density $(u(x,y))_{x,y\in F}$ of $X$ by setting $$\label{udef}
u(x,y)=\int_0^\infty e^{-t} p_t(x,y)dt.$$ It then holds that $u(x,y)<\infty$ for all $x,y\in F$. Furthermore, $$\label{hitlap}
E_x\left(e^{-\tau_y}\right)=\frac{u(x,y)}{u(y,y)},$$ where $\tau_y:=\inf\{t>0:\:X_t=y\}$ is the hitting time of $y$ by $X$, and also $$\label{ufluc}
|u(x,y)-u(x,z)|^2\leq u(x,x)R(y,z)$$ for all $x,y,z\in F$.\
(b) The process $X$ admits jointly measurable local times $(L_t(x))_{x\in F,t\geq 0}$ that satisfy, $P_x$-a.s. for any $x$, $$\label{occdens}
\int_0^t \mathbf{1}_A(X_s)ds = \int_AL_t(y)\mu(dy)$$ for all measurable subsets $A\subseteq F$ and $t\geq 0$.]{}
To prove part (a), we essentially follow the proof of [@Barlow Theorem 7.20], and then apply results from [@MR]. First, observe that the definition of the resistance metric at (\[resdef\]) readily implies $$\label{rfluc}
\left|f(x)-f(y)\right|^2\leq \mathcal{E}(f,f)R(x,y)$$ for all $f\in\mathcal{F}$, $x,y\in F$. Hence $$f(x)^2\leq 2f(y)^2+2\left|f(x)-f(y)\right|^2\leq 2f(y)^2+2\mathcal{E}(f,f)R(x,y).$$ Using the local compactness of $(F,R)$, for any point $x\in F$, we can integrate the above over a compact neighbourhood of $x$ to obtain $f(x)^2\leq c\mathcal{E}_1(f,f)$ for any $f\in\mathcal{D}$, where $\mathcal{E}_1$ was defined at . We thus have that $f\mapsto f(x)$ is a bounded linear operator on the Hilbert space $(\mathcal{D},\mathcal{E}_1^{1/2})$, and so by the Riesz representation theorem there exists a function $u(x,\cdot)\in\mathcal{D}$ such that $$\label{repro}
\mathcal{E}_1(u(x,\cdot),f)=f(x)$$ for all $f\in\mathcal{D}$. From , we immediately obtain that $u(x,x)=\mathcal{E}_1(u(x,\cdot),u(x,\cdot))<\infty$. In combination with , this implies and the finiteness of $u(x,y)$ everywhere. Furthermore, if we define an operator on $L^2(F,\mu)$ by setting $Uf(x):=\int_Fu(x,y)f(y)\mu(dy)$, then by arguing exactly as in the proof of [@Barlow Theorem 7.20], one can check $\mathcal{E}_1(Uf,g)=\int_Ffg d\mu$ for every $f\in C_0(F)$ and $g\in \mathcal{D}$. It follows that $U$ agrees with the resolvent of $X$ on $C_0(F)$, i.e. $Uf(x):=E_x\int_0^\infty e^{-t}f(X_t)dt$ for all $f\in C_0(F)$, and extending the latter statement to all $f\in L^2(F,\mu)$ is elementary. By the continuity of the transition density in this setting, this implies that the function $u$ can alternatively be defined via . To complete the proof of part (a), we note that is proved in [@MR Theorem 3.6.5].
From part (a), we know that $E_x(e^{-\tau_y})$ is a jointly continuous function of $x,y\in F$. Thus, because we also know that all points of $F$ are regular for $X$, we can immediately apply the first part of [@GK Theorem 1] to obtain that $X$ admits jointly measurable local times $(L_t(x))_{x\in F,t\geq 0}$. Furthermore, since $X$ has a transition density, it holds that $\mu$ is a reference measure for $X$, i.e. $\mu(A)=0$ if and only if $U\mathbf{1}_A(x)=\int_0^\infty e^{-t}P_x(X_t\in A)dt=0$ for all $x\in F$ (see [@BG Definition V.1.1]). Thus we can apply the second part of [@GK Theorem 1] to confirm (\[occdens\]) holds.
We now describe background on time-changes of the Hunt process $X$ from [@FOT Section 6.2]. First suppose $\nu$ is an arbitrary positive Radon measure on $(F,R)$. As at (\[atdef\]), define a continuous additive functional $(A_t)_{t\geq 0}$ by setting $A_t:=\int_FL_t(x)\nu(dx)$, and let $(\tau(t))_{t\geq 0}$ be its right-continuous inverse, i.e. $\tau(t):=\inf\left\{s>0:\:A_s>t\right\}$. If $G\subseteq F$ is the closed support of $\nu$, then $((\tilde{X})_{t\geq 0},\:P_x,\: x\in G)$ is also a strong Markov process, where $\tilde{X}_t:=X_{\tau(t)}$; this is the *trace* of $X$ on $G$ (with respect to $\nu$). We also define a trace of the Dirichlet form $(\mathcal{E},\mathcal{D})$ on $G$, which we will denote by $(\tilde{\mathcal{E}},\tilde{\mathcal{D}})$, by setting $$\label{formtrace}
\tilde{\mathcal{E}}(g,g):=\inf\left\{\mathcal{E}(f,f):\:f\in\mathcal{D}_e,\:f|_{G}=g\right\},$$ $$\label{domaintrace}
\tilde{\mathcal{D}}:=\left\{g\in L^2(G,\nu):\:\tilde{\mathcal{E}}(g,g)<\infty\right\},$$ where $\mathcal{D}_e$ is the extended Dirichlet space associated with $(\mathcal{E},\mathcal{D})$, i.e. the family of $\mu$-measurable functions $f$ on $F$ such that $|f|<\infty$, $\mu$-a.e. and there exists an $\mathcal{E}$-Cauchy sequence $(f_n)_{n\geq 0}$ in $\mathcal{D}$ such that $f_n(x)\rightarrow f(x)$, $\mu$-a.e. Connecting these two notions is the following result.
[\[trace\] It holds that $(\tilde{\mathcal{E}},\tilde{\mathcal{D}})$ is a regular Dirichlet form on $L^2(G,\nu)$, and the associated Hunt process is $\tilde{X}$.]{}
Finally, we note a result that, in the recurrent case, characterises the trace of our Dirichlet form on a compact set. Note that the Dirichlet form $(\mathcal{E},\mathcal{D})$ is said to be recurrent if and only if $1\in \mathcal{D}_e$ and $\mathcal{E}(1,1)=0$.
[\[goodtrace\] If $(\mathcal{E},\mathcal{D})$ is recurrent and $G$ is compact, then $(\tilde{\mathcal{E}},\tilde{\mathcal{D}})$ is a regular resistance form on $G$, with associated resistance metric $R|_{G\times G}$.]{}
Since $(\mathcal{E},\mathcal{D})$ is recurrent, we have that $\mathcal{D}_e=\mathcal{F}$ (see [@KL Proposition 2.13]). Thus $$\label{etrace}
\tilde{\mathcal{E}}(g,g)=\inf\left\{\mathcal{E}(f,f):\:f\in\mathcal{F},\:f|_{G}=g\right\},$$ and also $\tilde{\mathcal{D}}=\{f|_G:\:f\in\mathcal{F}\}\cap L^2(G,\nu)$. By (\[rfluc\]), we moreover have that $\{f|_G:\:f\in\mathcal{F}\}\subseteq C(G)\subseteq L^2(G,\nu)$, and so $$\label{dtrace}
\tilde{\mathcal{D}}=\left\{f|_G:\:f\in\mathcal{F}\right\}.$$ Finally, we observe that and give that $(\tilde{\mathcal{E}},\tilde{\mathcal{D}})$ is the trace of the resistance form $(\mathcal{E},\mathcal{F})$ on $G$ in the sense of [@Kig Definition 8.3]. Since $G$ is closed, by [@Kig Theorem 8.4], this implies $(\tilde{\mathcal{E}},\tilde{\mathcal{D}})$ is also a regular resistance form on this set, with associated resistance metric $R|_{G\times G}$.
Gromov-Hausdorff-vague topology {#ghpsec}
-------------------------------
In this section we introduce the Gromov-Hausdorff-vague topology and an extension that we require. For more details regarding such metrics, see [@ADH; @ALWtop]. We start by defining a topology on $\mathbb{F}_c$, which is the subset of $\mathbb{F}$ containing elements $(F,R,\mu,\rho)$ such that $(F,R)$ is compact. In particular, for two elements $(F,R,\mu,\rho),(F',R',\mu',\rho')\in\mathbb{F}_c$, we set $\Delta_c((F,R,\mu,\rho),(F',R',\mu',\rho'))$ to be equal to $$\begin{aligned}
\label{ghpmetric}
\inf_{M,\psi,\psi'}\left\{d_M^H\left(\psi(F),\psi'(F)\right)+d_M^P\left(\mu\circ\psi^{-1},\mu'\circ\psi'^{-1}\right)+d_M(\rho,\rho')\right\},\end{aligned}$$ where the infimum is taken over all metric spaces $M=(M,d_M)$ and isometric embeddings $\psi:(F,R)\rightarrow (M,d_M)$, $\psi':(F',R')\rightarrow (M,d_M)$, and we define $d_M^H$ to be the Hausdorff distance between compact subsets of $M$, and $d_M^P$ to be the Prohorov distance between finite Borel measures on $M$. It is known that $\Delta_c$ defines a metric on the equivalence classes of $\mathbb{F}_c$ (where we say two elements of $\mathbb{F}_c$ are equivalent if there is a measure and root preserving isometry between them), see [@ADH Theorem 2.5].
To extend $\Delta_c$ to a metric on the equivalence classes of $\mathbb{F}$, we consider bounded restrictions of elements of $\mathbb{F}$. More precisely, for $(F,R,\mu,\rho)\in \mathbb{F}$, define $(F^{(r)},R^{(r)},\mu^{(r)},\rho^{(r)})$ by setting: $F^{(r)}$ to be the closed ball in $(F,R)$ of radius $r$ centred at $\rho$, i.e. $\bar{B}_R(\rho,r)$; $R^{(r)}$ and $\mu^{(r)}$ to be the restriction of $R$ and $\mu$ respectively to $F^{(r)}$, and $\rho^{(r)}$ to be equal to $\rho$. By assumption, $(F^{(r)},R^{(r)})$ is compact, and so to check that $(F^{(r)},R^{(r)},\mu^{(r)},\rho^{(r)})\in\mathbb{F}_c$ it will suffice to note that: $R^{(r)}$ is a resistance metric on $F^{(r)}$, the associated resistance form $(\mathcal{E}^{(r)},\mathcal{F}^{(r)})$ is regular, and $(\mathcal{E}^{(r)},\mathcal{F}^{(r)})$ is moreover a recurrent regular Dirichlet form. (These claims follow from Theorem \[trace\] and Lemma \[goodtrace\].)
As in [@ADH Lemma 2.8], we can check the regularity of the restriction operation with respect to the metric $\Delta_c$ to show that, for any two elements of the space $\mathbb{F}$, the map $r\mapsto \Delta_c((F^{(r)},R^{(r)},\mu^{(r)},\rho^{(r)}),(F'^{(r)},R'^{(r)},\mu'^{(r)},\rho'^{(r)}))$ is cadlag. (NB. In [@ADH Lemma 2.8], the metric spaces are assumed to be length spaces, but it is not difficult to remove this assumption.) This allows us to define a function $\Delta$ on $\mathbb{F}^2$ by setting $$\begin{aligned}
\lefteqn{\Delta\left((F,R,\mu,\rho),(F',R',\mu',\rho')\right)}\nonumber\\
&:=&\int_0^\infty e^{-r}\left(1\wedge\Delta_c((F^{(r)},R^{(r)},\mu^{(r)},\rho^{(r)}),(F'^{(r)},R'^{(r)},\mu'^{(r)},\rho'^{(r)}))
\right) dr,\label{ghvdef}\end{aligned}$$ and one can check that this is a metric on (the equivalence classes of) $\mathbb{F}$, cf. [@ADH Theorem 2.9], and also [@ALWtop Proof of Proposition 5.12]. The associated topology is the Gromov-Hausdorff-vague topology, as defined at [@ALWtop Definition 5.8]. From [@ALWtop Proposition 5.9], we have the following important consequence of convergence in this topology.
[\[embeddings\] Suppose $(F_n,R_n,\mu_n,\rho_n)$, $n\geq 1$, and $(F,R,\mu,\rho)$ are elements of $\mathbb{F}$ such that $(F_n,R_n,\mu_n,\rho_n)\rightarrow(F,R,\mu,\rho)$ in the Gromov-Hausdorff-vague topology. It is then possible to embed $(F_n,R_n)$, $n\geq 1$, and $(F,R)$ isometrically into the same (complete, separable, locally compact) metric space $(M,d_M)$ in such a way that, for Lebesgue-almost-every $r\geq 0$, $$\label{embedconv}
d_M^H\left(F_n^{(r)},F^{(r)}\right)\rightarrow 0,\qquad d_M^P\left(\mu_n^{(r)},\mu^{(r)}\right)\rightarrow 0,\qquad d_M(\rho_n^{(r)},\rho^{(r)})\rightarrow 0,$$ where we have identified the various objects with their embeddings.]{}
We next note that the measure bounds of UVD transfer to limits under the Gromov-Hausdorff-vague topology. The proof, which is an elementary consequence of the previous result, is omitted.
[Suppose $(F,R,\mu,\rho)\in\mathbb{F}$ is the limit with respect to the Gromov-Hausdorff-vague topology of a sequence $(F_n,R_n,\mu_n,\rho_n)_{n\geq 1}$ in $\mathbb{F}$ that satisfies UVD. It is then the case that $$\label{uvdforlimit}
c_1v(r)\leq \mu\left(B_{R}(x,r)\right)\leq c_2v(r),\qquad\forall x\in F,\:r\in[r_0,r_\infty+1],$$ where $r_0:= \inf_{x,y\in F,\:x\neq y}R(x,y)$ and $r_\infty:=\sup_{x,y\in F}R(x,y)$.]{}
Finally, we define an extended version of the Gromov-Hausdorff-vague topology for elements of the form $(F,R,\mu,\nu,\rho)$, where $(F,R,\mu,\rho)\in\mathbb{F}$, and $\nu$ is another locally finite Borel regular measure on $(F,R)$ (not necessarily of full support). We do this in the obvious way: for elements $(F,R,\mu,\nu,\rho)$ and $(F',R',\mu',\nu',\rho')$ such that $(F,R)$ and $(F',R')$ are compact, we include the term $d_M^P\left(\nu\circ\psi^{-1},\nu'\circ\psi'^{-1}\right)$ in the definition of $\Delta_c$ at (\[ghpmetric\]); in the general case, we use this version of $\Delta_c$ to define $\Delta((F,R,\mu,\nu,\rho),(F',R',\mu',\nu',\rho'))$ as at (\[ghvdef\]); the induced topology is then the extended Gromov-Hausdorff-vague topology. It is straightforward to check that the natural adaptation of Lemma \[embeddings\] that includes the convergence $d_M^P(\nu_n^{(r)},\nu^{(r)})$ also holds, where $\nu_n^{(r)}$, $\nu^{(r)}$ is the restriction of $\nu_n$, $\nu$ to $F_n^{(r)}$, $F^{(r)}$, respectively.
Local time continuity
---------------------
Key to our arguments is the following equicontinuity result for the local times of a sequence satisfying the UVD property. Since the proof is similar to the discrete time version proved for graphs in [@CroyLT Theorem 1.2], we only provide a sketch.
[\[ltcont\] If $(F_n,R_n,\mu_n,\rho_n)_{n\geq 1}$ is a sequence in $\mathbb{F}_c$ satisfying $\sup_nr_\infty(n)<\infty$ and also UVD, then, for each $\varepsilon>0$ and $T>0$, $$\lim_{\delta\rightarrow0}\sup_{n\geq 1}\sup_{x\in F_n}{P}^{n}_x\left(\sup_{\substack{y,z\in F_n:\\R_n(y,z)\leq\delta}}\sup_{0\leq t\leq T}{\left|L^{n}_{t}(y)-L_t^{n}(z)\right|}\geq \varepsilon\right)=0.$$]{}
We start by checking the commute time identity for a resistance form. In particular, if $(F,R,\mu,\rho)\in\mathbb{F}_c$, then we claim that $$\label{commute}
E_x\left(\tau_y\right)+E_y\left(\tau_x\right)=R(x,y)\mu(F)\qquad\forall x,y\in F,$$ where $\tau_z$ is the hitting time of $z$ by $X$. Indeed, fix $x,y\in F$. As in the proof of [@Kum Proposition 4.2], there exists a function $g_{\{x\}}(y,\cdot)\in \mathcal{F}$ such that: $\mathcal{E}(g_{\{x\}}(y,\cdot),f)=f(y)$ for every $f\in\mathcal{F}$ such that $f(x)=0$; $g_{\{x\}}(y,y)=\mathcal{E}(g_{\{x\}}(y,\cdot),g_{\{x\}}(y,\cdot))=R(x,y)$; and also $g_{\{x\}}(y,x)=0$. By symmetry, we deduce that $$\mathcal{E}\left(g_{\{x\}}(y,\cdot)+g_{\{y\}}(x,\cdot),f\right)=\mathcal{E}\left(g_{\{x\}}(y,\cdot),f-f(x)\right)+\mathcal{E}\left(g_{\{y\}}(x,\cdot),f-f(y)\right)=0$$ for every $f\in \mathcal{F}$. It follows that $g_{\{x\}}(y,\cdot)+g_{\{y\}}(x,\cdot)$ is constant, and so satisfies $$g_{\{x\}}(y,\cdot)+g_{\{y\}}(x,\cdot)\equiv g_{\{x\}}(y,x)+g_{\{y\}}(x,x)=R(x,y).$$ Moreover, as at [@Kum (4.7)], we have that $g_{\{x\}}(y,\cdot)$ is the occupation density for $X$, started at $y$ and killed at $x$, and so $E_y(\tau_x)=\int_Fg_{\{x\}}(y,z)\mu(dz)$. Combining the latter two results, the identity at (\[commute\]) follows.
We now suppose $(F_n,R_n,\mu_n,\rho_n)_{n\geq 1}$ is a sequence in $\mathbb{F}_c$ as in the statement of the lemma, and consider the associated local time processes. From [@BG (V.3.28)], we have that $$\label{tailforlt}
{P}^{n}_x\left(\sup_{0\leq t\leq T}{\left|L^{n}_{t}(y)-L_t^{n}(z)\right|}\geq \varepsilon\right)\leq 2e^Te^{-\varepsilon/2\delta_n(x,y)},$$ where $$\delta_n(x,y)^2:=1-E_x^n\left(e^{-\tau^n_y}\right)E_y^n\left(e^{-\tau^n_y}\right)\leq E_x^n\left(\tau^n_y\right)+E_y^n\left(\tau^n_x\right)=R_n(x,y)\mu_n(F_n),$$ and the final equality is a consequence of (\[commute\]). Hence we obtain that $$\sup_{x,y,z\in F_n}{P}^{n}_x\left(\sup_{0\leq t\leq T}\frac{\left|L^{n}_{t}(y)-L_t^{n}(z)\right|}{\sqrt{R_n(y,z)\mu_n(F_n)}}\geq \varepsilon\right)\leq 2e^Te^{-\varepsilon/2}.$$ Thus if we set $$\Gamma_n:=\int_{F_n}\int_{F_n}
\exp\Big(\frac{\sup_{0\leq t\leq T}\left|L^{n}_{t}(y)-L_t^{n}(z)\right|}{4\sqrt{R_n(y,z)\mu_n(F_n)}}\Big)\mu_n(dy)\mu_n(dz),$$ then it follows that $$\begin{aligned}
\label{Gamest}
\lim_{\lambda\rightarrow\infty}\sup_{n\geq 1}\sup_{x\in F_n}P_x^n\left(\Gamma_n> \lambda \mu_n(F_n)^2\right)=0.\end{aligned}$$ The result now follows from a standard argument involving Garsia’s lemma, as originally proved in [@Garsia], see also [@GRR]; applications to local times appear in [@BP; @CroyLT], for example. We simply highlight the differences. Choose $y,z\in F_n$ and $t\in[0,T]$. Then let $(K_i)_{i=0}^\infty$ be a sequence of balls $K_i=B_n(y,2^{1-2i}R_n(y,z))$, so that $K_0$ contains both $y$ and $z$, and $\cap_{i\geq 0}K_i=\{y\}$. Write $f_{K_i}:=\mu_n(K_i)^{-1}\int_{K_i}L^n_t(w)\mu_n(dw)$, and then we deduce that $$\begin{aligned}
\lefteqn{e^{|f_{K_i}-f_{K_{i-1}}|/16\sqrt{2^{-2i}R_n(y,z)\mu_n(F_n)}}}\\
&\leq& \frac{1}{\mu_n(K_i)\mu_n(K_{i-1})}\int_{K_i} \int_{K_{i-1}} e^{|L^n_t(w)-L^n_t(w')|/4\sqrt{R_n(w,w')\mu_n(F_n)}}\mu_n(dw)\mu_n(dw')\\
&\leq & cv(2^{1-2i}R_n(y,z))^{-2}\Gamma_n,\end{aligned}$$ where the first inequality is an application of Jensen’s inequality, and the second is obtained from UVD and the definition of $\Gamma_n$. Summing over $i$ and repeating for a sequence decreasing to $z$ yields $$\begin{aligned}
\label{lteee}
\left|L^{n}_{t}(y)-L_t^{n}(z)\right|\leq 16\sqrt{R_n(y,z)\mu_n(F_n)}\sum_{i=0}^\infty2^{-i}\log\left(\frac{c\Gamma_n}{v(2^{1-2i}R_n(y,z))^{2}}\right).\end{aligned}$$ Now, suppose $\Gamma_n\leq \lambda \mu_n(F_n)^2$. The UVD property then gives $\Gamma_n\leq c\lambda v(r_\infty(n))$. Together with the doubling property of $v$ and the assumption that $M=\sup_nr_\infty(n)<\infty$ we thus find that $$\begin{aligned}
{\left|L^{n}_{t}(y)-L_t^{n}(z)\right|}
&\leq & 16\sqrt{R_n(y,z)v(r_\infty(n))}\sum_{i=0}^\infty2^{-i}\log\left(\frac{c\lambda v(r_\infty(n))^2}{v(2^{1-2i}R_n(y,z))^{2}}\right)\nonumber\\
&\leq & c\sqrt{R_n(y,z)v(M)}\max\{1,\log \lambda^{1/c}M, \log R_n(y,z)^{-1}\},\nonumber\end{aligned}$$ uniformly over $y,z\in F_n$ and $t\in[0,T]$. Combining this estimate with (\[Gamest\]) completes the proof.
Note that we also have continuity of the limiting local times.
[\[ltcontlimit\] If $(F,R,\mu,\rho)\in\mathbb{F}_c$ satisfies (\[uvdforlimit\]), then the local times $(L_t(x))_{x\in F,t\geq 0}$ of the associated process are continuous in $x$, uniformly over compact intervals of $t$, $P_y$-a.s. for any $y\in F$.]{}
Arguing as for (\[Gamest\]), we have that $$\Gamma:=\int_{F}\int_{F}e^{\sup_{0\leq t\leq T}\left|L_{t}(y)-L_t(z)\right|/4\sqrt{R(y,z)\mu(F)}}\mu(dy)\mu(dz)$$ is a finite random variable, $P_y$-a.s., for any $T<\infty$. Hence, by applying the estimate (\[lteee\]), we obtain the result.
Convergence of processes {#copsec}
========================
Compact case {#compsec}
------------
In this section, we prove the first part of Theorem \[main1\] in the case that the metric spaces $(F_n,R_n)$, $n\geq 1$, and $(F,R)$ are all compact (see Proposition \[compactcase\] below). Throughout, we assume that Assumption \[a1\] holds. Note that, by Lemma \[embeddings\], under this Gromov-Hausdorff-vague convergence assumption, it is possible to suppose that $(F_n,R_n)$, $n\geq 1$, and $(F,R)$ are isometrically embedded into a common metric space $(M,d_M)$ such that $$\label{hconv}
d_M^H\left(F_n,F\right)\rightarrow 0,\qquad d_M^P\left(\mu_n,\mu\right)\rightarrow 0,\qquad d_M(\rho_n,\rho)\rightarrow 0,$$ where we have identified the various objects with their embeddings. Throughout this section, we fix one such collection of embeddings.
Our argument will depend on approximating the processes $X^n$, $n\geq 1$, and $X$ by processes on finite state spaces. We start by describing such a procedure in the limiting case. Let $(x_i)_{i\geq 1}$ be a dense sequence of points in $F$ with $x_1=\rho$. For each $k$, it is possible to choose $\varepsilon_k$ such that $$\label{kcov}
F\subseteq\cup_{i=1}^kB_M(x_i,\varepsilon_k),$$ (where $B_M(x,r)$ represents a ball in $(M,d_M)$,) and moreover one can do this in such a way that $\varepsilon_k\rightarrow 0$ as $k\rightarrow\infty$. Choose $\varepsilon^k_1,\varepsilon^k_2,\dots,\varepsilon^k_k\in [\varepsilon_k,2\varepsilon_k]$ such that $(B_M(x_i,\varepsilon_i^k))_{i=1}^k$ are continuity sets for $\mu$ (i.e. $\mu(\bar{B}_M(x_i,\varepsilon_i^k)\backslash B_M(x_i,\varepsilon_i^k))=0$); such a choice is possible because, for any $x\in M$, the map $r\mapsto \mu(B_M(x,r))$ has a countable number of discontinuities. Define sets $K^k_1,K^k_2,\dots,K^k_k$ by setting $K^k_1=\bar{B}_M(x_1,\varepsilon_1^k)$ and $$K^k_{i+1}=\bar{B}_M(x_{i+1},\varepsilon^k_{i+1})\backslash \cup_{j=1}^i \bar{B}_M(x_{j},\varepsilon^k_j).\label{eq:setdec}$$ In particular, the elements of the collection $(K^k_i)_{i=1}^k$ are measurable, disjoint continuity sets, and cover $F$. We introduce a corresponding measurable mapping $\phi^{(k)}:F\rightarrow \{x_1,\dots,x_k\}$ by setting $\phi^{(k)}(x)=x_i$ if $x\in K^k_i$, and a related measure $\mu^{(k)}=\mu\circ (\phi^{(k)})^{-1}$. Of course, the image of $\phi^{(k)}$ might not be the whole of $\{x_1,\dots,x_k\}$ since some of the $K_i^k$ might be empty. So, to better describe it, we introduce the notation $I_k:=\{i:\:K_i^k\neq\emptyset\}$ and $V_k:=\{x_i:\:i\in I_k\}$. (We will often implicitly use the fact that the points $(x_i)_{i\in I_k}$ are distinct, which follows from the definition.) The following simple lemma establishes that the measure $\mu^{(k)}$ charges all the points of $V_k$.
[\[muksupport\] The support of the measure $\mu^{(k)}$ is equal to $V_k$.]{}
Suppose $i\in \{1,\dots,k\}$ and $\mu^{(k)}(\{x_i\})=0$. Then by definition $$0=\mu(K_i^k)=\mu\left(\bar{B}_R(x_{i},\varepsilon^k_{i})\backslash \cup_{j=1}^{i-1} \bar{B}_R(x_{j},\varepsilon^k_j)\right)=\mu\left({B}_R(x_{i},\varepsilon^k_{i})\backslash \cup_{j=1}^{i-1} \bar{B}_R(x_{j},\varepsilon^k_j)\right),$$ where we use that ${B}_R(x_{i},\varepsilon^k_{i})$ is a continuity set for $\mu$. Now, ${B}_R(x_{i},\varepsilon^k_{i})\backslash \cup_{j=1}^{i-1} \bar{B}_R(x_{j},\varepsilon^k_j)$ is an open set. Thus, because $\mu$ has full support, the fact that the latter set has zero measure implies that it is empty. Hence ${B}_R(x_{i},\varepsilon^k_{i})\subseteq \cup_{j=1}^{i-1} \bar{B}_R(x_{j},\varepsilon^k_j)$. Since the right-hand side is closed, it follows that $\bar{B}_R(x_{i},\varepsilon^k_{i})\subseteq \cup_{j=1}^{i-1} \bar{B}_R(x_{j},\varepsilon^k_j)$, and therefore $K_i^k=\emptyset$. Thus $i\not\in I_k$. In particular, we have established that the support of $\mu^{(k)}$ contains $V_k$. Since the reverse inclusion is trivial, this completes the proof.
Next observe that $\sup_{x\in F}R(x,\phi^{(k)}(x))\leq 2\varepsilon_k\rightarrow 0$, and hence $\mu^{(k)}\rightarrow \mu$ weakly as measures on $F$. This will allow us to check that a family of associated time-changed processes $X^{(k)}$ converge to $X$. Indeed, set $$A^{(k)}_t=\int_FL_t(x)\mu^{(k)}(dx).$$ The continuity of the local times $L$ (see Lemma \[ltcontlimit\]) then implies that, $P_\rho$-a.s., for each $t$, $$A^{(k)}_t\rightarrow \int_FL_t(x)\mu(dx)=t.$$ Since the processes are increasing, this convergence actually holds uniformly on compact intervals (cf. the proof of Dini’s theorem). Setting $\tau^{(k)}(t):=\inf\{s>0:\:A^{(k)}_t>s\}$, it follows that, $P_\rho$-a.s., $\tau^{(k)}(t)\rightarrow t$ uniformly on compact intervals. Composing with the process $X$ to define $X^{(k)}_t:=X_{\tau^{(k)}(t)}$, we thus obtain that $X^{(k)}_t\rightarrow X_t$ for all $t\geq 0$ such that $X$ is continuous at $t$, $P_\rho$-a.s. In particular, denoting by $T_X$ the set of times $t$ such that $P_\rho(X\mbox{ is continuous at }t)=1$, this implies the following finite dimensional convergence result.
[\[l1\] If $t_1,\dots,t_m\in T_X$, then $d_M(X^{(k)}_{t_i},X_{t_i})\rightarrow0$ for each $i=1,\dots,m$, as $k\to\infty$, $P_\rho$-a.s.]{}
We next adapt the approximation argument to the processes $X^n$, $n\geq1$. By (\[hconv\]), it is possible to choose $x_i^n\in F_n$ such that $d_M(x_i^n,x_i)\rightarrow 0$, with the particular choice $x_1^n=\rho_n$. Moreover, by (\[kcov\]), it is possible to suppose that for each $k$ there exists an integer $n_k$ such that, for $n\geq n_k$, $F_n\subseteq \cup_{i=1}^kB_M(x_i,\varepsilon_k)$. Thus, for each $k$ and $n\geq n_k$ we can define a map $\phi^{n,k}:F_n\rightarrow \{x_1^n,\dots,x_k^n\}$ by setting $\phi^{n,k}(x)=x_i^n$ if $x\in K_i^k$. Note that $$\label{closepoints}
\lim_{k\rightarrow\infty}\limsup_{n\rightarrow\infty}\sup_{x\in F_n}R_n(x,\phi^{n,k}(x))\leq \lim_{k\rightarrow\infty}\limsup_{n\rightarrow\infty}\left(2\varepsilon_k+\sup_{i=1,\dots,k}d_M(x_i^n,x_i)\right)=0.$$ We define $\mu_n^{(k)}=\mu_n\circ(\phi^{n,k})^{-1}$, and set $$A^{n,k}_t=\int_{F_n}L^n_t(x)\mu_n^{(k)}(dx).$$ Moreover, let $\tau^{n,k}(t)=\inf\{s>0:\:A^{n,k}_t>s\}$, and define $X^{n,k}_t:=X^{n}_{\tau^{n,k}(t)}$. It is then straightforward to deduce the following lemma.
[\[l2\] The law of $X^{n,k}$ under $P^n_{\rho_n}$ converges weakly to the law of $X^{(k)}$ under $P_\rho$ as probability measures on the space $D(\mathbb{R}_+,M)$. In particular, the finite-dimensional distributions converge for any collection of times $t_1,\dots,t_m\geq 0$, $m\in\mathbb{N}$.]{}
Fix $k$, and define $V_k$ as above Lemma \[muksupport\]. Our first step is to characterise the Dirichlet form $(\mathcal{E}^{(k)},\mathcal{D}^{(k)})$ of the Markov chain $X^{(k)}$, which by Theorem \[trace\] is given by (\[formtrace\]), (\[domaintrace\]) with $G=V_k$ and $\nu=\mu^{(k)}$. Since $F$ is compact, we have that $(\mathcal{E},\mathcal{D})=(\mathcal{E},\mathcal{F})$ (see [@Kig p. 35]), and so $(\mathcal{E},\mathcal{D})$ is recurrent. Hence we have from Lemma \[goodtrace\] that $(\mathcal{E}^{(k)},\mathcal{D}^{(k)})$ is also a resistance form with associated resistance metric $R^{(k)}:=R|_{V_k\times V_k}$. In particular, we obtain that $$\mathcal{E}^{(k)}(f,f)=\frac12\sum_{x,y\in V_k}c^{(k)}(x,y)(f(y)-f(x))^2,$$ where the conductances $(c^{(k)}(x,y))_{x,y\in V_k}$ are uniquely determined by the resistance $R^{(k)}$ [@Kigdendrite Theorem 1.7].
We similarly have that the Dirichlet form $(\mathcal{E}^{n,k},\mathcal{D}^{n,k})$ of the Markov chain $X^{n,k}$ is given by $$\mathcal{E}^{n,k}(f,f)=\frac12\sum_{x,y\in V_{n,k}}c^{n,k}(x,y)(f(y)-f(x))^2,$$ where $V_{n,k}:=\{x^n_i:\:i\in I_k\}$, and we note that for large $n$ we have that the cardinality of $V_{n,k}$ and $V_k$ are both equal. We will now check that $$\label{condlim}
\left(c^{n,k}(x_i^n,x_j^n)\right)_{i,j\in I_k}\rightarrow \left(c^{(k)}(x_i,x_j)\right)_{i,j\in I_k}.$$ Observe that, from the definition of the resistance metric, we have $c^{n,k}(x_i^n,x_j^n)\leq R_n(x_i^n,x_j^n)^{-1}$. Hence we find that $$\limsup_{n\rightarrow\infty}\max_{\substack{i,j\in I_k:\\i\neq j}}c^{n,k}(x_i^n,x_j^n)\leq
\max_{\substack{i,j\in I_k:\\i\neq j}}R(x_i,x_j)^{-1}<\infty.$$ In particular, for any subsequence $(c^{n_m,k}(x_i^{n_m},x_j^{n_m}))_{i,j\in I_k}$, we have a convergent subsubsequence $(c^{n_{m_l},k}(x_i^{n_{m_l}},x_j^{n_{m_l}}))_{i,j\in I_k}$ with limit $(\tilde{c}(x_i,x_j))_{i,j\in I_k}$. Define an associated form $(\tilde{\mathcal{E}},\tilde{\mathcal{D}})$ by setting $$\tilde{\mathcal{E}}(f,f)=\frac12\sum_{x,y\in V_k}\tilde{c}(x,y)(f(y)-f(x))^2,$$ and $\tilde{\mathcal{D}}:=\{f:V_k\rightarrow \mathbb{R}\}$, and let $\tilde{R}$ be the associated resistance (which may *a priori* be infinite between pairs of vertices). It is then an elementary exercise to check that $c^{n_{m_l},k}\rightarrow \tilde{c}$ implies $(R^{n_{m_l}}(x_i^{n_{m_l}},x_j^{n_{m_l}}))_{i,j\in I_k}\rightarrow (\tilde{R}(x_i,x_j))_{i,j\in I_k}$. However, we also know $(R^{n_{m_l}}(x_i^{n_{m_l}},x_j^{n_{m_l}}))_{i,j\in I_k}\rightarrow ({R}^{(k)}(x_i,x_j))_{i,j\in I_k}$, and so it must be the case that $\tilde{R}=R^{(k)}$. In turn, this implies $\tilde{c}=c^{(k)}$ (see [@Kigdendrite Theorem 1.7]), and the conclusion at (\[condlim\]) follows as desired.
Next, note that for each $i\in I_k$ $$\label{mconv}
\mu_n^{(k)}(\{x_i^n\})=\mu_n\left(K_i^k\right)\rightarrow \mu\left(K_i^k\right)=\mu^{(k)}(\{x_i\})>0,$$ where we have applied that $\mu_n\rightarrow\mu$ weakly, and that $K_i^k$ is a continuity set for the limiting measure. The fact that the limit is strictly positive was proved in Lemma \[muksupport\]. These observations will allow us to check convergence of the generators. Specifically, the generator of $X^{(k)}$ is given by $$\Delta^{(k)}f(x_i)=\frac{1}{\mu^{(k)}(\{x_i\})}\sum_{j\in I_k}c^{(k)}(x_i,x_j)(f(x_j)-f(x_i)).$$ Similarly, if we define $\pi_{n,k}:V_{n,k}\rightarrow V_k$ by $x_i^n\mapsto x_i$ (which is a bijection for large $n$), then the generator of $\pi_{n,k}(X^{n,k})$ is given by $$\Delta^{n,k}f(x_i)=\frac{1}{\mu_{n}^{(k)}(\{x_i^n\})}\sum_{j\in I_k}c^{n,k}(x_i^n,x_j^n)(f(x_j)-f(x_i)).$$ Hence, (\[condlim\]) and (\[mconv\]) imply that $$\max_{i\in I_k}\left|\Delta^{(k)}f(x_i)-\Delta^{n,k}f(x_i)\right|\rightarrow 0$$ for any $f:V_k\rightarrow \mathbb{R}$. Since the starting points of the processes satisfy $\pi_{n,k}(X^{n,k}_0)= X^{(k)}_0=\rho$ (as local time accumulates immediately), this generator convergence is enough to establish the distributional convergence $\pi_{n,k}(X^{n,k})\rightarrow X^{(k)}$ (see [@Kall Theorem 19.25]). To complete the proof of the first claim, it is thus enough to recall that $d_M(x_i^n,\pi_{n,k}(x_i^n))=d_M(x_i^n,x_i)\rightarrow 0$ for each $i$.
For the claim regarding finite-dimensional distributions, one notes that convergence in the space $D(\mathbb{R}_+,M)$ implies convergence of finite-dimensional distributions at times $t_1,\dots,t_m$ that are continuity times for the process $X^{(k)}$, i.e. times at which $X^{(k)}$ is continuous, $P_\rho$-a.s. Furthermore, it is elementary to check that every $t\geq 0$ is a continuity time for the finite state space continuous time Markov chain $X^{(k)}$.
The remaining ingredient we need to establish the result of interest is the following lemma.
[\[l3\] The laws of $X^n$ under $P_{\rho_n}^n$, $n\geq 1$, form a tight sequence in $D(\mathbb{R}_+,M)$. Moreover, for any $\varepsilon>0$ and $t\geq 0$, $$\label{merging}
\lim_{k\rightarrow\infty}\limsup_{n\rightarrow\infty}P^{n}_{\rho_n}\left(R_n\left(X^n_t, X^{n,k}_t\right)>\varepsilon\right)=0.$$]{}
To verify tightness, it will suffice to check Aldous’ tightness criteria (see, for example, [@Kall Theorem 16.11]): for any bounded sequence of $X^n$ stopping times $\sigma_n$ and any sequence $\delta_n\rightarrow 0$, it holds that, for $\varepsilon>0$, $P_{\rho_n}^n(R_n(X^n_{\sigma_n},X^n_{\sigma_n+\delta_n})>\varepsilon)$. Applying the strong Markov property, to establish this it will be enough to show that $$\label{tightness}
\sup_{x\in F_n}P_{x}^n\left(R_n(x,X^n_{\delta_n})>\varepsilon\right)\rightarrow 0.$$ To do this, we note that the UVD condition implies the following exit time estimate $$\label{etest}
\sup_{x\in F_n}P^{n}_{x}
\left(\sup_{0\leq t\leq \delta}R_n\left(x,X^{n}_t\right)>\varepsilon\right)\leq c_1 e^{-\frac{c_2 \varepsilon}{v^{-1}(\delta/\varepsilon)}},$$ uniformly in $n$, where $v$ is the function appearing in the definition of UVD (see [@Kum Proposition 4.2 and Lemma 4.2]). Moreover, the doubling property of $v$ implies that $v(r)\geq c_3r^{c_4}$ for $r\leq 1$, and so $v^{-1}(\delta/\varepsilon)\leq c_5(\delta/\varepsilon)^{c_6}$ for $\delta$ suitable small. The result at (\[tightness\]) follows.
To prove (\[merging\]), first note that $$\begin{aligned}
\left|A^{n,k}_t-t\right|&=&\left|\int_{F_n}L^n_t(x)\mu_n^{(k)}(dx)-\int_{F_n}L^n_t(x)\mu_n(dx)\right|\\
&\leq&\int_{F_n}\left|L^n_t(\phi^{n,k}(x))-L^n_t(x)\right|\mu_n(dx)\\
&\leq & \mu_n(F_n)\sup_{x\in F_n}\left|L^n_t(\phi^{n,k}(x))-L^n_t(x)\right|.\end{aligned}$$ Now, by (\[hconv\]), $\mu_n(F_n)\rightarrow\mu(F)$, and the compactness of the space $(F,R)$ implies that the latter is a finite limit. Hence, also applying (\[closepoints\]) and the local time equicontinuity result of Lemma \[ltcont\], it follows that $$\lim_{k\rightarrow\infty}\limsup_{n\rightarrow\infty}
{P}^n_{\rho_n}\left(\sup_{0\leq t\leq T}\left|A^{n,k}_t-t\right|>\varepsilon\right)=0.$$ Taking inverses, we thus find that $$\lim_{k\rightarrow\infty}\limsup_{n\rightarrow\infty}
{P}^n_{\rho_n}\left(\sup_{0\leq t\leq T}\left|{\tau^{n,k}(t)}-t\right|>\varepsilon\right)=0.$$ From this, we see that, for any $t,\varepsilon,\delta\geq 0$, $$\begin{aligned}
{\lim_{\delta\rightarrow0}\limsup_{n\rightarrow\infty}P^{n}_{\rho_n}
\left(R_n\left(X^n_t, X^{n,k}_t\right)>\varepsilon\right)}&\leq&\lim_{\delta\rightarrow0}\limsup_{n\rightarrow\infty}P^{n}_{\rho_n}
\left(\sup_{s\in[t-\delta,t+\delta]\cap\mathbb{R}_+}R_n\left(X^n_t, X^{n}_s\right)>\varepsilon\right)\\
&\leq&\lim_{\delta\rightarrow0}\limsup_{n\rightarrow\infty}\sup_{x\in F_n}
P^{n}_{x}
\left(\sup_{s\in[0,2\delta]}R_n\left(x, X^{n}_s\right)>\varepsilon\right)\\
&=&0,\end{aligned}$$ where to deduce the second inequality, we apply the Markov property at time $\max\{0,t-\delta\}$, and (\[etest\]) to deduce the equality.
Piecing together Lemmas \[l1\] and \[l2\], and (\[merging\]), we obtain that the finite-dimensional distributions of $X^n$ converge to those of $X$ for any collection of times $t_1,\dots,t_m\in T_X$, where we recall that $T_X$ is the set of continuity times of $X$ (see [@Kall Theorem 4.28]). Together with the tightness of $X^n$, as established in Lemma \[l3\], we arrive at the desired conclusion by applying [@Kall Theorem 16.10].
[\[compactcase\] The law of $X^{n}$ under $P^n_{\rho_n}$ converges weakly to the law of $X$ under $P_\rho$ as probability measures on the space $D(\mathbb{R}_+,M)$.]{}
Locally compact case
--------------------
In this section, we explain how to extend from the compact case to the locally compact case. The proof will involve considering the trace of the relevant processes on bounded subsets (cf. the proof of [@BCK Theorem 1.4]). Key to this approach is the following lemma, which is an immediate consequence of Lemma \[goodtrace\]. (Recall that we are assuming $(\mathcal{E},\mathcal{D})$ is recurrent for $(F,R,\mu,\rho)\in\mathbb{F}$.)
[\[rtrace\] Let $(F,R,\mu,\rho)\in\mathbb{F}$. For $r\geq 0$, let $(\mathcal{E}^{(r)},\mathcal{D}^{(r)})$ be the trace of $(\mathcal{E},\mathcal{D})$ on $F^{(r)}$ with respect to the measure $\mu^{(r)}$. Then $(\mathcal{E}^{(r)},\mathcal{D}^{(r)})$ is a resistance form on $F^{(r)}$ with associated resistance metric $R^{(r)}$.]{}
A second key ingredient for our argument is the following uniform exit time estimate for sequences of resistance forms satisfying UVD.
[\[exittimes\] Suppose $(F_n,R_n,\mu_n,\rho_n)_{n\geq 1}$ is a sequence in $\mathbb{F}$ satisfying UVD, then, for any $T<\infty$, $$\lim_{r\rightarrow\infty}\sup_{n\geq 1}P_{\rho_n}^n\left(\sup_{0\leq t\leq T}R_n(\rho_n,X_t^n)> r\right)=0.$$]{}
Similarly to (\[etest\]), we have by [@Kum Proposition 4.2 and Lemma 4.2] that $$\sup_{n\geq 1}P^{n}_{\rho_n}
\left(\sup_{0\leq t\leq T}R_n\left(x,X^{n}_t\right)>r\right)\leq c_1 e^{-\frac{c_2 r}{v^{-1}(T/r)}}.$$ Letting $r\rightarrow\infty$ establishes the result.
We are now ready to prove the main result of this section, which establishes the first claim of Theorem \[main1\].
[\[procconv\] Suppose $(F_n,R_n,\mu_n,\rho_n)$, $n\geq 1$, and $(F,R,\mu,\rho)$ satisfy Assumption \[a1\]. It is then possible to embed $(F_n,R_n)$, $n\geq 1$, and $(F,R)$ isometrically into the same metric space $(M,d_M)$ in such a way that the law of $X^n$ under $P^n_{\rho_n}$ converges weakly to the law of $X$ under $P_{\rho}$ as probability measures on the space $D(\mathbb{R}_+,M)$.]{}
Under the assumption of the proposition, it is possible to suppose all the objects of the discussion have been isometrically embedded into a common metric space $(M,d_M)$ in the way described in Lemma \[embeddings\]. Define $(\mathcal{E}^{(r)},\mathcal{D}^{(r)})$ as in the statement of Lemma \[rtrace\]. By Theorem \[trace\], we have that this is a regular Dirichlet form on $L^2(F^{(r)}, \mu^{(r)})$, and the associated process $X^{(r)}$ is given by a time-change according to the additive functional $$A^{(r)}_t:=\int_FL_t(x)\mu^{(r)}(dx).$$ By monotonicity and the fact that the various additive functionals are increasing, we have that $A^{(r)}_t\rightarrow \int_FL_t(x)\mu(dx) =t$ uniformly on compact time intervals, $P_\rho$-a.s. Similarly to the proof of Lemma \[l1\], it follows that if $t_1,\dots,t_m\in T_X$ for any $m\in\mathbb{N}$ (where again we denote by $T_X$ the continuity times of $X$), then $P_\rho$-a.s., $d_M(X^{(r)}_{t_i},X_{t_i})\rightarrow 0$ for each $i=1,\dots,m$. Moreover, writing $\tau^{(r)}$ for the right-continuous inverse of $A^{(r)}$, we have that for any bounded sequence of $X^{(r)}$ stopping times $\sigma_r$ and any sequence $\delta_r\rightarrow0$, it holds that, for any $\varepsilon,\delta>0$, $$\begin{aligned}
P_\rho\left(R\left(X^{(r)}_{\sigma_r},X^{(r)}_{\sigma_r+\delta_r}\right)>\varepsilon\right)
&=&P_\rho\left(R\left(X_{\tau^{(r)}(\sigma_r)},X_{\tau^{(r)}(\sigma_r+\delta_r)}\right)>\varepsilon\right)\\
&\leq&\sup_{x\in F}P_x\left(\sup_{s\leq \delta}R\left(x,X_s\right)>\varepsilon\right)+o(1),\end{aligned}$$ as $r\rightarrow\infty$. (We note that $\tau^{(r)}(\sigma_r)$ is a stopping time for $X$.) Since by (\[uvdforlimit\]) we know that the limiting space $(F,R,\mu,\rho)$ satisfies uniform volume doubling, we can again apply [@Kum Proposition 4.2 and Lemma 4.2] as at (\[etest\]) to deduce that the probability above is bounded by $c_1e^{-c_2\varepsilon/v^{-1}(\delta/\varepsilon)}$. Letting $\delta\rightarrow 0$, we obtain that Aldous’ tightness criteria holds (cf. the proof of Lemma \[l3\]), and so the laws of $X^{(r)}$ under $P_\rho$ are tight in $D(\mathbb{R}_+,M)$. Combining this with the above convergence of finite dimensional distributions, we obtain that, under $P_\rho$, $X^{(r)}$ converges to $X$ in distribution in the space $D(\mathbb{R}_+,M)$.
Next, let $X^{n,r}$ be the trace of $X^n$ on $F_n^{(r)}$ with respect to $\mu_n^{(r)}$. By Lemma \[rtrace\], the Dirichlet form of this process, $(\mathcal{E}_n^{(r)},\mathcal{D}_n^{(r)})$ say, is actually a resistance form with associated resistance metric $R_n^{(r)}$, cf. the corresponding result in the limiting case. Hence, recalling we have embedded all the relevant objects into $M$ in the way described by Lemma \[embeddings\], Proposition \[compactcase\] yields that, for Lebesgue-almost-every $r\geq 0$, we have that the law of $X^{n,r}$ under $P^n_{\rho_n}$ converges weakly to the law of $X^{(r)}$ under $P_\rho$ as probability measures on the space $D(\mathbb{R}_+,M)$.
Finally, we observe that if $\sup_{0\leq t\leq T+1}R_n(\rho_n,X^n_t)\leq r$, then the time-change functional describing $X^{n,r}$ satisfies $$A^{n,r}_t=\int_{F_n}L_t^n(x)\mu_n^{(r)}(dx) =\int_{F_n}L_t^n(x)\mu_n(dx)=t$$ for $t\leq T+1$. It follows that $X^{n,r}_t=X^n_t$ for $t\leq T$. Thus we find that, for any $\varepsilon>0$, $$P_{\rho_n}^n\left(\sup_{0\leq t\leq T}R_n(X^n_t,X^{n,r}_t)>\varepsilon\right)
\leq P_{\rho_n}^n\left(\sup_{0\leq t\leq T+1}R_n(\rho_n,X^n_t)> r\right).$$ By Lemma \[exittimes\], this converges to $0$ as $r\rightarrow\infty$, uniformly in $n\geq 1$. Combining this with the conclusions of the previous two paragraphs completes the result.
Convergence of local times {#ltconvsec}
--------------------------
Again we suppose that the spaces $(F_n,R_n)$, $n\geq 1$, and $(F,R)$ are isometrically embedded into a common metric space $(M,d_M)$ in such a way that the conclusion of Lemma \[embeddings\] holds. Given the convergence result of Proposition \[procconv\] (and [@Kall Theorem 4.30], for example), it is further possible to suppose that $X^n$ started from $\rho_n$ and $X$ started from $\rho$ are coupled so that $X^n\rightarrow X$ in $D(\mathbb{R}_+,M)$, almost-surely. We will suppose that this is the case throughout this section, and write the joint probability measure as $P$. To prove the finite dimensional convergence of local times as at (\[ltconv1\]), we will follow an approximation argument, based on averaging over small balls. To this end, it is useful to introduce the following functions: for $x\in M$, $\delta>0$, $$f_{\delta,x}(y):=\max\left\{0,\delta-d_M(x,y)\right\}.$$ An immediate consequence of the continuity of local times of $X$ is the following lemma.
[\[ltl1\] $P$-a.s., for any $x\in F$ and $T\geq0$, as $\delta\rightarrow0$, $$\sup_{t\in[0,T]}\left|\frac{\int_0^tf_{\delta,x}(X_s)ds}{\int_Ff_{\delta,x}(y)\mu(dy)}- L_t(x)\right|\rightarrow 0.$$]{}
For the case when $F$ is compact, the result follows easily from Lemma \[ltcontlimit\] (and the occupation density formula of ). In the case when $F$ is only locally compact, we note that on the event $\sup_{t\in[0,T]}R(\rho,X_s)\leq r$ it is the case that the local times of $X$ are identical to the local times of $X^{(r)}$ up to time $T$. Since the latter are continuous functions for each $t>0$, almost-surely, then so are the local times of $X$ for $t\in[0,T]$, almost-surely on $\sup_{t\in[0,T]}R(\rho,X_s)\leq r$. Taking $r\rightarrow\infty$, and then $T\rightarrow\infty$, we deduce that the local times of $X$ are continuous functions for each $t>0$, almost-surely, and the result follows in this case as well.
[\[ltl2\]$P$-a.s., for any $x\in F$, $T\geq0$ and $\delta>0$, as $n\rightarrow\infty$, $$\sup_{t\in[0,T]}\left|\frac{\int_0^tf_{\delta,x}(X^n_s)ds}{\int_{F_n}f_{\delta,x}(y)\mu_n(dy)}-\frac{\int_0^tf_{\delta,x}(X_s)ds}{\int_Ff_{\delta,x}(y)\mu(dy)}\right| \rightarrow 0.$$]{}
Fix $x\in F$. It is then possible to choose $r$ such that $\bar{B}_M(x,\delta)\cap{F}\subseteq F^{(r)}$ for every $\delta<1$. Moreover, since our choice of embeddings satisfies the conclusions of Lemma \[embeddings\], we may further suppose that $\mu_n^{(r)}\rightarrow \mu^{(r)}$ weakly as probability measures on $M$. It follows that $${\int_{F_n}f_{\delta,x}(y)\mu_n(dy)}={\int_{M}f_{\delta,x}(y)\mu_n^{(r)}(dy)}
\rightarrow {\int_{M}f_{\delta,x}(y)\mu^{(r)}(dy)}={\int_{F}f_{\delta,x}(y)\mu(dy)}>0,$$ where the strict positivity of the limit is a simple consequence of the fact that $\mu$ has full support. Thus it remains to show that, for any $T\geq 0$, $$\label{c0}
\sup_{t\in [0,T]}\left|\int_0^tf_{\delta,x}(X^n_s)ds-\int_0^tf_{\delta,x}(X_s)ds\right|\rightarrow0.$$ To begin with, suppose that $X$ is continuous at time $t$. It is then the case that, for each $n$, there exists a homeomorphism $\lambda_n:[0,t]\rightarrow[0,t]$ with $\lambda_n(0)=0$ such that $$\label{c1}
\sup_{s\in[0,t]}\left|s-\lambda_n(s)\right|\rightarrow 0,$$ and also $$\label{c2}
\sup_{s\in[0,t]}d_M\left(X^n_{\lambda_n(s)},X_s\right)\rightarrow 0.$$ Now, $$\begin{aligned}
\left|\int_0^tf_{\delta,x}(X^n_{s})ds-\int_0^tf_{\delta,x}(X_{s})ds\right|&\leq&
\int_0^t\left|f_{\delta,x}(X^n_{\lambda_n(s)})-f_{\delta,x}(X_{s})\right|d\lambda_n(s)\\
&&+\int_0^t\left|f_{\delta,x}(X_{\lambda^{-1}_n(s)})-f_{\delta,x}(X_{s})\right|ds.\end{aligned}$$ The first term in the upper bound here converges to zero by (\[c2\]). As for the second term, we have from (\[c1\]) that $d_M(X_{\lambda^{-1}_n(s)},X_s)\rightarrow 0$ whenever $X$ is continuous at $s$. Since the times at which $X$ is not continuous is at most countable, the dominated convergence theorem yields that the second term also converges to zero, thereby establishing the limit (\[c0\]) pointwise at times at which $X$ is continuous. To extend to the full result is straightforward, using again that the times at which $X$ is not continuous is countable, as well as the monotonicity and continuity of the limit.
[\[ltl3\] For any $x\in F$ and $T\geq0$, if $x_n\in F_n$ is such that $d_M(x_n,x)\rightarrow 0$, then $$\lim_{\delta\rightarrow 0}\limsup_{n\rightarrow\infty}P\left(\sup_{t\in[0,T]}\left|\frac{\int_0^tf_{\delta,x}(X^n_s)ds}{\int_{F_n}f_{\delta,x}(y)\mu_n(dy)}-L^n_t(x_n)\right|>\varepsilon\right)=0.$$]{}
For large $n$, we have that by the occupation density formula $$\left|\frac{\int_0^tf_{\delta,x}(X^n_s)ds}{\int_{F_n}f_{\delta,x}(y)\mu_n(dy)}-L^n_t(x_n)\right|\leq
\sup_{y,z\in \bar{B}_n(x_n,2\delta)}\left|L^n_t(y)-L^n_t(z)\right|.$$ Thus if the sequence $(F_n,R_n,\mu_n,\rho_n)_{n\geq 1}$ satisfies $\sup_n r_\infty(n)<\infty$, then the result follows from the local time equicontinuity result of Lemma \[ltcont\]. In the general case, it is possible to obtain the result by considering the restriction to bounded subsets as in the last part of the proof of Proposition \[procconv\].
From Lemmas \[ltl1\], \[ltl2\] and \[ltl3\], we deduce that for any $x\in F$ and $T\geq 0$, if $x_n\in F_n$ is such that $d_M(x_n,x)\rightarrow 0$, then $(L_t^n(x_n))_{t\in[0,T]}\rightarrow(L_t(x))_{t\in[0,T]}$ in $P$-probability in $C([0,T],\mathbb{R})$. This result immediately extends to finite collections of points, which is enough to establish (\[ltconv1\]).
Time-changed processes
----------------------
In this section, we prove Corollary \[maincor\], starting by showing convergence of the time-change additive functionals.
[\[atn\] If Assumption \[a2\] holds, then $(A^n_t)_{t\geq0}\rightarrow (A_t)_{t\geq0}$ in distribution in the space $C(\mathbb{R}_+,\mathbb{R})$, simultaneously with the convergence of processes $X^n\rightarrow X$ in $D(\mathbb{R}_+,\mathbb{R})$, where we assume that $X^n$ is started from $\rho_n$, and $X$ is started from $\rho$. ]{}
We first prove the result in the case that the underlying spaces are compact, i.e. when $(F_n,R_n,\mu_n,\rho_n)\in\mathbb{F}_c$, $n\geq 1$, and $(F,R,\mu,\rho)\in\mathbb{F}_c$. Suppose all the objects are isometrically embedded into a common space in the way described at (\[hconv\]), and let $(x_i)_{i\geq 1}$ be as in Section \[compsec\]. Moreover, for each $k$, define $(K_i^k)_{i=1}^k$ as in , but with each set chosen to be a continuity set for the measure $\nu$, rather than for $\mu$. Then, for any $T\geq 0$, we have from the continuity of local times (Lemma \[ltcontlimit\]) that, $P_{\rho}$-a.s., $$\sup_{t\in[0,T]}\left|A_t-\sum_{i=1}^kL_t(x_i)\nu(K_i^k)\right|\leq \nu(F)\sup_{t\in[0,T]}\sup_{\substack{y,z\in F:\\R(y,z)\leq 4\varepsilon_k}}\left|L_t(y)-L_t(z)\right|\rightarrow 0.$$ Next, from (\[ltconv1\]), we deduce that $$\left(\sum_{i=1}^kL^n_t(x_i^n)\nu^n(K_i^k)\right)_{t\geq0}\rightarrow \left(\sum_{i=1}^kL_t(x_i)\nu(K_i^k)\right)_{t\geq0}$$ in distribution in $C(\mathbb{R}_+,\mathbb{R})$, where $x_i^n$ are also chosen as in Section \[compsec\]. Furthermore, for large $n$, we have that $$\sup_{t\in[0,T]}\left|A_t^n-\sum_{i=1}^kL^n_t(x_i^n)\nu^n(K_i^k)\right|\leq \nu_n(F_n)\sup_{t\in[0,T]}\sup_{\substack{y,z\in F_n:\\R_n(y,z)\leq 4\varepsilon_k}}\left|L^n_t(y)-L^n_t(z)\right|.$$ Under $P_{\rho_n}^n$, this converges to zero in probability as $n\rightarrow\infty$ and then $k\rightarrow\infty$ by Lemma \[ltcont\]. Noting that, from Theorem \[main1\], the convergence of local times at (\[ltconv1\]) occurs simultaneously with the convergence of processes, the desired result follows.
For the general case, one again proceeds by considering the restriction to bounded subsets similarly to the proof of Proposition \[procconv\]. For this, it is useful to note that it is enough to consider radii $r$ that are continuity sets for both $\mu$ and $\nu$, since the collections of points of discontinuity of the maps $r\mapsto \mu(B_R(\rho,r))$ and $r\mapsto \nu(B_R(\rho,r))$ are both countable.
We next check the divergence of the additive functional $(A_t)_{t\geq0}$, as defined at (\[atdef\]).
[For $(F,R,\mu,\rho)\in\mathbb{F}$, and $\nu$ a locally finite Borel regular measure on $(F,R)$ with $\nu(F)>0$, we have $A_t\rightarrow\infty$, $P_x$-a.s. for any $x\in F$.]{}
First note that, by [@FukuChen Theorem 5.2.16], we have that $(\mathcal{E},\mathcal{D})$ is an irreducible Dirichlet form (see [@FukuChen Section 2.1] for a definition). Since $(\mathcal{E},\mathcal{D})$ is recurrent, we can apply [@FukuChen Theorem 3.5.6(ii)] to deduce that $P_x(\tau_y<\infty)=1$, for all $x,y\in F$. Moreover, by [@MR Theorem 3.6.5], we have that $E_x(\int_0^\infty e^{-t}dL_t(x))=u(x,x)>0$, for all $x\in F$, where $(u(x,y))_{x,y\in F}$ is the potential density of $X$, as defined at (\[udef\]). Combining these two observations, following the proof of [@Croyrogt Lemma 2.3] allows us to deduce that $\lim_{t\rightarrow\infty}\inf_{x\in F^{(r)}}L_t(x)=\infty$ for any $r\geq0$, $P_y$-a.s. for any $y\in F$. This readily yields the result.
Note that the previous lemma implies that $\tau(t):=\inf\{s>0:A_s>t\}$ remains finite for all $t\geq 0$, and so confirms that $X_{\tau(t)}$ has an infinite lifetime. We are now in a position to complete the proof of Corollary \[maincor\].
First, suppose we are in case (a); in particular, $\nu(F)$ has full support. Moreover, suppose that we have embedded all the objects of the discussion into a common metric space $(M,d_M)$ in the way described by Lemma \[embeddings\], and that the various processes are coupled so that $X^n\rightarrow X$ in $D(\mathbb{R}_+,M)$, and $A^n\rightarrow A$ in $D(\mathbb{R}_+,\mathbb{R}_+)$, almost-surely. As in Section \[ltconvsec\], denote the probability measure corresponding to the coupling by $P$. Now, note that, $P$-a.s., for any $t,\delta>0$ we have that $\int_{F}(L_{t+\delta}(x)-L_t(x))\mu(dx)=\delta>0$, and so, applying the continuity of local times, we can find an $\varepsilon>0$ such that $L_{t+\delta}(x)-L_t(x)\geq\varepsilon$ on a non-empty open set. Since $\nu(F)$ has full support, it readily follows that $(A_t)_{t\geq 0}$ is strictly increasing, $P$-a.s. Thus we can apply [@Whittpaper Theorem 7.2], to deduce that $\tau^n\rightarrow \tau$ in $D(\mathbb{R}_+,\mathbb{R}_+)$, where the limiting function is strictly increasing and continuous, $P$-a.s. (Recall that $\tau^n$ is the right-continuous inverse of $A^n$, and $\tau$ is the right-continuous inverse of $A$.) Together with the convergence $X^n\rightarrow X$, this implies (see [@Whittpaper Theorem 3.1]) that $X^{n,\nu_n}\rightarrow X^\nu$ in $D(\mathbb{R}_+,M)$, $P$-a.s., which confirms the result.
The proof of part (b) is essentially the same, but involves different topologies. In particular, from $A^n\rightarrow A$ in $D(\mathbb{R}_+,\mathbb{R}_+)$, it is only possible in general to suppose $\tau^n\rightarrow \tau$ with respect to the Skorohod $M_1$ topology [@Whittpaper Theorem 7.1]. Given this convergence holds simultaneously with $X^n\rightarrow X$ in $D(\mathbb{R}_+,M)$, where $X$ is assumed to be continuous, we can apply the straightforward generalisation of [@CroyMuir Lemma A.6] to deduce the result.
Liouville Brownian motion
=========================
Given an element in $(F,R,\mu,\rho)\in \mathbb{F}$, the associated Liouville Brownian motion is the process $X^\nu$, defined as at (\[xnudef\]), where $\nu$ is the Liouville measure. To define this, let us first introduce the Gaussian free field on $F$, $(\gamma(x))_{x\in F}$ say, which we will suppose is pinned at $\rho$, and built on a probability space with probability measure $\mathbf{P}$ and expectation $\mathbf{E}$. In particular, we define $(\gamma(x))_{x\in F}$ to be a centred Gaussian field (i.e. $\mathbf{E}\gamma(x)=0$ for all $x\in F$), with covariances given by $${\rm{Cov}}(\gamma(x),\gamma(y))=g(x,y),\qquad \forall x,y\in F,$$ where $g(x,y)$ is the Green’s function of $X$ killed on hitting $\rho$ (cf. the notation $g_{\{\rho\}}$ in the proof of Lemma \[ltcont\]). Note that these assumptions imply that $\gamma(\rho)=0$, $\mathbf{P}$-a.s., and yield that an alternative way to characterise the covariances is via the formula $$\mathbf{E}\left(\left(\gamma(x)-\gamma(y)\right)^2\right)=R(x,y),\qquad \forall x,y\in F.$$ (To deduce the latter identity, it is useful to observe that $2g(x,y)=R(\rho,x)+R(\rho,y)-R(x,y)$, see [@Kig Theorem 4.3].) Thus we have from standard estimates for Gaussian random variables that $$\label{tailforgff}
\mathbf{P}\left(\left|\gamma(x)-\gamma(y)\right|\geq \varepsilon\right)\leq 2 e^{-\frac{\varepsilon^2}{2R(x,y)}},$$ and substituting this for the estimate (\[tailforlt\]), one can follow the proof of Lemma \[ltcont\] to deduce that, if $(F,R,\mu,\rho)$ satisfies the volume doubling estimates of (\[uvdforlimit\]), then $(\gamma(x))_{x\in F}$ is a continuous function, $\mathbf{P}$-a.s. (To check this continuity property, one might alternatively note that (\[uvdforlimit\]) yields an estimate for the size of a $\varepsilon$-cover of $F^{(r)}$ of the form $c_1\varepsilon^{-c_2}$, and from this the result is an application of [@MRa Theorem 8.6], for example.) In this case, for $\kappa>0$ fixed, setting (similarly to (\[2dlmeasure\])) $$\nu(dx)=e^{\kappa\gamma(x)-\frac{\kappa^2}{2}\mathbf{E}(\gamma(x)^2)}\mu(dx),$$ yields a locally finite, Borel regular measure on $(F,R)$ of full support, $\mathbf{P}$-a.s. (Note also that this choice of normalisation yields $\mathbf{E}\nu(dx)=\mu(dx)$.) Thus, for $\mathbf{P}$-a.e. realisation of $\nu$, we can define $X^\nu$ by the procedure at (\[xnudef\]). Since under $P_x$ the starting point of $X^\nu$ is $x$, the corresponding *quenched law* of $X^\nu$ started from $x\in F$ is well-defined; we will denote this by $P^\nu_x$. Moreover, we can define the *annealed law* of the Liouville Brownian motion $X^\nu$ by integrating out the Liouville measure, i.e.$$\label{anndef}
\mathbb{P}^{\rm LBM}_x\left(\cdot\right):=\int P_x^\nu\left(\cdot\right)\mathbf{P}(d\nu).$$
The principal aim of this section is to show that if we have $(F_n,R_n,\mu_n,\rho_n)\rightarrow (F,R,\mu,\rho)$ in $\mathbb{F}$ and the UVD property holds (i.e. Assumption \[a1\] is satisfied), then the associated Liouville measures and Liouville Brownian motions converge. To this end, we start by noting the equicontinuity of the Gaussian free fields in the sequence, which we will denote by $(\gamma_n(x))_{x\in F_n}$, $n\geq 1$. As for the continuity of $\gamma$, the proof of this result is identical to that of the local time equicontinuity result of Lemma \[ltcont\], with replaced by , and so is omitted.
[\[gammanequi\] If $(F_n,R_n,\mu_n,\rho_n)_{n\geq 1}$ is a sequence in $\mathbb{F}$ satisfying UVD, then, for each $\varepsilon>0$ and $r>0$, $$\lim_{\delta\rightarrow0}\sup_{n\geq 1}\mathbf{P}\left(\sup_{\substack{y,z\in F^{(r)}_n:\\R_n(y,z)\leq\delta}}{\left|\gamma_n(y)-\gamma_n(z)\right|}\geq \varepsilon\right)=0.$$]{}
We can now deduce convergence of Liouville measures under Assumption \[a1\]; the following result can be interpreted as a distributional version of Assumption \[a2\]. We write $\nu_n$ for the Liouville measure associated with $(F_n,R_n,\mu_n,\rho_n)$.
[\[nunconv\] Suppose Assumption \[a1\] holds, and that $(F_n,R_n,\mu_n,\rho_n)$, $n\geq 1$, and $(F,R,\mu,\rho)$ are isometrically embedded into a common (complete, separable, locally compact) metric space $(M,d_M)$ so that the conclusion of Lemma \[embeddings\] holds. It is then the case that $\nu_n\rightarrow \nu$ in distribution with respect to the vague topology for locally finite Borel measures on $(M,d_M)$.]{}
By [@Kall Theorem 16.16], it will suffice to show that $$\label{fint}
\int_Mf(x)\nu_n(dx)\rightarrow\int_Mf(x)\nu(dx)$$ in distribution, for any non-negative, continuous, compactly supported $f:M\rightarrow \mathbb{R}$. For each such $f$, we note that the support of $f$ is contained in ${B}_M(\rho,r/2)$ for some $r>0$ for which holds. Moreover, under the assumptions of the lemma, we have that $(f(x)e^{\kappa\gamma(x)-\frac{\kappa^2}{2}\mathbf{E}(\gamma(x)^2)})_{x\in F}$ is a continuous function on $F$, $\mathbf{P}$-a.s., and Lemma \[gammanequi\] implies the equicontinuity of the functions $(f(x)e^{\kappa\gamma_n(x)-\frac{\kappa^2}{2}\mathbf{E}(\gamma_n(x)^2)})_{x\in F_n}$. Consequently, the result at can be proved in the same way as Proposition \[atn\] if we can show the analogue of (\[ltconv1\]) in this setting, i.e. if $(x_i^n)_{i=1}^k$ in $F_n$, $n\geq 1$, are such that $d_M(x_i^n,x_i)\rightarrow 0$ for some $(x_i)_{i=1}^k$ in $F$, then it holds that $$\left(\gamma_n\left(x_i^n\right)\right)_{i=1,\dots, k}\rightarrow \left(\gamma\left(x_i\right)\right)_{i=1,\dots, k},$$ in distribution in $\mathbb{R}^k$. However, this is straightforward, since all the random variables above are centred Gaussian random variables, and $$\left({\rm{Cov}}(\gamma_n(x_i^n),\gamma(x_j^n))\right)_{i,j=1,\dots,k}\rightarrow\left({\rm{Cov}}(\gamma(x_i),\gamma(x_j))\right)_{i,j=1,\dots,k},$$ where to deduce the latter convergence, it is again helpful to apply the identity $2g(x,y)=R(\rho,x)+R(\rho,y)-R(x,y)$ from [@Kig Theorem 4.3].
From this convergence of Liouville measures, a simple adaptation of Corollary \[maincor\] yields the convergence of Liouville Brownian motion in this setting. In particular, by the separability of the space of locally finite Borel measures on $(M,d_M)$ under the vague topology, we can suppose $\nu_n$ and $\nu$ are coupled so that the conclusion of Lemma \[nunconv\] holds $\mathbf{P}$-a.s. (see [@Kall Theorem 4.30], for example). Under this coupling, the proof of Corollary \[maincor\] yields the almost-sure convergence of quenched laws, i.e. $P_{\rho_n}^{n,\nu_n}\rightarrow P_{\rho}^{\nu}$, weakly as probability measures on $D(\mathbb{R}_+,M)$, $\mathbf{P}$-a.s., where, for $\nu_n$ given, $P_{\rho_n}^{n,\nu_n}$ is the law of $X^{n,\nu_n}$ started from $\rho_n$. Integrating out the above result with respect to $\mathbf{P}$ then gives the following, where $\mathbb{P}^{{\rm LBM}_n}_{\rho_n}$ is the annealed law of $X^{n,\nu_n}$ started from $\rho_n$.
[\[lbmconv\] Suppose Assumption \[a1\] holds. It is then possible to isometrically embed $(F_n,R_n,\mu_n,\rho_n)$, $n\geq 1$, and $(F,R,\mu,\rho)$ into a common metric space $(M,d_M)$ so that $$\mathbb{P}^{{\rm LBM}_n}_{\rho_n}\rightarrow\mathbb{P}^{{\rm LBM}}_{\rho}$$ weakly as probability measures on $D(\mathbb{R}_+,M)$.]{}
\[lbmexamples\]
As simple examples, one might consider graphical approximations to tree-like and low-dimensional fractal spaces. In the following, we briefly introduce some of these.
\(i) The most basic example would be to set $F_n$ to be the integer lattice $\mathbb{Z}$ equipped with the rescaled Euclidean distance $R_n(x,y)=n^{-1}|x-y|$, counting measure $\mu_n(A):=n^{-1}|A|$, and distinguished vertex $\rho_n=0$. Then Assumption \[a1\] is satisfied with limit $(F,R,\mu,\rho)$ given by $F=\mathbb{R}$, $R$ the Euclidean metric, $\mu$ one-dimensional Lebesgue measure, and $\rho=0$.
\(ii) More generally than the previous example, one might consider a family of graph trees $(G_n)_{n\geq1}$ for which there exist scaling factors $(a_n)_{n\geq 1}$, $(b_n)_{n\geq1}$ such that $(G_n,a_nR_n,b_n\mu_n,\rho_n)_{n\geq1}$ satisfies Assumption \[a1\] for some limiting tree $(F,R,\mu,\rho)\in \mathbb{F}$, where: $R_n$ is the resistance metric associated with unit resistances along edges of $G_n$; $\mu_n$ is the counting measure on $G_n$; and $\rho_n$ is a distinguished vertex. (Here and in the following, we call $G$ a graph tree if it is connected and contains no cycle.) In this setting the resistance metric is identical to the usual shortest path graph distance, for which the assumptions are generally easier to check. For example, it is elementary to check the result for the graphs approximating the Vicsek set, as shown in Figure \[vicsek\], with $a_n=3^{-n}$, $b_n=5^{-n}$.
\(iii) It is known that the resistance metric on the graphs approximating nested fractals, again when unit resistors are placed along edges, can be rescaled to yield a resistance metric on the limiting fractal (this is an application of the homogenisation result of [@peir Theorem 3.8], for example). In Section \[rcmfractalsec\], we introduce a more general class of fractals, so leave details until later (alternatively, see [@lind] for background on nested fractals). However, as an illustrative example, we note that the Assumption \[a1\] applies to the sequence $(G_n,a_nR_n,b_n\mu_n,\rho_n)$, $n\geq 1$, where: $G_n$ is the $n$th level Sierpiński gasket graph, as introduced at the end of Section $1$ (see Figure \[sg\]); $R_n$ is the associated resistance metric; $\mu_n$ is the counting measure on vertices; $\rho_n$ can be chosen arbitrarily as long as $(\rho_n)_{n\geq1}$ converges in $\mathbb{R}^2$; and the scaling factors are given by $a_n=(3/5)^n$, $b_n=3^{-n}$. Furthermore, we note that the results in this section also establish that $$a_n^{1/2}\sup_{x\in G_n}\gamma_{G_n}(x)\rightarrow \sup_{x\in F}\gamma(x)$$ in distribution, where $\gamma_{G_n}$ is the Gaussian free field on $G_n$, and $\gamma$ the Gaussian free field on the limiting fractal; this refines the result of [@KZ Theorem 2.2] for these examples.
\(iv) Finally, another fractal for which our present setting is appropriate is the two-dimensional Sierpiński carpet and its graphical approximations, as shown in Figure \[sc\]. (Again, the results would apply to other low-dimensional carpets.) Whilst the exact resistance scaling factor is not known in this case, previous results allow us to control the resistance in terms of the graph distance (cf. the comments in [@CroyLT Section 5.4]). It follows that there exist subsequences $(G_{n_i},a_{n_i}R_{n_i},b_{n_i}\mu_{n_i},\rho_{n_i})$ (where again $R_n$ is the resistance metric on $G_n$ with unit edge resistances, and $\mu_n$ is counting measure) that satisfy Assumption \[a1\] with $a_n=\gamma^n$ for some $\gamma\in (0,1)$, and $b_n=8^{-n}$.
Bouchaud trap model {#bouchsec}
===================
We start this section by introducing the (symmetric) Bouchaud trap model (BTM) on a locally finite, connected graph $G=(V_G,E_G)$. To do this, we first introduce a trapping landscape $\xi=(\xi_x)_{x\in V_G}$, which is a collection of independent and identically distributed strictly-positive random variables built on a probability space with probability measure $\mathbf{P}$. Conditional on $\xi$, the dynamics of the BTM are then given by a continuous-time $V_G$-valued Markov chain $X^\xi=(X^\xi_t)_{t\geq 0}$ with jump rate from $x$ to $y$ given by $1/\xi_x$ if $\{x,y\}\in E_G$, and jump rate 0 otherwise. The quenched law of $X^\xi$ started from $x$ (i.e. the law given $\xi$) will be denoted $P_x^\xi$, and the corresponding annealed law, obtained by integrating out $\xi$, by $\mathbb{P}^{\rm BTM}_x$ (cf. (\[anndef\])).
To put the BTM into the framework of this article, we note that it can be obtained as a time-change of the continuous time simple random walk on $G$. In particular, let $R_G$ be the resistance metric on $V_G$ obtained by placing unit resistors along edges, and $\mu_G$ be the counting measure on $V_G$ (i.e. $\mu_G(\{x\})=1$ for all $x\in V_G$). As in Section \[rfsec\], we can naturally associate a process $X$ with the triple $(V_G,R_G,\mu_G)$, and it is an elementary exercise to check that this is the continuous-time $V_G$-valued Markov chain with unit jump rate along edges. Moreover, by time-changing $X$ as at (\[xnudef\]) according to the measure $\nu_G$ defined by setting $\nu_G(\{x\})=\xi_x$, we obtain $X^\xi$.
Similarly to the previous section, our goal is to present scaling limits of the BTM for sequences of recurrent graphs which, when equipped with resistance metrics and counting measure, satisfy UVD and converge in the Gromov-Hausdorff-vague topology to a limit in $\mathbb{F}$. We will do this in the case when the trapping environment is heavy-tailed. More specifically, in this section we make the following assumption.
[\[btmassu\] Suppose $(G_n)_{n\geq 1}$ is a sequence of locally finite, connected graphs with vertex sets $V_n$, resistance metrics $R_n$ (as above, here we assume that individual edges have unit resistance), counting measures $\mu_n$, and distinguished vertices $\rho_n$. Suppose further that each graph $G_n$ is recurrent, so that $(V_n,R_n,\mu_n,\rho_n)\in \mathbb{F}$. Moreover, assume that there exist scaling factors $(a_n)_{n\geq 1}$, $(b_n)_{n\geq 1}$ such that $(V_n,a_nR_n,b_n\mu_n,\rho_n)_{n\geq 1}$ satisfy Assumption \[a1\], where the measure $\mu$ of the limiting space $(F,R,\mu,\rho)\in \mathbb{F}$ is non-atomic. Finally, we suppose that each $G_n$ is equipped with a trapping landscape $\xi^n=(\xi^n_x)_{x\in V_n}$ such that $$\label{alphatail}
\mathbf{P}\left(\xi^n_x>u\right)\sim u^{-\alpha}$$ for some fixed $\alpha\in(0,1)$, where $f (u)\sim g (u)$ means $\lim_{u\to\infty}f(u)/g(u)=1$.]{}
We next describe the limits of the trapping landscape and BTM under the above assumption. We will show that the former is given by the natural generalisation to (\[onedfinmeasure\]) obtained by setting $$\nu(dx):=\sum_{i}v_i\delta_{x_i}(dx),$$ where $(v_i,x_i)$ are the points of a Poisson process on $(0,\infty)\times F$ with intensity $\alpha v^{-1-\alpha}dv\mu(dx)$, and $\delta_{x}$ is the probability measure on $F$ that places all its mass at $x$; note that this is a locally finite, Borel regular measure on $(F,R)$ of full support, $\mathbf{P}$-a.s. (where we suppose $\nu$ is also built on the probability space with probability measure $\mathbf{P}$). Moreover, the latter is given by $X^{\nu}$, that is, the time-change of the process $X$ naturally associated with $(F,R,\mu)$ by the measure $\nu$. Reflecting the terminology for the corresponding one-dimensional object, we will call this the $\alpha$-FIN process on $(F,R,\mu)$. In general, this is not a diffusion, but under our assumptions it will be whenever $X$ is, and in this case we will call it the $\alpha$-FIN diffusion. Given $\nu$, the quenched law of $X^\nu$ started from $x$ will be denoted by $P^{\nu}_x$, and the associated annealed law $\mathbb{P}^{\rm FIN}_x$.
The following lemma establishes convergence of the trapping landscapes. We write $\nu_n$ for the measure on $V_n$ induced by the trapping landscape $\xi^n$.
[\[nunconv2\] Suppose Assumption \[btmassu\] holds, and $(V_n,a_nR_n,b_n\mu_n,\rho_n)$, $n\geq 1$, and $(F,R,\mu,\rho)$ are isometrically embedded into a common (complete, separable, locally compact) metric space $(M,d_M)$ so that the conclusion of Lemma \[embeddings\] holds. It is then the case that $b_n^{1/\alpha}\nu_n\rightarrow \nu$ in distribution with respect to the vague topology for locally finite Borel measures on $(M,d_M)$.]{}
By [@Kall Theorem 16.16] (and the fact that measures of disjoint sets are independent under both $\nu_n$ and $\nu$), it will suffice to show that $b_n^{1/\alpha}\nu_n(B)\rightarrow \nu(B)$ in distribution, for every relatively compact set $B\subseteq M$ such that $B$ is a continuity set for $\nu$, $\mathbf{P}$-a.s. Since $\nu(B)=0$, $\mathbf{P}$-a.s., if and only if $\mu(B)=0$, the latter requirement is equivalent to supposing $B$ is a continuity set for $\mu$. For such a $B$, we have by assumption that $b_n\mu_n(B)\rightarrow \mu(B)$. Hence we have by an elementary computation that $$\label{laplaceconv}
\mathbf{E}\left(e^{-\lambda b_n^{1/\alpha}\nu_n(B)}\right)=\mathbf{E}\left(e^{-\lambda b_n^{1/\alpha}\sum_{x\in B}\xi^n_x}\right)=\left(1-\lambda^\alpha b_n \Gamma(1-\alpha)+o(b_n)\right)^{\mu_n(B)}\rightarrow e^{-\lambda^\alpha \Gamma(1-\alpha)\mu(B)}.$$ Moreover, it is a simple application of Campbell’s theorem [@Kingman (3.17)] that $$\mathbf{E}\left
(e^{-\lambda\nu(B)}\right)
=\mathbf{E}\left(e^{-\sum_{i:x_i\in B}\lambda v_i}\right)=e^{-\int_{(0,\infty)\times B}\left(1-e^{-\lambda v}\right)\alpha v^{-1-\alpha}dv\mu(dx)}=e^{-\lambda^\alpha\Gamma(1-\alpha)\mu(B)},$$ and so we are done.
In light of Lemma \[nunconv2\], and incorporating the scaling factors where appropriate, we can proceed exactly as for the proof of Proposition \[lbmconv\] to deduce convergence of the rescaled BTMs. We write $\mathbb{P}^{{\rm BTM}_n}_x$ for the annealed law of the BTM $X^{n,\xi^n}$ on $G_n$.
[\[btmresult\] Suppose Assumption \[btmassu\] holds. It is then possible to isometrically embed $(V_n,a_nR_n,b_n\mu_n,\rho_n)$, $n\geq 1$, and $(F,R,\mu,\rho)$ into a common metric space $(M,d_M)$ so that $$\mathbb{P}^{{\rm BTM}_n}_{\rho_n}\left(\left(X^{n,\xi^n}_{t/a_nb_n^{1/\alpha}}\right)_{t\geq 0}\in\cdot\right)
\rightarrow\mathbb{P}^{{\rm FIN}}_{\rho}\left(\left(X^{\nu}_{t}\right)_{t\geq 0}\in\cdot\right)$$ weakly as probability measures on $D(\mathbb{R}_+,M)$.]{}
[\[SG55ex\] [As applications of Proposition \[btmresult\], we can consider the same spaces as discussed in Example \[lbmexamples\]. For instance, for the BTM on the graph approximations to the Sierpiński gasket $(G_n)_{n\geq1}$ of Example \[lbmexamples\](iii), we have the convergence of the annealed law of $(X^{n,\xi^n}_{3^{n/\alpha}(5/3)^nt})_{t\geq 0}$ to the annealed law of the $\alpha$-FIN diffusion on the Sierpiński gasket.]{}]{}
Random conductance model {#rcmsec}
========================
We now recall from the introduction the random conductance model; this is defined similarly to the BTM, but with random weights now assigned to the edges rather than to the vertices. As in the previous section, let $G=(V_G,E_G)$ be a locally finite, connected graph. Let $\omega=(\omega_e)_{e\in E_G}$ be a collection of independent and identically distributed strictly-positive random variables built on a probability space with probability measure $\mathbf{P}$; these are the so-called *random conductances*. (Actually, for our model of self-similar fractals, we will allow some local dependence.) Conditional on $\omega$, we define the *variable speed random walk (VSRW)* $X^\omega=(X^\omega_t)_{t\geq 0}$ to be the continuous-time $V_G$-valued Markov chain with jump rate from $x$ to $y$ given by $\omega_{xy}$ if $\{x,y\}\in E_G$, and jump rate 0 otherwise. We obtain the associated *constant speed random walk (CSRW)* $X^{\omega,\nu}=(X^{\omega,\nu}_t)_{t\geq 0}$ by setting the jump rate along edge $x$ to $y$ to be $\omega_{xy}/\nu(\{x\})$, where $$\nu\left(\{x\}\right):=\sum_{e\in E_G:\:x\in e}\omega_e;$$ note that this is the time-change of $X^\omega$ according to the measure $\nu$, and has unit mean holding times at each vertex.
An important observation is that the VSRW and CSRW experience different trapping behaviour on edges of large conductance. In particular, if we have an edge of conductance $\omega_e\gg1$ (surrounded by other edges of conductance close to 1), then both the VSRW and CSRW cross the edge order $\omega_e$ times before escaping. However, each crossing only takes the VSRW a time of $1/\omega_e$, meaning that it is only trapped for a time of order 1, whereas each crossing for the CSRW takes a time of order 1, and so the latter process is trapped for a total time of order $\omega_e$. In particular, when the weights are bounded below, we might typically expect the VSRW associated with the conductances $\omega$ to behave like the VSRW on the unweighted graph, which in each of the examples we consider converges under scaling to Brownian motion on the limiting space. Moreover, we might expect the CSRW to behave like the Bouchaud trap model with trapping environment described by $\nu$, and therefore we expect to see FIN-type scaling limits for this process when the conductances are heavy-tailed.
The aim of this section is to make the heuristics of the previous paragraph rigourous, in the sense that we will show for the random conductance model on certain sequences of graphs that, if the weights are chosen to satisfy (similarly to ) $$\label{omegatail}
\mathbf{P}\left(\omega_e>u\right)\sim u^{-\alpha}$$ for some fixed $\alpha\in(0,1)$, then the rescaled VSRW $X^\omega$ converges to the canonical Brownian motion on the limit space, and the rescaled CSRW $X^{\omega,\nu}$ converges to the $\alpha$-FIN diffusion. The two classes we discuss are graph trees, and a family of self-similar fractals.
Random conductance model on trees {#rcmtreesec}
---------------------------------
In this section, we will study the scaling limit of the VSRW and CSRW for the random conductance model on sequences of graph trees; our main result is Proposition \[rcmtreeresult\]. As for the Bouchaud trap model, we will need to show that the associated time-change measures converge. The additional part of the argument will be to check that we also have homogenisation of the resistance metric when random conductances are placed along edges. In this setting, this is straightforward, since we can apply the law of large numbers along paths. We start by stating the main assumption of this section, which closely matches Assumption \[btmassu\]. The restriction to compact spaces is only for convenience of presentation, and not essential.
[\[rcmtreeassu\] Suppose $(T_n)_{n\geq 1}$ is a sequence of finite graph trees with vertex sets $V_n$, edge sets $E_n$, resistance metrics $R_n$ (here we assume that individual edges have unit resistance), counting measures $\mu_n$, and distinguished vertices $\rho_n$. In particular, $(V_n,R_n,\mu_n,\rho_n)\in \mathbb{F}_c$. Moreover, assume that there exist scaling factors $(a_n)_{n\geq 1}$, $(b_n)_{n\geq 1}$ such that $\sum_{n\geq 1}a_n^2<\infty$ and $(V_n,a_nR_n,b_n\mu_n,\rho_n)_{n\geq 1}$ satisfy Assumption \[a1\], where the limit space $(F,R,\mu,\rho)$ is in $\mathbb{F}_c$, and the measure $\mu$ is non-atomic. Finally, we suppose that each $T_n$ is equipped with random conductances $\omega^n=(\omega^n_e)_{e\in E_n}$ such that (\[omegatail\]) holds.]{}
We start by considering the resistance metrics on the weighted graph trees. In particular, given the conductances $\omega^n$, we define $R_n^\omega$ to be the associated resistance metric on $V_n$. In the following lemma, we show that, for large $n$, these random metrics are uniformly close to a scaled copy of $R_n$. The scaling factor is given by $\varrho:=\mathbf{E}\omega_e^{-1}$.
[\[resconv\] Suppose Assumption \[rcmtreeassu\] holds. It is then the case that, $\mathbf{P}$-a.s., $$\sup_{x,y\in V_n}a_n\left|R_n^\omega(x,y)-\varrho R_n(x,y)\right|\rightarrow 0.$$]{}
Suppose $(V_n,a_nR_n)$ and $(F,R)$ are embedded into the same space $(M,d_M)$ such that holds. Define $(x_i^n)_{i,n\geq1}$ and $(x_i)_{i\geq1}$ as in Section \[compsec\], so that $a_nR_n(x_i^n,x_j^n)\rightarrow R(x_i,x_j)$, for all $i,j\geq1$. Since $R_n^\omega(x_i^n,x_j^n)$ is the sum of $(\omega^n_e)^{-1}$ along the $R_n(x_i^n,x_j^n)$ edges in the path from $x_i^n$ to $x_j^n$, we obtain from (a fourth moment version of) the strong law of large numbers that, $\mathbf{P}$-a.s., ${R^\omega_n(x_i^n,x_j^n)}/{R_n(x_i^n,x_j^n)}\rightarrow \varrho$, for every $i,j\geq1$ (it is for this that the assumption $\sum_{n\geq 1}a_n^2<\infty$ is needed). In particular, the combination of the two previous observations implies that, $\mathbf{P}$-a.s., $$\sup_{i,j\leq k}a_n\left|R_n^\omega(x_i^n,x_j^n))-\varrho R_n(x_i^n,x_j^n))\right|\rightarrow 0,\qquad \forall k\geq1.$$ Since the resistances of unit edges satisfy $(\omega^n_e)^{-1}\leq 1$, we also have that $R_n^{\omega}\leq R_n$. It thus follows that, $\mathbf{P}$-a.s., $$\begin{aligned}
\lefteqn{\limsup_{n\rightarrow\infty}\sup_{x,y\in V_n}a_n\left|R_n^\omega(x,y)-\varrho R_n(x,y)\right|}\\
&\leq&
\limsup_{n\rightarrow\infty}\left\{\sup_{i,j\leq k}a_n\left|R_n^\omega(x_i^n,x_j^n)-\varrho R_n(x_i^n,x_j^n)\right|+2\sup_{x\in V_n}\inf_{i\leq k}a_n R_n(x,x_i^n)\right\}\\
&\leq&2\varepsilon_k,\end{aligned}$$ where $\varepsilon_k$ is defined as in Section \[compsec\]. In particular, since $\varepsilon_k\rightarrow0$ as $k\rightarrow\infty$, the result follows.
Similarly to Lemma \[nunconv2\], we next check convergence of the measures $\nu_n$, where we define $\nu_n(\{x\})=\sum_{e\in E_n:\:x\in e}\omega_e^n$ for $x\in V_n$. The limiting measure $\nu$ is the FIN measure on $F$, defined as in the previous section.
[\[nunconv3\] Suppose Assumption \[rcmtreeassu\] holds, and $(V_n,a_nR_n,b_n\mu_n,\rho_n)$, $n\geq 1$, and $(F,R,\mu,\rho)$ are isometrically embedded into a common (complete, separable, locally compact) metric space $(M,d_M)$ so that the conclusion of Lemma \[embeddings\] holds. It is then the case that $2^{-1}b_n^{1/\alpha}\nu_n\rightarrow \nu$ in distribution with respect to the vague topology for locally finite Borel measures on $(M,d_M)$.]{}
We first note that, if $\tilde{\mu}_{n}$ is a measure on $V_n$ defined by setting $\tilde{\mu}_{n}(\{x\})={\rm deg}_n(x)$, i.e. the usual graph degree of $x$ in $T_n$, then it is an elementary exercise to check that $d_{T_n}^P(\tilde{\mu}_{n},2{\mu}_{n})\leq2$, where $d_{T_n}^P$ is the Prohorov metric for measures on $T_n$. In particular, it follows that $b_n\tilde{\mu}_{n}\rightarrow 2\mu$ weakly as measures on $M$.
We next show that, for all $x\in M$, $r>0$ such that $B_M(x,r)$ is a continuity set for $\mu$, $$\label{ballmeasconv}
2^{-1}b_n^{1/\alpha}\nu_n\left(B_M(x,r)\right)\rightarrow \nu\left(B_M(x,r)\right)$$ in distribution. Writing $B=B_M(x,r)$, we have that $$\nu_n\left(B\right)=2\sum_{e\in E_n: e\subseteq B}\omega_e+\sum_{e\in E_n:\: e\cap B\neq \emptyset,\:e\not\subseteq B}\omega_e.$$ If we denote by $E_n^1(B)$ and $E_n^2(B)$ the subsets over which the two sums are taken, respectively, then we claim that $$\label{enlimits}
b_n\left|E_n^1(B)\right|\rightarrow \mu(B),\qquad b_n\left|E_n^2(B)\right|\rightarrow 0.$$ Indeed, for the second limit, we note that the edges in $E_n^2(B)$ each connect to a distinct vertex in the annulus $\bar{B}_M(x,r+a_n)\backslash B$. It follows that, for any $\varepsilon>0$, $$\limsup_{n\rightarrow\infty}b_n\left|E_n^2(B)\right|\leq \limsup_{n\rightarrow\infty} b_n\mu_n\left(\bar{B}_M(x,r+a_n)\backslash B\right)\leq \mu\left(\bar{B}_M(x,r+\varepsilon)\backslash B\right).$$ Since $B$ is a continuity set for $\mu$, the right-hand side can be made arbitrarily small by adjusting $\varepsilon$ as appropriate, which confirms the desired result. Given this, the first limit at (\[enlimits\]) is a simple consequence of the identity $\tilde{\mu}_n(B)=2|E_n^1(B)|+|E_n^1(B)|$, and the conclusion of the first paragraph. Thus, exactly as for (\[laplaceconv\]), we have that $\mathbf{E}(e^{-\lambda 2^{-1} b_n^{1/\alpha}\nu_n(B)})\rightarrow e^{-\lambda^\alpha \Gamma(1-\alpha)\mu(B)}$, which establishes (\[ballmeasconv\]).
With the same techniques, it is straightforward to extend (\[ballmeasconv\]) to the result that $$\left(2^{-1}b_n^{1/\alpha}\nu_n\left(B_i\right)\right)_{i=1}^k\rightarrow \left(\nu\left(B_i\right)\right)_{i=1}^k$$ in distribution, where each set $B_i$ is a finite unions of balls that are continuity sets for $\mu$. In particular, since the collection of such sets forms a separating class (see [@Kall p. 317]), this implies the result (see [@Kall Theorem 16.16 and Exercise 16.11]).
From Lemmas \[resconv\] and \[nunconv3\], we are able to prove the main result of this section. We write $P^{{\rm VSRW}_n}_x$ for the quenched law of the VSRW $X^{n,\omega}$ on the tree $T_n$ with conductances $\omega^n$, started from $x$. We write $\mathbb{P}^{{\rm CSRW}_n}_{x}$ for the annealed law of the corresponding CSRW $X^{n,\omega,\nu}$. We write $P_x$ for the law of the Brownian motion on $(F,R,\mu)$ started from $x$, and $\mathbb{P}^{{\rm FIN}}_{\rho}$ is the annealed law of the associated $\alpha$-FIN diffusion, defined as in Section \[bouchsec\].
[\[rcmtreeresult\] Suppose Assumption \[rcmtreeassu\] holds. It is then possible to isometrically embed $(V_n,a_nR_n,b_n\mu_n,\rho_n)$, $n\geq 1$, and $(F,R,\mu,\rho)$ into a common metric space $(M,d_M)$ so that, $\mathbf{P}$-a.s., $$\label{vsrwresult}
P^{{\rm VSRW}_n}_{\rho_n}\left(\left(X^{n,\omega}_{\varrho t/ a_nb_n}\right)_{t\geq 0}\in\cdot\right)\rightarrow P_\rho\left(\left(X_{t}\right)_{t\geq 0}\in\cdot\right)$$ weakly as probability measures on $D(\mathbb{R}_+,M)$. Moreover, $$\mathbb{P}^{{\rm CSRW}_n}_{\rho_n}\left(\left(X^{n,\omega,\nu}_{2\varrho t/ a_nb_n^{1/\alpha}}\right)_{t\geq 0}\in\cdot\right)
\rightarrow\mathbb{P}^{{\rm FIN}}_{\rho}\left(\left(X^{\nu}_{t}\right)_{t\geq 0}\in\cdot\right)$$ weakly as probability measures on $D(\mathbb{R}_+,M)$.]{}
The proof is essentially the same as that of Theorem \[main1\] and Corollary \[maincor\], but with care needed as the random metric $R_n^\omega$ is different to the metric $R_n$ used for the embedding. (We suppose throughout that the embeddings of $(V_n,a_nR_n,b_n\mu_n,\rho_n)$, $n\geq 1$, and $(F,R,\mu,\rho)$ into $(M,d_M)$ satisfy the conclusion of Lemma \[embeddings\].)
We first note that, since $R_n^\omega\leq R_n$ and Lemma \[resconv\] holds, we have from the UVD assumption for the underlying space that, $\mathbf{P}$-a.s., $$b_n\mu_n\left(B_{n}^\omega(x,r)\right)\geq c_1v(r),\qquad\forall x\in V_n,\:r\in[r^\omega_0(n),r^\omega_\infty(n)+1],$$ and, for every $\varepsilon>0$, $$b_n\mu_n\left(B_{n}^\omega(x,r)\right)\leq c_2v(r),\qquad\forall x\in V_n,\:r\in[\max\{r^\omega_0(n),a_n^{-1}\varepsilon\},r^\omega_\infty(n)+1],$$ where distances are defined with respect to the metric $R_n^\omega$ (note the truncation at $a_n^{-1}\varepsilon$ in the upper bound). These bounds are enough to repeat the proof of Lemma \[ltcont\] (cf. the weaker version of UVD in [@CroyLT]) to deduce the equicontinuity of the rescaled local times $(a_nL^{n,\omega}_{\varrho t/a_nb_n}(x))_{x\in V_n}$ of the VSRW $X^{n,\omega}$ with respect to the distance $a_nR_n^\omega$, and, by Lemma \[resconv\] again, the equicontinuity of these local times with respect to $a_nR_n$. We also claim that the above volume bounds yield that in place of (\[etest\]) we have, for $\delta,\tilde{\varepsilon}>0$, $$\label{exptailest}
\limsup_{n\rightarrow\infty}\sup_{x\in V_n}P^{{\rm VSRW}_n}_{x}
\left(\sup_{0\leq t\leq \delta}a_nR_n\left(x,X^{n,\omega}_{\varrho t/ a_nb_n}\right)>\tilde{\varepsilon}\right)\leq c_1 e^{-\frac{c_2 \tilde{\varepsilon}}{v^{-1}(\delta/\tilde{\varepsilon})}},\qquad \mathbf{P}\mbox{-a.s.}$$ Checking this requires only a minor adaptation of results from [@Kum]. Indeed, writing $\tau^{n,\omega}(x,r):=\inf\{t>0:\:a_nR_n(x,X^{n,\omega}_{\varrho t/ a_nb_n})> r\}$ and $h(r) = rv(r)$, the proof of [@Kum Proposition 4.2] gives the existence of constants $c_1, c_2$ such that $$E^{{\rm VSRW}_n}_{x}\left(\tau^{n,\omega}(y,r)\right)\leq c_1h(r), \qquad E^{{\rm VSRW}_n}_{x}\left(\tau^{n,\omega}(x,r)\right)\geq c_2h(r),$$ for all $x,y\in V_n$, $r\in [\max\{a_nr^\omega_0(n),\varepsilon\},a_n(r^\omega_\infty(n)+1)]$ and $n\geq 1$, and from this it readily follows that $$P^{{\rm VSRW}_n}_{x}\left(\tau^{n,\omega}(x,r)\leq s\right)\leq 1-c_3+\frac{c_4s}{h(r)}$$ for every $x\in V_n$, $r\in [\max\{a_nr^\omega_0(n),\varepsilon\},a_n(r^\omega_\infty(n)+1)]$, $s\geq 0$ and $n\geq 1$, cf. proof of [@Kum Lemma 4.2]. To obtain the exponential estimate of , we then follow the chaining argument of [@Kum Lemma 4.2]. This requires us to apply the previous exit time tail estimate for radii no smaller than $c_5v^{-1}(\tilde{\varepsilon}/\delta)$ (with respect to the metric $a_nR_n$). Noting that $a_nr^\omega_0(n)\rightarrow 0$, $\mathbf{P}$-a.s., one can thus adjust $\varepsilon$ so that the relevant estimates hold for large $n$.
Applying the conclusions of the previous paragraph, the proof of Proposition \[compactcase\] can be followed exactly to yield the result at (\[vsrwresult\]). Moreover, since we have local time equicontinuity and the distributional convergence of time-change measures given by Lemma \[nunconv3\], we also obtain the convergence of local times as at (\[ltconv1\]), and the convergence of the CSRW $X^{n,\omega,\nu}$ under the annealed measure (cf. the proof of Proposition \[lbmconv\] again).
[\[exa6-5\] [As a first application of Proposition \[rcmtreeresult\], one might consider the random conductance model on the Vicsek set example of Example \[lbmexamples\](ii). For this, we obtain the quenched convergence of the VSRW, $$\left(X^{n,\omega}_{\varrho 15^nt}\right)_{t\geq 0}\rightarrow \left(X_t\right)_{t\geq0},$$ where $X$ is the Brownian motion on the Vicsek set, and also the annealed convergence of the CSRW, $$\left(X^{n,\omega,\nu}_{2\varrho5^{n/\alpha}3^n t}\right)_{t\geq 0}\rightarrow\left(X^{\nu}_t\right)_{t\geq 0},$$ where $X^{\nu}$ is the $\alpha$-FIN diffusion on the Vicsek set.]{}]{}
Random conductance model on self-similar fractals {#rcmfractalsec}
-------------------------------------------------
In this section, we study the random conductance model on a class of self-similar fractals, extending the homogenisation results of [@Kum2; @kk] greatly. After introducing the model in Section \[ufrsec\], we then go on to study the renormalisation and homogenisation of associated discrete Dirichlet forms in Sections \[renormsec\] and \[homogsec\], respectively, and derive our main scaling results in Section \[concsec\].
### Uniform finitely ramified graphs {#ufrsec}
For $\beta>1$ and $I=\{1,2,\cdots, N\}$, let $(\Psi_i)_{i\in I}$ be a family of contraction maps on ${{\mathbb R}}^d$ such that $\Psi_i{\bf x}=\beta^{-1}U_i{\bf x} +\gamma_i,~{\bf x}\in {{\mathbb R}}^d$, where $U_i$ is a unitary map and $\gamma_i\in {{\mathbb R}}^d$. Assume that $(\Psi_i)_{i\in I}$ satisfies the open set condition, i.e., there is a non-empty, bounded open set $W$ such that $(\Psi_i (W))_{i\in I}$ are disjoint and $\cup_{i\in I} \Psi_i (W)\subset W$. As $(\Psi_i)_{i\in I}$ is a family of contraction maps, there exists a unique non-void compact set $F$ such that ${F}
=\cup_{i\in I}\Psi_i ({F})$. We assume $F$ is connected.
Let $Fix$ be the set of fixed points of the maps $\Psi_{i}$, $i\in I$. A point $x \in Fix$ is called an [*essential fixed point*]{} if there exist $i,j \in I,~i \ne j$ and $y\in Fix$ such that $\Psi_{i}(x)=\Psi_{j}(y)$. Let $I_{Fix}:=\{i\in I:$ the fixed point of $\Psi_i$ is an essential fixed point$\}$. We write ${V}_0$ for the set of essential fixed points. Denote $\Psi_{i_{1},\dots,i_{n}}=\Psi_{i_{1}}\circ\dots\circ\Psi_{i_{n}}$. We further assume a finitely ramified property, i.e., if $\{i_{1},\dots,i_{n}\}, \{j_{1},\dots,j_{n}\}$ are distinct sequences, then $$\Psi_{i_{1},\dots,i_{n}}({ F}) \bigcap\Psi_{j_{1},\dots,j_{n}}
({F})=\Psi_{i_{1},\dots,i_{n}}({V}_0)\bigcap\Psi_{j_{1},
\dots,j_{n}}({ V}_0);$$ note that, for each $n\ge 0$ and $i_1,\cdots,i_{n}\in I$, we call a set of the form $\Psi_{i_1,\cdots, i_n}({ V}_0)$ an $n$-cell. A compact *uniform finitely ramified (u.f.r.) fractal* ${F}$ is a set determined by $(\Psi_i)_{i\in I}$ satisfying the above assumptions with $|{V}_0|\ge 2$. Throughout, we assume without loss of generality that $\Psi_1({\bf x})=\beta^{-1}{\bf x}$ and ${\bf 0}$ belongs to ${V}_0$. We observe that u.f.r. fractals, first introduced in [@HK], form a class of fractals which is wider than nested fractals ([@lind]), and is included in the class of p.c.f. self-similar sets ([@kig1]). In particular, the Sierpiński gasket is an example of a u.f.r. fractal.
We next introduce the sequence of u.f.r. graphs approximating $F$. In particular, let $${V}_n=\cup_{i_1,\cdots,i_n\in I}\Psi_{i_{1},\dots,i_{n}}({V}_0),$$ noting that ${F}$ is the closure of $\cup_{n=0}^\infty{V}_n$. Moreover, denote by $E_n$ the collection of pairs of distinct points $x,y\in V_n$ such that $x$ and $y$ are in the same $n$-cell, and let $\mu_n$ be the counting measure on $V_n$ (placing mass one on each vertex). We will be interested in the scaling behaviour of $(V_n,R_n^\omega,\mu_n,\rho_n)$ (where $\rho_n$ is some distinguished vertex) and the associated VSRW and CSRW when $R_n^\omega$ is the resistance metric determined by placing random conductances along edges in $E_n$; in this section we generalise slightly from the i.i.d. conductance assumption to allow dependencies within the same $n$-cell.
For some of our results, it will be convenient to work in terms of the unbounded u.f.r. fractal and graphs; we define these now. We call ${\hat F}:=\cup_{n=1}^{\infty} \beta^n {F}$ an unbounded uniform finitely ramified fractal. Let ${\hat V}={\hat V}_0=\cup_{n=0}^{\infty}\beta^{n} {V}_n$, and ${\hat V}_{n}=\beta^{-n} {\hat V}$ for $n \in {\mathbb Z}$. We define $n$-cells for $n\in {\mathbb Z}$ as in the compact case, and denote by ${\hat E}_n$ the edges of the unbounded graph, connecting vertices within the same $n$-cell.
Finally for this section, we introduce some useful index spaces. In particular, let $$\begin{aligned}
\Xi &=&\{\eta \in I^{{\mathbb Z}}: \mbox { there exists $n\in {\mathbb Z}$ such that }
\eta_k=1, k \ge n\},\\
\Xi_+ &=&\{\eta \in I^{{\mathbb N}}: \mbox { there exists $n\in {\mathbb N}$ such that }
\eta_k=1, k \ge n\}.\end{aligned}$$ There is then a continuous map $\pi: \Xi\to {\mathbb R}^D$ such that $$\pi(\eta)= \lim_{n\to \infty} \beta^n
\Psi_{\eta_n}(\Psi_{\eta_{n-1}}(\cdots(\Psi_{\eta_{-n}}
({\bf 0}))\cdots)).$$ It is easy to see ${\hat F}=\pi (\Xi)$. For any $\eta\in \Xi
_+$ and $i\in I_{Fix}$, define $[\eta,i]\in \Xi$ as follows; $$[\eta,i](k)=\left\{\begin{array}{rl}\eta_k,&~~k\ge 1\\i,&~~k\le
0.\end{array}\right.$$ Then, ${\hat V}=\{\pi([\eta,i]): \eta \in \Xi_+, i\in I_{Fix}\}.$
### Renormalisation of forms {#renormsec}
In this section, we introduce notation and basic properties for Dirichlet forms and associated renormalisation maps on uniform finitely ramified graphs. To begin with, let ${\mbox{$\mathcal Q$}}$ be the set of $Q=(Q_{ij})_{i,j\in I_{Fix}}$ such that $$Q_{ij}=Q_{ji},~~\forall i,j\in I_{Fix},\qquad\sum_{j\in I_{Fix}}
Q_{ij}=0, ~~\forall i\in I_{Fix}.$$ Observe that ${\mbox{$\mathcal Q$}}$ is a vector space, with an inner product $(\cdot,\cdot)_{{\mbox{$\mathcal Q$}}}$ given by $$(Q,Q')_{{\mbox{$\mathcal Q$}}}=\sum_{j,k\in I_{Fix}} Q_{jk}Q'_{jk}=\mbox {Trace}~ Q{}^tQ',\qquad
Q,Q'\in {\mbox{$\mathcal Q$}}.$$ Let ${\mbox{$\mathcal Q$}}_+=\{Q\in {\mbox{$\mathcal Q$}}: {S}_Q(\xi,\xi)\ge 0$ for any $\xi \in l(I_{Fix})\}$, where $${S}_Q(\xi,\xi)=-\sum_{i,j\in I_{Fix}}Q_{ij}\xi_i \xi_j=\frac 12 \sum_{i,j\in I_{Fix}}Q_{ij}(\xi_i-\xi_j)^2,$$ and we define $l(A)=\{f:A\to {\mathbb R}\}$ for a set $A$. Set $$\|Q\|^2=\sup_{\xi\in l(I_{Fix})}\frac{{S}_Q(\xi,\xi)}{\sum_{i\in I_{Fix}}\xi_i^2}.$$ Note that $c_{1}\|Q\|^2\le (Q,Q)_{{\mbox{$\mathcal Q$}}}\le c_{2}\|Q\|^2$ for all $Q\in {\mbox{$\mathcal Q$}}_+$. Let $$\begin{aligned}
{\mbox{$\mathcal Q$}}_M&:=&\{Q\in {\mbox{$\mathcal Q$}}: Q_{ij}\ge 0, \forall i,j\in I_{Fix}, i \ne j\},\\
\mbox{Int} ({\mbox{$\mathcal Q$}}_M)&:=&\{Q\in {\mbox{$\mathcal Q$}}: Q_{ij}> 0, \forall i,j\in I_{Fix}, i \ne j\},\\
{\mbox{$\mathcal Q$}}_{irr}&:=&\{Q\in {\mbox{$\mathcal Q$}}_M: {S}_Q(\xi,\xi)=0 \Leftrightarrow \xi \mbox{ is
constant}\}.\end{aligned}$$ Note that $\mbox{Int} ({\mbox{$\mathcal Q$}}_M)\subset {\mbox{$\mathcal Q$}}_{irr} \subset {\mbox{$\mathcal Q$}}_M \subset {\mbox{$\mathcal Q$}}_+$. Take $Q_*\in\mbox{Int} ({\mbox{$\mathcal Q$}}_M)$, and let $$\begin{aligned}
{\mbox{$\Theta$}}_+:=C(\Xi_+,{\mbox{$\mathcal Q$}}_+),~~
{\mbox{$\Theta$}}_M:=C(\Xi_+,{\mbox{$\mathcal Q$}}_M),~~{\mbox{$\Theta$}}_{irr}:= C(\Xi_+,{\mbox{$\mathcal Q$}}_{irr}).\end{aligned}$$ Then ${\mbox{$\Theta$}}_+$ and ${\mbox{$\Theta$}}_M$ are convex cones. For any $\theta\in {\mbox{$\Theta$}}_+$, define ${\hat S}_\theta$ by $${\hat S}_\theta(u,u)=\frac 12\sum_{\eta\in\Xi_+}{S}_{\theta(\eta)}(u(\pi([\eta,\cdot])), u(\pi([\eta,\cdot]))),~~u\in {{\mathbb L}^2 ({\hat V},{\hat \mu}_0)},$$ where ${\hat \mu}_0$ is the counting measure on ${\hat V}$. If $\theta\in {\mbox{$\Theta$}}_M$, then ${\hat S}_\theta$ is a Dirichlet form on ${{\mathbb L}^2 ({\hat V},{\hat \mu}_0)}$. So, there is an associated Markov process $((X^\theta_t)_{t\geq 0}, (P^\theta_x)_{x\in \hat{V}})$. We introduce an order relation $\le$ in ${\mbox{$\Theta$}}_+$ as follows: $$\theta\le {\theta'} ~\mbox { if }~ {\hat S}_\theta(u,u) \le {\hat S}_{\theta'} (u,u) ~\mbox { for all }~ u \in {{\mathbb L}^2 ({\hat V},{\hat \mu}_0)}.$$ The norm on ${\mbox{$\Theta$}}_+$ is given by $\|\theta\|^2=\sup_{u\in {{\mathbb L}^2 ({\hat V},{\hat \mu}_0)}}{\hat S}_\theta(u,u)/\|u\|_{{{\mathbb L}^2 ({\hat V},{\hat \mu}_0)}}^2$.
We now define the renormalisation map $\bar{\Phi}$. For any $\theta\in {\mbox{$\Theta$}}_+$, let ${\hat S}_\theta^{(1)}:{{\mathbb L}^2 ({\hat V},{\hat \mu}_0)}\to [0,\infty)$ be given by $${\hat S}_\theta^{(1)}(u)=\inf \{{\hat S}_\theta(v,v):v\in {{\mathbb L}^2 ({\hat V},{\hat \mu}_0)},~
v(\beta x)=u(x),~x\in {\hat V}\}.$$ By the self-similarity of ${\hat F}$, there is a renormalisation map ${\bar \Phi}:{\mbox{$\Theta$}}_+ \to {\mbox{$\Theta$}}_+$ defined by setting ${\hat S}_\theta^{(1)}(u)={\hat S}_{{\bar \Phi}(\theta)} (u,u)$ for all $\theta\in {\mbox{$\Theta$}}_+$ and $u \in {{\mathbb L}^2 ({\hat V},{\hat \mu}_0)}$. Let $\iota: {\mbox{$\mathcal Q$}}_+\to {\mbox{$\Theta$}}_+$ be such that $\iota (Q)(\eta)=Q$ for all $\eta\in\Xi_+$ and $Q\in {\mbox{$\mathcal Q$}}_+$. Define a renormalisation map ${\tilde \Phi}:{\mbox{$\mathcal Q$}}_+\to{\mbox{$\mathcal Q$}}_+$ as ${\tilde \Phi}(Q)={\bar \Phi}(\iota(Q))(\eta)$ for $\eta\in \Xi_+$. Note that it is independent of the choice of $\eta\in\Xi_+$. By Schauder’s fixed point theorem, we know that there exists $Q_*\in {\mbox{$\mathcal Q$}}_M$ (with $(Q_*)_{ij}>0$ for some $i\ne j$) and $\varrho_{Q_*}>0$ such that ${\tilde \Phi}(Q_*)=\varrho_{Q_*}^{-1} Q_*$. Henceforth, we assume the following.
\[thm:ass1\] (1) For each $Q\in {\mbox{$\mathcal Q$}}_{irr}$, there exists $n_0=n_0(Q)\in {\mathbb N}$ such that ${\tilde \Phi}^n(Q)\in \mbox{Int} ({\mbox{$\mathcal Q$}}_M)$ for all $n\ge n_0$.\
(2) There exists $Q_0\in \mbox{Int} ({\mbox{$\mathcal Q$}}_M)$ and $\varrho_{Q_0}>0$ such that ${\tilde \Phi}(Q_0)=\varrho_{Q_0}^{-1} Q_0$.
In the following, we take one $Q_0\in \mbox{Int} ({\mbox{$\mathcal Q$}}_M)$, as given by Assumption \[thm:ass1\](2), and fix it.
[(1) It is known that $\varrho_{Q_0}>0$ is uniquely determined, i.e. if $Q_1,Q_2\in {\mbox{$\mathcal Q$}}_{irr}$ satisfy ${\tilde \Phi}(Q_j)=\varrho_{Q_j}^{-1} Q_j$ ($j=1,2$) with $\varrho_{Q_1}, \varrho_{Q_2}>0$, then $\varrho_{Q_1}=\varrho_{Q_2}=\varrho_{Q_0}$. In the class of fractal graphs we consider, we can prove $\varrho_{Q_0}>1$ (see [@kig1], for example).\
(2) Every nested fractal satisfies Assumption \[thm:ass1\].]{}
Under Assumption \[thm:ass1\], we set $\Phi=\varrho_{Q_0} {\bar \Phi}: {\mbox{$\Theta$}}_+\to{\mbox{$\Theta$}}_+$, $\hat\Phi=\varrho_{Q_0} {\tilde \Phi}$ and ${\hat S}_\theta^{\Phi}(u)=\varrho_{Q_0}{\hat S}_\theta^{\bar \Phi}(u)$ for $u \in {{\mathbb L}^2 ({\hat V},{\hat \mu}_0)}$.
### Homogenisation of forms {#homogsec}
In this subsection, we will describe the homogenisation of the discrete Dirichlet forms associated with the random conductance model on u.f.r. fractals, see Theorem \[thm:conth\] for the main result. First, we give some further definitions for later use. Let ${V}_0=\{a_i: i\in I_{Fix}\}$. For $Q^*\in \mbox{Int} ({\mbox{$\mathcal Q$}}_M)$, $k\in I$, we define a matrix $A_{k,Q^*}\in l(I_{Fix}^2)$ by setting $$(A_{k,Q^*})_{ij}=P^{Q^*}_{\Psi_k(a_i)}\big(X^1_{\tau_{{ V}_0}}=a_j\big),$$ where $X^1$ is a discrete time Markov chain on ${V}_1$ whose transition probabilities are determined by the Dirichlet form obtained by placing a copy of $Q^*$ on each $1$-cell, and $\tau_{{V}_0}=\inf\{n\ge 0: X^1_n\in {V}_0\}$. Then, it is easy to see that the following holds for u.f.r. graphs; $0<(A_{k,Q^*})_{ij}<1$ if $k\ne i$ and $(A_{k,Q^*})_{kj}=\delta_{kj}$.
We now define a liberalisation of the renormalisation map around the fixed point $Q_*$. For any $\theta\in {\mbox{$\Theta$}}_+$ and $Q_*\in \mbox{Int} ({\mbox{$\mathcal Q$}}_M)$ with ${\hat \Phi}(Q_*)=Q_*$, define $${\hat S}_\theta^{(2)}(u)=\varrho_{Q_0}{\hat S}_\theta(v,v)\qquad\mbox { for }~~~u \in {{\mathbb L}^2 ({\hat V},{\hat \mu}_0)},$$ where $v\in {{\mathbb L}^2 ({\hat V},{\hat \mu}_0)}$ satisfies $v(\beta x)=u(x),~x\in {\hat V}$, and $v$ is $Q_*$-harmonic on ${\hat V}\setminus \beta {\hat V}$, i.e., $$v(\pi([\eta\cdot i,j]))=
\sum_{k\in I_{Fix}} (A_{i,Q_*})_{jk}u(\pi([\eta,k]))\qquad\mbox { for }
~~~i\in I,j\in I_{Fix}.$$ Here $\eta\cdot i\in \Xi_+$ is given by $(\eta\cdot i)_n=\eta_{n-1},~n\ge2$ and $(\eta\cdot i)_1=i$. It is easy to see that ${\hat S}_\theta^{(2)}(u)={\hat S}_{{H_{Q_*}}(\theta)}(u,u)$ for all $\theta\in {\mbox{$\Theta$}}_+$ and $u \in {{\mathbb L}^2 ({\hat V},{\hat \mu}_0)}$, where $${H_{Q_*}}(\theta)(\eta)=\varrho_{Q_0}\sum_{k\in I}
{}^tA_{k,Q_*} \theta(\eta\cdot k)A_{k,Q_*}.$$ Similarly, we define a linear map $\hat H_{Q_*}: {\mbox{$\mathcal Q$}}_+\to {\mbox{$\mathcal Q$}}_+$ by ${\hat H_{Q_*}}(Q)=\varrho_{Q_0}\sum_{k\in I}
{}^tA_{k,Q_*} QA_{k,Q_*}.$ Note that $\hat H_{Q_*}(Q_*)=Q_*$. The following properties of $\Phi$ and $H_{Q_*}$ are easy, but important. Note that the corresponding results hold for ${{\hat \Phi}}$.
Let $Q_*\in \mbox{Int} ({\mbox{$\mathcal Q$}}_M)$ satisfy ${\hat \Phi}(Q_*)=Q_*$ and $\theta,{\theta'}\in {\mbox{$\Theta$}}_+$.\
(1) If $\theta\le {\theta'}$, then $\Phi(\theta)\le \Phi({\theta'})$, $H_{Q_*}(\theta)\le
H_{Q_*}({\theta'})$ and $\Phi(\theta)\le H_{Q_*}(\theta)$.\
(2) For $a,b\ge 0$, $\Phi(a\theta+b{\theta'})\ge a\Phi(\theta)+b\Phi({\theta'})$ and $H_{Q_*}(a\theta+b{\theta'})= aH_{Q_*}(\theta)+bH_{Q_*}({\theta'})$.
We are now ready to introduce a probability measure $\mathbf{P}$ on $\Theta_M$ to describe our random conductance model in this setting. In particular, we now write $\theta$ for a $\Theta_M$-valued random variable, and suppose that, under $\mathbf{P}$, the elements $(\theta(\eta))_{\eta\in \Xi_+}$ are independently identically distributed ${\mbox{$\mathcal Q$}}_M$-valued random variables such that $C_1 Q_0\le \theta(\eta)$ for $\eta\in \Xi_+$. Note that in [@Kum2; @kk] it was assumed that $\mathbf{P}(\{\theta\in {\mbox{$\Theta$}}_M: C_1 Q_0\le \theta(\eta) \le C_2 Q_0,$ for $\eta\in \Xi_+\})=1$ for some $C_1,C_2>0$. Here we do not assume such a uniform ellipticity condition from above. We note the following further property of $\Phi$: $$\label{basic10}
\mathbf{E}(\Phi(\theta))\le \Phi(\mathbf{E}(\theta)),$$ where the expectation is taken for each element of the matrix in ${\mbox{$\mathcal Q$}}_M$.
Let $\Phi^n$ be the $n$-th iteration of $\Phi$. We make the following further assumption, which is possible to verify in the case of nested fractals when the distribution of the individual conductances does not have too heavy a tail at infinity.
\[thm:ass2\] There exists $n_0\in {\mathbb N}$ such that $$\mathbf{E}\left[(\Phi^{n_0}(\theta)(\eta)_{ij})^2\right]<\infty,\qquad \forall i, j\in I_{Fix},\:\eta\in\Xi_+.$$
Note that under Assumption $\ref{thm:ass2}$ we have, for all $i\ne j\in I_{Fix}$, $\eta\in\Xi_+$, $$\mathbf{E}[\Phi^{n_0}(\theta)(\eta)_{ij}]
\le (\mathbf{E}[(\Phi^{n_0}(\theta)(\eta)_{ij})^2])^{1/2}<\infty,$$ so by (\[basic10\]), $\mathbf{E}[\Phi^{n}(\theta)(\eta)_{ij}]<\infty$ for all $i\ne j\in I_{Fix}$, $\eta\in\Xi_+$, $n\ge n_0$. We next give a sufficient condition for Assumption $\ref{thm:ass2}$ to hold. For $x,y\in V_n$, define $$h_n(x,y)=\min\{k: K_1,\cdots K_k \mbox{ are $n$-cells}, x\in K_1,y\in K_k, K_i\cap K_{i+1}\ne \emptyset, \forall i=1,\dots,k-1\}.$$
Suppose that there exists $n_0\in {\mathbb N}$ such that $\min_{x,y\in V_0, x\ne y}h_{n_0}(x,y)\ge 2$. Suppose also that the law of $\theta(\eta)_{ij}$ has at most polynomial decay at infinity for all $i\ne j\in I_{Fix}$, $\eta\in\Xi_+$, namely there exists $c_1,\gamma_{ij}>0$ such that $\mathbf{P} (\theta(\eta)_{ij}\ge s)\le c_1s^{-\gamma_{ij}}$. Then Assumption \[thm:ass2\] holds. In particular, Assumption \[thm:ass2\] holds for nested fractal graphs if the law of the random conductances has at most polynomial decay at infinity.
First, suppose we have two edges with conductance $\omega_1$, $\omega_2$ such that $\mathbf{P}(\omega_i\ge s)\le c_is^{-\gamma_i}$ for $i=1,2$. If the edges are connected in parallel, then the effective conductance is $\omega_1+\omega_2$, which satisfies $$\label{eq:biebi11}
\mathbf{P}(\omega_1+\omega_2\ge s)\le \mathbf{P}(\omega_1\ge s/2)+\mathbf{P}(\omega_2\ge s/2)\le 2(c_1\vee c_2)s^{-\gamma_1\wedge \gamma_2},\qquad\forall s\ge 1.$$ Similarly, connect the two conductances in series, and assume that $\omega_1$ and $\omega_2$ are independent. Then the effective conductance is $(\omega_1^{-1}+\omega_2^{-1})^{-1}$, and we have $$\begin{aligned}
\mathbf{P}((\omega_1^{-1}+\omega_2^{-1})^{-1}\ge s)
&=&\mathbf{P}(\omega_1^{-1}+\omega_2^{-1}\le s^{-1})\nonumber\\
&\le &\mathbf{P}(\omega_1^{-1}\le s^{-1})\mathbf{P}(\omega_2^{-1}\le s^{-1})\nonumber\\
&\le &c_1c_2s^{-\gamma_1- \gamma_2},
\qquad\forall s\ge 1.\label{eq:biebi22}\end{aligned}$$ Next, note that by the assumption we have $\min_{x,y\in V_0, x\ne y}h_{ln_0}(x,y)\ge 2^l$ for all $l\ge 1$. Let $a_i\in V_0$ be the fixed point of $\Psi_i$. Consider the network on $\beta^{ln_0}V_{ln_0}$ and fix $a_i\ne a_j\in V_0$. Define $H_m=\{z\in \beta^{ln_0}V_{ln_0}: h_{ln_0}(a_i,\beta^{-ln_0}z)=m\}$ for $1\le m \le h_{ln_0}(a_i,a_j)-1$, and $H_{h_{ln_0}(a_i,a_j)}=\{z\in \beta^{ln_0}V_{ln_0}: h_{ln_0}(a_i,\beta^{-ln_0}z)\ge h_{ln_0}(a_i,a_j)\}$. Now short all the vertices that are in the same $H_m$ for $1\le m \le h_{ln_0}(a_i,a_j)$, and let $C_{ij}$ be the effective conductance between $a_i$ and $a_j$ for the induced network. By Rayleigh’s monotonicity principle for electric networks, we see that $\Phi^{ln_0}(\theta)(\eta)_{ij}\le C_{ij}$, where $\eta=(1,1,1,\dots)$. Applying and repeatedly, we see that $C_{ij}^2$ is integrable when $l$ is large enough. Therefore Assumption $\ref{thm:ass2}$ holds in this case. Finally, note that the condition $\min_{x,y\in \hat V_0, x\ne y}h_{n_0}(x,y)\ge 2$ holds for nested fractal graphs due to [@kus Lemma $(2.8)$], so the last assertion holds.
We are now ready to state the main result of this section.
\[thm:conth\] Under Assumptions $\ref{thm:ass1}$ and $\ref{thm:ass2}$, there exists $Q_{\mathbf{P}}
\in\mbox{Int}({\mbox{$\mathcal Q$}}_{M})$ such that, for all $\eta\in \Xi_+$, $$\label{eq:daiji}
Q_{\mathbf{P}}=\lim_{n\to\infty}\Phi^n(\theta)(\eta),\qquad
\mbox{in }{L}^1({\mbox{$\mathcal Q$}}_M,\mathbf{P}).$$
The rest of this subsection is devoted to proving Theorem \[thm:conth\]. The next proposition is a restricted version of the result by Peirone [@peir], whose original ideas come from Sabot [@sab]; see [@Kum2 Appendix $A$] for the proof.
\[thm:peir\] Under Assumption $\ref{thm:ass1}$, for each $M\in {\mbox{$\mathcal Q$}}_{irr}$, there exists $Q_M \in \mbox{Int}({\mbox{$\mathcal Q$}}_M)$ such that $Q_M=\lim_{n\to\infty} {{\hat \Phi}}^n(M)$.
The next lemma is an adaptation of [@kk Lemma 4.1], but the proof requires serious modification from the latter work to cover our more general setting. We denote by $H^n_{Q_*}$ the $n$-th iteration of $H_{Q_*}$.
\[thm:EHX\] Let $Q_*\in \mbox{Int}({\mbox{$\mathcal Q$}}_M)$ satisfy ${{\hat \Phi}}(Q_*)=Q_*$. Under Assumption \[thm:ass2\], there exist $c_{1}>0$ and $0<\varepsilon <1$ such that $$\label{eq:neow}
\mathbf{E}[\Vert H^n_{Q_*}(\Phi^{n_0}(\theta))(\eta)-H^n_{Q_*}(\mathbf{E}[\Phi^{n_0}(\theta)])(\eta)\Vert^2]\le c_{1} (1-\varepsilon)^n,~~~\forall\eta\in \Xi_+, n\ge 1.$$ In particular, it $\mathbf{P}$-a.s. holds that $$\lim_{n\to\infty}\Vert H^n_{Q_*}(\Phi^{n_0}(\theta))(\eta)-H^n_{Q_*}(\mathbf{E}[\Phi^{n_0}(\theta)])(\eta)\Vert= 0,\qquad\forall\eta\in \Xi_+.$$
Let the left hand side of $(\ref{eq:neow})$ be $f(n,\eta)$ and set $\theta'=\Phi^{n_0}(\theta)$. Further, let $$\begin{aligned}
\theta'^{(1)}_{i_1,\cdots,i_n}(\eta)&=&{}^tA_{i_n}\cdots{}^tA_{i_1}
\theta'(\eta\cdot i_1\cdots\cdot i_n)A_{i_1}\cdots A_{i_n},\\
\theta'^{(2)}_{i_1,\cdots,i_n}(\eta)&=&{}^tA_{i_n}\cdots{}^tA_{i_1}
\mathbf{E}[\theta'(\eta\cdot i_1\cdots\cdot i_n)]A_{i_1}\cdots A_{i_n},\\
\theta'_{i_1,\cdots,i_n}(\eta)&=&\theta'^{(1)}_{i_1,\cdots,i_n}(\eta)
-\theta'^{(2)}_{i_1,\cdots,i_n}(\eta),\end{aligned}$$ where we set $A_i:=A_{i,Q_*}$. Then we have $$\begin{aligned}
f(n,\eta) &\leq& c\varrho_{Q_0}^{2n}
\mathbf{E}\left[\mbox{Trace}\left[\left(\sum_{i_1,\cdots,i_n}\theta'_{i_1,\cdots,i_n}(\eta)\right)^2\right]\right]\\
&=&c
\varrho_{Q_0}^{2n}\sum_{i_1,\cdots,i_n}
\mathbf{E}\left[\mbox{Trace}\left[\left(\theta'_{i_1,\cdots,i_n}(\eta)\right)^2\right]\right]\\
&= &c\varrho_{Q_0}^{2n}\sum_{i_1,\cdots,i_n}
\left(\mathbf{E}\left[\mbox{Trace}\left[\left(\theta'^{(1)}_{i_1,\cdots,i_n}(\eta)\right)^2\right]\right]-\mbox{Trace}\left[\left(\theta'^{(2)}_{i_1,\cdots,i_n}(\eta)\right)^2\right]\right)\\
&\le& c\varrho_{Q_0}^{2n}
\sum_{i_1,\cdots,i_n}\mathbf{E}\left[\left(\mbox{Trace}\:\theta'^{(1)}_{i_1,\cdots,i_n}(\eta)\right)^2\right],\end{aligned}$$ where the first equality is because $\theta'_{i_1,\cdots,i_n}(\eta)$ and $\theta'_{j_1,\cdots,j_n}(\eta)$ are independent (because of the finitely ramified property) and mean zero for $({i_1,\cdots,i_n})\ne (j_1,\cdots,j_n)$, and the last inequality is because $\mbox{Trace } (B^2) \le (\mbox{Trace } B)^2$ for any non-negative definite symmetric matrix $B$.
Set $A=(a_{ij})=A_{i_1}\cdots A_{i_n}$, $(x_{ij})
=\theta'(\eta\cdot i_1\cdots\cdot i_n)$. Then we have $$\begin{aligned}
(\theta'^{(1)}_{i_1,\cdots,i_n}(\eta))_{ij}=\sum_{k,l}a_{ki}a_{lj}x_{kl}
=-\frac 12\sum_{k,l: k\ne l}(a_{ki}-a_{li})(a_{kj}-a_{lj})x_{kl},\end{aligned}$$ because $x_{ll}=-\sum_{k:k\ne l}x_{kl}$. Thus, denoting $Q_*=(q_*)_{ij}$, we have $$\begin{aligned}
\lefteqn{\mathbf{E}\left[\left(\mbox{Trace}\:\theta'^{(1)}_{i_1,\cdots,i_n}(\eta)\right)^2\right]}\\
&=&\mathbf{E}\left[\left(-\frac 12\sum_i\sum_{k,l: k\ne l}(a_{ki}-a_{li})^2x_{kl}\right)^2\right]\\
&=&\frac 14\sum_i\sum_{k,l: k\ne l}\sum_{i'}\sum_{k',l': k'\ne l'}
(a_{ki}-a_{li})^2(a_{k'i'}-a_{l'i'})^2\mathbf{E}[x_{kl}x_{k'l'}]\\
&\le&\frac 14\sum_i\sum_{k,l: k\ne l}\sum_{i'}\sum_{k',l': k'\ne l'}
(a_{ki}-a_{li})^2(a_{k'i'}-a_{l'i'})^2(\mathbf{E}[x_{kl}^2])^{1/2}(\mathbf{E}[x_{k'l'}^2])^{1/2}\\
&=&\frac 14\left(\sum_i\sum_{k,l: k\ne l}(a_{ki}-a_{li})^2(\mathbf{E}[x_{kl}^2])^{1/2}\right)^2\\
&\le& c_1\left(\sum_i\sum_{k,l: k\ne l}(a_{ki}-a_{li})^2
(q_*)_{kl}\right)^2,\end{aligned}$$ where the last inequality is because there exists $c_*>0$ such that $\mathbf{E}[(\Phi^{n_0}(\theta)(\eta)_{ij})^2]\le c_*$ for all $i, j\in I_{Fix}$, which is due to Assumption $\ref{thm:ass2}$. In particular, we obtain that $$\begin{aligned}
f(n,\eta) &\le&c_2\varrho_{Q_0}^{2n}\sum_{i_1,\cdots,i_n}
\{\mbox{Trace }{}^tA_{i_n}\cdots{}^tA_{i_1}
Q_* A_{i_1}\cdots A_{i_n}\}^2.\end{aligned}$$ Now, from the proof of [@kus Proposition (5.5)], we have $$\varrho_{Q_0}^{n}~ {}^tA_{i_n}\cdots{}^tA_{i_1}Q_* A_{i_1}\cdots A_{i_n} \le(1-\varepsilon)^n Q_*$$ for some $0 <\varepsilon<1$. (Note that in [@kus] it is assumed that $\sum_{k\ne i}{^tA_kA_k}$ is strictly positive for all $i$, but this assumption is satisfied in our setting; see [@kus Proposition $(7.2)$].) Combining this with $\hat H_{Q_*}(Q_*)=Q_*$, we obtain $$f(n,\eta) \le c'_* (1-\varepsilon)^n (\mbox{Trace }Q_*)^2\le c_1 (1-\varepsilon)^n.$$
Let $\phi_m=\mathbf{E}[\Phi^{m+n_0}(\theta)(\eta)]$ ($\phi_m$ is independent of $\eta$). By Assumption $\ref{thm:ass2}$ and Proposition \[thm:peir\], for each $m\in {\mathbb N}$, there exists $Q_m\in \mbox{Int} ({\mbox{$\mathcal Q$}}_M)$ such that $\lim_{n\to\infty} {{\hat \Phi}}^n(\phi_m)=Q_m$ and ${{\hat \Phi}}(Q_m)=Q_m$. On the other hand, similarly to (\[basic10\]) we see $$\label{eq:rnrm}
{{\hat \Phi}}^n (\phi_m)\ge \phi_{n+m}\qquad\forall m,n\in {\mathbb N}\cup\{0\},$$ so that $Q_m\ge Q_{n+m}$. Denote the limit of $(Q_m)_{m\geq 0}$ by $Q_+$; then ${{\hat \Phi}}(Q_+)=Q_+$. (Note that $Q_+\in \mbox{Int}({\mbox{$\mathcal Q$}}_M)$ due to Assumption \[thm:ass1\](1) and the assumption $\mathbf{P}(\{\theta\in {\mbox{$\Theta$}}_M: C_1 Q_0\le \theta(\eta)$ for $\eta\in \Xi_+\})=1$.) For any $\varepsilon>0$, there exists $N_{\varepsilon}\in {\mathbb N}$ such that $$\label{eq:12-add}
(1+\varepsilon)Q_+\ge \phi_m\qquad \forall m\ge N_{\varepsilon}.$$ Indeed, if this does not hold, then because there exists $C_*>0$ such that $(\phi_m)_{ij}\le C_*$ for all $i\ne j\in I_{Fix}$ and all $m\in {\mathbb N}$, there exists a subsequence $(l_j)_{j\geq 0}$ such that $\phi_{l_j}\ge (1+\varepsilon)Q_+$ and $\lim_{j\to\infty}\phi_{l_j}=:{\bar \phi}$ exists. On the other hand, by (\[eq:rnrm\]), we have ${{\hat \Phi}}^{l_{j'}-l_j}(\phi_{l_j})\ge \phi_{l_{j'}}$ for all $j'\ge j$ so that $Q_+\ge {\bar \phi}$, which is a contradiction. By the definition of $Q_m$, for each $m$ and $\varepsilon>0$, there exists $L_{m,\varepsilon}$ such that $(1-\varepsilon)Q_m\le {{\hat \Phi}}^n (\phi_m)$ for all $n\ge L_{m,\varepsilon}$. Combining these facts and noting ${{\hat H}}_{Q_+}^n(\phi_m)\ge {{\hat \Phi}}^n(\phi_m)$, we have $$\label{eq:12-1}
(1-\varepsilon)Q_+\le {{\hat H}}_{Q_+}^n (\phi_m)\le (1+\varepsilon)Q_+\qquad \forall n\ge L_{m,\varepsilon}, m\ge N_{\varepsilon}.$$ On the other hand, by Lemma \[thm:EHX\], we have $\mathbf{P}$-a.s. that $$\lim_{n\to\infty}\|H_{Q_+}^n(\Phi^{m+n_0}(\theta))(\eta)-{{\hat H}}_{Q_+}^n(\phi_m)\|=\lim_{n\to\infty}\|H_{Q_+}^n(\Phi^{m+n_0}(\theta))(\eta)- H_{Q_+}^n(\iota(\phi_m))(\eta)\|= 0$$ for all $\eta\in \Xi_+$, $m\geq0$. Since $H_{Q_+}^n(\Phi^{m+n_0}(\theta))(\eta)\ge \Phi^{n+m+n_0}(\theta)(\eta)$, we see that the following holds $\mathbf{P}$-a.s.: for some $N'_{\varepsilon,\eta}\in{\mathbb N}$, $$\label{eq:12-3}
(1+\varepsilon)Q_+\ge \Phi^{m+n_0}(\theta)(\eta),\qquad \forall \eta\in \Xi_+, m\ge N'_{\varepsilon,\eta}.$$ We now establish some more properties of ${{\hat H}}_{Q_+}$. It is easy to see $\sup_{n}|||{{\hat H}}^n_{Q_+}|||<\infty$, where $|||{{\hat H}}^n_{Q_+}|||:=\sup_{Q\in \mathcal{Q}_M, \|Q\|=1}\|{{\hat H}}^n_{Q_{+}}(Q)\|$, see [@kk Lemma 4.3]. Using this, we see that the size of each Jordan cell corresponding to the largest eigenvalue of ${{\hat H}}_{Q_+}$ is $1$. We thus obtain that there exists an orthogonal projection $P_0: {\mbox{$\mathcal Q$}}_M\to{\mbox{$\mathcal Q$}}_M$ so that for each $k\in {\mathbb N}$, there exists $n_k\in {\mathbb N}$ such that $$\label{eq:12-4}
|||{{\hat H}}^{n_k}_{Q_+}-P_0|||\le 2^{-k}.$$ By (\[eq:12-1\]) and (\[eq:12-4\]), we have $\phi_m\ge P_0\phi_m\ge (1-\varepsilon)Q_+$ for all $m\ge N_{\varepsilon}$. Together with (\[eq:12-add\]), we have $$\label{eq:12-5}
\lim_{n\to\infty}\phi_n=Q_+.$$ Now, by Fatou’s lemma and (\[eq:12-5\]), $$\label{eq:fatoo}
\mathbf{E}\big[\limsup_{n\to\infty}S_{\Phi^n(\theta)(\eta)}(u,u)\big]
\ge \limsup_{n\to\infty}S_{\phi_n}(u,u)= S_{Q_+}(u,u),$$ for all $\eta\in \Xi_+,~u\in l(V_{\eta})$, where $V_{\eta}:=\{\pi([\eta,i]): i\in I_{Fix}\}$ is a $0$-cell whose address is $\eta$. (Note that we can use Fatou’s lemma thanks to (\[eq:12-3\]).) By (\[eq:12-3\]) and (\[eq:fatoo\]), we have $$\limsup_{n\to\infty}S_{\Phi^n(\theta)(\eta)}(u,u)= S_{Q_+}(u,u),$$ $\mathbf{P}$-a.s. for all $\eta\in \Xi_+,~u\in l(V_{\eta})$. Applying [@kk Lemma 4.2] with $Y_n=S_{\Phi^n(\theta)(\eta)}(u,u)$ and $Y=S_{Q_+}(u,u)$ (note that $\sup_n \mathbf{E}[Y_n^2]<\infty$ due to Assumption \[thm:ass2\]), we have $$\lim_{n\to\infty}\mathbf{E}\big[|S_{\Phi^n(\theta)(\eta)}(u,u)
-S_{Q_+}(u,u)|\big]
= 0,\qquad\forall \eta\in \Xi_+,~u\in l(V_{\eta}).$$ Since $l(V_{\eta})$ is finite dimensional, we obtain (\[eq:daiji\]) where $Q_{\mathbf{P}}=Q_+$.
### Application to the random conductance model {#concsec}
We are now ready to explain the application of the homogenisation results of the previous section to the random conductance model; see Proposition \[ssfrcmresult\] for the main result. For the setting, we recall the graphs $(V_n,E_n)$, and the associated counting measure $\mu_n$, from Section \[ufrsec\]. We further suppose each graph is equipped with a collection of random conductances $(\omega^n_e)_{e\in E_n}$ such that the conductances within each $n$-cell, $(\omega^n_e)_{e\subseteq \Psi_{i_1,\dots,i_n}(V_0)}$, are independent, and identically distributed as $(\omega^0_e)_{e\in E_0}$ (and built on a probability space with probability measure $\mathbf{P}$). The associated random resistance metric will be denoted by $R_n^\omega$.
Note that this family of random graphs can be coupled with the framework of the previous section. In particular, suppose that $(\theta(\eta)_{ij})_{i,j=1}^{I_F}$ is distributed as $(\omega^0_{a_i,a_j})_{i,j=1}^{I_F}$, independently for each $\eta$. Then we easily see that the random weighted graph $(V_n,E_n,\omega^n)$ is identical in distribution to that given by the conductances associated with $\theta$ on $\beta^nV_n\subseteq\hat{V}$. We will fix this identification throughout the section, and typically suppose that Assumptions \[thm:ass1\] and \[thm:ass2\] are satisfied accordingly. This means that we can define the $Q_\mathbf{P}$ for which the conclusion of Theorem \[thm:conth\] holds.
We next describe the limiting object. First, let $R_n$ be the resistance metric on $V_n$ induced by placing conductances according to $Q_{\mathbf{P}}$ along edges of $n$-cells, i.e. setting the conductance from $\Psi_{i_1,\dots,i_n}(a_i)$ to $\Psi_{i_1,\dots,i_n}(a_j)$ to be $(Q_\mathbf{P})_{ij}$. From the fact that $\hat{\Phi}(Q_{\mathbf{P}})=Q_{\mathbf{P}}$, it follows that there exists a resistance metric $R$ on $V_*:=\cup_{n\geq 0}V_n$ defined by setting $R:=\varrho_{Q_0}^{-n}R_n$ on $V_n$, where $\varrho_{Q_0}>1$ is the scaling factor given by Assumption \[thm:ass1\]. Moreover, by [@kig1 Theorems 2.3.10 and 3.3.4], taking the completion of the metric space $(V_*,R)$ yields a resistance metric $R$ on the u.f.r. fractal $F$, which is topologically equivalent to the Euclidean metric. It is moreover an elementary exercise to check that $N^{-n}\mu_n\rightarrow \mu$, where $\mu$ is the (unique up to a constant multiple) self-similar measure on $F$, placing equal weight on each $1$-cell; this measure is non-atomic and has full-support. We observe that, for any $\rho_n\in V_n$ such that $\rho_n\rightarrow \rho$ (with respect to $R$, or equivalently the Euclidean metric), we have that $(V_n,a_nR_n,b_n\mu_n,\rho_n)\rightarrow(F,R,\mu,\rho)$ in $\mathbb{F}_c$ with respect to the Gromov-Hausdorff-vague topology for $a_n=\varrho_{Q_0}^{-n}$, $b_n=N^{-n}$. We moreover note that $(V_n,a_nR_n,b_n\mu_n,\rho_n)_{n\geq 1}$ satisfies UVD (see [@HK Lemma 3.2]).
As in Section \[rcmtreesec\], to get from the convergence of the previous paragraph to the convergence of the VSRW associated with the random conductances $(\omega^n_e)_{e\in E_n}$, we need to establish the convergence of the random metric $R_n^\omega$. This is the aim of the following lemma.
[\[ssfresconv\] Suppose Assumptions \[thm:ass1\] and \[thm:ass2\] hold, and that the conductances $(\omega_e^0)_{e\in E_0}$ are uniformly bounded from below (i.e. there exists a constant $c>0$ such that $\omega_e^0\geq c$, $\mathbf{P}$-a.s.). Then it is the case that in $\mathbf{P}$-probability $$\sup_{x,y\in V_n}a_n\left|R_n^\omega(x,y)-R_n(x,y)\right|\rightarrow 0.$$]{}
Translating Theorem \[thm:conth\] into the present notation, and noting that, for a finite network, convergence of edge conductances implies convergence of the resistance metric (cf. the proof of Lemma \[l2\]), we obtain for any $x,y\in V_*$ that, in $\mathbf{P}$-probability, $a_nR_n^\omega(x,y)\rightarrow R(x,y)$. Moreover, the fact that conductances are uniformly bounded below implies that there exists a constant $c_1$ such that $R_n^\omega\leq c_1R_n$, $\mathbf{P}$-a.s. From these two facts, one can deduce the result by following the argument of Lemma \[resconv\].
To prove convergence of the CSRW to the $\alpha$-FIN diffusion, we introduce the random time-change measures $\nu_n$, as given by $\nu_n(\{x\})=\sum_{e\in E_n:\:x\in e}\omega_e^n$. In Lemma \[nunconvssf\], we will prove convergence to the limiting FIN measure $\nu$, again obtained from a Poisson process on $(0,\infty)\times F$ with intensity $\alpha v^{-1-\alpha}dv\mu(dx)$, under the following assumption. We note, in this setting, it makes sense to state convergence results with respect to the original Euclidean topology, since the objects already have a natural (non-isometric) embedding there. Moreover, we observe that the assumption is satisfied for i.i.d. edge weights, each with tails satisfying the same distributional asymptotics.
[\[ssfassu\] There exists a constant $c>0$ such that the random conductance distribution satisfies $$\mathbf{P}\left(\sum_{e\in E_0}\omega^0_e>u\right)
\sim c u^{-\alpha}$$ for some $\alpha\in(0,1)$.]{}
[\[nunconvssf\] Suppose Assumption \[ssfassu\] holds, then there exists a constant $c_0>0$ such that $c_0^{-1}b_n^{1/\alpha}\nu_n\rightarrow \nu$ in distribution with respect to the weak topology for finite measures on $\mathbb{R}^d$.]{}
The proof is again similar to the tree case (Lemma \[nunconv3\]). In particular, it is an easy exercise to check that there exists a constant $c_0>0$ such that, for any $i_1,\dots,i_m\in\{1,\dots,N\}$, $b_n^{1/\alpha}\nu_n(\Psi_{i_1,\dots,i_m}(F))\rightarrow c_0\nu(\Psi_{i_1,\dots,i_m}(F))$ in distribution. From this, the result again follows from [@Kall Theorem 16.16].
From Lemmas \[ssfresconv\] and \[nunconvssf\], we are able to prove the main result of this section by a similar argument to the proof of Proposition \[rcmtreeresult\]; we thus state it without proof. We write $\mathbb{P}^{{\rm VSRW}_n}_x$ for the annealed law of the VSRW $X^{n,\omega}$ on the graph $V_n$ with conductances $\omega^n$, started from $x$. (Note that in Proposition \[rcmtreeresult\] convergence of VSRW is shown $\mathbf{P}$-a.s., but here we have only annealed convergence since the convergence in Theorem \[thm:conth\] is in the ${L}^1$-sense.) We write $\mathbb{P}^{{\rm CSRW}_n}_{x}$ for the annealed law of the corresponding CSRW $X^{n,\omega,\nu}$. We write $P_x$ for the law of the Brownian motion on $(F,R,\mu)$ started from $x$, and $\mathbb{P}^{{\rm FIN}}_{\rho}$ is the annealed law of the associated $\alpha$-FIN diffusion, defined as in Section \[bouchsec\].
[\[ssfrcmresult\] Suppose Assumptions \[thm:ass1\] and \[thm:ass2\] hold, and that the conductances $(\omega_e^0)_{e\in E_0}$ are uniformly bounded from below. It is then the case that $$\mathbb{P}^{{\rm VSRW}_n}_{\rho_n}\left(\left(X^{n,\omega}_{t/ a_nb_n}\right)_{t\geq 0}\in\cdot\right)\rightarrow {P}_\rho\left(\left(X_{t}\right)_{t\geq 0}\in\cdot\right)$$ weakly as probability measures on $D(\mathbb{R}_+, \mathbb{R}^d)$. Moreover, if Assumption \[ssfassu\] also holds, then $$\mathbb{P}^{{\rm CSRW}_n}_{\rho_n}\left(\left(X^{n,\omega,\nu}_{c_0 t/ a_nb_n^{1/\alpha}}\right)_{t\geq 0}\in\cdot\right)
\rightarrow\mathbb{P}^{{\rm FIN}}_{\rho}\left(\left(X^{\nu}_{t}\right)_{t\geq 0}\in\cdot\right)$$ weakly as probability measures on $D(\mathbb{R}_+,
\mathbb{R}^d)$.]{}
[\[exa6-18\] [To continue with the example of the Sierpiński gasket graphs from previous sections, one can also apply Proposition \[ssfrcmresult\] for this collection. In particular, assuming that the conductances are uniformly bounded below and have at most polynomial decay at infinity, we know that nested fractals satisfy both Assumption \[thm:ass1\] and \[thm:ass2\], and so we obtain the annealed convergence of the VSRW on the Sierpiński gasket graphs as at (\[sgvsrw\]). Moreover, if it is further the case that the tail behaviour at infinity of the conductances satisfies Assumption \[ssfassu\], then we also have the annealed convergence of the CSRW as at (\[sgcsrw\]).]{}]{}
[\[finitemoments\] [When $\mathbf{E}\omega_e^0<\infty$ for each $e\in E_0$, one obtains in place of Lemma \[nunconvssf\] (via the same argument) that there exists a constant $c_0$ such that $c_0^{-1}b_n\nu_n\rightarrow\mu$. Consequently, if Assumption \[ssfassu\] is replaced by the assumption of finite first moments, then one can check the annealed limit of the CSRW is Brownian motion, rather than the FIN diffusion that appears in the second statement of Proposition \[ssfrcmresult\]. A similar remark pertains to Proposition \[btmresult\] and the second statement of Proposition \[rcmtreeresult\].]{}]{}
[10]{}
R. Abraham, J.-F. Delmas, and P. Hoscheit, *A note on the [G]{}romov-[H]{}ausdorff-[P]{}rokhorov distance between (locally) compact metric measure spaces*, Electron. J. Probab. **18** (2013), no. 14, 21.
S. Andres and L. Hartung, *Diffusion processes on branching [B]{}rownian motion*, in preparation.
S. Andres and N. Kajino, *Continuity and estimates of the [L]{}iouville heat kernel with applications to spectral dimensions*, Probab. Theory Related Fields, to appear.
S. Athreya, W. Löhr, and A. Winter, *The gap between [G]{}romov-vague and [G]{}romov-[H]{}ausdorff-vague topology*, preprint.
M. T. Barlow, *Diffusions on fractals*, Lectures on probability theory and statistics ([S]{}aint-[F]{}lour, 1995), Lecture Notes in Math., vol. 1690, Springer, Berlin, 1998, pp. 1–121.
M. T. Barlow and J. [Č]{}ern[ý]{}, *Convergence to fractional kinetics for random walks associated with unbounded conductances*, Probab. Theory Related Fields **149** (2011), no. 3-4, 639–673.
M. T. Barlow, D. A. Croydon, and T. Kumagai, *Subsequential scaling limits of simple random walk on the two-dimensional uniform spanning tree*, Ann. Probab., to appear.
M. T. Barlow and E. A. Perkins, *Brownian motion on the [S]{}ierpiński gasket*, Probab. Theory Related Fields **79** (1988), no. 4, 543–623.
G. Ben Arous, M. Cabezas, J. [Č]{}ern[ý]{}, and R. Royfman, *Randomly trapped random walks*, Ann. Probab. **43** (2015), no. 5, 2405–2457.
G. Ben Arous and J. [Č]{}ern[ý]{}, *Bouchaud’s model exhibits two different aging regimes in dimension one*, Ann. Appl. Probab. **15** (2005), no. 2, 1161–1192.
N. Berestycki, *Diffusion in planar [L]{}iouville quantum gravity*, Ann. Inst. Henri Poincaré Probab. Stat. **51** (2015), no. 3, 947–964.
R. M. Blumenthal and R. K. Getoor, *Markov processes and potential theory*, Pure and Applied Mathematics, Vol. 29, Academic Press, New York-London, 1968.
J. [Č]{}ern[ý]{}, *On two-dimensional random walk among heavy-tailed conductances*, Electron. J. Probab. **16** (2011), no. 10, 293–313.
Z.-Q. Chen and M. Fukushima, *Symmetric [M]{}arkov processes, time change, and boundary theory*, London Mathematical Society Monographs Series, vol. 35, Princeton University Press, Princeton, NJ, 2012.
D. A. Croydon, *Scaling limits for simple random walks on random ordered graph trees*, Adv. in Appl. Probab. **42** (2010), no. 2, 528–558.
[to3em]{}, *Moduli of continuity of local times of random walks on graphs in terms of the resistance metric*, Trans. London Math. Soc. **2** (2015), no. 1, 57–79.
D. A. Croydon, B. M. Hambly, and T. Kumagai, *Heat kernel estimates for [FIN]{} diffusions associated with resistance forms*, in preparation.
D. A. Croydon and S. Muirhead, *Functional limit theorems for the [B]{}ouchaud trap model with slowly varying traps*, Stochastic Process. Appl. **125** (2015), no. 5, 1980–2009.
B. Duplantier and S. Sheffield, *Liouville quantum gravity and [KPZ]{}*, Invent. Math. **185** (2011), no. 2, 333–393.
L. R. G. Fontes, M. Isopi, and C. M. Newman, *Random walks with strongly inhomogeneous rates and singular diffusions: convergence, localization and aging in one dimension*, Ann. Probab. **30** (2002), no. 2, 579–604.
M. Fukushima, Y. Oshima, and M. Takeda, *Dirichlet forms and symmetric [M]{}arkov processes*, extended ed., de Gruyter Studies in Mathematics, vol. 19, Walter de Gruyter & Co., Berlin, 2011.
C. Garban, R. Rhodes, and V. Vargas, *Liouville [B]{}rownian motion*, Ann. Probab., to appear.
A. M. Garsia, *Continuity properties of [G]{}aussian processes with multidimensional time parameter*, Proceedings of the [S]{}ixth [B]{}erkeley [S]{}ymposium on [M]{}athematical [S]{}tatistics and [P]{}robability ([U]{}niv. [C]{}alifornia, [B]{}erkeley, [C]{}alif., 1970/1971), [V]{}ol. [II]{}: [P]{}robability theory, Univ. California Press, Berkeley, Calif., 1972, pp. 369–374.
A. M. Garsia, E. Rodemich, and H. Rumsey, Jr., *A real variable lemma and the continuity of paths of some [G]{}aussian processes*, Indiana Univ. Math. J. **20** (1970/1971), 565–578.
R. K. Getoor and H. Kesten, *Continuity of local times for [M]{}arkov processes*, Compositio Math. **24** (1972), 277–303.
B. M. Hambly and T. Kumagai, *Heat kernel estimates for symmetric random walks on a class of fractal graphs and stability under rough isometries*, Fractal geometry and applications: a jubilee of [B]{}enoît [M]{}andelbrot, [P]{}art 2, Proc. Sympos. Pure Math., vol. 72, Amer. Math. Soc., Providence, RI, 2004, pp. 233–259.
N. Kajino, *[N]{}eumann and [D]{}irichlet heat kernel estimates in inner uniform domains for local resistance forms*, in preparation.
O. Kallenberg, *Foundations of modern probability*, second ed., Probability and its Applications (New York), Springer-Verlag, New York, 2002.
J. Kigami, *Harmonic calculus on limits of networks and its application to dendrites*, J. Funct. Anal. **128** (1995), no. 1, 48–86.
[to3em]{}, *Analysis on fractals*, Cambridge Tracts in Mathematics, vol. 143, Cambridge University Press, Cambridge, 2001.
[to3em]{}, *Resistance forms, quasisymmetric maps and heat kernel estimates*, Mem. Amer. Math. Soc. **216** (2012), no. 1015, vi+132.
J. F. C. Kingman, *Poisson processes*, Oxford Studies in Probability, vol. 3, The Clarendon Press, Oxford University Press, New York, 1993, Oxford Science Publications.
T. Kumagai, *Heat kernel estimates and parabolic [H]{}arnack inequalities on graphs and resistance forms*, Publ. Res. Inst. Math. Sci. **40** (2004), no. 3, 793–818.
[to3em]{}, *Homogenization on finitely ramified fractals*, Stochastic analysis and related topics in [K]{}yoto, Adv. Stud. Pure Math., vol. 41, Math. Soc. Japan, Tokyo, 2004, pp. 189–207.
T. Kumagai and S. Kusuoka, *Homogenization on nested fractals*, Probab. Theory Related Fields **104** (1996), no. 3, 375–398.
T. Kumagai and O. Zeitouni, *Fluctuations of maxima of discrete [G]{}aussian free fields on a class of recurrent graphs*, Electron. Commun. Probab. **18** (2013), no. 75, 12.
S. Kusuoka, *Statistical mechanics and fractals*, ch. Lecture on diffusion processes on nested fractals, pp. 39–98, Springer Berlin Heidelberg, Berlin, Heidelberg, 1993.
T. Lindstr[ø]{}m, *Brownian motion on nested fractals*, Mem. Amer. Math. Soc. **83** (1990), no. 420, iv+128.
P. Maillard, R. Rhodes, V. Vargas, and O. Zeitouni, *Liouville heat kernel: regularity and bounds*, Ann. Inst. Henri Poincaré Probab. Stat., to appear.
M. B. Marcus and J. Rosen, *Sample path properties of the local times of strongly symmetric [M]{}arkov processes via [G]{}aussian processes*, Ann. Probab. **20** (1992), no. 4, 1603–1684.
[to3em]{}, *Markov processes, [G]{}aussian processes, and local times*, Cambridge Studies in Advanced Mathematics, vol. 100, Cambridge University Press, Cambridge, 2006.
R. Peirone, *Convergence and uniqueness problems for [D]{}irichlet forms on fractals*, Boll. Unione Mat. Ital. Sez. B Artic. Ric. Mat. (8) **3** (2000), no. 2, 431–460.
C. Sabot, *Existence and uniqueness of diffusions on finitely ramified self-similar fractals*, Ann. Sci. École Norm. Sup. (4) **30** (1997), no. 5, 605–673.
W. Whitt, *Some useful functions for functional limit theorems*, Math. Oper. Res. **5** (1980), no. 1, 67–85.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'J. Devin$^{(1)}$'
- 'F. Acero$^{(2)}$'
- 'J. Ballet$^{(2)}$'
- 'J. Schmid$^{(2)}$'
bibliography:
- 'article.bib'
date: 'Received: 13 March 2018 / Accepted: 25 May 2018'
title: 'Disentangling hadronic from leptonic emission in the composite SNR G326.3$-$1.8'
---
[G326.3$-$1.8 (also known as MSH 15$-$5*6*) has been detected in radio as a middle-aged composite supernova remnant (SNR) consisting of an SNR shell and a pulsar wind nebula (PWN), which has been crushed by the SNR’s reverse shock. Previous $\gamma$-ray studies of SNR G326.3$-$1.8 revealed bright and extended emission with uncertain origin. Understanding the nature of the $\gamma$-ray emission allows probing the population of high-energy particles (leptons or hadrons) but can be challenging for sources of small angular extent.]{} [With the recent $\textit{Fermi}$ Large Area Telescope data release Pass 8 providing increased acceptance and angular resolution, we investigate the morphology of this SNR to disentangle the PWN from the SNR contribution. In particular, we take advantage of the new possibility to filter events based on their angular reconstruction quality. ]{} [We perform a morphological and spectral analysis from 300 MeV to 300 GeV. We use the reconstructed events with the best angular resolution (PSF3 event type) to separately investigate the PWN and the SNR emissions, which is crucial to accurately determine the spectral properties of G326.3$-$1.8 and understand its nature.]{} [The centroid of the $\gamma$-ray emission evolves with energy and is spatially coincident with the radio PWN at high energies (E $>$ 3 GeV). The morphological analysis reveals that a model considering two contributions from the SNR and the PWN reproduces the $\gamma$-ray data better than a single-component model. The associated spectral analysis using power laws shows two distinct spectral features, a softer spectrum for the remnant ($\Gamma$ = 2.17 $\pm$ 0.06) and a harder spectrum for the PWN ($\Gamma$ = 1.79 $\pm$ 0.12), consistent with hadronic and leptonic origin for the SNR and the PWN respectively. Focusing on the SNR spectrum, we use one-zone models to derive some physical properties and, in particular, we find that the emission is best explained with a hadronic scenario in which the large target density is provided by radiative shocks in H[i]{} clouds struck by the SNR.]{}
Introduction
============
Supernova remnants (SNRs) and pulsar wind nebulae (PWNe) have long been considered potential sources of Galactic cosmic rays and have therefore been investigated over a wide range of energies. In SNRs, the fast shock wave propagating into the interstellar medium (ISM) or the circumstellar medium is thought to accelerate particles (electrons and protons), which gain energy through first order Fermi acceleration [@Bell:1978] also known as the Diffusive Shock Acceleration mechanism (DSA). In core-collapse SNRs, a very fast rotating and highly magnetized pulsar can give rise to a PWN, in which electrons and positrons from the pulsar wind are re-accelerated to relativistic energies at a termination shock. These Galactic accelerators have mostly been studied independently while in case of core-collapse SNRs, the SNR, the PWN and the pulsar are part of the same object. However, for systems with angular sizes smaller or comparable to the instrument point-spread function (PSF), it can be difficult to assess the origin of the emission, in particular at $\gamma$-ray energies where the angular resolution is comparatively much larger than in the radio and X-ray ranges. Although it can be challenging to understand their origin, these $\gamma$ rays allow probing the population of high-energy particles, such as accelerated electrons interacting with the Cosmic Microwave Background (CMB) or other target photons by Inverse Compton (IC) scattering, and also accelerated protons interacting with gas that produce neutral pions which decay into $\gamma$ rays. Morphological studies complementing purely spectral analyses have the potential to help identify multiple particle acceleration regions in one object such as interaction regions with surrounding clouds or the different emissions coming from the SNR and/or the PWN in composite objects.
With an SNR shell and a PWN seen at radio wavelengths, the Galactic SNR G326.3$-$1.8 is a prototype of the so-called composite SNRs [@Mills; @Milne]. Its distance is estimated between 3.1 kpc [@Goss:1972] and 4.1 kpc [@Rosado:1996] as established by the H[i]{} absorption profile and H$\alpha$ velocity measurements respectively. @Temim:2013 estimated this SNR to be 16,500 years old with a shock velocity of 500 km s$^{-1}$, expanding in an ISM density of $n_{0}$ = 0.1 cm$^{-3}$.
Figure \[fig:MOST\], obtained from radio observations [@MOSTcat:1996], shows a symmetric SNR shell with 0.3° radius and a PWN trailing the putative pulsar and likely crushed by the remnant’s reverse shock. Non-thermal radio emission has been reported with a spectral index of $\alpha$ = 0.34 for the shell and $\alpha$ = 0.18 for the nebula where $S_{\nu}$ $\propto$ $\nu^{-\alpha}$ [@Dickel:2000]. Optical H$\alpha$ filaments were observed in the southwest and northeast parts of the remnant, and appear to spatially correlate with the shell [@Vandenbergh:1979; @Dennefeld:1980] indicating the presence of neutral material at the shock front. The PWN component is highly polarized with a luminosity of L($10^{7} − 10^{11}$ Hz) ∼ $5 × 10^{34}$ erg s$^{−1}$ [@Dickel:2000]. The associated pulsar has not been detected but *Chandra* maps have revealed a point source embedded in the X-ray PWN located in the southwest of the radio nebula [@Temim:2013]. SNR G326.3$-$1.8 was also detected in X-rays by *ROSAT* [@Kassim:1993] and *ASCA* [@Plucinsky:1998] showing a complete shell that spatially correlates with the radio SNR while the width of the PWN in X-rays shrinks near the compact object. At higher energies, previous $\gamma$-ray studies have revealed emission with uncertain origin [@Temim:2013] and SNR G326.3$-$1.8 has only recently been found to be extended with the $\textit{Fermi}$-LAT data [@First_SNRcat:2016].
The latest Large Area Telescope data release Pass 8 [@Pass8:Atwood:2013] allows not only a claim of significant extension of the $\gamma$-ray emission, but also a study of the PWN and SNR contributions separately. This distinction might be crucial for understanding the underlying emission mechanisms and potentially distinguishing between hadronic and leptonic nature of the constituents. In this paper, we briefly describe the latest data release Pass 8 before presenting a morphological study of SNR G326.3$-$1.8. In particular, we investigate its energy-dependent morphology and model the emission with different templates. We also report a spectral analysis of our best models using two spatial components for the $\gamma$-ray emission and derive physical properties using one-zone models for the SNR spectrum.
*Fermi*-LAT and Pass 8 description \[sec:Fermi\]
================================================
The Large Area Telescope (LAT) on board the $\textit{Fermi}$ satellite is a pair-conversion instrument sensitive to $\gamma$ rays in the energy range from 30 MeV to more than 300 GeV.\
Since the launch in August 2008, the $\textit{Fermi}$-LAT event reconstruction algorithm has been progressively upgraded to make use of the increasing understanding of the instrument performance as well as the environment in which it operates. Following Pass 7, released in August 2011, Pass 8 is the latest version of the $\textit{Fermi}$-LAT data release [@Pass8:Atwood:2013]. The enhanced reconstruction and classification algorithms result in improvements of the effective area, the PSF and the energy resolution. One major advance with respect to previous releases is the classification of detected photon events according to their reconstruction quality. The data set is hence divided into types of events with different energy or angular reconstruction qualities. The PSF selection divides the data into four parts: from PSF0 to PSF3, the latter being the quartile with the best angular resolution (68% containment radius of 0.4° at 1 GeV compared to 0.8° without selection). This type of event selection, combined with the large amount of data collected by the LAT since its launch, makes Pass 8 $\gamma$ rays a powerful tool to identify and study extended $\gamma$-ray sources.
Data analysis\[sec:ata\]
========================
We perform a binned analysis using 6.5 years of data collected from August 4, 2008 to January 31, 2015, within a 10° $\times$ 10° region around the position of SNR G326.3$-$1.8. Since the object remains significant with $25\%$ of the data (more than 24$\sigma$ between 300 MeV and 300 GeV), we take advantage of the new PSF3 selection to limit contamination between the PWN and the SNR components as well as that from the Galactic plane. We select events between 300 MeV and 300 GeV, with a maximum zenith angle of 100° to reduce the contamination of the bright Earth limb. Time intervals during which the rocking angle of the satellite was more than 52° are excluded as well as those during which it passed through the South Atlantic Anomaly. We set the pixel size to 0.05° and divide the whole energy range (300 MeV – 300 GeV) into 30 bins. We use version 10 of the `Science Tools` (v10r0p5) and the `P8R2_V6` Instrument Response Functions (IRFs) with the `SOURCE` event class for the following analysis[^1]. The resulting count map of the 10° $\times$ 10° region centered on the position of the SNR is shown in Figure \[fig:Cmap\].
The $\gamma$-ray data around the source are modeled starting with the $\textit{Fermi}$-LAT 3FGL source catalog [@3FGL], complemented by the extended source FGES J1553.8$-$5325[^2] [@FGES:Pass8]. We first fit the point sources and extended sources within a 10° radius (additionally accounting for the most significant sources between 10° and 15°) simultaneously with the Galactic and isotropic diffuse emissions described by the files `gll_iem_v06.fits` [@GIEM_Fermi:2016] and `iso_P8R2_SOURCE_V6_v06_PSF3.txt` respectively [^3]. We then compute a residual test statistic (TS) map to search for additional sources.
The TS is defined to test the likelihood of one hypothesis $\mathcal{L}_1$ (including a source) against the null hypothesis $\mathcal{L}_0$ (absence of source), such that: $$\hspace{2.5cm} \textrm{TS} = 2 \times (\log{\mathcal{L}_1 } - \log{\mathcal{L}_0}) \label{eq:TS}$$ This can be directly interpreted in terms of significance of hypothesis $1$ with respect to the null hypothesis $0$ in which the TS follows a $\chi^2$-law with $n$ degrees of freedom for $n$ additional parameters. To evaluate the significance of putative new sources, we compute a two-dimensional residual TS map that tests the hypothesis of a point source with a generic $E^{-2}$ spectrum against the null hypothesis at each point in the sky. The positions where the TS values exceed 25 (corresponding to a significance of more than $4\sigma$) are used as seeds to identify $\gamma$-ray sources in addition to the 3FGL. In that way, we iteratively add eleven sources in the 10° $\times$ 10° region and we fix their spectral parameters to their best-fit values found with `gtlike`. Figure \[fig:TSmap\_all\_sky\] shows the final residual TS map including all the sources. Note that the apparent diffuse residual emission (for which TS$_{\rm{max}}$ $\approx$ 17) disappears above 500 MeV.
![Residual TS map from 300 MeV to 300 GeV of a 10° $\times$ 10° region centered on the SNR and using the PSF3 events. The pixel size is 0.25° and the radio contours of the SNR are overlaid in white. The white circle is an FGES extended source. The white crosses are the 3FGL point sources and the red crosses are the sources added to the model. \[fig:TSmap\_all\_sky\]](TSmap_ALL_Sky_40pix.png)
Morphological analysis\[sec:Morphological\_analysis\]
-----------------------------------------------------
### Extension\[sec:Extension\]
The 3FGL catalog compiled by the $\textit{Fermi}$-LAT collaboration [@3FGL] has two point-like sources tentatively associated with the SNR, which we remove for our analysis. Since the creation of the first SNR catalog [@First_SNRcat:2016], G326.3$-$1.8 has been known to show extended $\gamma$-ray emission, and its radius has been determined to be $0.21^\circ$ using an extended uniform disk model, somewhat smaller than the $0.31^\circ$ radius of the radio shell but larger than the radio PWN. However, that analysis was based on only three years of data and made use of the former Pass 7 data release.
With the latest Pass 8 data and using the PSF3 event type, we revisit the morphology of this SNR to understand the nature of the $\gamma$-ray emission. We start by finding the best position of a point source, modeling its emission as a power law and using the `pointlike` framework [@Kerr:2010] from 300 MeV to 300 GeV. Then, we investigate the extension of the $\gamma$-ray emission using a 2D-symmetric Gaussian and a disk with a uniform brightness. Table \[tab:Fit\_Pointlike\] shows the respective best-fit position and extension – if extended – for the different spatial models. The significance of a source extension is expressed in terms of the test statistic TS$_{\rm ext}$, where the hypothesis of the best-fit extended spatial model is tested against the null hypothesis of the best-fit point-like source. Given that in both hypotheses the localization of the source is optimized, the extended source model adds one degree of freedom – the source size – with respect to the point-source model. Thus, the significance can be directly interpreted as the square root $\sqrt{\rm{TS_{ext}}}$. As reported in Table \[tab:Fit\_Pointlike\], the $\gamma$-ray emission is extended with more than 13$\sigma$ confidence level and the uniform disk radius is found to be $r$ = $0.266^\circ$ $\pm$ 0.012$^\circ$. Figure \[fig:Best\_Disk\_Gaussian\] (left) shows the best-fit position and extension (68% containment radius $r_{68}$) for the disk and the Gaussian, plotted on the radio image, with the associated uncertainties. The centroid of each extended model is slightly shifted toward the radio PWN but is not coincident with its position. No significant residual emission appears in the residual TS map (not shown here) including either the disk or the Gaussian in the model.
[l|ccccccc]{}
------------------------------------------------------------------------
Spatial model & RA$_{\rm{J2000}}$ ($^\circ$)& Dec$_{\rm{J2000}}$ ($^\circ$) & $r$ or $\sigma$ ($^\circ$) & $r_{68}$ ($^\circ$) & TS & TS$_{\rm ext}$\
------------------------------------------------------------------------
Point source & 238.167 $\pm$ 0.009 & $-$56.181 $\pm$ 0.008 & — & — & 689.5 & —\
------------------------------------------------------------------------
Disk & 238.170 $\pm$ 0.012 & $-$56.152 $\pm$ 0.012 & 0.266 $\pm$ 0.012 & 0.218 $\pm$ 0.010 & 866.5 & 177.0\
------------------------------------------------------------------------
Gaussian & 238.157 $\pm$ 0.013 & $-$56.166 $\pm$ 0.012 & 0.134 $\pm$ 0.009 & 0.202 $\pm$ 0.014 & 863.4 & 173.9\
{width="40.00000%"} {width="40.00000%"}
### Energy-dependent morphology\[sec:E\_dpd\_morpho\]
Although the $\gamma$-ray emission can be adequately described with a one-component model, either a disk or a 2D-symmetric Gaussian, this stands in slight tension to the discovery of a hard point-like $\gamma$-ray source above 50 GeV [@2FHL] at the location of the PWN, clearly displaced from the center of the SNR shell. To investigate the morphology in more detail, we divide the data into five logarithmically spaced energy bins from 300 MeV to 300 GeV that we subsequently fit individually with `pointlike` using a 2D-symmetric Gaussian. Because the PSF width depends strongly on energy up to $\sim$ 10 GeV, and our energy bins are quite broad (half a decade), we need to adopt a specific spectral model for the source. We describe it with a power law with free spectral index. The normalizations of the source, the Galactic and isotropic diffuse emissions are let free while the spectral parameters of the other sources are fixed to their best-fit values. Figure \[fig:Best\_Disk\_Gaussian\] (right) depicts the results of the fitting procedure in the individual energy bands. At low energies (300 MeV – 1 GeV), the PSF ($r_{68}$ $\sim$ 0.4$^{\circ}$ at 1 GeV) is larger than the SNR radius. An extended source of the SNR size is compatible with the data but is not significantly better than a point source. The associated best-fit position lies outside the radio PWN. Between $1$ and $3$ GeV, the significance of the extension is more than 5$\sigma$ (the values are reported in Figure \[fig:TS\_and\_TSext\]) and the position of the Gaussian appears to be fairly consistent with the center of the radio SNR. At higher energies (from 3 to 30 GeV), the $\gamma$-ray morphology is still significantly extended (more than 5$\sigma$) and the centroid of the best-fit Gaussian gets closer to the radio PWN. Above 30 GeV, the $\gamma$-ray emission is not significantly extended and the best-fit position lies inside the radio PWN.
### Building a more detailed model
This energy-dependent source morphology clearly requires a more detailed investigation beyond a one-component modeling. Since the PSF below 1 GeV is not small enough to resolve the SNR, the following morphological analysis uses data between 1 and 300 GeV.
Electrons and positrons, accelerated in the PWN, that radiate by synchrotron emission are expected to also radiate in the GeV band by IC scattering on photon fields. Since this SNR is relatively young, particles are well-confined inside the PWN. [@Temim:2013] estimate the magnetic field to be $B_{\rm{PWN}}$ $\approx$ 34 $\mu$G when the SNR has already begun significantly compressing the PWN at an age of 19,000 years. The emission seen in the radio band should track the older accelerated electrons and we expect that the extension of the $\gamma$-ray emission should not be larger than the radio emission. We thus model the $\gamma$-ray emission from the PWN using its radio template (see Figure \[fig:Templates\], left panel), knowing that the magnetic field spatial distribution inside the PWN should only moderately impact our model given the small size of the PWN compared to the $\textit{Fermi}$-LAT PSF.
![\[fig:TS\_and\_TSext\]Test statistic of the source (black bars) and of extension (colored bars) for the best-fit Gaussian in individual energy bands.](TS_TSext_bis-crop.pdf)
{width="50.00000%"} {width="50.00000%"}
Since the best-fit position of the $\gamma$-ray emission is consistent with that of the PWN at high energies, we first assume that the $\gamma$-ray emission comes only from the PWN. We model its spectrum as a power law and we perform a likelihood fit where the spectral parameters of the PWN, those of the nearest point source, the Galactic and isotropic diffuse emissions are free during the fit. The spectral parameters of the other sources are fixed to their best-fit values since they are further than $\sim$ 2° from G326.3$-$1.8. The fit gives a TS value of TS = 593.4, as reported in Table \[tab:LL\], with the number of additional free parameters compared to the model without source.
[l|ccc]{}
------------------------------------------------------------------------
Spatial models & TS & N$_{\rm{dof}}$ & TS$_{\rm{PWN}}$\
------------------------------------------------------------------------
Radio PWN & 593.4 & 2 & —\
------------------------------------------------------------------------
Point source & 503.3 & 4 & —\
------------------------------------------------------------------------
Point source + radio PWN & 661.4 & 6 & 158.1\
------------------------------------------------------------------------
Disk & 681.8 & 5 & —\
------------------------------------------------------------------------
Disk + radio PWN & 694.8 & 7 & 13.0\
------------------------------------------------------------------------
Radio SNR & 667.3 & 2 & —\
------------------------------------------------------------------------
Radio SNR + radio PWN & 683.0 & 4 & 15.7\
------------------------------------------------------------------------
SNR mask & 670.3 & 2 & —\
------------------------------------------------------------------------
SNR mask + radio PWN & 696.4 & 4 & 26.1\
Figure \[fig:TSmap\_PWN\] (left) depicts the 1° $\times$ 1° residual TS map from 1 to 300 GeV obtained by fixing the spectral parameters of the radio PWN and the nearest point source to their best-fit values. The TS map tests a putative point source. It shows qualitatively where there is missing signal, and extended emission can only be more significant than the peak of the TS map. The maximum TS value of the map is TS $\approx$ 60 indicating that this residual emission is clearly significant. The radio template of the PWN is thus not sufficient to describe the data. This confirms our previous results in the Section \[sec:E\_dpd\_morpho\], that show that the emission below 3 GeV lies outside of the radio contours of the PWN.
To model the contribution of an additional component that seems to give rise to the low-energy part, we test several templates using first a simple disk component and then physically motivated templates (derived from the radio map of the SNR). We first use the `pointlike` framework to find the best position and extension of an additional source, described by a disk, when the PWN is already included in the model. The fit localizes the position near the center of the SNR at RA$_{\rm{J2000}}$ = $238.169^\circ$ $\pm$ $0.013^\circ$ and Dec$_{\rm{J2000}}$ = $-56.133^\circ$ $\pm$ $0.014^\circ$ with a radius $r$ = $0.295^\circ$ $\pm$ $0.013^\circ$, similar to the radio extension of the SNR ($0.31^\circ$). The significance of the extension is 5.8$\sigma$ (TS$_{\rm ext}$= 33.4), calculated with the TS value of the model including the point source and the radio PWN (reported in Table \[tab:LL\]). This rules out the hypothesis of a point-like source being responsible for the additional emission such as an active galactic nucleus behind the SNR.
We further use the radio observations to derive two other templates for the SNR. First, we use the radio map replacing the contribution of the nebula by the average value of the radio emission around it (labeled “radio SNR”; see Figure \[fig:Templates\], center). From that, we also create another template, following the radio shock and filled homogeneously (called here “SNR mask”; see Figure \[fig:Templates\], right).
[l|c|cc|cc]{}
------------------------------------------------------------------------
Values from the fit & Disk only & radio PWN & Disk & radio PWN & SNR mask\
------------------------------------------------------------------------
$\Phi$ (ph cm$^{-2}$ s$^{-1}$) $\times$ $10^{-8}$ & 2.70 $\pm$ 0.15 & 0.23 $\pm$ 0.11 & 2.43 $\pm$ 0.23 & 0.33 $\pm$ 0.11 & 2.31 $\pm$ 0.23\
------------------------------------------------------------------------
$\Gamma$ & 2.07 $\pm$ 0.04 & 1.74 $\pm$ 0.15 & 2.16 $\pm$ 0.06 & 1.79 $\pm$ 0.12 & 2.17 $\pm$ 0.06\
 
For all components, the $\gamma$-ray emission is described by a power law and the free spectral parameters are those of the components, the nearest point source, and those of the Galactic and isotropic diffuse emissions. The results from our maximum likelihood fit are given in Table \[tab:LL\] with the numbers of free parameters associated to the models (spectral and/or spatial).\
First, the TS values obtained using the one-component models (disk, radio SNR or SNR mask) alone are clearly higher than that obtained using only the radio PWN, indicating that the fit prefers a model more extended than the radio PWN. For comparison, Figure \[fig:TSmap\_PWN\] (right) depicts the residual TS map when only the SNR mask is included to the model, showing a residual emission coincident with the position of the PWN. In terms of test statistic, when adding a second component, the model with one component becomes the null hypothesis to test the significance of the second component. In Table \[tab:LL\], we compare the TS values obtained with each of the one-component models to the two-component models testing the improvement of the fit when adding the radio template of the PWN. The difference TS$_{\rm{PWN}}$ can be converted to a significance since in the null hypothesis (no PWN emission) it behaves as a $\chi^{2}$-law with two degrees of freedom.
For all our extended models, the significance of adding the PWN lies between 3 and 4$\sigma$ and the maximum TS values are obtained for the model including the radio PWN with either the disk or the SNR mask. In terms of significance, our best model involves the radio PWN and the SNR mask since it requires fewer free parameters during the fit than the disk whose spatial components have been optimized. The lower TS value using the two radio templates (for the SNR and the PWN) indicates that the $\gamma$-ray emission does not entirely follow the synchrotron distribution and the fit prefers a more homogeneous structure for the shell, keeping in mind that this conclusion depends on the model we choose for the PWN.
Spectral analysis\[sec:spectral\_analysis\]
-------------------------------------------
To understand the underlying emission processes, we perform a spectral analysis from 300 MeV to 300 GeV using our best models found in the previous section: the disk alone and the radio PWN with either the SNR mask or the disk. Here the $\gamma$-ray emissions are still described with power laws since other spectral representations did not improve the fit. We do not take into account the energy dispersion (this induces a bias $\sim$ $-$ 5% on flux[^4]). Using `gtlike`, we perform a maximum likelihood fit leaving the same spectral parameters free as before. Table \[tab:Fit\] reports the results obtained from the fit.
When we use the disk alone to describe the $\gamma$-ray emission, the photon index is found to be close to 2. Using differentiated models for the PWN and the SNR, the fit leads to a spectral separation between the two components: a softer spectrum for the remnant ($\Gamma \approx 2.16$ using the disk and $\Gamma \approx 2.17$ using the SNR mask) and a harder spectrum for the nebula ($\Gamma \approx 1.74$ and $\Gamma \approx 1.79$). The choice of the model for the remnant (either the disk or the SNR mask) has a very slight impact on the spectral study.
To compute the spectral energy distribution (SED), we divide the whole energy range (300 MeV – 300 GeV) into six bins and impose a TS threshold of 1 per energy bin for the flux calculation; otherwise an upper limit is calculated. In each bin, the photon indexes of the sources of interest are fixed to 2 to avoid any dependence on the spectral models. The fluxes of the PWN and the SNR components are let free during the fit as well as those of the Galactic and isotropic diffuse emissions. All other sources are fixed to their best global model.
Figure \[fig:SEDs\] shows the SED of the uniform disk (left) and the SED of the best-fit two-component model (right) using the radio PWN and the SNR mask. The colored error bars represent the statistical errors while the quadratic sums of the statistical and systematic errors [calculated using eight alternative Galactic diffuse emission models as explained in the first $\textit{Fermi}$-LAT supernova remnant catalog, @First_SNRcat:2016] are represented with black horizontal bars. The systematic errors are never dominant and are comparable to the statistical ones only in the first band. Note that the effective area uncertainty also induces systematic errors (10% between 100 MeV and 100 GeV). These SEDs clearly emphasize that two distinct morphologies give rise to two distinct spectral signatures, while the different emissions seem to be mixed when we use a single component model. The TS values in each energy bin highlight the different contributions of the two components: at low energy (E $<$ 10 GeV), the emission is dominated by the SNR while the contribution of the PWN becomes important above 10 GeV, bringing out the spatial and spectral distinctions between these two nested objects.
Results and discussion
======================
For this entire section we assume the distance to the SNR is 4.1 kpc [@Temim:2013].
SNR spectrum
------------
To understand the observed $\gamma$-ray spectrum of the SNR, we perform multi-wavelength modeling using the one-zone models provided by the `naima` package [@naima]. From [@Dickel:2000] we take the five radio flux measurements of the shell. As there is no associated synchrotron emission in the X-ray domain, we use the *ROSAT* thermal flux reported by [@Kassim:1993] as an upper limit. We also use the TeV upper limit derived from H.E.S.S. with 14 hours of observational live time and assuming a photon index of 2.3 [@HESS:SNRpop:2018].
Assuming the Sedov phase, we derive the kinetic energy released by the supernova: $$\hspace{0.5cm}
\frac{E_{\rm{SN}}}{10^{51} \hspace{0.1cm} \rm{erg}} = R_{12.5}^5 \times \Big(\frac{n_{0}}{ \textrm{cm}^{-3}}\Big) \times t_{4}^{-2} = 0.5$$ where $R_{12.5}$ = $R$/(12.5 pc) and $t_{4}$ = $t$/(10,000 yrs) taking $R$ = 21 pc, $t$ = 16,500 yrs and $n_{0}$ = 0.1 cm$^{-3}$ for a distance of 4.1 kpc. We take the inputs from [@Temim:2013] but do not derive the same explosion energy. Here we use the common values of $\xi$ = 2.026 (for $\gamma$ = 5/3) and $\rho_{0}$ = 1.4$m_{\rm{H}}n_{0}$ in the usual Sedov equation: $R^{5}$ = $\xi$ ($E_{\rm{SN}}$/$\rho_{0}$) $t^{2}$.
The explosion energy and age depend on the distance, which is still uncertain. In addition, the density and temperature (the latter provides the shock speed estimate) were derived from a small region in the south of the SNR and it is not yet clear whether this region is representative of the rest of the SNR. (A Large Program with *XMM-Newton* on G326.3$-$1.8 is currently ongoing and will provide more constraints on the thermal emission across the SNR.) In this regard, the multi-wavelength modeling presented in this section is not to be viewed as a precise measurement of the properties of the accelerated particles but rather as showing that a simple self-consistent model can reproduce the observations.
We describe the electron population as a broken power law spectrum with spectral indexes $\Gamma_{\rm{e,1}}$/$\Gamma_{\rm{e,2}}$ with an exponential cut-off. The break at energy $E_{\rm{b}}$ is assumed to be due to cooling so we set $\Gamma_{\rm{e,2}}$ = $\Gamma_{\rm{e,1}}$ + 1 while the cut-off defines the maximum attainable energy of the particles $E_{\rm{max,e}}$. The proton spectrum is described as a power law with spectral index $\Gamma_{\rm{p}}$ with an exponential cut-off $E_{\rm{max,p}}$. In our models we consider by default the CMB as the only photon seed for IC scattering.\
### Leptonic scenario
We first investigate the leptonic scenario for which we vary the values of the magnetic field $B$, the total energy budget in electrons $W_{\rm{e}}$ and protons $W_{\rm{p}}$, and the break and maximum energy of the particles. Figure \[fig:Fit\_naima\_SNR\_Lepto\] (left) shows one of the combinations that simultaneously fits the radio and the $\gamma$-ray data. Since this solution is not unique, we report in Table \[tab:Phys\_param\] the range of permitted values of these parameters. For clarity, we fix the total energy in protons to W$_{\rm{p}}$ = 5 $\times$ 10$^{49}$ erg (corresponding to 10% of $E_{\rm{SN}}$) and we report the range of permitted values of the electron-proton ratio $K_{\rm{e-p}}$. Since the maximum energy of protons is always higher than that of electrons, which suffer synchrotron losses, we use the maximum value of $E_{\rm{max,e}}$ as a lower limit for $E_{\rm{max,p}}$.
To reproduce the radio spectral shape, we need a hard index for the electrons $\Gamma_{\rm{e,1}}$ = 1.8 (taking thus $\Gamma_{\rm{e,2}}$ = 2.8), while we keep $\Gamma_{\rm{p}}$ = 2 due to the lack of observational constraints. For a $B$ field between 10 and 20 $\mu$G, the $\gamma$-ray data can only be explained if the total energy in electrons reaches $W_{\rm{e}}$ = (2.5 – 7) $\times$ $10^{49}$ erg, which is clearly unreasonable since that requires a $K_{\rm{e-p}}$ between 0.5 and 1.4. If in order to reduce $W_{\rm{e}}$, we increase $B$ to higher values than expected for the compressed ISM, the $\gamma$-ray data cannot be fitted and the IC spectrum lies one order of magnitude below the data. Even if infrared and optical photon fields with energy density 0.26 eV cm$^{-3}$ each (the same as the CMB value, which is a reasonable estimate 100 pc below the Galactic plane) are added, an unrealistically large $W_{\rm{e}}$ is still required to fit the $\gamma$-ray data.
Another inconsistency of that model is that the values of $E_{\rm{b}}$, $E_{\rm{max,e}}$ and $E_{\rm{max,p}}$ reported in Table \[tab:Phys\_param\] are not consistent with the magnetic field. Following @Parizot:2006, we use the synchrotron loss time: $$\centering
\hspace{0.5cm} \tau_{\rm{sync}}=(1.25 \times 10^{3}) \times E_{\rm{TeV}}^{-1} B_{100}^{-2} \hspace{0.2cm} \textrm{yrs}$$ and the acceleration timescale: $$\hspace{0.5cm} t_{\rm{acc}}= 30.6 \times \frac{3r^2}{16(r-1)} \times k_{\rm{0}}(E) \times E_{\rm{TeV}}B_{100}^{-1}u_{\rm{sh,3}}^{-2} \hspace{0.2cm} \textrm{yrs}$$ with $r$ being the shock compression ratio, $k_0$ the ratio between the mean free path and the gyroradius, $B_{\rm{100}}$ and $u_{\rm{sh,3}}$ the magnetic field and the shock velocity in units of 100 $\mu$G and 1000 km s$^{-1}$, respectively. $k_{\rm{0}}$ $\geq$ 1 can be interpreted as the ratio of the total magnetic energy density to that in the turbulent field ($B_{\rm{tot}}^{2}/B_{\rm{turb}}^{2}$) and $k_0$ $\approx$ 1 has been found for young SNRs [@Uchiyama:2007]. For evolved systems, we expect the turbulent magnetic field to be smaller than the large-scale component (so that $k_0$ $>$ 1) and we adopt here $k_0$ = 10 for the highest-energy electrons. Taking $r$ = 4, we thus calculate $E_{\rm{b}}$ (equating $\tau_{\rm{sync}}$ = $t_{\rm{age}}$), $E_{\rm{max,e}}$ ($t_{\rm{acc}}$ = min$\{\tau_{\rm{sync}}, t_{\rm{age}}\}$) and $E_{\rm{max,p}}$ ($t_{\rm{acc}}$ = $t_{\rm{age}}$). For $B$ = 20 $\mu$G, we obtain $E_{\rm{b}}$ = 1.9 TeV, $E_{\rm{max,e}}$ = 2.3 TeV and $E_{\rm{max,p}}$ = 2.7 TeV. Figure \[fig:Fit\_naima\_SNR\_Lepto\] (right) shows the corresponding spectrum which implies an IC cut-off at too high energy and does not fit the $\gamma$-ray data.
One noteworthy aspect of this source concerns the difficulty to explain its very high radio flux [114 Jy at 1 GHz, @Dickel:2000]. High radio fluxes are also found in middle-aged SNRs interacting with molecular clouds, such as W44 [230 Jy at 1 GHz, @Castelletti07:W44] and IC 443 [160 Jy at 1 GHz, @Milne71:IC443], where the highly compressed gas enhances the synchrotron emission.
The particularly high total energy required in electrons to reproduce the SNR spectrum and the impossibility to fit the data with consistent values rule out a leptonic origin of the $\gamma$-ray emission and lead us to investigate the hadronic scenario.
### Hadronic scenario
[@Vandenbergh:1979] has reported H$\alpha$ emission in the northeast and southwest regions of the SNR (see Figure \[fig:Naima\_Hadro\_Halpha\], left panel) and [@Dennefeld:1980] obtained a spectrum indicating an \[S II\]/H$\alpha$ ratio characteristic of a radiative shock. This is evidence of the interaction of the shock with neutral material where some regions of the SNR are entering the radiative phase while other parts are freely expanding in the ISM. As a consequence, we suggest to model the SNR spectrum with two contributions:
- a radiative shock arising from the presence of clouds in the surroundings of the SNR
- a main shock with a velocity of $u_{\rm{sh}}$ = 500 km s$^{-1}$ and expanding in an ISM density of $n_{\rm{0}}$ = 0.1 cm$^{-3}$ [@Temim:2013]
Below we calculate the physical parameters associated with the radiative component. [@Uchiyama:2010] studied the non-thermal emission from crushed clouds in SNRs where re-acceleration of pre-existing cosmic rays can explain the observed GeV emission powered by hadronic interactions.\
Following this work, the strong shock driven into the clouds has a velocity of: $$\hspace{0.5cm}
u_{\rm{sh,cl}} = k \sqrt{\frac{n_0}{n_{\rm{0,cl}}}} \times u_{\rm{sh}}$$ where $k$ = 1.3 is adopted as in [@Uchiyama:2010], $n_{0}$ and $n_{\rm{0,cl}}$ being the upstream ISM and clouds density respectively. For the upstream magnetic field in the clouds, we have: $$\hspace{0.5cm}
B_{\rm{0,cl}} = b \sqrt{\frac{n_{\rm{0,cl}}}{\textrm{cm}^{-3}}} \hspace{0.1cm} \mu \textrm{G} \label{Bup}$$ where $b$ = $v_{\rm{A}}$/(1.84 km s$^{-1}$) with $v_{\rm{A}}$ being the Alfvén velocity, the mean value of which is thought to be roughly equal to the velocity dispersion observed in molecular clouds ($\sim$ 0.5 – 5 km s$^{-1}$), implying $b$ $\sim$ 0.3 – 3 [@HM:89]. As in [@Uchiyama:2010], we assume that the magnetic pressure in the cooled gas is equal to the shock ram pressure and we have: $$\hspace{0.5cm}
\frac{B_{\rm{m}}^2}{8 \pi} = k^2 n_{\rm{0}} \mu_{\rm{H}} u_{\rm{sh}}^2 \label{press_eq}$$ where $\mu_{\rm{H}}$ is the mass per hydrogen nucleus and $B_{\rm{m}}$ the downstream magnetic field in the cooled regions: $$\hspace{0.5cm}
B_{\rm{m}} = \sqrt{\frac{2}{3}} \times \Big(\frac{n_{\rm{m}}}{n_{\rm{0,cl}}}\Big) \times B_{\rm{0,cl}} \label{Bcool}$$ with $n_{\rm{m}}$ being the downstream density in the cooled regions. In this model, the compressed magnetic field is fixed to 158 $\mu$G by the pressure in the SNR (Eq \[press\_eq\]). This requires a large $W_{\rm{e}}$ in the clouds for the synchrotron emission to be consistent with the bright observed radio flux. We set $K_{\rm{e-p}}$ in the clouds to 0.03, which is large but still reasonable, since a lower $K_{\rm{e-p}}$ would result in an uncomfortably large W$_{\rm{p}}$. In any case, the cosmic-ray energy in the shocked clouds must be high. Thus, fitting simultaneously the $\gamma$-ray data and the radio data implies that the compressed density should be relatively low. From Eqs (\[Bup\], \[press\_eq\], \[Bcool\]), we have: $$\hspace{0.5cm}
n_{\rm{m}} = \sqrt{\frac{3}{\pi \mu_{\rm{H}}}} \times \frac{B_{\rm{m}}^2}{4 b \times u_{\rm{sh,cl}}}$$ We adopt here the highest reasonable values $b$ = 3 and $u_{\rm{sh,cl}}$ = 150 km s$^{-1}$ (above which the shock would have no time to become radiative). Thus the downstream density in the cooled regions is $n_{\rm{m}}$ = 88.3 cm$^{-3}$. Taking $n_{\rm{0}}$ = 0.1 cm$^{-3}$, $u_{\rm{sh}}$ = 500 km s$^{-1}$ [@Temim:2013], $u_{\rm{sh,cl}}$ = 150 km s$^{-1}$ and $b$ = 3, we obtain for the upstream density and magnetic field in the clouds $n_{\rm{0,cl}}$ = 1.88 cm$^{-3}$ and $B_{\rm{0,cl}}$ = 4.11 $\mu$G. This relatively low density is in agreement with the non-detection of CO lines close to this SNR. The densities encountered in G326.3$-$1.8 (cloud and intercloud medium) would then be very similar to the Cygnus Loop [@Raymond:1988].
The electrons accelerated in the clouds will rapidly cool due to the strong magnetic field in the dense regions for which we derive the break energy of the particles by equating $\tau_{\rm{sync}}$ = $t_{\rm{age}}$/2 (time since the clouds were shocked). At the shock front, the downstream magnetic field is $B_{\rm{d,cl}}$ = $\sqrt{11}B_{\rm{0,cl}}$ = 13.6 $\mu$G, assuming a randomly directed field, and we derive the corresponding maximum energy of the particles, using $k_{\rm{0}}$ = 10 and $u_{\rm{sh,cl}}$ = 150 km s$^{-1}$ when equating $t_{\rm{acc}}$ = min$\{\tau_{\rm{sync}},t_{\rm{age}}/2\}$. For particles trapped in the clouds, we thus find $E_{\rm{b}}$ = 15.2 GeV and $E_{\rm{max,e}}$ = $E_{\rm{max,p}}$ = 82.7 GeV.
Figure \[fig:Naima\_Hadro\_Halpha\] (right) shows the corresponding spectrum with the contributions from the main shock (solid lines) and the radiative shock (dashed lines). With such high magnetic field and density in the cooled regions, the radiative shock dominates the synchrotron and the $\gamma$-ray emission. Setting $K_{\rm{e-p}}$ = 0.03, the observed spectrum can be explained with $W_{\rm{p}}$ = 1.9 $\times$ $10^{49}$ erg (and thus $W_{\rm{e}}$ = 5.7 $\times$ $10^{47}$ erg), corresponding to 3.8% of $E_{\rm{SN}}$ transmitted to the re-accelerated protons in the clouds. To reproduce the radio spectral shape, we use harder indexes for the electrons at the radiative shock $\Gamma_{\rm{e,1}}$/$\Gamma_{\rm{e,2}}$ = 1.8/2.8 which is also observed in other radiative SNRs [@Ferrand:SNRcat][^5]. The $\gamma$-ray cut-off implied by $E_{\rm{max,p}}$ = 82.7 GeV fits the observed spectrum well. This is however largely coincidental. $k_{\rm{0}}$ is unconstrained, $t_{\rm{acc}}$ is unknown (we do not know when the clouds were shocked). [@Uchiyama:2010] predict an increase of the maximum energy by a factor of (($n_{\rm{m}}$/$n_{\rm{0}}$)/4)$^{1/3}$ = 2.27 due to adiabatic compression, which we did not enter into $E_{\rm{max,p}}$. The damping of Alfvén waves due to ion-neutral collisions also implies a break in the proton spectrum that we did not take into account because, with $B_{\rm{0,cl}}$ = 4.11 $\mu$G and $n_{\rm{0,cl}}$ = 1.88 cm$^{-3}$, it occurs around 100 GeV. Observationally, $E_{\rm{max,p}}$ must range between 30 and 100 GeV, which is in between other radiative SNRs such as W 44 or IC 443 [respectively 22 and 239 GeV, @Ackermann:13].
Since our model predicts that radiative shocks can explain the entire spectrum, we cannot assess observational constraints at the main shock. We take $B$ = 10 $\mu$G, implying $B_{\rm{ISM}}$ $\approx$ 3 $\mu$G (with $r$ = 4), to stay consistent with $B_{\rm{0,cl}}$ $>$ $B_{\rm{ISM}}$ but $B_{\rm{ISM}}$ could have been lower. We also use the typical 10% of $E_{\rm{SN}}$ going into protons but this and the value of $K_{\rm{e-p}}$ could also be reduced. For the particle spectra, we keep $\Gamma_{\rm{e,1}}$ = $\Gamma_{\rm{p}}$ = 2 and $\Gamma_{\rm{e,2}}$ = 3 since we have simple acceleration at the main shock and no observational constraints. The corresponding break and maximum energy are calculated following [@Parizot:2006] with $u_{\rm{sh}}$ = 500 km s$^{-1}$ and $k_{\rm{0}}$ = 10 as we did for the leptonic-dominated scenario. All the values used for the plot are reported in Table \[tab:Phys\_param\].
The entire SNR spectrum can thus be explained by the emission from radiative shocks. Although there is no clear correlation between the H$\alpha$ and the radio maps, this difference can be explained by the orientation of the magnetic field: where $B$ is perpendicular to the shock velocity, the synchrotron emission is largest (compression of the tangential component of the field) whereas optical emission should be enhanced when $B$ is parallel since the compression is no longer limited by the magnetic field. Quantitatively, the total energy required in the cosmic rays at the radiative shocks is large. Assuming 20% of the pressure in the radiative shocks is in the form of cosmic rays (the rest is mostly magnetic), it requires a surface covering factor close to 50% (consistent with the fact that we see little deviation from a uniform disk). This may be tested by deep H$\alpha$ imaging.\
PWN spectrum
------------
![\[fig:PWN\_Temim\_comparison\] Comparison of the $\gamma$-ray PWN spectra where the model derived in [@Temim:2013] is multiplied by a factor of 0.45 to fit our data.](PWN_Comparison_Temim13_with_HESS_UL_final.pdf)
We find that the largest fraction of the $\gamma$-ray emission comes from the SNR, presumably from the hadronic process. Nevertheless the PWN appears to contribute as well. We briefly and qualitatively discuss the impact of the PWN flux diminution on the physical parameters derived in [@Temim:2013] who assumed the entire $\gamma$-ray emission originates in the PWN, and based their analysis on the previous data release (Pass 7). Figure \[fig:PWN\_Temim\_comparison\] compares the two $\gamma$-ray spectra where the model of [@Temim:2013], who assumed a fully leptonic origin of the emission, is scaled to fit our data. The current flux corresponds to 45% of the previous one.\
If we approximate $W_{\rm{e}}$ $\approx$ $\int_{0}^{t_{\rm{age}}} \dot{E} dt$, $\dot{E}$ being the energy loss rate of the pulsar, we obtain: $$\hspace{3cm} W_{\rm{e}} \approx \dot{E_{0}}\frac{\tau_{\rm{0}} t_{\rm{age}}}{\tau_{\rm{0}} + t_{\rm{age}}}$$ where $\tau_0$ is the initial spin-down timescale of the pulsar and $\dot{E_{0}}$ the initial spin-down power. [@Temim:2013] derived $\tau_{0} \approx 2.1 \times 10^{4}$ years and $\dot{E_{0}}$ = 3 $\times$ 10$^{38}$ erg s$^{-1}$. We now require less than half of $W_{\rm{e}}$ leading to $\dot{E_{0}}$ = 1.35 $\times$ 10$^{38}$ erg s$^{-1}$ for the same age and initial spin-down time scale of the pulsar.\
In their 1-D model, the observed SNR radius is reached at an age of 19 kyrs for which they estimate the PWN magnetic field to be $B_{\rm{PWN}}$ = 34 $\mu$G. The decrease in $W_{\rm{e}}$ would thus also imply a higher magnetic field to still stay consistent with the radio flux of the PWN.
However, a more nuanced interpretation is required given the complexity of this object. This will require more investigations and detailed modeling which are beyond the scope of this paper. Note also that the PWN spectrum derived in this analysis is model-dependent when considering the assumption made for the SNR. In any case, its flux is reduced compared to previous studies since the SNR contributes most of the $\gamma$-ray emission.
Conclusions
===========
We perform an analysis from 300 MeV to 300 GeV of the composite SNR G326.3$-$1.8 with the $\textit{Fermi}$-LAT Pass 8 data. We take advantage of the new PSF3 event class by selecting the events with the best angular reconstruction to limit mixture between the SNR and the PWN contributions and also emission from the Galactic plane. Using the `pointlike` and the `gtlike` frameworks, we confirm that the emission is significantly extended (more than 13$\sigma$) between 300 MeV and 300 GeV. We perform an analysis in five energy bands which shows that the morphology evolves with energy and the size shrinks towards the radio PWN at high energies (E $>$ 3 GeV). We thus investigate a more detailed morphology using the radio map of the PWN as a starting point. We find that it is clearly not sufficient to describe the $\gamma$-ray data and that an additional extended component is needed. We then test different models for an additional contribution such as a uniform disk, the radio map of the remnant and its homogeneously filled radio template, called here the SNR mask. Using the maximum likelihood fitting procedure starting at 1 GeV, we find that the model with the SNR mask and the radio PWN reproduces the $\gamma$-ray emission best.
Modeling both $\gamma$-ray emissions by a power law from 300 MeV to 300 GeV, we obtain a spectral separation between the two components: a softer spectrum for the remnant ($\Gamma=2.17$ $\pm$ $0.06$) and a harder spectrum for the nebula ($\Gamma=1.79$ $\pm$ $0.12$). The corresponding SEDs also highlight their different contributions: the SNR dominates the low-energy part (300 MeV – 10 GeV) while the PWN protrudes at higher energies (E $>$ 10 GeV).
Concerning the PWN spectrum, we briefly discuss the impact of the flux diminution (about 55%) compared to previous studies that assumed that the entire $\gamma$-ray emission may come from the PWN.
The spectral modeling of the SNR emission disproves the leptonic scenario since it requires an unrealistic high energy budget in the electrons to fit the $\gamma$-ray data ($W_{\rm{e}}$ of several $10^{49}$ erg). As H$\alpha$ emission has been reported in this SNR, we suggest a spectral modeling where the main contribution arises from regions entering the radiative phase. The high magnetic field and density in the cooled regions lead to enhanced synchrotron and GeV emission that dominates the entire spectrum. The best-fit model involves a compressed magnetic field of 10 $\mu$G and 158 $\mu$G at the main and the radiative shock respectively. With 3.8% of the kinetic energy released by the supernova going into particles at the radiative shock, we find that an electron-proton ratio of $K_{\rm{e-p}}$ = 0.03 can adequately reproduce the observed spectrum. Although this ratio is slightly higher than one would expect, this is the most appropriate and consistent model we find that can simultaneously explain the high radio and $\gamma$-ray emissions from this SNR. In the future, CTA (Cherenkov Telescope Array) will give more insight into the properties of this source, providing better sensitivity above 30 GeV.
[^1]: The Science Tools package and related documentation are distributed by the *Fermi* Science Support Center at https://fermi.gsfc.nasa.gov/ssc
[^2]: From the *Fermi* Galactic Extended Source catalog
[^3]: Available at <https://fermi.gsfc.nasa.gov/ssc/data/access/lat/BackgroundModels.html>
[^4]: See <https://fermi.gsfc.nasa.gov/ssc/data/analysis/documentation/Pass8_edisp_usage.html>
[^5]: Radio spectral indexes of some radiative SNRs, such as W 44 or IC 443, can be found at <http://www.physics.umanitoba.ca/snr/SNRcat/>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Learning-assisted automated reasoning has recently gained popularity among the users of , , and . In this paper, we present an add-on to the proof assistant and an adaptation of the system that provides machine learning-based premise selection and automated reasoning also for . We efficiently record the dependencies and extract features from the theorem statements, which form a basis for premise selection. transforms the statements in the various TPTP-ATP proof formats, which are then processed by the ATPs.
We discuss the different evaluation settings: ATPs, accessible lemmas, and premise numbers. We measure the performance of on the standard library. The results are combined accordingly and compared with the experiments, showing a comparably high quality of predictions. The system directly benefits users by automatically finding proofs dependencies that can be reconstructed by .
bibliography:
- 'biblio.bib'
title: Premise Selection and External Provers for HOL4
---
HOL4; higher-order logic; automated reasoning; premise selection
Introduction
============
The proof assistant [@hol4] provides its users with a full ML programming environment in the LCF tradition. Its simple logical kernel and interactive interface allow safe and fast developments, while the built-in decision procedures can automatically establish many simple theorems, leaving only the harder goals to its users. However, manually proving theorems based on its simple rules is a tedious task. Therefore, general purpose automation has been developed internally, based on model elimination ( [@meson]), tableau ( [@blast]), or resolution ( [@metis]). Although essential to developers, the methods are so far not able to compete with the external ATPs [@eprover; @KovacsCAV13] optimized for fast proof search with many axioms present and continuously evaluated on the TPTP library [@TPTP] and updated with the most successful techniques. The TPTP (Thousands of Problems for Theorem Provers) is a library of test problems for automated theorem proving (ATP) systems. This standard enables convenient communication between different systems and researchers.
On the other hand, the system provides a functionality to search the database for theorems that match a user chosen pattern. The search is semi-automatic and the resulting lemmas are not necessarily helpful in proving the conjecture. An approach that combines the two: searching for relevant theorems and using automated reasoning methods to (pseudo-)minimize the set of premises necessary to solve the goal, forms the basis of “hammer” systems such as [@sledgehammer10] for , [@holyhammer] for or for [@DBLP:journals/corr/KaliszykU13b]. Furthermore, apart from syntactic similarity of a goal to known facts, the relevance of a fact can be learned by analyzing dependencies in previous proofs using machine learning techniques [@malarea], which leads to a significant increase in the power of such systems [@mash].
In this paper, we adapt the system to the system and test its performance on the standard library. The libraries of and are exported together with proof dependencies and theorem statement features; the predictors learn from the dependencies and the features to be able to produce lemmas relevant to a conjecture. Each problem is translated to the TPTP FOF format. When an ATP finds a proof, the necessary premises are extracted. They are read back to as proof advice and given to for reconstruction.
An adapted version of the resulting software is made available to the users of in interactive session, which can be used in newly developed theories. Given a conjecture, the function computes every step of the interaction loop and, if successful, returns the conjecture as a theorem:
(interactive call)
` load `“`holyHammer`”`; `\
` val it = (): unit `\
` holyhammer “1+1=2“;`\
` Relevant theorems: ALT_ZERO ONE TWO ADD1`\
` metis: r[+0+6]#`\
` val it = |- 1 + 1 = 2: thm`\
` `
The prover already benefits from export to SMT solvers such as [@WeberTjark], [@Z3reconstruction] and [@tgckckmn-paar14-accepted]. These methods perform best when solving problems from the supported theories of the SMT solver. Comparatively, is a general purpose tool as it relies on ATPs without theory reasoning and it can provide easily[^1] re-provable problem to .
The standard distribution has since long been equipped with proof recording kernels [@Wong95recordingand; @DBLP:conf/itp/KumarH12]. We first considered adapting these kernels for our aim. But as machine learning only needs the proof dependencies and the approach based on full proof recording is not efficient, we perform minimal modifications to the original kernel.
#### Contributions
We provide learning assisted automated reasoning for and evaluate its performance in comparison to that in . In order to do so, we :
- Export the data
Theorems, dependencies, and features are exported by a patched version of the kernel. It can record dependencies between theorems and keep track on how their conjunctions are handled along the proof. We export the standard libraries (58 types, 2305 constants, 11972 theorems) with respect to a strict name-space rule so that each object is uniquely identifiable, preserving if possible its original name.
- Reprove
We test the ability of a selection of external provers to reprove theorems from their dependencies.
- Define accessibility relations
We define and simulate different development environments, with different sets of accessible facts to prove a theorem.
- Experiment with predictors
Given a theorem and a accessibility relation, we use machine learning techniques to find relevant lemmas from the accessible sets. Next, we measure the quality of the predictions by running ATPs on the translated problems.
The rest of this paper is organized as follows. In Section \[sec:palibs\] we describe the export of the and data into a common format and the recording of dependencies in . In Section \[sec:setting\], we present the different parameters: ATPs, proving environments, accessible sets, features, and predictions. We select some of them for our experiments and justify our choice. In Section \[sec:experiments\] we present the results of the experiments, relate them to previous experiments and explain how this affects the users. Finally in Section \[sec:concl\] we conclude and present an outlook on the future work.
Sharing HOL data between , and {#sec:palibs}
===============================
In order to process and data in a uniform way in , we export objects from their respective theories, as well as dependencies between theorems into a common format. The export is available for any and development. We shortly describe the common format used for exporting both libraries and present in more detail our methods for efficiently recording objects (types, constants and theorems) and precise dependencies in . We will refer to [@holyhammer] for the details on recording objects and dependencies for formalizations.
and share a common logic (higher-order logic with implicit shallow polymorphism), however their implementations differ both in terms of the programming language used ( and respectively), data structures used to represent the terms and theorems (higher-order abstract syntax and de Bruijn indices respectively), and the exact inference rules provided by the kernel. As has been initially implemented in as an extension of , we need to export all the data and read it back into , replacing its type and constant tables. The format that we chose is based on the TPTP THF0 format [@tptpthf0] used by higher-order ATPs. Since formulas contains polymorphic constants which is not supported by the THF0 format, we will present an experimental extension of this format where the type arguments of polymorphic constants are given explicitly.
(experimental template)
` tt(name, role, formula) `\
` `
The field name is the object’s name. The field role is “ty” if the object is a constant or a type, and “ax” if the object is a theorem. The field formula is an experimental THF0 formula.
(Object export from to an experimental format)
- Type
` (list,1) `$\rightarrow$` tt(list, ty, $t > $t).`
- Constant
`(HD,“:’a list -> :’a“) `$\rightarrow$` `\
` tt(HD, ty, ![A:$t]: (list A > A).`\
`(CONS,“:’a -> :’a list -> :’a list“) `$\rightarrow$` `\
` tt(CONS ,ty, ![A:$t]: (A > list A > list A).`\
` `
- Theorem
`(HD,“`$\forall$` n:int`\
`t:list[int].`\
`HD (CONS n t) = n“) `$\rightarrow$` `\
` tt(HD0, ax, (![n:int, t:(list int)]:`\
` ((HD int) ((CONS int) n t) = n).`\
` `
In this example, `$t` is the type of all basic types.
All names of objects are prefixed by a namespace identifier, that allow identifying the prover and theory they have been defined in. For readability, the namespace prefixes have been omitted in all examples in this paper.
Creation of a theory
--------------------
In , types and constants can be created and deleted during the development of a theory. These objects are named at the moment they are created. A theorem is a value of type $thm$ and can be derived from a set of basic rules, which is an instance of a typed higher-order classical logic. To distinguish between important lemmas and theorems created by each small steps, the user can name and delete theorems (erase the name). Each named object still present at the end of the development is saved and thus can be called in future theories.
There are two ways in which an object can be lost in a theory: either it is deleted or overwritten. As proof dependencies for machine learning get more accurate when more intermediate steps are available, we decided to record all created objects, which results in the creation of slightly bigger theories. As the originally saved objects can be called from other theories, their names are preserved by our transformation. Each lost object whose given name conflicts with the name of a saved object of the same type is renamed.
#### Deleted objects
The possibility of deleting an object or even a theory is mainly here to hide internal steps or to make the theory look nicer. We chose to remove this possibility by canceling the effects of the deleting functions. This is the only user-visible feature that behaves differently in our dependency recording kernel.
#### Overwritten objects
An object may be overwritten in the development. As we prevent objects from being deleted, the likelihood of this happening is increased. This typically happens when a generalized version of a theorem is proved and is given the same name as the initial theorem. In the case of types and constants, the internal mechanism already renames overwritten objects. Conversely, theorems are really erased. To avoid dependencies to theorems that have been overwritten, we automatically rename the theorems that are about to be overwritten.
Recording dependencies
----------------------
Dependencies are an essential part of machine learning for theorem proving, as they provide the examples on which predictors can be trained. We focus on recording dependencies between named theorems, since they are directly accessible to a user. The time mark of our method slows down the application of any rules by a negligible amount.
Since the statements of 951 theorems are conjunctions, sometimes consisting of many toplevel conjuncts, we have refined our method to record dependencies between the toplevel conjuncts of named theorems.
(Dependencies between conjunctions)
`ADD_CLAUSES: 0 + m = m `$\wedge$` m + 0 = m `$\wedge$` `\
`SUC m + n = SUC (m + n) `$\wedge$` m + SUC n = SUC (m + n)`\
\
`ADD_ASSOC depends on:`\
` ADD_CLAUSES_c1: 0 + m = m`\
` ADD_CLAUSES_c3: SUC m + n = SUC (m + n)`\
` ...`
The conjunct identifiers of a named theorem `T` are noted `T_c1`, $\ldots$, `T_cN`.
In certain theorems, a toplevel universal quantifier shares a number of conjuncts. We will also split the conjunctions in such cases recursively. This type of theorem is less frequent in the standard library (203 theorems).
(Conjunctions under quantifier)
`MIN_0: `$\forall$` n. (MIN n 0 = 0) `$\wedge$` (MIN 0 n = 0)`
By splitting conjunctions we expect to make the dependencies used as training examples for machine learning more precise in two directions. First, even if a theorem is too hard to prove for the ATPs, some of its conjuncts might be provable. Second, if a theorem depends on a big conjunction, it typically depends only on some of its conjuncts. Even if the precise conjuncts are not clear from the human-proof, the reproving methods can often minimize the used conjuncts. Furthermore, reducing the number of conjuncts should ease the reconstruction.
Implementation of the recording {#sec:record}
-------------------------------
The type of theorems $thm$ includes a tag field in order to remember which oracles and axioms were necessary to prove a theorem. Each call to an oracle or axiom creates a theorem with the associated tag. When applying a rule, all oracles and axioms from the tag of the parents are respectively merged, and given to the conclusion of the rule. To record the dependencies, we added a third field to the tag, which consists of a dependency identifier and its dependencies.
\[exampletag\] (Modified tag type)
` type tag = ((dependency_id, dependencies), `\
` oracles, axioms)`\
` type thm = (tag, hypotheses, conclusion)`\
` `
Since the name of a theorem may change when it is overwritten, we create unmodifiable unique identifiers at the moment a theorem is named.
It consists of the name of the current theory and the number of previously named theorems in this theory. As a side effect, this enables us to know the order in which theorems are named which is compatible by construction with the pre-order given by the dependencies. Every variable of type $thm$ which is not named is given the identifier $unnamed$. Only identifiers of named theorems will appear in the dependencies.
We have implemented two versions of the dependency recording algorithm, one that tracks the dependencies between named theorems, other one tracking dependencies between their conjuncts. For the named theorems, the dependencies are a set of identified theorems used to prove the theorem. The recording is done by specifying how each rule creates the tag of the conclusion from the tag of its premises. The dependencies of the conclusion are the union of the dependencies of the unnamed premises with its named premises.
This is achieved by a simple modification of the `Tag.merge` function already applied to the tags of the premises in each rule.
When a theorem $\vdash A\,\wedge\,B$ is derived from the theorems $\vdash A$ and $\vdash B$, the previously described algorithm would make the dependencies of this theorem the union of the dependencies of the two. If later other theorems refer to it, they will get the union as their dependencies, even if only one conjunct contributes to the proof. In this subsection we define some heuristics that allow more precise tracking of dependencies of the conjuncts of the theorems.
In order to record the dependencies between the conjuncts, we do not record the conjuncts of named theorems, but only store their dependencies in the tags. The dependencies are represented as a tree, in which each leaf is a set of conjunct identifiers (identifier and the conjunct’s address). Each leaf of the tree represents the respective conjunct $c_i$ in the theorem tree and each conjunct identifier represents a conjunct of a named goal to prove $c_i$.
(An example of a theorem and its dependencies)
`Th0 (named theorem): A `$\wedge$` B`\
`Th1: C `$\wedge$` (D `$\wedge$` E) `\
` with dependency tree Tree([Th0],[Th0_c2])`\
\
`This encodes the fact that:`\
` C depends on Th0. `\
` D `$\wedge$` E depends Th0_c2 which is B.`\
` `
Dependencies are combined at each inference rule application and dependencies will contain only conjunct identifiers. If not specified, a premise will pass on its identifier if it is a named conjunct (conjunct of a named theorem) and its dependency tree otherwise. We call such trees passed dependencies. The idea is that the dependencies of a named conjunct should not transmit its dependencies to its children but itself. Indeed, we want to record the direct dependencies and not the transitive ones.
For rules that do not preserve the structure of conjunctions, we flatten the dependencies, i.e. we return a root tree containing the set of all (conjunct) identifiers in the passed dependencies. We additionally treat specially the rules used for the top level organization of conjunctions: `CONJ`, `CONJUNCT1`, `CONJUNCT2`, `GEN`, `SPEC`, and `SUBST`.
- `CONJ`: It returns a tree with two branches, consisting of the passed dependencies of its first and second premise.
- `CONJUNCT1` (`CONJUNCT2`): If its premise is named, then the conjunct is given a conjunct identifier. Otherwise, the first (second) branch of the dependency tree of its premise become the dependencies of its conclusion.
- `GEN` and `SPEC`: The tags are unchanged by the application of those rules as they do not change the structure of conjunctions. Although we have to be careful when using `SPEC` on named theorems as it may create unwanted conjunctions. These virtual conjunctions are not harmful as the right level of splitting is restored during the next phase.
(Creation of a virtual conjunction from a named theorem)
$\forall$` x.x `$\vdash$` `$\forall$` x.x `\
` SPEC [A `$\wedge$` B]`\
$\forall$` x.x `$\vdash$` A `$\wedge$` B `\
` CONJUNCT1`\
$\forall$` x.x `$\vdash$` A`
- `SUBST`: Its premises consist of a theorem, a list of substitution theorems of the form $(A=B)$ and a template that tells where each substitution should be applied. When `SUBST` preserves the structure of conjuncts, the set of all identifiers in the passed dependencies of the substitution theorems is distributed over each leaf of the tree given by the passed dependencies of the substituted theorems. When it is not the case the dependency should be flattened. Since the substitution of sub-terms below the top formula level does not affect the structure of conjunctions, it is sufficient (although not necessary) to check that no variables in the template is a predicate (is a boolean or returns a boolean).
The heuristics presented above try to preserve the dependencies associated with single conjuncts whenever possible. It is of course possible to find more advanced heuristics, that would give more precise human-proof dependencies. However, performing more advanced operations (even pattern matching) may slow down the proof system too much; so we decided to restrict to the above heuristics.
Before exporting the theorems, we split them by recursively distributing quantifiers and splitting conjunctions. This gives rise to conflicting degree of splitting, as for instance, a theorem with many conjunctions may have been used as a whole during a proof. Given a theorem and its dependency tree, each of its conjunctions is given the set of identifiers of its closest parent in this tree. Then, each of these identifiers is also split maximally. In case of a virtual conjunction (see the `SPEC` rule above), the corresponding node does not exist in the theorem tree, so we take the conjunct corresponding to its closest parent. Finally, for each conjunct, we obtain a set of dependencies by taking the union of the split identifiers.
(Recovering dependencies from the named theorem Th1)
`Th0 (named theorem): A `$\wedge$` B`\
`Th1 (named theorem): C `$\wedge$` (D `$\wedge$` E)`\
` with dependency tree Tree([Th0],[Th0_c1])`\
\
` Recovering dependencies of each conjunct`\
`Th1_c0: Th0 `\
`Th1_c1: Th0_c1 `\
`Th1_c2: Th0_c1`\
` Splitting the dependencies`\
`Th1_c0: Th0_c1 Th0_c2`\
`Th1_c1: Th0_c1 `\
`Th1_c2: Th0_c1`
Evaluation {#sec:setting}
==========
In this section we describe the setting used in the experiments: the ATPs, the transformation from HOL to the formats of the ATPs, the dependencies accessible in the different experiments, and the features used for machine learning.
ATPs and problem transformation
-------------------------------
supports the translation to the formats of various TPTP ATPs: FOF, TFF1, THF0, and two experimental TPTP extensions. In this paper we restrict ourselves to the first order monomorphic logic, as these ATPs have been the most powerful so far and integrating them in already poses an interesting challenge. The transformation that uses is heavily influenced by previous work by Paulson [@PaulsonS07] and Harrison [@meson]. It is described in detail in [@holyhammer], here we remind only the crucial points. Abstractions are removed by $\beta$-reduction followed by $\lambda$-lifting, predicates as arguments are removed by introducing existentially quantified variables and the apply functor is used to reduce all applications to first-order. By default uses the tagged polymorphic encoding [@BlanchetteBPS13]: a special tag taking two arguments is introduced, and applied to all variable instances and certain applications. The first argument is the first-order flattened representation of the type, with variables functioning as type variables and the second argument is the value itself.
Prover Version Premises
-------- --------- ----------
2.6 96
1.8 128
4.32 32
1.3 128
3.5 32
1.0 128
2.3 32
: \[tab:provers\]ATP provers, their versions and arguments
The initially used provers, their versions and default numbers of premises are presented in Table \[tab:provers\]. The experiments [@holyhammer] showed, that different provers perform best with different given numbers of premises. This is particularly visible for the ATP provers that already include the relevance filter SInE [@sine], therefore we preselect a number of predictions used with each prover. Similarly, the strategies that the ATP provers implement are often tailored for best performance on the TPTP library, for the annual CASC competition [@Sutcliffe14]. For ITP originating problems, especially for different strategies are often better, so we run it under the alternate scheduler `Epar` [@DBLP:journals/corr/abs-1301-2683].
Accessible facts
----------------
As has initially been designed for , it treats accessible facts in the same way as the accessibility relation defined there: any fact that is present in a theory loaded chronologically before the current one is available. In there are explicit theory dependencies, and as such a different accessibility relation is more natural. The facts present in the same theory before the current one, and all the facts in the theories that the current one depends on (possibly in a transitive way) are accessible. In this subsection we discuss the four different accessible sets of lemmas, which we will use to test the performance of on.
#### Exact dependencies (reproving)
They are the closest named ancestors of a theorem in the proof tree. It tests how much could reprove if it had perfect predictions. In this settings no relevance filtering is done, as the number of dependencies is small.
#### Transitive dependencies
They are all the named ancestors of a theorem in the proof tree. It simulates proving a theorem in a perfect environment, where all recorded theorems are a necessary step to prove the conjecture. This corresponds to a proof assistant library that has been refactored into little theories [@Farmer92littletheories].
#### Loaded theorems
All theorems present in the loaded theories are provided together with all the theorems previously built in the current theory. This is the setting used when proving theorems in , so it is the one we use in our interactive version presented and evaluated in Section \[sec:interactive\].
#### Linear order
For this experiment, we additionally recorded the order in which the theories were built, so that we could order all the theorems of the standard library in a similar way as theorems are ordered. All previously derived theorems are provided.
Features
--------
Machine learning algorithms typically use features to define the similarity of objects. In the large theory automated reasoning setting features need to be assigned to each theorem, based on the syntactic and semantic properties of the statement of the theorem and its attributes.
represents features by strings and characterizes theorems using lists of strings. Features originate from the names of the type constructors, type variables, names of constants and printed subterms present in the conclusion. An important notion is the normalization of the features: for subterms, their variables and type variables need to be normalized. Various scenarios for this can be considered:
- All variables are replaced by one common variable.
- Variables are replaced by their de Bruijn index numbers [@US+08].
- Variables are replaced by their (variable-normalized) types [@holyhammer].
The union of the features coming from the three above normalizations has been the most successful in the experiments, and it is used here as well.
Predictors
----------
In all our experiments we have used the modified k-NN algorithm [@ckju-pxtp13]. This algorithm produces the most precise results in the experiments for [@holyhammer]. Given a fixed number ($k$), the k-nearest neighbours learning algorithm finds $k$ premises that are closest to the conjecture, and uses their weighted dependencies to find the predicted relevance of all available facts. All the facts and the conjecture are interpreted as vectors in the $n$-dimensional feature space, where $n$ is the number of all features. The distance between a fact and the conjecture is computed using the Euclidean distance. In order to find the neighbours of the conjecture efficiently, we store an association list mapping features to theorems that have those features. This allows skipping the theorems that have no features in common with the conjecture completely.
Having found the neighbours, the relevance of each available fact is computed by summing the weights of the neighbours that use the fact as a dependency, counting each neighbour also as its own dependency
Experiments {#sec:experiments}
===========
In this section, we present the results of several experiments and discuss the quality of the advice system based on these results. The hardware used during the reproving and accessibility experiments is a 48-core server (AMD Opteron 6174 2.2 GHz. CPUs, 320 GB RAM, and 0.5 MB L2 cache per CPU). In these experiments, each ATPs is run on a single core for each problem with a time limit of 30 seconds. The reconstruction and interactive experiments were run on a laptop with a Intel Core processor (i5-3230M 4 x 2.60GHz with 3.6 GB RAM).
Reproving
---------
We first try to reprove all the 9434 theorems in the libraries with the dependencies extracted from the proofs. This number is lower than the number of exported theorems because definitions are discarded. Table \[Reproving\_unsplit\] presents the success rates for reproving using the dependencies recorded without splitting. In this experiment we also compare many provers and their versions. For [@Schulz:LPAR-2013], we also compare its different scheduling strategies [@DBLP:journals/corr/abs-1301-2683]. The results are used to choose the best versions or strategies for the selected few provers. Apart from the success rates, the unique number of problems is presented (proofs found by this ATP only), and [@Barrett:2011:CVC:2032305.2032319] seems to perform best in this respect. The translation used by default by is an incomplete one (it gives significantly better results than complete ones), so some of the problems are counter-satisfiable.
Prover Version Theorem(%) Unique CounterSat
-------- ---------- ------------ -------- ------------
`Epar` 3 44.45 3 0
`Epar` 1 44.15 9 0
`Epar` 2 43.95 9 0
`Epar` 0 43.52 2 0
1.3 42.71 44 0
4.32 41.96 8 5
4.40 41.65 1 6
1.8 41.37 14 0
2.6 41.10 14 0
1.8 38.34 6 0
4.40q 35.19 11 5
3.0 34.82 0 0
3.5 31.67 0 0
2.3 29.98 0 0
1.0 25.52 2 35
total 50.96 38
: Reproving experiment on the 9434 unsplit theorems of the standard libary[]{data-label="Reproving_unsplit"}
From this point on, experiments will be performed only with the best versions of three provers: , [@KovacsCAV13], and [@DeMoura:2008:ZES:1792734.1792766]. They have a high success rate combined with an easy way of retrieving the unsatisfiable core. The same ones have been used in the experiments for .
In Table \[Reproving\_split\], we try to reprove conjuncts of these theorems with the different recording methods described in Section \[sec:record\]. First, we notice that only benefits from the tracking of more accurate dependencies. More, removing the unnecessary conjuncts worsen the results of and . One reason is that and do well with large number of lemmas and although a conjunct was not used in the original proof it may well be useful to these provers.Suprisingly, the percentage of reproved facts did not increase compared to Table \[Reproving\_unsplit\], as this was the case for experiments. By looking closely at the data, we notice the presence of the `quantHeuristics` theory, where 85 theorems are divided into 1538 conjuncts. As the percentage of reproving in this theory is lower than the average (16%), the overall percentage gets smaller given the increased weight of this theory. Therefore, we have removed the quantHeuristic theory in the Basic\* and Optimized\* experiments for a fairer comparison with the previous experiments. Finally, if we compare the Optimized experiment with the similar reproving experiment on 14185 problems [@holyhammer], we notice that we can reprove three percent more theorems in . This is mostly due to a 10 percent increase in the performance of on problems.
Basic Optimized Basic\* Optimized\*
------- ------- ----------- --------- -------------
42.43 42.41 46.23 45.91
39.79 39.32 43.24 42.41
39.59 40.63 43.78 44.18
total 46.74 46.76 50.97 50.55
: Success rates of reproving (%) on the 13910 conjuncts of the standard library with different dependency tracking mechanism.[]{data-label="Reproving_split"}
In Table \[Reproving\_theories\] we have compared the success rates of reproving in different theories, as this may represent a relative difficulty of each theory and also the relative performance of each prover. We observe that performs best on the theories `measure` and `probability`, `list` and `finite_map`, whereas and have a higher success rate on the theories `arithmetic`, `real`, `complex` and `sort`. Overall, the high success rate in the `arithmetic` and `real` theories confirms that can already tackle this type of theorems. Nonetheless, it would still benefit from integrating more SMT-solvers’ functionalities on advanced theories based on `real` and `arithmetic`.
------- ------- ------- ------- -------
61.29 72.97 91.22 27.01
59.74 69.57 77.19 20.85
51.42 64.46 86.84 31.27
total 63.63 75.31 92.10 32.70
------- ------- ------- ------- -------
: Percentage (%) of reproved theorems in the theories `arithmetic`, `real`, `complex`, `measure`, `probability`, `list`, `sorting` and `finte_map`. []{data-label="Reproving_theories"}
------- ------- ------- ------- -------
42.16 23.56 34.54 33.07
37.34 21.96 32.72 27.16
54.21 25.62 25.45 43.70
total 55.42 26.77 40.00 45.27
------- ------- ------- ------- -------
: Percentage (%) of reproved theorems in the theories `arithmetic`, `real`, `complex`, `measure`, `probability`, `list`, `sorting` and `finte_map`. []{data-label="Reproving_theories"}
With different accessible sets
------------------------------
In Table \[accessible\_sets\] we compare the quality of the predictions in different proving environments. We recall that only the transitive dependencies, loaded theories and linear order settings are using predictions and that the number of these predictions is adapted to the ability of each provers. The exact dependencies setting (reproving), is copied from Table \[Reproving\_split\] for easier comparison.
ED TD LT LO
------- ------- ------- ------- -------
42.41 33.10 43.58 43.64
39.32 29.56 38.46 38.54
40.63 24.66 31.22 31.20
total 46.76 37.54 50.54 50.68
: Percentage (%) of proofs found using different accessible sets: exact dependencies (ED), transitive dependencies (TD), loaded theories (LT), and linear order (LO)[]{data-label="accessible_sets"}
We first notice the lower success rate in the transitive dependencies setting. There may be two justifications. First, the transitive dependencies provide a poor training set for the predictors; the set of samples is quite small and the available lemmas are all related to the conjecture. Second, it is very unlikely that a lemma in this set will be better than a lemma in the exact dependencies, so we cannot hope to perform better than in the reproving experiment.
We now focus on the loaded theories and linear order settings, which are the two scenarios that correspond to the regular usage of a “hammer” system in a development: given all the previously known facts try to prove the conjecture. The results are surprisingly better than in the reproving experiment. First, this indicates that the training data coming from a larger sample is better. Second, this shows that the library is dense and that closer dependencies than the exact one may be found by the predictors. It is quite common that large-theory automated reasoning techniques find alternate proofs. Third, if we look at each ATP separately, we see a one percent increase for , a one percent decrease for , and 9 percent decrease for . This correlates with the number of selected premises. Indeed, it is easy to see that if a prover performs well with a large number of selected premises, it has more chance to find the relevant lemmas. Finally, we see that each of the provers enhanced the results by solving different problems.
We can summarize the results by inferring that predictors combined with ATPs are most effective in large and dense developments.
The linear order experiments was also designed to make a valid comparison with a similar experiment where 39% of theorems were proved by combining 14 methods This number was later raised to 47% by improving the machine learning algorithm. Comparatively, the current 3 methods can prove 50% of the theorems. This may be since the machine learning methods have improved, since the ATPs are stronger now or even because the theories contain a more linear (less dense) development than the libraries, which makes it harder for automated reasoning techniques.
Reconstruction {#sec:reconstruction}
--------------
Until now all the ATP proved theorems could only be used as oracles inside . This defeats the main aim of the ITP which is to guarantee the soundness of the proofs. The provers that we use in the experiments can return the unsatisfiable core: a small set of premises used during the proof. The HOL representation of these facts can be given to in order to reprove the theorem with soundness guaranteed by its construction. We investigate reconstructing proofs found by on the loaded theories experiments (used in our interactive version of ). We found that could reprove, with a one second time limit, 95.6% of these theorems. This result is encouraging for two reasons: First, we have not shown the soundness of our transformations, and this shows that the found premises indeed lead to a valid proof in HOL. Second, the high reconstruction rate suggest that the system can be useful in practice.
Case study
----------
Finally, we present two sets of lemmas found by advised on the loaded libraries. We discuss the difference with the lemmas used in the original proof.
The theorem `EULER_FORMULE` states that any complex number can be represented as a combination of its norm and argument. In the human-written proof script ten theorems are provided to a rewriting tactic. The user is mostly hindered by the fact that she could not use the commutativity of multiplication as the tactic would not terminate. Free of these constraints, the advice system returns only three lemmas: the commutativity of multiplication, the polar representation `COMPLEX_TRIANGLE`, and the Euler’s formula `EXP_IMAGINARY`.
(In theory `complex`)
`Original proof:`\
`val EULER_FORMULE = store_thm(`“`EULER_FORMULE`”`,`\
` “!z:complex. modu z exp (i arg z) = z“,`\
` REWRITE_TAC[complex_exp, i, complex_scalar_rmul, `\
` RE, IM, REAL_MUL_LZERO, REAL_MUL_LID, EXP_0, `\
` COMPLEX_SCALAR_LMUL_ONE, COMPLEX_TRIANGLE]);`\
` `\
`Discovered lemmas:`\
`COMPLEX_SCALAR_MUL_COMM COMPLEX_TRIANGLE `\
`EXP_IMAGINARY`
The theorem `LCM_LEAST` states that any number below the least common multiple is not a common multiple. This seems trivial but actually the least common multiple ($lcm$) of two natural numbers is defined as their product divided by their greatest common divisor. The user has proved the contraposition which requires two calls. The discovered lemmas seem to indicate a similar proof, but it requires more lemmas, namely `FALSITY` and `IMP_F_EQ_F` as the false constant is considered as any other constant in and uses the combination of `LCM_COMM` and `NOT_LT_DIVIDES` instead of `DIVIDES_LE`.
Interactive version {#sec:interactive}
-------------------
In our previous experiments, all the different steps (export, learning/predictions, translation, ATPs) were performed separately, and simultaneously for all the theorems. Here, we compose all this steps to produce one step, that given a conjecture proves it, usable in any development in an interactive advice loop. It proceeds as follows: The conjecture is exported along with the currently loaded theories. Features for the theorems and the conjecture are computed, and dependencies are used for learning and selecting the theorems relevant to the conjecture. translates the problem to the formats of the ATPs and uses them to prove the resulting problems. If successful, the discovered unsatisfiable core, consisting of the theorems used in the ATP proof, is then read back to , returned as a proof advice, and replayed by . In the last experiment, we evaluate the time taken by each steps on two conjectures, which are not already proved in the libraries. The first tested goal $C_1$ is $gcd\
(gcd\ a\ a)\ (b + a) = (gcd\ b\ a)$, where $gcd\ n\ m$ is the greatest common divisor of $n$ and $m$. It can be automatically proved from three lemmas about $gcd$. The second goal is $C_2$ is $Im(i*i) = 0$, where $Im$ the imaginary part of a complex number. It can be automatically proved from 12 lemmas in the theories `real`, `transc` and `complex`.
In Table \[tab:Timer\], the time taken by the export and import phase linearly depends on the number of theorems in the loaded libraries (given in parenthesis), as expected by the knowledge of our data and the complexity analysis of our code.
The time shown in the fourth column (“Predict”) includes the time to extract features, to learn from the dependencies and to find 96 relevant theorems. The time needed for machine learning is relatively short. The time taken by shows that the second conjecture is harder. This is backed by the fact that we could not tell in advance what would be the necessary lemmas to prove this conjecture. The overall column presents the time between the interactive call and the display of advised lemmas. The low running times support the fact that our tool is fast enough for interactive use.
-------------- ------ ------ ------ ------ ------ --
$C_1$ (2224) 0.38 0.20 0.29 0.01 0.97
$C_2$ (4056) 0.67 0.43 0.59 1.58 3.42
-------------- ------ ------ ------ ------ ------ --
: Time (in seconds) taken by each step of the advice loop[]{data-label="tab:Timer"}
Conclusion {#sec:concl}
==========
In this paper we present an adaptation of the system for , which allows for general purpose learning-assisted automated reasoning. As uses machine learning for relevance filtering, we need to compute the dependencies, define the accessibility relation for theorems and adapt the feature extraction mechanism to . Further, as we export all the proof assistant data (types, constants, named theorems) to a common format, we define the namespaces to cover both and .
We have evaluated the resulting system on the standard library toplevel goals: for about 50% of them a sufficient set of dependencies can be found automatically. We compare the success rates depending on the accessibility relation and on the treatment of theorems whose statements are conjunctions. We provide a command that translates the current goal, runs premise selection and the ATP, and if a proof has been found, it returns a call needed to solve the goal. The resulting system is available at <https://github.com/barakeel/HOL>.
Future Work
-----------
The libraries of and are currently processed completely independently. We have however made sure that all data is exported in the same format, so that same concepts and theorems about them can be discovered automatically [@tgck-cicm14]. By combining the data, one might get goals in one system solved with the help of theorems from the other, which can then be turned into lemmas in the new system. A first challenge might be to define a combined accessibility relation in order to evaluate such a combined proof assistant library.
The format that we use for the interchange of and data is heavily influenced by the TPTP formats for monomorphic higher-order logic [@tptpthf0] and polymorphic first-order logic [@jasmincade2013]. It is however slightly different from that used by Sledgehammer’s `fullthf`. By completely standardizing the format, it would be possible to interchange problems between and .
In , theorems include the information about the theory they originate from and other attributes. It would be interesting to evaluate the impact of such additional attributes used as features for machine learning on the success rate of the proofs. Finally, most users call its web interface [@ckju-msc-hh-accepted], rather than locally install the necessary prover modifications, proof translation and the ATP provers. It would be natural to extend the web interface to support .
We would like to thank Josef Urban and Michael Färber for their comments on the previous version of this paper. This work has been supported by the Austrian Science Fund (FWF): P26201.
[^1]: reconstruction rate is typically above 90%
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this work we compute the families of classical Hamiltonians in two degrees of freedom in which the Normal Variational Equation around an invariant plane falls in Schrödinger type with polynomial or trigonometrical potential. We analyze the integrability of Normal Variational Equation in Liouvillian sense using the Kovacic’s algorithm. We compute all Galois groups of Schrödinger type equations with polynomial potential. We also introduce a method of algebrization that transforms equations with transcendental coefficients in equations with rational coefficients without changing essentially the Galoisian structure of the equation. We obtain Galoisian obstructions to existence of a rational first integral of the original Hamiltonian via Morales-Ramis theory.'
author:
- 'Primitivo Acosta-Humánez & David Blázquez-Sanz'
title: 'Non-integrability of some Hamiltonians with rational potentials'
---
[**Key Words:**]{} Picard-Vessiot Theory, Hamiltonian systems, Integrability, Kovacic’s Algorithm.
Introduction
============
In [@MorSimo], C. Simó and J. Morales-Ruiz find the complete list of two degrees of freedom classical hamiltonians, with an invariant plane $\Gamma=\{x_2=y_2=0\}$ with Normal Variational Equation of Lamé type along generic curves in $\Gamma$. In those computations they use systematically differential equations satisfied by coefficients of Lamé equation. It allow us to develop a method to find the families of hamiltonians with invariant plane $\Gamma$, with the property of having NVE satisfying certain conditions. It allow us to find the families of hamiltonians with generic NVE of type Mathieu, Shrödinger with polynomial potential of odd degree, quantum harmonic oscillator, and some other simpler examples.
Then, we analyze those families of linear differential equations in the context of Picard-Vessiot theory. For equations with rational coefficients, we use Kovacic’s Algorithm. In particular we give a complete descripion of Galois groups of Schrödinger type equations with polynomial potential (theorem \[polynint\]). For equations with transcendental coefficients, we develop a method of algebrization. We characterize equations that can be algebrized through change of variables of certain type (hamiltonian). We prove the following result.
*[**Algebrization algorithm.**]{} The differential equation $\ddot{y}=r(t)y$ is algebrizable through a hamiltonian change of variable $x=x(t)$ if and only if there exists $f,\alpha$ such that ${\alpha'\over\alpha},\quad {f\over \alpha}\in \mathbf{C}(x),\text{ where } f(x(t))=r(t),
\quad \alpha(x)=2(H-V(x))=\dot x^2.$ Furthermore, the algebraic form of the equation $\ddot{y}=r(t)y$ is $$y''+{1\over2}{\alpha'\over \alpha}y'-{f\over\alpha}y=0.$$*
Once we get the complete analysis of linearized equations, we apply a theorem of Morales-Ramis, obtaining the following results on the non integrability of those hamiltonians for generic values of the parameters.
[**Non-integrability Results.**]{} *Hamiltonians,*
1. $\frac{y_1^2+y_2^2}{2} +
\frac{\lambda_4}{(\lambda_2 + 2\lambda_3 x_1)^2} + \lambda_0 -
\lambda_1 x_2^2 - \lambda_2 x_1x_2^2 - \lambda_3 x_1^2 x_2^2 +
\beta(x_1,x_2)x_2^3,$ with $\lambda_3\neq 0$;
2. $\frac{y_1^2+y_2^2}{2} + \lambda_0 + Q(x_1)x_2^2 +
\beta(x_1,x_2)x_2^3,$ where $Q(x_1)$ is a non-constant polynomial;
3. $\frac{y_1^2+y_2^2}{2}+\mu_0 + \mu_1x_1 +\frac{\omega^2
x_1^2}{2}- \lambda_0x_2^2 - \lambda_1x_1x_2^2 + \beta(x_1,x_2)x_2^3,$ with $\omega \neq 0$.
4. $\frac{y_1^2+y_2^2}{2} + \mu_0 +
\frac{\mu_1}{(\lambda_1+2\lambda_2x_1)^2}+\frac{\lambda_1\omega^2 x_1}{8\lambda_2}
+ \frac{\omega^2 x_1^2}{8} + -\lambda_0 x_2^2 - \lambda_1x_1x_2^2 - \lambda_2x_1^2x_2^2 +
\beta(x_1,x_2)x_2^3,$ with $\omega\neq 0$, and $\lambda_2\neq 0$.
*where $\beta(x_1,x_2)$ is any analytical function around $\Gamma = \{x_2 = y_2 = 0\}$, do no admit any additional rational first integral.*
[**Corollary.**]{} *Every integrable (by rational functions) polynomial potentials with invariant plane $\Gamma=\{x_2=y_2=0\}$ can be written in the following form*
$$V=Q_1(x_1,x_2)x_2^3+\lambda_1x_2^2+\lambda_0,\quad \lambda_0,\lambda_1\in \mathbf{C}.$$
Note that for $\beta(x_1, x_2)$ polynomial and some values of the parameters we fall in the case of homogeneous polynomial potentials. Those cases had been analyzed deeply in [@Mac; @Nak].
Picard-Vessiot theory
---------------------
The Picard-Vessiot theory is the Galois theory of linear differential equations. In the classical Galois theory, the main object is a group of permutations of the roots, while in the Picard-Vessiot theory is a linear algebraic group. For polynomial equations we want a solution in terms of radicals. From classical Galois theory it is well known that this is possible if and only if the Galois group is solvable.
An analogous situation holds for linear homogeneous differential equations. For more details see [@VanSinger]. The following definition is true for matrices $n\times n$, but for simplicity we are restricting to matrices $2\times 2.$
An algebraic group of matrices $2\times 2$ is a subgroup $G\subset
GL(2,\mathbf{C})$, defined by algebraic equations in its matrix elements. That is, there exists a set of polynomials $$\{P_i(x_{11},x_{12},x_{21},x_{22})\}_{i\in I},$$ such that $$\left(\begin{array}{cc} x_{11} & x_{12} \\ x_{21} & x_{22}
\end{array}\right)\in G \quad\Leftrightarrow\quad \forall i\in I,
P_i(x_{11},x_{12},x_{21},x_{22}) = 0.$$
In this case we say that $G$ is an algebraic manifold provided of an structure of group. In the remainder of this paper we only work, as particular case, with linear differential equations of second order $$y''+ay'+by=0,\quad a,b\in \mathbf{C}(x).$$
Suppose that $y_1, y_2$ is a fundamental system of solutions of the differential equation. This means that $y_1, y_2$ are linearly independent over $\mathbf{C}$ and every solution is a linear combination of these two. Let $L = \mathbf{C}(x)\langle y_1, y_2
\rangle = \mathbf{C}(x)(y_1, y_2, y_1', y_2')$, that is the smallest differential field containing to $\mathbf{C}(x)$ and $\{y_{1},y_{2}\}.$
The group of all differential automorphisms of $L$ over $\mathbf{C}(x)$ is called the [*G*alois group]{} of $L$ over $\mathbf{C}(x)$ and denoted by $Gal(L/\mathbf{C}(x))$ or also by $Gal^L_{\mathbf{C}(x)}$. This means that for $\sigma\colon L\to
L$, $\sigma(a')=\sigma'(a)$ and $\forall a\in \mathbf{C}(x),$ $\sigma(a)=a$.
If $\sigma \in Gal(L/\mathbf{C}(x))$ then $\sigma y_1, \sigma y_2$ is another fundamental system of solutions of the linear differential equation. Hence there exists a matrix
$$A=
\begin{pmatrix}
a & b\\
c & d
\end{pmatrix}
\in GL(2,\mathbf{C}),$$ such that $$\sigma
\begin{pmatrix}
y_{1}\\
y_{2}
\end{pmatrix}
=
\begin{pmatrix}
\sigma y_{1}\\
\sigma y_{2}
\end{pmatrix}
=A
\begin{pmatrix}
y_{1}\\
y_{2}
\end{pmatrix}
.$$ This defines a faithful representation $Gal(L/\mathbf{C}(x))\to GL(2,\mathbf{C})$ and it is possible to consider $Gal(L/\mathbf{C}(x))$ as a subgroup of $GL(2,\mathbf{C})$. It depends on the choice of fundamental system $y_1$, $y_2$, but only up to conjugate.
One of the fundamental results of the Picard-Vessiot theory is the following theorem.
The Galois group $G=Gal(L/\mathbf{C}(x))$ is an algebraic subgroup of $GL(2,\mathbf{C})$.
Now we are interested in the differential equation $$\label{LDE}
\xi''=r\xi,\quad r\in \mathbf{C}(x).$$
We recall that equation can be obtained from the general second order linear differential equation $$y''+ay'+by=0,\quad a,b\in\mathbf{C}(x),$$ through the change of variable $$y=e^{-{1\over 2}\int_{}^{}a}\xi,\quad r={a^2\over 4}+{a'\over 2}-b.$$
On the other hand, through the change of variable $v=\xi'/\xi$ we get the associated Riccatti equation to equation $$\label{Riccatti}
v'=r-v^2,\quad v={\xi'\over \xi}.$$
For the differential equation , $G=Gal(G/\mathbf{C}(x))$ is an algebraic subgroup $SL(2,\mathbf{C})$.
Recall that an algebraic group $G$ has a unique connected normal algebraic subgroup $G^0$ of finite index. This means that the identity component $G^0$ is the biggest connected algebraic subgroup of $G$ containing the identity.
Let $F$ be a differential extension of $\mathbf{C}(x)$, and let be $\eta$ solution of the differential equation $$y''+ay'+by=0,\quad a,b\in F$$
1. $\eta$ is [*algebraic*]{} over $F$ if $\eta$ satisfies a polynomial equation with coefficients in $F$, i.e. $\eta$ is an algebraic function of one variable.
2. $\eta$ is [*primitive*]{} over $F$ if $\eta' \in F$, i.e. $\eta = \int f$ for some $f \in F$.
3. $\eta$ is [*exponential*]{} over $F$ if $\eta' /\eta \in F$, i.e. $\eta = e^{\int f}$ for some $f \in F$.
A solution $\eta$ of the previous differential equation is said to be [*Liouvillian*]{} over $F$ if there is a tower of differential fields $$F = F_0 \subset F_1 \subset ... \subset F_m = L,$$ with $\eta \in L$ and for each $i = 1,...,m$, $F_i =
F_{i-1}(\eta_i)$ with $\eta_i$ either algebraic, primitive, or exponential over $F_{i-1}$. In this case we say that the differential equation is integrable.
Thus a Liouvillian solution is built up using algebraic functions, integrals and exponentials. In the case $F=\mathbf{C}(x)$ we get, for instance logarithmic, trigonometric functions, but not special functions such that the Airy functions.
We recall that a group $G$ is called solvable if and only if there exists a chain of normal subgroups $$e=G_0\triangleleft G_1 \triangleleft \ldots \triangleleft G_n=G$$ such that the quotient $G_i/G_j$ is abelian for all $n\geq i\geq j\geq 0$.
\[LK\] The equation is integrable (has Liouvillian solutions) if and only if for $G=Gal(L/\mathbf{C}(x))$, the identity component $G^0$ is solvable.
Using the theorem \[LK\], Kovacic in 1986 introduced an algorithm to solve the differential and show that is integrable if and only if the solution of the equation is a rational function (case 1), is a root of polynomial of degree two (case 2) or is a root of polynomial of degree 4, 6, or 12 (case 3) (see Appendix A). Based in the Kovacic’s algorithm we have the following *key* result.
*The Galois group of differential equation with $r=Q_k(x)$ a polynomial of degree $k>0$ is a non-abelian connected group*.
For complete result, details and proof see Appendix A.
Morales-Ramis theory
--------------------
Morales-Ramis theory [@MorRamis1; @MorRamis2], see also [@MorMonograph], relates the integrability of hamiltonian systems to the integrability of linear differential equations. In this approach analyze the linearization (variational equations) of hamiltonian systems along some known particular solution. If the hamiltonian system is integrable, then we expect that the linearized equation has good properties in the sense of Picard-Vessiot theory. Exactly, for integrable hamiltonian systems, the Galois group of the linearized equation must be virtually abelian. It gives us the best non-integrability criterion known for hamiltonian systems. This approach has been extended to higher order variational equations in [@MorRamSimo].
### Integrability of hamiltonian systems
A symplectic manifold (real or complex), $M_{2n}$ is a $2n$-dimensional manifold, provided with a closed 2-form $\omega$. This closed $2$-form gives us a natural isomorphism between vector bundles, $\flat\colon TM\to
T^*M$. Given a function $H$ on $M$, there is an unique vector field $X_H$ such that, $$\flat(X_H) = dH$$ this is *the hamiltonian vector field of $H$*. Furthermore, it gives an structure of *Poisson algebra* in the ring of differentiable functions of $M_{2n}$ by defining: $$\{H,F\} = X_H F.$$ We say that $H$ and $F$ are *in involution* if $\{H,F\} = 0$. From our definition, it is obvious that $F$ is a *first integral* of $X_H$ if and only if $H$ and $F$ are in involution. In particular $H$ is always a first integral of $X_H$. Moreover, if $H$ and $F$ are in involution, then their flows commute.
The equations of the flow of $X_H$, in a system of canonical coordinates, $p_1,\ldots,p_n,q_1,\ldots,q_n$ (i.e. such that $\omega_2 = \sum_{i=1}^n p_i\wedge q_i$), are written $$\dot q = \frac{\partial H}{\partial p} \left( = \{H, q\}\right),
\quad \dot p = - \frac{\partial H}{\partial q} \left(= \{ H,
p\}\right),$$ and they are known as *Hamilton equations*.
Let $X_H$ be a hamiltonian defined on a real symplectic manifold $M_{2n}$. Assume that there are $n$ functionally independent first integrals $F_1,\ldots, F_n$ in involution. Let $M_a$ be a non-singular (i.e. $dF_1,\ldots, dF_n$ are independent on very point of $M_a$) level manifold, $$M_a = \{p\colon F_1(p)=a_1,\ldots,
F_n(p)=a_n\}.$$ Then,
1. If $M_a$ is compact and connected, then it is a torus $M_a\simeq \mathbf R^n/\mathbf Z^n$.
2. In a neighborhood of the torus $M_a$ there are functions $I_1,\ldots I_n,\phi_1,\ldots,\phi_n$ such that $$\omega_2 = \sum_{i=1}^n d I_i\wedge d\phi_i,$$ and $\{H,I_j\} = 0$ for $j = 1,\ldots, n$.
From now on, we will consider ${\mathbf C}^{2n}$ as a complex symplectic manifold. Lioville-Arnold theorem gives us a notion of integrability for Hamiltonian systems. A hamiltonian $H$ in ${\mathbf C}^{2n}$ is called *integrable in the sense of Liouville* if there exist $n$ independent first integrals of $X_H$ in involution. We will say that $H$ in integrable *by terms of rational functions* if we can find those first integrals between rational functions, or whatever.
### Variational equations
We want to relate integrability of hamiltonian systems with Picard-Vessiot theory. We deal with non-linear hamiltonian systems. But, given a hamiltonian $H$ in ${\mathbf C}^{2n}$, and $\Gamma$ an integral curve of $X_H$, we can consider the *first variational equation* (VE), as $$\mathcal{L}_{X_H}\xi=0,$$ in which the linear equation induced in the tangent bundle ($\xi$ represents a vector field supported on $\Gamma$).
Let $\Gamma$ be parameterized by $\gamma\colon t\mapsto (x(t),y(t))$ in such way that $$\frac{d x_i}{d t} = \frac{\partial H}{\partial y_i}, \quad
\frac{d y_i}{dt} = - \frac{\partial H}{\partial x_i}.$$
Then the VE along $\Gamma$ is the linear system, $$\left(\begin{array}{c} \dot \xi_i \\ \dot\eta_i, \end{array}\right) =
\left(\begin{array}{cc}
\frac{\partial^2 H}{\partial y_i\partial x_j}(\gamma(t)) & \frac{\partial^2 H}{\partial y_i \partial y_j}(\gamma(t)) \\
- \frac{\partial^2 H}{\partial x_i \partial x_j}(\gamma(t)) &
-\frac{\partial^2 H}{\partial x_i \partial y_j}(\gamma(t))
\end{array}\right)
\left(\begin{array}{c}\xi_i
\\ \eta_i \end{array}\right).$$
From the definition of Lie derivative, comes that $$\xi_i(t) = \frac{\partial H}{\partial y_i} (\gamma(t)),\quad
\eta_i(t) = - \frac{\partial H}{\partial x_i}(\gamma(t)),$$ is a solution of the VE. We can use it to reduce, using some generalization of D’Alambert method [@MorRamis1; @MorRamis2] (see also [@MorMonograph]) our equation, obtaining the *Normal Variational Equation* (NVE), which is a linear system of rank $2(n-1)$. In the case of $2$-degrees of freedom hamiltonian systems, those NVE can be seen as second order linear homogeneous differential equation.
### Non-integrability tools
Morales-Ramis theory, concerns several results that relate the existence of first integrals of $H$ with the Galois group of the variational equations. Along this paper we will use systematically the following one:
\[:MR\] Let $H$ be a Hamiltonian in ${\mathbf C}^{2n}$, and $\gamma$ a particular solution such that the [NVE]{} has irregular singularities at the points of $\gamma$ at infinity. Then, if $H$ is completely integrable by terms of rational functions, then the Identity component of Galois Group of the [NVE]{} is abelian.
Here, the field of coefficients of the NVE is the field meromorphic functions on $\gamma$.
Some results on linear differential equations
=============================================
Algebrization of Linear Differential Equations
----------------------------------------------
For some differential equations it is useful, if is possible, to replace the original differential equation over a compact Riemann surface, by a new differential equation over the Riemann sphere $\mathbb{P}^1$ (i.e., with rational coefficients) by a change of the independent variable. This equation on $\mathbb{P}^1$ is called the [**algebraic form**]{} or [**algebrization**]{} of the original equation. Kovacic’s algorithm can be applied over the algebraic form to solve the original equation. In a more general way we will consider the effect of a finite ramified covering on the Galois group of the original differential equation. In [@MorMonograph; @MorRamis1] the following theorem is given.
\[moramis\] Let $X$ be a (connected) Riemann surface. Let $f : X'\to X)$ be a finite ramified covering of $X$ by a Riemann surface $X'$. Let $\nabla$ be a meromorphic connection on $X$. We set $\nabla' =
f^*\nabla$. Then we have a natural injective homomorphism $Gal
(\nabla')\to Gal (\nabla)$ of differential Galois groups which induces an isomorphism between their Lie algebras.
We observe that, in terms of the differential Galois groups, this theorem means that the identity component of the differential Galois group is invariant by the covering. In other words, if the original differential equation over the Riemann surface $X'$ is transformed by a change of the independent variable in a differential equation over the Riemann surface $X$, then both equations have the same identity component of the differential Galois group.
\[pr1\] Let us consider the following equation, with coefficients in $\mathbf C(x)$: $$\label{eq1}
y''+a(x)y'+b(x)y=0,\quad y'={dy\over dx}$$ and $\mathbf C(x)\hookrightarrow L$ the corresponding Picard-Vessiot extension. Let $(K,\delta)$ be a differential field with $\mathbf C$ as field of constants. Let $\xi\in K$ be a non-constat. Then, by the change of variable $x = \xi$, is transformed in, $$\label{eq2}
\ddot{y}+\left(a(\xi)\dot{\xi}-{\ddot{\xi}\over
\dot{\xi}}\right)\dot{y}+b(\xi)(\dot{\xi})^2y=0, \quad \dot z = \delta z.$$ Let, $K_0\subset K$ be the smallest differential field containing $\xi$ and $\mathbf C$. Then is a differential equation with coefficients in $K_0$. Let $K_0\hookrightarrow L_0$ be the corresponding Picard-Vessiot extension. Assume that $$\mathbf C(x) \to K_0,\quad x\mapsto \xi$$ is an algebraic extension, then $$Gal(L_0 / K_0)^0 = Gal(L / \mathbf C(x))^0.$$
By the chain rule we have $$\frac{d}{dx} = \frac{1}{\dot\xi}\delta,$$ and, $$\frac{d^2}{dx^2}=\frac{1}{(\dot{\xi})^2}\delta^2 - \frac{\ddot\xi}{(\dot{\xi})^3}\delta,$$ now, changing $y'$, $y''$ in and making monic this equation we have $$\ddot{y}+\left(a(\xi)\dot{\xi}-{\ddot{\xi}\over
\dot{\xi}}\right)\dot{y}+b(\xi)(\dot{\xi})^2y=0.$$ In the same way we can obtain through .
If $K_0$ is an algebraic extension of $\mathbf C(x)$, then we can identify $K_0$ with the ring of meromorphic functions on a compact $X$ Riemann surface, and $\xi$, is a finite ramified covering of the Riemann sphere, $$X\xrightarrow{\xi} S^1.$$ If we consider $\nabla$ the meromorphic connection in $S^1$ induced by equation , then $\xi^*(\nabla)$, is the meromorphic connection induced by equation in $X$, and finally by theorem \[moramis\], we conclude.
Recently Manuel Bronstein in [@Bronstein] has implemented an algorithm to solve differential equation over $\mathbf{C}(t,e^{\int_{}^{}f(t)})$ without algebrize the equation. As immediate consequence of the proposition we have the following corollary.
Let be $f\in\mathbf{C}(t),$ the differential equation
$$\label{exp}
\ddot{y}-\left(f(t)+{\dot{f}(t)\over f(t)}
-f(t)e^{\int_{}^{}f(t)}a\left(e^{\int_{}^{}f(t)}\right)\right)\dot{y}+\left(f(t)\left(e^{\int_{}^{}f(t)}\right)\right)^2b\left(e^{2\int_{}^{}f(t)}\right)y=0,$$
by the change $x=e^{\int_{}^{}f(t)}$ is algebrizable and its algebraic form is given by $$y''+a(x)y'+b(x)y=0.$$
In this corollary, we have the following cases.
1. $f=n{h'\over h}$, for a rational function $h$, $n\in\mathbf{Z}_+$, we have the trivial case, both equations are over the Riemann sphere and they have the same differential field, so that the equation do not need be algebrized.
2. $f={1\over n}{h'\over h}$, for a rational function $h$, $n\in\mathbf{Z}^+$, the equation is defined over an algebraic extension of $\mathbf{C}(t)$ and so that this equation is not necessarily over the Riemann sphere.
3. $f\neq q{h'\over h}$, for any rational function $h$, $q\in\mathbf{Q}$, the equation is defined over an transcendental extension of $\mathbf{C}(t)$ and so that the this equation is not over de Riemann sphere.
In first and second case, we can apply proposition \[pr1\], taking $K_0 = \mathbf C(t,e^{\int_{}^{}f(t)})$, so that the identity component of the algebrized equation es conserved. The conservation of Galois group in the third case, requires further analysis, and will not be discussed here. We just remark that the Galois group corresponding to original equation is a subgroup of the Galois group of the algebrized equation.
\[def1\] The point $x=x_0$ is called regular singular point (or regular singularity) of the equation if and only if $x=x_0$ is not an ordinary point and
$$(x-x_0)a(x),\quad (x-x_0)^2b(x),$$ are both analytic in $x=x_0.$ If $x=x_0$ is not a regular singularity, then is an [**irregular singularity**]{}.
To study asymptotic behaviors (or the point $x=\infty$) in we can take $x(t)={1\over t}$ and to analyze in the behavior in $t=0$. That is, to study the behavior of the equation in $x=\infty$ we should study the behavior in $t=0$ of the equation $$\label{eq3}
\ddot{y}+\left({2\over t}-\left({1\over t^2}\right)a\left({1\over
t}\right)\right)\dot{y}+{1\over t^4}b\left({1\over t}\right)y=0.$$ In this way, by the definition \[def1\], we say that $x=\infty$ is a regular singularity of the equation if and only if $t=0$ is a regular singularity of the equation $\eqref{eq3}$.
To algebrize second order linear differential equations is easier when the term in $\dot y$ is absent and the change of variable is [*[hamiltonian]{}*]{}, that is, RLDE $\ddot{y}=r(t)y.$
\[def2\] A change of variable $x=x(t)$ is called hamiltonian if and only if $(x(t),\dot
x(t))$ is a solution curve of the autonomous hamiltonian system $$H=H(x,y)={y^2\over 2}+V(x).$$
Assume that we algebrize equation through a hamiltonian change of variables, $x = \xi(t)$. Then, $K_0 = \mathbf C(\xi, \dot\xi, \ldots)$, but, we have the algebraic relation, $$(\dot\xi) ^2 = 2h - 2V(\xi), \quad h = H(\xi,\dot \xi) \in \mathbf C,$$ so that $K_0 = \mathbf C(\xi,\dot \xi)$ is an algebraic extension of $\mathbf C(x)$. We can apply proposition \[pr1\], and then the identity component of the Galois group is conserved.
\[pr2\] The differential equation $$\ddot{y}=r(t)y$$ is algebrizable through a hamiltonian change of variable $x=x(t)$ if and only if there exists $f,\alpha$ such that $${\alpha'\over\alpha},\quad {f\over \alpha}\in \mathbf{C}(x),\text{ where } f(x(t))=r(t),\quad \alpha(x)=2(H-V(x))=\dot x^2.$$ Furthermore, the algebraic form of the equation $\ddot{y}=r(t)y$ is $$\label{eq4}
y''+{1\over2}{\alpha'\over \alpha}y'-{f\over\alpha}y=0.$$
Because $x=x(t)$ is a hamiltonian change of variable for the differential equation $\ddot{y}=r(t)y$ so that $\dot x=y$, $\dot
y=\ddot x=-V'(x)$ and there exists $f,$ $\alpha$ such that $\ddot{y}=f(x(t))y$ and $\dot x^2=2(H-V(x))=\alpha(x)$. By the proposition \[pr1\] we have $-f(x)=b(x)\dot{x}^2$ and $a(x)\dot
x-\ddot x/\dot x =0,$ therefore $a(x)=\ddot x/\dot x^2$ and $b(x)=-f(x)/\alpha(x)$. In this way $\ddot y=r(t)y$ is algebrizable if and only if $a(x), b(x) \in\mathbf{C}(x).$ As $2(H-V(x))=\alpha(x)$ then $\alpha'(x)=-2V'(x)=2\ddot x,$ and therefore $a(x)={1\over 2}{\alpha'(x)\over \alpha (x)}.$ In this way, we obtain the equation .
As consequence of the proposition \[pr2\] we have the following result.
\[cor1\] Let be $r(t)=g(x_1,\cdots, x_n)$, where $x_i=e^{\lambda_i t}$, $\lambda_i\in\mathbf{C}$. The differential equation $\ddot{y}=r(t)y$ is algebrizable if and only if $${\lambda_i\over \lambda_j}\in \mathbf{Q},\quad 1\leq i\neq j\leq n,\quad g\in \mathbf{C}(x).$$ Furthermore, we have $\lambda_i=c_i\lambda$, where $\lambda\in\mathbf{C}$ and $c_i\in \mathbf{Q}$ and one change of variable is $$x=e^{\lambda t\over q},\text{ where } c_i={p_i\over q_i}, \text{
}\gcd(p_i,q_i)=1 \text{ and } q=mcm(q_1,\cdots,q_n).$$
To algebrize the differential equation $\ddot y=r(t)y$ it should keep in mind the following steps.
Step 1
: Find a hamiltonian change of variable $x=x(t)$.
Step 2
: Find $f$ and $\alpha$ such that $r(t)=f(x(t))$ and $(\dot x (t))^2=\alpha(x(t))$.
Step 3
: Write $f(x)$ and $\alpha(x)$.
Step 4
: Verify if $f(x)/\alpha (x)\in\mathbf{C}(x)$ and $\alpha'(x)/\alpha
(x)\in\mathbf{C}(x)$ to see if the RLDE is algebrizable or not.
Step 5
: If the RLDE is algebrizable, write the algebraic form of the original equation such as follows $$y''+{1\over2}{\alpha'\over \alpha}y'-{f\over\alpha}y=0.$$
When we have algebrized second order linear differential equation we study its integrability and its Galois groups.
\[monic\] Let be the reduced linear differential equation (RLDE) $$\ddot y=\left(\sum_{k=0}^nc_kt^k\right)y,\quad c_k\in\mathbf{C}, \quad k=1,\cdots,n.$$ By the algebrization algorithm we can take $x=\mu t$, $\mu\in\mathbf{C}$, so that $$\dot x=\mu,\quad \alpha(x)=\mu^2,\quad \alpha'(x)=0 \text{ and } f(x)=\sum_{k=0}^nc_k\left({x\over \mu}\right)^k,\quad c_k,\mu\in\mathbf{C}.$$ Now, by the equation the new differential equation is $$y''=\left(\sum_{k=0}^n\left({c_k\over \mu^{k+2}}\right)x^k\right)y,\quad c_k,\mu\in\mathbf{C}.$$
In general, for $\mu=\sqrt[n+2]{c_{n}}$ we can obtain the equation
$$y''=\left(x^n+\sum_{k=0}^{n-1}d_kx^k\right)y,\quad d_k=\left({c_k\over \mu^{k+2}}\right),\quad k=0,\cdots, n-1.$$
Furthermore, we can observe, by definition \[def1\] and equation , that the point at $\infty$ is an irregular singularity for the differential equation with non-constant polynomial coefficients because using equation we can see that zero is not an ordinary point and neither is a regular singularity for the differential equation.
\[Mathiewobs\] The Extended Mathieu differential equation is $$\label{Mathew}
\ddot{y}=(a+b\sin t+c\cos t)y,$$
in particular, when $b=0$ or $c=0$ and $|a|+|b|\neq 0,$ we have the so called Mathieu equation. Applying the corollary \[cor1\] and the steps of the algorithm we have $x=e^{it}$, $\dot x=ix,$ therefore $$f(x)={(b+c)x^2+2ax+c-b\over 2x},\quad \alpha(x)=-x^2,\quad \alpha'(x)=-2x,$$
so that the algebraic form of the equation is $$\label{AlMat}
y''+{1\over x}y'+{(b+c)x^2+2ax+c-b\over 2x^3}y=0,$$ Making the change $x=1/z$ in the equation we obtain $$\label{matirr}
\ddot{\zeta}+\left({1\over
z}\right)\dot{\zeta}+\left({(c-b)z^2+2az+(b+c)\over
2z^3}\right)\zeta=0.$$ We can observe, by definition \[def1\] and equation , that $z=0$ is an irregular singularity for the equation and therefore $x=\infty$ is an irregular singularity for the equations and .
Now, we compute the Galois group and the integrability in equation . So that, the RLDE is given by $$\label{matred}
\xi''=-\left({(b+c)x^2+(2a+1)x+c-b\over 2x^3}\right)\xi.$$ Applying the Kovacic’s algorithm, see Appendix A, we can see that for $b\neq -c$ this equation falls in case 2: $(c_3,\infty_3),$ $E_0=\{3\}$, $E_{\infty}=\{1\}$ and so that $D=\emptyset$ because $m=-1\notin \mathbf{Z}_+$. In this way we have that the equation is not integrable, the Galois Group is the connected group $SL(2,\mathbf{C}),$ and finally, by the theorem \[moramis\] the identity component of the Galois group for equation is exactly $SL(2,\mathbf{C})$ that is not an abelian group. In the same way, for $b=-c$ we have the equations $$\ddot{y}=(a-be^{-it})y,\quad y''+{1\over x}y'+{2ax-2b\over 2x^3}y=0,$$ in which $\infty$ continues being an irregular singularity. Now, Its RLDE is given by $$\xi''=-\left({(2a+1)x-2b\over 2x^3}\right)\xi,$$ applying Kovacic’s algorithm we can see that this equation falls in case 2: $(c_3,\infty_2),$ $E_0=\{3\}$, $E_{\infty}=\{0,2,4\}$, so that $D=\emptyset$ because $1/2(e_{\infty}-e_0)\notin
\mathbf{Z}_+.$ This means that the Galois group continues being $SL(2,\mathbf{C}).$ Using this result, taking $\epsilon$ instead of $i$, we can say that in the case of [*[harmonic oscillator with exponential waste]{}*]{} $$\ddot{y}=(a+be^{-\epsilon t})y, \quad \epsilon>0,$$ the point at $\infty$ is an irregular singularity and the identity component of the Galois group is $SL(2,\mathbf{C})$.
Galois groups of Schrödinger equations with polynomial potential {#:sspoly}
----------------------------------------------------------------
Let be the reduced linear differential equation (RLDE) $$\ddot \xi=\left(\sum_{k=0}^nc_kt^k\right)\xi,\quad a_k\in\mathbf{C}.$$
By the remark \[monic\], through the change of variable $x=\mu t,$ $\mu=\sqrt[n+2]{c_k},$ this equation become in
$$\xi''=P_n(x)\xi,\quad P_n(x)\text{ is a monic polynomial of degree }n.$$
Kovacic in [@Kov] remarked that for $n=2k+1$ there are not liouvillian solutions and therefore the Galois group of the RLDE is $SL(2,\mathbf{C})$. Using the Kovacic’s algorithm (see Appendix), we can see that for $n=2k$ the equation falls in case 1, specifically in $c_0$ (because has not poles) and $\infty_3$ (because $\circ
r_{\infty}=-2k$), that is $\{c_0,\infty_3\}$.
\[cosq\] Every monic polynomial of degree even can be written in one only way completing squares, that is $$Q_{2n}(x)=x^{2n}+\sum_{k=0}^{2n-1}q_kx^k=\left(x^n+\sum_{k=0}^{n-1}a_kx^k\right)^2+\sum_{k=0}^{n-1}b_kx^k.$$
Firstly, we can see that $$\left(x^n+\sum_{k=0}^{n-1}a_kx^k\right)^2=x^{2n}+2\sum_{k=0}^{n-1}a_kx^{n+k}+\sum_{k=0}^{n-1}\sum_{j=0}^{n-1}a_ka_jx^{k+j},$$ so that by indeterminate coefficients we have
$$a_{n-1}={q_{2n-1}\over 2},\quad a_{n-2}={q_{2n-2}-a^2_{n-1}\over 2},\quad a_{n-3}={q_{2n-3}-2a_{n-1}a_{n-2}\over 2},\cdots,$$
$$a_0={q_n-2a_1a_{n-1}-2a_2a_{n-2}-\cdots\over 2},\quad
b_0=q_0-a_0^2,\quad b_1=q_1-2a_0a_1,\quad \cdots,$$ $$b_{n-1}=q_{n-1}-2a_0a_{n-1}-2a_1a_{n-2}-\cdots,$$ therefore, we has proven the lemma.
By case 1, $\{c_0,\infty_3\}$, of the kovacic’s algorithm, remark \[rkov2\] \[I5\] (See Appendix), remark \[monic\] and by lemma \[cosq\] we have proven the following theorem.
\[polynint\] Let us consider the equation, $$\ddot \xi = Q(x)\xi,$$ with $Q(x)$ a polynomial of degree $k>0$. Then, it falls in one of the following cases:
1. $k=2n$ is even, $\pm b_{n-1}-n=2m,$ $m\in \mathbf{Z}_+$, and there exist a monic polynomial $P_m$ of degree $m$ satisfying $$P_m'' + 2\left(x^n+\sum_{k=0}^{n-1}a_kx^k\right)P_m' +
\left(nx^{n-1}+\sum_{k=0}^{n-2}(k+1)a_{k+1}x^{k} +
\sum_{k=0}^{n-1}b_kx^k\right)P_m = 0,$$ the solutions are given by $$\begin{aligned}
\xi_1&=&P_me^{{x^{n+1}\over n+1}+\sum_{k=0}^{n-1}{a_kx^{k+1}\over k+1}}, \\
\quad \xi_2&=&P_me^{{x^{n+1}\over
n+1}+\sum_{k=0}^{n-1}{a_kx^{k+1}\over k+1}}
\int_{}^{}{dx\over P_m^2e^{2\left({{x^{n+1}\over n+1}+\sum_{k=0}^{n-1}{a_kx^{k+1}\over k+1}}\right)}}.\end{aligned}$$ and the Galois group is $\mathbf{C}^*\ltimes\mathbf{C}$ (non-abelian, resoluble, connected group).
2. The equation has not liouvillian solutions and its Galois group is $SL(2,\mathbf{C})$.
\[:rkm2\]
consider the case where $r$ is a polynomial of degree two: $$\ddot y = \left(At^2 + Bt+C\right)y.$$ There are no poles and the order at $\infty$, $\circ r_{\infty}$, is $-2$, so we need to follow case 1 in $\{c_0,\infty_3\}$ of the algorithm. Now, by remark \[monic\] and by lemma \[cosq\] we have $$y''=\left((x+a)^2+b\right)y.$$ We find that $$\begin{aligned}
[\sqrt{r}]_{\infty}&= x + a \\
\alpha_\infty^\pm &= {1\over 2}\left(\pm b - 1\right)\\
m &= \alpha_\infty^+\quad\text{or}\quad\alpha_\infty^-.\end{aligned}$$ If $b$ is not an odd integer then $m$ cannot be an integer so case 1 cannot hold so the RLDE in $x$ has no Liouvillian solutions. If $b$ is an odd integer than we can complete steps 2 and 3 and actually (only) find a solution it which is $$y=P_me^{{x^2\over 2}+ax}.$$
In particular, the quantum harmonic oscillator $$y''=(x^2+\lambda)y$$ is integrable when $\lambda$ is an odd integer and the only one solution obtained by means of kovacic’s algorithm is given by $$y=H_me^{x^2\over 2},$$ where $H_m$ denotes the classical Hermite’s polynomials.
For another approach to this problem see Vidunas [@Vidunas] and Zoladek in [@Zoladek].
Determining families of Hamiltonians with specific NVE
======================================================
Let us consider a two degrees of freedom classical hamiltonian, $$H = \frac{y_1^2+y_2^2}{2} + V(x_1,x_2).$$ $V$ is the *potential function*, and it is assumed to be analytical in some open subset of ${\mathbf C}^2$. The evolution of the system is determined by Hamilton equations:
$$\dot x_1 = y_1,\quad \dot x_2 = y_2, \quad \dot y_1 =
-\frac{\partial V}{\partial x_1},\quad \dot y_2 = -\frac{\partial
V}{\partial x_2}.$$
Let us assume that the plane $\Gamma = \{x_2=0, y_2 = 0\}$ is an invariant manifold of the hamiltonian. We keep in mind that the family of integral curves lying on $\Gamma$ is parameterized by the energy $h = H|_\Gamma$, but we do not need to use it explicitly. We are interested in studying the linear approximation of the system near $\Gamma$. Since $\Gamma$ is an invariant manifold, we have $$\left.\frac{\partial V}{\partial x_2}\right|_\Gamma = 0,$$ so that the Normal Variational Equation for a particular solution $$t\mapsto\gamma(t) =(x_1(t), y_1 = \dot x_1(t), x_2 = 0, y_2 = 0),$$ is written, $$\dot \xi = \eta,\quad \dot \eta = -\left[\frac{\partial^2 V}{\partial x_2^2}(x_1(t),0)\right] \xi.$$
Let us define, $$\phi(x_1) = V(x_1, 0),\quad \alpha(x_1) =
- \frac{\partial^2 V}{\partial x_2^2}(x_1,0),$$ and then we write the second order Taylor series in $x_2$ for $V$, obtaining the following expression for $H$ $$\label{:Hgeneral}
H = \frac{y_1^2+y_2^2}{2} + \phi(x_1) - \alpha(x_1)\frac{x_2^2}{2}
+ \beta(x_1,x_2)x_2^3,$$ which is the *general form of a classical a Hamiltonian, with invariant plane $\Gamma$*.The NVE associated to any integral curve lying on $\Gamma$ is, $$\label{:NVE}
\ddot \xi = \alpha(x_1(t)) \xi.$$
General Method
--------------
We are interested in compute Hamiltonians of the family (\[:Hgeneral\]), such that its NVE (\[:NVE\]) belongs to a specific family of Linear Differential Equations. Then we can apply our results about the integrability of this LDE, and Morales-Ramis theorem to obtain information about the non-integrability of such Hamiltonians.
From now on, we will write $a(t) = \alpha(x_1(t))$, for a generic curve $\gamma$ lying on $\Gamma$, parameterized by $t$. Then, the NVE is written $$\label{:NVEa}
\ddot \xi = a(t)\xi.$$
[**Problem.** ]{}*Assume that $a(t)$ is a root of a $Q(a, \dot a, \ddot a, \ldots ) \in {\mathbf C}[a, \dot a,
\ddot a,\ldots]$. We want to compute all hamiltonians in satisfying such a condition.*
In this section we give a method to compute, for any given $Q(a,\dot a,\ldots)$, the family of classical hamiltonians with invariant plane $\Gamma$ such that, for any integral curve lying on $\Gamma$, the coefficient $a(t)$ of the NVE satisfies, $$\label{:Q}
Q(a,\dot a,\ddot a,\ldots) =0,$$ by solving certain differential equations. This method lies under the calculus done by J. Morales and C. Simó in [@MorSimo].
We should notice that, for a generic integral curve $\gamma(t) = (x_1(t),y_1 = \dot x_1(t))$ lying on $\Gamma$, equation (\[:NVEa\]) depends only of the values of functions $\alpha$, and $\phi$. It depends of $\alpha(x_1)$, since $a(t) = \alpha(x_1(t))$. We observe that the curve $\gamma(t)$ is a solution of the restricted Hamiltonian, $$\label{:restrictedH}
h = \frac{y_1^2}{2} + \phi(x_1)$$ whose associated Hamiltonian vector field is, $$\label{:Xh}
X_h = y_1\frac{\partial}{\partial x_1} - \frac{d\phi}{d x_1}\frac{\partial}{\partial y_1},$$ thus $x_1(t)$ is a solution of the differential equation, $\ddot x_1 = -\frac{d\phi}{d x_1},$ and then, the relation of $x_1(t)$ is given by $\phi$.
Since $\gamma(t)$ is an integral curve of $X_h$, for any function $f(x_1,y_1)$ defined in $\Gamma$ we have $$\frac{d}{d t}\gamma^*(f) = \gamma^*(X_h f),$$ where $\gamma^*$ denote the usual pullback of functions. Then, using $a(t) = \gamma^*(\alpha)$, we have for each $k\geq 0$, $$\frac{d^k a}{d t^k} = \gamma^*(X_h^k\alpha),$$ so that, $$Q(a,\dot a, \ddot a,\ldots ) = Q(\gamma^*(\alpha),\gamma^*(X_h\alpha),\gamma^*(X_h^2\alpha),\ldots).$$
There is an integral curve of the Hamiltonian passing through each point of $\Gamma$, so that we have proven the following.
\[:prop2\] Let $H$ be a Hamiltonian of the family [(\[:Hgeneral\])]{}, and $Q(a,\dot a,\ddot a,\ldots)$ a differential polynomial with constants coefficients. Then, for each integral curve lying on $\Gamma$, the coefficient $a(t)$ of the NVE [(\[:NVEa\])]{} verifies $Q(a,\dot a, \ddot a, \ldots,) = 0$, if and only if the function $$\hat Q(x_1,y_1) = Q(\alpha, X_h\alpha, X_h^2 \alpha, \ldots),$$ vanish on $\Gamma$.
In fact the NVE of a integral curve depends on the parameterization. Our criterion does not depend on any choice of parameterization of the integral curves. This is simple, the NVE corresponding to different parameterizations of the same integral curve are related by a translation of time $t$. We need just observe that a polynomial $Q(a,\dot a, \ddot a,\ldots)$ with constant coefficients is invariant of the group by translations of time. Then if the coefficient $a(t)$ of the NVE (\[:NVEa\]) for certain parameterization of an integral curve $\gamma(t)$ satisfied $\{Q=0\}$, then it is also satisfied for any other right parameterization of the curve.
Next, we will see that $\hat Q(x_1,y_1)$ is a polynomial in $y_1$ and its coefficients are differential polynomials in $\alpha, \phi$. If we write down the expressions for successive Lie derivatives of $\alpha$, we obtain $$\label{:armonico}
X_h \alpha = y_1\frac{d \alpha}{d x_1},$$
$$\label{:airy}
X_h^2 \alpha = y_1^2\frac{d^2\alpha}{d x_1^2} - \frac{d \phi}{d x_1}\frac{d \alpha}{ d x_i}$$
$$\label{:qarmonic}
X_h^3\alpha = y_1^3\frac{d^3\alpha}{dx_1^3} - y_1\left(
\frac{d}{dx_1}\left(\frac{d\phi}{dx_1}\frac{d\alpha}{dx_1}\right)
+ 2\frac{d\phi}{dx_1}\frac{d^2\alpha}{dx_1^2}\right)$$
$$X_h^4\alpha = y_1^4\frac{d^4\alpha}{dx_1^4}-y_1^2\left(\frac{d}{dx_1}\left(
\frac{d}{dx_1}\left(\frac{d\phi}{dx_1}\frac{d\alpha}{dx_1}\right) + 2\frac{d\phi}{dx_1}\frac{d^2\alpha}{dx_1^2}\right)
+3\frac{d^3\alpha}{dx_1^3}\frac{d\phi}{dx_1}\right) +$$
$$+\left(
\frac{d}{dx_1}\left(\frac{d\phi}{dx_1}\frac{d\alpha}{dx_1}\right) + 2\frac{d\phi}{dx_1}\frac{d^2\alpha}{dx_1^2}\right)
\frac{d\phi}{dx_1}.$$
In general form we have, $$\label{:induction}
X_h^{n+1}\alpha = y_1\frac{\partial X_h^{n}\alpha}{d x_1} - \frac{d\phi}{d x_1}\frac{\partial X_h^{n}}{\partial y_1},$$ it inductively follows that they all are polynomial in $y_1$ with coefficients differential polynomials in $\alpha,\phi$. If we write it down explicitly, $$\label{:Ex}
X_h^n\alpha = \sum_{n\geq k \geq 0} E_{n,k}(\alpha,\phi)y_1^k$$ we can see that the coefficients $E_{n,k}(\alpha,\phi) \in
{\mathbf C}\left[\alpha,\phi,\frac{d^r\alpha}{d
x_1^r},\frac{d^s\phi}{dx_1^s}\right]$, satisfies the following recurrence law, $$\label{:rlaw}
E_{n+1,k}(\alpha,\phi) = \frac{d}{d x_1} E_{n,k-1}(\alpha,\phi) - (k+1)E_{n,k+1}(\alpha,\phi)\frac{d\phi}{d x_1}$$ with initial conditions, $$\label{:condition}
E_{1,1}(\alpha,\phi) = \frac{d \alpha}{d x_1},\quad E_{1,k}(\alpha,\phi) = 0 \,\,\,\forall k\neq 1.$$
\[:remark\] The recurrence law (\[:rlaw\]) and initial conditions (\[:condition\]) determine the coefficients $E_{n,k}(\alpha,\phi)$. We can compute the value of some of them easily:
- $E_{n,n}(\alpha,\phi) = \frac{d^n\alpha}{d x_1^n}$ for all $n \geq 1$.
- $E_{n,k}(\alpha,\phi)= 0$ if $n-k$ is odd, or $k<0$, or $k>n$.
Some Examples
-------------
Here we compute families of hamiltonians (\[:Hgeneral\]) such give rise to specific NVE. Although, in order to this computations, we need to solve Polynomial Differential Equations, we will see that we can deal with this in a branch of cases. Particularly, when $Q$ is a Differential Linear Operator, we will obtain equations that involve products of few Linear Differential Operator.
*Harmonic oscillator* equation is $$\label{:eqarmonico}
\ddot \xi = c_0\xi,$$ with $c_0$ constant. Then, a hamiltonian of type (\[:Hgeneral\]) gives such NVE if $\dot a = 0$. Looking at formula (\[:armonico\]), it follows that $\frac{d \alpha}{d x_1}=0$, so that $\alpha$ is a constant. We conclude that the general form of a Hamiltonian (\[:Hgeneral\]) which give rise to NVE of the type (\[:eqarmonico\]) is, $$H = \frac{y_1^2+y_2^2}{2} + \phi(x_1) + \lambda_0x_2^2 + \beta(x_1,x_2)x_2^3,$$ being $\lambda_0$ a constant, and $\phi,\beta$ arbitrary analytical functions.
In [@Audin], M. Audin notice that the hamiltonian, $$\frac{y_1^2+y_2^2}{2} + x_1x_2^2$$ gives an example of a simple non-integrable classical hamiltonian, since its NVE along any integral curve in $\Gamma$ is an *Airy equation*. Here we compute the family of classical hamiltonians that have NVE of type Airy for integral curves lying on $\Gamma$. General form of Airy equation is $$\label{:eqairy}
\ddot\xi = (c_0 + c_1 t)\xi$$ with $c_0,c_1\neq 0$ two constants. If follows that a hamiltonian gives rise to NVE of this type if $\ddot a = 0$, and $\dot a \neq
0$. The equation $\ddot a = 0$ gives, by proposition \[:prop2\] as we see in formula (\[:airy\]), the following system: $$\frac{d^2\alpha}{d x_1^2} = 0,\quad \frac{d\phi}{d x_1}\frac{d \alpha}{d x_1} = 0.$$ It split in two independent systems, $$\frac{d\alpha}{dx_1} = 0,\quad \left\{\begin{array}{c}
\frac{d^2\alpha}{dx_1^2} = 0 \\
\frac{d\phi}{dx_1} = 0
\end{array}\right.$$ Solutions of the first one fall into the previous case of *harmonic oscillator*. Then, taking the general solution of the second system, we conclude that the general form of a classical hamiltonian of type (\[:Hgeneral\]) with Airy NVE is: $$H = \frac{y_1^2+y_2^2}{2}+\lambda_0 + \lambda_1 x_2^2 + \lambda_2 x_1x_2^2 + \beta(x_1,x_2)x_2^3,$$ with $\lambda_2 \neq 0$.
### NVE quantum harmonic oscillator
Let us consider now equations with $\frac{d^3 a}{dt^3}=0$, and $\frac{d^2 a}{dt^2}\neq 0$, it is $$\label{:quadratic}
\ddot \xi = (c_0 + c_1 t + c_2 t^2)\xi$$ with $c_2\neq 0$. Those equation can be reduced to a *quantum harmonic oscillator equation* by an affine change of $t$, and its integrability has been studied using Kovacic’s Algorithm. Using proposition \[:prop2\] and formula (\[:qarmonic\]), we obtain the following system of differential equations for $\alpha$ and $\phi$: $$\frac{d^3\alpha}{dx_1^3}=0,\quad \frac{d^\alpha}{d x_1}\frac{d^2\phi}{dx_1^2}+ 3\frac{d^2\alpha}{d x_1^2}
\frac{d\phi}{dx_1}=0.$$ General solution of the first equation is $$\alpha = \frac{\lambda_1}{2} + \frac{\lambda_2}{2}x_1 + \frac{\lambda_3}{2}x_1^2,$$ and substituting it into the second equation we obtain a Linear Differential Equation for $\phi$, $$\frac{d^2\phi}{dx_1^2} + 3\frac{2\lambda_3}{\lambda_2+2\lambda_3 x_1}\frac{d\phi}{d x_1} = 0,$$ this equation is integrated by two quadratures, and its general solution is $$\phi = \frac{\lambda_4}{(\lambda_2 + 2\lambda_3 x_1)^2} + \lambda_0.$$ We conclude that the general formula for hamiltonians of type (\[:Hgeneral\]) with NVE (\[:quadratic\]) for any integral curve lying on $\Gamma$ is $$H = \frac{y_1^2+y_2^2}{2} +$$ $$\label{:Hqarmonic}
\frac{\lambda_4}{(\lambda_2 + 2\lambda_3 x_1)^2} + \lambda_0 -
\lambda_1 x_2^2 - \lambda_2 x_1x_2^2 - \lambda_3 x_1^2 x_2^2 +
\beta(x_1,x_2)x_2^3,$$ with $\lambda_3\neq 0$.
Formula (\[:Hqarmonic\]) is the first example, in this paper, in which we find non-linear dynamics in the invariant plane $\Gamma$. Notice that this dynamic is continuously deformed to linear dynamics when $\lambda_4$ tends to zero. In general case, for a fixed energy $h$, we have the general integral of the equation: $$8\lambda_3^2h^2(t-t_0)^2 = h(\lambda_2 + 2\lambda_3x_1)^2 - \lambda_4.$$
### NVE with polynomial coefficient $a(t)$ of odd degree
Let us consider for $n>0$ the following differential polynomial, $$Q_m(a,\dot a,\ldots) = \frac{d^m a}{dt^m}.$$ It is obvious that $a(t)$ is polynomial of degree $n$ if and only if $Q_n(a,\dot a,\ldots)\neq 0$ and $Q_{n+1}(a,\dot a,\ldots) = 0$.
Looking a proposition \[:prop2\], we see that a hamiltonian (\[:Hgeneral\]) has NVE along a generic integral curves lying on $\Gamma$, $$\label{:epol}
\ddot \xi = P_n(t)\xi,$$ with $P_n(t)$ of degree $n$, if and only if $X_h^n\alpha\neq 0$ and $X^{n+1}_h\alpha$ vanish on $\Gamma$. Let us remember expression (\[:Ex\]), $X^{n+1}_h\alpha$ vanish in $\Gamma$ if and only if $(\alpha,\phi)$ is a solution of the differential system, $$R_{n+1} = \{E_{n+1,0}(\alpha,\phi)=0,\ldots, E_{n+1,n+1}(\alpha,\phi)=0\}.$$ Using the recurrence law (\[:rlaw\]) defining the differential polynomials $E_{n,k}(\alpha,\phi)$ we are going to compute the family of hamiltonians giving rise to $a(t)$ polynomial of odd degree.
Let $(\alpha,\phi)$ be a solution of $R_{2m}$. Then, if $\frac{d\phi}{dx_1}\neq 0$, then $(\alpha,\phi)$ is a solution of $R_{2m-1}$.
By remark \[:remark\] $E_{2m-1,2k}(\alpha,\phi) = 0$ for all $m-1\geq k\geq 0$. Then let us proof that $E_{2m-1,2k+1}(\alpha,\phi)$ for all $m-2\geq k\geq 0$.
In the first step of the recurrence law defining $R_{2m}$, $$0 = E_{2m,0}(\alpha,\phi) = \frac{dE_{2m-1,1}}{dx_1}(\alpha,\phi) - \frac{d\phi}{d x_1}E_{2m-1,1}(\alpha,\phi),$$ we use $\frac{d\phi}{dx_1}\neq 0$, and remark \[:remark\], $E_{2m-1,-1}(\alpha,\phi) = 0$ to obtain, $$E_{2m-1,1}(\alpha,\phi) = 0.$$ If we assume $E_{2m-1, 2k+1}(\phi,\alpha) = 0$, substituting it in the recurrence law $$E_{2m,2k+1}(\alpha,\phi) = \frac{d E_{2m-1,2k}}{dx_1}(\alpha,\phi) -
2(k+1)\frac{d\phi}{dx_1}E_{2m-1,2(k+1)}(\alpha,\phi),$$ we obtain that $$E_{2m-1,2(k+1)}(\alpha,\phi)=0,$$ and we conclude by finite induction.
Let $H$ be a classical hamiltonian of type [(\[:Hgeneral\])]{}, then the following statements are equivalent,
1. The NVE for generic integral curve [(\[:NVEa\])]{} lying on $\Gamma$ has polynomial coefficient $a(t)$ of degree $2m-1$.
2. $H$ is written, $$\label{:odd} H = \frac{y_1^2+y_2^2}{2} + \lambda_0 - P_{2m-1}(x_1)x_2^2 +
\beta(x_1,x_2)x_2^3,$$ for $\lambda_0$ constant, and $P_{2m-1}(x_1)$ polynomial of degree $2m-1$.
It is clear that condition *1.* is satisfied if and only if $(\alpha,\phi)$ is a solution of $R_{2m}$ and it is not a solution of $R_{2m-1}$. By the previous lemma, it implies $\frac{d\phi}{dx_1}= 0$, and then the system $R_{2m}$ is reduced to $\frac{d^{2m}\alpha}{dx_1^{2m}}$ and then, $\phi$ is a constant and $\alpha$ is a polynomial of degree at most $2m-1$.
### NVE Mathieu extended
This is the standard Mathieu equation, $$\label{:Mathieu}
\ddot \xi =(c_0 + c_1\cos(\omega t))\xi, \quad \omega \neq 0.$$ We can not apply our method to compute the family of hamiltonians corresponding to this equation, because $\{c_0+c_1\cos(\omega t)\}$ is not the general solution of any differential polynomial with constant coefficients. But, let us consider $$\label{:Mequation}
Q(a) = \frac{d^3a}{dt^3}+\omega^2\frac{da}{dt},$$ the general solution of $\{Q(a)=0\}$ is $$a(t) = c_0+c_1\cos(\omega t)+c_2\sin(\omega t).$$ Just notice that, $$c_1\cos(\omega t)+c_2\sin(\omega t) = \sqrt{c_1^2+c_2^2}
\cos\left(\omega t+\arctan\frac{c_2}{c_1}\right),$$ thus NVE , when $a$ is a solution of , is reducible to Mathieu equation (\[:Mathieu\]) by a translation of time.
Using Proposition \[:prop2\], we find the system of differential equations that determine the family of hamiltonians, $$\frac{d^3\alpha}{dx_1^3}=0,\quad
\frac{d\alpha}{dx_1}\frac{d^2\phi}{dx_1^2}+3\frac{d^2\alpha}{dx_1}\frac{d\phi}{dx_1}-\omega^2\frac{d\alpha}{dx_1}=0.$$ General solution of the first equation is $$\alpha = \lambda_0 + \lambda_1x_1 + \lambda_2x_1^2.$$ substituting it in the second equation, and writing $y =
\frac{d\phi}{dx_1}$, we obtain a non homogeneous linear differential equation for $y$, $$\label{:yeq}
\frac{dy}{dx_1}+\frac{6\lambda_2y}{\lambda_1+2\lambda_2x_1}=\omega^2.$$ We must distinguish two cases depending on the parameter. If $\lambda_2 = 0$, then we just integrate the equation by trivial quadratures, obtaining $$\phi = \mu_0 + \mu_1x_1 + \frac{\omega^2x_1^2}{2}$$ and then, $$\label{:Mhamiltonian1}
H = \frac{y_1^2+y_2^2}{2}+\mu_0 + \mu_1x_1 +\frac{\omega^2
x_1^2}{2}- \lambda_0x_2^2 - \lambda_1x_1x_2^2 + \beta(x_1,x_2)x_2^3,$$ If $\lambda_2\neq 0$, then we can reduce the equation to separable using $$u = \frac{6\lambda_2y}{\lambda_1+2\lambda_2x_1},$$ obtaining $$\frac{3du}{3\omega^2-4u} = \frac{6\lambda_2 dx}{\lambda_1+2\lambda_2 x_1},\quad u = \frac{3\omega^2}{4} + \frac{3 \mu_1}{4(\lambda_1+2\lambda_2x_1)^4},$$ and then $$y = \frac{1}{8\lambda_2}\left(\omega^2\lambda_1 + 2\omega^2\lambda_2x_1 + \frac{\mu_1}{(\lambda_1 + 2\lambda_2x_1)^3}\right),$$ and finally we integrate it to obtain $\phi$, $$\phi = \int y dx_1 = \mu_0 -
\frac{\mu_1}{32\lambda_2^2}\frac{1}{(\lambda_1+2\lambda_2x_1)^2}+\frac{\omega^2\lambda_1x_1}{8\lambda_2}+\frac{\omega^2x_1^2}{8},$$ scaling the parameters adequately we write down the general formula for the hamiltonian, $$H = \frac{y_1^2+y_2^2}{2} + \mu_0 +
\frac{\mu_1}{(\lambda_1+2\lambda_2x_1)^2}+\frac{\lambda_1\omega^2 x_1}{8\lambda_2}
+ \frac{\omega^2 x_1^2}{8} +$$ $$\label{:Mhamiltonian2}
-\lambda_0 x_2^2 - \lambda_1x_1x_2^2 - \lambda_2x_1^2x_2^2 +
\beta(x_1,x_2)x_2^3.$$
Application of non-integrability criteria
-----------------------------------------
### NVE with $a(t)$ polynomial
According to theorem \[polynint\] (see Appendix), the Galois group corresponding to equations,
$$\ddot \xi = P(t)\xi$$
where $P(t)$, is a non-constant polynomial, is a connected non-abelian group. So that, we can apply theorem \[:MR\], and we get the following result.
Hamiltonians, $$H = \frac{y_1^2+y_2^2}{2} + x^2Q(x_1) + \beta(x_1,x_2)x_1^2$$ where $Q(x_1)$ is a non-constant polynomial, and $\beta(x_1,x_2)$ analytic function around $\Gamma$, do no admit any additional rational first integral.
Every integrable (by rational functions) polynomial potential with invariant plane $\Gamma=\{x_2=y_2=0\}$ is written in the following form $$V=Q_1(x_1,x_2)x_2^3+\lambda_1x_2^2+\lambda_0,\quad \lambda_0,\lambda_1\in \mathbf{C}.$$
### NVE reducible to quantum harmonic oscillator
We have seen, that hamiltonians , has generic NVE along curves in $\Gamma$ of type (\[:quadratic\]). Once again, we apply theorem \[polynint\].
Hamiltonians of the family do not admit any additional rational first integral.
We can also discuss, the Picard-Vessiot integrability of those equations . First, we shall notice that by just an scaling of $t$, $$t = \frac{\tau}{\sqrt[4]{c_2}}-\frac{c_1}{2c_2}$$ we reduce it to an *quantum harmonic oscillator equation*,
$$\label{:E}
\frac{d^2\xi}{d\tau^2} = (\tau^2 - E)\xi, \quad\quad E =
\frac{c_1^2-4c_0c_1}{4\sqrt{c_2^3}}.$$
In appendix (remark \[:rkm2\]) we analyze this equation. It is *Picard-Vessiot integrable* if and only if $E$ is an odd positive number. Then, let us compute de parameter $E$ associated to NVE of integral curves of Hamiltonians \[:Hqarmonic\].
Let us keep in mind that the family of those curves is parameterized by $$h = \frac{y_1^2}{2}+\frac{\lambda_4}{(\lambda_2+2\lambda_3x_1)^2}$$ In order to fix the parameterization of those curves, let us assume that time $t=0$ corresponds to $x_1 = 0$. The NVE corresponding to a curve, depending on energy $h$, is written: $$\ddot \xi = (c_0(h)+c_1(h)t + c_2(h)t^2)\xi.$$ We compute these coefficients $c_i(h)$ using, $$y_1 = \frac{\sqrt{2h(\lambda_2+2\lambda_3x_1)^2-2\lambda_4}}{\lambda_2+2\lambda_3x_1}
\xrightarrow{t\to
0}\frac{\sqrt{2h\lambda_2^2-2\lambda_4}}{\lambda_2},$$ and then, by applying the hamiltonian field, $$c_0(h) = \frac{\lambda_1}{2},\quad c_1(h) =
\sqrt{\frac{h\lambda_2^2-\lambda_4}{2}},\quad c_2(h) = \lambda_3h,$$ and then, it is reducible to equation (\[:E\]) with parameter, $$E = \frac{1}{8\sqrt{\lambda_3^3}}\left(\frac{\lambda_2^2-4\lambda_1\lambda_3}{\sqrt{h}}-\frac{\lambda_4}{\sqrt{h^3}} \right),$$
If $\lambda_2 = 4\lambda_1\lambda_3$ and $\lambda_4 = 0$, then parameter $E$ vanish for every integral curve in $\Gamma$. For any other case, $E$ is a non-constant analytical function of $h$.
*We have proven that those NVE are, generically not Picard-Vessiot integrable for any hamiltonian of the family.*
### NVE Mathieu
In order to apply theorem \[:MR\], we just need to make some remarks on the field of coefficients. Let $\gamma$ be a generic integral curve of , or . Those curves are, in general, Riemann spheres. The field of coefficients $\mathcal M_\gamma$, is generated by $x_1, y_1$, so that it is $\mathbf C(x_1,\dot x_1)$. We also have, $$a = \lambda_0 + \lambda_1 x_1 + \lambda_2 x_1^2, \quad \dot a = \dot x_1(\lambda_1 + 2\lambda_2 x_1),$$ so that, for $\lambda_2 = 0$, we have $$\mathcal M_\gamma = \mathbf(\alpha,\dot\alpha) = \mathbf C(\sin t, cos t) = \mathbf {\mathbf C}(e^{it}).$$ and, for $\lambda_2 \neq 0$, $$\mathbf {\mathbf C}(e^{it}) \hookrightarrow \mathcal M_\gamma$$ is an algebraic extension.
So that, in our algebrization algorithm, the field of coefficients of Mathieu equation is taken, $\mathbf {\mathbf C}(e^{it})$. For $\lambda_2 = 0$, we can apply directly theorem \[:MR\], and for $\lambda_2 \neq 0$, we can apply theorem \[moramis\] and then theorem \[:MR\].
Non trivial equations of type Mathieu, with field of coefficients $\mathbf C(e^{it})$, analyzed in remark \[Mathiewobs\], have Galois group $SL(2,\mathbf C)$. Thus for computed families of hamiltonians with NVE of type Mathieu, we get:
Hamiltonians of the families and if $\lambda_1 \neq 0$ and $(\lambda_1,\lambda_2) \neq (0,0)$ respectively, do not admit any additional rational first integral.
Thanks and acknowledgements {#thanks-and-acknowledgements .unnumbered}
---------------------------
We want to thank Juan Morales-Ruiz for his valuable help, advices, suggestions, and also for proposing the initial problem. We also acknowledge Sergi Simón for his valuable suggestions in the first stage of developing our method. We are also indebted with Jackes Arthur Weil by his suggestions on Kovacic’s algorithm.
This research is supported by project [**BFM2003-09504-C02-02**]{} (Spanish Government).
Appendix {#appendix .unnumbered}
========
Kovacic’s Algorithm
===================
This algorithm is devoted to solve the reduced linear differential equation (RLDE) $\xi''=r\xi$ and is based on the algebraic subgroups of $SL(2,\mathbf{C}).$ For more details see [@Kov]. Improvements for this algorithm is given in [@UlmerWeil], in where is not necessary to reduce the equation. Here, we follows the original version given by Kovacic.
\[subgroups\] Let $G$ be an algebraic subgroup of $SL(2,\mathbf{C})$. Then one of the following four cases can occur.
1. $G$ is triangularizable.
2. $G$ is conjugate to a subgroup of infinite dihedral group (also called meta-abelian group) and case 1 does not hold.
3. Up to conjugation $G$ is either of following finite groups: Tetrahedral group, Octahedral group or Icosahedral group, and cases 1 and 2 do not hold.
4. $G = SL(2,\mathbf{C})$.
Each case in the Kovacic algorithm is related with each one of the algebraic subgroups of $SL(2,\mathbf{C})$ and the associated Riccatti equation $$\theta^{\prime}=r-\theta ^{2}=\left( \sqrt{r}-\theta\right)
\left( \sqrt{r}+\theta\right),\quad\theta={\xi'\over \xi}.$$
According to theorem \[subgroups\], there are four cases in the Kovacic algorithm. Only for cases 1, 2 and 3 we can solve the differential equation RLDE, but for the case 4 we have not Liouvillian solutions for RLDE. Is possible that the Kovacic algorithm only can provide us only one solution ($y_1$), so that we can obtain the second solution ($y_2$) through $$\label{second}
y_2=y_1\int\frac{dx}{y_1^2}.$$
[**Notations.**]{} For the equation RLDE with $$r={s\over t},\quad s,t\in \mathbf{C}[x]$$ we use the following notations.
1. Denote by $\Gamma'$ be the set of (finite) poles of $r$, $\Gamma^{\prime}=\left\{ c\in\mathbf{C}:t(c)=0\right\}$.
2. Denote by $\Gamma=\Gamma^{\prime}\cup\{\infty\}$.
3. By the order of $r$ at $c\in \Gamma'$, $\circ(r)_c$, we mean the multiplicity of $c$.
4. By the order of $r$ at $\infty$, $\circ\left(
r_{\infty}\right) ,$ we mean the order of $\infty$ as a zero of $r$. That is $\circ\left( r_{\infty }\right) =deg(t)-deg(s)$.
5. By the order of $r$ at $c\in \Gamma'$, $\circ(r)_c$, we mean the multiplicity of $c$.
The four cases
--------------
[**Case 1.**]{} In this case $\left[ \sqrt{r}\right] _{c}$ and $\left[ \sqrt{r}\right] _{\infty}$ means the Laurent series of $\sqrt{r}$ at $c$ and the Laurent series of $\sqrt{r}$ at $\infty$ respectively. Furthermore, we define $\varepsilon(p)$ as follows: if $p\in\Gamma,$ then $\varepsilon\left( p\right) \in\{+,-\}.$ Finally, the complex numbers $\alpha_{c}^{+},\alpha_{c}^{-},\alpha_{\infty}^{+},\alpha_{\infty}^{-}$ will be defined in the first step. If the differential equation has not poles it only can fall in this case.
[**Step 1.**]{} Search for each $c \in \Gamma'$ and for $\infty$ the corresponding situation such as follows:
$(c_{0})$
: If $\circ\left( r_{c}\right) =0$, then $$\left[ \sqrt {r}\right] _{c}=0,\quad\alpha_{c}^{\pm}=0.$$
$(c_{1})$
: If $\circ\left( r_{c}\right) =1$, then $$\left[ \sqrt {r}\right] _{c}=0,\quad\alpha_{c}^{\pm}=1.$$
$(c_{2})$
: If $\circ\left( r_{c}\right) =2,$ and $$r= \cdots
+ b(x-c)^{-2}+\cdots,\quad \text{then}$$ $$\left[ \sqrt {r}\right]_{c}=0,\quad \alpha_{c}^{\pm}=\frac{1\pm\sqrt{1+4b}}{2}.$$
$(c_{3})$
: If $\circ\left( r_{c}\right) =2v\geq4$, and $$r=
(a\left( x-c\right) ^{-v}+...+d\left( x-c\right)
^{-2})^{2}+b(x-c)^{-(v+1)}+\cdots,\quad \text{then}$$ $$\left[
\sqrt {r}\right] _{c}=a\left( x-c\right) ^{-v}+...+d\left(
x-c\right) ^{-2},\quad\alpha_{c}^{\pm}=\frac{1}{2}\left(
\pm\frac{b}{a}+v\right).$$
$(\infty_{1})$
: If $\circ\left( r_{\infty}\right) >2$, then $$\left[\sqrt{r}\right] _{\infty}=0,\quad\alpha_{\infty}^{+}=0,\quad\alpha_{\infty}^{-}=1$$
$(\infty_{2})$
: If $\circ\left( r_{\infty}\right) =2,$ and $r= \cdots + bx^{2}+\cdots$, then $$\left[
\sqrt{r}\right] _{\infty}=0,\quad\alpha_{\infty}^{\pm}=\frac{1\pm\sqrt{1+4b}}{2}$$
$(\infty_{3})$
: If $\circ\left( r_{\infty}\right) =-2v\leq0$, and $$r=\left( ax^{v}+...+d\right) ^{2}+ bx^{v-1}+\cdots,\quad \text{then}$$ $$\left[ \sqrt{r}\right] _{\infty}=ax^{v}+...+d,\quad
\text{and}\quad \alpha_{\infty}^{\pm }=\frac{1}{2}\left(
\pm\frac{b}{a}-v\right).$$
[**Step 2.**]{} Find $D\neq\emptyset$ defined by $$D=\left\{
m\in\mathbf{Z}_{+}:m=\alpha_{\infty}^{\varepsilon
(\infty)}-{\displaystyle\sum\limits_{c\in\Gamma^{\prime}}}
\alpha_{c}^{\varepsilon(c)},\forall\left( \varepsilon\left(
p\right) \right) _{p\in\Gamma}\right\} .$$ If $D=\emptyset$, then we should start with the case 2. Now, if $\#D>0$, then for each $m\in D$ we search $\omega$ $\in\mathbf{C}(x)$ such that $$\omega=\varepsilon\left(
\infty\right) \left[ \sqrt{r}\right] _{\infty}+{\displaystyle\sum\limits_{c\in\Gamma^{\prime}}}
\left( \varepsilon\left( c\right) \left[ \sqrt{r}\right] _{c}+{\alpha_{c}^{\varepsilon(c)}}{(x-c)^{-1}}\right).$$
[**Step 3**]{}. For each $m\in D$, search for a monic polynomial $P_m$ of degree $m$ with $$P_m'' + 2\omega P_m' + (\omega' + \omega^2 - r) P_m = 0.$$
If success is achieved then $\xi_1=P_m e^{\int\omega}$ is a solution of the differential equation RLDE. If not, then case 1 cannot hold.
[**Case 2.**]{} Search for each $c \in \Gamma'$ and for $\infty$ the corresponding situation such as follows:
[**Step 1.**]{} Search for each $c\in\Gamma^{\prime}$ and $\infty$ the sets $E_{c}\neq\emptyset$ and $E_{\infty}\neq\emptyset.$ For each $c\in\Gamma^{\prime}$ and for $\infty$ it is define $E_{c}\subset\mathbf{Z}$ and $E_{\infty}\subset\mathbf{Z}$ as follows:
($c_1$)
: If $\circ\left( r_{c}\right)=1$, then $E_{c}=\{4\}$
($c_2$)
: If $\circ\left( r_{c}\right) =2,$ and $r= \cdots +
b(x-c)^{-2}+\cdots ,\ $ then $$E_{c}=\left\{
2+k\sqrt{1+4b}:k=0,\pm2\right\}.$$
($c_3$)
: If $\circ\left( r_{c}\right) =v>2$, then $E_{c}=\{v\}$
$(\infty_{1})$
: If $\circ\left( r_{\infty}\right) >2$, then $E_{\infty }=\{0,2,4\}$
$(\infty_{2})$
: If $\circ\left( r_{\infty}\right) =2,$ and $r= \cdots + bx^{2}+\cdots$, then $$E_{\infty }=\left\{
2+k\sqrt{1+4b}:k=0,\pm2\right\}.$$
$(\infty_{3})$
: If $\circ\left( r_{\infty}\right) =v<2$, then $E_{\infty }=\{v\}$
[**Step 2.**]{} Find $D\neq\emptyset$ defined by $$D=\left\{
m\in\mathbf{Z}_{+}:\quad m=\frac{1}{2}\left( e_{\infty}-
{\displaystyle\sum\limits_{c\in\Gamma^{\prime}}} e_{c}\right)
,\forall e_{p}\in E_{p},\text{ }p\in\Gamma\right\}.$$ If $D=\emptyset,$ then we should start the case 3. Now, if $\#D>0,$ then for each $m\in D$ we search a rational function $\theta$ defined by $$\theta=\frac{1}{2}
{\displaystyle\sum\limits_{c\in\Gamma^{\prime}}}
\frac{e_{c}}{x-c}.$$
[**Step 3.**]{} For each $m\in D,$ search a monic polynomial $P_m$ of degree $m$, such that $$P_m^{\prime\prime\prime}+3\theta
P_m^{\prime\prime}+(3\theta^{\prime}+3\theta
^{2}-4r)P_m^{\prime}+\left( \theta^{\prime\prime}+3\theta\theta^{\prime}+\theta^{3}-4r\theta-2r^{\prime}\right)P_m=0.$$ If $P_m$ there is not exists, then the case 2 cannot hold. If such a polynomial is found, set $\phi = \theta + P'/P$ and let $\omega$ be a solution of $$\omega^2 + \phi \omega + {1\over2}\left(\phi' + \phi^2 -2r\right)=
0.$$
Then $\xi_1 = e^{\int\omega}$ is a solution of the differential equation RLDE.
[**Case 3.**]{} Search for each $c \in \Gamma'$ and for $\infty$ the corresponding situation such as follows:
[**Step 1.**]{} Search for each $c\in\Gamma^{\prime}$ and $\infty$ the sets $E_{c}\neq\emptyset$ and $E_{\infty}\neq\emptyset.$ For each $c\in\Gamma^{\prime}$ and for $\infty$ it is define $E_{c}\subset\mathbf{Z}$ and $E_{\infty}\subset\mathbf{Z}$ as follows:
$(c_{1})$
: If $\circ\left( r_{c}\right) =1$, then $E_{c}=\{12\}$
$(c_{2})$
: If $\circ\left( r_{c}\right) =2,$ and $r= \cdots +
b(x-c)^{-2}+\cdots$, then $$E_{c}=\left\{ 6+k\sqrt{1+4b}:\quad
k=0,\pm1,\pm2,\pm3,\pm4,\pm5,\pm6\right\}.$$
$(\infty)$
: If $\circ\left( r_{\infty}\right) =v\geq2,$ and $r=
\cdots + bx^{2}+\cdots$, then $$E_{\infty }=\left\{
6+{12k\over n}\sqrt{1+4b}:\quad
k=0,\pm1,\pm2,\pm3,\pm4,\pm5,\pm6\right\},\quad n\in\{4,6,12\}.$$
[**Step 2.**]{} Find $D\neq\emptyset$ defined by $$D=\left\{
m\in\mathbf{Z}_{+}:\quad m=\frac{n}{12}\left(
e_{\infty}-{\displaystyle\sum\limits_{c\in\Gamma^{\prime}}}
e_{c}\right) ,\forall e_{p}\in E_{p},\text{
}p\in\Gamma\right\}.$$ In this case we start with $n=4$ to obtain the solution, afterwards $n=6$ and finally $n=12$. If $D=\emptyset$, then the differential equation has not Liouvillian solution because falls in the case 4. Now, if $\#D>0,$ then for each $m\in D$ with its respective $n$, it is search a rational function $$\theta={n\over 12}{\displaystyle\sum\limits_{c\in\Gamma^{\prime}}}
\frac{e_{c}}{x-c}$$ and a polynomial $S$ defined as $$S=
{\displaystyle\prod\limits_{c\in\Gamma^{\prime}}} (x-c).$$
[**Step 3**]{}. Search for each $m\in D$, with its respective $n$, a monic polynomial $P_m=P$ of degree $m,$ such that its coefficients can be determined recursively by $$\bigskip P_{-1}=0,\quad P_{n}=-P,$$ $$P_{i-1}=-SP_{i}^{\prime}-\left( \left( n-i\right)
S^{\prime}-S\theta\right) P_{i}-\left( n-i\right) \left(
i+1\right) S^{2}rP_{i+1},$$ where $i\in\{0,1\ldots,n-1,n\}.$ If $P$ there is not exists, then the differential equation has not Liouvillian solution because falls in the case 4. Now, if $P$ there exists, it is search $\omega$ such that $${\displaystyle\sum\limits_{i=0}^{n}} \frac{S^{i}P}{\left(
n-i\right) !}\omega^{i}=0,$$ then a solution of the differential equation RLDE is given by $$\xi=e^{\int \omega},$$ where $\omega$ is solution of the previous polynomial of degree $n$.
Some remarks on Kovacic’s algorithm
-----------------------------------
Along of this section we assume that RLDE falls only in one of the four cases.
\[rkov2\] If RLDE fall in case 1, then its Galois group is given by any of the following groups:
I1
: $e$ when the algorithm provides two rational solutions or only one rational solution and the second solution obtained by equation has not logarithmic term. $$e=\left\{\begin{pmatrix}1&0\\0&1\end{pmatrix}\right\},$$ this group is connected and abelian.
I2
: $\mathbf{G}_k$ when the algorithm provides only one algebraic solution $\xi$ such that $\xi^k\in\mathbf{C}(x)$ and $\xi^{k-1}\notin\mathbf{C}(x)$. $$\mathbf{G}_k=\left\{\begin{pmatrix}\lambda&d\\0&\lambda^{-1}\end{pmatrix}:\quad \lambda\text{ is a $k$-root of the unity, }d\in\mathbf{C}\right\},$$ this group is disconnected and its identity component is abelian.
I3
: $\mathbf{C}^*$ when the algorithm provides two non algebraic solutions. $$\mathbf{C}^*=\left\{\begin{pmatrix}c&0\\0&c^{-1}\end{pmatrix}:c\in \mathbf{C}^*\right\},$$ this group is connected and abelian.
I4
: $\mathbf{C}^{+}$ when the algorithm provides one rational solution and the second solution is not algebraic. $$\mathbf{C}^{+}=\left\{\begin{pmatrix}1&d\\0&1\end{pmatrix}:d\in \mathbf{C}\right\}, \quad {\xi}\in\mathbf{C}(x),$$ this group is connected and abelian.
I5
: $\mathbf{C}^*\ltimes\mathbf{C}^{+}$ when the algorithm only provides one solution $\xi$ such that $\xi$ and its square are not rational functions. $$\mathbf{C}^*\ltimes\mathbf{C}^{+}=\left\{\begin{pmatrix}c&d\\0&c^{-1}\end{pmatrix}:c\in\mathbf{C}^*,d\in \mathbf{C}\right\}, \quad \xi\notin\mathbf{C}(x),\quad {\xi}^2\notin\mathbf{C}(x).$$ This group is connected and non-abelian.
I6
: $SL(2,\mathbf{C})$ if the algorithm do not provides any solution. This group is connected and non-abelian.
If RLDE fall in case 2, then the Kovacic’s Algorithm can provides one or two solutions. This depends of $r$ such as follows:
II1
: if $r$ is given by $$r={2\phi'+2\phi-\phi^2\over 4},$$ then there exists only one solution,
II2
: if $r$ is given by $$r\neq{2\phi'+2\phi-\phi^2\over 4},$$ then there exists two solutions.
II3
: The identity component of the Galois group for this case is abelian.
If RLDE falls in case 3, then its Galois group is given by any of the following groups:
III1
: [**Tetrahedral group**]{} when $\omega$ is obtained with $n=4.$ This group of order 24 is generated by $$\begin{pmatrix}
e^{\frac{k\pi i}{3}} & 0\\
0 & e^{-\frac{k\pi i}{3}}
\end{pmatrix}
, \quad\frac{1}{3}\left( 2e^{\frac{k\pi i}{3}}-1\right)
\begin{pmatrix}
1 & 1\\
2 & -1
\end{pmatrix},\quad k\in \mathbf{Z}.$$
III2
: [**Octahedral group**]{} when $\omega$ is obtained with $n=6.$ This group of order 48 is generated by $$\begin{pmatrix}
e^{\frac{k\pi i}{4}} & 0\\
0 & e^{-\frac{k\pi i}{4}}
\end{pmatrix}, \quad \frac{1}{2}e^{\frac{k\pi i}{4}}\left( e^{\frac{k\pi
i}{2}}+1\right)
\begin{pmatrix}
1 & 1\\
1 & -1
\end{pmatrix},\quad k\in \mathbf{Z}.$$
III3
: [**Icosahedral group**]{} when $\omega$ is obtained with $n=12.$ This group of order 120 is generated by $$\begin{pmatrix}
e^{\frac{k\pi i}{5}} & 0\\
0 & e^{-\frac{k\pi i}{5}}
\end{pmatrix}
,\quad
\begin{pmatrix}
\phi & \psi\\
\psi & -\phi
\end{pmatrix},\quad k\in \mathbf{Z},$$ being $\phi$ and $\psi$ defined as $$\phi=\frac{1}{5}\left( e^{\frac{3k\pi
i}{5}}-e^{\frac{2k\pi i}{5} }+4e^{\frac{k\pi i}{5}}-2\right),
\quad \psi=\frac{1}{5}\left( e^{\frac{3k\pi i}{5}}+3e^{\frac{2k\pi
i}{5}}-2e^{\frac{k\pi i}{5}}+1\right)$$
III4
: The identity component of the Galois group for this case is abelian.
[99]{}
*Les systèmes Hamiltoniens et leur integrabilité,* Société Mathématique de France, Marseille (2001).
*Solving linear ordinary differential equations over $C(x,e\sp {\int f(x)dx})$.* Proceedings of the 1999 International Symposium on Symbolic and Algebraic Computation (Vancouver, BC), 173–179 (electronic), ACM, New York, (1999). *An Algorithm for Solving Second Order Linear Homogeneus Differential Equations,* J. Symb. Comput. [**2**]{} (1986), 3–43.
*A list of all integrable 2D homogeneous polynomial potentials with a polynomial integral of order at most 4 in the momenta* Preprint. arXiv:nlin.SI/018051 v1 Aug 2001
*Darboux points and integrability of Hamiltonian systems with homogeneous polynomial potential,* J. Math. Phys. [**46**]{}, 062901 (2005)
*Differential Galois Theory and Non-Integrability of Hamiltonian systems,* Progress in Mathematics [**179**]{}, Birkhäuser (1999).
*Non-integrability criteria for hamiltonians in the case of Lamé Normal Variational Equations,* J. Diff. Eq. [**129**]{} (1996), 111–135.
, *Galoisian obstructions to integrability of Hamiltonian systems, I*, Methods and Applications of Analysis **8** (2001), 33–95.
, *Galoisian obstructions to integrability of Hamiltonian systems, II*, Methods and Applications of Analysis **8** (2001), 97–112.
*Integrability of Hamiltonian Systems and Differential Galois Groups of Higher Variational Equations*, to appear.
*Note on kovacic’s algorithm*. J. Symb. Comp. **22**, (1996), 179–200.
, *Galois Theory in Linear Differential Equations*, Springer Verlag, New York, (2003).
, *Differential equations of order two with one singular point*, J. Symb. Comp. **28**, (1999), 495–520.
, *Polynomial Riccati equations with algebraic solutions*, Differential Galois theory (Bedlewo, 2001), 219–231, Banach Center Publ., **58**, Polish Acad. Sci., Warsaw, (2002)
| {
"pile_set_name": "ArXiv"
} |
---
title: 'Une correspondance de Jacquet-Langlands $p$-adique'
---
Gaëtan CHENEVIER
Ecole normale supérieure, DMA [^1]
Université Paris 7
[*Abstract:*]{} In this paper, we extend the Jacquet-Langlands’correspondence, between Hecke-modules of usual modular forms and quaternionic modular forms, to overconvergent $p$-adic forms of finite slope. We show that this correspondence respects $p$-adic families and is induced by an isomorphism between some associated eigencurves.
AMS classification: 11F85 (11F12, 11F72, 14G22)
Introduction
============
Soit $D/{\mathbb{Q}}$ l’algèbre de quaternions définie de discriminant $d$, $p$ un nombre premier et $N$ un entier tels que $(Np,d)=(N,p)=1$, nous établissons dans ce texte un transfert “à la Jacquet-Langlands” bijectif entre les formes modulaires $p$-adiques surconvergentes de pente finie ([@K], [@Col2]), propres, cuspidales et nouvelles en $d$, et les formes modulaires $p$-adiques quaternioniques pour $D$ de pente finie ([@Buz2]), en niveau modéré $N$ et poids-caractère quelconque (théorème \[jlp\]). Ce transfert est Hecke-équivariant et coïncide avec la correspondance de Jacquet-Langlands usuelle lorsqu’on le restreint aux formes modulaires “classiques” de part et d’autre. Mieux, il respecte les familles $p$-adiques, et nous prouvons qu’il provient d’un isomorphisme rigide-analytique entre les courbes de Hecke (“the eigencurves”) correspondantes (théorème \[jlpfamille\]).
Notre preuve est basée sur l’existence de systèmes de modules de Banach de part et d’autre (\[banachsystem\]), et de leur comparaison. On démontre en fait un énoncé général (théorème \[general\]) sur la comparaison de ces systèmes. Des arguments de Zariski-densité de points “classiques” sur certaines hypersurfaces de Fredholm et sur les variétés de Hecke y jouent un rôle important. On en déduit notre correspondance en utilisant la correspondance de Jacquet-Langlands usuelle, et les assertions de “classicité en petite pente” de part et d’autre.
Un des charmes de cette correspondance est que les objets qu’elle compare sont, dans une large mesure, de natures assez différentes. Les formes modulaires $p$-adiques surconvergentes usuelles sont des sections de fibrés surconvergentes sur le lieu ordinaire des courbes modulaires, alors que les formes modulaires quaternioniques $p$-adiques sont purs produits de la théorie des groupes, spécialement des représentations $p$-adiques de $GL_2({\mathbb{Q}}_p)$. Nous y voyons plusieurs intérêts, l’un d’entre eux étant la possibilité d’introduire des méthodes de théorie des représentations $p$-adiques des groupes de Lie $p$-adiques dans la théorie des formes modulaires surconvergentes. Ceci est par exemple en accord avec une philosphie de Langlands purement $p$-adique (encore non formulée à notre connaissance). Un autre intêret vient de qu’il reste une multitude de questions non résolues concernant “the eigencurve” (voir l’introduction de [@eigen]), l’origine plus combinatoire des formes quaternioniques peut permettre, sinon de les résoudre, de faciliter tout au moins les expériences numériques. Nous discutons de certaines conséquences de notre correspondance, ainsi que de problèmes encore ouverts, dans la dernière partie du texte.
Dans [@ch], des variétés de Hecke sont construites pour tous les groupes algébriques $G$ sur ${\mathbb{Q}}$ tels que $G({\mathbb{R}})=U_n({\mathbb{C}})$, $G({\mathbb{Q}}_p)=GL_n({\mathbb{Q}}_p)$, et on peut se demander de manière générale si les transferts usuels se prolongent, et comment, aux formes modulaires surconvergentes pour ces groupes. En ce qui concerne les correspondances de Jacquet-Langlands éventuelles entre de tels groupes, nos résultats s’y appliquent sans modification, du moment que la correspondance classique entre les deux groupes en question est connue pour suffisament de poids (voir le théorème \[general\]).
Ce travail repose substantiellement sur ceux de Buzzard, Coleman, et Mazur, nous les en remercions. La question de l’existence ou non d’un morphisme entre les deux courbes de Hecke avait notamment été posée par K.Buzzard.
[**Notations**]{}: $p$ est un nombre premier. ${\mathbb{C}}_p$ est le complété d’une clôture algébrique de ${\mathbb{Q}}_p$, muni de la norme telle que $|p|=1/p$, $v(.)$ est la valuation associée. Si $X/{\mathbb{C}}_p$ est un espace rigide, $A(X)$ est l’anneau des fonctions rigides analytiques sur $X$; si $X$ est de plus affinoide réduit, la norme sup. sur $X$ munit $A(X)$ d’une structure de ${\mathbb{C}}_p$-algèbre de Banach. ${\mathbb{A}}^1$ désigne la droite rigide sur ${\mathbb{C}}_p$. ${\mathbb{A}}$ (resp. ${\mathbb{A}}_f$, resp. ${\mathbb{A}}_f^{(d)}$) est l’anneau des adèles (resp. adèles finies, resp. adèles hors de $\{d,\infty\}$) de ${\mathbb{Q}}$. Si $d$ et $n$ sont des entiers, on note $d{\, |\!| \,}n$ si $d$ divise $n$ et $(d,n/d)=1$. Si $B$ est un anneau, $B^*$ désigne le groupe multiplicatif de ses inversibles. ${\textrm{diag}}(a,b)$ désigne la matrice $\left( \begin{array}{cc} a & 0 \\ 0 & b \end{array}
\right)$.
Systèmes de modules de Banach
=============================
Modules de Banach {#banachsystem}
-----------------
La notion de systèmes de modules de Banach orthonormalisables est introduite dans [@eigen] §4.3.
On appellera [*système d’espaces de Banach*]{} la donnée d’un ensemble $E=\{ E_n, i_n, n \in {\mathbb{N}}\}$ de ${\mathbb{C}}_p$-espaces de Banach orthonormalisables $E_n$, et d’applications ${\mathbb{C}}_p$-linéaires compactes $i_{n}: E_n \rightarrow E_{n+1}$, on notera $E^{\dagger}$ la limite inductive des $E_n$ selon les $i_n$. Si ${\mathcal{W}}/{\mathbb{C}}_p$ est un espace rigide réduit, on appellera [*faisceau de modules de Banach orthonormalisables sur ${\mathcal{W}}$*]{} la donnée d’un faisceau $B$ sur ${\mathcal{W}}$ tel que pour tout ouvert affinoide $V \subset {\mathcal{W}}$, $B(V)$ ait une structure de $A(V)$-module de Banach orthonormalisable, et tel que si $V \subset V'$ sont des ouverts affinoides de ${\mathcal{W}}$, l’application canonique $B(V')\widehat{\otimes}_{A(V')}A(V) \rightarrow B(V)$ soit un isomorphisme de $A(V)$-modules de Banach. Un [*système de modules de Banach sur ${\mathcal{W}}$*]{} est la donnée d’un ensemble $E=\{E_n,i_n, n\in {\mathbb{N}}\}$ où $E_n$ est un faisceau de modules de Banach orthonormalisables sur ${\mathcal{W}}$, et $i_n: E_n \rightarrow E_{n+1}$ est un morphisme de faisceaux tel que si $V$ est un ouvert affinoide de ${\mathcal{W}}$, $i_n(V): E_n(V) \rightarrow E_{n+1}(V)$ soit $A(V)$-linéaire compacte. On notera alors $E(V,n):=E_n(V)$, et on abrégera le tout en $E=(E(V,n))$. Enfin, on dispose d’une notion évidente de sous-système de modules de Banach sur ${\mathcal{W}}$.
[*Exemple:*]{} Si $E=\{ E_n, i_n, n \in {\mathbb{N}}\}$ est un système d’espaces de Banach, on peut lui associer un système de modules de Banach sur ${\mathcal{W}}$ comme suit. Soit $E_{{\mathcal{W}},n}$ le faisceau de modules de Banach orthonormalisables associé à $E_n$ au sens de [@ch] 3.7.2. On rappelle que si $V$ est ouvert affinoide, on a $E_{{\mathcal{W}},n}(V):=A(V)\widehat{\otimes}_{{\mathbb{C}}_p} E_n$ que l’on notera encore $E(V,n)$. $i_n$ nous fournit de plus par extension des scalairs une application compacte $A(V)$-linéaire $i_n(V): E(V,n) \rightarrow E(V,n+1)$. On appellera $(E(V,n))$ [*le système de modules de Banach sur ${\mathcal{W}}$ associé à $E$*]{}.
Fixons $E=(E(V,n))$ un système de modules de Banach sur ${\mathcal{W}}$. Si $x \in {\mathcal{W}}({\mathbb{C}}_p)$, on notera $E_{x}$ le système d’espaces de Banach déduit de $E$ par évaluation en $x$: $E_{x,n}:=E(V,n)\widehat{\otimes}_{A(V)}{\mathbb{C}}_p$ pour tout $V$ tel que $x \in
V({\mathbb{C}}_p)$ ($A(V) \rightarrow {\mathbb{C}}_p$ étant l’évaluation en $x$), $i_{x,n}: E_{x,n}
\rightarrow E_{x,n+1}$ l’application compacte déduite. Dans nos applications, les $i_{x,n}$ seront toujours injectives. C’est par exemple le cas si $E_{{\mathcal{W}}}$ est le système de modules de Banach sur ${\mathcal{W}}$ associé a un système d’espaces de Banach $\{E_n, \, i_n, \, n \in {\mathbb{N}}\}$ tel que les $i_n$ soient injectives.
On appellera endomorphisme de $E$ la donnée, pour chaque $V \subset {\mathcal{W}}$ ouvert affinoide, et pour tout $n$ assez grand ($V$ étant fixé), d’un endomorphisme continu $A(V)$-linéaire $U(V,n)$ de $E(V,n)$, tel que les $U(V,n)$ commutent aux applications $E(V,n) \rightarrow E(V,n+1)$, et aux changements de bases ouverts affinoides $E(V,n) \rightarrow E(V',n)$ lorsqu’ils sont définis. On dira que $U:=(U(V,n))$ est un endomorphisme de $E$ et on identifiera deux endomorphismes $U$ et $U'$ si pour tout $V$ et $n$ assez grand, $U(V,n)=U'(V,n)$. L’ensemble ${\textrm{End}}(E)$ de ces endomorphismes est alors une ${\mathbb{C}}_p$-algèbre de manière naturelle. Si $H$ est un anneau (resp. $G$ un moinoide), une représentation de $H$ (resp. de $G$) sur $E$ est la donnée d’un morphisme de ${\mathbb{C}}_p$-algèbre $H \rightarrow
{\textrm{End}}(E)$ (resp. ${\mathbb{C}}_p[G] \rightarrow {\textrm{End}}(E)$). On parlera alors de système de $H$-modules de Banach sur ${\mathcal{W}}$.
On note ${\textrm{Comp}}(E)$ l’idéal bilatère de ${\textrm{End}}(E)$ composé des éléments ayant la propriété suivante: pour $V \subset {\mathcal{W}}$ un ouvert affinoide fixé, il existe pour tout $n$ assez grand un endomorphisme $A(V)$-linéaire continu $T(V,n): E(V,n+1) \rightarrow E(V,n)$ tel que le diagramme suivant soit commutatif:
$$\xymatrix{ E(V,n) \ar@{->}[d]_{i_{n}(V)} \ar@{->}[rr]^{U(V,n)}
& & E(V,n) \ar@{->}[d]^{i_{n}(V)} \\
E(V,n+1) \ar@{->}[rr]_{U(V,n+1)} \ar@{->}[urr]^{T(V,n)} & & E(V,n+1) }$$
En particulier, $U(V,n)$ est compact, et [@BMF] A.2.3 assure que $\det(1-U(V,n)_{|E(V,n)}) \in 1+TA(V)\{\{T\}\}$ est indépendant de $n$ assez grand. La formation des séries de Fredholm commutant aux changements de base ouverts affinoides, toutes ces séries proviennent par restriction d’une unique série de Fredholm $_E$($U$) sur ${\mathcal{W}}$, $_E$($U$) $\in
1+TA({\mathcal{W}})\{\{T\}\}$.
L’espace des poids-caractères {#poids}
-----------------------------
${}^{}$
On fixe $p>2$ dans ce qui suit, $\Lambda$ est l’anneau local complet ${\mathbb{Z}}_p[[(1+p{\mathbb{Z}}_p)]]$, ${\mathcal{W}}$ la boule ouverte rigide de centre $1$ et rayon $1$, ${\mathcal{W}}({\mathbb{C}}_p):=\{z \in {\mathbb{C}}_p, \, \, |z-1|<1\}$. L’application $\Lambda \rightarrow
A({\mathcal{W}})$ définie par $[1+p] \mapsto Z$ identifie topologiquement $\Lambda$ aux fonctions analytiques ${\mathbb{Q}}_p$-valuées sur ${\mathcal{W}}({\mathbb{Q}}_p)$ bornées par $1$ sur tout ${\mathcal{W}}({\mathbb{C}}_p)$. L’application $\kappa \mapsto \kappa(1+p)$ induit une bijection: $$\textrm{Hom}_{gr-cont}(1+p{\mathbb{Z}}_p,{\mathbb{C}}_p^*) \simeq {\mathcal{W}}({\mathbb{C}}_p)$$
On dispose d’un caractère “universel” continu déterminé par $\kappa^{univ}(1+p)=Z$: $$\kappa^{univ}: 1+p{\mathbb{Z}}_p
\rightarrow A({\mathcal{W}})^*$$
On note $\mu_{p^{\infty}}:=\{\zeta \in {\mathbb{C}}_p^*, \, \, \exists r \in {\mathbb{N}},
\zeta^{p^r}=1\}$, $\mu_{p^{\infty}} \subset {\mathcal{W}}({\mathbb{C}}_p)$. Si $\zeta
\in {\mathbb{C}}_p^*$ est tel que $\zeta^{p^r}=1$, $\zeta.(1+p)^k \in {\mathcal{W}}({\mathbb{C}}_p)$ correspond au caractère $x \rightarrow x^k \chi(x)$, $\chi$ étant le caractère d’ordre fini de $1+p{\mathbb{Z}}_p$ tel que $\chi(1+p)=\zeta$, caractère que l’on notera $(k,\chi)$. Dans ce cas, on appellera conducteur de $\kappa$, noté ${\textrm{cond}}(\chi)$, le plus petit entier naturel $r$ tel que $\chi$ soit trivial sur $1+p^r{\mathbb{Z}}_p$ .
On aura besoin en \[quat\] d’introduire pour $r \geq 1$, la boule rigide ${\mathcal{W}}_r \subset {\mathcal{W}}$: $${\mathcal{W}}_r({\mathbb{C}}_p):=\{ \kappa \in {\mathcal{W}}({\mathbb{C}}_p), \, \,
|\kappa-1|< p^{-\frac{1}{p^{r-1}(p-1)}} \}$$ On a par exemple $(1+p)^k\zeta \in
{\mathcal{W}}_r$ quand $\zeta^{p^{r-1}}=1$. Tout $\kappa$ dans ${\mathcal{W}}_r({\mathbb{C}}_p)$ est un caractère de $1+p{\mathbb{Z}}_p$ de restriction analytique à $1+p^r{\mathbb{Z}}_p$. Mieux, si $V \subset {\mathcal{W}}_r$ est ouvert affinoide, la restriction de $\kappa^{univ}$ à $A(V)$ est un caractère $A(V)$-valué de $1+p{\mathbb{Z}}_p$ dont la restriction à $1+p^r{\mathbb{Z}}_p$ est analytique.
On note $\tau: ({\mathbb{Z}}/p{\mathbb{Z}})^* \rightarrow {\mathbb{Z}}_p^*$ le caractère de Teichmüller. On verra en général les caractères de $1+p{\mathbb{Z}}_p$ comme des caractères de ${\mathbb{Z}}_p^*$, en les étendant trivialement sur les racines de l’unité.
Formes modulaires $p$-adiques {#formes}
=============================
Nous allons dans ce qui suit rappeler les acteurs essentiels de ce papier. On fixe un nombre premier $p\geq 5$, des entiers $N$ et $d$, $(N,p)=(N,d)=(p,d)=1$, $d$ sans facteur carré.
Soit $\mathcal{H}$ la ${\mathbb{Z}}$-algèbre commutative de polynômes sur les lettres $S_l$, $T_l$, si $l$ est premier et $(l,Ndp)=1$, et $U_l$ si $l$ premier divise $Ndp$. On fixe $\varepsilon=\varepsilon_p\varepsilon_N$ un caractère de $({\mathbb{Z}}/p{\mathbb{Z}}\times
{\mathbb{Z}}/N{\mathbb{Z}})^*$. On notera aussi, si $n \geq 1$, $T_n$ l’élément de ${\mathcal{H}}$ obtenu par les formules usuelles:
$$\sum_{n\geq 1} T_n n^{-s} =\prod_{l|Npd} (1-U_ll^{-s})^{-1}\prod_{l
\nmid Npd} (1-T_ll^{-s}+l S_ll^{-2s})^{-1}$$
Formes modulaires surconvergentes
---------------------------------
${}^{}$
On fait des rappels sur certaines constructions faites dans [@eigen] §2.1, [@BMF] §4.3 .
Soit $X_1(Np,d)$ la courbe propre et plate sur ${\mathbb{Z}}_p$ paramétrant les courbes elliptiques généralisées avec structure de niveau de type $\Gamma_1(Np)\cap\Gamma_0(d)$, $\omega$ le faisceau inversible habituel sur $X_1(Np,d)$. Si $p>3$, on rappelle que la série d’Eisenstein normalisée de niveau $1$, $E_{p-1}$, est une section globale de $\omega^{p-1}$ sur $X_1(Np,d)/{\mathbb{Z}}_p$ qui relève l’invariant de Hasse. Si $0\leq v<1 \in {\mathbb{Q}}$, on définit $X_1(Np,d)(v)$ comme étant l’ouvert affinoide de $X_1(Np,d)^{rig}$ sur lequel $|E_{p-1}|\geq p^{-v}$, et $Z_1(Np,d)(v)$ la composante connexe affinoide de $\infty$ dans $X_1(Np,d)(v)$ ([@eigen] §2.1, §4.3). Pour tout $m \geq 1$, on s’intéressera plus généralement ([@eigen] §2) à la courbe rigide analytique (lisse irréductible) $Z_1(Np^m,d)(v)$, avec $0\leq v <
1$, qui est la composante connexe affinoide de $\infty$ dans $X_1(Np^m,d)(v)$. $$M_k(p^m,d,v):=H^0(Z_1(Np^m,d)(v),\omega^k)$$ est le ${\mathbb{C}}_p$-espace de Banach des formes modulaires $v$-surconvergentes de poids $k$, de niveau $\Gamma_1(Np^m)\cap \Gamma_0(d)$, il est muni d’une action naturelle de ${\mathcal{H}}$ ([@eigen] §3.2, [@BMF] B5). On fixe ici, et pour tout le texte, une suite réelle décroissante $(v_n)_{n\in {\mathbb{N}}}$, telle que $\forall n \in {\mathbb{N}}, \, v_n=p^{-n}v_0 \in [\frac{p}{p^{n+1}(p+1)},\frac{p}{p^{n}(p+1)}[\cap
{\mathbb{Q}}$. La construction de nos systèmes de modules de Banach dépend de $v_0$ d’une manière qui nous importera peu, pour alléger les notations nous omettrons cette dépendance dans ce qui suit. On pourrait fixer $v_0=\frac{1}{p+1}$.
Si $\kappa=(k,\chi) \in {\mathcal{W}}({\mathbb{C}}_p)$ est de conducteur $m$, on dispose d’une série d’Eisenstein $E_{\kappa}$, qui est une forme modulaire de poids $k$ sur $X_1(Np^m,d)(v)$ de caractère trivial hors de $p$, $\chi \tau^{-k}$ en $p$. Si $0\leq v<\frac{p}{p^{m-1}p+1}$, $Z_1(Np,d)(v)$ s’identifie canoniquement au quotient de $Z_1(Np^m,d)(v)$ par l’action des diamants de $(1+p{\mathbb{Z}}_p)/(1+p^m{\mathbb{Z}}_p)$, permettant de voir $M_0(p,d,v)$ comme le sous-espace de $M_0(p^m,d,v)$ de caractère trivial sous $(1+p{\mathbb{Z}}_p)/(1+p^m{\mathbb{Z}}_p)$. La multiplication par $E_{\kappa}$ est un isomorphisme de ${\mathbb{C}}_p$-Banach de $M_0(p^m,d,v)$ sur $M_k(p^m,d,v)$ multipliant le caractère en $p$ par $\chi \tau^{-k}$. Le $q$-développement à l’infini de $E_{\kappa}$ est la spécialisation en $\kappa$ d’un $q$-développement “abstrait” ${\mathbb{E}}(q) \in 1+q\Lambda[[q]]$ appelé famille d’Eisenstein restreinte ([@eigen], §2.2).
On pose $F:=(F(V,n))$, le système de modules de Banach sur ${\mathcal{W}}$ associé au système d’espaces de Banach $\{ A(Z_1(Np,d)(v_n)), {\textrm{res}}_n, n\in {\mathbb{N}}\}$, les applications $${\textrm{res}}_n: A(Z_1(Np,d)(v_n)) \rightarrow
A(Z_1(Np,d)(v_{n+1}))$$ étant les restrictions compactes naturelles. Par définition, si $V \subset {\mathcal{W}}$ est ouvert affinoide, $F(V,n)=A(V \times
Z_1(Np,d)(v_n))$. $F$ est de manière naturelle une représentation de ${\mathcal{H}}$, $U_p \in {\textrm{Comp}}(F)$ ([@eigen] 3.2, [@BMF] B5). Si $\kappa=(k,\chi) \in {\mathcal{W}}({\mathbb{C}}_p)$ est de conducteur $m$, $n\geq m-1$, $F_{\kappa,n}$ s’identifie, comme représentation de ${\mathcal{H}}$, au sous-espace de $M_k(p^m,d,v_n)$ de caractère $\chi
\tau^{-k}$ sous $({\mathbb{Z}}/p^m{\mathbb{Z}})^*$.
En chaque pointe $c$ dans $Z_1(Np,d)(0)$, la théorie des courbes de Tate (voir [@K] A1.2) nous fournit un $q_c$-développement: $F(V,n) \hookrightarrow
A(V)[[q_c]]$. De plus, $({\mathbb{Z}}/Np{\mathbb{Z}})^*$ agit par automorphismes naturels sur $X_1(Np,d)$ en préservant $Z_1(Np,d)(v)$ et donc l’ensemble de ses pointes. Notons $A(Z_1(Np^m,d)(v_n))^{0,\varepsilon}$ le sous-espace de $A(Z_1(Np^m,d)(v_n))$ composé des fonctions de $q_c$-développement nul pour tout $c$ dans $Z_1(Np,d)(0)$, et sur lequel $({\mathbb{Z}}/Np{\mathbb{Z}})^*$ agit par le caractère $\varepsilon$ fixé plus haut. On s’intéresse au système de modules de Banach $F^{0,\varepsilon}$ associé au système d’espaces de Banach $$\{ A(Z_1(Np,d)(v_n))^{0,\varepsilon},
{{\textrm{res}}_n}_{|A(Z_1(Np,d)(v_n))^{0,\varepsilon}}, n \in {\mathbb{N}}\}$$ C’est un sous-système de modules de Banach de $F$, préservé par ${\mathcal{H}}$. Son évaluation $F^{0,\varepsilon}_{\kappa,n}$ en $\kappa=(k,\chi) \in
{\mathcal{W}}({\mathbb{C}}_p)$ de conducteur $m$ tel que $n \geq m-1$ s’identifie au sous-espace de $M_k(p^m,d,v_n)$ composé des formes de caractère $\varepsilon \tau^{-k} \chi $ s’annulant aux pointes de $Z_1(Np^m,d)(0)$ (noter que $E_\kappa$ est non nulle en chacune de ces pointes).
On définit dans ce qui suit le sous-module de Banach de $F^{0,\varepsilon}$ composé des familles de formes surconvergentes “nouvelles en $d$” (voir aussi [@BMF] B5). Soit $l$ premier divisant $d$, $ld'=d$, on dispose de morphismes canoniques finis et plats $$\pi_l: X_1(Np^m,d) \rightarrow X_1(Np^m,d')$$ oubliant le sous-groupe d’ordre $l$. Ces morphismes induisent des morphismes rigides analytiques $\pi_{l}: Z_1(Np^m,d)(v) \rightarrow Z_1(Np^m,d')(v)$, finis et plats de degré $l+1$ (étale hors des pointes). L’isomorphisme canonique $\pi_l^*(\omega)\simeq \omega$ permet d’en considérer la trace, ${\pi_l}_*\omega \rightarrow \omega$, d’où en particulier sur les sections sur $Z_1(Np^m,d')(v)$: $$tr(\pi_l)_k: M_k(p^m,d,v) \rightarrow M_k(p^m,d',v)$$
On étend $tr(\pi_l)_0: A(Z_1(Np,d)(v)) \rightarrow
A(Z_1(Np,d')(v))$, linéairement en un $A(V)$-morphisme $${\textrm{Tr}}_l: A(V\times Z_1(Np,d)(v)) \rightarrow A(V \times Z_1(Np,d')(v))$$
${\textrm{Tr}}_l$ définit ainsi un endomorphisme du module de Banach $F^{0,\varepsilon}$, il est bien défini pour tous les couples $(V,n)$.
Soit $\kappa=(k,\chi) \in {\mathcal{W}}({\mathbb{C}}_p)$, $m=\textrm{cond}(\kappa)$, $n\geq m-1$, le diagramme suivant est commutatif:
$$\xymatrix{ F_{\kappa,n} \ar@{->}[d]_{{\textrm{Tr}}_l} \ar@{->}[r]^{\hspace{- 1 cm}\times E_{\kappa}}
& M_k(p^m,d,v_n) \ar@{->}[d] \ar@{->}[d]^{tr(\pi_l)_k} \\
F_{\kappa,n} \ar@{->}[r]^{\hspace{-1 cm} \times E_{\kappa}}
& M_k(p^m,d',v_n) }$$
[*Preuve:*]{} Il faut montrer que si $f \in M_0(p^m,d,v)$, $tr(\pi_l)_k(E_{\kappa}f)=E_{\kappa}tr(\pi_l)_0(f)$, ce qui découle immédiatement de ce que $E_{\kappa}$ est de niveau premier à $l$. $\square$
On vérifie aisément que ${\textrm{Tr}}_l$ commute aux correspondances de Hecke hors de $l$, aux opérateurs diamants, et que ${\textrm{Tr}}_l^2=(l+1){\textrm{Tr}}_l$ (cf. [@BMF] B5.1). Notons $W_l$ l’involution d’Atkin sur $X_1(Np^m,d)/{\mathbb{Z}}_p$ définie modulairement par $(E,H,\alpha) \mapsto (E/H,E[l]/H,\pi\cdot \alpha)$, $H$ étant un sous-groupe d’ordre $l$ de $E$, $\pi$ l’isogénie $E \rightarrow E/H$, et $\alpha$ une structure de niveau de type $\Gamma_1(Np^m)\cap
\Gamma_0(d')$. Elle préserve les $Z_1(Np^m,d)(v)$ et induit par extension des scalaires une involution encore notée $W_l$ sur les $A(Z_1(Np,d)(v)\times
V)$.
[**Définition:**]{} Soit $\kappa \in {\mathcal{W}}({\mathbb{C}}_p)$, $f \in
F_{\kappa,n}^{0,\varepsilon}$ est dite nouvelle en $d$ si pour tout $l$ divisant $d$, ${\textrm{Tr}}_l(f)={\textrm{Tr}}_l(W_l(f))=0$.
On notera $F^{0,\varepsilon,d}$ le système de modules de Banach sur ${\mathcal{W}}$ associé au système d’espaces de Banach $ \{
A(Z_1(Np,d)(v_n))^{0,\varepsilon,d}, res_n , n \in {\mathbb{N}}\}$, où $$A(Z_1(Np,d)(v_n))^{0,\varepsilon,d}:=A(Z_1(Np,d)(v_n))^{0,\varepsilon}\cap (\bigcap_{l|d} {\textrm{Ker}}({\textrm{Tr}}_l)\cap
{\textrm{Ker}}({\textrm{Tr}}_l \cdot
W_l))$$
$F^{0,\varepsilon,d}$ est un sous-système de modules de Banach de $F^{0,\varepsilon}$ stable par l’action de ${\mathcal{H}}$.
[*Preuve*]{}: Ça n’est pas complètement évident en ce qui concerne les $U_l$ avec $l|d$. Il suffit de vérifier que pour $\kappa=(k,\chi) \in {\mathcal{W}}({\mathbb{C}}_p)$, $m={\textrm{cond}}(\kappa)$, $n \geq m-1$, $U_l$ préserve $F_{\kappa,n}^{0,\varepsilon,d}$ dans $F_{\kappa,n}^{0,\varepsilon}$. Sur $M_k(p^m,d,v_n)$, on dispose d’un endomorphisme $W_{l,k}$ défini par $W_{l,k}(f)(E,H,\alpha,\omega)=f(E/H,E[l]/H,\pi.\alpha,
(\pi^{\vee})^*(\omega) )$, avec les notations évidentes. On vérifie que sur le sous-espace de caractère $\varepsilon\chi\tau^{-k}$, on a $W_{l,k}^2=\varepsilon(l)\kappa(l)$ et $$tr(\pi_l)_k(f)=f+l\varepsilon(l)^{-1}\kappa(l)^{-1}U_l(W_{l,k}(f)), \, \,
tr(\pi_l)_k(W_{l,k}(f))=f+l\cdot U_l(f)$$ De plus, si $g \in
F_{\kappa,n}^{0,\varepsilon}$ et $e_l$ désigne la fonction inversible sur $X_1(Np,d)(v_n)$ de $q$-développement $\frac{E_{\kappa}(q^l)}{E_{\kappa}(q)}$, alors un calcul montre que $$E_{\kappa}^{-1}W_{l,k}(gE_{\kappa})=e_lW_l(g)$$ En particulier, la multiplication par $E_{\kappa}$ envoie $F_{\kappa,n}^{0,\varepsilon,d}$ isomorphiquement sur l’intersection des ${\textrm{Ker}}({\textrm{tr}}(\pi_l)_k)\cap
{\textrm{Ker}}({\textrm{tr}}(\pi_l)_k\cdot W_{l,k})$ dans le sous-espace de $M_k(p^m,d,v_n)$ de caractère $\varepsilon\chi\tau^{-k}$. Les relations ci-dessus montrent que ce sous-espace est stable par $W_{l,k}$ puis par $U_l$, qui coïncide avec $-l^{-1}W_{l,k}$ sur ce dernier. $\square$
Si $\kappa=(k,\chi) \in {\mathcal{W}}({\mathbb{C}}_p)$, $n\geq {\textrm{cond}}(\chi)-1$, $F^{0,\varepsilon,d}_{\kappa,n}$ est ${\mathbb{C}}_p\otimes_{{\mathbb{Z}}}{\mathcal{H}}$-isomorphe au sous-espace de Banach de $M_k(p^m,d,v_n)$ constitué des formes s’annulant aux pointes de $Z_1(Np^m,d)(0)$, de caractère $\varepsilon \chi \tau^{-k}$, et annulées par les ${\textrm{tr}}(\pi_l)_k$ et ${\textrm{tr}}(\pi_l)_k\cdot W_{l,k}$. Ces formes seront appelées [*nouvelles en $d$*]{}. Sur le sous-espace de $M_k(p^m,d,v_n)$ composé des restrictions à $Z_1(Np^m,d)(v_n)$ des formes convergentes sur tout $X_1(Np^m,d)$, cette condition d’être nouvelle en $d$ est précisément la condition usuelle.
[*Remarques*]{}: i) Soit $\kappa=(k,\chi) \in {\mathcal{W}}({\mathbb{C}}_p)$ de conducteur $m$, $r \in {\mathbb{N}}$, $f \in
F_{\kappa,r}^{0,\varepsilon,d}$ propre pour tout ${\mathcal{H}}$, on note $\chi: {\mathcal{H}}\rightarrow {\mathbb{C}}_p$ le caractère défini par $T_n(f):=\chi(T_n)f$. $f$ a un $q$-développement sur la pointe $\infty$ de la forme $\sum_{n\geq 1} a_n q^n$ tel que $a_n=\chi(T_n)a_1$. Le principe du $q$-développement (i.e l’irréductibilité de $Z_1(Np^m,d)(0)$) montre alors que $a_1 \neq 0$, et donc que l’on peut supposer $a_1=1$. Ceci définit donc une bijection entre caractères de ${\mathcal{H}}$ dans $F_{\kappa,r}^{0,\varepsilon,d}$ et les éléments de ce dernier qui sont propres et de $q$-développement normalisé à $1$ (“multiplicité $1$ faible”).
ii\) Les ${\mathbb{C}}_p$-espaces de Banach $A(Z_1(Np,d)(v))$, ainsi que tous leurs sous-espaces considérés dans ce paragraphe, sont orthonormalisables. En effet, d’après [@Ser] 1.1, tout espace de Banach provenant par extension des scalairs d’un espace de Banach sur un corps local est orthonormalisable. De plus, par [@Ser] 1.2, les inclusions $A(Z_1(Np,d)(v_n)) \supset
A(Z_1(Np,d)(v_n))^{0,\varepsilon} \supset
A(Z_1(Np,d)(v_n))^{0,\varepsilon,d}$ sont scindées dans la catégorie des ${\mathbb{C}}_p$-espaces de Banach.
Formes modulaires quaternioniques $p$-adiques
---------------------------------------------
${}^{}$
Dans ce qui suit, nous nous référerons à [@Buz2].
### Séries principales $p$-adiques de l’Iwahori {#rep}
Introduisons tout d’abord quelques notations de théorie des groupes: $$L \textrm{ désigne le Borel inférieur de } GL_2({\mathbb{Q}}_p)$$ $$N \textrm{ les unipotents supérieurs de } GL_2({\mathbb{Z}}_p)$$ $$I(m) \textrm{ le sous-groupe de $GL_2({\mathbb{Z}}_p)$ composé des éléments triangulaires
supérieurs modulo $p^{m}$ }$$ $$I:=I(0) \textrm{ l'Iwahori, }\, \, \, u:={\textrm{diag}}(1,p)$$ $${\mathbb{M}}(m) \textrm{ est le sous-monoide de $GL_2({\mathbb{Q}}_p)\cap M_2({\mathbb{Z}}_p)$ engendré par
$I(m)$ et $u$}, \, \, \, {\mathbb{M}}:={\mathbb{M}}(1)$$
La décomposition d’Iwahori s’écrit $I=(L\cap I)\times N$. On identifie $N$ à ${\mathbb{Z}}_p$ via $$\left( \begin{array}{cc} 1 & t \\ 0 & 1 \end{array} \right) \mapsto t$$ La grosse cellule de Bruhat $L\backslash LI \subset
L\backslash GL_2({\mathbb{Q}}_p)={\mathbb{P}}^1({\mathbb{Q}}_p)$ est stable par multiplication à droite par ${\mathbb{M}}$. On note $T$ la coordonnée sur ${\mathbb{Z}}_p$, l’action de ${\mathbb{M}}$ par translation à droite sur les fonctions sur $L\backslash
LI ={\mathbb{Z}}_p$ s’écrit alors $$\left( \begin{array}{cc} a & b \\ c & d \end{array} \right).T = \frac{b+dT}{a+cT}$$
Soit $n \in {\mathbb{N}}$, on note ${\mathcal{C}}_{n}$ la ${\mathbb{C}}_p$-algèbre de Banach des fonctions sur ${\mathbb{Z}}_p$ de restriction analytique aux $a+p^n{\mathbb{Z}}_p$, ${\mathcal{C}}_n$ est stable par l’action de ${\mathbb{M}}$ et ${\mathcal{C}}_0$ s’identifie à l’algèbre de Tate ${\mathbb{C}}_p\!\!<\!T\!>$. Les restrictions naturelles $i_n: {\mathcal{C}}_n \rightarrow
{\mathcal{C}}_{n+1}$ sont compactes injectives , ce qui fait de ${\mathcal{C}}=\{{\mathcal{C}}_n,i_n, n\in {\mathbb{N}}\}$ un système d’espaces de Banach tel que $u \in
{\textrm{Comp}}({\mathcal{C}})$. On note ${\mathcal{C}}_{{\mathcal{W}}}$ le système de modules de Banach sur ${\mathcal{W}}$ associé à ${\mathcal{C}}$. Il sera commode de poser ${\mathcal{C}}_{-n}:={\mathcal{C}}_0$ et ${\mathcal{C}}(V,-n):={\mathcal{C}}(V,0)$ si $n \in
{\mathbb{N}}$, $V \subset {\mathcal{W}}$ ouvert affinoide.
Soit $j: I \rightarrow {\mathcal{C}}_0^*$ le 1-cocycle défini par $$j( \left( \begin{array}{cc} a & b \\ c & d \end{array} \right) ):=
\tau^{-1}(a)(a+cT) \in 1\!+\!p{\mathbb{Z}}_p\!<\!T\!> \, \subset \, {\mathcal{C}}_0^*$$ Il se prolonge à ${\mathbb{M}}$ en le prenant trivial sur $u$. Si $\kappa \in {\mathcal{W}}({\mathbb{C}}_p)$, $\gamma \in
{\mathbb{M}}$, on définit un cocycle $\kappa(j): {\mathbb{M}}\rightarrow \bigcup_{n \in {\mathbb{N}}}
{\mathcal{C}}_{n}^*$ par $$(\kappa(j)(\gamma))(t):=\kappa(j(\gamma)(t)), \, \, \, t \in {\mathbb{Z}}_p$$ Si $\gamma \in {\mathbb{M}}(m)$ et $\kappa \in {\mathcal{W}}_r({\mathbb{C}}_p)$, $\kappa(j)(\gamma) \in
{\mathcal{C}}_{r-m}^*$. On notera $\rho_{\kappa}$ la représentation de ${\mathbb{M}}$ sur l’espace $\bigcup_{n \in {\mathbb{N}}}{\mathcal{C}}_n$ tordue par $\kappa(j)$, i.e $\rho_{\kappa}(v):=\kappa(j)(\gamma)\gamma.v$. Cette torsion n’affecte pas $u$ et si $\kappa \in {\mathcal{W}}_r({\mathbb{C}}_p)$, ${\mathbb{M}}(m)$ préserve ${\mathcal{C}}_n$ dès que $n\geq
r-m$. Plus généralement, si $V \subset {\mathcal{W}}$ est un ouvert affinoide, on dispose d’un $1$-cocycle $\kappa^{univ}(j): {\mathbb{M}}\rightarrow
\bigcup_{n \in {\mathbb{N}}}{\mathcal{C}}(V,n)^*$, tel que $\kappa^{univ}(j)_{\kappa}=\kappa(j)$ si $\kappa \in {\mathcal{W}}({\mathbb{C}}_p)$ et $\kappa^{univ}(j)(\gamma) \in {\mathcal{C}}_{r-m}^*$ si $\gamma \in {\mathbb{M}}(m)$ et $V \subset {\mathcal{W}}_r$. On dispose donc d’une représentation $\rho^{univ}$ de ${\mathbb{M}}$ sur $\bigcup_{n \in {\mathbb{N}}} {\mathcal{C}}(V,n)$, en tordant par $\kappa^{univ}(j)$ la représentation naturelle obtenue par extension complète des scalairs à $A(V)$ de celle sur ${\mathcal{C}}_n$. Si $V \subset
{\mathcal{W}}_r$, ${\mathbb{M}}(m)$ préserve ${\mathcal{C}}(V,n)$ dès que $n \geq r-m$.
Ainsi, le système de modules de Banach ${\mathcal{C}}_{{\mathcal{W}}}$ sur ${\mathcal{W}}$ est muni d’une représentation $\rho^{univ}: {\mathbb{M}}\rightarrow
{\textrm{End}}({\mathcal{C}}_{{\mathcal{W}}})$. Si $V \subset {\mathcal{W}}_r$, ${\mathbb{M}}(m)$ préserve ${\mathcal{C}}(V,n)$ dès que $n \geq r-m$, et on a $u \in {\textrm{Comp}}({\mathcal{C}}_{{\mathcal{W}}})$. On appelle ${\mathcal{C}}_{{\mathcal{W}}}$ [*la famille analytique des séries principales $p$-adiques de $I$*]{}.
[*Remarques:*]{} i) ${\mathcal{C}}^{\dagger}=\bigcup_{n \in {\mathbb{N}}}{\mathcal{C}}_n$ est l’espace des fonctions localement analytiques sur ${\mathbb{Z}}_p$, muni de sa topologie localement convexe c’est un espace de type compact au sens de [@ST] §1 [^2]. La représentation $\rho_{\kappa}$ de $I$ sur ${\mathcal{C}}^{\dagger}$ est la série principale localement analytique de $I$ de caractère $\kappa$ ([@ST] §5, noter que leur $B$ est l’Iwahori opposé à $I$).
ii\) Les assertions de ce paragraphe sont détaillées dans [@Buz2] §4, §7, noter qu’il a des actions à droites, non à gauche. A cette modification près, on a dans ses notations: $\mathcal{A}_{\kappa,p^{-n}}={\mathcal{C}}_{\kappa,n}$ et $M_{m} \supset {\mathbb{M}}(m)$; d’autre part si $\kappa \in {\mathcal{W}}_r({\mathbb{C}}_p)\backslash {\mathcal{W}}_{r-1}({\mathbb{C}}_p)$, “$m$ is good for $(\kappa,p^{-n})$” équivaut à $n \geq r-m$.
### Formes modulaires {#quat}
Soit $D({\mathbb{Q}})$ une algèbre de quaternions sur ${\mathbb{Q}}$, on fixe $D({\mathbb{Z}})$ un ordre maximal de $D({\mathbb{Q}})$, et on note $D$ le schéma en anneaux sur ${\mathbb{Z}}$ associé, $D^*$ son groupe des inversibles. On suppose que $D$ est définie. Soit $d={\textrm{disc}}(D)$ le produit des premiers ramifiés, on fixe un isomorphisme au dessus de ${\mathbb{Z}}[1/d]$, $$\varphi: D/{\mathbb{Z}}[1/d] \simeq \mathbb{M}_2/{\mathbb{Z}}[1/d]$$ Si $M$ est un entier premier à $d$, on notera $U_1(M)$ (resp. $U_0(M)$) le sous-groupe ouvert compact de $D^*({\mathbb{A}}_f) \simeq D^*({\mathbb{A}}_{f}^{(d)}) \times \prod_{l | d}
D^*({\mathbb{Q}}_l)$, décomposé selon ce produit, valant le compact maximal $D^*({\mathbb{Z}}_l)$ en les premiers $l$ divisant $d$, égal à $\varphi^{-1}(\Gamma_1(M))$ (resp. $\varphi^{-1}(\Gamma_0(M))$) sur l’autre facteur, avec:
$$\Gamma_1(M)=\{ g \in GL_2(\widehat{{\mathbb{Z}}[1/d]}), \, \, g \equiv \left(
\begin{array}{cc} 1 & * \\ 0 & * \end{array} \right)\,\bmod M\},$$ $$\Gamma_0(M)=\{ g \in GL_2(\widehat{{\mathbb{Z}}}), \, \, g \equiv
\left( \begin{array}{cc} *
& * \\ 0 & * \end{array} \right)\,\bmod M \}$$
On verra les caractères de $({\mathbb{Z}}/M{\mathbb{Z}})^*$ comme des caractères de $U_0(M)$ par l’étoile supérieure gauche. Si $M\geq 5$, $D^*({\mathbb{Q}}) \times
U_1(M)$ agit librement sur $D^*({\mathbb{A}}_f)$, et $D^*({\mathbb{Q}})\backslash
D^*({\mathbb{A}}_f)/U_1(M)$ est fini de cardinal noté $h_1(M)$. On choisit $M=Np$ comme en \[formes\].
On note ${F}^{D}$ le système de modules de Banach sur ${\mathcal{W}}$ tel que si $V
\subset {\mathcal{W}}$ est ouvert affinoide, $n \in {\mathbb{N}}$, $F^{D}(V,n)$ est le $A(V)$-module de Banach des fonctions $f: D^*({\mathbb{Q}})\backslash D^*({\mathbb{A}}_f)\rightarrow {\mathcal{C}}(V,n)$ satisfaisant $$f(xu)=j(u_p^{-1})^{-2}\rho^{univ}(u_p^{-1})f(x), \, \, \forall (x,u) \in D^*({\mathbb{A}}_f) \times
U_1(Np)$$ et $i_n$ est l’application déduite de la restriction canonique ${\mathcal{C}}(V,n)
\rightarrow {\mathcal{C}}(V,n+1)$. L’orthonormalisabilité vient de ce que $F^D(V,n)$ s’identifie à ${\mathcal{C}}(V,n)^{h_1(Np)}$ car $Np\geq 5$ (voir aussi [@Buz2] §7, §4).
On a une représentation naturelle sur $F^D$ de $U_0(Np)/U_1(Np)\simeq
({\mathbb{Z}}/Np{\mathbb{Z}})^*$, définie par $(<\gamma>.f)(x)=j(\gamma)^{-2}\rho^{univ}(\gamma)f(x\gamma)$. On notera $F^{D,\varepsilon}$ le sous-système de modules de Banach de $F^D$ sur lequel $({\mathbb{Z}}/Np{\mathbb{Z}})^*$ agit par $\varepsilon^{-1}$. On dispose d’une représentation naturelle ${\mathcal{H}}\rightarrow {\textrm{End}}(F^{D,\varepsilon})$ telle que $U_p \in {\textrm{Comp}}(F^{D,\varepsilon})$. Les doubles classes considérées ici pour les opérateurs de Hecke sont (avec les abus évidents) les ${\textrm{diag}}(1,l)$, pour $T_l$ et $U_l$ si $l\nmid d$, ${\textrm{diag}}(l,l)$ pour $S_l$ si $l \nmid Npd$, comme en \[JLclass\] pour $U_l$ avec $l|d$. Si $\kappa \in {\mathcal{W}}({\mathbb{C}}_p)$, $F^{D,\varepsilon}_{\kappa,n}$ s’identifie à l’espace des fonctions $D^*({\mathbb{Q}})\backslash D^*({\mathbb{A}}_f)
\rightarrow {\mathcal{C}}_n$ telles que $$f(xu)=\varepsilon^{-1}(u)j(u_p^{-1})^{-2}\rho_{\kappa}(u_p^{-1})f(x), \,
\,\, \forall \, (x,u) \in D^*({\mathbb{A}}_f) \times U_0(Np)$$
$(F^{D,\varepsilon}_{\kappa})^{\dagger}$, vu avec sa structure de ${\mathcal{H}}$-module, est [*l’espace des formes modulaires $p$-adiques quaternioniques de poids-caractère $\kappa$, de niveau modéré $N$, de caractère $\varepsilon$*]{}.
Soient $n,\, m,\, r$ des entiers tels que $n\geq r-1 \geq 0$, $n\geq m-1 \geq
0$, et $\kappa \in {\mathcal{W}}_r({\mathbb{C}}_p)$, considérons l’espace annexe $F_{\kappa,n}^{D,\varepsilon}[m]$ des fonctions $f:
D^*({\mathbb{Q}})\backslash D^*({\mathbb{A}}_f)\rightarrow
{\mathcal{C}}_{\kappa,n}$ satisfaisant $$f(xu)=\varepsilon^{-1}(u)j(u_p^{-1})^{-2}\rho_{\kappa}(u_p^{-1})f(x), \, \,
\forall (x,u) \in
D^*({\mathbb{A}}_f^{(d)}) \times U_0(Np^m)$$
C’est un ${\mathcal{H}}$-module, en prenant cette fois-ci des doubles classes par rapport à $U_0(Np^m)$ (ce qui ne change que $U_p$, encore défini par la double classe de ${\textrm{diag}}(1,p)$). D’après [@Buz2] (§5 lemme 3,iv), $F_{\kappa,n-m+1}^{D,\varepsilon}[m]$ et $F_{\kappa,n}^{D,\varepsilon}$ sont isomorphes comme ${\mathcal{H}}\otimes_{{\mathbb{Z}}}{\mathbb{C}}_p$-modules. Si de plus $\kappa=(k,\chi)$ est de conducteur $m$, et $k\geq 2$, $F_{\kappa,0}^{D,\varepsilon}[m]$ contient comme sous-${\mathcal{H}}\otimes_{{\mathbb{Z}}}{\mathbb{C}}_p$-module l’espace des fonctions à valeurs polynomiales en $T$ de degré $\leq k-2$, ce dernier est isomorphe comme ${\mathcal{H}}\otimes_{{\mathbb{Z}}} {\mathbb{C}}_p$-module à l’espace des fonctions $f: D^*({\mathbb{Q}})\backslash D^*({\mathbb{A}}_f)\rightarrow \textrm{Sym}^{k-2}({\mathbb{C}}_p^2) $ telles que $$f(xu)=\varepsilon(u)^{-1}\chi(u_p)^{-1}\tau^{k}(u_p)u_p^{-1}f(x), \, \, \forall
\, \, (x,u) \, \, \in \, \, D^*({\mathbb{A}}_f) \times
U_0(Np^m)$$
Ce ${\mathcal{H}}\otimes_{{\mathbb{Z}}} {\mathbb{C}}_p$ module est $S_{k}^D(Np^m,\varepsilon\chi\tau^{-k},{\mathbb{C}}_p)$ dans la notation du §\[pstructure\].
Préliminaires de théorie spectrale
==================================
Semi-simplification en dimension infinie {#semi}
----------------------------------------
Soit $K$ un corps valué complet non archimédien (non discret), $V$ un $K$-espace de Banach orthonormalisable, $U$ un endomorphisme compact de $V$. La série caractéristique $P(T)=\det(1-TU)$ de $U$ se décompose sous la forme $$P(T)=\prod_{i \geq 0}
P_i(T)^{n_i}$$ où les $P_i(T)$ sont des irréductibles de $1+TK[T]$ deux à deux distincts tels que $|P_i(T)-1|
\underset{i \rightarrow \infty}{\rightarrow} 0$ pour la norme $|.|$ sur $K[T]$ du sup des coefficients. Par [@Ser], on sait que $V$ est somme directe topologique de $Ker(U)$ et des espaces de dimension finie $V(P_i):={\textrm{Ker}}(P^*_i(u)^{n_i})$, $Q^*(T)$ désignant le polynôme réciproque de $Q(T)$, sur lesquels $U$ a pour polynôme caractéristique $P^*_i(T)^{n_i}$. Soit $H$ une $K$-algèbre, $\rho: H \rightarrow {\textrm{End}}_K(V)$ une représentation telle que $\rho(H)$ contienne $U$ et lui commute, alors $\rho(H)$ stabilise les $V(P_i)$, que l’on peut semi-simplifier.
[**Définition:**]{} On notera $\mathcal{X}_U(V)$ l’ensemble des représentations irréductibles de $H$ apparaissant dans la réunion des semi-simplifications des $V(P_i)$, comptées avec multiplicité (qui sont nécessairement finies).
Notons que $\mathcal{X}_U(V)$ dépend de $U$, mais que $\mathcal{X}_U(V)=\mathcal{X}_{U'}(V)$ si $U' \in \rho(H)$ est un autre endomorphisme compact de $V$ commutant à $\rho(H)$ et à $U$ tel que ${\textrm{Ker}}(U')={\textrm{Ker}}(U)$. Si $\mathcal{X}$ est un ensemble de représentations d’une algèbre, on notera $|\mathcal{X}|$ l’ensemble des classes d’isomorphie de représentations apparaissant dans $\mathcal{X}$ (autrement dit, on oublie les multiplicités).
\[semisimpl\] Soient $(\rho_1,V_1)$ et $(\rho_2,V_2)$, des représentations d’une $K$-algèbre $H$ dans les endomorphismes continus de $V_1$ et $V_2$, telles que $V_i$ ($i=1,2$) est muni d’un endomorphisme compact $U_i$ commutant à $\rho_i(H)$.
On suppose de plus que pour tout $h \in H$, $\det(1-T\rho_1(h)U_1)=\det(1-T\rho_2(h)U_2) \in K\{\{T\}\}$. Alors $\mathcal{X}_{U_1}(V_1)=\mathcal{X}_{U_2}(V_2)$.
[*Preuve:*]{} Soit $\alpha \in {\mathbb{R}}$, $i \in \{1,2\}$, $V_i^{\alpha}$ le plus grand sous-espace de dimension finie de $V_i$ stable par $U_i$ sur lequel le polygone de Newton du polynôme caractéristique de $U_i$ est de pente $\alpha$. Soit $h \in H$, $\rho_i(h)$ stabilise les $V_i^{\alpha}$, et ses valeurs propres sur ces derniers sont toutes bornées par $||\rho_i(h)||< \infty$, $||.||$ désignant la norme d’opérateur sur $V_i$. Soit $x \in K^*$ tel que $|x|<1$, on peut trouver par conséquent un $N \in {\mathbb{N}}$ tel que $\rho_i(1+x^Nh)$ ait toutes ses valeurs propres de norme $1$ sur les $V_i^{\alpha}$ ($i=1,2$). On pose $h'=1+x^Nh \in H$. En co-trigonalisant $U_i$ et $\rho_i(h')$ sur $V_i^{\alpha}$, on en déduit: $$\det(1-TU_i\rho_i(h'))^{\alpha}=\det(1-TU_i\rho_i(h')_{|V_i^{\alpha}})$$ Si $Q \in 1+TK\{\{T\}\}$, $Q^{\alpha}$ désigne le polynôme $\in 1+TK[T]$ divisant $Q$ tel que le polygone de Newton de $Q/Q^{\alpha}$ n’a pas la pente $\alpha$. Ainsi, $\det(1-T\rho_1(h')U_1)=\det(1-T\rho_2(h')U_2)$ donne $$\det(1-T\rho_1(h'){U_1}_{|V_1^{\alpha}})= \det(1-T\rho_2(h'){U_2}_{|V_2^{\alpha}})$$ $\rho_i(h)U_i$ étant injectif sur $V_i^{\alpha}$, notons que cela implique que $\dim(V_1^{\alpha})=\dim(V_2^{\alpha})$, et que $\rho_1(h'){U_1}_{|V_1^{\alpha}}$ et $\rho_2(h'){U_2}_{|V_2^{\alpha}}$ ont même polynôme caractéristique. Ceci reste vrai pour les même raisons en remplaçant $h'$ par $h'+\lambda$ pour (une infinité de) $\lambda \in K$ assez petit. Si $A$ et $B$ sont deux endomorphismes qui commutent d’un $K$-espace vectoriel de dimension finie $r$, avec $A$ inversible, la donnée de $$\det(X.1-A(B+Y.1))=\prod_{i=1}^r(X-a_iY-b_i) \in \overline{K}[X,Y]$$ permet de retrouver $\det(X.1-B)=\prod_{i=1}^r(X-(b_i/a_i))$ (ici $\overline{K}$ est une clôture algébrique de $K$). On en déduit $\det(T-\rho_1(h')_{|V_1^{\alpha}})=\det(T-\rho_2(h')_{|V_2^{\alpha}})$, puis la même chose en remplaçant $h'$ par $h$. Ainsi, $V_1^{\alpha}$ et $V_2^{\alpha}$ sont deux représentations de $H$ ayant même polynômes caractéristiques, on sait alors que leurs semi-simplifications sont isomorphes, ce qui conclut. $\square$
Terminons cette partie par une légère amélioration de \[semisimpl\]. Soient $E^i$, $i=1,2$, deux systèmes d’espaces de Banach, munis de représentations $\rho_i: H \rightarrow \textrm{End}(E^i)$ et d’endomorphismes compacts $U_i \in \textrm{Comp}(E^i)$, $U_i \in
\rho_i(H)$ commutant à $\rho_i(H)$. $E^{i,\dagger}$ est muni d’une opération de $\rho_i(H)$ et $U_i$. Si $\alpha \in {\mathbb{R}}$ est fixé et $E_n^{i,\alpha}(U_i)$ désigne le sous-espace de $E^i_n$ sur lequel $U_i$ est de pente $\alpha$, $i_n$ induit pour tout $n$ assez grand une bijection $E_n^{i,\alpha}(U_i) \rightarrow E_{n+1}^{i,\alpha}(U_i)$ qui commute à $U_i$. Cet espace définit donc un sous-espace de dimension finie $E^{i,\alpha}(U_i)$ de $E^{i,\dagger}$ qui hérite d’une représentation de $H$. Ceci permet de définir à nouveau $\mathcal{X}_{U_i}(E^i)$ comme étant l’ensemble des représentations (comptées avec multiplicités) de $H$ apparaissant dans les semi-simplifications des $E^{i,\alpha}(U_i)$, $\alpha$ variant dans ${\mathbb{R}}$.
\[corsemisimpl\] Sous ces hypothèses, supposons que pour tout $h \in H$, $\det(1-T\rho_1(h)U_1)=\det(1-T\rho_2(h)U_2) \in K\{\{T\}\}$. Alors $\mathcal{X}_{U_1}(E^1)=\mathcal{X}_{U_2}(E^2)$.
[*Preuve:*]{} Fixons $\alpha \in {\mathbb{R}}$, $E^{i,\alpha}(U_i)$ est de dimension finie, $H$ agit donc sur $E^{1,\alpha}(U_1)\oplus E^{2,\alpha}(U_2)$ à travers un quotient de dimension finie. Considérons une sous-$K$-algèbre $H_0$ de $H$ de type fini sur $K$ engendrant toute l’image de $H$ dans ce quotient; pour $n$ assez grand, $H_0$ et $U_i$ agissent alors par endomorphismes sur $E^i_n$, $i=1,2$. On applique la proposition \[semisimpl\] avec $V_i=E^i_n$, il vient $\mathcal{X}_{U_1}(E^{1,\alpha}(U_1))=\mathcal{X}_{U_2}(E^{2,\alpha}(U_2))$, ce qui conclut. $\square$
Un critère de densité {#critere}
---------------------
On fixe $M$ un système de modules de Banach comme en §\[banachsystem\], $\textrm{dim}({\mathcal{W}})>0$, muni d’une action d’une ${\mathbb{C}}_p$-algèbre $H$, et d’un endomorphisme compact $U$ lui commutant, on notera $(M,H,U)$ une telle donnée. On rappelle que si $T/{\mathbb{C}}_p$ est un espace rigide, un sous-ensemble $X\subset
T({\mathbb{C}}_p)$ est dit Zariski-dense si pour tout fermé analytique $F$ de $T$ tel que $X \subset F({\mathbb{C}}_p)$, alors $F({\mathbb{C}}_p)=T({\mathbb{C}}_p)$. Soit $X \subset {\mathcal{W}}({\mathbb{C}}_p)$ un sous-ensemble Zariski-dense, tel que pour tout $x \in X$ et tout voisinage ouvert affinoide irréductible $V$ de $x$ dans ${\mathcal{W}}$, $V({\mathbb{C}}_p) \cap X$ est Zariski-dense dans $V$. On dira alors que $X$ est [**très Zariski-dense dans ${\mathcal{W}}$**]{}. C’est par exemple le cas des points de la forme $\zeta(1+p)^k$, avec $\zeta^{p^m}=1$ et $k,m \in {\mathbb{N}}$, dans la boule ouverte de centre $1$ de rayon $1$ de ${\mathbb{C}}_p$.
On se fixe de plus une “structure classique sur $X$”, on entendra par là la donnée pour tout $x \in X$ d’un sous-espace vectoriel de dimension finie $M_x^{cl}$ de $M_x^{\dagger}$ stable par l’action de $H$. Soit $\alpha \in {\mathbb{R}}$, on notera $M_{x}^{\leq \alpha}$ (resp. $M_x^{\alpha}$) le sous-espace de dimension finie de $M_x^{\dagger}$ sur lequel $U$ est de pente au plus $\alpha$ (resp. exactement $\alpha$). On fera de plus l’hypothèse de “contrôle”:
\[Cl\] (Cl)
\[det\] Soient $(M_1,H,U)$ et $(M_2,H,U)$ deux systèmes de modules de Banach sur ${\mathcal{W}}$ relativement factoriel. On se donne $X \subset {\mathcal{W}}({\mathbb{C}}_p)$ un ensemble très Zariski-dense, et une structure très classique sur $X$ pour $M_1$ et $M_2$, chacune de ces structures satisfaisant (Cl). Soit $h \in H$, supposons $$\forall x \in X, \, \, \, \, \, \det(1-ThU_{|M_{1,x}^{cl}})=\det(1-ThU_{|M_{2,x}^{cl}})\in {\mathbb{C}}_p[T]$$ Alors $_{M_1}(hU)$=$_{M_2}(hU)$.
[*Preuve:*]{} Soit $Z_i \subset {\mathcal{W}}\times {\mathbb{A}}^{1}$ l’hypersurface de Fredholm de $P_i$:=$_{M_i}(hU)$, $p_i$ la première projection $Z_i \rightarrow
{\mathcal{W}}$. On dira que $z \in Z_i({\mathbb{C}}_p)$ est classique si $z=(x,\lambda)$ avec $x \in X$ et si $\lambda^{-1}$ est valeur propre de $hU$ sur $M_{i,x}^{cl}$. On montrera plus bas que les points classiques sont Zariski-denses dans $Z_i({\mathbb{C}}_p)$, admettons le pour l’instant. Par hypothèse, $P_1$ s’annule sur les points classiques de $Z_2({\mathbb{C}}_p)$, donc sur $Z_2({\mathbb{C}}_p)$ par Zariski-densité. Par symétrie, il vient $Z_1^{{\textrm{red}}}=Z_2^{{\textrm{red}}}$ et on déduit de [@Con] 4.3.2 que $P_1$ et $P_2$ ont même facteurs irréductibles, il reste à prouver que ces derniers ont même multiplicités.
Soit $\Pi$ un de ces facteurs irréductibles, de multiplicité $n_i$ dans $P_i$, $Z(\Pi)_i \subset Z_i$ la composante irréductible associée. Soit $W_i$ l’ouvert de $Z(\Pi)_i$ dont le complémentaire est l’ensemble points de $Z(\Pi)_i$ qui sont dans au moins deux composantes irréductibles de $Z_i$. Admettons pour l’instant que l’on puisse trouver $x \in X$ et $z=(x,\lambda) \in
W_1({\mathbb{C}}_p)=W_2({\mathbb{C}}_p)$ tels que $\lambda$ soit une racine de $\det(1-ThU_{|M_{i,x}^{cl}})$ mais pas de $P_i(x,T)/\det(1-ThU_{|M_{i,x}^{cl}})$. Par le choix de $W_i$, $\lambda$ est une racine de $P_i(x,T)$ qui est en fait une racine de $\Pi(x,T)$ mais pas des autres facteurs irréductibles. La multiplicité de $\lambda$ comme racine de $P_i(x,T)$ est donc de la forme $n_in$ où $n$ est la multiplicité de $\lambda$ comme racine de $\Pi(x,T)$. Mais, par le choix de $z$, $nn_i$ est aussi la multiplicité de $\lambda$ comme racine de $\det(1-ThU_{|M_{i,x}^{cl}})$, qui ne dépend pas de $i$. Ainsi, $n_1n=n_2n$, puis $n_1=n_2$.
Il reste à trouver un tel $z$ et à prouver que les points classiques sont Zariski-denses dans $Z_i({\mathbb{C}}_p)$. Par hypothèse sur ${\mathcal{W}}$ et [@Con] 4.3.2, les composantes irréductibles de $Z_i$ sont des hypersurfaces de Fredholm. Ces dernières étant d’image Zariski-ouverte dans ${\mathcal{W}}$, elles contiennent toutes des points d’image (par $p_i$) dans $X$. Soit $z_i$ un de ces points, appartenant a une composante irréductible $T_i$ de $Z_i$, et soit $\Omega_i \in \mathcal{C}(Z_i)$ contenant $z_i$. $\mathcal{C}(Z_i)$ désigne le recouvrement canonique de l’hypersurface de Fredholm $Z_i$ (voir la discussion au début de \[unicite\]). $\Omega_i$ étant fini et plat sur son image $V_i \subset {\mathcal{W}}$ (que l’on peut supposer irréductible), chacune de ses composantes irréductibles se surjecte sur $V_i$. $hU$ et $U$ sont des endomorphismes de $M_i(V_i,n)$ pour un $n$ assez grand que l’on fixe, et $P_i(T)_{|V_i}=\det(1-T{hU}_{|M_i(V_i,n)})\in
1+A(V_i)T\{\{T\}\}$. Par choix de $\Omega_i
\in \mathcal{C}(Z_i)$ et un théorème de Coleman ([@eigen] §7.1, [@BMF] A.4.3), on peut trouver un sous-$A(V_i)$-module $N_i$ de $M_i(V_i,n)$ localement libre de rang fini, stable par $U$ et $hU$, tel que $\Omega_i$ soit le fermé des zéros de $\det(1-{ThU}_{|N_i})$ dans $V_i \times
{\mathbb{A}}^1$. $hU$ est un endomorphisme inversible de $N_i$, ainsi donc que $U$, et ils sont automatiquement continus ([@BGR] 3.7.3 proposition 2). En particulier, les valeurs propres de $U^{-1}$ sur les $N_{i,x}$, $x \in V_i({\mathbb{C}}_p)$, sont bornées par une constante ne dépendant que des $V_i$. Ceci et $(Cl)$ impliquent que pour tout $x \in X \cap V_i({\mathbb{C}}_p)$ sauf peut-être un nombre fini d’entre eux, $\Omega_i({\mathbb{C}}_p) \cap p_i^{-1}(\{x\})$ est composé de points classiques. $X\cap V_i({\mathbb{C}}_p)$ étant Zariski-dense dans $V_i({\mathbb{C}}_p)$, et $\Omega_i
\rightarrow V_i$ étant fini et plat, on en déduit que les points classiques sont Zariski-denses dans chaque composante irréductible de $\Omega_i$, et en particulier dans $Z_i({\mathbb{C}}_p)$ et $T_i({\mathbb{C}}_p)$ ([@Con] 2.2.3). Plus exactement, on en déduit que l’ensemble des points $z=(x,\lambda)$ tels que $M_{i,x}^{hU=\lambda^{-1}} \subset M_{i,x}^{cl}$ est Zariski-dense dans $T_i({\mathbb{C}}_p)$. $M_{i,x}^{hU=\lambda^{-1}}$ désigne ici le sous-espace de $M_{i,x}$ qui est l’espace caractéristique pour la valeur propre $\lambda^{-1}$ de l’endomorphisme compact $hU$.
On appliquant cela à $T_i:=Z(\Pi)_i$. $\Omega_i \cap W_i$ est un ouvert de $\Omega_i$ et contient donc des points du type précédent. Si $(x,\lambda)$ en est un, $\lambda$ est une racine de $\det(1-ThU_{|M_{i,x}^{cl}})$ mais pas de $P_i(x,T)/\det(1-ThU_{|M_{i,x}^{cl}})$, c’est juste ce qui nous manquait pour conclure. $\square$
La correspondance de Jacquet Langlands “$p$-adique”
===================================================
Rappels sur la correspondance classique {#JLclass}
---------------------------------------
Si $M \in {\mathbb{N}}$, on pose :
$$K_1(M)=\{ g \in GL_2(\widehat{{\mathbb{Z}}}), \, \, g \equiv \left( \begin{array}{cc} * & * \\ 0 & 1 \end{array}
\right)\,\bmod M\},$$ $$K_0(M)=\{ g \in GL_2(\widehat{{\mathbb{Z}}}), \, \, g \equiv \left( \begin{array}{cc} *
& * \\ 0 & * \end{array} \right)\,\bmod M \}$$
On pourra voir, par l’étoile inférieure droite, les caractères complexes de $({\mathbb{Z}}/M{\mathbb{Z}})^*$ comme des caractères de $K_0(M)$. On fixe $\varepsilon$ un tel caractère, ainsi qu’un entier $k \geq 2$.
Le ${\mathbb{C}}$-espace vectoriel $S_k(M,\varepsilon)$ des formes modulaires paraboliques de poids $k$, de niveau $M$, de caractère $\varepsilon: ({\mathbb{Z}}/M{\mathbb{Z}})^*
\rightarrow {\mathbb{C}}$ s’identifie à l’espace des fonctions complexes sur $GL_2({\mathbb{Q}})\backslash GL_2({\mathbb{A}})/K_1(M)$, de caractère central $|.|^{2-k}\varepsilon^{-1}$, de caractère $\varepsilon^{-1}$ sous $K_0(M)$, satisfaisant les conditions usuelles de croissance aux pointes et holomorphie, en associant à $f$ (voir par exemple [@hida] §3.1 pour des détails) [^3] $$g=(g_{\infty},g_f) \in GL_2^{+}({\mathbb{R}}) \times K_0(M) \mapsto
|\det(g)|^{1-k/2}f_{|_k g_{\infty}}(i)\varepsilon^{-1}(g_f)$$
Cette identification préserve l’action des opérateurs de Hecke usuels (non renormalisés du côté adélique). Soit $l$ premier, l’opérateur de Hecke $T_l$ si $(l,M)=1$, $U_l$ sinon, est donné par la double classe $U_0(M){\textrm{diag}}(l,1)U_0(M)$; si $(l,M)=1$, $S_l=l^{k-2}\varepsilon(l)$ est l’action de ${\textrm{diag}}(l,l)$ (ici, on voit ${\textrm{diag}}(a,b) \in GL_2({\mathbb{A}})$ partout trivial sauf en $l$ où il vaut effectivement ${\textrm{diag}}(a,b)$).
On notera $S_k(M,\varepsilon)^{d-new}$ le sous-espace de $S_k(M,\varepsilon)$ composé des formes $d$-nouvelles au sens usuel.
On fixe un entier $d$ sans facteur carré, ayant un nombre impair de diviseurs premiers, tel que $d{\, |\!| \,}M$ et que $\varepsilon$ soit trivial sur $({\mathbb{Z}}/d{\mathbb{Z}})^*$. On reconsidère $D({\mathbb{Q}})$ l’algèbre de quaternions de discriminant $d$ introduite en \[quat\]. $K_1(M)$ (resp. $K_0(M)$) est le compact ouvert de $D^*({\mathbb{A}}_f)$ décomposé place par place, qui vaut $D({\mathbb{Z}}_p)^*$ aux places $p|d$, $\varphi^{-1}(K_1(M))$ (resp. $\varphi^{-1}(K_0(M))$) hors de $d$ (cf. \[quat\]), noter que $K_1(M) \neq U_1(M)$.
On note $S_k^{*,D}(M,\varepsilon)$ le ${\mathbb{C}}$-espace vectoriel des fonctions complexes sur $D^*({\mathbb{Q}})\backslash D^*({\mathbb{A}})/K_1(M)$, de caractère $\varepsilon^{-1}$ sous $K_0(M)$, de caractère central $|\det(g)|^{2-k} \varepsilon^{-1}$, engendrant sous $D^*({\mathbb{R}})$ un multiple du dual de la représentation $\textrm{Sym}^{k-2}({\mathbb{C}}^2)$. Si $k=2$ et $\varepsilon=1$, on note $S$ la droite des fonctions constantes sur $D^*({\mathbb{A}})$ dans $S_2^{*,D}(M,1)$, elle est stable par $D^*({\mathbb{A}})$.
(Arthur, Jacquet, Langlands) \[JL\] Si $k>2$ ou $\varepsilon\neq 1$, les ${\mathbb{C}}$-espaces vectoriels $S_k^{*,D}(M,\varepsilon)$ et $S_k(M,\varepsilon)^{d-new}$ sont isomorphes en tant que modules sous l’algèbre de Hecke de $GL_2({\mathbb{A}}_f^{(d)})$. Si $k=2$, $\varepsilon=1$, c’est encore vrai en remplaçant $S_2^{*,D}(M,1)$ par $S_2^{*,D}(M,1)/S$.
L’action des opérateurs de Hecke dans cette correspondance se précise de plus en $l|d$, en faisant correspondre à $U_l$, l’opérateur de Hecke de $S_k^{*,D}(M,\varepsilon)$ donné par la double classe $K_0(M)\pi_lK_0(M)$, où $\pi_l \in D^*({\mathbb{A}}_f)$ est partout trivial sauf en $l$ où il vaut une (quelconque) uniformisante de $D({\mathbb{Z}}_l)$. On notera encore $U_l$ cet opérateur de Hecke pour $D^*$.
Il se trouve que nous n’allons pas considérer exactement l’espace $S_k^{*,D}(M,\varepsilon)$, mais un autre légèrement différent (ce qui explique la notation $*$ provisoire), qui lui est isomorphe comme module sous l’algèbre de Hecke. Soit $$\omega_M:= \left(
\begin{array}{cc} 0 & 1 \\ M & 0 \end{array} \right) \in GL_2({\mathbb{Q}})$$ on le voit comme un élément de $D^*({\mathbb{A}})$ trivial aux places divisant $d$ et à l’infini, diagonalement $\omega_M$ dans $D^*({\mathbb{A}}_f^{(d)})=GL_2({\mathbb{A}}_f^{(d)})$. $\omega_M$ normalise $K_0(M)$ et agit par ${\textrm{diag}}(a,b) \mapsto {\textrm{diag}}(b,a)$ sur le tore diagonal de $GL_2({\mathbb{A}}_f^{(d)})$.
L’application $f \mapsto \omega_M.f $ induit un isomorphisme ${\mathbb{C}}$-linéaire de $S_k^{*,D}(M,\varepsilon)$ sur l’espace des fonctions complexes sur $D^*({\mathbb{Q}})\backslash D^*({\mathbb{A}})/U_1(M)$ de caractère $\varepsilon^{-1}$ sous $U_0(M)$, de caractère central $|\det(g)|^{2-k}\varepsilon^{-1}$, engendrant sous $D^*({\mathbb{R}})$ un multiple du dual de $\textrm{Sym}^{k-2}({\mathbb{C}}^2)$.
On note $S^{D}_k(M,\varepsilon)$ ce ${\mathbb{C}}$-espace vectoriel muni de l’action de l’algèbre de Hecke obtenue par transport. Explicitement, si $l$ est premier ne divisant pas $d$, l’opérateur de Hecke $T_l$ si $(l,M)=1$, $U_l$ sinon, est donné par la double classe $U_0(M){\textrm{diag}}(1,l)U_0(M)$; si $(l,M)=1$, $S_l=l^{k-2}\varepsilon(l)$ est l’action de ${\textrm{diag}}(l,l)$, $U_l$ est inchangé si $l|d$.
Comme en \[formes\], ${\mathcal{H}}$ désigne la ${\mathbb{Z}}$-algèbre de polynômes sur les lettres $T_l$, $S_l$ si $l$ premier ne divise pas $M$, $U_l$ si $l|M$. Par ce que l’on vient de dire plus haut, $S_k^D(M,\varepsilon)$ et $S_k(M,\varepsilon)^{d-new}$ sont deux ${\mathcal{H}}$-modules isomorphes, avec la même exception pour $k=2$ que dans \[JL\].
Structures $p$-adiques des espaces de formes classiques {#pstructure}
-------------------------------------------------------
${}^{}$
Nous allons introduire une ${\mathbb{Q}}_p$-structure sur les espaces $S_k^{D}(M,\varepsilon)$ et $S_k(M,\varepsilon)$. On fixe pour cela $p$ premier, $\iota: {\mathbb{C}}_p \rightarrow {\mathbb{C}}$ un isomorphisme de corps, $K \subset {\mathbb{C}}_p$ un sous-corps complet, et $M=Npd$ avec $N,p$ et $d$ comme en §\[formes\].
La courbe modulaire $X_1(M)$ a une structure naturelle sur ${\mathbb{Q}}_p$, préservée par les correspondances de Hecke. Soit $S_k(M,\varepsilon,K)$ le sous $K$-espace vectoriel de $H^0(X_1(M)/K,\omega^k)$ composé des formes paraboliques de caractère $\varepsilon$. Notons que par “$GAGA$ rigide analytique”, ce ${\mathcal{H}}$-module est canoniquement isomorphe à son analogue sur $X_1(M)^{rig}/K$. La formation du ${\mathcal{H}}$-module $S_k(M,\varepsilon,K)$ commute à l’extension des scalaires sur $K$ et la donné de $\iota$ identifie $S_k(M,\varepsilon,{\mathbb{C}}_p)$ et $S_k(M,\varepsilon)$.
$S^{D}_k(M,\varepsilon,K)$ le $K$-espace vectoriel des fonctions $f: D^*({\mathbb{Q}})\backslash D^*({\mathbb{A}}_f) \rightarrow
\textrm{Sym}^{k-2}(K^2)$ satisfaisant $$f(xu)=\varepsilon(u)^{-1}u_p^{-1}f(x), \, \,
\, \, \forall (x,u) \in D^*({\mathbb{A}}_f)\times U_0(M)$$ c’est un ${\mathcal{H}}$-module de manière naturelle. La formation du ${\mathcal{H}}$-module $S_k^D(M,\varepsilon,K)$ commute à l’extension des scalaires sur $K$ et la donnée de $\iota$ identifie $S_k^D(M,\varepsilon,{\mathbb{C}}_p)$ et $S_k^D(M,\varepsilon)$, on rappelle comment dans ce qui suit.
A une fonction $f \in S_k^{D}(M,\varepsilon)$ est associé par définition un morphisme $D^*({\mathbb{R}})$-équivariant $\varphi_f$ de $\textrm{Sym}^{k-2}({\mathbb{C}}^2)^*$ vers l’espace des fonctions complexes sur $D^*({\mathbb{Q}})\backslash
D^*({\mathbb{A}})/U_1(M)$. Si $x \in D^*({\mathbb{A}})$, $$v \mapsto
\varphi_f(v)(x)$$ définit par dualité un unique élément $F_f(x) \in
\textrm{Sym}^{k-2}({\mathbb{C}}^2)$. On considère alors l’application qui à $f$ associe l’élément de $S_k^{D}(M,\varepsilon,{\mathbb{C}}_p)$ défini par $$x_f
\mapsto x_p^{-1}\iota^{-1}(F_f(1 \times x_f))$$ c’est l’isomorphisme cherché comme on le vérifie immédiatement.
Si $S_k(M,\varepsilon,{\mathbb{Q}}_p)^{d-new}$ désigne le sous-espace de $S_k(M,\varepsilon,{\mathbb{Q}}_p)$ composé des formes nouvelles en $d$ de $S_k(M,\varepsilon,{\mathbb{Q}}_p)$, il vient
Si $k>2$ ou $\varepsilon \neq 1$, les ${\mathcal{H}}\otimes_{{\mathbb{Z}}}{\mathbb{Q}}_p$-modules $S_k(M,\varepsilon,{\mathbb{Q}}_p)^{d-new}$ et $S_k^{D}(M,\varepsilon,{\mathbb{Q}}_p)$ sont isomorphes. $S_2(M,1,{\mathbb{Q}}_p)$ est ${\mathcal{H}}$-isomorphe au quotient de $S_2^{D}(M,1,{\mathbb{Q}}_p)$ par la droite des fonctions constantes.
La correspondance $p$-adique à poids-caractère fixé {#cla}
---------------------------------------------------
On reprend les notations du §\[formes\], et on s’intéresse aux systèmes de modules de Banach sur ${\mathcal{W}}$, $$M^1:={F^{0,\varepsilon,d}}\, \, \, \, et \, \, \,
\, M^2:={F^{D,\varepsilon}}$$ Comme on l’a vu, ils sont munis d’une action de l’algèbre $H:=\mathcal{H}$, et $U_i:=\rho_i(U_p) \in
\textrm{Comp}(M^i)$ commute à $\rho_i(H)$. On pose $X:=\{ \zeta(1+p)^k, k\geq
2, \zeta \in \mu_{p^{\infty}} \} \subset {\mathcal{W}}({\mathbb{C}}_p)$, il est très Zariski-dense dans ${\mathcal{W}}$. On munit $M^1$ et $M^2$ de structures classiques sur $X$ en posant pour $M^{i,cl}_{\zeta (1+p)^k}$, $k\geq 3$, $\zeta^{p^m}=1$,
- le sous-espace de $F_{\zeta(1+p)^k,n}^{0,\varepsilon,d}$, avec $n$ quelconque tel que $n \geq m-1$, des restrictions à $X_1(Np^m,d)(v_n)$ des sections de $\omega^k$ convergentes sur tout $X_1(Np^m,d)/{\mathbb{C}}_p$, s’annulant à [*toutes*]{} les pointes de $X_1(Np^m,d)/{\mathbb{C}}_p$, si $i=1$.
- le sous-espace $S_k^D(Np^m,\varepsilon\chi\tau^{-k},{\mathbb{C}}_p)$ de $F_{\zeta (1+p)^k,n}^{D,\varepsilon}$, $n \geq m-1$ quelconque, défini à la fin du paragraphe \[quat\].
Ce sont bien des structures classiques, satisfaisant $(Cl)$ en $(1+p)^k\zeta$ dès que $\alpha
< k-1 $ par les assertions connues de classicité en pente petite devant le poids ([@Col2] §8, [@Col3] 1.1,[@Buz2] §4). Il faut noter qu’une forme modulaire de poids $k$ sur $X_1(Np^m,d)$ qui est propre pour $U_p$ de pente strictement inférieure à $k\!-\!1$, et qui s’annule en toutes les pointes de $Z_1(Np^m,d)(0)$, est en fait parabolique. On est alors dans les hypothèses de la proposition \[det\] par le théorème \[JL\]. On en déduit le
\[seriecaracteristique\] Soit $h \in {\mathcal{H}}$, alors ${\textrm{Fred}}_{F^{0,\varepsilon,d}}(hU_p)={\textrm{Fred}}_{F^{D,\varepsilon}}(hU_p)$
[*Remarques:*]{}
- Il est aisé de voir que le membre de droite de cette égalité est en fait dans $1+T\Lambda\{\{T\}\}$, ainsi donc que le premier.
- On aurait pu se restreindre, dans notre choix de l’ensemble $X$, à celui de $\{(1+p)^k, k\geq 3\}$ pour obtenir le même résultat \[seriecaracteristique\]. Via le théorème \[jlpfamille\] (qui n’utilise que le résultat de \[seriecaracteristique\]), on peut alors déduire de la propriété $(Cl)$ pour les formes quaternionique, et de [@Col2], une nouvelle preuve du théorème de contrôle de [@Col3].
Soit $\kappa \in {\mathcal{W}}({\mathbb{C}}_p)$, on dispose de deux systèmes d’espaces de Banach $F^{0,\varepsilon,d}_{\kappa}$ et $F^{D,\varepsilon}_{\kappa}$, munis de représentations de ${\mathcal{H}}$. Le corollaire \[corsemisimpl\] implique le
\[jlp\] $\mathcal{X}_{U_p}({F^{0,\varepsilon,d}})=\mathcal{X}_{U_p}({F^{D,\varepsilon}})$
Autrement dit, l’espace des formes modulaires paraboliques surconvergentes de pente finie, de poids-caractère $\kappa$, de niveau modéré $Nd$, nouvelles en $d$, de caractère $\varepsilon$ et l’espace des formes modulaires quaternioniques $p$-adiques de pente finie, poids-caractère $\kappa$, de niveau modéré $N$, de caractère $\varepsilon$, ont même semi-simplification comme ${\mathcal{H}}$-module.
La correspondance en familles $p$-adiques
=========================================
Une propriété d’unicité pour les variétés spectrales {#unicite}
----------------------------------------------------
Soit ${\mathcal{W}}/{\mathbb{C}}_p$ un espace rigide réduit, soit $M$ la donnée d’un système de modules de Banach orthonormalisables sur ${\mathcal{W}}$, muni d’une représentation d’une ${\mathbb{C}}_p$-algèbre commutative $\rho: H \rightarrow \textrm{End}(M)$, et d’un $U \in H$ tel que $\rho(U) \in \textrm{Comp}(M)$. A une telle donnée $(M,U,H)$, on peut attacher la série de Fredholm $_M$($U$) $\in 1+TA({\mathcal{W}})\{\{T\}\}$, et l’hypersurface de Fredholm associée $Z \subset {\mathcal{W}}\times
{\mathbb{A}}^1$ définie par $_M$($U$)=0, munie de ses deux projections $(pr_1,pr_2): Z \rightarrow {\mathcal{W}}\times {\mathbb{A}}^1$. On sait alors que $Z$ a un recouvrement admissible canonique $\mathcal{C}:=\mathcal{C}$($_M(U)$), composé des ouverts affinoides $\Omega \subset Z$ finis et plats sur leur image $pr_1(\Omega)$, et ouverts fermés dans $pr_1^{-1}(pr_1(\Omega))$ ([@BMF] A5.8, ce qui suffit pour nos applications, [@Buz1] §4 pour le cas général). On sait de plus construire par recollement à l’aide de $\mathcal{C}$, un espace rigide $D$ ([@eigen] §7 [^4] puis [@ch] §6), la “variété spectrale attachée à $(M,U,H)$”, muni d’un morphisme fini $\pi: D \rightarrow Z$. On dispose de plus d’un morphisme d’anneaux $a: H \rightarrow
A(D)$, ainsi que d’un diagramme commutatif:
$$\xymatrix{ D \ar@{->}^{\kappa}[dd] \ar@{->}[dr]_{\pi}
\ar@{->}[drr]^{a(U)^{-1}}
\\ & Z \ar@{->}_{pr_2}[r] \ar@{->}^{pr_1}[dl] & {\mathbb{A}}_{rig}^1
\\ {\mathcal{W}}}$$
$H$, $U \in H$ et ${\mathcal{W}}$ étant fixés, on note $\mathcal{E}$ la catégorie dont les objets sont les couples $(\pi,a)$ formés d’un morphisme d’espaces rigides $\pi: D \rightarrow Z$ au dessus de ${\mathcal{W}}$, ainsi que d’un morphisme d’anneaux $a: H \rightarrow A(D)$ tel que $a(U)^{-1}=pr_2\cdot \pi$. Les morphismes $\textrm{Hom}((\pi_1,a_1),(\pi_2,a_2))$ sont ceux $(\varphi_D, \varphi_Z): \pi_1 \rightarrow \pi_2$ au dessus de ${\mathcal{W}}$ tels que $\forall h \in H, a_2(h) \cdot \varphi_D= a_1(h)$. Si $X=(\pi,a)$ est un objet de $\mathcal{E}$, on notera $D(X)$ (resp. $Z(X)$) l’espace rigide $D$ source (resp. $Z$ but) de $\pi$, $a(X):=a$, $\pi(X):=\pi$. Si $(M,U,H)$ est comme plus haut, on notera $\mathcal{E}(M)$ l’objet de $\mathcal{E}$ associé à $M$ par la construction précédente. On notera de plus $\mathcal{E}(M)^{{\textrm{red}}}$, la réduction de $\mathcal{E}(M)$, l’objet $(\pi^{{\textrm{red}}},a^{{\textrm{red}}})$ de $\mathcal{E}$, défini par $$\pi^{{\textrm{red}}}: D^{{\textrm{red}}}
\overset{can}{\hookrightarrow} D \overset{pi}{\longrightarrow} Z , \, \, \,
a^{{\textrm{red}}}: H \overset{a}{\longrightarrow}A(D)
\overset{can}{\longrightarrow} A(D)/\textrm{Nilrad}(A(D))$$
\[spectral\]Soient $(M^1,U,H)$ et $(M^2,U,H)$ comme plus haut, on suppose $$\forall h \in H, {\textrm{Fred}}_{M^1}(hU)={\textrm{Fred}}_{M^2}(hU)$$ Alors $\mathcal{E}(M^1)^{{\textrm{red}}}$ et $\mathcal{E}(M^2)^{{\textrm{red}}}$ sont canoniquement isomorphes.
[*Preuve:*]{} Par hypothèse, $Z(\mathcal{E}(M^1))=Z(\mathcal{E}(M^2))$, on la note $Z$, elle est munie de sa première projection $pr_1: Z \rightarrow {\mathcal{W}}$, on pose de plus $D_i=D_i(\mathcal{E}(M^i))$, $\mathcal{E}(M^i)=(\pi_i,a_i)$. Soit $\mathcal{C}$ le recouvrement canonique de $Z \rightarrow {\mathcal{W}}$, $\Omega \in
\mathcal{C}$, $V:=pr_1(\Omega)$ est un ouvert affinoide de ${\mathcal{W}}$, et par hypothèse $\Omega \rightarrow pr_1^{-1}(V)$ est un ouvert fermé fini sur $V$. Pour $n$ assez grand, $M^i(V,n)$ contient alors un sous-$A(V)$-module localement libre de rang fini, “indépendant de $n$”, dont on note $M^i(\Omega)$ l’image dans $M_i(V)^{\dagger}$. $M^i(\Omega)$ hérite d’une action de $H$, $\rho_i(U)$ ayant pour polynôme caractéristique le polynôme associé à la donnée de $\Omega$ (cf. par exemple [@ch] 6.3.3). Soit $H_V:=H \otimes_{{\mathbb{C}}_p} A(V)$, par construction $D_i(\Omega)$ est l’affinoide d’algèbre l’image de $H_V$ dans ${\textrm{End}}_{A(V)}(M^i(\Omega))$.
Soit $h \in H$, montrons que si $P_{i,h}(X):=\det(h_{|M^i(\Omega)}-X.1)$, alors $P_{1,h}(X)=P_{2,h}(X)$. $V \subset
{\mathcal{W}}$ étant réduit, il suffit de le faire après évaluation en tout $x \in
V({\mathbb{C}}_p)$. Soit $x \in V({\mathbb{C}}_p)$, alors le corollaire \[corsemisimpl\] s’applique à $(H,U)$ agissant sur les systèmes de modules de Banach $M^i_x$, on en conclut que les même caractères de $H$ apparaissent dans les semi-simplifications de ces deux espaces (avec multiplicités). $M^i(\Omega)_x$ est par définition le sous-espace de $\bigcap_{n\in {\mathbb{N}}} M^i_{x,n}$ sur lequel $\rho_i(U)$ a ses valeurs propres d’inverse dans $pr_1^{-1}(\{x\})\cap \Omega$, et ces dernières ne dépendent pas de $i$, la série caractéristique de $U$ n’en dépendant pas. Il vient donc $P_{1,h}(X)(x)=P_{2,h}(X)(x)$, puis $P_{1,h}(X)=P_{2,h}(X)$. Si plus généralement $h \in
H_V$, on a encore $P_{1,h}(X)=P_{2,h}(X)$ car par l’argument précédent on a cette égalité après évaluation en tout $x \in V({\mathbb{C}}_p)$.
Soit $I_i$ l’idéal de $H_V$ noyau du morphisme $H_V \rightarrow {\textrm{End}}_{A(V)}(M^i(\Omega))$, prouvons que $\sqrt{I_1}$=$\sqrt{I_2}$. Soit $f \in I_1$, alors $P_{1,h}(X)=X^d$, $d$ étant le rang de $M^1(\Omega)$ (égal au rang de $M^2(\Omega)$), on en déduit que $P_{2,h}(X)=X^d$ puis que $I_1^d \subset
I_2$. Il vient $\sqrt{I_1} \subset \sqrt{I_2}$, puis par symétrie $\sqrt{I_1}=\sqrt{I_2}$, ce que l’on voulait. On en déduit l’existence d’un isomorphisme d’anneau $H_V$-linéaire: $\varphi^*(\Omega):
A(D_2(\Omega)^{{\textrm{red}}}) \rightarrow A(D_1(\Omega)^{{\textrm{red}}})$. Un tel morphisme $H_V$-linéaire est nécessairement unique s’il existe, il est au dessus de $A(\Omega)$.
Soit alors $\varphi(\Omega) : D_1(\Omega)^{{\textrm{red}}} \rightarrow
D_2(\Omega)^{{\textrm{red}}}$ l’isomorphisme induit au dessus de $\Omega$. Vérifions que si $\Omega' \subset \Omega \in \mathcal{C}$, $\varphi_{\Omega}$ envoie $D_1(\Omega')^{{\textrm{red}}}$ dans $D_2(\Omega')^{{\textrm{red}}}$. Soit $V=pr_1(\Omega)
\subset {\mathcal{W}}$, $V'=pr_1(\Omega')$, alors $\Omega_{V'}:=\Omega
\cap pr_1^{-1}(V') \in \mathcal{C}$ et $D_i(\Omega_{V'})$ est l’ouvert $D_i(\Omega)^{{\textrm{red}}} \times_V V'$ de $D_i(\Omega)^{{\textrm{red}}}$. $\varphi(\Omega)$ induit un $H_{V'}$-isomorphisme $D_1(\Omega_{V'})=D_1(\Omega)^{{\textrm{red}}}\times_V V' \rightarrow
D_2(\Omega)^{{\textrm{red}}}=D_2(\Omega')^{{\textrm{red}}}$, ce qui conclut le cas $\Omega'=\Omega_{V'}$. Il reste donc le cas $V=V'$. On a $D_i(\Omega)^{{\textrm{red}}}=D_i(\Omega')^{{\textrm{red}}}\coprod D_i(\Omega \backslash
\Omega')^{{\textrm{red}}}$. $\varphi_{\Omega}$ étant $A(\Omega)$ et $H_V$-linéaire sur les fonctions, elle envoie $D_1(\Omega')^{{\textrm{red}}}$ isomorphiquement sur $D_2(\Omega')^{{\textrm{red}}}$ au dessus de $H_V$ et $A(\Omega')$. On conclut par unicité d’un tel morphisme que les $\varphi_{\Omega}$ se recollent en un isomorphisme $D_1^{{\textrm{red}}} \rightarrow
D_2^{{\textrm{red}}}$ au dessus de $Z$ par [@BGR] 9.3.3/1. $\square$
[*Remarque:*]{} Il est bien sur faux en général que sous les hypothèses de la proposition \[spectral\], $M^1$ et $M^2$ soient des $H$-modules isomorphes; \[spectral\] est la généralisation naturelle de \[semisimpl\].
Nous pouvons énoncer un résultat général combinant \[det\], \[spectral\] et \[corsemisimpl\]. Fixons ${\mathcal{W}}$ réduit de dimension $>0$ et relativement factoriel, $(M^1,H,U)$ et $(M^2,H,U)$ des systèmes de $H$-modules de Banach sur ${\mathcal{W}}$ munis de structures classiques sur un sous-ensemble très Zariski-dense $X
\subset {\mathcal{W}}({\mathbb{C}}_p)$ comme en \[critere\], satisfaisant (Cl):
\[general\] Supposons pour tout $h \in H$, $x \in X$, $$\det(1-ThU_{|M_x^{1,cl}})=\det(1-ThU_{|M_x^{2,cl}}) \in {\mathbb{C}}_p[T]$$ alors,
- ${\textrm{Fred}}_{M^1}(hU)={\textrm{Fred}}_{M^2}(hU) \in 1+TA({\mathcal{W}})\{\{T\}\}$,
- $\mathcal{E}(M^1)$ est canoniquement isomorphe à $\mathcal{E}(M^2)$,
- Pour tout $x \in {\mathcal{W}}({\mathbb{C}}_p)$, $\mathcal{X}_U(M^1_x)=\mathcal{X}_U(M^2_x)$.
Nous allons énoncer, pour terminer ce paragraphe, un critère sur $(M,U,H)$ assurant que $D(\mathcal{E}(M))$ est réduit. Ce passage peut être omis en première lecture, et apparaît ici par manque de référence satisfaisante. On fera les hypothèses suivantes sur ${\mathcal{W}}$: ${\mathcal{W}}$ est relativement factoriel ([@Con] §4), de dimension $>0$, et pour tout $x \in {\mathcal{W}}({\mathbb{C}}_p)$, $\widehat{{\mathcal{O}}_{X,x}}$ est intègre. On suppose de plus que l’on dispose de $X \subset {\mathcal{W}}({\mathbb{C}}_p)$ un ensemble très Zariski dense, et une structure classique sur $X$ au sens de §\[critere\], satisfaisant $(Cl)$. On fait de plus l’hypothèse de type “multiplicité $1$” suivante: $$\textrm{ "Soit $\alpha \in {\mathbb{R}}$, pour
presque tout $x \in X$, $H$ agit de manière semi-simple sur $M_x^{cl}\cap
M_x^{\leq \alpha}$}"$$
\[reduit\] Sous ces hypothèses, $D(\mathcal{E}(M))$ est réduit.
[*Rappels:*]{} Si $A$ est une algèbre affinoide, $x \in
\textrm{Specmax}(A)$, on notera $A_x$ (resp. $A^{rig}_x$) le localisé Zariski (resp. rigide) en $x$ ([@BGR] 7.3.2). Ils sont tous deux locaux noethériens, ont des complétés canoniquement isomorphes, et $A_x, A_x^{rig}$ et $\widehat{A_x}$ sont simultanément réduits ([@BGR] 7.3.2/8). Notons que si $A \rightarrow A'$ est un morphisme plat entre anneaux noethériens, alors si $A'$ n’a pas d’idéaux premiers associés immergés, il en va de même pour $A$. Ceci vaut en particulier pour $A_x \rightarrow A_x^{rig} \rightarrow
\widehat{A_x^{rig}}=\widehat{A_x}$. Par exemple, si tous les $A^{rig}_x$ sont sans composantes associées immergées, alors $A$ est sans composante associée immergée. Enfin, pour qu’un anneau noethérien $A$ sans composante associée immergée soit réduit, il suffit que pour un ensemble $\{x_1,...,x_n\} \subset \textrm{Specmax}(A)$ tel que chaque composante irréductible de $\textrm{Spec}(A)$ contienne un des $x_i$, chacun des $A_{x_i}$ soit réduit. En effet, sous ces hypothèses l’application canonique $A
\rightarrow \oplus_{i=1}^r A_{x_i}$ est injective. En particulier, si un tel anneau $A$ a son spectre irréductible, soit il est réduit, soit aucun des $A_x$, $x \in
\textrm{Specmax}(A)$ n’est réduit.
[*Preuve:*]{} On reprend les notations du début du §\[unicite\], un point $z \in D({\mathbb{C}}_p)$ sera dit classique si $\kappa(z) \in X$, et si le caractère de $H$ obtenu par évaluation en $z$ sur $D({\mathbb{C}}_p)$ est dans la semi-simplification du $H$-module $M_{\kappa(z)}^{cl}$. Les points classiques sont alors très Zariski denses car ${\mathcal{W}}$ est relativement factoriel de dimension $>0$, et par (Cl).
Soit $\Omega \in \mathcal{C}$ tel que $D(\Omega)$ contienne un point classique, il en contient alors un ensemble Zariski-dense. Soit $M(\Omega)$ le $A(V)$-module projectif de type fini associé à $\Omega$ comme dans la preuve de \[spectral\]. Soit $u \in A(D(V))\subset
{\textrm{End}}_{A(V)}(M(\Omega))$ nilpotent, par hypothèse de multiplicité $1$, les évaluations de $u$ sont nulles en presque tout $x$ de $X\cap V({\mathbb{C}}_p)$, et $u$ est donc nul; $D(V)$ est donc réduit s’il contient un point classique.
Soit $\Omega \in \mathcal{C}$ quelconque, montrons que les complétés des anneaux locaux aux points fermés de $D(\Omega)$ n’ont pas de composantes associées immergées. C’est en fait une propriété générale des sous-$A$-algèbres $B$ de ${\textrm{End}}_A(P)$ où $A$ est intègre noethérien tel que les $\widehat{A_m}$, $m \in \textrm{Max}(A)$, soient intègres, et $P$ projectif de type fini sur $A$. En effet, si $m \in \textrm{Max}(A)$ est fixé, $A':=\widehat{A_m}$, la platitude de $A \rightarrow A'$ entraîne que $B_m:=B
\otimes_A A'$ est canoniquement isomorphe à son image dans ${\textrm{End}}_{A'}(P\otimes_A A')\simeq M_r(A')$, où $r:=\textrm{rg}_A(P)$. $A'$ étant hensélien, $B_m$ est un produit d’algèbres locales $B_m^i$ finies sur $A'$, sans $A'$-torsion, car incluse dans $M_r(A')$. En particulier si $Q$ est un idéal premier associé de $B_m^i$, $Q\cap A'$ est un idéal premier de $A'$ de même hauteur que $Q$ annulant un élément de $B_m^i$, donc $Q\cap A'=0$ et cette hauteur commune est nulle, ce que l’on voulait.
Les $D(\Omega)$ recouvrant $D$ de manière admissible, tous les ${\mathcal{O}}_{D,x}^{rig}$, $x \in D$, sont sans composantes immergées associées, on conclut par le lemme suivant. $\square$
\[reduc\] Soit $X/{\mathbb{C}}_p$ un espace rigide dont les anneaux locaux n’ont pas de composantes associées immergées, et ayant un ensemble Zariski dense de $x$ tels que ${\mathcal{O}}_{X,x}^{rig}$ soit réduit, alors $X$ est réduit.
[*Preuve:*]{} Soit $X^0$ l’ensemble des $x$ de $X$ qui sont dans une seule composante irréductible de $X$, $X^0$ est un ouvert admissible de $X$. Soit ${\textrm{Red}}(X):=\{x \in X, {\mathcal{O}}_{X,x}^{rig} \textrm{ est réduit
}\}$, c’est (sans condition sur $X$) un ouvert admissible de $X$, on définit de même ${\textrm{Red}}(X^0)$. ${\textrm{Red}}(X)$ est Zariski-dense dans $X$, comme il est ouvert il y est aussi très Zariski-dense. On en déduit que ${\textrm{Red}}(X^0)$ est Zariski-dense dans $X^0$. Les rappels ci-dessus assurent que ${\textrm{Red}}(X^0)$ est un ouvert fermé admissible de $X^0$, on a donc $X^0={\textrm{Red}}(X^0)$. Si $V$ est un ouvert affinoide de $X$, $V\cap X^0$ est Zariski-dense dans $V$, et les rappels montrent que $V \subset {\textrm{Red}}(X)$. $\square$.
L’isomorphisme rigide analytique $_p$
-------------------------------------
Soient ${\mathcal{W}},N,p,d$ comme en §\[formes\], $${D}^{0,\varepsilon,d}:=D(\mathcal{E}(F^{0,\varepsilon,d}))^{{\textrm{red}}}, \, \, \,
\,
{D}^{D,\varepsilon}:=D(\mathcal{E}(F^{D,\varepsilon}))^{{\textrm{red}}}$$
On note $Z \subset {\mathcal{W}}\times {\mathbb{A}}^1$ l’hypersurface de Fredholm associée à $_{F^{0,\varepsilon,d}}(U_p)$= $_{F^{D,\varepsilon}}(U_p)$ (théorème \[seriecaracteristique\]), $a:=a(\mathcal{E}(F^{0,\varepsilon,d}))$, $a_D:=a(\mathcal{E}(F^{D,\varepsilon}))$, on dispose aussi de morphismes naturels ${D^{0,\varepsilon,d}}\rightarrow {\mathcal{W}}$ et ${D^{D,\varepsilon}}\rightarrow {\mathcal{W}}$ que l’on notera par le même nom $\bf{\kappa}$.
\[jlpfamille\] Il existe un unique isomorphisme rigide analytique ${\textrm{JL}}_p$: $D^{D,\varepsilon} \rightarrow D^{0,\varepsilon,d}$ au-dessus de ${\mathcal{W}}$, coïncidant avec la correspondance de Jacquet-Langlands sur les points classiques différents du point spécial. Il satisfait $\forall h \in
{\mathcal{H}}, \, \, a(h).{\textrm{JL}}_p=a_D(h)$.
Avant de prouver ce théorème, rappelons qu’un point $x$ de $D^{0,\varepsilon,d}({\mathbb{C}}_p)$ (resp. $D^{D,\varepsilon}({\mathbb{C}}_p)$) est dit [**classique**]{} si :
i\) $\kappa(x)=(1+p)^k\zeta$, où $k \geq 2$ est un entier, $\zeta \in \mu_{p^{\infty}}$,
ii\) le caractère ${\mathcal{H}}\rightarrow {\mathbb{C}}_p$ obtenu par l’évaluation en $x$ apparaît dans la semi-simplification du ${\mathcal{H}}$-module $F_{(1+p)^k\zeta}^{0,\varepsilon,d,cl}$ (resp. $F_{(1+p)^k\zeta}^{D,\varepsilon,cl}$).
On notera $x_0$ le point de $D^{D,1}({\mathbb{C}}_p)$ correspondant au caractère de ${\mathcal{H}}$ sur la droite des fonctions constantes dans $F_{(1+p)^2}^{D,1,cl}$, on l’appellera le [**point spécial**]{}.
[*Preuve:*]{} L’assertion d’unicité de \[jlpfamille\] découle de la Zariski-densité des points classiques de $D^{D,\varepsilon}$. Le théorème \[seriecaracteristique\] combiné à la proposition \[spectral\] assure l’existence d’un isomorphisme canonique $\phi: \mathcal{E}(F^{D,\varepsilon})^{{\textrm{red}}} \rightarrow
\mathcal{E}(F^{0,\varepsilon,d})^{{\textrm{red}}}$. On définit alors $${\textrm{JL}}_p: {D^{D,\varepsilon}}\rightarrow {D^{0,\varepsilon,d}}$$ comme étant l’isomorphisme induit par $\phi$. Par construction, il est au dessus de ${\mathcal{W}}$ et satisfait $a={\textrm{JL}}_p.a_D$.
Prouvons que ${\textrm{JL}}_p$ induit la correspondance de Jacquet-Langlands sur les points classiques non spéciaux. Rappelons (cf. par exemple [@ch] 6.2.4, 6.2.5) que si $w \in {\mathcal{W}}({\mathbb{C}}_p)$, l’application qui a un point $x
\in {D^{0,\varepsilon,d}}({\mathbb{C}}_p)$ (resp. ${D^{D,\varepsilon}}({\mathbb{C}}_p)$) tel que ${\bf \kappa}(x)=w$ associe le caractère ${\mathbb{C}}_p$-valué de ${\mathcal{H}}$ d’évaluation en $x$ induit une bijection entre ${ \bf \kappa}^{-1}(w)$ et $|\mathcal{X}_{U_p}(F_{w}^{0,\varepsilon,d})|$ (resp. $|\mathcal{X}_{U_p}(F_{w}^{D\varepsilon})|$, voir §\[semi\] pour la notation $|.|$). En particulier, si $x \in {D^{D,\varepsilon}}({\mathbb{C}}_p)$ est un point classique non spécial, la relation $a={\textrm{JL}}_p.a_D$ assure que ${\textrm{JL}}_p(x)$ correspond au même caractère de ${\mathcal{H}}$ que $x$. Or la correspondance de Jacquet-Langlands usuelle nous assure l’existence d’une forme classique de poids $\kappa(x)$ ayant ce caractère sous ${\mathcal{H}}$; par unicité (i.e par la bijection rappelée ci-dessus) le point de ${D^{0,\varepsilon,d}}({\mathbb{C}}_p)$ correspondant est nécessairement ${\textrm{JL}}_p(x)$, ce qui prouve le théorème. $\square$
[*Remarques:*]{} i) Par des techniques usuelles, on peut montrer que l’adhérence $\overline{{\mathcal{H}}}$ de la $\Lambda$-algèbre engendrée par l’image de ${\mathcal{H}}$ dans $A(D^{D,\varepsilon})$ est compacte. De ceci et de l’existence des représentations galoisiennes attachées aux formes modulaires classiques résulte facilement (voir par exemple [@ch] §7) l’existence d’un unique pseudo-caractère continu de dimension $2$ $$T: \textrm{Gal}(\overline{{\mathbb{Q}}}/{\mathbb{Q}})_{Npd}
\longrightarrow \overline{{\mathcal{H}}} \subset A(D^{0,\varepsilon,d})$$ tel que si $l$ est premier avec $(l,Npd)=1$, $T(\textrm{Frob}_l)=a(T_l)$.
ii\) ${\textrm{JL}}_p(x_0)$ correspond à la forme modulaire parabolique surconvergente de poids $2$, nouvelle en $d$, de $q$-développement $q+\sum_{n\geq 2}a_n q^n$ avec $a_l=l+1$ si $(l,pd)=1$, $1$ si $l|d$, et $p$ si $l=p$. Elle n’est pas classique au sens strict précédent, mais elle est quand même convergente sur tout $X_1(Np,d)$, bien que non cuspidale. Elle est “critique”, car de pente $1$ et de poids $2$.
En prenant $X:=\{(1+p)^k, k\geq 2 \in {\mathcal{W}}({\mathbb{C}}_p)\}$, les structures classiques sur $X$ ci-dessus, $N=1$, l’hypothèse de multiplicité $1$ pour ${\mathcal{H}}$ agissant sur $M_{(1+p)^k}^{cl}\cap M_{(1+p)^k}^{\leq \alpha}$ sont vérifiées dès que $\frac{k-1}{2}
> \alpha $ et \[reduit\] donne la
Si $N=1$, $D(\mathcal{E}(F^{0,\varepsilon,d}))$ et $D(\mathcal{E}(F^{D,\varepsilon}))
$ sont réduits.
Quelques conséquences, remarques et questions
=============================================
Opérateurs thêta
----------------
Soit $\kappa=(k,\chi)$ de conducteur $m$, $k \geq 2$, il existe un opérateur d’entrelacements surjectif $${\mathcal{C}}_{(1+p)^{-2}\kappa}^{\dagger} \longrightarrow
({\mathcal{C}}_{(1+p)^{-2}{\kappa^*}}^{\dagger})\otimes {\det}^{k-1}, \, \, \, f \mapsto (\frac{d}{dT})^{k-1}(f), \, \, \, \,
\kappa^*:=(2-k,\chi)$$ en tant que représentations de ${\mathbb{M}}$ (cf. [@Buz2] §6, voir aussi [@ST] 5.5). Son noyau est l’espace des fonctions localement polynomiales de degré $\leq k-2$. Il induit un opérateur $\Theta^{1-k}:
F^{D,\varepsilon}_{\kappa,0}[m] \rightarrow F^{D,\varepsilon}_{\kappa,0}[m]$ avec les notations de \[quat\], tel que $\forall n \in {\mathbb{N}}, \, \, \, \,
T_n(\Theta^{1-k}(f))=n^{1-k}\Theta^{1-k}(f)$. Le noyau de $\Theta^{1-k}$ est le ${\mathcal{H}}$-module $S_k^D(Np^m,\varepsilon\chi\tau^{-k},{\mathbb{C}}_p)$.
Du côté des courbes modulaires, on peut définir via l’application de Kodaira-Spencer ([@Col2] §4, [@Col3]) un opérateur $M_{2-k}(Np^m,d)^{\dagger} \rightarrow
M_{k}(Np^m,d)^{\dagger}$ noté $\Theta^{k-1}$, qui agit sur les $q$-développements par $(\frac{d}{dq})^{k-1}$, et satisfait donc $\forall n \in {\mathbb{N}}, \, \,
T_n(\Theta^{k-1}(f))=n^{k-1}\Theta^{k-1}(T_n(f))$. Les formes modulaires classiques propres, de pente finie, de poids $k$ et de niveau $Np^md$ ne sont pas dans l’image de $\theta^{1-k}$.
Si $x \in D^{D,\varepsilon}({\mathbb{C}}_p)$ est non classique mais de poids $\kappa$, il existe une forme $f_x \in F_{\kappa}^{D,\varepsilon}[m]$ propre, non classique, ayant pour système de valeurs propres celui associé à $x$. $\Theta^{1-k}(f_x) \in F_{\kappa^*}^{D,\varepsilon}[m]$ est propre et non nul, et correspond donc à un point que l’on notera $\Theta^{1-k}(x) \in
D^{D,\varepsilon}({\mathbb{C}}_p)$. On définit de même $\Theta^{k-1}(x)$ pour tout $x \in D^{0,\varepsilon,d}({\mathbb{C}}_p)$ de poids $\kappa^*$, \[jlpfamille\] implique la:
Soit $x \in D^{D,\varepsilon}({\mathbb{C}}_p)$ non classique de poids $\kappa(x)=(k,\chi)$, alors $$\Theta^{k-1}({\textrm{JL}}_p(\Theta^{1-k}(x)))=x$$
Questions
---------
${}^{}$
(Q1) Nous n’avons pas montré que les systèmes de modules de Banach $F^{0,\varepsilon,d}$ et $F^{D,\varepsilon}$ sont isomorphes. Est-il vrai, par exemple, que si $\kappa \in {\mathcal{W}}({\mathbb{C}}_p)$, les sous-espaces de $(F_{\kappa}^{D,\varepsilon})^{\dagger}$ et de $(F_{\kappa}^{0,\varepsilon,d})^{\dagger}$ composés des vecteurs de pente finie sont des ${\mathcal{H}}$-modules isomorphes ? Cette version “non semi-simplifiée” de notre correspondance serait par exemple intéressante pour des questions de multiplicité $1$ mieux comprises du côté $GL_2$ (essentiellement par la présence du $q$-développement).
(Q2) Existe-t’il une réalisation géométrique de la correspondance donnée ? Nous espérons revenir sur ce point dans un travail ultérieur.
(Q3) Plusieurs autres espaces de “formes modulaires $p$-adiques” naturels peuvent être définis autant au niveau quaternionique que pour $GL_2$. Par exemple, on peut remplacer les séries principales de $I$ dans notre définition des formes quaternioniques $p$-adiques par des restrictions des séries cuspidales de $GL_2({\mathbb{Q}}_p)$, qui apparaissent aussi en familles sur ${\mathcal{W}}$ et contiennent les représentations de dimension finie usuelles aux poids-caractères arithmétiques: à quoi correspondent-elles du côté $GL_2$ ?
Nous espérons revenir aussi sur l’étude d’une autre famille d’espaces de Banach (construite du côté elliptique cette fois-ci), obtenue en considérant les sections de $\omega^k$ sur la réunion finie des disques supersinguliers de $X_1(N)/{\mathbb{C}}_p$ ($N$ premier à $p$ disons). Ces espaces sont aussi liés à des espaces de formes modulaires quaternioniques pour $D$ ramifiée en $p$ et l’infini cette fois-ci (cf. une lettre de Serre à Tate [@Ser2] pour une correspondance modulo $p$).
[9999]{}
\
En préparation (2002)
\
Disponible à l’adresse http://www.ma.ic.ac.uk/$\thicksim$kbuzzard/maths/research/papers/index.html
\
Springer Verlag, Grundlehren der mathematischen Wissenschaften [**261**]{}
\
Annales de l’institut Fourier, [**49**]{}, 905-919 (1999).
\
Inventiones math. [**124**]{}, 214-241 (1996)
\
Inventiones math. [**127**]{}, 417-479 (1997)
\
Journal de théorie des nombres de Bordeaux [**9**]{}, 395-403 (1997)
\
Disponible à l’adresse http://www.dma.ens.fr/$\thicksim$chenevie/.
\
Proc. Durham, 1996. London Math. Soc. Lecture Note Ser. [**254**]{}, (1998)
\
Cambridge university press [**69**]{}.
\
Springer lecture notes [**114**]{}, (1970)
\
Modular functions of one variable 3, Springer lecture notes [**350** ]{}, (1972)
\
Preprint.
\
Publications Math. I.H.E.S. [**12**]{} (1962)
\
Oeuvres complètes, $IV$.\
Journal AMS [**15** ]{}, 443-468 (2002)
[^1]: [email protected]
[^2]: Strictement, il faudrait plutôt prendre des fonctions $K$-valuées avec $K$ sphériquement complet.
[^3]: $f_{|k g}(z):=f(g(z))j(g,z)^{-k}\det(g)^{k/2}, \, \, g \in GL_2({\mathbb{R}})^+$, en particulier, si $g \in Z({\mathbb{R}})$, $f_{|k g}=f$
[^4]: Il s’agit de la construction “D” de la courbe de Hecke dans [@eigen], uniquement basée sur de la théorie spectrale, nous ne nous occupons pas dans ce texte de la “C”
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Nowadays there is compelling evidence for the existence of dark matter in the Universe. A general consensus has been expressed on the need for a directional sensitive detector to confirm, with a complementary approach, the candidates found in conventional searches and to finally extend their sensitivity beyond the limit of neutrino-induced background. We propose here the use of a detector based on nuclear emulsions to measure the direction of WIMP-induced nuclear recoils. The production of nuclear emulsion films with nanometric grains is established. Several measurement campaigns have demonstrated the capability of detecting sub-micrometric tracks left by low energy ions in such emulsion films. Innovative analysis technologies with fully automated optical microscopes have made it possible to achieve the track reconstruction for path lengths down to one hundred nanometers and there are good prospects to further exceed this limit. The detector concept we propose foresees the use of a bulk of nuclear emulsion films surrounded by a shield from environmental radioactivity, to be placed on an equatorial telescope in order to cancel out the effect of the Earth rotation, thus keeping the detector at a fixed orientation toward the expected direction of galactic WIMPs. We report the schedule and cost estimate for a one-kilogram mass pilot experiment, aiming at delivering the first results on the time scale of six years.'
address:
- 'INFN Sezione di Bari, Bari, Italy'
- 'INFN Sezione di Napoli, Napoli, Italy'
- 'INFN Sezione di Padova, Padova, Italy'
- 'INFN Sezione di Roma, Roma, Italy'
- 'INFN-Laboratori Nazionali del Gran Sasso, Assergi (L’Aquila), Italy '
- 'INFN-Laboratori Nazionali di Frascati, Frascati (Roma), Italy '
- 'Dipartimento di Fisica dell’Università di Bari, Italy'
- 'Dipartimento di Fisica dell’Università Federico II di Napoli, Napoli, Italy'
- 'Dipartimento di Fisica e Astronomia dell’Università di Padova, Padova, Italy'
- 'Dipartimento di Fisica dell’Università di Roma, Rome, Italy'
- 'Nagoya University and KM Institute, Nagoya, Japan '
- 'Chiba University, Chiba, Japan '
- 'JINR-Joint Institute for Nuclear Research, Dubna, Russia '
- 'SINP MSU-Skobeltsyn Institute of Nuclear Physics of Moscow State University, Russia'
- 'LPI-Lebedev Physical Institute of the Russian Academy of Sciences, Moscow, Russia'
- 'METU-Middle East Technical University, Ankara, Turkey'
author:
- 'A. Aleksandrov'
- 'A. Anokhina'
- 'T. Asada'
- 'D. Bender'
- 'I. Bodnarchuk'
- 'A. Buonaura'
- 'S. Buontempo'
- 'M. Chernyavskii'
- 'A. Chukanov'
- 'L. Consiglio'
- 'N. D’Ambrosio'
- 'G. De Lellis'
- 'M. De Serio'
- 'A. Di Crescenzo'
- 'N. Di Marco'
- 'S. Dmitrievski'
- 'T. Dzhatdoev'
- 'R. A. Fini'
- 'S. Furuya'
- 'G. Galati'
- 'V. Gentile'
- 'S. Gorbunov'
- 'Y. Gornushkin'
- 'A. M. Guler'
- 'H. Ichiki'
- 'C. Kamiscioglu'
- 'M. Kamiscioglu'
- 'T. Katsuragawa'
- 'M. Kimura'
- 'N. Konovalova'
- 'K. Kuge'
- 'A. Lauria'
- 'P. Loverre'
- 'S. Machii'
- 'A. Managadze'
- 'P. Monacelli'
- 'M. C. Montesi'
- 'T. Naka'
- 'M. Nakamura'
- 'T. Nakano'
- 'A. Pastore'
- 'D. Podgrudkov'
- 'N. Polukhina'
- 'F. Pupilli'
- 'T. Roganova'
- 'G. Rosa'
- 'O. Sato'
- 'T. Shchedrina'
- 'S. Simone'
- 'C. Sirignano'
- 'A. Sotnikov'
- 'N. Starkov'
- 'P. Strolin'
- 'Y. Tawara'
- 'V. Tioukov'
- 'A. Umemoto'
- 'M. Vladymyrov'
- 'M. Yoshimoto'
- 'S. Zemskova'
title: '[ LNGS-LOI 48/15]{} NEWS: Nuclear Emulsions for WIMP Search (NEWS Collaboration) '
---
Introduction {#sec:Intro}
============
Compelling evidence for an abundant, non-baryonic, non-luminous (dark) matter component was collected in the last decades [@PdG]. Yet, the nature of the dark matter (DM) remains totally unknown, and the quest for an answer ranks as one of the main issues of the experimental particle physics, astrophysics and cosmology.
Weakly Interacting Massive Particles (WIMPs) [@ref1; @ref2] are creditable, theoretically appealing DM candidates. If these massive relics of the early universe do exist, they are expected to be gravitationally bound to the baryonic visible matter. A direct search for WIMPs in the mass range from a few GeV/c$^2$ to a few TeV/c$^2$ could be based on the detection of nuclear recoils induced by WIMP elastic scattering. Cross-sections are not expected to exceed those of weak processes.
The kinetic energy of scattered nuclei and consequently their range in dense matter would be determined by the WIMP mass and by its velocity relative to a terrestrial target. In the Standard Halo Model the WIMP speed in the galaxy is supposed to follow a Maxwellian distribution, showing null average values of all the velocity components. The motion of the Solar System through the galaxy, however, creates an apparent wind of dark matter particles, blowing opposite to the direction of the Sun’s motion toward the Cygnus constellation. The intensity of this wind, i.e. the WIMP flux, is expected to be time-modulated due to the Earth motion in the Solar System, with an annual period and a maximum rate in summer [@Spergel]. The speed of the Earth in the Solar System is anyway small compared to the speed of the Sun in the Milky Way, so the amplitude of the annual modulation is of the order of a few percent. The DAMA experiment [@DAMA] at LNGS has indeed reported a signal with a very clear evidence of annual modulation, as a possible indication of DM induced signal. This signal, although statistically extremely significant ($>8$ standard deviations), is controversial because many experiments have already partially or totally excluded the region allowed by DAMA. Therefore DAMA results remain an intriguing puzzle. Figure \[fig:StateOfArt\] shows the upper limits and contour regions for the WIMP spin-independent cross sections, normalized to the scattering on a single nucleon, as function of the WIMP mass. The constraints from SUSY models with the inclusion of LHC results are also shown. The figure was made with the `dmtools` web page [@dmtool].
![WIMP cross sections (normalized to a single nucleon) for spin-independent couplings versus mass. The DAMA/LIBRA [@DAMA] and CoGeNT [@COGENT] contour regions indicate possible signal events. The 90$\%$ C.L. upper limits for the CRESST-II [@CRESST], CDMS+EDELWEISS [@EDELWEISS], XENON100 [@Xenon100] and LUX [@LUX] experiments are shown as solid curves. The green region indicates the predictions from the Minimal Supersymmetrized Standard Model (MSSM) integrated with constraints set by LHC experiments [@MSSM]. []{data-label="fig:StateOfArt"}](figs/StateOfArt_dmtools){width="0.65\linewidth"}
On the other hand, the angular distribution of the scattered nuclei is peaked around the direction of the apparent dark matter wind. The directional modulation is expected to be stronger than the annual modulation, with a rate of forward-scattered nuclei one order of magnitude higher than the backward-scattered nuclei. Since background sources are expected to be isotropic, the detection of a signal with a preferred direction would provide a powerful discrimination. Directional experiments intend to exploit this effect by measuring the direction of nuclear recoils, and hence the WIMP direction.
In the above sketched WIMP scenario, the key points for the design of an experiment searching for DM with a directional approach are the expected event rate and the expected angular and energy distribution of recoiling nuclei. The expected event rate does not exceed 1 event/kg/year. Such extremely low rates require strong background suppression. The WIMPs mean velocity inside our galaxy is a few hundred kilometers per second at the location of our Solar System. For these velocities, WIMPs interact with ordinary matter mainly via elastic scattering on nuclei. With expected WIMP masses in the range from 10 GeV to 10 TeV, typical nuclear recoil energies are of the order of 1 $\div$ 100 keV. The expected nuclear recoil energy spectrum decreases almost exponentially with energy. To exploit directionality with light-medium mass scattered nuclei, the required spatial accuracy is in the sub-mm domain for gaseous detectors and in the sub-$\mu$m range for solid detectors. In the first case the low event rate sets the requirement of very large volumes while in the second case an extremely high resolution is required in order to cope with the very short range of the recoil nuclei.
Experiments for dark matter searches based on solid or liquid targets are not able to measure the direction of nuclear recoils. They search for a WIMP signal as an excess of events over the expected background with possibly an annual modulation of the event rate, if sensitive enough. The gaseous detectors, on the other hand, are capable of reconstructing the three-dimensional tracks of nuclear recoils, but their mass and the corresponding sensitivity are rather limited. Current gas-based detectors as DRIFT [@DRIFT], NEWAGE [@NEWAGE], DMTPC [@DMTPC] and MIMAC [@MIMAC] make use of low-pressure CF$_4$ with a fiducial volume ranging from 3 to 140 g [@DRIFT], thus providing limits only on the spin-dependent WIMP-proton cross-section.
The use of a solid target would allow to explore lower cross sections in the phase space indicated by recent limits drawn by direct search experiments, the challenge being the shorter track length, $O$(100nm), resulting in the WIMP-nucleus scattering.
The Nuclear Emulsions for WIMP Search (NEWS) project presented here aims at the direct detection of dark matter candidates by measuring the direction of WIMP-induced nuclear recoils. For this challenge, the detector exploits new generation nuclear emulsions with nanometric grains. An R$\&$D conducted by the Nagoya University in collaboration with the Fujifilm Company has established the production of films with nanometric grains for an ultra-high spatial resolution. We do report the results of this R$\&$D and the corresponding development of new fully automated scanning systems capable of detecting such short tracks, with improved optical technologies overcoming the diffraction limit of conventional systems. We have studied the detection efficiency of nanometric tracks, using ion implantation systems to reproduce nuclear tracks of the same length as expected from WIMP-induced nuclear recoils. A paragraph of this document is devoted to the measurements performed on the neutron yield from intrinsic film radioactivity and more in general to the discussion of potential background sources. Given that nuclear emulsions are time insensitive, the detector will be placed on a standard equatorial telescope to keep its orientation fixed toward the Cygnus constellation. The choice of appropriate shielding materials and detector layout is also discussed.
Finally we propose the design and construction of a one-kilogram detector for a pilot experiment, acting as a demonstrator of the technology and aiming at scaling it up to a larger scale experiment. The construction, run and data analysis are planned on the time scale of about six years. This experiment will demonstrate the potentiality of the technique and will start constraining the parameter space outlined by the DAMA experiment.
NIT: Nano Imaging Tracker {#sec:NIT}
=========================
After decades of remarkable experimental applications, nuclear emulsions still mantain their attraction as ionizing particle detectors of unmatched spatial and angular resolution. The first application of fully automated scanning systems to large-scale experiment was for the CHORUS experiment [@CHORUS]. Impressive achievements with new generation systems, more than one order of magnitude faster, allowed the design of the OPERA experiment [@OPERA] while current developments of the technology still inspire the design of high-statistics neutrino experiment with large active target [@SHIP].
Nuclear emulsions are made of silver halide crystals embedded in a gelatine matrix. When light falls on the emulsion, or ionizing particles pass through it, some of the halide crystals are modified in such a way that they are turned into grains of silver when the emulsion is immersed in a reducing bath (the so-called *developer*). The modifications in the grains caused by the action of light or radiation are invisible and the effect is referred to as *formation of latent image*. After development, a silver halide emulsion is placed in a second bath, called *fixer*, which dissolves the unaffected grains of silver halide but leaves the small black granules of silver. Finally, the plate is washed and dried [@Emulsion1; @Emulsion2; @Chap2HandbookOf]. The primary function of the gelatine is to provide a three-dimensional framework which serves to locate the small crystals of the halide and to prevent them migrating during development and fixation. The three-dimensional trajectory of passing through particles can be reconstructed with an optical microscope by connecting all the silver grains produced after development.
The size of silver halide crystals in standard emulsion ranges from 0.1 $\mu$m to 1 $\mu$m. The sensitivity of the emulsion strongly depends on the size of the crystals: the larger the grain, the higher the emulsion sensitivity to ionising radiation. Due to the low recoil energy of a WIMP-scattered nucleus, the expected track length is of the order of a few hundred nanometers. State-of-the-art emulsions produced by the Fuji Co. [@OPERAemulsion] for the OPERA experiment, with a linear dimension of the crystals of 200 nm, are therefore not suitable for Dark Matter searches.
The R$\&$D performed at Nagoya University, in collaboration with Fuji Co. experts, led to the production of novel emulsion films with grain diameters down to a few tens of nm, one order of magnitude smaller than conventional ones. The so-called Nano Imaging Trackers (NIT) and Ultra-Nano Imaging Trackers (U-NIT), have grains of 44.2 and 24.8 nm diameter respectively (see Figure \[fig:grains\]). NIT films have a linear density of crystals of about 11 crystals/$\mu$m [@NIT] while U-NIT show 29 crystals/$\mu$m [@U-NIT]. They make the reconstruction of trajectories with path lengths shorter than 100 nm possible, if analyzed by means of microscopes with enough resolution.\
![Distribution of the crystal diameter measured with an electron microscope for NIT (left) and U-NIT (right) emulsions. The measurements refer to three different batches.[]{data-label="fig:grains"}](figs/NIT_grain_distribution_2){width="1.0\linewidth"}
![NIT gel production machine.[]{data-label="fig:production-machine"}](figs/production-machine_old){width="0.8\linewidth"}
NIT are produced in three steps using a dedicated machine (see Figure \[fig:production-machine\]). First, the AgBr crystal growth is obtained by mixing in a thermostatic bath AgNO$_3$ and NaBr exploiting the following reaction: $$\mbox{AgNO}_3 + \mbox{NaBr} \rightarrow \mbox{AgBr} + \mbox{Na}^+ + \mbox{NO}_3^-$$ Polyvinyl alcol (PVA) is then added to ensure the uniformity of the grain size of the crystals. NaI, with a concentration of 4% mol, is also used in order to increase the quantum efficiency in the activation of the crystals. Next, in the desalination phase, AgBr crystals are mixed with the gelatin while the residual extra ions (Na$^+$,NO$_3^-$) are extracted by means of a reduction process. A homogeneous crystal distribution is obtained with a centrifugation process at 1000 rpm and $50^\circ$ C.
Finally, the emulsion gel obtained with this procedure (see Figure \[fig:gel\], left) is mixed with ultra-pure water and poured on a rigid support (usually plastic or glass) as shown in the right picture of Figure \[fig:gel\]. The production machine is able to produce up to 3 kg of NIT emulsion gel per week.
The mass fractions of NIT constituents and the chemical composition of NIT emulsions are reported in Tables \[tab:composition\] and \[tab:constituents\], respectively. The emulsion composition has been carefully determined for light elements by an elemental analyser (YANACO MT-6) with an uncertainty of 0.3 %. The mass fraction of silver and bromine has been measured by an energy dispersive X-ray analysis with an uncertainty of 2%. The density amounts to 3.43 g/cm$^{3}$.
![Left: emulsion gel. Right: emulsion gel poured on a glass support.[]{data-label="fig:gel"}](figs/gel1){width="1.0\linewidth"}
Constituent Mass Fraction
------------- ---------------
AgBr-I 0.78
Gelatin 0.17
PVA 0.05
: Constituents of NIT emulsions[]{data-label="tab:composition"}
Element Mass Fraction Atomic Fraction
--------- --------------- -----------------
Ag 0.44 0.10
Br 0.32 0.10
I 0.019 0.004
C 0.101 0.214
O 0.074 0.118
N 0.027 0.049
H 0.016 0.410
S 0.003 0.003
: Elemental composition of NIT emulsions.[]{data-label="tab:constituents"}
During the whole lifetime of the emulsion and before the development, due to thermal excitation, sensitive crystals can be randomly activated thus resulting in the production of random dark grains: the so-called *fog* (of the order of $1 \div 10 \slash(10 \mu$m$)^3$ for OPERA emulsions) represents a potentially dangerous source of background when looking for very short track length ($O$(100nm)) made of only two consecutive dark grains. In this case, indeed, if the fog density is too high, the probability that two fog grains are close enough to mimic a signal track is not negligible. A recent R$\&$D led to a new chemical development procedure resulting in a suppression of the fog density of one order of magnitude: using a low-temperature ($5^\circ$C) developer based on MAA-1 a fog density of $\sim0.1 \slash(10 \mu$m$)^3$ has been achieved. Moreover, fog grains show a rather different contrast and shape with respect to radiation sensitized grains. These important features can be exploited to enhance the signal to background ratio, as it will be explained in Section \[sec:read-out\].
Experimental concept {#sec:expConcept}
====================
NEWS is a very innovative approach for a high sensitivity experiment aiming at the directional detection of WIMPs: the detector is based on recent developments of the nuclear emulsions technology allowing to reach an extremely high spatial resolution.
The detector is conceived as a bulk of nuclear emulsions acting both as a target and as a tracking device surrounded by a shield (see Section \[sec:set-up\]) to reduce the external background. The detector will be placed on an equatorial telescope in order to absorb the earth rotation, thus keeping the orientation towards the Cygnus constellation fixed. The emulsion films will lie with their surface permanently parallel to the expected average WIMP wind direction. Figure \[fig:wimp\_direction\] shows the distribution of the WIMP incoming angle, in the laboratory frame, projected on a plane containing the average WIMP wind direction. The majority of WIMPs are directed forward with a peak at zero. The superimposed red curve shows the same angle if one is not sensitive to the forward/backward direction. The angular distribution of the trajectories of WIMP-scattered nuclei is therefore expected to be anisotropic.
![WIMP 2-dim angle distribution on a plane containing the average WIMP wind direction (blue curve). The red curve shows the same angle if one is not sensitive to the forward/backward direction.[]{data-label="fig:wimp_direction"}](figs/wimp_angle_2D){width="0.6\linewidth"}
The presence in the emulsion gel of lighter nuclei such as carbon, oxygen and nitrogen, in addition to the heavier nuclei of silver and bromine, is a key feature of the NEWS project, resulting in a good sensitivity to WIMPs with both light and heavy masses. The sensitivity indeed strongly depends on the minimum detectable track length. The path length of the recoiled track depends in turn on the kinetic energy of the scattered nucleus, being the kinematics determined both by the mass of the incident WIMP and by that of the target nucleus. The correlation between the track length of the recoiled nucleus and its kinetic energy is shown in Figure \[fig:correlation\] for the different target nuclei. WIMP with a mass of about 100 GeV/c$^2$ prefers Ag and Br as target, producing e.g. Br recoils with an average kinetic energy of about 50 keV. Although Ag and Br are the most effective targets for WIMP masses in this range, the detection capability is reduced since their ranges are shorter than lighter elements at the same energy. Instead, for a WIMP with a mass around 10 GeV/c$^2$, the kinematics favours lighter nuclei that, for a given kinetic energy, have a longer range. Therefore, the contribution of the C, N and O ions is essential for WIMP masses around 10 GeV/c$^2$.
![Correlation between the track length of the recoiled nuclei and their kinetic energy, for different target nuclei in NIT emulsions.[]{data-label="fig:correlation"}](figs/correlation){width="0.6\linewidth"}
The estimated WIMP rates are of the order of 1 event$\slash$kg$\slash$year, much lower than the usual radioactive backgrounds. For this reason, the detector has to be placed underground to be protected from cosmic-ray induced background. Moreover, a careful control of the radioactive contamination of the materials used for the detector construction and a precise estimation of the corresponding induced background are needed. We will discuss in detail the most relevant background sources for the WIMP search with an emulsion based detector on the mass scale of a few kilograms.
After the exposure, the emulsion films composing the target will be developed and the whole detector volume will be analyzed by using fully automated scanning systems. The read-out (see Section \[sec:read-out\]) is performed in two phases. In the first phase a fast scanning is performed (see Section \[sec:optical-read-out\]) by means of an improved version of the optical microscope used for the scanning of the OPERA films ([@ESS; @S-UTS]). By this step a fast pre-selection of the candidate signal tracks with a relatively low spatial resolution (200 nm) can be achieved. In order to resolve the nanometric grains belonging to a signal tracks and to enhance the signal to background ratio, a further scanning of the pre-selected tracks with a ultra-high resolution scanning system is foreseen (see Section \[sec:plasmon\]). The final resolution for the reconstruction of nuclear recoil tracks is estimated to be between $10$ and $20$ nm in position and better than $15 ^\circ$ in angle.
Read-out technique {#sec:read-out}
==================
In the NEWS experiment the expected WIMP signal will consist of short-path, anisotropically distributed, nuclear recoils over an isotropically distributed background. The search for signal candidates requires the scanning of the whole emulsion volume.
The read-out system has therefore to fulfill two main requirements: a fast, completely automated, scanning system is needed to analyse the target volume over a time scale comparable with the exposure; the spatial resolution has to be improved by more than one order of magnitude compared to that achieved with standard emulsion films, reaching the challenging value of a few tens of nanometers, in order to ensure high efficiency and purity in the selection of signal candidates.
The analysis of NIT emulsions is performed with a two-step approach: a fast scanning with a state-of-the-art resolution for the signal preselection followed by a pin-point check of preselected candidates with unprecedented nanometric resolution to further enhance the signal to noise ratio and perform very accurate measurements of the range and the recoil direction. These two steps are discussed in the next sub-sections.
Optical microscopy for candidate selection {#sec:optical-read-out}
------------------------------------------
[![Optical scanning systems modified for the analysis of NIT. \[fig:mic\_Nagoya\] Prototype installed at Nagoya University. \[fig:mic\_LNGS\] Prototype installed at LNGS and Naples scanning laboratories. \[fig:mics\]](figs/mic_Nagoya.jpg "fig:"){width="0.5\linewidth"}]{}
The members of the NEWS Collaboration own state-of-the-art experience of large-scale fast automated scanning with a spatial resolution of about $1\mu$m and an angular resolution of about 1 mrad, as currently applied in the OPERA experiment [@OPERAhowTo]: the European Scanning System (ESS [@ESS]) in Europe and the Super-Ultra Track Selector (S-UTS [@S-UTS]). In the last years an R&D program aimed at improving the ESS performances was carried by INFN groups, leading to prototypes with resolution improved by one order of magnitude, achieving a speed of almost 200 cm$^2$/h [@ESS-new]. A new system is being developed in Japan (the Super-Ultra Track Selector), aiming at increasing the scanning speed up to 5000 cm$^2$/h.
Stepping into the nano imaging domain requires substantial upgrades of the OPERA-style scanning systems. New prototypes (see Figure \[fig:mics\]) were already set-up both in Japan and in Italy, featuring:
- higher magnification of the objectives lens, from 50x to 100x
- higher numerical aperture, from 0.8 to 1.45
- higher optical contrast (illumination by reflected light instead of transmitted light).
- light with green or blue wavelength to improve the resolution
- high pixel to micron ratio ($\sim$ 28 nm/pixel), one order of magnitude better than the systems used in OPERA
- high resolution (4Mpx) and high speed (563 fps) CMOS camera.
In parallel with the hardware improvements, the development of a new acquisition software and a new tracking algorithm has been carried out: the high data rate (1.7 GB/s), a factor 4 higher than the ESS one due to the improved sensor resolution, has required the use of last generation acquisition boards (Matrox Radient eCL SFCL/DFCL). As a consequence, a more powerful computing system, exploiting a GPU (Graphic Processing Unit) based architecture, has been implemented.
In order to evaluate the performances of the new scanning systems, extensive tests were performed with exposures of NIT to slow ions and neutron beams. Results are discussed here.
The starting point of the emulsion scanning is the image analysis to collect clusters made of dark grains at several depths across the emulsion plate thickness. Given the intrinsic resolution of the optical microscope ($\sim$ 200 nm), the sequence of several grains making a track of a few hundred nanometers, appears as a single cluster. Therefore, the key element to distinguish clusters made of several grains from clusters made of a single grain produced by thermal excitation (fog) is the analysis of their shape. A cluster made of several grains indeed tends to have an elliptical shape with the major axis along the direction of the trajectory, while a cluster produced by a single grain tends to have a spherical shape.
In order to simulate the effect of a WIMP-induced nuclear recoil and to measure the efficiency and the resolution of the new optical prototype, a test beam with low velocity ions was performed. We used both a Kr ion beam with energies of 200 and 400 keV [@ShapeAnalysis] and a C ion beam with energies of 60, 80 and 100 keV. Kr and C ions of such energies produce in emulsion tracks with a length in the range 100$\div$300 nm. These ions were implanted in the emulsions using an implantation facility of low speed ions at Nagoya University.
When analysed with the optical microscope, submicrometric tracks produced by Kr and C ions appear as shown in Figure \[fig:KrIon-ShapeAnalysis\]. Although silver grains belonging to the tracks are not distinguishable and appear as a single cluster, the elongated form of the cluster is clearly visible [@ShapeAnalysis2]. An elliptical fit of the cluster shape allows a clear separation between fog grains and signal tracks: the latter ones are expected to have ellipticity larger than a given threshold, typically 1.25 or higher (see left plot of Figure \[fig:shapeAnalysis60\] and \[fig:shapeAnalysis80\]).
The angular distributions of 60 and 80 keV C ions are reported in the right plot of Figure \[fig:shapeAnalysis60\] and Figure \[fig:shapeAnalysis80\], respectively. A peak corresponding to the direction of the implanted ions is clearly visible; the width of the distribution corresponds to the angular resolution, amounting to 360 mrad. The angular resolution is given by the convolution of the intrinsic resolution with the angular deviations caused by the scattering in the material. For low energy (below 100 keV) tracks, the scattering cannot be neglected.
In order to evaluate the intrinsic angular resolution of the scanning system we analysed an emulsion sample exposed to a 2.8 MeV neutron beam at the Fusion Neutronics Source (FNS) of the Japan Atomic Energy Agency (JAEA). In this case the track length distribution of neutron-induced proton recoils shows a wider range, up to a few hundred of micrometers. A sample of tracks with length of the order of few tens of micrometers and made by a sequence of several elliptical clusters was selected, being the scattering effect negligible for them. The same ellipticity cut applied in the previous analysis was used for the selection of the clusters. For each cluster, the angular difference $\Delta \theta$ between its major axis and the fitted track was evaluated (see Figure \[fig:angularResMethod\]). The distribution of $\Delta \theta$ shows a gaussian shape, as shown in Figure \[fig:angularResAndrey\] with a width corresponding to the intrinsic angular resolution and amounting to 230 mrad. This value represents the intrinsic angular resolution achieved with fully automated scanning systems, by far the best resolution achieved with direction sensitive detectors in this energy range. The simulation shows that this result is compatible with the measurement reported above when the scattering contribution is included.
![Kr ions implanted on NIT films. The image is taken with an optical microscope. The selection of candidate tracks is based on the elliptic fit of the clusters[]{data-label="fig:KrIon-ShapeAnalysis"}](figs/KrIon-ShapeAnalysis_2){width="0.6\linewidth"}
![Left: scatter plot of major and minor axes for clusters analysed with an elliptical fit in a 60 keV C ion test beam. Signal tracks are shown as red dots, fog grains in blue. Right: angular distribution of 60 keV C ion tracks selected by the ellipticity cut.[]{data-label="fig:shapeAnalysis60"}](figs/shapeAnalysis60){width="1.0\linewidth"}
![Left: scatter plot of major and minor axes for clusters analysed with an elliptical fit in a 80 keV C ion test beam. Signal tracks are shown as red dots, fog grains in blue. Right: angular distribution of 80 keV C ion tracks selected by the ellipticity cut.[]{data-label="fig:shapeAnalysis80"}](figs/shapeAnalysis80){width="1.0\linewidth"}
Tracks selected with the shape analysis were validated using the X-ray microscope [@NakaX-ray]. This technique features a higher resolution (of the order of 60 nm) but a slower scanning speed when compared with the optical microscopy. The analysis of a few hundred $\mu$m$^2$ takes about 100 s. The X-ray microscopy can therefore be used only to check a sample of already selected candidate tracks: X-ray analysis was used to demonstrate the principle of selection by elliptical shape analysis and measure the efficiency achievable with the optical microscopy.
The comparison of optical and X-ray images of candidate tracks is reported in Figure \[fig:x-ray\_confirmation\]: the high resolution of the X-ray microscope allows to resolve grains belonging to submicrometric tracks thus providing the final discrimination between signal and background.
In Figure \[fig:eff\_vs\_length\] the detection efficiency of the optical system as a function of the track length is shown: the efficiency is obtained first selecting a set of multi-grain tracks with the X-ray microscope and then scanning them with the optical one and applying the shape analysis. In this test an optical microscope with a pixel to micron ratio of 55 nm/pixel was used. Results show that the efficiency reaches 100$\%$ above 180-200 nm.
In Figure \[fig:eff\_vs\_energy\] the efficiency as a function of the recoil energy for C ions of 60, 80 and 100 keV, is shown: MC simulations (red line) well describes the data (blue points). It is worth noting that the capability of reconstructing low energy tracks (E $<$ 40 keV), corresponding to shorter path lengths, although with a lower efficiency, could significantly enhance the sensitivity to low WIMP mass regions.
The scanning speed of the prototype currently used for the shape analysis is of about 25 mm$^2$/h.
![Comparison between reconstructed tracks of a few hundred nanometers length with the optical microscope and with the X-ray microscope.[]{data-label="fig:x-ray_confirmation"}](figs/x-ray_confirmation_2){width="0.8\linewidth"}
![Efficiency of the elliptical fit analysis versus the track length when an ellipticity of 1.25 is used as a threshold. []{data-label="fig:eff_vs_length"}](figs/eff){width="0.6\linewidth"}
![Efficiency of the elliptical fit analysis versus the C ion energy when an ellipticity of 1.25 is used as a threshold. MC simulation (red line) well describes the data (blue points).[]{data-label="fig:eff_vs_energy"}](figs/Eff-vs-Energy){width="0.8\linewidth"}
Beyond the limits of the optical scanning for candidate validation {#sec:plasmon}
------------------------------------------------------------------
The use of optical microscopes allows the reconstruction of tracks down to 200 nm. The X-ray microscopy can overcome this limit though being extremely slow if compared with automated optical systems. Being the speed an issue in the analysis of a large mass detector, NEWS aims at the improvement of the spatial resolution enhancing the optical microscopy without using X-ray microscopes.
The basic idea is to exploit the resonance effect occurring when nanometric metal grains are dispersed in a dielectric medium [@ResonantLightScattering]. The polarization dependence of the resonance frequencies strongly reflects the shape anisotropy and can be used to infer the presence of non-spherical nanometric silver grains. Figure \[fig:resonantLight\] shows the results of the resonant light scattering from individual Ag nanoparticles [@ResonantLightScattering]: spherical particles do not show any different response as a function of the incident polarization, while a deformed sphere is sensitive to the polarization.
![Scattered-light spectra from individual Ag particles with spherical (left) and spheroidal (right) shape [@ResonantLightScattering]. The inset shows the 300 $\times$ 300 nm$^2$ SEM image of the particle. Arrows indicate the polarization of the incident light. A dependence of the response on the light polarization is observed for particles with ellipsoidal shape.[]{data-label="fig:resonantLight"}](figs/resonantLight){width="1.0\linewidth"}
NEWS will use this technology to retrieve track information in NIT emulsions beyond the optical resolution. Images of the same cluster taken with different polarization angles will show a displacement of the position of its barycenter. The analysis of the displacements allows to distinguish clusters made of a single grain from those made of two (or more) grains.
![Schematic view of the optical path instrumented with a polarizer to obtain a nanometric resolution with optical microscopes. []{data-label="fig:plasmon_prototype"}](figs/plasmon_prototype){width="1.0\linewidth"}
![Application of resonant light scattering to an elliptical cluster with ellipticity 1.27. Left plot: $dx$ and $dy$ are the displacements of the cluster barycenter for a given polarization in pixel units (1 pixel = 55 nm). Right plot: track slope fit and its length of about 90 nm.[]{data-label="fig:plasmon_analysis1"}](figs/plasmon_analysis1){width="1.0\linewidth"}
![Position accuracy of $x$ (left) and $y$ (right) coordinates of about 10 nm with the resonant light scattering.[]{data-label="fig:plasmon_resolution"}](figs/plasmon_resolution){width="1.0\linewidth"}
In order to study the polarized light effect, several tests have been performed on NIT samples exposed to 100 keV C ions. Optical microscopes have been equipped with a polarization filter as shown in Figure \[fig:plasmon\_prototype\]. The polarization direction can be changed by rotating the polariser. The rotation is at the moment done by hand while its automation is being designed. Images of the same clusters were taken by rotating the polarizer of 180$^\circ$ with steps of 10$^\circ$. An example of the analysis performed on a cluster with ellipticity 1.27 is reported in Figure \[fig:plasmon\_analysis1\]. For all the images, the displacement ($dx$, $dy$) of the cluster barycenter in $x$ and $y$ coordinates is measured in terms of pixel units (1 pixel $=$ 55 nm). A displacement exceeding the position accuracy of a single grain is the evidence for a cluster made of two consecutive grains and therefore produced by a signal track. From the analysis of $dy$ versus $dx$ it is possible to retrieve the track length and slope. In this case, the measured track length is 1.5 pixel, corresponding to about 90 nm.
The evaluation of the position accuracy was performed by analysing images of single grains. The unprecedented accuracy of about 10 nm can be achieved in both coordinates, as shown in Figure \[fig:plasmon\_resolution\].
The test performed demonstrates that this technology is very promising and that it can replace the X-ray microscope. The resonant light scattering has, in fact, the big advantage to achieve a nanometric resolution with optical microscopes. The validation of the candidates identified by the shape analysis will be performed in the same scanning laboratory, without moving the samples to a dedicated laboratory for the X-ray analysis. Moreover, optical microscopes are characterized by a much faster scanning speed with respect to X-ray microscopes, since they profit of all the R$\&$D performed in the last decades both for the OPERA and the NEWS experiments.
Expected Background {#sec:expected-bkg}
===================
The final sensitivity of low-energy rare event searches is strongly limited by the background induced by radioactivity. Two main categories have to be taken into account: the environmental or external background and the intrinsic one. The flux of the former can be significantly reduced by placing the detector underground, to absorb the cosmic radiation, and designing an appropriate shield against the natural radioactivity. The latter is an irreducible source of radiation: it is therefore crucial to control the radioactivity of the materials used for the construction of both the detector and of the shield as well as of the structure of the apparatus.
Background sources for dark matter searches are $\alpha$ and $\beta$ particles, $\gamma$-rays and neutron induced recoils, while NIT are essentially not sensitive to minimum ionizing particles (MIP).
The main sources of $\alpha$-particles are U and Th radioactive chains and Radon. The $\alpha$-particles produced in those processes have energies of the order of MeV and their range in emulsion is of the order of tens of microns, by far longer than WIMP-induced nuclear recoils. $\alpha$-particles can therefore be identified and discarded in the emulsions by an upper cut on the track length. Anyway Radon progeny $^{214}$Pb, $^{214}$Bi and $^{210}$Bi emit energetic $\beta$ and $\gamma$ radiation. To prevent Radon contamination, the detector has to be kept sealed from the air and continuously flushed with boil-off nitrogen.
The $\gamma$ radiation due to environmental radioactivity constitutes a non-negligible contribution to the total background budget. In Figure \[fig:gamma-bkg\] the measured $\gamma$ flux in the LNGS underground halls is shown [@BrunoPhDThesis; @arneodo; @wulandari]. Passive or active shielding (usually water, copper or lead) can be used to suppress the external $\gamma$-radiation down to the level of ppb or ppt. The thickness *l* required to reduce the external flux by a factor $f > 1$ can be estimated assuming exponential damping $\emph{l} = \lambda (E_\gamma) \times \log f$, where $\lambda (E_\gamma)$ is the energy-dependent attenuation length and $E_\gamma$ is the $\gamma$-ray energy.
A relevant source of background is represented by $\beta$-rays produced in $^{14}$C decay. Given the carbon content in the emulsions and the $^{14}C$ activity, a rejection power R$_{\beta}\leq10^{-8}$ is required in order to make it negligible (i.e. less than one background track/kg/year). The current rejection power for tracks made by two crystals is R$_{\beta}=10^{-6}$. In order to further improve the rejection, three possibile improvements are under investigation. The first one is based on the different energy deposition per path length of WIMP induced recoils and electrons [@gamma-response]: the response of emulsions can be tuned by dedicated chemical treatments (e.g. Tetrazorium compound [@tetraz]). The second possibility is to exploit the response of $\beta$-rays to the polarized light scattering: indeed grains induced by $\beta$-rays might be less sensitive to polarized light. Finally, a reduction of the background can be achieved by performing a cryogenic exposure and by exploiting the phonon effect. Preliminary tests at $\sim 100$ K show an upper limit of R$_{\beta}<10^{-7}$ for tracks made by two crystals.
![$\gamma$-flux measured in the underground LNGS halls [@BrunoPhDThesis; @arneodo; @wulandari].[]{data-label="fig:gamma-bkg"}](figs/gamma-flux){width="0.7\linewidth"}
Neutron induced recoils rank as the main background source because they are not distinguishable from the expected WIMP signal, except for the isotropic angular distribution and for the typical track length, largely exceeding the range expected for WIMP-induced recoils. Indeed, while neutron-induced proton recoils can be as long as few hundred microns, the maximum length of a WIMP-induced nuclear recoil is smaller than $1\mu$m even for large ($O$(TeV)) WIMP masses. Three types of neutron sources affect underground experiments: radiogenic neutrons in the MeV range produced in ($\alpha$, n) and spontaneous fission reactions in the detector due to its intrinsic radioactive contaminants, cosmogenic neutrons with energy spectrum extending to GeV energies induced by muons penetrating underground through the rock, neutrons induced by environmental radioactivity.
In Figure \[fig:neutron-flux\] the measured neutron flux in the LNGS underground halls is shown [@BrunoPhDThesis]: for a neutron energy of the order of a few MeV (*fast* neutrons) the flux ranges from $10^{-6}$ to $10^{-10}$ cm$^{-2}$ s$^{-1}$ MeV$^{-1}$. Light materials are effective moderators for fast neutrons: polyethylene (PE, C$_2$H$_4$) is commonly used to reduce the external neutron flux.
![The neutron flux measured in the underground LNGS halls [@BrunoPhDThesis].[]{data-label="fig:neutron-flux"}](figs/neutron-flux){width="0.7\linewidth"}
While the external neutron flux can be reduced to a reasonable level with an appropriate shielding, the intrinsic emulsion radioactivity would be responsible of an irreducible neutron yield through ($\alpha$, n) and $^{238}$ U spontaneous fission reaction. In order to estimate this contribution, a sample of each component of the nuclear emulsion (AgBr, Gelatin and PVA) has been analysed by the Chemistry Service in Laboratori Nazionali del Gran Sasso (LNGS, Italy), with the Inductively Coupled Plasma Mass Spectrometry (ICP-MS) technique [@ICP-MS] and at the low background facility STELLA (SubTErranean Low Level 125 Assay) of the LNGS [@STELLA] with germanium detectors. The complementary use of these techniques allows to determine both the Uranium and Thorium activities and to verify the secular equilibrium hypothesis.
The measured activities are reported in Table \[tab:activities-MS\] for all the constituents. The upper limits on PVA are evaluated at 95$\%$ CL.
By weighting the measured activity of each constituent for its mass fraction, the total activity of nuclear emulsion can be calculated. Using the contamination measured with the mass spectrometry, the $^{238}$U activity amounts to $23\pm 7$ mBq kg$^{-1}$, while the $^{232}$Th one is $5.1\pm 1.5$ mBq kg$^{-1}$. The reported errors are dominated by the 30$\%$ uncertainty in the radioactive contamination measurements. By assuming a null contribution from PVA, the previous contaminations are reduced by $\sim 2\%$.
The $\gamma$ spectrometry gives comparable results for the AgBr sample. For the gelatin the measurements provide comparable results for the $^{232}$Th chain, while the measured concentrations of $^{226}$Ra in the $^{238}$U chain is about 20 times smaller than the parent isotope, with a measured value of $2.4\pm 0.6$ mBq kg$^{-1}$. This measurement suggests a break in the secular equilibrium of the decay chain at this point. Therefore the secular equilibrium is assumed for the upper part of this chain, using the activity measured by mass spectrometry, while, for the lower part, nuclides are considered in equilibrium with $^{226}$Ra and the activity measured with $\gamma$-spectroscopy is used. The nuclear emulsion activity for nuclides of the $^{226}$Ra sub-chain is therefore $15\pm 5$ mBq kg$^{-1}$ [@intrisicBkgPaper].
[c|c|c]{} Nuclide & Contamination \[10$^{-9}$ g g$^{-1}$\] & Activity \[mBq kg$^{-1}$\]\
\
$^{232}$Th & 1.0 & 4.1\
$^{238}$U & 1.5 & 18.5\
\
$^{232}$Th & 2.7 & 11.0\
$^{238}$U & 3.9 & 48.1\
\
$^{232}$Th & $< 0.5$ & $< 2.0$\
$^{238}$U & $< 0.7$ & $< 8.6$\
The measured activity was used to determine the neutron yield both through a semi-analitical calculation [@refCalcFabio1; @refCalcFabio2] and a MC simulation based on the SOURCES code [@SOURCES]. Results are reported in Table \[tab:resNeutronYield\]. The two approaches give comparable results and the flux due to the intrinsic radioactive contamination is expected to be of the order of $1.2 \pm 0.4$ neutron per year per kilogram of nuclear emulsion. The energy spectrum of the produced neutrons, as calculated with SOURCES, is reported in Figure \[fig:SOURCES-spectrum\].
![Total neutron energy spectrum (black line); in red the contribution from $^{238}$ U spontaneous fission is shown, while in blue and green the contributions from ($\alpha$,n) reactions due to nuclides in the $^{238}$U and $^{232}$Th chains respectively are displayed [@intrisicBkgPaper].[]{data-label="fig:SOURCES-spectrum"}](figs/neutron_spectrum){width="0.7\linewidth"}
In order to estimate the detectable background due to radiogenic neutrons produced by the intrinsic radioactive contamination of the nuclear emulsions, a GEANT4 based simulation was performed. Simulated neutrons have an isotropic angular distribution and are uniformly distributed in a target where emulsion are arranged in a stack with a surface of $25 \times 25$ cm$^2$ and a thickness of 0.5 cm; their energy spectrum was generated according to Figure \[fig:SOURCES-spectrum\]. The fraction of interacting neutrons is 20.4$\%$: they can produce either a proton a nuclear recoil. In the former case the track length in emulsion extend up to several hundreds $\mu$m (see Figure \[fig:proton\_recoils\]) while nuclear recoils show shorter track lengths, not exceeding 3 $\mu$m for light nuclei (C, N, O) and 1 $\mu$m for heavy nuclei (Ag, Br, I) (see Figure \[fig:nuclear\_recoils\]). The overall fraction of neutron-induced recoils contributing to the background is computed by accounting for recoil tracks with lengths above the read-out threshold. Moreover an upper limit on the track length can be introduced since the signal is expected to be below 1 $\mu$m even for large ($O$(TeV)) WIMP masses (see Figure \[fig:maximumRange-vs-WIMPmass\]). The fractions of neutron-induced recoils below this cut, as a function of the read-out threshold, are reported in Table \[tab:recoils1\]: only fraction, from 5% to 10%, contributes to the background. A further reduction of $\sim 70\%$ of the neutron-induced background can be achieved exploiting the directionality information with the cut $-1 < \phi < 1$. Under these assumptions, the detectable neutron-induced background would be 0.02 $\div$ 0.03 per year per kilogram.
![Track length (left) and energy spectrum (right) for proton recoils produced by elastic (blue curve) and inelastic (red curve) processes.[]{data-label="fig:proton_recoils"}](figs/proton_recoils_G4){width="1.0\linewidth"}
![Track length (left) and energy spectrum (right) for heavy (blue curve) and light (red curve) nuclei.[]{data-label="fig:nuclear_recoils"}](figs/nuclear_recoils_G4){width="1.0\linewidth"}
![Maximum range expected for nuclear recoils as a function of the WIMP mass for the various nuclei.[]{data-label="fig:maximumRange-vs-WIMPmass"}](figs/maximum_range_vs_WIMPmass){width="0.7\linewidth"}
------------------------------------- ------------------------ -----------------------------
Process SOURCES simulation Semi-analytical calculation
\[kg$^{-1}$ y$^{-1}$\] \[kg$^{-1}$ y$^{-1}$\]
($\alpha$, n) from $^{232}$Th chain 0.12 $0.10 \pm0.03$
($\alpha$, n) from $^{238}$U chain 0.27 $0.26 \pm 0.08$
Spontaneous fission 0.79 $0.8 \pm 0.3$
Total flux 1.18 $1.2 \pm 0.4$
------------------------------------- ------------------------ -----------------------------
: Neutrons per kilogram per year due to ($\alpha$, n) and spontaneous fission reactions in the nuclear emulsion, evaluated with the SOURCES code and semi-analytical calculation using the measured $^{238}$U and $^{232}$Th contaminations as input.[]{data-label="tab:resNeutronYield"}
The neutron-induced background due to the intrinsic radioactive contamination allows the design of an emulsion detector with an exposure of about 10 kg year. A careful selection of the emulsion components and a better control of their production could further increase the radiopurity, thus extending the detector mass and exposure time. In particular, since the activity of the gelatin is higher than that of the other emulsion components (see Table \[tab:activities-MS\]) and since PVA shows a very low radioactive level, we are studying a possible replacement of gelatin with PVA.
In nuclear emulsion-based detectors the instrumental background is due to the so called *fog* grains, i.e. dark grains produced by thermal excitation. The fog density determines the probability of random coincidences of two or more fog grains mimicking a WIMP-induced nuclear recoil. The measured value of the fog density for current NIT samples is about 0.1 grains/(10$\mu$m)$^3$.
The number of background tracks due to random coincidences of fog grains depends on the minimum number of grains required to build a track and increases with the track length, as shown in the left plot of Figure \[fig:combinatorial\_bkg\], where the instrumental background for 1 kg emulsion target is reported. In NIT (U-NIT) emulsions a track made of 2 grains has an average length of about 100 nm (50 nm). The number of background tracks corresponding to this track length amounts to 10$^4$ (10$^3$), as outlined by red arrows on the plot.\
In order to make the combinatorial background smaller than one, the coincidence of at least 3 grains has to be required. In NIT (U-NIT) emulsions a track made of a sequence of 3 grains has on average a path length of about 200 nm (100 nm): the corresponding background level is 0.3 tracks ($4\times10^{-3}$ tracks).\
The right plot in Figure \[fig:combinatorial\_bkg\] shows the number of background tracks as function of the fog density in NIT emulsions, if 2-grain tracks are accepted: the background can be considered as negligible only reducing the fog density to from the current value to 10$^{-3}$ grains/(10$\mu$m)$^3$. Preliminary tests show that a value of 0.03 grains/(10$\mu$m)$^3$ can be obtained using purified gelatine. Further purification might lead to lower fog values. This research line will be followed in collaboration with the firm producing the gelatine.\
In order to further reduce the fog density, two possible improvements are under study. The first one exploits the response of fog grains to the polarized light scanning: fog grains show indeed both a different image contrast and a different size with respect to the grains sensitized by a nuclear recoil. This effect is essentially due to different $dE/dx$ of the two processes and offers a powerful discrimination of such kind of background. Moreover, a reduction of the fog density can be achieved operating the detector at low temperature (from a simple refrigeration down to a cryogenic regime of $\sim 80$ K) or by applying dedicated chemical treatments.
![Left: number of background tracks in 1 kg of NIT emulsions as function of the track length for tracks made by two (continuous red line) and three fog grains (dashed blue line). Right: number of background tracks in 1 kg of NIT emulsions as function of the fog density for 50 nm (continuous green line), 100 nm (dashed black line) and 200 nm (dotted-dashed magenta line) threshold in the track length. Only tracks made by two grains are considered here. []{data-label="fig:combinatorial_bkg"}](figs/combinatorial_bkg_arrow){width="1.0\linewidth"}
Finally, the requirement of a background-free experiment sets the necessity of operating in a clean environment in order to avoid surface contamination. Moreover in order to reduce the activation risk of detector materials, an underground location for the emulsion production and handling facilities is required. The construction of a (dark) clean room in the Gran Sasso underground Laboratory is therefore needed.
Threshold \[nm\] Fraction
------------------ ----------
50 0.100
100 0.075
150 0.060
200 0.052
: Fraction of detectable neutron-induced recoils as a function of the read-out threshold.[]{data-label="tab:recoils1"}
Experimental set-up {#sec:set-up}
===================
As a first phase of the project, we plan to perform a pilot experiment with a detector of 1 kg exposed for one year. Details of the related schedule will be examined in Section \[sec:schedule\].
A detector with one kg mass of NIT can be made of 50 $\mu$m thick-films assembled in a stack of 100 planes with a surface of $25 \times 25$ cm$^2$. We are considering the option of embedding OPERA-like films between two consecutive NIT planes: OPERA-like films would act as a monitoring system to register, with micrometric accuracy and high sensitivity, all the radiation integrated by the detector along the exposure. As composed of the same raw materials, the intrinsic radioactivity of the OPERA-like films would be of the same order of magnitude of that of NIT, therefore tolerable for a 1 kg detector.
The emulsion planes are placed with their surface parallel to the expected WIMP wind direction. We might consider to place an equivalent amount of emulsion films in an orthogonal plane. These films would act as a control sample. In case a signal would be found in thefirst sample, and only in this case, the scanning of these films would be performed to demonstrate that the signal found is not an artefact.
To maintain the detector with a fixed orientation towards the Cygnus constellation it will be installed on an Equatorial Telescope (see Figure \[fig:detector\]) allowing to cancel out the effect of the Earth rotation thus keeping the detector pointed on a fixed position in the sky.
An equatorial telescope has two axes: the so called Polar Axis, parallel to the rotation axis of the Earth and pointed to the North celestial pole, and the Declination Axis, perpendicular to the polar one. The motion of the Earth can be canceled out by driving at a constant speed the Polar Axis synchronised with the apparent daily motion of the sky. The Polar Axis will be motorized and both axes will be equipped with precise encoders to constantly check the position of the mechanics with high accuracy. The detector will be therefore pointed towards the Cygnus constellation and kept in that direction with an accuracy better than 1 degree.
A calibration procedure of the telescope will be performed before the installation in the underground laboratory. To ensure a precise syncronization of the mount with the apparent daily motion of the sky it is necessary to tune the response of the mechanics and to correct for any possible periodic error. The calibration procedure foresees several steps. The mount will at first be tested in the external laboratory using an optical telescope mounted on it and aligned with the Polar Axis: the telescope will be used, during the night, to point a star in the Cygnus constellation. Using a specific software and an imaging CCD camera attached to the prime focus of the telescope, the mount will be guided to keep the star centered in the field of view of the CCD camera. The software will record all the guiding parameters, as the position of the star and the corrections applied to the Polar and Declination Axis. This procedure will be repeated during several nights and all the data collected will be analyzed in order to get and apply the necessary correction to the mechanics and to the electronic system.
In a second phase the mount will be used throughout the whole day to compensate the apparent daily motion: the position during the night will be then measured in order to evaluate the pointing accuracy given by the difference between the nominal and measured positions. This measurement will provide a fine tuning on the position of both the Polar and the Declination axes.
Finally the mount will be moved underground in its final position: profiting of the presence, in the underground halls, of already existing high precision reference points the mount will be aligned with high accuracy in the north-south direction in order to align the Polar Axis parallel to the rotation axis of the Earth.
The design and construction of the equatorial telescope will be carried-out in collaboration with specialized firms. A screening of all the materials used in the construction of the telescope is foreseen in order to evaluate their intrinsic radioactivity. A detailed simulation of all the components of the telescope is planned.
A large telescope supporting both the target and the surrounding shield is considered (see Figure \[fig:detector\]). This configuration allows to build a light shield while ensuring a low contamination of the background originating from the telescope itself.
![Schematic view of the detector structure.[]{data-label="fig:detector"}](figs/Mount_1){width="0.7\linewidth"}
In Figure \[fig:detector\] a schematic view of the detector structure is shown: a stack of NIT films is placed at the center of a plexiglass sphere with a diameter of 30 cm. A sphere of polyethylene will act as a shield against the external neutron background. The target and the shielding are installed on the equatorial telescope. The target emulsions are arranged in such a way to have the film surface parallel to the WIMP wind.
From a preliminary evaluation a layer of 50 cm of polyethylene will reduce the external neutron flux by a factor of the order of $10^{4}$: considering an integrated flux of the order of $\phi_n \sim 2 \times 10^{-6}$ cm$^{-2}$ s$^{-1}$, for a target with an exposed surface of $25 \times 25$ cm$^2$ and a thickness of 0.5 cm this corresponds to a residual flux of the order of 1 neutron/kg/year, the same order of magnitude of the intrinsic neutron contamination. More accurate evaluations of the polyethylene thickness sufficient to provide the required background rejection power are under study. The addition of a thin ($1\div 2$ cm) layer of Cadmium to capture thermalised neutrons is under study.
As explained in Section \[sec:expected-bkg\] NIT have a high electron rejection power: a proper chemical treatment allows to reach a reduction factor of the order of 10$^{-6}$ in the sensitivity to electrons. For this reason the use of high-Z shielding materials (Pb and Cu) against the external $\gamma$ flux is not foreseen at the moment.
Both the passive shield and the emulsion target will be enclosed in a sealed plexiglass box maintained in High Purity (HP) Nitrogen atmosphere in slight overpressure with respect to the external environment to prevent radon contamination.
The shape of the shield surrounding the detector will be optimized in order to design the lighter and efficient structure. Two solutions are under study, either a parallelepiped box containing the shielding and the emulsion target, or a spherical one. In the first case the weight of the PE layer is of the order of 1.8 ton, while in the spherical option the weight is $\sim$1.14 ton. Even if the latter option ensures a lighter and symmetrical shielding, the final choice will depend on the cost and on the technical implementation of the design.
Nevertheless, we are also investigating a completely different approach based on the use of water as shielding material against the external background. A preliminary layout is shown in Figure \[fig:WaterOption\]: the emulsion detector is hermetically enclosed inside a spherical container made of low-Z material (teflon or polyethylene) with a diameter of 55 cm. The inner volume is flushed with N$_2$. The container is mounted on a long shaft and positioned in the centre of a tank (diameter 5 m, height 5 m) filled with ultra-pure water. The shaft is made of light, low-radioactive material (i.e. aluminum) aligned with Earth’s rotation axis. The constant orientation of the target with respect to the Cygnus constellation is kept thanks to the slow rotation of the shaft with the period of one sidereal day. All the mechanics needed to keep the orientation and the rotation of the shaft is mounted outside the water tank. The immersed part can be constructed in a way to keep the mean density close to 1 g/cm$^3$. In this way the mechanical load becomes negligible, thus simplifying the design and providing a big flexibility for materials selection.
This solution can be more flexible and cheaper, allowing to hold much larger masses without changing neither the mechanics of the telescope nor the shielding. A detailed simulation of the shielding and a study of the mechanics requirements, together with an estimation of the costs, are ongoing.
![Schematic view of the detector structure for the water shielding: the detector holder is placed in the centre of tank. Its orientation toward the Cygnus constellation is kept by the rotating pivot mounted with one edge above the water surface. Only pure and low-Z materials are used for the immersed part.[]{data-label="fig:WaterOption"}](figs/Mount_2){width="1.0\linewidth"}
![Sketch of the planimetry of the NIT production and development facility.[]{data-label="fig:emulsion-facility"}](figs/layout_CleanRoom_100m2_v3){width="0.85\linewidth"}
![A picture of the existing OPERA CS facility in hall B.[]{data-label="fig:CSemulsion-facility"}](figs/CSfacility){width="0.7\linewidth"}
![Planimetry the existing OPERA CS facility in hall B.[]{data-label="fig:CSemulsion-facility-planimetry"}](figs/planimetria_csFacility){width="0.7\linewidth"}
Emulsion production and development facility
--------------------------------------------
The layout of the facility we intend to build is shown in Figure \[fig:emulsion-facility\]. The total surface is about 100 m$^2$ and it is divided in four parts: emulsion gel production, emulsion gel pouring, film development and chemical solution preparation.\
Once produced, the gel will be sealed in an envelop flushed and filled with N$_2$ and moved to the pouring station, where a glove box flushed with HP Nitrogen will be installed. After the pouring, the films will be sealed in an envelop flushed and filled with N$_2$ and stored underground until the exposure.\
All the operations involving the emulsion production and development require a dark room environment.\
In order to minimize the surface contamination and the activation risk, the facility will be hosted in a clean room located underground. A class 1’000 clean room is required for the emulsion production, pouring and the chemical solutions preparation; a class 100’000 will be installed for the area devoted to the development.\
An air conditioning system will be installed in order to stabilize and monitor the temperature, ($20 \pm 1)^\circ$, and the humidity, ($60\pm 5)\%$, of the clean room. A demineralized water treatment plant and a chemical waste system are also required.
For the film development and the pouring activity foreseen in the first year of the project an excellent starting point is the existing OPERA emulsion handling facility shown in Figure \[fig:CSemulsion-facility\]. The facility, currently hosted in Hall B, is made of three rooms: a control room, a handling room and a development room, for a total surface of $\sim$ 50 m$^2$ (see Figure \[fig:CSemulsion-facility-planimetry\]). The handling room will be equipped with a pouring station and a development station. The installation of two systems for the temperature control is also foreseen.
The scanning of the exposed films will be performed in the existing OPERA scanning facilities in Italy, Russia, Turkey and Japan. In Italy the scanning laboratories are located at LNGS, Naples and Bari with 13, 5 and 3 OPERA microscopes respectively. Few more microscopes are currently located in Russian and Turkish scanning laboratories. Moreover at LNGS and Naples two additional microscopes, partially upgraded for the scanning of NIT and the polarized light analysis, are available. An equivalent scanning power is hosted at Nagoya University.
Physics reach
=============
The 90$\%$ C.L. upper limit in case of null observation is shown in Figure \[fig:sensitivity1Kg\] for an exposure of 1 kg$\cdot$year of NIT emulsions, with a minimum detectable track length ranging from 200 nm down to 50 nm and in the hypothesis of zero background. Even not including the directionality discrimination of the signal and assuming to reach a negligible background level, such an experiment would cover a large part of the parameter space indicated by the DAMA/LIBRA results with a small (1 kg) detector mass, using a powerful and complementary approach.
It is worth noting that the sensitivity strongly depends on the final detection threshold: as explained in Section \[sec:expected-bkg\] the current threshold value is limited to 200 nm only by the fog density. A reduction of the fog density or its discrimination through the use of the optical microscope with polarized light, would allow to lower the threshold to 100 nm. In order to lower the threshold down to 50 nm the use of the U-NIT technology is needed. Moreover we are conservatively assuming zero efficiency below the threshold value while, as shown in Figure \[fig:eff\_vs\_length\], the efficiency is not negligible even for shorter tracks. This would enhance the sensitivity to low WIMP masses. This effect will be taken into account.
![The 90$\%$ C.L. upper limits for a NIT detector with an exposure of 1 kg $\times$ year, a threshold ranging from 200 nm down to 50 nm, in the zero background hypothesis. The directionality information is not included.[]{data-label="fig:sensitivity1Kg"}](figs/NEWS_sensitivity1Kg_JP){width="0.6\linewidth"}
Schedule, Cost Estimate, Organization
=====================================
Time schedule {#sec:schedule}
-------------
{width="1.2\linewidth"}
On a time scale of six years we intend to perform the first exposure with a target mass of 1 kg and the corresponding analysis of the data taken. In Figure \[fig:gantt\] a detailed plan of all the phases of the project is reported.
In the beginning of 2016 we plan to construct a prototype shield and the equipment for the emulsion pouring. The above mentioned phases have to be completed in nine months in order to perform a first test to benchmark the level of intrinsic radioactivity of emulsions. For this measurement, we will use the gelatine produced at Nagoya University and perform the pouring underground. We will perform an exposure of a 10 g detector surrounded by the prototype shield. The detector exposure together with the analysis of the emulsion films will last nine months. The results of this test will provide a measurement of the background, intended to cross-check the estimates based on simulation and measurements of intrinsic radioactivity. In parallel, tests with radioactive sources are foreseen to characterize the response to external radioactivity.
We do consider the possibility of getting raw materials for the emulsion production within European countries, provided that their intrinsic radioactivity does not exceed the level measured in Japanese samples. This would allow a reduction of the activation processes induced during transportation. This activity will take place in 2016.
The measurement of intrinsic radioactivity of the different emulsion components and the prototype shield materials will be performed from June 2016 to the end of 2017. Tests of the activation due to cosmic rays during the transportation will be performed by bringing samples back and forth between Italy and Japan.
The design of the gel production machine will start in September 2016 while the design of the clean room will be carried on with the help of specialized firms, starting from January 2017.
The construction of the clean room, the pouring facility and the gel production machine will start from January 2018 and will last six months. As soon as the film production machine will be operational in the underground laboratory and the gelatine will be produced, the measurement of intrinsic radioactivity will be performed. If satisfying the required radioactivity level, the pouring of the gelatine will be performed.
The design of the equatorial telescope and the choice of the materials is supposed to start soon in 2016 and last 18 months. The construction of the telescope will start in the beginning of 2018 and last six months. In the second part of 2018 the surface calibration measurements and the underground telemetry will be carried out.
The construction of the shield and the target holding will start in 2018. Once the equatorial telescope installation will be finalised, the detector commissioning will start.
We plan to finalize the upgrade of the read-out system on a prototype microscope, exploiting in particular the resonant light scattering technique. This activity will start at the beginning of 2016.
In June 2017 a Technical Design Report will be submitted.
The upgrade of all the available OPERA systems will start in the second half of 2018 and last 27 months. By March 2020 we plan to have the final equipment installed on a number of microscopes adequate for the analysis of 1 kg of NIT emulsion in one year. Once the whole film production will be completed, the run with 1 kg mass detector will start. The data taking will last one year: from October 2019 to October 2020. The emulsion films will be developed soon after the exposure. The scanning and the analysis of the emulsion films will start once the upgrade of all the read-out systems will be complered and it is supposed to be completed by the end of 2021.
Costs
-----
The cost for the constructions of the clean room (75 m$^2$ class 1’000 and 25 m$^2$ class 100’000) is estimated to be around 200 k. As explained in Section \[sec:set-up\] the clean room will host the production machines, the pouring and the development facilities.\
The cost of the production machine is of the order of 200 k. The pouring and the development facilities will cost about 18 kand 50 k, respectively. The above mentioned costs will be shared according to a MoU to be signed between parties. In case the production will be carried out at Nagoya University, Japan will cover the corresponding costs. Japan will cover the costs for the all the emulsion components.\
A first estimate of the cost for the equatorial telescope is 15 k for the design and 240 k; the cost for the construction of the shielding amounts to about 15 k.\
Finally the upgrade of the read-out systems will be needed. Japan will cover the cost for the realization of their own scanning systems. The construction of the microscope prototype in Europe costs about 300 k; the hardware and computing upgrade of each OPERA microscope amounts to about 30 k. Depending on the final scanning speed, from 10 to 14 systems will be modified for the high resolution and high speed scanning of NIT for a total cost ranging from $\sim$ 300 kto $\sim$ 420 k.\
An expense of 80 kis expected for the maintenance of the microscopes and 120 kfor the consumables.
In Table \[tab:costs\] a summary of the expected costs is reported.
[c | c | c]{} Category & Cost \[k\] & Assignment\
Clean Room & 200 & EU\
NIT production machine & 200 & JP\
Pouring facility & 18 & EU\
Development facility & 50 & EU\
Equatorial Telescope & 255 & EU\
Shielding & 15 & EU\
EU Prototype Microscope & 30 & EU\
EU Microscopes Upgrade & 300 & EU\
EU Microscope Maintenance & 80 & EU\
JP Microscopes Upgrade & 300 & JP\
Consumables & 120 & EU\
TOTAL & 1468 &\
Collaboration
-------------
NEWS is at present a collaboration between Italy, Japan and Russia and Turkey.
The involved groups are:
- University and INFN Bari, Italy
- Lab. Naz. Gran Sasso, Italy
- University and INFN Naples, Italy
- University and INFN Rome, Italy
- Nagoya University and KM Institute, Japan
- Chiba University, Japan
- JINR Dubna, Russia
- Moscow State University, Moscow, Russia
- Lebedev Physical Institute, Moscow, Russia
- METU, Ankara, Turkey
All the above mentioned groups are leaders in the emulsion scanning having gathered the experience of the emulsion analysis in in the OPERA experiment. The scanning and the analysis of the exposed emulsions will be shared according to the available scanning power of each group.
The development of the prototype, both for hardware and software, of the new read-out system is shared between LNGS and Naples while it is entirely carried out at Nagoya University for the Japanese one.
The LNGS and Napoli groups are in charge of the design of the telescope, the construction of the prototype and the calibration measurements. The same groups will perform the intrinsic background measurements, the studies about the environmental background and the design of the detector shielding and structure.
The Russian groups will perform radioactive studies.
The design and the realization of the local underground facilities will be shared between LNGS and Japan.
The simulation of the detector response, efficiency and resolution as well as and the expected sensitivity is shared between LNGS, Naples and Nagoya.
The responsibility about the emulsion production, development and handling is currently assigned to the Nagoya group.
Conclusions and outlook
=======================
![Sensitivity at 90$\%$ C.L, in the zero background hypothesis for an experiment with a mass of 10 kg (green) and 100 kg (blue) for two value of detection threshold: 100 nm (dashed lines) and 50 nm (solid line). []{data-label="fig:NEWSsensitivity_10-100Kg_50-100nm_JP"}](figs/NEWSsensitivity_10-100Kg_50-100nm_JP_2){width="0.8\linewidth"}
NEWS is meant to be the first experiment with a solid target for directional dark matter searches: the use of a nuclear emulsion based detector, acting both as target and tracking device, would allow to explore the low cross section region in the phase space indicated by DAMA.
The novel emulsion technology, based on the use of nuclear emulsion with nanometric AgBr crystals (NIT), makes it possible to record the sub-micrometric tracks produced by the WIMP scattering off a target nucleus. The presence, in the emulsion components, of light and heavy nuclei results in an enhanced sensitivity to both light and heavy WIMP masses.
The read-out of tracks with length of the order of 100 nm, is possible thanks to an R$\&$D carried out on the scanning systems currently used for the analysis of the OPERA emulsions. The use of improved optics and mechanics allowed to reach a spatial and angular resolution of the order of 100 nm and 235 mrad, respectively, with a tracking efficiency approaching 100$\%$ for tracks with lengths longer than 180 nm. The new optical microscope has a scanning speed of about 25 mm$^2$/h allowing to perform a fast preselection of the candidate signal tracks with the shape analysis method.
The final signal confirmation is obtained with powerful optical microscope equipped with a light polarizer: exploiting the different response of non spherical grain clusters to different polarization angles, the unprecedented spatial resolution of 10 nm is obtained. This resolution allows to resolve grains belonging to a few hundred of nanometer long tracks thus providing the final signal confirmation with a very high signal to noise ratio.
The intrinsic radioactivity of nuclear emulsions has been measured and a detailed MC simulation has been performed: the estimated neutron yield allows to design an experiment with masses of the order of 10 kg while keeping this background negligible. A careful evaluation of the external background sources has been performed allowing to design a proper shielding. The final experimental set-up foresees the use of an equatorial telescope holding both the emulsion target and the shielding.
We plan to perform a pilot experiment with a 1 kg mass target on a time scale of six years: even using a rather small detector mass we would be able to explore the region indicated by the DAMA experiment with a powerful and complementary approach (see Figure \[fig:sensitivity1Kg\]).
The actual intrinsic radioactive level allows to scale the target mass and exposure time up to one order of magnitude. A careful selection of the emulsion components and a better control of their production could further increase the radiopurity, thus allowing larger detector mass. The reduction of the fog density and further developments of the optical microscopy with polarized light would allow to reduce the detection threshold down to 50 nm. Improvements both in the mechanics (use of piezoelectric-driven objective) and in the image acquisition (use of multiple image sensors) envisage already now the possibility to analyse with such a resolution a volume of 100 kg or larger. Moreover further improvements both in the microscope hardware and in the analysis software will permit to fully exploit the intrinsic emulsion capability of recording 3D tracks.
In Figure \[fig:NEWSsensitivity\_10-100Kg\_50-100nm\_JP\] the upper limit in case of null observation for an experiment with a mass of 10 (green) and 100 (blue) kg and for a detection threshold of 50 (dashed lines) and 100 (solid lines) nm is shown at 90 $\%$ C.L and in the zero background hypothesis.
The proposed program would open a new window in the DM search. The developments done will likely have impact on the nano-imagining applications in physics, biology and medicine.
[00]{}
K.A. Olive et al. (Particle Data Group), Chin. Phys. C, **38** (2014) 090001,\
Plank Collaboration, *Plank 2015 results. XI. CMB power spectrum, likelihoods, and robustness of parameters*, arXiv:1507.02704. G. Bertone, Particle Dark Matter, Cambridge University Press, 2010. M. W. Goodman and E. Witten, Phys. Rev. **D31** (1985) 3059. D. N. Spergel, Phys. Rev. **D37** (1988) 1353. R. Bernabei et al., Eur. Phys. J. **C56** (2008) 333. DMTOOLS site: `http://dmtools.brown.edu:8080/`. C. E. Aalseth et al., CoGeNT Collaboration, *CoGeNT: A search for low-mass dark matter using p-type point contact germanium detectors*, Phys. Rev. **D88** (2013) 012002. G. Angloher et al., CRESST-II Collaboration, *Results on low mass WIMPs using an upgraded CRESST-II detector*, Eur. Phys. J. **C74** 12 (2014) 3184. Z. Ahamed et al., *Combined limits on WIMPs from the CDMS and EDELWEISS Experiments*, Phys. Rev. D 84 (2011) 011102 E. Aprile et al., XENON100 Collaboration, *Dark matter results from 225 live days of XENON100 data*, Phys. Rev. Lett. **109** (2012) 181301. D. S. Akerib et al., *First results from the LUX dark matter experiment at the Sanford Underground Research Facility*, Phys. Rev. Lett. **112** (2014) 091303. O. Buchmueller et al., *Implications of initial LHC searches for Supersymmetry*, Eur.Phys.J. C71 (2011) 1634. S.P. Ahlen et al., *Time-projection-chambers with optical readout for dark matter, double beta decay, and neutron measurements*, Int. J. Mod. Phys. **A25** (2010) 4525. K. Miuchi, H. Nishimura et al., *First underground results with NEWAGE-$0.3$a direction-sensitive dark matter detector*, Phys. Lett. **B686(1)** (2010) 11. S. Ahlen, J. Battat et al., *First dark matter search results from a surface run of the 10-L DMTPC directional dark matter detector*, Phys. Lett. **B695(1-4)** (2011) 124. J. Billard, F. Mayet et al., *Directional detection of dark matter with MIMAC: WIMP identification and track reconstruction*, J. Phys. Conf. Ser. **309** (2011) 012015. E. Eskut et al., *The CHORUS experiment to search for muon-neutrino $\to$ tau-neutrino oscillation*, Nucl. Instrum. Meth. **A401** (1997) 7.\
S. Aoki, E. Barbuto, C. Bozza, J. Fabre, W. Flegel, et al., *Nuclear emulsions in a large, hybrid experiment (CHORUS) to search for $\nu_\mu \to \nu_\tau$*, Nucl. Instrum. Meth. **A447** (2000) 361. OPERA Collaboration, *The OPERA experiment in the CERN to Gran Sasso neutrino beam*, JINST [**4**]{} (2009) P04018. SHiP Collaboration, *A facility to Search for Hidden Particles (SHiP) at the CERN SPS*, arXiv:1504.04956 \[physics.ins-det\]. P.H. Fowler, D.H. Perkins and C.F. Powell, *The study of elementary particles by the photographic method*, Pergamon Press (1959). W.H. Barkas, *Nuclear research emulsion*, Academic Press, New York, 1973. G. De Lellis, A. Ereditato and K. Niwa, Nuclear Emulsions, in Handbook of Physics, Vol. 2; C.W. Fabjan and H. Schopper (Eds.), (2011) Springer Publishers. T. Nakamura, A. Ariga, T. Ban, T. Fukuda et al., *The OPERA film: New nuclear emulsion for large-scale, high-precision experiments*, Nucl.Instrum.Meth. **A556** (2006) 80-86. M. Natsume et al., *Low-velocity ion tracks in fine grain emulsion*, Nucl. Instr. Meth. **A575** (2007) 439. T. Naka et al., *Fine grained nuclear emulsion for higher resolution tracking detector*, Nucl. Instrum. Meth. **A718** (2013) 519. N. Armenise et al., Nucl. Instr. Meth. **A551** (2005) 261\
L. Arrabito et al., Nucl. Instr. Meth. **A568** (2006) 578\
M. De Serio et al., Nucl. Instr. Meth. **A554** (2005) 247\
L. Arrabito et al., *JINST* **2** (2010) P05004\
C. Bozza et al., *Nucl. Instrum. Meth. A* **703** (2013) 204 K. Morishima and T. Nakano, *Development of a new automatic nuclear emulsion scanning system, S-UTS, with continuous 3D tomographic image read-out*, *JINST* **5** (2010) P04011.\
S. Aoki et al., *The fully automated emulsion analysis system*, *Nucl. Instrum. Meth. B* **51** (1990) 466\
T. Nakano, *Automatic analysis of nuclear emulsion*, Ph.D. thesis, Nagoya University, Japan (1997).\
OPERA Collaboration, JINST [**4**]{} (2009) P06020. A. Alexandrov, V. Tioukov, M. Vladymyrov, *Further progress for a fast scanning of nuclear emulsions with Large Angle Scanning System*, JINST **9** (2014) C02034. T. Naka, et al., *R$\&$D Status of Nuclear Emulsion For Directional Dark Matter Search*, EAS Publ. Ser. **53** (2012) 51-58. M. Kimura and T. Naka, *Submicron track readout in fine grained nuclear emulsions using optical microscopy*, Nucl. Instrum. Meth. **A680** (2012) 12-17 Naka et al., Rev. Sci. Instrum. **86** (2015) 073701 H. Tamaru et al., *Resonant light scattering from individual Ag nanoparticles and particle pairs*, Applied Phys. Lett. **80** (2002) 1826. G. Bruno, *Neutron Background studies for direct dark matter searches in the Gran Sasso Underground Laboratory*, PhD Thesis, L’Aquila University (2012). Arneodo et al., *Neutron background measurements in the Hall C of the Gran Sasso Laboratory*, Nuovo Cim. **A112** (1999) 819. Wulandari et al., *Neutron flux underground revisited*, Astropart.Phys. **22** (2004) 313. K. I. Nagao and T. Naka, Prog. Theor. Exp. Phys. (2012) 043B02. T. Habu, N. Mii, K. Kuge, H. Manto, Y. Takamuki, J. Imaging Sci. **35** (1991) 202. J. S. Becker, Inorganic Mass Spectrometry - Principles and Applications (Wiley, 2007), ISBN 9780470012000. M. Laubenstein et al., Appl. Radiat. Isot. **61** (2004) 167. F. Pupilli et al., *Intrinsic neutron background of nuclear emulsions for directional Dark Matter searches*, submitted to Atrophys. J., arXiv:1507.03532 (astro-ph). J. K. Shultis, R. E. Faw, Fundamentals of Nuclear Science and Engineering (CRC press, 2007), p. 141, ISBN 1420051369. R. Heaton, H. Lee, P. Skensved and B. C. Robertson, Nucl. Instrum. Meth. **A276** (1989) 529. W. B. Wilson et al., *SOURCES 4A: A Code for Calculating ($\alpha$,n), Spontaneous Fission, and Delayed Neutron Sources and Spectra*, LA-13639-MS, Los Alamos (1999).
| {
"pile_set_name": "ArXiv"
} |
[Theory and Phenomenology of $\mu$ in $M$ theory\
]{} Bobby Samir Acharya$^{1,2}$, Gordon Kane$^1$, Eric Kuflik$^1$, Ran Lu $^1$\
^1^[*Michigan Center for Theoretical Physics, University of Michigan, Ann Arbor, MI 48109*]{}\
^2^[*Abdus Salam International Centre for Theoretical Physics, Trieste, Italy* ]{}
We consider a solution to the $\mu$-problem within $M$ theory on a $G_2$-manifold. Our study is based upon the discrete symmetry proposed by Witten that forbids the $\mu$-term and solves the doublet-triplet splitting problem. We point out that the symmetry must be broken by moduli stabilization, describing in detail how this can occur. The $\mu$-term is generated via Kahler interactions after strong dynamics in the hidden sector generate a potential which stabilizes all moduli and breaks supersymmetry with $m_{3/2} \sim 20 - 30 \operatorname{TeV}$. We show that $\mu$ is suppressed relative to the gravitino mass, by higher dimensional operators, $\mu \sim 0.1 m_{3/2} \sim 2-3 \operatorname{TeV}$. This necessarily gives a Higgsino component to the (mostly Wino) LSP, and a small but non-negligible LSP-nucleon scattering cross-section. The maximum, spin-independent cross-sections are not within reach of the current XENON100 experiment, but are within reach of upcoming runs and upgrades.
Introduction
============
In the Minimal Supersymmetric Standard Model (MSSM) [@Martin:1997ns], the only low energy supersymmetric parameter with mass dimension one is the $\mu$ parameter. Through the $\mu$-term, $W \supset \mu H_u H_d $, it gives mass to Higgsinos and also generates scalar potential couplings for Higgs fields. The size of $\mu$ plays an important role in phenomenology. In particular, it affects properties of potential dark matter particles. LEP searches for the charged Higgsino require $\mu \gtrsim 100$ GeV, while arguments against fine tuning of the mass of the Z-boson suggest that $\mu$ should not be too large. On the other hand, one might expect, with ignorance of the high scale theory, that $\mu \sim m_{GUT}$, the natural UV cutoff. Solving the $\mu$-problem [@Kim:1983dt] presumably requires an understanding of the fundamental theory that generates the scale of the $\mu$ parameter. Thus the $\mu$-problem is exceptionally important–a high scale theory cannot be qualitatively complete without addressing it, and its solution will have significant implications for dark matter, Higgs physics, and fine-tuning issues.
The most promising framework for a complete fundamental theory that incorporates low energy supersymmetry is string theory. Within string theory, many explanations for the small value of $\mu$ have been proposed. In most scenarios the $\mu$-term is forbidden at the high scale. Then, it is somehow dynamically generated at a lower scale. In many cases, the $\mu$-term is forbidden by a continuous or discrete symmetry, which is spontaneously broken at a smaller, dynamically generated scale ($\ll m_{GUT}$), and perhaps related to supersymmetry breaking [@Antoniadis:1994hg; @Nath:2002nb]. Some examples of the above include NMSSM scenarios [@Suematsu:1994qm; @Cvetic:1995rj; @Cvetic:1997ky; @Lebedev:2009ag; @RamosSanchez:2010sc; @Ratz:2010zz] and approximate $R$-symmetric models [@Casas:1992mk; @Kappl:2008ie]. Others scenarios have the $\mu$-term forbidden by stringy selection rules, and are broken by non-perturbative instanton effects that produce exponentially suppressed mass scales [@Ibanez:2007tu; @Ibanez:2008my; @Green:2009mx; @Cvetic:2009yh].
It has long been suspected that the MSSM unifies the strong and electroweak forces [@Langacker:1991an] into a single $SU(5)$ grand unified group. Each family of quarks and leptons are organized into a $\bf{10} \oplus \bf{\bar{5}}$ representation of $SU(5)$. The remaining MSSM fields, the Higgs doublets, do not form a complete $SU(5)$ representation. Minimally, the Higgs doublets can be assigned to a $\bf{5}\oplus \bf{\bar{5}}$ representation, but require the introduction of a pair of Higgs color triplets. The Higgs triplets can mediate baryon and lepton violating processes, and thus should be very heavy, $m_{T} \gtrsim 10^{14}$ GeV, to avoid rapid proton decay [@Murayama:2001ur]. Additionally, they should be heavy to ensure gauge coupling unification in the minimal model. If Higgs triplet masses are very heavy, then an $SU(5)$ symmetric theory would require that the Higgs doublets masses be the same as the triplet mass, $\mu = m_T \sim M_{GUT} $, but it was just argued that is this a factor $10^{13}$ too large. A string theoretic solution to the $\mu$-problem is inevitably related to the solution of the doublet-triplet problem of grand unified theories.
Therefore, it is paramount that the symmetry that protects the $\mu$-term not forbid the triplet masses if both problems are to be solved. This restriction leads to an elegant, perhaps unique solution to the $\mu$-problem in $M$ theory; the symmetry which protects $\mu$ from being generated at the unification scale iwas originally proposed by Witten [@Witten:2001bf]. Although Witten did not discuss how this symmetry would be broken, we argue that the symmetry would–indeed must– be broken by moduli stabilization. Then by including the mechanism for stabilizing the moduli proposed in [@Acharya:2006ia], we will show that $\mu \sim 0.1\,m_{3/2}$. Finally, the implications for dark matter discovery are discussed, where we conclude that the XENON100 experiment should not observe a dark matter signal, but may do so in its next upgrade (Figure \[fig:xsecuniversal\]).
$M$ theory
==========
Matter and Gauge Theory
-----------------------
In $M$ theory compactified on a $G_2$ manifold, $ADE$ gauge symmetries ($SU(n)$, $SO(2n)$ and $E_6$, $E_7$, $E_8$) are localized along three dimensional submanifolds of orbifold singularities [@Acharya:2000gb; @Acharya:1998pm]. Chiral matter, charged under the $ADE$ gauge theory, is localized at conical singularities in the seven dimensional $G_2$ manifold, at points where the $ADE$ singularity is enhanced [@Witten:2001uq; @Acharya:2001gy; @Acharya:2004qe]. Matter will additionally be charged under the $U(1)$ symmetry, corresponding to the vanishing $2$-cycle that enhances the singularity. Hence, all chiral matter will charged under at least one $U(1)$ symmetry. Bi-fundamental matter, charged under two non-Abelian gauge groups, is also possible, but will not be considered here.
As argued by [@Pantev:2009de], the additional $U(1)$ symmetries are never anomalous. Therefore, there is no Green-Schwarz mechanism [@Green:1984sg] needed for anomaly cancellation, and GUT-scale FI $D$-terms are not present in the theory. This will be important later, since it removes a possibility for generating large scalar vacuum expectation values (vevs) for charged matter fields.
Two gauge theories will generically only have precisely the same size gauge coupling if they arise from the same orbifold singularities. Therefore, if gauge coupling unification is to be motivated theoretically, and not an approximation or accident, the gauge group of the $ADE$ singularity should be a simple group containing the Standard Model gauge group, which we will take (for simplicity) to be $SU(5)$. Any larger group containing $SU(5)$ will give results similar to those we find below. To obtain the Standard Model gauge group, $SU(5)$ needs to be broken. Perhaps the 4D gauge symmetry can be broken spontaneously, but only representations smaller than the adjoint are realizable in $M$ theory–the $\bf{10}$ and $\bf{5}$ representations (and their conjugates) in $SU(5)$. This leaves only “flipped $SU(5)$” [@Barr:1981qv; @Derendinger:1983aj; @Antoniadis:1987dx] as a possible mechanism to break the GUT group and solve doublet-triplet splitting. Given the difficulty in constructing a realistic flipped $SU(5)$ model [@Kuflik:2010dg], it will not be considered here. The remaining possibility is to break the higher dimensional gauge theory by Wilson lines and will be discussed below.
Moduli Stabilization
--------------------
In the mid-80’s it was realized that, classically, string vacua contain a plethora of moduli fields. The standard lore was that, after supersymmetry breaking, the moduli fields would obtain masses and appropriate vacuum expectation values. Part of this lore was also the idea that strong dynamics in a hidden sector would be responsible for breaking supersymmetry at, or around, the TeV scale. Though some progress was made, it was not until recently that it has been clearly demonstrated that these ideas can be completely realized in string/$M$ theory: in $M$ theory compactified on a $G_2$-manifold (without fluxes) strong gauge dynamics can generate a potential which stabilizes all moduli and breaks supersymmetry at a hierarchically small scale [@Acharya:2006ia; @Acharya:2007rc]. These vacua will be the starting point for our considerations.
In these vacua, the gravitino mass (and therefore also the moduli masses [@Acharya:2010af]) $m_{3/2} \sim {\Lambda^3 \over m_{pl}^2}$, where $\Lambda$ is the strong coupling scale of the hidden sector gauge interaction. This is parametrically of order $\Lambda \sim e^{ -2\pi /( \alpha_{h}b)} m_{pl}$, where $\alpha_h$ is the coupling constant of the hidden sector and $b$ is a beta-function coefficient. The vacuum expectation values of the moduli fields are also determined in terms of $\alpha_h$: Roughly speaking, one has: s\^A \~1/\_h where the modulus here is dimensionless and not yet canonically normalized. The physical meaning of the vevs of $s^A$ is that it characterizes the volumes in eleven dimensional units of 3-cycles in the extra dimensions, e.g., the 3-cycle that supports the hidden sector gauge group. Thus, self-consistently when the hidden sector is weakly coupled in the UV, the moduli are stabilized at large enough volumes in order to trust the supergravity potential which only makes sense in this regime. In general, the rough formula exhibits the scaling with $\alpha_h$ and, numerically the moduli vevs in the vacua considered thus far range from about $1 \leq s^A \leq 5/\alpha_h$.
In order to incorporate the moduli vevs into the effective field theory in an $M$ theory vacuum, we have to consider the normalized dimensionful vevs which appear in the Einstein frame supergravity Lagrangian. For obtaining the normalization it suffices to consider the moduli kinetic terms alone: m\_[pl]{}\^2 [12]{} g\_[AB]{} \_ s\^[A]{} \^ s\^[B]{} \[modulikin\] where $s^A$ are the dimensionless moduli described above and $g_{AB}$ is the (Kahler) metric on the moduli space. From the fact that the extra dimensions have holonomy $G_2$, it follows that each component of $g_{AB}$ is homogeneous of degree [*minus*]{} two in the moduli fields g\_[AB]{} = \_A \_B K = \_A \_B(-3 V\_7+ …) because the volume of $X$, $V_7$ is homogeneous of degree $7/3$.
For isotropic $G_2$-manifolds, i.e. those which receive similar order contributions to their volume from each of the $N$ moduli, studying examples shows that, not only is the metric of order ${1 \over s^2}$, but also of order $1/N$: g \~
Therefore in a given vacuum the order of magnitude of the entries of $g_{AB}$ are g \~
Therefore, a dimensionless modulus vev of order $1/\alpha_h$ translates into a properly normalized dimensionful vev \~ \~0.1 m\_[pl]{} for $N\sim 100$, which is a typical expectation for the number of moduli [@joyce][^1].
This can lead to a suppression of the effective couplings which generate the $\mu$-term, once the symmetry forbidding $\mu$ is broken. More precise calculations of the moduli vevs can be found in [@Acharya:2007rc; @Acharya:2008zi]. Clearly, however, a $G_2$-manifold with less than ten or so moduli will not have suppressed, normalized moduli vevs; such cases are presumably unlikely candidates for $G_2$-manifolds with realistic particle spectra and will not be considered further.
We briefly also discuss the spectrum of Beyond Standard Model (BSM) particles which arise from the $M$ theory vacuum. Classically, it is well known that string/$M$ theory has no vacuum with a positive cosmological constant (de Sitter minimum). From the effective field theory point of view, this is the statement that moduli fields tend to have potentials which, in the classical limit have no de Sitter minimum. If we now consider quantum corrections to the moduli potential, which [*only*]{} involve the moduli fields – if they are computed in a perturbative regime – they tend to be small and hence are unlikely to generate de Sitter vacua. Positive, larger sources of vacuum energy must therefore arise from other, non-moduli fields. This is indeed the case in the $M$ theory vacua described in [@Acharya:2007rc]. Here the dominant contribution to the vacuum energy arises from a [*matter*]{} field in the hidden sector (where it can be shown that, without the matter field, no de Sitter vacuum exists). This is important for the following reasons.
Adopting supersymmetric terminology, this suggests that the fields with the dominant $F$-terms are not moduli. Hence, the moduli $F$-terms are suppressed relative to the dominant contribution (in fact, in $M$ theory the suppression is of order $\alpha_h$). This affects the spectrum of BSM particles. In string/$M$ theory, gaugino masses are generated through $F$-terms of moduli vevs (because the gauge coupling function is a superfield containing volume moduli). Hence, at leading order these will be suppressed relative to, say, scalar masses which receive order $m_{3/2}$ contributions from all $F$-terms in the absence of accidental symmetries. Therefore, in the $G_2$-MSSM (and presumably other classes of string vacua) the scalar superpartners and moduli fields will have masses of order $m_{3/2}$ whereas the gaugino’s will have masses which are suppressed; in fact in the $G_2$-MSSM the gaugino masses at the GUT scale are at least two orders of magnitude below $m_{3/2}$. This is what makes the anomaly mediated contributions to gaugino masses relevant to the $G_2$-MSSM and also why the models often contain a Wino LSP [@Acharya:2008zi]. Important for our considerations below will be the fact that the suppression of the gaugino masses is greater than the suppression of moduli vevs discussed above by one order of magnitude (at the GUT scale), at least for $G_2$-manifolds with less than O($10^4$) moduli.
Geometric Symmetries and Moduli Transformations
-----------------------------------------------
Compact, Ricci-flat manifolds with finite fundamental groups, such as manifolds with holonomy $G_2$ or $SU(3)$ can not have continuous symmetries. They can, however, have [*discrete*]{} symmetries. Witten was considering just such a discrete symmetry ($G$) of a $G_2$-manifold when he proposed the symmetry which prevents $\mu$. Assuming the simplest possibility of an Abelian discrete symmetry, let us consider $G = {\bf Z_N}$, which acts on $X$: : X X As a result of this, it will also act naturally on the fields on $X$. In particular ${\bf Z_N}$ will act on the set of harmonic forms on $X$. Our interest here is $H^3 (X, {\bf R})$ the set of harmonic 3-forms on $X$, since this locally represents the moduli space of $G_2$-manifolds. A $G_2$-manifold, with moduli at a point $\langle s^S \rangle = s_0^A$ is determined by a harmonic (locally) $G_2$ invariant 3-form $\varphi$ as = s\_0\^A \_A where $\beta_A$ are a basis for $H^3 (X, {\bf R})$. If the point $s^A_0$ is such that ${\bf Z_N}$ is a symmetry, then $\varphi$ will be invariant under ${\bf Z_N}$, because invariance of $\varphi$ is equivalent to invariance of the metric. The three-forms $\beta_A$ transform in a representation of ${\bf Z_N}$, which is a real representation because the 3-forms are real on a $G_2$-manifold. Hence, : \_A M\_A\^B\_B where $M$ is defined by this equation.
The fact that the particular $G_2$-manifold, characterized by the particular point in moduli space $s_0^A$, is ${\bf Z_N}$-invariant is simply the statement that: s\_0\^B M\_A\^B = s\_0\^A i.e., the $s_0^A$ are an eigenvector of $M$ with unit eigenvalue. Clearly, this will not be true for a generic vector $s^A$; hence, for a generic point in the moduli space, the entire ${\bf Z_N}$ symmetry will be broken. Since the representation of ${\bf Z_N}$ defined by the matrix $M$ is real, it must be the sum of a complex representation plus its conjugate. Thus, the basis $\beta_B$ can be chosen such that the complex representation is spanned by [*complex*]{} linear combinations of moduli fields. For instance, there might be a linear combination S = \^1 + i \^2 \[complexmoduli\] which we choose to write in-terms of the dimensionful fields ($\hat{s}$), that transforms as S e\^[2i/N]{} S.
Since we usually consider complex representations of discrete symmetries acting on the matter fields in effective field theories, it will be precisely the linear combinations of moduli (those in the form (\[complexmoduli\])) which span ${\bf r_C}$ which will appear in the “symmetry breaking sector” of the effective Lagrangian. In other words, the moduli will appear in complex linear comibinations such as (\[complexmoduli\]) in the Kahler potential operators containing other fields that transform under the ${\bf Z_N}$. Note that in (\[complexmoduli\]) we are abusing notation in the sense that the “$i$” which appears is in general an $N$-by-$N$ matrix whose square is minus the identity.
Witten’s Solution
=================
In heterotic and type-II string theories doublet-triplet splitting is often solved via orbifold compactifications [@Hosotani:1983vn; @Witten:1985xc]. In these theories, higher (space-time) dimensional gauge symmetries are broken by the Wilson lines in an orbifold compactification, while the Kaluza-Klein zero mode Higgs triplets are absent due to non-trivial transformations under the orbifold symmetry. On the contrary, matter fields in $M$ theory are co-dimension 7, that is, the fields live only in four dimensions, and are not zero modes of a KK tower of fields, so this solution to the $\mu$-problem will not work. Other possibilities, such as NMSSM realizations or string instanton effects, will also not work since the symmetry that forbids $\mu$ (a $U(1)$ or stringy selection rules) would also forbid the triplet mass, thus spoiling doublet-triplet splitting.
One may also consider the possibility that a discrete $R$-symmetry can forbid the $\mu$-term while solving doublet-triplet splitting. Requiring the symmetry to be anomaly free, and that it commutes with the gauge theory can lead to a unique symmetry [@Lee:2010gv]. However, this symmetry will also forbid the triplet mass and spoil doublet triplet splitting unless the triplets are absent from the four dimensional theory. For most string theories, this can be accomplished by a Wilson line in the higher dimensional theory, but in $M$ theory, this is not possible since matter only exists in four dimensions.
Therefore, an alternative approach is needed to solve doublet-triplet splitting in $M$ theory. The only known possibility, originally discussed by Witten, is to construct a discrete $ {\bf Z_N}$ symmetry of the geometry, that will act on both matter fields and moduli-fields. When combined with a discrete Wilson line thats breaks $SU(5)$, this symmetry need not commute with the $SU(5)$, thus allowing components of a single $SU(5)$ representation to have different ${\bf Z_N}$ charges. Since the above arguments demonstrate that there must be a symmetry that acts differently on doublets and triplets, so far this is the only approach known to work, and maybe be the only solution.
The minimal $SU(5)$ matter content contains three generations of matter descending from three copies of $\bf{10}_M \oplus \bf{\bar{5}}_M$. There is also a $\bf{5}_H \oplus \bf{\bar{5}}_{H}$ pair containing the MSSM Higgs doublets, $H_u \oplus H_d$, and a vector-like pair of Higgs triplets, $T_u \oplus T_d$. Here a doublet and a triplet from one of the Higgs representations can transform differently under the ${\bf Z_N}$ symmetry group. Without loss of generality or phenomenology, this field is taken to be the $\bf{\bar{5}}_{H}$ field, with the following charges for the fields $$\begin{array}{rc|c}
\multicolumn{2}{l}{\mbox{Field}} & {\bf Z_N} \\ \hline
\multicolumn{2}{c|}{ \mathbf{10}_M } & \eta^\sigma \\
\multicolumn{2}{c|}{ \mathbf{\overline{5}}_M } & \eta^\tau \\
\multirow{2}{*}{$\mathbf{5}_H $} & T_u & \eta^\alpha \\
& H_u & \eta^\alpha \\
\multirow{2}{*}{$\mathbf{\overline{5}}_{\overline{H}} $} & T_d & \eta^\gamma \\
& H_d & \eta^\delta
\end{array}$$ where $\eta\equiv e^{2\pi i / N}$.
These charges are constrained by the requirement that the $\bf Z_N$-symmetry does not forbid necessary terms in the superpotential, such as Yukawa couplings, Majorana neutrino masses, and the Higgs triplet masses $$\begin{array}{rc|c}
\multicolumn{2}{c}{\mbox{Coupling}\;\;\;\;\;} & \mbox{Constraint }\\ \hline
\mbox{Up Yukawa Coupling} & \mathbf{10}_M \mathbf{10}_M H_u & 2 \sigma + \alpha = 0 \mod N\\
\mbox{Down Yukawa Coupling} & \mathbf{10}_M \mathbf{\overline{5}}_M H_d & \sigma + \tau + \delta = 0 \mod N\\
\mbox{Majorana Neutrino Masses} & H_d H_d \mathbf{\overline{5}}_M \mathbf{\overline{5}}_M & 2 \alpha + 2 \tau = 0 \mod N\\
\mbox{Triplet Masses} & T_u T_d & \alpha + \gamma= 0 \mod N.\\
\end{array}$$ The solution to these equations are $$\begin{array}{ccl}
\alpha &=& -2 \sigma \\
\gamma &=& 2 \sigma \\
\delta &=& -3 \sigma + N/2 \\
\tau &=& 2 \sigma + N/2 \\
\sigma &=& \sigma.
\end{array}$$ Inherently, the $\bf Z_N$ should forbid the $\mu$-term, and if possible, other dangerous terms, such as dimension-5 proton decay operators and dimensions 3 and 4 R-parity violation. $$\begin{array}{rc|c}
\multicolumn{2}{c}{\mbox{Coupling}} & \mbox{Constraint }\\ \hline
\mu-\mbox{term} & H_d H_u & -5 \sigma + N/2 \ne 0 \mod N\\
\mbox{D-5 Proton Decay} & \mathbf{10}_M \mathbf{10}_M \mathbf{10}_M \mathbf{\overline{5}}_M & 5\sigma - N/2 \ne 0 \mod N\\
\mbox{D-3 R-Parity} & \mathbf{5}_H \mathbf{\overline{5}}_M & N/2 \ne 0 \mod N\\
\mbox{D-4 R-Parity} & \mathbf{10}_M \mathbf{\overline{5}}_M \mathbf{\overline{5}}_M & 5\sigma \ne 0 \mod N.\\
\end{array}$$ Doublet-triplet splitting occurs if $5 \sigma \ne N/2 \mod N $. If one only wants to solve doublet-triplet splitting while forbidding the $\mu$-term, then there is a solution for $N=2$ and $\sigma=1$. Forbidding all the dangerous operators above can be accomplished with a ${\bf Z_4}$ symmetry.
An essential point is that the existing bounds coming from the LEP experiments assert that the masses of charged Higgsinos are at least 100 GeV, hence an effective $\mu$-term must be generated. In our context here this implies that the ${\bf Z_N}$ symmetry must be broken, an aspect not discussed in [@Witten:2001bf]. This symmetry breaking is the subject of the next section.
Generating $\mu$
================
As discussed in above, the ${\bf Z_N}$ symmetry is a geometric symmetry of the internal $G_2$ manifold, under which the moduli are charged. The $G_2$ moduli [@Acharya:2006ia] reside in chiral supermultiplets whose complex scalar components, z\^j =t\^j + i s\^j ,are formed from the geometric moduli of the manifold[^2], $s_i$, and axionic components of the three-from $C$-field, $t_i$. We expect the moduli to break the discrete symmetry just below Planck scale when their vevs are stabilized [@Acharya:2007rc; @Acharya:2008zi] (see Section (2.2)), \~0.1 m\_p . \[modulivev\] Likewise, the moduli $F$ terms are expected to give gaugino masses in the usual way, so that m\_[1/2]{} m\_p \[moduliFvev\] . where $m_{1/2}$ is the tree level gaugino mass at the GUT scale. The axion shift symmetries $t_i \rightarrow t_i + a_i $ require that only imaginary parts of the moduli appear in perturbative interactions. The superpotential, being holomorphic in the fields, will not contain polynomial terms that explicitly depend on the moduli. The $\mu$-term can then only be generated via Kahler interactions when supersymmetry is broken via a Guidice-Massiero like mechanism [@Giudice:1988yz], i.e., from Kahler potential couplings quadratic in the Higgs fields.
To understand the size of $\mu$ (and $B\!\mu$) we we first find a combination of moduli fields (or product of moduli fields), invariant under the axion symmetries, that transform under (a complex representation of ) $\bf Z_N$ with charge $5 \sigma - N/2$ S\^1 = \^[i]{} + i \^[j]{} \[S1\] and another with charge $-5 \sigma - N/2$ S\^2 = \^[m]{} + i \^[n]{} \[S2\] . These fields have the correct charge to break the symmetry and generate the $\mu$-term which has total ${\bf Z_N} $ charge $-5 \sigma - N/2$.
In a general supergravity theory [@Wess:1992cp; @Brignole:1997dp] the fermion mass matrix is m\^\_[ij]{} = m\_[pl]{}\^3 e\^[G/2]{} ( \_i G\_j + G\_i G\_j ) \[fermmasses\] and the holormorphic components of the scalar mass matrix are m\^[ 2]{}\_[ij]{} = m\_[pl]{}\^4 e\^[G]{} ( \_i G\_j + G\^k \_i \_j G\_k ) \[scalarmasses\] where $G = m_{pl}^{-2} K + \ln (m_{pl}^{-6} |W|^2)$ and subscripts on $G$ denote derivatives with respect to the scalar fields $\phi_i$ or their conjugates $\phi^{*}_{\bar{i}}$. Respectively, (\[fermmasses\]) and (\[scalarmasses\]) can be used to find $\mu$ = m\_[3/2]{} K\_ - F\^[|[k]{}]{} K\_[|[k]{}]{} and $B\mu$
[ccl]{} B&=& 2 m\_[3/2]{}\^2 [K]{}\_ - m\_[3/2]{}F\^ [K]{}\_[|[k]{}]{} + m\_[3/2]{} F\^m K\_[m ]{}\
&-& ( m\_[3/2]{} F\^m K\^[n ]{} K\_[ m ]{} K\_[n ]{} +( ))\
&-& F\^n F\^[|[m]{}]{} ( K\_[n|[m]{}]{} - K\^[j ]{} K\_[ n ]{} K\_[j |[m]{} ]{} +( ))
where the indices run over the moduli fields and we have used that the superpotential does not contribute to either mass. Leading contributions come from Kahler potential terms K H\_u H\_d + H\_u H\_d + h.c. \[Kahlermu\] where the coefficients $\alpha,\beta$ are expected to be $\mathcal{O}(1)$. Plugging the Kahler potential ( \[Kahlermu\]) into the formulas for $\mu$ and $B\!\mu$ gives the $\mu$-term
[ccl]{} &=& m\_[3/2]{} +\
B&=& 2 m\_[3/2]{}\^2 + m\_[3/2]{} + m\_[3/2]{}. \[muterm\]
However, as a result of (\[modulivev\]), (\[moduliFvev\]) and the suppression of $m_{1/2}$ by order two orders of magnitude in the $G_2$-MSSM, $\langle S^i \rangle m_{3/2} \simeq 10 \;\;\langle F^{S^i} \rangle $, the contribution to the masses coming from $F$-terms are sub-dominant, at least if we assume that $N\ll10^4$. Therefore, to a good approximation B2 m\_[3/2]{} a fact which will have significant phenomenological consequences[^3] .
The coefficients of the operators in (\[Kahlermu\]) are in principle determined from $M$ theory, but is not yet known how to calculate them precisely. It is natural to assume that the coupling coefficients are of $\mathcal{O}(1)$. When combined with a model of moduli stabilization, such as in the $G_2$-MSSM described in [@Acharya:2006ia; @Acharya:2007rc; @Acharya:2008zi] and briefly reviewed section (2.2), $\mu$ and $B\!\mu$ can be approximately determined. Since the real and imaginary components of the complex fields, $S^1$ (\[S1\]) and $S^2$ (\[S2\]), are expected to have similar, but not necessarily identical vevs, $\mu$ will generically have a phase, that will be unrelated to the phases that enter the gaugino masses. But, $B\mu$ and $\mu$ will have the same phase since both are proportional to $S^1$ and the same coupling constant.
Before moving on to the next section we discuss the possibility that other matter fields may be charged under the ${\bf Z_N}$ symmetry, spontaneously break the ${\bf Z_N}$ symmetry, and generate $\mu$. Consider an $SU(5)$-singlet matter field $X$ that generates the $\mu$-term via the superpotential coupling $X H_u H_d$. Since $X$ is a matter field, $M$ theory requires that it is charged under least one $U(1)$ symmetry. Then $H_u H_d$ is not invariant under the $U(1)$, and consequently, the triplet mass term $T_d T_u$ is not invariant, spoiling doublet-triplet splitting. Thus, such contributions should not occur.
Alternatively, the $\mu$-term may be generated by a $U(1)$ invariant combination of two fields, for example by the operator H\_u H\_d. Requiring $\mu \gtrsim 10^3$ GeV, and taking $\Lambda \sim M_{GUT}$ this would require $\sqrt{ \langle X_1 X_2 \rangle } \gtrsim 10^{10}$ GeV. Radiative symmetry breaking will generally give a vev $\sim m_{3/2}$– usually large vevs are associated with FI $D$-terms. But since FI $D$-terms are absent in $M$ theory, it may be difficult for such large vevs to arise from here. The recent results of [@Acharya:2011kz] do suggest that the $F$-term potential can generate large matter field vevs, however in that case the vevs are too large to be relevant for the $\mu$ problem. Therefore, we very tentatively conclude that a matter field spurion is not responsible for breaking the ${\bf Z_N}$ symmetry and giving a physically relevant $\mu$-term.
Finally, we comment on a potential domain wall problem. The moduli are stabilized away from a ${\bf Z_N}$ point, which implies that the ${\bf Z_N}$ symmetry was really only an approximate symmetry of the $G_2$-manifold. The moduli stabilization serves to parameterize the amount that the $G_2$-manifold differs from a ${\bf Z_N}$ symmetric manifold. Therefore, since the ${\bf Z_N}$ symmetry is not an exact symmetry of the $G_2$ manifold, the Lagrangian will explicitly break the ${\bf Z_N}$ symmetry, and domains walls would not have formed in the early universe.
Origin of $R$-Parity in $M$ theory
==================================
In the Standard Model, the Yukawa couplings and Higgs potential form the most general set of renormalizable couplings consistent with the gauge symmetries. In this sense, baryon (B) and lepton (L) number are accidental symmetries of the theory. However, this is not the case in supersymmetric theories, which allow for the B and L violating renormalizable couplings[^4] W\_ = \^ L L e\^c + \^ L Q d\^c + \^ u\^c d\^c d\^c + L h\_u . \[WRbreaking\] If the squark masses are not of order the GUT scale (which presumably they are not), these operators can lead to too rapid proton decay if not heavily suppressed. Hence one usually introduces $R$-parity, where the Standard Model fields have $R$-parity $+1$, while their supersymmetric partners have $R$-parity $-1$. This forbids all the couplings in (\[WRbreaking\]).
Additionally, $R$-parity invariance insures the stability of the LSP, and the absence of an $R$-parity can eliminate the LSP as a dark matter candidate. Therefore, in this section we will discuss the origin of $R$-parity in $M$ theory, or at least an approximate $R$-parity that leaves the proton and LSP very long lived. Of course from a theoretical point of view an $R$-parity or equivalent symmetry should emerge from the theory and not be put in by hand.
The ${\bf Z_N}$ symmetry constructed in Section 3 contains $R$-parity, but for generic moduli charges the complete ${\bf Z_N}$ symmetry, including any $R$-parity subgroup, will be spontaneously broken. Although the ${\bf Z_N}$ symmetry will prevent the superpotential couplings in (\[WRbreaking\]) from being invariant, supersymmetry breaking will revitalize the operators just as in the case of the $\mu$-term, from Kahler potential operators K\_ L L e\^c + L Q d\^c + u\^c d\^c d\^c + L h\_u \[kahlerbad1\] where the $\tilde{S}^{ \dagger}$’s symbolically represent the moduli and need not all be the same.
Just as the $\mu$-term was generated from the Kahler potential as a result of moduli stabilization, the effective superpotential can be calculated from the supersymmetry breaking contribution from (\[kahlerbad1\]) to
[ccl]{} \_[i j k]{} && [m\_[pl]{}\^[-2]{}]{} ( m\_[3/2]{} +F\_ )( K\_ )\_[i jk]{}\
&& m\_[pl]{}\^[-1]{} ( m\_[3/2]{} +F\_)( K\_ )\_[L h\_u]{} \[kappalambda\]
for $\lambda = \lambda^{\prime}, \lambda^{\prime\prime}, \lambda^{\prime\prime\prime}$ and where $ i,j,k$ run over the matter fields. Comparing (\[kappalambda\]) to (\[muterm\]), one easily sees that $\kappa \sim \mu$, since both are generated the same way.
Then using that $\mu \sim \kappa $, the superpotential can be rewritten as W\_ (L L e\^c + L Q d\^c + u\^c d\^c d\^c )+ L h\_u. The trilinear couplings are suppressed but the lepton violating bilinear coupling is large and of order the $\mu$-term–this is simply a consequence of $\kappa$ not being suppressed. After rotating away the $L h_u$ term using the approximation (\[muyukawa\]), the superpotential simplifies to W\_ \~y\_e L L e\^c + y\_d L Q d\^c + u\^c d\^c d\^c \[WRbad1\] where smaller terms in $\lambda^{\prime},\lambda^{\prime\prime},\lambda^{\prime\prime\prime}$ have been dropped. Thus the lepton number violating trilinears pick up large contributions from the bilinear term, even if they were originally suppressed.
The proton lifetime for the decay mode $p \rightarrow e^{+} \pi^0$ is estimated to be \_[p e\^[+]{} \^0]{} . The current bounds on this partial decay width is $\tau_{p \rightarrow e^{+} \pi^0} > 1.6 \times 10^{33}$ years [@Amsler:2008zzb]. For scalar masses in the $G_2$-MSSM ($\sim 10 \operatorname{TeV}$ see [@Acharya:2008zi]) this gives the experimental bound \^ \^ 10\^[-24]{} which clearly excludes the superpotential (\[WRbad1\]), since $ \lambda^{\prime\prime } \sim y_e \sim 10^{-5}$ and $ \lambda^{\prime\prime \prime} \sim \mu / m_{pl} \sim 10^{-14}$. Therefore, proton stability requires an additional form of $R$-parity invariance beyond the discrete symmetry proposed.
One possible way to preserve the $R$-parity is to simply assume that the $G_2$-manifold in the vacuum is $R$-parity invariant, though not ${\bf Z_N}$ invariant i.e. the vacuum partially breaks ${\bf Z_N}$ to an $R$-parity subgroup. For example, take $N=6$, then
[rc|c]{} &[**Z\_6**]{}\
- & H\_d H\_u & \^4\
& M\_[10]{} M\_[|[5]{}]{} M\_[|[5]{}]{} & \^5\
& M\_[|[5]{}]{} H\_u & \^3
for $\eta \equiv e^{i 2 \pi /6}$. If all moduli transform under the ${\bf Z_3}$ subgroup of ${\bf Z_6}$, then ${\bf Z_6}$ is broken to ${\bf Z_2}$ $R$-Parity, since no $R$-parity couplings can be generated. This is technically satisfactory, but is presumably “non-generic”. It could certainly emerge from $M$ theory, but we will not consider it further here.
Alternatively, $R$-parity may manifest itself as matter-parity, a conserved remnant of a local, continuous $U(1)$ symmetry. As is well known, matter parity arises naturally in $SO(10)$ theories. When embedded into an $SO(10)$ unified theory, the Standard Model matter fields belong to a different representation than the Higgs fields– a generation of matter is contained in a $\bf{16}$ of $SO(10)$, while a pair of Higgs doublets comes from a $\bf{10}$ of $SO(10)$.
When $SO(10)$ is broken to $SU(5)\times U(1)_{\chi}$, for example by a discrete Wilson line, the Higgs fields and matter fields are charged differently under $U(1)_{\chi}$:
[ccl]{} SO(10) & & SU(5)U(1)\_\
**[16]{} & & **[10]{}\_[-1]{} **[|[5]{}]{}\_[3]{} **[1]{}\_[-5]{}\
**[10]{} & & **[5]{}\_[2]{} **[|[5]{}]{}\_[-2]{}.**************
where the subscript is the $U(1)_{\chi}$ charge.
The vacuum expectation values of the Higgses, which are contained in the $\bf{5}_{2}$ and $\bf{\bar{5}}_{-2}$ multiplets, will break the $U(1)_{\chi}$ symmetry into a discrete ${\bf Z_2}$ subgroup. This is because the Lagrangian is no longer invariant under the full local transformation $\Phi \rightarrow e^{i \alpha(x) q_d } \Phi$, but only the subgroup of transformations given by $\alpha(x)=\pi$. In terms of the $U(1)_{\chi}$ charges $q_{\chi}$, the chiral multiplets have ${\bf Z_2}$-parity $e^{i \pi q_{\chi}}$. Thus chiral superfields with even $U(1)_{\chi}$ charge will have parity $+1$ and fields with odd $U(1)_{\chi}$ charge will have parity $-1$. The ${\bf Z_2}$ symmetry is exactly $R$-parity.
The only $SU(5)$ singlet with $U(1)_{\chi}$ charge is the $\bf{1}_{-5}$ field (and its conjugate), and thus this is the only field that can break $U(1)_{\chi}$ without breaking the SM gauge group. But since it has odd $U(1)_{\chi}$ charge, its vev will break $R$-parity. Therefore an $SO(10)$ completion of $U(1)_{\chi}$ will not contain an unbroken $R$-parity, but perhaps when combined with the ${\bf Z_N}$ symmetry, $R$-parity violating operators may be sufficiently suppressed to allow a long lived proton and LSP. Next we estimate these lifetimes.
The singlet field $\bf{1}_{-5}$ can be considered to be the right-handed neutrino, $\nu^c$, since it has the right quantum numbers to make the operator $ \nu^c h_u L$ invariant under $U(1)_{\chi}$. However, if $\langle \nu^c \rangle \ne 0 $, all baryon and lepton violating operators in (\[WRbreaking\]) will be generated via the superpotential W\_ \~\^c L L e\^c + \^c L Q d\^c + \^c u\^c d\^c d\^c + \^c h\_u L . \[WRbad2\] The operators in (\[WRbad2\]) should be suppressed and can be forbidden by the ${\bf Z_N}$ symmetry. The story will be the same as above and the Kahler potential operators will generate (\[WRbad2\]) , but with additional suppression coming from $U(1)_{\chi}$ breaking W\_ = ()( ) (L L e\^c + L Q d\^c + u\^c d\^c d\^c )+ () ( ) L h\_u. Diagonalizing away the $L h_u $ term, and using (\[muterm\]) gives W\_ \~y\_e L L e\^c + y\_d L Q d\^c + u\^c d\^c d\^c. \[WRbadc\] where again large lepton violating trilinear terms are induced by the rotation.
![Decays of the LSP. Only the lepton number violating diagrams are shown, since the lepton number violating couplings– $\lambda^{\prime}$ (in the first line) and $\lambda^{\prime\prime}$ (in the second line)– receive large contributions (compared to the baryon number violating couplings) when the bilinear $R$-parity violating term, $h_u L$, is rotated away. Primes on the $L$ indicate that the lepton flavor is different than the slepton flavor. Figures from [@Martin:1997ns]. \[fig:rparityviolation\]](LSP.png){width="100.00000%"}
To be conservative in our estimates, we can take $\langle \nu^c \rangle \sim$ TeV, which may be expected from radiative symmetry breaking [@Ambroso:2010pe]. In this limit, proton decay constraints are safe from $R$-parity violation, but there are more stringent constraints coming from the LSP lifetime. Current bounds on the LSP lifetime are slightly model dependent, but for the most part are [@Dreiner:1997uz] \_[LSP]{} 1 \_[LSP]{} 10\^[25]{} . The first bound excludes the region where the LSP decays would ruin the successful predictions of big bang nucleosynthesis on light nuclei abundances [@Reno:1987qw; @Ellis:1990nb]. The other region is excluded by indirect dark matter detection experiments that search for energetic positrons and anti-protons coming from decaying or annihilating relics [@Berezinsky:1991sp; @Baltz:1997ar; @Arvanitaki:2009yb; @Shirai:2009fq].
The LSP lifetime can be calculated in terms of the general $R$-parity violating superpotential couplings (\[WRbreaking\]). Diagrams in Figure 1 lead to an LSP lifetime ( )\^4 ( )\^5 where $\lambda = \lambda^{\prime} ,\lambda^{\prime\prime}, \lambda^{\prime\prime\prime}$ and $m_0$ is the mass of the sfermion mediating the decay. Taking $\lambda = \frac{\langle \nu^c \rangle}{m_p} \sim 10^{-15}$, $m_0 \sim 10$ TeV, and $m_{LSP} \sim 100 $ GeV gives \_ \~10\^[17]{} , about the age of the universe. The $R$-parity violating couplings still need to be about $10^{-4}\sim 10^{-5} $ smaller to have an LSP lifetime greater than $10^{25}$ seconds.
There are several ways additional suppressions might arise. We have not yet discussed the possibility of there being a horizontal family structure to the couplings. This could appear as a Froggett-Nielson symmetry, or a symmetry relating the locations of the matter singularities on the $G_2$ manifold, and would be responsible for forging the quark and lepton hierarchy. It may also suppress the LSP decay width pass the astrophysical bounds. Family symmetries arise naturally from the $E_8$ structure [@King:2010mq], which can also explain why the Standard Model has three generation, and this may hint towards a larger gauge theory. We leave this issue to future work..
If the family symmetry is not the answer, then it may be the case that resolution of the $E_8$ singularity to $SU(5)$ preserves a $U(1)$ symmetry–whose charges are necessarily given as a linear combination of four $U(1)$s belonging to the coset group $E_8/SU(5)$–and is broken to an exactly conserved $R$-parity. There are two well known examples, $U(1)_\chi$ and $U(1)_\psi$ , defined as the symmetries coming from the breaking $SO(10)\rightarrow SU(5) \times U(1)_\chi$ and $E_6\rightarrow SO(10) \times U(1)_\psi$. However, $U(1)_\chi$ does not contain a field that can break $U(1)_\chi$ to $R$-Parity, and $U(1)_\psi$ forbids Higgs triplet masses, spoiling doublet-triplet splitting, so neither of these choices give a conserved $R$-parity.
However, there is a possibility that $U(1)$ symmetry is similar to $U(1)_\chi$, in that the MSSM fields and right handed handed neutrinos have the same charge assignment as in $U(1)_\chi$, but has additional $SU(5)$ singlet fields with even charges[^5]. These theories can then be broken to a [*conserved*]{} $R$-parity, when the additional singlets get vevs. It is easy to construct such a linear combination, though it is unclear why from a purely theoretical perspective why $G_2$ compactifications would favor this $U(1)$ symmetry. For instance, if $U(1)_a \times U(1)_b$ is the cartan subgroup of $SU(3)$ in the breaking pattern $E_8\rightarrow E_6\times SU(3)$, then the $U(1)$ given by the linear combination of charges $$q_\chi + 5(q_a-q_b)$$ allows for conical singularities that give rise to MSSM and right handed neutrino fields with $U(1)_\chi$ charges, but with additional $SU(5)$ singlets with charges $q=\pm 10$. The vevs of the additional singlets will break the $U(1)$ to a $Z_{10}$ symmetry that contains a $Z_2$ $R$-parity.
Finally we note (for the non string duality oriented reader) that $E_8 \times E_8$ is well motivated theoretically if the $G_2$-manifold is a $K3$ fibration. This is because the intersection matrix of 2-cycles inside $K3$ contain the Cartan matrix of $E_8 \times E_8$. It is in this case–that the gauge-theory of $M$ theory matches the gauge theory of $E_8 \times E_8$ Heterotic string theory– in which M theory on a $K3$-fibered $G2$-manifold and the heterotic string theory on a $T^3$- fibered Calabi-Yau threefold are dual.
To summarize, we find that incorporating the $\mu$ parameter into the structure of $M$ theory compactified on a $G_2$-manifold, with stabilized moduli, can lead to a broken discrete symmetry allowing $\mu$ to be non-zero. $R$-parity is slightly broken, giving an LSP lifetime long enough to be the dark matter, but not quite long enough to evade satellite detector constraints. The theoretical structure allows for family symmetries, or an embedding of $R$-parity into $E_8$, both of which stabilize the LSP lifetime to be consistent with the experimental constraints. An example of the latter case is given above, so this is indeed a possibility. Either case will lead to the same dark matter phenomenology. The $R$-parity completion of this story is an interesting avenue for further investigation.
Phenomenology
=============
The $M$ theory framework, along with moduli stabilization in the $G_2$-MSSM, allows one to estimate the high-scale SUSY breaking masses and $\mu$ to within a factor of a few. This allows $M$ theory to make many phenomenological predictions. For some cases even small variations in the high-scale theory can have significant phenomenological consequences. In particular, the low-scale values of $\mu $ and $\tan\!\beta$ have significant implications for dark matter properties, and thus it is crucial to have a good understanding of their low-scale values while considering the $M$ theory predictions of the high-scale masses.
Electroweak Symmetry Breaking
-----------------------------
The first and foremost phenomenological constraint is that the theory accurately produce electroweak symmetry breaking (EWSB). That is, the theory must give a stable potential (bounded from below), break the electroweak symmetry and allow for the correct Z-boson mass. Respectively, these three conditions can be quantified by the following tree level constraints at the EWSB scale
[ccc]{} | B| & & ( m\_[H\_u]{}\^2 + m\_[H\_d]{}\^2 ) + ||\^2\
| B|\^2 & & ( m\_[H\_u]{}\^2 + ||\^2)( m\_[H\_d]{}\^2 + ||\^2)\
M\_Z\^2 & = & -2 ||\^2 + 2 \[EWSBconstraints\]
where $\tan\!\beta$ is not an independent parameter, but is determined by 2= . \[bmutanbeta\] and m\_A\^2 = m\_[H\_u]{}\^2 + m\_[H\_d]{}\^2 +2 ||\^2 \[mA\]. where $A$ is the pseudoscalar Higgs boson.
To get a feeling for $\tan\!\beta$, we plug in the expected values (at the unification scale and with degenerate scalars) of $B\!\mu \simeq 0.2 m_{3/2}^2$ and $m_A^2 \simeq 2 m_{3/2}^2$, into ($\ref{bmutanbeta}$) which gives $\tan\!\beta \simeq 10$. On the other hand, RGE flow will lower the values of both $B\!\mu $ and $m_A^2$, resulting in variations around $\tan\!\beta \simeq 10$. In Section $7.1.1$, a numerical scan will show a lower bound of $\tan\!\beta \gtrsim 5$, when scalars are taken to be degenerate at the unification scale.
The lowest values of $\tan\!\beta$ occur for the smallest values of $m_A^2$. The EW scale value for the mass depends on the running of the Higgs scalar masses, and in turn is very sensitive to the values of the squark masses. For specific non-degenerate values of scalar masses at the unification scale, $m_A^2$ can be of order $B\!\mu$ at the EW scale, resulting in values of $\tan\!\beta < 5$. We will consider this situation in Section $7.1.2$.
At tree level the mass of the Z-boson is determined by the four Higgs parameters $$M_Z (m_{H_u}^2, \; m_{H_d}^2, \; |\mu |^2, \; \tan\!\beta).$$ These parameters not only depend on their respective values at the high-scale, but also on other masses as a result of RGE-flow. Assuming that the scalar masses are much larger than the gaugino masses, $M_Z$ has strongest dependence on the Higgs mass parameters and stop masses M\_Z( \_[H\_u]{}\^2, \_[H\_d]{}\^2, , | |\^2, \_[Q\_3]{}\^2, \_[U\_3]{}\^2 ) .where hatted ($~\hat{}~$) masses refer to GUT scale values.
Interestingly the cancellation between the soft scalars masses contributing to $M_Z$ can be significant, even in the case in which the scalar masses are unified at the GUT scale $$\hat{m}_{H_u}^2 = \hat{m}_{H_d}^2 = \hat{m}_{Q_3}^2 = \hat{m}_{U_3}^2.$$ Naively what one thought was a large fine-tuning between the Higgs soft-masses and $\mu$ in eq. (\[EWSBconstraints\]) for $M_Z$, is in fact smaller. This is evident (see Figure \[fig:unifiedmutanbeta\]) from the fact that the scalar masses can be of order the gravitno mass at unification and $\mu$ can be an order of magnitude smaller, but the cancellation in eq. (\[EWSBconstraints\]) for $M_Z$ still occurs. In this sense, the ratio $\mu / m_{3/2}$, shown in Figure \[fig:unifiedmutanbeta\], might be considered a measure of the fine-tuning involved in EWSB. In other words, the smaller the ratio, the less fine tuning there will be of $\mu$ against the scalar masses in order to have the correct value for $M_Z$.
### Degenerate Scalars
A numerical scan was performed over $M$ theory parameter space described in [@Acharya:2008zi] using SOFTSUSY [@Allanach:2001kg][^6]. We allow for the following variation in the $G_2$-MSSM parameters,
- $10 \operatorname{TeV}\le m_{3/2} \le 20 \operatorname{TeV}$ – the gravitino mass
- $10\le V_7 \le 40$ – the volume of the $G_2$-manifold in units of the eleven-dimensional Planck length
- $-10 \le \delta \le 0$ – the size of the threshold corrections to the (unified) gauge coupling, $\alpha^{-1}_{\text{GUT}}$. [^7]
An interested reader is referred to Section V of [@Acharya:2008zi] for variations in the spectra of $G_2$-MSSM models. In addition, order one variations are allowed for the coefficients in (\[muterm\]) for the formula for $\mu$, while it is imposed that $B\!\mu$ is in the range. 1m\_[3/2]{} < B< 3m\_[3/2]{}. \[bmubound\]
The results are shown in Figure \[fig:unifiedmutanbeta\]. As is evident from the plot, values of $\mu$ much smaller than the gravitino mass are allowed under all the constraints, signaling a non-imposed cancellation among the scalars contributing to $M_Z$. Of note is the fact that $\tan\!\beta$ and $\mu$ are inversely correlated, which will play a significant role in limiting the maximum spin-independent scattering cross-section, when scalar masses are unified at the high scale.
![$\mu/m_{3/2}$ vs. $\tan\!\beta$. The upper band scans over the $G_2$-MSSM parameter space with [*degenerate scalars*]{} at the unification scale. The lower region on the left (low $\tan\!\beta$) scans over the $G_2$-MSSM parameter space where the scalar mass ratio $\hat{m}^2_{H_u}\!:\! \hat{m}^2_{U_3} \!:\! \hat{m}^2_{Q_3} = 3\!:\! 2\!:\! 1$ is required to be accurate within $20\%$. The black points show models that correctly break the EW symmetry, but are inconsistent with constraint $1\,\mu m_{3/2} < B\!\mu < 3\,\mu m_{3/2}$, so we expect them to not be valid solutions. The red points satisfy the constraint on $B\!\mu$ as given in the legend. The empty space on the plot, between the two regions, is expected to be filled in with complete scan over possible non-degenerate scalar mass parameter space. All points have EWSB.[]{data-label="fig:unifiedmutanbeta"}](mutanb2.png){width="100.00000%"}
### Non-Degenerate Scalars and Low
We also consider the possibility that $M$ theory allows for scalar unification to be somewhat perturbed (at the factor of two to three level). Since eventually we will be interested in calculating the largest possible spin-independent scattering cross sections we will only consider high-scale scalar masses that give $\tan\!\beta\lesssim 3 $–since the scattering cross sections decrease with increasing $\tan\!\beta$
Consider the 1-Loop RGE equations, where only terms proportional to $\lambda_t$ are kept, while neglecting the $\lambda_t$ running. The RGE equations of the relevant scalars are:
[ccl]{} [8\^2]{} &= &3 | \_t |\^2 ( m\_[H\_u]{}\^2 + m\_[Q\_3]{}\^2 + m\_[U\_3]{}\^2 + | A\_t |\^2)\
[8\^2]{} &= &2 | \_t |\^2 ( m\_[H\_2]{}\^2 + m\_[Q\_3]{}\^2 + m\_[U\_3]{}\^2 + | A\_t |\^2)\
[8\^2]{} &=& 1 | \_t |\^2 ( m\_[H\_2]{}\^2 + m\_[Q\_3]{}\^2 + m\_[U\_3]{}\^2 + | A\_t |\^2)\
[8\^2]{} &=& 6\_t\^2 A\_t \[eqn:rge\]
whose solution is
[ccl]{} m\^2\_[H\_u]{} &=& ( \^2\_[H\_u]{} -\^2\_[U\_3]{} - \^2\_[Q\_3]{} + e\^( |\_t|\^2 (-1+ e\^) + \^2\_[H\_u]{} + \^2\_[U\_3]{}+ \^2\_[Q\_3]{}) )\
m\^2\_[U\_3]{} &=& (- \^2\_[H\_u]{}+ 2\^2\_[U\_3]{} - \^2\_[Q\_3]{} + e\^(|\_t|\^2 (-1+ e\^) + \^2\_[H\_u]{} + \^2\_[U\_3]{}+ \^2\_[Q\_3]{}) )\
m\^2\_[Q\_3]{} &=& (- \^2\_[H\_u]{} - \^2\_[U\_3]{} + 5 \^2\_[Q\_3]{} + e\^(|\_t|\^2 (-1+ e\^) + \^2\_[H\_u]{} + \^2\_[U\_3]{}+ \^2\_[Q\_3]{} ) )\
A\_t\^2 &=& \_t\^2 e\^\
\[rgesolutions\]
where hatted ($\hat{}$) masses indicate GUT scale mass.
Since $m^2_{H_d}$ barely runs for low $\tan\!\beta$ and it is predicted that $\hat{\mu}^2$ is over an order of magnitude smaller than $m^2_{H_d}$ , the cancellation in $M_Z$ (\[EWSBconstraints\]) should occur between $m^2_{H_u}$ and $m^2_{H_d}$. Therefore, $m^2_{H_u}$ needs to stay positive at the EWSB scale. Ignoring the exponentially suppressed terms in (\[rgesolutions\]), we see that there are no choices of $\{ \hat{m}^2_{H_u}, \hat{m}^2_{Q_3},\hat{m}^2_{U_3} \} $ that leave all low-scale masses positive. On the other hand, there is a fixed point solution to the above RGEs \^2\_[H\_u]{}: \^2\_[U\_3]{} : \^2\_[Q\_3]{} = 3 : 2: 1 \[fixedpoint\] where the non-exponentially suppressed terms are identically zero, insuring that if the trilinears are of order the scalars as expected in the $G_2$-MSSM, all three masses will stay positive. This fixed point is analogous to the focus point solution in minimal supergravity (mSUGRA) theories [@Feng:1999zg; @Chan:1997bi], as it minimizes the fine tuning of EWSB. However, unlike the focus point region of mSUGRA where the Higgs scalars run small due to RGE flow, here the scalars remain heavy, and are close to the gravitino mass.
Near this region, low $\tan\!\beta$ parameter space with EWSB can be realized. Results on the numerical scan can be seen in Figure \[fig:unifiedmutanbeta\].
The Nature of the LSP
---------------------
As explained in detail in[@Acharya:2008zi], the $G_2$-MSSM framework gives rise to mostly Wino LSPs (as opposed to Bino LSPs). The tree level gaugino masses are degenerate at the GUT scale, but are suppressed by $F$-terms of the moduli relative to the gravitino mass to be of order the gaugino masses from the anomaly mediation contribution. The additional contribution from the anomaly lifts $M_1$ over $M_2$, leading to mostly Wino LSP models. In the original $G_2$-MSSM scenario, where it was simply that $\mu \sim m_{3/2}$, there were additional contributions to the gaugino masses from supersymmetric Higgs loops, proportional to $\mu$ [@Pierce:1996zz], that for some choices of high scale parameters, could re-lift $M_2$ over $M_1$. These models are disfavored by precision gauge coupling unification [@Acharya:2008zi], and occur less frequently in parameter space here than the original models since $\mu \not{\!\!\!\sim}\, m_{3/2}$. However, smaller $\mu$ will tend to introduce a small Higgsino admixture into the mostly Wino LSP - a fact which has significant implications on dark matter discovery (Section 6.3). All these considerations combine to strongly suggest that a Wino-like LSP with mass $\sim 140 - 200$ GeV constitutes a significant fraction of the dark matter.
As emphasized in [@Acharya:2008bk; @Acharya:2009zt], in order to obtain about the right relic density from the moduli decays, the LSP must be a Wino-like particle, with a large annihilation cross section of about $3 \times 10^{−24} \operatorname{cm}^2$. A non-thermal history dominated by moduli and a wino LSP give a consistent picture for dark matter from the compactified string theory. Also encouraging is the fact that the PAMELA satellite data on positrons and antiprotons can be consistently described by a Wino LSP [@Grajek:2008jb; @Hisano:2008ti; @Kane:2009if; @Feldman:2009wv; @Chen:2010yi; @Chen:2010kq]. More recently, by also considering Wino annihilations into photons and Z-bosons one finds a cross-section of about $10^{-26} \operatorname{cm}^2$ – a fact relevant for future Fermi data.
Direct Detection of Dark Matter
-------------------------------
In December 2009, CDMS reported at most two possible WIMP candidate events, with a high likelihood of being background [@Ahmed:2009zw]. Combining with their previous data, this amounts to a bound on the spin-independent scattering cross-section of $\sigma_{si} \lesssim 6 \times 10^{-44} \text{ cm}^2 $ for a WIMP of mass around $200$ GeV. More recently, the XEXON100 experiment [@Aprile:2010um] reported observing no events after their first 11 days of running, slightly strengthening the CDMS bound. In the near future, XEXON100 is expected to report results that will probe much smaller scattering cross sections $\sigma_{SI} \sim 2\times 10^{-45} \text{cm}^2 $. We will see that even this region is out of reach given the $M$ theory predictions we calculate.
In the decoupling limit, defined when the pseudoscalar mass is much larger that the Z-boson mass, $m_{A^0} \gg M_{Z}$, the charged and heavy CP-even Higgses are also heavy, $m_{H^{\pm}} \simeq m_{H^{0}} \simeq m_{A^{0}} $. The other Higgs boson $h^0$ remains light and behaves in the same way as the SM Higgs boson. The lower bound on its mass, corresponds to the same bound on the SM Higgs boson, namely 114 GeV[^8][@Barate:2003sz]. All the models consistent with all the theoretical and phenomenological constraints have light Higgs mass close to thia LEP limit. Since the squarks are also heavy in $G_2$-MSSM, the light Higgs boson exchange will give the only substantial contribution to the spin-independent scattering cross sections. The scattering of the LSP off nuclei is via the Higgsino component. While the LSP will be mostly Wino-like, the prediction that $\mu$ is of order the TeV scale implies that the LSP wavefunction can have non-trivial Higgsino mixing.
Following [@Cohen:2010gj] we estimate the size of the direct detection cross section in the decoupling limit to be $$\sigma_{\rm SI} \left( \chi N \rightarrow \chi N \right) \approx 5 \times 10^{-45} \text{cm}^2 \left( \frac{115\operatorname{GeV}}{m_h} \right)^4
\left(\frac{Z_{H_u}\sin\beta - Z_{H_d}\cos\beta}{0.1}\right)^2 \left( Z_W - \tan\theta_W Z_B \right)^2
\label{eqn:tim}$$ where the $Z$’s give the composition of the LSP Z\_B+Z\_W+Z\_[H\_d]{}\_d+Z\_[H\_u]{}\_u.
This gives us an estimate of the largest direct detection scattering cross sections, which naively may seem that for $Z_{H_u} \sim 0.1$ can be very close to the reach of XENON. Eq. (\[eqn:tim\]) can further be simplified, with the aid of analytical expressions for the neutralino mass matrix eigenvalues and eigenvectors [@ElKheishen:1992yv; @Barger:1993gh; @Bertone:2004pz]. Taking the limit $ M_1 = M_2 $, which maximizes the scattering cross section for fixed $\mu$ and $\tan\!\beta$, (\[eqn:tim\]) becomes $$\sigma^{\rm MSSM}_{\rm SI} \left( \chi N \rightarrow \chi N \right) \approx 6 \times 10^{-45} \text{cm}^2 \left( \frac{115 \operatorname{GeV}}{m_h} \right)^4
\left(\frac{1 \operatorname{TeV}}{\mu}\right)^2 \left( \frac{\sin2\beta+M_{2}/\mu}{1 - (M_2/\mu)^2} \right)^2
\label{eqn:upperlimit}$$ which falls off both with $\tan\!\beta$ and $\mu$. Allowing for the variation in $M_1$ and $M_2$ in the $G_2$-MSSM will only decrease this fraction. The value $M_2/\mu$ is typically around $.1\sim.2$. The parameters for three different models, along with their scattering cross sections, can be seen in Table 1 and are appropriately labeled in Figure 3.
However, as shown in the previous section, when considering degenerate scalar masses at the unification scale EWSB imposes that small $\mu$ corresponds to large $\tan\!\beta$, and small $\tan\!\beta$ corresponds to large $\mu$. Hence, large cross-sections, of order the XEXON100 reach are not attainable for this region. To verify this we perform a scan of parameter space, using DarkSUSY [@Gondolo:2004sc]. The results are show in Figure \[fig:xsecuniversal\] where it is seen that the largest scattering cross-sections are $\sim 1 \times 10^{-45} \text{cm}^2$, close to, but slightly beyond the reach of XENON100.
In Figure \[fig:xsecuniversal\] we also scan over the $G_2$-MSSM parameter space, while requiring that the ratio $ \hat{m}^2_{H_u}: \hat{m}^2_{U_3} : \hat{m}^2_{Q_3} = 3 : 2: 1$ be accurate within $20\%$. The spin-independent scattering cross-section reaches an upper limit of $1 \times 10^{-45}\operatorname{cm}^2$, just beyond the XENON100 reach. Since this is the region where largest cross-sections appear, we can conclude that if the solution of the $\mu$-problem proposed, along with moduli-stabilization in the $G_2$-MSSM, is the model of nature, the XENON100 experiment will not observe a dark matter signal soon, but its next run and upgraded detectors may do so.
![Spin-independent scattering cross-sections vs $\tan\!\beta$. The region shown scans over the $G_2$-MSSM parameter space where the scalar mass ratio $\hat{m}^2_{H_u}\!:\! \hat{m}^2_{U_3} \!:\! \hat{m}^2_{Q_3} = 3\!:\! 2\!:\! 1$ is required to be accurate within $20\%$. All points satisfy the constraint $\mu m_{3/2} < B\!\mu < 3\,\mu m_{3/2}$, have a SM-like Higgs with mass $m_h \ge 110$ GeV, and have EWSB. We also list the parameters for the 3 models in Table 1. In the region where EWSB, supergravity, and phenomenological constraints are satisfied, the upper limit on $\sigma_{SI}$ is robust, but the lower limit can decrease if the sign of $\mu$ is reversed.[]{data-label="fig:xsecuniversal"}](mu_new.pdf){width="100.00000%"}
[2]{}
Model 1 Model 2 Model 3
-------------------------------------- --------------------- ---------------------- ----------------------
$M_{\text{3/2}}$ 17.8 TeV 18.1 TeV 17.9 TeV
$\sqrt{B\mu_{\text{GUT}}}$ 9.75 TeV 10.4 TeV 9.29 TeV
$ \mu_{\text{GUT}}$ 3.79 TeV 2.10 TeV 1.69 TeV
$ M_1 $ 151\. GeV 153\. GeV 150\. GeV
$ M_2 $ 145\. GeV 143\. GeV 138\. GeV
$ \mu$ 3.89 TeV 2.15 TeV 1.77 TeV
$ M_A $ 18.8 Tev 19.0 TeV 18.2 TeV
$ m_h $ 110\. GeV 110\. GeV 115\. GeV
$ M_{\chi_1 } $ 141 GeV 143 GeV 141 GeV
$ M_{\chi_2 } $ 143 GeV 147 GeV 145 GeV
$ M_{\chi_1^\pm } $ 141 GeV 144 GeV 142 GeV
$ Z_{\tilde{W}}$ 0.94 0.91 0.91
$ Z_{\tilde{B}} $ -0.35 -0.41 -0.41
$ Z_{\tilde{H}_d} $ -0.02 -0.04 -0.05
$ Z_{\tilde{H}_u} $ 0.01 0.02 0.02
$ \tan\beta $ 2.53 2.37 2.87
$ \sigma _{\text{SI}} [\text{cm}^2]$ $3.\times 10^{-46}$ $ 9.\times 10^{-46}$ $ 1.\times 10^{-45}$
$ \sigma _{\text{SD}} [\text{cm}^2]$ $5.\times 10^{-45}$ $ 5.\times 10^{-44}$ $ 1.\times 10^{-43}$
: High scale and low scale parameters for 3 models with larger spin independent scattering cross sections. All models shown belong to the parameter space where the scalar mass ratio $\hat{m}^2_{H_u}\!:\! \hat{m}^2_{U_3} \!:\! \hat{m}^2_{Q_3} = 3\!:\! 2\!:\! 1$ is accurate within $20\%$. We assume that details of the calculation and software outputs are sufficiently uncertain to allow $m_h \gtrsim 110$ GeV to be consistent with LEP bounds.
\
\
Conclusions
===========
We have argued that if our universe is described by $M$ theory compactified on a manifold of $G_2$ holonomy, with doublet-triplet splitting solved in the way originally proposed by Witten [@Witten:2001bf], then there is a simple solution to the $\mu$-problem: strong coupling dynamics in the the hidden sector will generate a non-perturbative potential for the moduli, which stabilizes all the moduli vevs, and breaks the symmetry forbidding $\mu$. Then, following the numerical analysis done in the $G_2$-MSSM [@Acharya:2008zi], the breaking will generate $\mu \sim \langle \frac{S}{m_{pl}} \rangle m_{3/2} \sim 0.1 ~m_{3/2} \sim 2 \operatorname{TeV}$. This then implies a non-zero Higgsino component of the mostly Wino LSP, with an upper limit, which in turn gives an upper limit of about $1 \times 10^{-45}\operatorname{cm}^2$ on the spin-independant scattering cross-section, somewhat below the reach of the XENON100 experiment, as well as a lower limit of about $10^{-46} \operatorname{cm}^2$. The Wino-like LSP also can account for the PAMELA positron and antiproton excesses [@Adriani:2008zr; @Kane:2009if], and gives about the desired relic density for a non-thermal cosmological history [@Acharya:2010af], as expected in theories with moduli.
Since the scalars are of order $m_{3/2} \gtrsim 20 $ TeV, the Higgs sector is decoupled, and the light Higgs boson behaves like a Standard Model one. It’s mass is predicted to be of order 110-120 GeV. If we insist on a good description of the Pamela data plus consistent compactification, we find an LSP mass from about 140-155 GeV, and an annihilation cross section $2-3.5 \times 10^{-24} \text{cm}^3/\text{s}$. The annihilation to $\gamma/Z$ ranges from $(0.7-1.2) \times 10^{-26}$.
Additionally, we noted that an eact $R$-parity could arise through ‘partial symmetry breaking’, though this isn’t obviously motivated by the theory itself. An alternative is that $R$-parity is either an exact remnant of a broken continuous gauge symmetry, or only an approximate symmetry of larger broken discrete and continuous groups. In either case, this requires the inclusion of additional $U(1)$ gauge symmetries, suggesting that the GUT group is larger than $SU(5)$, and may originate from an $E_8$ singularity.
Acknowledgments {#acknowledgments .unnumbered}
===============
We appreciate helpful conversations with Tim Cohen, Daniel Feldman, Piyush Kumar, Paul Langacker, Joseph Marsano, Aaron Pierce, and Lian-Tao Wang. B.A.. is grateful to the University of Michigan Physics Department and MCTP for support, and E.K. is grateful for a String Vacuum Project Graduate Fellowship funded through NSF grant PHY/0917807. This work was supported by the DOE Grant \#DE-FG02-95ER40899.
Appendix A: Largest Spin Independent Cross Sections {#appendix-a-largest-spin-independent-cross-sections .unnumbered}
===================================================
Following [@Cohen:2010gj], the spin-independent cross section for the LSP scattering off a nucleon, is given in the decoupling limit ($M_Z \ll M_A$) by the approximation $$\sigma_{\rm SI} \left( \chi N \rightarrow \chi N \right) \approx 5 \times 10^{-45} \text{cm}^2 \left( \frac{115\operatorname{GeV}}{m_h} \right)^4
\left(\frac{Z_{H_u}\sin\beta - Z_{H_d}\cos\beta}{0.1}\right)^2 \left( Z_W - \tan\theta_W Z_B \right)^2
\label{eqn:tim}$$ where the $Z$’s give the composition of the LSP $$\chi \equiv Z_B\,\tilde{B}+Z_W\,\tilde{W}\,+Z_{H_d}\,\tilde{H}_d+Z_{H_u}\,\tilde{H}_u.$$
Consider the neutralino mass matrix [@Frere:1983dd]: $$\mathcal{M} =
\begin{pmatrix}
M_1 & 0 & -M_Z \cos\beta\sin\theta_W & M_Z\sin\beta\sin\theta_W \\
0 & M_2 & M_Z \cos\beta\cos\theta_W & -M_Z\sin\beta\cos\theta_W \\
-M_Z\cos\beta\sin\theta_W & M_Z\cos\beta\cos\theta_W & 0 & -\mu \\
M_Z\sin\beta\sin\theta_W & -M_Z\sin\beta\cos\theta_W & -\mu & 0
\end{pmatrix}
\label{eqn:massmatrix}$$ in the $\{\tilde{B},\tilde{W},\tilde{H}_d,\tilde{H}_u\}$ basis.The analytical expression [@ElKheishen:1992yv; @Barger:1993gh; @Bertone:2004pz] for the components in the LSP can be written as: $$\begin{aligned}
\alpha Z_{B} &= z_B = - \sin\theta_W \nonumber\\
\alpha Z_{W} &= z_W = \cos\theta_W \frac{M_1-M_\chi}{M_2-M_\chi} = \cos\theta_W \frac{\left( M_1- M_\chi \right)^2}{\Delta} \nonumber \\
\alpha Z_{H_d} &= z_ {H_d} = \frac{\mu\left( M_1 - M_\chi \right)\left( M_2 - M_\chi \right) + M_Z^2\sin\beta\cos\beta\left( \left( M_1 - M_2 \right)\cos^2\theta_W +M_2 - M_\chi \right)}{M_Z\left( M_2 - M_\chi \right)\left( -\mu\cos\beta + M_\chi\sin\beta \right)} \nonumber\\
\alpha Z_{H_u} &= z_{H_u} = \frac{M_\chi \left( M_1 - M_\chi \right)\left( M_2 - M_\chi \right) + M_Z^2\cos^2\beta\left( \left( M_1 - M_2 \right)\cos^2\theta_W +M_2 - M_\chi \right)}{M_Z\left( M_2 - M_\chi \right)\left( -\mu\cos\beta + M_\chi\sin\beta \right)}
\label{eqn:components}\end{aligned}$$ where $ \alpha = \sqrt{ z_{B}^2+z_{W}^2+z_{H_d}^2+ z_{H_u}^2}$ is an overall normalization factor and $\Delta\equiv(M_\chi -M_1)(M_\chi-M_2)$.
The combination $Z_{H_u}\sin\beta-Z_{H_d}\cos\beta$, that appears in the scattering cross section takes an especially simple form $$Z_{H_u}\sin\beta-Z_{H_d}\cos\beta = \frac{\left( M_\chi\sin\beta - \mu\cos\beta \right)\left( M_1-M_\chi \right)\left( M_2-M_\chi \right)}{M_Z\left( M_2 - M_\chi \right)\left( -\mu\cos\beta+M_\chi\sin\beta \right)} = \frac{M_1-M_\chi}{M_Z}.
\label{eqn:components2}$$ It is clear from (\[eqn:components2\]) that as $M_1 - M_\chi$ increases, $Z_{H_u}\sin\beta-Z_{H_d}\cos\beta$ grows slower than the $Z_{W}$ component. Thus after normalization both the $\tilde{H}_u$ and the $\tilde{H}_d$ components will decrease. So the maximum of $Z_{H_u}\sin\beta-Z_{H_d}\cos\beta$ is realized when $M_1 - M_\chi$ is minimal.
The eigenvalues of the neutralino mass matrix (\[eqn:massmatrix\]) are given by the solutions to: $$\left( x - M_1 \right)\left( x - M_2 \right)\left( x - \mu \right)\left( x + \mu \right)+\left(M_1\cos^2\theta_W +M_2\sin^2\theta\right)M_Z^2\mu\sin2\beta = 0
\label{eqn:eigenvalue}$$ Then the LSP mass, corresponding to $M_{\chi} \equiv x$, can be found by taking the limit $M_\chi \ll \mu$, so that (\[eqn:eigenvalue\]) is simply a quadratic equation. Then it is easy to see that the minimal value of $M_1 - M_\chi$, which maximizes $Z_{H_u}\sin\beta-Z_{H_d}\cos\beta$, corresponds to the situation when $M_1 - M_2$ is also minimized. Additionally, when $M_1 = M_2$, the term $Z_W - \tan\theta_W Z_B$ also reaches its maximum. Thus the maximum scattering cross sections will occur when $M_1 = M_2$.
To normalize the expressions in (\[eqn:components\]) (i.e. finding $\alpha$) is tedious. Instead, a new basis is defined where $\tilde{\gamma} = \cos\theta_W \tilde{B} + \sin\theta_W \tilde{W}$ and $\tilde{Z} = -\sin\theta_W \tilde{B} + \cos\theta_W \tilde{W}$, where in the supersymmetric limit, these are the superpartners of the photon and $Z$-boson, respectively. The new mass matrix, in the $\{ \tilde{\gamma},\tilde{Z},\tilde{H}_d,\tilde{H}_u \}$ is $$\mathcal{M} =
\begin{pmatrix}
M_1{\cos\theta_W}^2 +M_2{\sin\theta_W}^2 & (M_2-M_1)\sin\theta_W\cos\theta_W & 0 & 0 \\
(M_2-M_1)\sin\theta_W\cos\theta_W & M_2{\cos\theta_W}^2 +M_1{\sin\theta_W}^2 & M_Z \cos\beta & -M_Z\sin\beta \\
0 & M_Z\cos\beta & 0 & -\mu \\
0 & -M_Z\sin\beta & -\mu & 0
\end{pmatrix}
\label{eqn:massmatrix2}$$ Taking the limit $M_1 = M_2 \equiv M$, one immediately one finds that $\tilde{\gamma}$ in an eigenvector with mass eigenvalue $M$. The next lightest eigenvector of the remaining $3\times 3$ sub-matrix will be mostly $\tilde{Z}$, and to leading order in $M_Z/\mu$, the mass is M\_M - ( -2 ) Next we will assume that the phases of $M$ and $\mu$ are such that absolute value of $ M_\chi$ is smaller than $|M|$, so that it is indeed the LSP. The other scenario, in which the LSP is mostly $\tilde{\gamma}$, will have negligible scattering cross-section.
Diagonalizing the remaining $3\times 3$ sub-matrix, the coefficients of $\tilde{H}_u$ and $\tilde{H}_d$ component to leading order is $$\begin{aligned}
Z_{H_d} &= \frac{1}{2}\left( \frac{\left(\sin\beta - \cos\beta\right)M_Z}{\mu+M} + \frac{\left(\sin\beta + \cos\beta\right)M_Z}{\mu - M} \right) = \frac{M_Z\left( \mu\sin\beta+M\cos\beta \right)}{\mu^2 - M^2} \nonumber \\
Z_{H_u} &= \frac{1}{2}\left( \frac{\left(\sin\beta - \cos\beta\right)M_Z}{\mu+M} -\frac{\left(\sin\beta + \cos\beta\right)M_Z}{\mu - M} \right) = -\frac{M_Z\left( \mu\cos\beta+M\sin\beta \right)}{\mu^2 - M^2}
\label{eqn:hu}\end{aligned}$$ and from the definition on $\tilde{Z}$, Z\_W - \_W Z\_B = [\_W]{}\^[-1]{}. \[eqn:zu\]
Finally, using (\[eqn:hu\]) and (\[eqn:zu\]) as inputs to (\[eqn:tim\]) the upper limit for the cross section is $$\sigma_{\rm SI} \left( \chi N \rightarrow \chi N \right) \approx 6 \times 10^{-45} \text{cm}^2 \left( \frac{115 \operatorname{GeV}}{m_h} \right)^4
\left(\frac{1 \operatorname{TeV}}{\mu}\right)^2 \left( \frac{\sin2\beta+M_{2}/\mu}{1 - (M_2/\mu)^2} \right)^2
\label{eqn:upperlimit}$$ From the discussion in the text we expect $M_2 /\mu \lesssim 0.2$, $\sin{2\beta} \lesssim 0.8$ and $\mu \gtrsim 1 \operatorname{TeV}$, giving largest scattering cross-sections around $\sigma_{\rm SI} \lesssim 6 \times 10^{-45} \text{cm}^2$. However, as discussed in Section 6.3, the constraints cannot all be satisfied simultaneously, so in practice only a cross section of about $10^{-45} \operatorname{cm}^2$ could be achieved.
[10]{}
S. P. Martin, [*[A Supersymmetry Primer]{}*]{}, [[hep-ph/9709356]{}](http://xxx.lanl.gov/abs/hep-ph/9709356).
J. E. Kim and H. P. Nilles, [*[The mu Problem and the Strong CP Problem]{}*]{}, [*Phys.Lett.*]{} [**B138**]{} (1984) 150.
I. Antoniadis, E. Gava, K. Narain, and T. Taylor, [*[Effective mu term in superstring theory]{}*]{}, [*Nucl.Phys.*]{} [**B432**]{} (1994) 187–204, \[[[hep-th/9405024]{}](http://xxx.lanl.gov/abs/hep-th/9405024)\].
P. Nath and T. R. Taylor, [*[Modular invariance, soft breaking, mu and tan beta in superstring models]{}*]{}, [*Phys.Lett.*]{} [**B548**]{} (2002) 77–87, \[[[hep-ph/0209282]{}](http://xxx.lanl.gov/abs/hep-ph/0209282)\].
D. Suematsu and Y. Yamagishi, [*[Radiative symmetry breaking in a supersymmetric model with an extra U(1)]{}*]{}, [*Int.J.Mod.Phys.*]{} [**A10**]{} (1995) 4521–4536, \[[[ hep-ph/9411239]{}](http://xxx.lanl.gov/abs/hep-ph/9411239)\].
M. Cvetic and P. Langacker, [*[Implications of Abelian extended gauge structures from string models]{}*]{}, [*Phys.Rev.*]{} [**D54**]{} (1996) 3570–3579, \[[[ hep-ph/9511378]{}](http://xxx.lanl.gov/abs/hep-ph/9511378)\].
M. Cvetic, D. A. Demir, J. Espinosa, L. Everett, and P. Langacker, [ *[Electroweak breaking and the mu problem in supergravity models with an additional U(1)]{}*]{}, [*Phys.Rev.*]{} [**D56**]{} (1997) 2861, \[[[hep-ph/9703317]{}](http://xxx.lanl.gov/abs/hep-ph/9703317)\].
O. Lebedev and S. Ramos-Sanchez, [*[The NMSSM and String Theory]{}*]{}, [ *Phys.Lett.*]{} [**B684**]{} (2010) 48–51, \[[[arXiv:0912.0477]{}](http://xxx.lanl.gov/abs/arXiv:0912.0477)\].
S. Ramos-Sanchez, [*[The mu-problem, the NMSSM and string theory]{}*]{}, [[arXiv:1003.1307]{}](http://xxx.lanl.gov/abs/arXiv:1003.1307).
M. Ratz, [*[Stringy Surprises]{}*]{}, [*Prog.Theor.Phys.Suppl.*]{} [**180**]{} (2010) 96–111, \[[[ arXiv:1003.0549]{}](http://xxx.lanl.gov/abs/arXiv:1003.0549)\].
J. Casas and C. Munoz, [*[A Natural solution to the mu problem]{}*]{}, [ *Phys.Lett.*]{} [**B306**]{} (1993) 288–294, \[[[hep-ph/9302227]{}](http://xxx.lanl.gov/abs/hep-ph/9302227)\].
R. Kappl, H. P. Nilles, S. Ramos-Sanchez, M. Ratz, K. Schmidt-Hoberg, [ *et. al.*]{}, [*[Large hierarchies from approximate R symmetries]{}*]{}, [ *Phys.Rev.Lett.*]{} [**102**]{} (2009) 121602, \[[[arXiv:0812.2120]{}](http://xxx.lanl.gov/abs/arXiv:0812.2120)\].
L. Ibanez and A. Uranga, [*[Instanton induced open string superpotentials and branes at singularities]{}*]{}, [*JHEP*]{} [**0802**]{} (2008) 103, \[[[arXiv:0711.1316]{}](http://xxx.lanl.gov/abs/arXiv:0711.1316)\].
L. Ibanez and . Richter, Robert, [*[Stringy Instantons and Yukawa Couplings in MSSM-like Orientifold Models]{}*]{}, [*JHEP*]{} [**0903**]{} (2009) 090, \[[[arXiv:0811.1583]{}](http://xxx.lanl.gov/abs/arXiv:0811.1583)\].
D. Green and T. Weigand, [*[Retrofitting and the mu Problem]{}*]{}, [[arXiv:0906.0595]{}](http://xxx.lanl.gov/abs/arXiv:0906.0595).
M. Cvetic, J. Halverson, and . Richter, Robert, [*[Realistic Yukawa structures from orientifold compactifications]{}*]{}, [*JHEP*]{} [**0912**]{} (2009) 063, \[[[ arXiv:0905.3379]{}](http://xxx.lanl.gov/abs/arXiv:0905.3379)\].
P. Langacker and M.-x. Luo, [*[Implications of precision electroweak experiments for M(t), rho(0), sin\*\*2-Theta(W) and grand unification]{}*]{}, [ *Phys.Rev.*]{} [**D44**]{} (1991) 817–822.
H. Murayama and A. Pierce, [*[Not even decoupling can save minimal supersymmetric SU(5)]{}*]{}, [*Phys.Rev.*]{} [**D65**]{} (2002) 055009, \[[[hep-ph/0108104]{}](http://xxx.lanl.gov/abs/hep-ph/0108104)\].
E. Witten, [*[Deconstruction, G(2) holonomy, and doublet triplet splitting]{}*]{}, [[ hep-ph/0201018]{}](http://xxx.lanl.gov/abs/hep-ph/0201018).
B. S. Acharya, K. Bobkov, G. Kane, P. Kumar, and D. Vaman, [*[An M theory Solution to the Hierarchy Problem]{}*]{}, [*Phys.Rev.Lett.*]{} [**97**]{} (2006) 191601, \[[[ hep-th/0606262]{}](http://xxx.lanl.gov/abs/hep-th/0606262)\].
B. S. Acharya, [*[On Realizing N=1 superYang-Mills in M theory]{}*]{}, [[hep-th/0011089]{}](http://xxx.lanl.gov/abs/hep-th/0011089).
B. S. Acharya, [*[M theory, Joyce orbifolds and super Yang-Mills]{}*]{}, [ *Adv. Theor. Math. Phys.*]{} [**3**]{} (1999) 227–248, \[[[hep-th/9812205]{}](http://xxx.lanl.gov/abs/hep-th/9812205)\].
E. Witten, [*[Anomaly cancellation on G(2) manifolds]{}*]{}, [[hep-th/0108165]{}](http://xxx.lanl.gov/abs/hep-th/0108165).
B. S. Acharya and E. Witten, [*[Chiral fermions from manifolds of G(2) holonomy]{}*]{}, [[ hep-th/0109152]{}](http://xxx.lanl.gov/abs/hep-th/0109152).
B. S. Acharya and S. Gukov, [*[M theory and singularities of exceptional holonomy manifolds]{}*]{}, [*Phys.Rept.*]{} [**392**]{} (2004) 121–189, \[[[hep-th/0409191]{}](http://xxx.lanl.gov/abs/hep-th/0409191)\].
T. Pantev and M. Wijnholt, [*[Hitchin’s Equations and M-Theory Phenomenology]{}*]{}, [[ arXiv:0905.1968]{}](http://xxx.lanl.gov/abs/arXiv:0905.1968).
M. B. Green and J. H. Schwarz, [*[Anomaly Cancellation in Supersymmetric D=10 Gauge Theory and Superstring Theory]{}*]{}, [*Phys.Lett.*]{} [**B149**]{} (1984) 117–122.
S. M. Barr, [*[A New Symmetry Breaking Pattern for SO(10) and Proton Decay]{}*]{}, [*Phys.Lett.*]{} [**B112**]{} (1982) 219.
J. Derendinger, J. E. Kim, and D. V. Nanopoulos, [*[Anti-SU(5)]{}*]{}, [ *Phys.Lett.*]{} [**B139**]{} (1984) 170.
I. Antoniadis, J. R. Ellis, J. Hagelin, and D. V. Nanopoulos, [ *[Supersymmetric Flipped SU(5) Revitalized]{}*]{}, [*Phys.Lett.*]{} [**B194**]{} (1987) 231.
E. Kuflik and J. Marsano, [*[Comments on Flipped SU(5) (and F-theory)]{}*]{}, [[1009.2510]{}](http://xxx.lanl.gov/abs/1009.2510).
B. S. Acharya, K. Bobkov, G. L. Kane, P. Kumar, and J. Shao, [*[Explaining the Electroweak Scale and Stabilizing Moduli in M Theory]{}*]{}, [*Phys.Rev.*]{} [**D76**]{} (2007) 126010, \[[[ hep-th/0701034]{}](http://xxx.lanl.gov/abs/hep-th/0701034)\].
B. S. Acharya, G. Kane, and E. Kuflik, [*[String Moduli Phenomenology, Cosmological History, Supersymmetry Breaking, and Dark Matter]{}*]{}, [[arXiv:1006.3272]{}](http://xxx.lanl.gov/abs/arXiv:1006.3272).
D. D. Joyce, [*Compact Manifolds with Special Holonomy*]{}. Oxford University Press, 2000.
B. S. Acharya, K. Bobkov, G. L. Kane, J. Shao, and P. Kumar, [*[The G(2)-MSSM: An M Theory motivated model of Particle Physics]{}*]{}, [ *Phys.Rev.*]{} [**D78**]{} (2008) 065038, \[[[arXiv:0801.0478]{}](http://xxx.lanl.gov/abs/arXiv:0801.0478)\].
Y. Hosotani, [*[Dynamical Gauge Symmetry Breaking as the Casimir Effect]{}*]{}, [*Phys.Lett.*]{} [**B129**]{} (1983) 193.
E. Witten, [*[Symmetry Breaking Patterns in Superstring Models]{}*]{}, [ *Nucl.Phys.*]{} [**B258**]{} (1985) 75.
H. M. Lee [*et. al.*]{}, [*[A unique $Z_4^R$ symmetry for the MSSM]{}*]{}, [[1009.0905]{}](http://xxx.lanl.gov/abs/1009.0905).
G. Giudice and A. Masiero, [*[A Natural Solution to the mu Problem in Supergravity Theories]{}*]{}, [*Phys.Lett.*]{} [**B206**]{} (1988) 480–484.
J. Wess and J. Bagger, [*[Supersymmetry and supergravity]{}*]{}, .
A. Brignole, L. E. Ibanez, and C. Munoz, [*[Soft supersymmetry-breaking terms from supergravity and superstring models]{}*]{}, [[hep-ph/9707209]{}](http://xxx.lanl.gov/abs/hep-ph/9707209).
B. S. Acharya and M. Torabian, [*[Supersymmetry Breaking, Moduli Stabilization and Hidden U(1) Breaking in M-Theory]{}*]{}, [[1101.0108]{}](http://xxx.lanl.gov/abs/1101.0108). \* Temporary entry \*.
Collaboration, C. Amsler [*et. al.*]{}, [*[Review of Particle Physics]{}*]{}, [*Phys.Lett.*]{} [**B667**]{} (2008) 1.
M. Ambroso and B. A. Ovrut, [*[The Mass Spectra, Hierarchy and Cosmology of B-L MSSM Heterotic Compactifications]{}*]{}, [[1005.5392]{}](http://xxx.lanl.gov/abs/1005.5392).
H. K. Dreiner, [*[An Introduction to explicit R-parity violation]{}*]{}, [[hep-ph/9707435]{}](http://xxx.lanl.gov/abs/hep-ph/9707435). To be published in ’Perspectives on Supersymmetry’, Ed. by G.L. Kane, World Scientific.
M. Reno and D. Seckel, [*[Primordial Nucleosynthesis: The Effects of Injecting Hadrons]{}*]{}, [*Phys.Rev.*]{} [**D37**]{} (1988) 3441.
J. R. Ellis, G. Gelmini, J. L. Lopez, D. V. Nanopoulos, and S. Sarkar, [ *[Astrophysical constraints on massive unstable neutral relic particles]{}*]{}, [*Nucl.Phys.*]{} [**B373**]{} (1992) 399–437.
V. Berezinsky, A. Masiero, and J. Valle, [*[Cosmological signatures of supersymmetry with spontaneously broken R-parity]{}*]{}, [*Phys.Lett.*]{} [ **B266**]{} (1991) 382–388.
E. A. Baltz and P. Gondolo, [*[Limits on R-parity violation from cosmic ray antiprotons]{}*]{}, [*Phys. Rev.*]{} [**D57**]{} (1998) 7601–7606, \[[[hep-ph/9704411]{}](http://xxx.lanl.gov/abs/hep-ph/9704411)\].
A. Arvanitaki [*et. al.*]{}, [*[Decaying Dark Matter as a Probe of Unification and TeV Spectroscopy]{}*]{}, [*Phys. Rev.*]{} [**D80**]{} (2009) 055011, \[[[0904.2789]{}](http://xxx.lanl.gov/abs/0904.2789)\].
S. Shirai, F. Takahashi, and T. T. Yanagida, [*[R-violating Decay of Wino Dark Matter and electron/positron Excesses in the PAMELA/Fermi Experiments]{}*]{}, [*Phys. Lett.*]{} [**B680**]{} (2009) 485–488, \[[[0905.0388]{}](http://xxx.lanl.gov/abs/0905.0388)\].
S. King, G. Leontaris, and G. Ross, [*[Family symmetries in F-theory GUTs]{}*]{}, [*Nucl.Phys.*]{} [**B838**]{} (2010) 119–135, \[[[arXiv:1005.1025]{}](http://xxx.lanl.gov/abs/arXiv:1005.1025)\].
B. Allanach, [*[SOFTSUSY: a program for calculating supersymmetric spectra]{}*]{}, [*Comput.Phys.Commun.*]{} [**143**]{} (2002) 305–331, \[[[hep-ph/0104145]{}](http://xxx.lanl.gov/abs/hep-ph/0104145)\].
D. Feldman, G. Kane, R. Lu, and B. D. Nelson, [*[Dark Matter as a Guide Toward a Light Gluino at the LHC]{}*]{}, [*Phys.Lett.*]{} [**B687**]{} (2010) 363–370, \[[[1002.2430]{}](http://xxx.lanl.gov/abs/1002.2430)\].
J. L. Feng, K. T. Matchev, and T. Moroi, [*[Focus points and naturalness in supersymmetry]{}*]{}, [*Phys. Rev.*]{} [**D61**]{} (2000) 075005, \[[[hep-ph/9909334]{}](http://xxx.lanl.gov/abs/hep-ph/9909334)\].
K. L. Chan, U. Chattopadhyay, and P. Nath, [*[Naturalness, weak scale supersymmetry and the prospect for the observation of supersymmetry at the Tevatron and at the CERN LHC]{}*]{}, [*Phys.Rev.*]{} [**D58**]{} (1998) 096004, \[[[hep-ph/9710473]{}](http://xxx.lanl.gov/abs/hep-ph/9710473)\].
D. M. Pierce, J. A. Bagger, K. T. Matchev, and R.-j. Zhang, [*[Precision corrections in the minimal supersymmetric standard model]{}*]{}, [*Nucl. Phys.*]{} [**B491**]{} (1997) 3–67, \[[[hep-ph/9606211]{}](http://xxx.lanl.gov/abs/hep-ph/9606211)\].
B. S. Acharya, P. Kumar, K. Bobkov, G. Kane, J. Shao, [*et. al.*]{}, [ *[Non-thermal Dark Matter and the Moduli Problem in String Frameworks]{}*]{}, [ *JHEP*]{} [**0806**]{} (2008) 064, \[[[arXiv:0804.0863]{}](http://xxx.lanl.gov/abs/arXiv:0804.0863)\].
B. S. Acharya, G. Kane, S. Watson, and P. Kumar, [*[A Non-thermal WIMP Miracle]{}*]{}, [*Phys.Rev.*]{} [**D80**]{} (2009) 083529, \[[[0908.2430]{}](http://xxx.lanl.gov/abs/0908.2430)\].
P. Grajek, G. Kane, D. J. Phalen, A. Pierce, and S. Watson, [*[Neutralino Dark Matter from Indirect Detection Revisited]{}*]{}, [[0807.1508]{}](http://xxx.lanl.gov/abs/0807.1508).
J. Hisano, M. Kawasaki, K. Kohri, and K. Nakayama, [*[Positron/Gamma-Ray Signatures of Dark Matter Annihilation and Big-Bang Nucleosynthesis]{}*]{}, [ *Phys.Rev.*]{} [**D79**]{} (2009) 063514, \[[[arXiv:0810.1892]{}](http://xxx.lanl.gov/abs/arXiv:0810.1892)\].
G. Kane, R. Lu, and S. Watson, [*[PAMELA Satellite Data as a Signal of Non-Thermal Wino LSP Dark Matter]{}*]{}, [*Phys.Lett.*]{} [**B681**]{} (2009) 151–160, \[[[ arXiv:0906.4765]{}](http://xxx.lanl.gov/abs/arXiv:0906.4765)\]. \* Brief entry \*.
D. Feldman, Z. Liu, P. Nath, and B. D. Nelson, [*[Explaining PAMELA and WMAP data through Coannihilations in Extended SUGRA with Collider Implications]{}*]{}, [*Phys.Rev.*]{} [**D80**]{} (2009) 075001, \[[[arXiv:0907.5392]{}](http://xxx.lanl.gov/abs/arXiv:0907.5392)\].
N. Chen, D. Feldman, Z. Liu, P. Nath, and G. Peim, [*[Positron and Photon Compliant Higgsino Dark Matter and LHC-7]{}*]{}, [[1010.0939]{}](http://xxx.lanl.gov/abs/1010.0939). \* Temporary entry \*.
N. Chen, D. Feldman, Z. Liu, P. Nath, and G. Peim, [*[Low Mass Gluino within the Sparticle Landscape, Implications for Dark Matter, and Early Discovery Prospects at LHC-7]{}*]{}, [[ 1011.1246]{}](http://xxx.lanl.gov/abs/1011.1246). \* Temporary entry \*.
Collaboration, Z. Ahmed [*et. al.*]{}, [*[Dark Matter Search Results from the CDMS II Experiment]{}*]{}, [*Science*]{} [**327**]{} (2010) 1619–1621, \[[[0912.3592]{}](http://xxx.lanl.gov/abs/0912.3592)\].
Collaboration, E. Aprile [*et. al.*]{}, [*[First Dark Matter Results from the XENON100 Experiment]{}*]{}, [[1005.0380]{}](http://xxx.lanl.gov/abs/1005.0380).
Collaboration, R. Barate [ *et. al.*]{}, [*[Search for the standard model Higgs boson at LEP]{}*]{}, [ *Phys. Lett.*]{} [**B565**]{} (2003) 61–75, \[[[hep-ex/0306033]{}](http://xxx.lanl.gov/abs/hep-ex/0306033)\].
T. Cohen, D. J. Phalen, and A. Pierce, [*[On the Correlation Between the Spin-Independent and Spin- Dependent Direct Detection of Dark Matter]{}*]{}, [ *Phys. Rev.*]{} [**D81**]{} (2010) 116001, \[[[1001.3408]{}](http://xxx.lanl.gov/abs/1001.3408)\].
M. El Kheishen, A. Aboshousha, and A. Shafik, [*[Analytic formulas for the neutralino masses and the neutralino mixing matrix]{}*]{}, [*Phys.Rev.*]{} [ **D45**]{} (1992) 4345–4348.
V. D. Barger, M. Berger, and P. Ohmann, [*[The Supersymmetric particle spectrum]{}*]{}, [*Phys.Rev.*]{} [**D49**]{} (1994) 4908–4930, \[[[hep-ph/9311269]{}](http://xxx.lanl.gov/abs/hep-ph/9311269)\].
G. Bertone, D. Hooper, and J. Silk, [*[Particle dark matter: Evidence, candidates and constraints]{}*]{}, [*Phys. Rept.*]{} [**405**]{} (2005) 279–390, \[[[hep-ph/0404175]{}](http://xxx.lanl.gov/abs/hep-ph/0404175)\].
P. Gondolo, J. Edsjo, P. Ullio, L. Bergstrom, M. Schelke, [*et. al.*]{}, [ *[DarkSUSY: Computing supersymmetric dark matter properties numerically]{}*]{}, [*JCAP*]{} [**0407**]{} (2004) 008, \[[[astro-ph/0406204]{}](http://xxx.lanl.gov/abs/astro-ph/0406204)\].
Collaboration, O. Adriani [*et. al.*]{}, [*[An anomalous positron abundance in cosmic rays with energies 1.5-100 GeV]{}*]{}, [*Nature*]{} [**458**]{} (2009) 607–609, \[[[arXiv:0810.4995]{}](http://xxx.lanl.gov/abs/arXiv:0810.4995)\].
G. L. Kane and J. M. Frere, [*[On The Possibility Of Finding Light Uncolored Supersymmetric Partners At Present And Future Machines]{}*]{}, [*Nucl. Phys.*]{} [**B223**]{} (1983) 331.
[^1]: Presumably, $N$ is of the same order as the number of renormalizable coupling constants of the effective low energy theory.
[^2]: Note section 2. The “$i$”’s are not the same in $S$ and $z$.
[^3]: We leave the case of $N \geq 10^4$ for further study.
[^4]: The final term in (\[WRbreaking\]) can be rotated away in superpotential by a unitary transformation on $(h_d, L)$. This rotation will induce additional contributions to the lepton violating coupling constants $\lambda^{\prime}$ and $\lambda^{\prime\prime}$ that are proportional to the Yukawa couplings. Assuming that $\mu \gtrsim \kappa $, their sizes are approximately \^ \~y\_e \^ \~y\_d \[muyukawa\]
[^5]: If this $U(1)$ symmetry is to be broken to $R$-parity, then requiring the symmetry to be flavor blind, allowing for Higgs triplet masses, and allowing an explanation for neutrino masses, basically constrains the charges of the MSSM and right handed neutrino fields to be the $U(1)_\chi$ charges.
[^6]: See [@Feldman:2010uv] for general phenomenological discussions.
[^7]: see Section IV of [@Acharya:2008zi] for the precise definition of $\delta$
[^8]: Since there are theoretical and calculational uncertainties with calculating the Higgs mass, we will consider models with $m_h \ge 110$ GeV.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We use functional, Fréchet, derivatives to quantify how thermodynamic outputs of a molecular dynamics (MD) simulation depend on the potential used to compute atomic interactions. Our approach quantifies the sensitivity of the quantities of interest with respect to the input *functions* as opposed to its parameters as is done in typical uncertainty quantification methods. We show that the functional sensitivity of the average potential energy and pressure in isothermal, isochoric MD simulations using Lennard-Jones two-body interactions can be used to accurately predict those properties for other interatomic potentials (with different functional forms) without re-running the simulations. This is demonstrated under three different thermodynamic conditions, namely a crystal at room temperature, a liquid at ambient pressure, and a high pressure liquid. The method provides accurate predictions as long as the change in potential does not significantly affect the region in phase space explored by the simulation. The functional uncertainty quantification approach can be used to estimate the uncertainties associated with constitutive models used in the simulation and to correct predictions if a more accurate representation becomes available.'
address: 'School of Materials Engineering, Birck Nanotechnology Center, Purdue University, West Lafayette, Indiana 47906, USA'
author:
- Samuel Reeve
- Alejandro Strachan
bibliography:
- 'bibtex/Reeve-Strachan-2016\_main.bib'
title: 'Error correction in multi-fidelity molecular dynamics simulations using functional uncertainty quantification'
---
uncertainty quantification ,functional derivative ,molecular dynamics ,interatomic potential ,free energy calculation
Introduction
============
Uncertainty quantification (UQ) is becoming increasingly important in predictive simulations of materials and devices [@Chernatynskiy2013]. While the majority of the early UQ work focused on solid and fluid mechanics, there is growing interest in applying and extending UQ techniques to simulations at the material level, including density functional theory [@Mortensen2005; @Anderson2012], molecular dynamics (MD) [@Rizzi2012p1; @Rizzi2013p1; @Angelikopoulos2013; @Karniadakis2015; @Patrone2016], and multiscale methods [@Koslowski2011; @Kim2013; @Vedula2013]. These efforts highlighted the importance of acknowledging uncertainties in model parameters, from measurement or averaging techniques, as well as intrinsic variability of the systems or processes under investigation. These studies, and most UQ work thus far, examined uncertainties in input parameters and specialized software packages have been developed for this task: examples include DAKOTA [@Adams2014], PUQ [@Hunt2015], and $\Pi$4U [@Hadjidoukas2015]. However, in many applications – especially those involving complex physics – accounting for uncertainty in parameters is not enough, as the functional forms of the constitutive models themselves are approximate. This is particularly true in materials modeling where lack of knowledge leads to unquantified or poorly quantified uncertainties. Examples of input functions with varying degrees of accuracy include exchange and correlation functionals used in density functional theory calculations [@Mortensen2005; @Aldegunde2016], interatomic potentials for molecular dynamics (MD) [@Rizzi2012p2; @Rizzi2013p2; @Farrell2015], generalized stacking faults used in dislocation dynamics [@Cao2015], and constitutive laws for micromechanical simulations. In this paper we use functional derivatives (FD), recently proposed as a mathematical framework to quantify uncertainties that arise from constitutive models used in simulations [@Strachan2013], to quantify and correct uncertainties that originate from the interatomic potential used in MD simulations.
Functional uncertainty quantification (FUQ) can, in principle, be used to assess the uncertainties originating from approximate input constitutive laws, correct predictions if a more accurate function becomes available, and rank when and where to replace a low-fidelity model used in a simulation with one of higher fidelity in order to reduce prediction error by running additional simulations. This paper introduces a computationally efficient method to compute FDs in MD simulations involving two-body interatomic potentials, extending ideas from thermodynamic integration and free-energy perturbation methods [@Frenkel2001; @Chipot2007]. The FD quantifies how a quantity of interest (QoI) – in this case the total potential energy or pressure computed from an MD simulation in the canonical ensemble – depends on the input function, here the Lennard-Jones (LJ) two-body pair potential. We further show that the FD with respect to the LJ potential can be used to compute accurate correction to the potential energy and pressure for a family of pair potentials without re-running the simulation. This is true as long as the discrepancy between the potentials remains within reasonable bounds and the phase space explored by the system with the new potential is not significantly different from that of the original.
The remainder of the paper is as follows. Section 2 describes the functional approach to UQ followed by simulation details and methods to calculate the functional derivative numerically in Section 3. In Section 4 we describe specific examples of error correction using FUQ with low and high-fidelity models. Sections 5 and 6 discuss results and conclude the paper.
Functional uncertainty quantification
=====================================
In general, a simulation predicts a quantity of interest, $Q$, given some set of input parameters $P_i$ and constitutive functions $f_j$, themselves functions of an independent variable $z$ and input parameters $N_k$: $$\label{eq:QoI}
Q=Q({P_i},{f_j(z,{N_k})})$$ In the cases of interest here, $Q$ will be the time averaged potential energy or pressure and $f(z)$ the LJ potential as a function of interatomic distance used in the MD simulation.
The most common form of UQ is concerned with uncertainty in simulation outputs arising from uncertainty in the input parameters. This approach, while very valuable, ignores the fact that the functional forms used as constitutive laws are, almost invariably, approximate and lead to errors in the simulation. To quantify uncertainties with respect to input functions we utilize functional derivatives (Fréchet derivatives) of the QoI with respect to the input functions. The FD can be written in differential form as: $$\label{eq:FD1}
\frac{\delta Q[f]}{\delta f(z)}(z_i) = \lim_{\epsilon \to 0} \frac{Q[f(z) + \epsilon\cdot\delta(z-z_i)] - Q[f(z)]}{\epsilon}$$ This characterizes the functional sensitivity of the QoI with respect to the input function. The definition in Eq. \[eq:FD1\] uses the Dirac delta function as the functional variation; as we describe below, to calculate the functional derivative numerically for the MD simulations we use narrow Gaussian distributions centered at $z_i$.
Besides quantifying uncertainties in the prediction given uncertain input functions, the functional derivative can be used to correct the error that arises from the use of a low-fidelity model $(f)$ if a higher-fidelity one $(g)$ becomes available. The correction is then obtained using the product of the functional sensitivity (the FD) and functional discrepancy $(g(z)-f(z))$. We call this product the functional error and the correction for the QoI is: $$\label{eq:correct}
\Delta Q = \int \frac{\delta Q[f]}{\delta f(z)} \cdot (g(z)-f(z)) dz$$ In the absence of a high-fidelity model Eq. \[eq:correct\] can alternatively be used for uncertainty propagation by replacing the discrepancy with the uncertainty in $f(z)$ and taking the absolute value of the functional derivative [@Strachan2013]. This procedure returns a first order bound of the uncertainty in Q.
Additionally, this equation can be used to rank high-fidelity simulations in order of their functional error to optimally reduce the error in the predicted QoI, recently demonstrated with calculation of the functional derivative for the restoring force in a multi-fidelity radio frequency MEMS switch simulation [@Strachan2013]. This functional derivative was used to rank model evaluations and minimize the necessary computational cost to maximize error correction in the simulation. In Section 3 we extend the formulation to a significantly more challenging problem: molecular dynamics simulations.
Functional derivatives in molecular dynamics
============================================
Systems of interest and simulation details
------------------------------------------
We demonstrate the FUQ method in molecular dynamics simulations where the input function is the pairwise interatomic potential and the QoIs are the potential energy and pressure (long-time averages) of the system. We take the Lennard-Jones 12-6 potential as the low-fidelity input function: $$\label{eq:LJ}
\phi_0 = 4\epsilon((\sigma/r)^{12}-(\sigma/r)^6)$$ and test the ability of FUQ to correct the prediction of the QoI for a family of high-fidelity potentials. The parameters for the low-fidelity LJ potential are designed to roughly describe copper $(\sigma=2.315$Å, $\epsilon=0.167 eV)$ [@Wolf1992], fit to the bulk melt temperature and room temperature lattice constant. We note that this is not an accurate potential for Cu as it ignores important many body effects critical to describe elastic constants and defect energetics. Also notable is the discrepancy of the liquid densities as compared to experiment [@Cahill1962]. However, the goal of the paper is to demonstrate the applicability of FUQ to an MD problem. In this spirit, the high-fidelity potentials are similarly designed for demonstration purposes only and do not represent a more accurate representation of an actual material. We construct seven pair potentials by additively modifying the LJ potential with sine functions in Table 1. This family of functions will be denoted Sine 1 to Sine 7.
Name Modification function
-------- ------------------------------------------
Sine 1 $0.44+0.46\sin(0.17(24.2+r))$
Sine 2 $-0.47\sin(-0.15(14+r))\cdot \exp(-r)$
Sine 3 $0.07\sin(1.2(-1.2+r))/r^2$
Sine 4 $-0.01+0.2\sin(0.3(14+r))\cdot \exp(-r)$
Sine 5 $0.7\sin(0.4(11+r))\cdot \exp(-r)$
Sine 6 $0.9\sin(0.4(11+r))\cdot \exp(-r)$
Sine 7 $1.1\sin(0.4(11+r))\cdot \exp(-r)$
: High-fidelity potential sine modification terms[]{data-label="tbl:sine"}
A final high-fidelity model tested is the Morse potential, defined as the sum of two exponentials: $$\label{eq:Morse}
\phi=D_0(e^{-2\alpha(r-r_0)} -2e^{-\alpha(r-r_0)})$$ We use $D_0=0.161 eV$, $\alpha=2.09$Å${}^{-1}$, and $r_0=2.62$Å to similarly roughly describe copper.
For all potentials a smoothing function is used to ensure stable dynamics near the cutoff. The function is of fourth order to create potential energy and force curves that both smoothly tend to zero: $$\label{eq:smooth}
s(r) = \frac{(\frac{(r - r_c)}{w})^4}{(1+\frac{(r - r_c)}{w})^4}$$ where $r_c$ is the cutoff distance and $w$ is the width of the smoothing. For all potentials the cutoff is 5.79Å ($2.5\sigma$, commonly used for LJ potentials) and smoothing width 1.5Å. Each potential curve is created by summing the base equation and the modification term and subsequently taking the product with the smoothing function. The force curve is then created with product rule differentiation of the potential. These potentials are shown in Figure \[fig:potentials\].
Each system consists of 500 atoms simulated under isothermal, isochoric conditions (canonical ensemble). Three physical states of the system were simulated: one solid at ambient temperature and pressure ($300 K$ with density of $9.02 g/cm^3$), one liquid slightly above the melt temperature at ambient pressure ($1300 K$ with density of $6.48 g/cm^3$) and another liquid at extreme temperature and pressure ($5000 K$ with density of $8.93 g/cm^3$, corresponding to $55 GPa$).
All MD simulations were performed using the LAMMPS package [@Plimpton1995]. For each high-fidelity potential we performed 16 independent simulations, each 1 ns long, sufficient for good statistical sampling of the quantities of interest. For each low-fidelity potential, from which the functional derivatives were calculated, the total sample time was 64 ns, similarly from 64 independent simulations of 1 ns. The convergence of this sampling is shown in the supplementary material.
Numerical functional derivatives and a perturbative approach for their calculation
----------------------------------------------------------------------------------
For the specific case of interest the functional derivative expression becomes: $$\label{eq:FD2}
\frac{\delta Q[\phi]}{\delta \phi}(r_i) = \lim_{\epsilon \to 0} \frac{Q[\phi_0(r) + \epsilon\cdot\phi'(r-r_i)] - Q[\phi_0(r)]}{\epsilon}$$ where $Q[\phi]$ denotes the average QoI – the potential energy or pressure of the system – using interatomic potential $\phi$. The unmodified LJ potential is $\phi_0$ and $\phi'$ a normalized Gaussian perturbation centered at interatomic distance $r=r_i$ with width $\sigma$.
The functional derivative in Eq. \[eq:FD2\] can be calculated by performing a set of simulations with the low-fidelity potential modified by perturbations of varying $\epsilon$ and computing a numerical derivative for each position $r_i$. These modified Lennard Jones potentials and resulting average potential energies as a function of size of the perturbation $\epsilon$ are shown in Fig. \[fig:calcFD\](a) and (b), respectively. However, this brute force approach is computationally very intensive considering that we may need to sample the functional derivative at 100 values of $r_i$. Even with only three values of $\epsilon$ for each separation distance one would need to perform 300 separate MD simulations. To alleviate the computational cost of the approach we now derive a perturbative approach to calculate the functional derivative that can be computed with little overhead with respect to the nominal simulation using the low-fidelity potential, $\phi_0$.
At the heart of the calculation of the FD is the evaluation of canonical averages of the QoIs with the interatomic potential and a Gaussian perturbation. Recognizing that the modified Hamiltonian of the system can be additively decomposed into the kinetic energy, the original LJ potential energy and the functional variation (or perturbation) for these two-body potentials we can write the canonical ensemble average of quantity Q as: $$\label{eq:ensavg}
\langle Q \rangle_H = \frac{\int Q \cdot e^{-\beta H_0} \cdot e^{-\beta H'}}{\int e^{-\beta H_0} \cdot e^{-\beta H'}}$$ where $H_0$ is the Hamiltonian with the LJ potential and $H'$ the potential energy resulting from the Gaussian perturbation following form from Eq. \[eq:FD2\]. Equation \[eq:ensavg\] can be re-written as the ratio between two canonical averages over the unmodified Hamiltonian by multiplying by additional factors of $\int e^{-\beta H_0}$ and rearranging: $$\label{eq:perturbative}
\langle Q \rangle_H = \langle Q \cdot e^{-\beta H'} \rangle_{H_0} \cdot
\frac{1}{\langle e^{-\beta H'} \rangle_{H_0}}$$ Since both canonical averages are over the unmodified potential only simulation with the low-fidelity potential is required. Therefore, computing the functional derivatives requires evaluating the Gaussian perturbation potentials on the trajectory obtained with the low-fidelity model. Such expressions are commonly used to compute free energies in thermodynamic integration and free energy perturbation approaches [@Straatsma1992; @Kollman1993].
Quantities for the canonical average in Eq. \[eq:perturbative\] are computed every 1 ps with a total of 48 ns of simulation time with the low-fidelity LJ potential, deemed well converged (see supplementary material). We evaluate these averages using a binned coordination number with a total of 2000 bins to compute $H'$ and $Q$ (as it contains contributions from the perturbation), described in more detail in the appendix. Use of the coordination number simply allows to further reduce computation and avoid modifying the MD code. Performing the calculation directly within the MD force loop with atomic positions would be equivalent excluding slight discrepancies from discretization.
Using Eq. \[eq:FD2\] with the first term in the difference calculated using Eq. \[eq:perturbative\] we compute the functional derivative with respect to Gaussian perturbations centered at $r_i$ with width $\sigma = 0.1$ Å and height $\epsilon=\pm0.00075$ and $\pm0.00375$ eV. The perturbation width was chosen to minimize (localize) the perturbation while retaining sufficient sampling. The heights were similarly minimized, due to the difficulty of converging exponentially weighted averages, while ensuring a measurable effect from the perturbations. The numerical derivative is then evaluated by computing the slope of the QoI with respect to $\epsilon$ (Fig. \[fig:calcFD\](b)) at a set of interatomic separations ranging from zero past the potential cutoff in increments of $0.05$ Å.
The thick blue lines in Figure \[fig:FD\] show the functional derivatives obtained in this manner for all physical conditions and both QoIs. These curves were averaged from multiple calculations of the FD with independent randomized samplings from the total of 64 ns of simulation time. The functional derivative displays significant information about the physics of the systems. The overall shapes correspond to the atomic shells as in the radial distribution function – notably more pointed for the solid case. At very low separation distances the FD goes to zero as atoms hit the soft-wall repulsion of the potentials; for the extreme liquid system the atoms occupy smaller separations.
To assess the accuracy of the calculation the functional derivative was also computed at equally spaced values of $r_i$ using the brute force approach described above. For each value of $r_i$ we perform four MD simulations with varying perturbations (identical to those in the perturbative approach) added to the LJ potential as shown in Fig. \[fig:calcFD\](a) and obtain the functional derivative with the same numerical derivatives as the perturbative approach. For each separate perturbation the simulation was run for 16 ns (split between 16 independent systems). The results of the brute force approach are shown in red in Fig. \[fig:FD\]; the two methods of calculating the functional derivative are nearly identical with only small numerical discrepancies.
Error correction using functional derivatives {#sec:results}
=============================================
To demonstrate our approach we now use the functional derivative calculated perturbatively in the previous section to correct the potential energy and pressure predicted with the low-fidelity LJ potential assuming a more accurate function is available. As discussed in subsection 3.1, a family of seven high-fidelity potentials were created by modifying the LJ potential with sine functions; see Figure \[fig:potentials\] and Table \[tbl:sine\]. These results are discussed in sub-section \[sec:resultsSine\]; the results for the Morse potential are discussed in sub-section \[sec:resultsMorse\].
Corrections for sine modified potentials {#sec:resultsSine}
----------------------------------------
In order to use Eq. \[eq:correct\] we need, in addition to the functional derivative, the discrepancy function $(\phi_{HF} (r) - \phi_{LF} (r))$, the difference between the high and low-fidelity potentials. Figure \[fig:correct\] illustrates the functional derivative (a), the discrepancy (b), and the product of the two, the functional error (c), each as a function of interatomic distance for one case, the ambient liquid with modified potential Sine 1. These results are shown for each physical case and potential in the supplementary material. Note that the functional error goes to zero both for small and large values of $r$; at small $r$ it goes to zero following the functional derivative (due to steep repulsion that keeps atoms from coming close to one another), while for large distances the discrepancy goes to zero as both potentials tend to zero at the same cutoff.
The functional error was numerically integrated using the trapezoid rule to obtain the total correction for the QoI. In order to verify these correction predictions we performed explicit MD simulations with the high-fidelity modified LJ potentials. Figure \[fig:results\] compares the potential energy (a) and pressure (b) differences explicitly simulated with low and high-fidelity potentials (gray) to the corrections obtained with FUQ (with colors matching those in Fig. \[fig:potentials\], as well as the hybrid LJ-Morse potential discussed in sub-section \[sec:resultsMorse\] in white). These results are shown in detail in Tables \[tbl:PE\] and \[tbl:P\] in the appendix.
For almost all cases, the FUQ predictions are in excellent agreement with the direct simulations. Excluding only the Sine 6 and 7 potentials for the ambient liquid (discussed below) average error is 0.600% and 1.70% for potential energy and pressure, respectively. We stress that these very accurate corrections are obtained only making use of the the simulation with the unmodified LJ potential, no additional MD simulations are required.
The ability of FUQ to correct the QoIs is impacted greatly by the degree of phase space overlap between the high and low-fidelity simulations. This can be shown simply by the overlap in histograms of probability distributions of differences in potential energy from the initial to final state, a common practice in free energy calculations [@Pohorille2010]. In this case, the states refer to the potential used and the distributions are taken from: $$\label{eq:hist0}
\Delta U_0 = (U(\phi_{HF}, \Gamma_{LF}) - U(\phi_{LF}, \Gamma_{LF}))$$ $$\label{eq:hist1}
\Delta U_1 = (U(\phi_{LF}, \Gamma_{HF}) - U(\phi_{HF}, \Gamma_{HF}))$$ where each term is the energy with potential $\phi$ and set of samples in phase space $\Gamma=(\mathbf{x_1}, ..., \mathbf{x_N})$ (each point dependent on the positions of the $N$ atoms) from the high or low-fidelity potential trajectory. Examples are shown in Fig. \[fig:hist\].
Lack of phase space overlap and difference in probability distributions is most significant for the the ambient liquid with the Sine 6 and 7 potentials, the same two cases with significantly larger error. Further inspection of the MD trajectory shows that the modified potential results in a structural transformation while the unmodified LJ does not. This is shown in snapshots included as insets in Fig. \[fig:hist\] comparing the atomic structures for the ambient liquid with Sine 1 and Sine 7: the liquid undergoes cavitation for Sine 7 (and to a lesser extent with Sine 5 and 6). Insets were created using the OVITO software package [@Stukowski2010]. Thus the low-fidelity LJ explores a vastly different region in phase space as compared to that of the high-fidelity Sine 7; however, even under such unfavorable conditions FUQ is able to provide some level of correction. In contrast, the solid cases have sharply peaked distributions with strong overlap and correspondingly highly accurate predictions, with the caveat of more difficult convergence of the results. Histograms for all physical cases and potentials are within the supplementary material.
The degree of similarity of phase space exploration can be further summarized by plotting the average of $\Delta U_0$ against the average of $\Delta U_1$. The closer to slope of unity, the more significant the distribution overlap. Figure \[fig:histpoints\] gives results that agree with the histograms in Fig. \[fig:hist\] and supplementary material; namely, for cases where the explored phase space for the low and high-fidelity trajectories is less similar and there is little histogram overlap, the prediction error increases, particularly for the ambient liquid. Results in this figure for the Morse (gray symbols) and hybrid LJ-Morse (white) potentials are discussed in sub-section \[sec:resultsMorse\].
Correction for the Morse potential and hybrid potential calculations {#sec:resultsMorse}
--------------------------------------------------------------------
Predicting properties for the Morse potential from LJ further highlights the challenge of attempting to correct a prediction made with a low-fidelity model. In this case the discrepancy between the two models is very large for short interatomic distances, leading to near zero phase space overlap and an inaccurate FUQ prediction for all physical cases. The red line in Figure \[fig:hybrid\] shows this discrepancy as a function of interatomic separation with the functional derivatives from both liquid conditions in blue to demonstrate the sensitivity to changes in the potential at these distances. The short range repulsion is described with an exponential for Morse while LJ uses an inverse power of 12. Thus the discrepancy between the two potentials diverges for short distances and the functional derivative correction, being first order, produces extremely large errors. These results are therefore omitted from Figure \[fig:results\].
Our approach can, nevertheless, be useful in such circumstances, but the high-fidelity model cannot be fully replaced by the low-fidelity one. Instead, we create a hybrid potential in Eq. \[eq:hybrid\] that smoothly switches from the low-fidelity LJ to the high-fidelity Morse between 2.8 and 2.4Å using error functions, leading to a discrepancy of reasonable magnitude shown in green in Figure \[fig:hybrid\].
$$\label{eq:hybrid}
\phi_{Hybrid} =
\begin{cases}
\phi_{LJ}(r) & r\geq 2.8 \\
\phi_{LJ}(r)\cdot [B(r)+0.5] + \phi_{Morse}(r)\cdot [-B(r)+0.5] & 2.8 > r > 2.4 \\
\phi_{Morse}(r) & 2.4 \geq r \\
\end{cases}$$
$$\label{eq:erf}
B(r) = 0.5\operatorname{erf}(8(r-2.6))$$
With the hybrid potential in Eq. \[eq:hybrid\] the high-fidelity model is used only where necessary. This is similar in spirit to adaptive sampling methods in multi-scale simulations, with examples in the literature [@Knap2008; @Barton2011; @Roehm2015] and codes available: the Co-design Embedded Visco-Plasticity Proxy Application (CoEVP) [@Dorr2014] and the Co-design Heterogeneous Multiscale Method Proxy Application (CoHMM) [@RouetLeduc2014].
It is then possible to use FUQ to correct a low-fidelity simulation run with this hybrid potential to the result with the high-fidelity Morse by using the discrepancy between the two potentials and the functional derivative of the hybrid potential. The hybrid potential reduces the discrepancy and the functional approach provides an accurate correction for all cases, as shown with the white bars in Figure \[fig:results\]. The solid case correction shows an additional example of increased error, again due to poor overlap in explored phase space between the two potentials (see supplementary material). Even with these significantly different systems FUQ provides the majority of the necessary correction. Further, all predictions for Morse from the hybrid potential (white symbols) lie near the unity line in Fig. \[fig:histpoints\] (describing good agreement between the phase space of each trajectory) in contrast to Morse from LJ (gray symbols), furthest from the line. We note that the high-fidelity model needs to be used during the actual simulation; however, only approximately 12% of the atomic force calculations fall in the range requiring the high-fidelity model. Thus, if the high-fidelity model was significantly more intensive than the low-fidelity the methodology would still greatly reduce computational cost.
Discussion
==========
The perturbative method to obtain the functional derivatives described here is similar to techniques in free energy methods: free energy perturbation and thermodynamic integration [@Bash1987; @Karplus1990]. One subset of these methods most comparable to FUQ and often used in biology simulations, referred to as computer alchemy, is utilized to calculate free energy changes along non-physical paths as potentials are slowly turned on or off for the various molecules or solvents of interest [@Gumbart2012; @He2014; @Wescott2002; @Lawrenz2012]. These methods share with the work described here the need to sample from the phase space trajectory of an initial state (often a simple reference state for free energy calculations, e.g. the Einstein crystal), ensure that phase space is not too dissimilar to that of the final state of interest, and to converge exponentially weighted averages. However, while free energy methods generally focus on a single path from initial to final state (or a bidirectional path), using FUQ, once the functional derivative is calculated for a given low-fidelity model, the corrections can be made with respect to any other high-fidelity model as long as minimum conditions are satisfied. If the low-fidelity simulation does not explore the regions of phase space relevant for the high-fidelity potential at the conditions of interest the results will be poor; this is most striking when switching potentials results in a structural phase transition. This phase space overlap should be checked as demonstrated here to ensure meaningful results as with free energy methods. Additionally, the functional derivative correction can only be used with a functional discrepancy such that its effect can be described to first order. We also note that the perturbative approach used here is applicable only to input functions that appear linearly in the Hamiltonian so the perturbation can be separated additively. This is not a general limitation of FUQ but of the specific approach used here to obtain the functional derivatives.
This work demonstrates FUQ in equilibrium simulations. Continuing work should include investigation of functional derivatives in non-equilibrium processes as an analog of the relationships between free energy perturbation and non-equilibrium work methods in free energy calculations. The method can be generally used in equilibrium MD simulations and numerous other physics problems including multi-scale simulations. Ongoing investigations will include using FUQ to correct predictions in solid mechanics with the plasticity model as the input constitutive law.
An important advantage of FUQ over uncertainty propagation in parameters is that it enables changes in the actual functional forms used in the simulation. In our example, we are not limited to Lennard-Jones potentials with different parameters but can change the shape of the potential (with the limitations discussed above).
Conclusions
===========
In summary we demonstrated the ability to calculate functional derivatives of a quantity of interest predicted by a non-trivial physics simulation with respect to its input laws. This information can be used to quantify the uncertainties originating in the simulation due to the use of the input function or to correct the prediction if a more accurate model becomes available. We developed a computationally efficient approach to compute the functional derivative in a MD simulations performed using the Lennard-Jones potential and shows that this information can be used to infer thermodynamic properties corresponding to various other potentials without re-running the simulation. The functional uncertainty quantification approach is quite generally applicable and we believe it will be useful to quantify uncertainties in a variety of materials models.
Thermodynamic quantities using coordination number
==================================================
The calculation of the functional derivative using the perturbative approach (Eq. \[eq:FD2\] and \[eq:perturbative\]) is expanded upon here, beginning with the contribution to the Hamiltonian from the perturbation. This quantity can be directly calculated with a sum over all pairs of atoms as the potential energy in any MD simulation: $$\label{eq:PEall}
H' = \sum_{i<j} \phi'(r_{ij})$$ This would require modification of the MD code to calculate this quantity within the force loop or summing over a saved atomic trajectory. This expression can be replaced: $$\label{eq:PEcoord}
H' = \frac{N}{2} \sum_{k} \phi'(r_k) \cdot c(r_k)$$ where $c(r)$ is the average coordination number at a given separation distance $r$, discretized into $k$ bins and $N$ the number of atoms. This is neither invasive to the code, nor requires significant storage or computation past the low-fidelity simulation.
With the potential energy as the QoI, Q is simply the sum of Eq. \[eq:PEcoord\] and the low-fidelity simulation potential energy. For the pressure we use the virial expression (without the small ideal gas contribution as it is already present in the low-fidelity pressure): $$\label{eq:Pallvec}
P' = \frac{1}{3V}\sum_{i<j} \textbf{f}_{ij} \textbf{r}_{ij}$$ Because we examine cases of two-body central forces this can be simplified: $$\label{eq:Pall}
P' = \frac{1}{3V}\sum_{i<j} f(r_{ij})r_{ij}$$ and with the same motivations as above, we rewrite in terms of the coordination number, again with $k$ bins in separation distance: $$\label{eq:Pcoord}
P' = \frac{N}{2} \frac{1}{3V}\sum_k f(r_k)\cdot r \cdot c(r_k)$$
This quantity, added to the low-fidelity simulation pressure, provides Q in Eq. \[eq:perturbative\] for the QoI pressure.
Correction results
==================
Tables \[tbl:PE\] and \[tbl:P\] show full results for all physical conditions for potential energy and pressure, respectively. Figure \[fig:results\] compares columns $\Delta Q^{Sim}$ and $\Delta Q^{FUQ}$. All results here are calculated with respect to the LJ low-fidelity potential, except the rows marked with (H). Those cases use the low-fidelity hybrid LJ-Morse potential.
Temp. Potential $Q_{LF}$ $Q_{HF}$ $Q_{LF} + \Delta Q^{FUQ}$ $\Delta Q^{Sim}$ $\Delta Q^{FUQ}$ $\Delta Q$ % Error
------- ----------- ---------- ---------- --------------------------- ------------------ ------------------ --------------------
(K) - -
300 Sine 1 -1.08 -1.28 -1.27 -0.197 -0.196 0.456
300 Sine 2 -1.08 -0.928 -0.927 0.151 0.152 0.784
300 Sine 3 -1.08 -1.03 -1.03 0.0518 0.0522 0.793
300 Sine 4 -1.08 -1.31 -1.31 -0.235 -0.235 0.0132
300 Sine 5 -1.08 -1.33 -1.34 -0.256 -0.259 0.988
300 Sine 6 -1.08 -1.41 -1.41 -0.330 -0.333 0.993
300 Sine 7 -1.08 -1.48 -1.49 -0.403 -0.407 0.995
300 Morse -1.08 -0.998 -1.02 0.0804 0.0628 21.9
300 Morse (H) -1.05 -0.998 -1.00 0.0552 0.0520 5.89
1300 Sine 1 -0.704 -0.850 -0.850 -0.146 -0.146 0.0158
1300 Sine 2 -0.704 -0.602 -0.601 0.102 0.103 1.23
1300 Sine 3 -0.704 -0.666 -0.666 0.0377 0.0379 0.588
1300 Sine 4 -0.704 -0.867 -0.868 -0.164 -0.164 0.160
1300 Sine 5 -0.704 -0.878 -0.876 -0.174 -0.172 1.08
1300 Sine 6 -0.704 -0.956 -0.925 -0.252 -0.221 12.1
1300 Sine 7 -0.704 -1.03 -0.974 -0.330 -0.271 18.0
1300 Morse -0.704 -0.655 -0.662 0.0485 0.0416 14.3
1300 Morse (H) -0.703 -0.673 -0.673 0.0295 0.0295 0.0295
5000 Sine 1 -0.398 -0.598 -0.598 -0.200 -0.200 0.136
5000 Sine 2 -0.398 -0.243 -0.241 0.155 0.157 0.896
5000 Sine 3 -0.398 -0.346 -0.345 0.0524 0.0527 0.477
5000 Sine 4 -0.398 -0.636 -0.637 -0.238 -0.239 0.441
5000 Sine 5 -0.398 -0.663 -0.665 -0.265 -0.267 0.932
5000 Sine 6 -0.398 -0.738 -0.741 -0.340 -0.343 0.959
5000 Sine 7 -0.398 -0.814 -0.818 -0.416 -0.420 1.01
5000 Morse -0.398 -0.321 -0.393 0.0774 0.00547 93.1
5000 Morse (H) -0.366 -0.320 -0.320 0.0454 0.0455 0.376
: Comparison of potential energy correction from functional derivatives and direct simulation.[]{data-label="tbl:PE"}
Temp. Potential $Q_{LF}$ $Q_{HF}$ $Q_{LF} + \Delta Q^{FUQ}$ $\Delta Q^{Sim}$ $\Delta Q^{FUQ}$ $\Delta Q$ % Error
------- ----------- ---------- ---------- --------------------------- ------------------ ------------------ --------------------
(K) - -
300 Sine 1 -0.116 -3.37 -3.25 -3.26 -3.13 3.89
300 Sine 2 -0.116 2.69 2.70 2.81 2.81 0.279
300 Sine 3 -0.116 0.514 0.538 0.630 0.654 3.75
300 Sine 4 -0.116 -4.12 -4.05 -4.01 -3.94 1.73
300 Sine 5 -0.116 -5.03 -5.07 -4.91 -4.95 0.763
300 Sine 6 -0.116 -6.43 -6.48 -6.32 -6.37 0.792
300 Sine 7 -0.116 -7.83 -7.90 -7.72 -7.78 0.805
300 Morse -0.116 1.43 -1.31 1.54 -1.19 177
300 Morse (H) 0.491 1.27 1.49 0.781 0.998 27.7
1300 Sine 1 0.129 -1.12 -1.13 -1.25 -1.26 1.44
1300 Sine 2 0.129 1.37 1.36 1.24 1.23 0.498
1300 Sine 3 0.129 0.563 0.562 0.434 0.433 0.219
1300 Sine 4 0.129 -1.58 -1.62 -1.71 -1.75 2.34
1300 Sine 5 0.129 -1.87 -2.03 -2.00 -2.16 7.70
1300 Sine 6 0.129 -1.75 -2.65 -1.88 -2.77 47.3
1300 Sine 7 0.129 -1.74 -3.26 -1.87 -3.391 81.5
1300 Morse 0.129 0.0733 -0.359 -0.0558 -0.488 774
1300 Morse (H) 0.124 0.469 0.463 0.344 0.338 1.68
5000 Sine 1 54.4 51.6 51.6 -2.77 -2.77 0.0187
5000 Sine 2 54.4 57.1 57.1 2.72 2.74 0.905
5000 Sine 3 54.4 55.1 55.1 0.7413 0.750 1.17
5000 Sine 4 54.4 50.6 50.6 -3.80 -3.81 0.161
5000 Sine 5 54.4 49.6 49.5 -4.81 -4.84 0.704
5000 Sine 6 54.4 48.2 48.2 -6.18 -6.23 0.845
5000 Sine 7 54.4 46.9 46.8 -7.53 -7.61 1.12
5000 Morse 54.4 39.8 36.8 -14.6 -17.5 20.1
5000 Morse (H) 39.0 39.8 39.8 0.796 0.794 0.279
: Comparison of pressure correction from functional derivatives and direct simulation.[]{data-label="tbl:P"}
Acknowledgments {#acknowledgments .unnumbered}
===============
This research was supported by the US Department of Energy (DOE), Office of Advanced Scientific Computing Research (ASCR) through the Exascale Co-Design Center for Materials in Extreme Environments (ExMatEx, exmatex.org). Computational resources of nanoHUB.org are gratefully acknowledged.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this work we present a mimetic spectral element discretization for the 2D incompressible Navier-Stokes equations that in the limit of vanishing dissipation exactly preserves mass, kinetic energy, enstrophy and total vorticity on unstructured grids. The essential ingredients to achieve this are: (i) a velocity-vorticity formulation in rotational form, (ii) a sequence of function spaces capable of exactly satisfying the divergence free nature of the velocity field, and (iii) a conserving time integrator. Proofs for the exact discrete conservation properties are presented together with numerical test cases on highly irregular grids.'
address:
- 'Eindhoven University of Technology, Department of Mechanical Engineering, P.O. Box 513, 5600 MB Eindhoven, The Netherlands'
- 'Delft University of Technology, Faculty of Aerospace Engineering, P.O. Box 5058, 2600 GB Delft, The Netherlands'
author:
- 'A. Palha'
- 'M. Gerritsma'
bibliography:
- './library\_clean.bib'
title: 'A mass, energy, enstrophy and vorticity conserving (MEEVC) mimetic spectral element discretization for the 2D incompressible Navier-Stokes equations'
---
energy conserving discretization ,mimetic discretization ,enstrophy conserving discretization ,spectral element method ,incompressible Navier-Stokes equations
Introduction
============
Relevance of structure preserving methods
-----------------------------------------
Structure-preserving discretizations are known for their robustness and accuracy. They conserve fundamental properties of the equations (mass, momentum, kinetic energy, etc). For example, it is well known that the application of conventional discretization techniques to inviscid flows generates artificial energy dissipation that pollutes the energy spectrum. For these reasons, structure-preserving discretizations have recently gained popularity.
For a long time, in the development of general circulation models used in weather forecast, it has been noticed that care must be taken in the construction of the discretization of physical laws. Phillips [@Phillips1959] verified that the long-time integration of non-linear convection terms resulted in the breakdown of numerical simulations independently of the time step, due to the amplification of weak instabilities. Later, Arakawa [@Arakawa1966] proved that such instabilities can be avoided if the integral of the square of the advected quantity is conserved (kinetic energy, enstrophy in 2D, for example). Staggered finite difference discretizations that avoid these instabilities have been introduced both by Harlow and Welch [@Harlow1965] and Arakawa and his collaborators [@Arakawa1977; @Mesinger1976]. Lilly [@Lilly1965] showed that these discretizations could conserve momentum, energy and circulation. At the same time, Piacseck and Williams [@Piacsek1970] connected the conservation of energy with the preservation of skew-symmetry of the convection operator at the discrete level. These ideas of staggering the discrete physical quantities and of preserving the skew-symmetry of the convection operator have been successfully explored by several authors aiming to construct more robust and accurate numerical methods.
In this work we focus on the incompressible Navier-Stokes equations, particularly convection-dominated flow problems (e.g. wind turbine wake aerodynamics). It is widely known that in the absence of external forces and viscosity these equations contain important symmetries or invariants, e.g. [@Arnold1966; @Arnold1992; @Majda2001; @Foias2001], such as conservation of kinetic energy. A straightforward discretization using standard methods does not guarantee the conservation of these invariants. Discrete energy conservation is important from a physical point of view, especially for turbulent flow simulations when Direct Numerical Simulation (DNS) or Large Eddy Simulation (LES) are used. In these cases, the accurate reproduction of the energy spectrum is essential to generate the associated energy cascade. The numerical diffusion inherent (and sometimes essential) to many discretizations can dominate both the molecular diffusion contribution (DNS) and the sub-grid model contribution (LES). This negatively affects the energy spectrum and consequently the energy cascade. Energy-conserving discretizations ensure that all diffusion is modelled and not a product of discretization errors. For this reason, many authors have shown that energy-conserving schemes are essential both for DNS (e.g. [@Perot1993; @Verstappen1995; @Le1997; @Verstappen1998; @Verstappen2003]) and LES simulations (e.g. [@Mahesh2004; @Mittal1997; @Benhamadouche2002; @Nagarajan2003; @Felten2006; @Ham2007]). In addition, discrete kinetic energy conservation provides a non-linear stability bound to the solution (e.g. [@Sadourny1975; @Sanderse2013]). This bound relaxes the characteristic stability constraints of standard methods, allowing the choice of mesh and time step size to be based only on accuracy requirements. This is particularly relevant for LES where mesh sizes have to be kept as large as possible, due to the still significant computational effort required for this approach. Energy-conserving methods generate well-behaved global errors and adequate physical behaviour, even on coarse meshes.
In two-dimensions, enstrophy is another conserved quantity of the flow when external forces and viscosity are not present. The classical Arakawa scheme [@Arakawa1966] and derived schemes exploit both the conservation of energy and enstrophy. By doing so, these methods have far better long time simulation properties than competing methods that do not conserve these two quantities.
Overview of structure preserving methods
----------------------------------------
Two of the most used approaches to construct energy preserving discretizations are: (i) staggering and (ii) skew-symmetrization of the convective operator.
Most staggered grid methods date back to the pioneering work of Harlow and Welch [@Harlow1965], and Arakawa and colleagues [@Arakawa1977; @Mesinger1976]. These methods employ a discretization that distributes the different physical quantities (pressure, velocity, vorticity, etc) at different locations in the mesh (vertices, faces, cell centres). It can be shown that, by doing so, important conservation properties can be maintained. Due to its success, much attention has been given to this approach and several extensions to the original work have been made by several authors. Morinishi [@Morinishi1998] has developed a high-order version on Cartesian grids. Extensions to non-uniform meshes have been presented for example in [@Wesseling1999] for structured quadrilateral meshes and by several authors [@Apanovich1988; @Choudhury1990; @Perot2000; @Mullen2009] for simplicial meshes.
The advantages of exploiting the skew-symmetry of the convection operator at the discrete level, were first identified by Piacseck [@Piacsek1970]. Following his ideas, Verstappen and Veldman [@Verstappen; @Verstappen1998; @Verstappen2003] and later Knicker [@Knikker2009] constructed high order discretizations on Cartesian grids with substantially improved properties. Kok [@Kok2009] presents another formulation valid on curvilinear grids. The extension to simplicial grids has been reported for example in the works of Vasilyev [@Vasilyev2000], Veldman [@Veldman2008] and van’t Hof [@VanHof2012].
At another level, Perot [@perot43discrete] suggested that a connection exists between the skew-self-adjoint formulation presented by Tadmor [@Tadmor1984] and box schemes. Box schemes or Keller box schemes are a class of numerical methods originally introduced by Wendroff [@wendroff1960] for hyperbolic problems and later popularized by Keller [@keller1971; @keller1978] for parabolic problems. This method is face-based and space and time are coupled by introducing a space-time control volume. This method is known to be physically accurate and successful application has been reported in [@Croisille2002; @Croisille2005; @Gustafsson2006; @Ranjan2013]. More recently, this method has been shown to be multisymplectic by Ascher [@Ascher2005] and Frank [@Frank2006]. Perot, [@Perot2007], also established a relation between box schemes and discrete calculus and generalized it to arbitrary meshes.
Although most of the literature on the simulation of incompressible Navier-Stokes flows has been performed using finite difference and finite volume methods, finite element methods have also been actively used and shared many of the developments already discussed for finite differences and finite volumes. One of the initial challenges in using finite elements for flow simulations lies in the fact that specific finite element subspaces must be used to discretize the different physical quantities. The finite element subspaces must satisfy the Ladyzhenskaya-Babuska-Brezzi (LBB) (or inf-sup) compatibility condition (see [@brezzi1991mixed]), otherwise instability ensues. Several families of suitable finite elements have been proposed in the literature, the most common being the Taylor-Hood family (Taylor and Hood [@TaylorHood1973]) and the Crouzeix-Raviart family (Crouzeix and Raviart [@Crouzeix]). The underlying idea behind these different finite element families is to use different polynomial orders for velocity and pressure. In essence, this is intimately related to the staggering approach already discussed in the context of finite differences and finite volumes. Staggering is explicitly mentioned in some finite element work, for example [@KoprivaKolias; @Liu2007; @Liu2008; @Chung2012; @Tavelli2015]. Another important aspect when using finite elements is the weak formulation used for the convection term. Some of these forms (rotational and skew-symmetric) can retain important symmetries of the original system of equations, this has been reported for example in [@Guevremont1990; @Blaisdell1996; @Ronquist1996; @Wilhelm2000]. Ensuring conservation in finite element discretizations is not straightforward, nevertheless examples in the literature exist, e.g. [@Liu2000; @Bernsen2006]. Also within the context of finite elements, Rebholz and co-authors, e.g. [@Rebholz2007; @Olshanskii2010], have developed several velocity-vorticity discretizations for the 3D Navier-Stokes equations. These discretizations are capable of conserving both energy and helicity.
One final aspect when discussing existing structure preserving methods is time integration. Many of the conserving methods reported in the literature present proofs regarding spatial discretization. Nevertheless, when discrete time evolution is taken into account many time integrators destroy the nice properties of the spatial discretization. The relevance of time integration for the construction of structure preserving discretizations has been analyzed by Sanderse [@Sanderse2013].
Overview of mimetic discretizations
-----------------------------------
Over the years numerical analysts have developed numerical schemes which preserve some of the structure of the differential models they aim to approximate, so in that respect the whole structure preserving idea is not new. One of the contributions of mimetic methods is to identify the proper language in which to encode these structures/symmetries is the language of differential geometry. Another novel aspect of mimetic discretizations is the identification of the metric-free part of differential models, which can (and should) be conveniently described in terms of algebraic topology. A general introduction and overview to spatial and temporal mimetic/geometric methods can be found in [@Christiansen2011; @perot43discrete; @Budd2003; @Hairer2006].
The relation between differential geometry and algebraic topology in physical theories was first established by Tonti [@tonti1975formal]. Around the same time Dodziuk [@Dodziuk76] set up a finite difference framework for harmonic functions based on Hodge theory. Both Tonti and Dodziuk introduce differential forms and cochain spaces as the building blocks for their theory. The relation between differential forms and cochains is established by the Whitney map ($k$-cochains $\rightarrow$ $k$-forms) and the de Rham map ($k$-forms $\rightarrow$ $k$-cochains). The interpolation of cochains to differential forms on a triangular grid was already established by Whitney, [@Whitney57]. These interpolatory forms are now known as the [*Whitney forms*]{}.
Hyman and Scovel [@HymanScovel90] set up the discrete framework in terms of cochains, which are the natural building blocks of finite volume methods. Later Bochev and Hyman [@bochev2006principles] extended this work and derived discrete operators such as the discrete wedge product, the discrete codifferential, the discrete inner product, etc.
In a finite difference/volume context Robidoux, Hyman, Steinberg and Shashkov, [@HymanShashkovSteinberg97; @HymanShashkovSteinberg2002; @HYmanSteinberg2004; @RobidouxAdjointGradients1996; @RobidouxThesis; @bookShashkov; @Steinberg1996; @SteibergZingano2009] used symmetry considerations to discretize diffusion problems on rough grids and with non-smooth anisotropic diffusion coefficients. In a more recent paper by Robidoux and Steinberg [@Robidoux2011] a discrete vector calculus in a finite difference setting is presented. Here the differential operators grad, curl and div are exactly represented at the discrete level and the numerical approximations are all contained in the constitutive relations, which are already polluted by modeling and experimental error. For mimetic finite differences, see also Brezzi et al. [@BrezziBuffaLipnikov2009; @brezzi2010].
The application of mimetic ideas to unstructured staggered grids has been extensively studied by Perot, [@Perot2000; @ZhangSchmidtPerot2002; @perot2006mimetic; @PerotSubramanian2007a; @PerotSubramanian2007]. Especially in [@perot43discrete] where the rationale of preserving symmetries in numerical algorithms is lucidly described. The most *geometric approach* is described in the work by Desbrun et al. [@desbrun2005discrete; @ElcottTongetal2007; @MullenCraneetal2009; @PavlovMullenetal2010] and the thesis by Hirani [@Hirani_phd_2003].
The *Japanese papers* by Bossavit, [@bossavit_japan_computational_1; @bossavit_japan_computational_2; @bossavit_japan_computational_3; @bossavit_japan_computational_4; @bossavit_japan_computational_5], serve as an excellent introduction and motivation for the use of differential forms in the description of physics and the use in numerical modeling. The field of application is electromagnetism, but these papers are sufficiently general to extend to other physical theories.
In a series of papers by Arnold, Falk and Winther, [@arnold:Quads; @arnold2006finite; @arnold2010finite], a finite element exterior calculus framework is developed. Higher order methods are also described by Rapetti [@Rapetti2007; @Rapetti2009] and Hiptmair [@hiptmair2001]. Possible extensions to spectral methods were described by Robidoux, [@robidoux-polynomial]. A different approach for constructing arbitrary order mimetic finite elements has been proposed by the authors [@Palha2014; @gerritsmaicosahom2012; @Rebelo2014; @palhaAdvectionIcosahom2014; @kreeft::stokes; @bouman::icosahom2009; @palha::icosahom2009].
Extensions of these ideas to polyhedral meshes have been proposed by Ern, Bonelle and co-authors in [@Bonelle2015; @Bonelle2014; @Bonelle2015a; @Bonelle2016] and by Brezzi and co-authors in [@BeiraodaVeiga2014; @Brezzi2014; @BeiraodaVeiga2016; @DaVeiga2015]. These approaches provide more geometrical flexibility while maintaining fundamental structure preserving properties.
Mimetic isogeometric discretizations have been introduced by Buffa et al. [@BuffaDeFalcoSangalli2011], Evans and Hughes [@Evans2013a], and Hiemstra et al. [@Hiemstra2014].
Another approach develops a discretization of the physical field laws based on a discrete variational principle for the discrete Lagrangian action. This approach has been used in the past to construct variational integrators for Lagrangian systems, e.g. [@Kouranbaeva2000; @Marsden2003]. Recently, Kraus and Maj [@Kraus2015] have used the method of formal Lagrangians to derive generalized Lagrangians for non-Lagrangian systems of equations. This allows to apply variational techniques to construct structure preserving discretizations on a much wider range of systems.
Outline of paper
----------------
In this work we present a new exact mass, energy, enstrophy and vorticity conserving (MEEVC) mimetic spectral element solver for the 2D incompressible Navier-Stokes equations. The essential ingredients to achieve this are: (i) a velocity-vorticity formulation in rotational form, (ii) a sequence of function spaces capable of exactly satisfying the divergence free nature of the velocity field, and (iii) a conserving time integrator. This results in a set of two decoupled equations: one for the evolution of velocity and another one for the evolution of vorticity.
In [Section \[sec::spatial\_discretization\]]{} we present the spatial discretization. We start by introducting the $({\vec{u}},{\omega})$ formulation based in the rotational form in [Section \[subsec::the\_v\_omega\_formulation\_in\_the\_rotational\_form\]]{} and then in [Section \[subsec::finite\_element\_discretization\]]{} the finite element discretization is discussed. This is followed by the temporal discretization in [Section \[sec::temporal\_discretization\]]{}. Once we have introduced the numerical discretization its conservation properties are proved in [Section \[sec::conservation\_properties\_and\_time\_reversibility\]]{}. In [Section \[sec::numerical\_test\_cases\]]{} the method is applied and tested on two test cases. We start by testing the accuracy of the method with a Taylor-Green vortex, for which the analytical solution is known. We finalize the test cases with an inviscid shear layer roll-up test to illustrate the conservation properties of the method. In [Section \[sec::conclusions\]]{} we conclude with a discussion of the merits and limitations of this solver and future extensions.
Spatial discretization {#sec::spatial_discretization}
======================
The $({\vec{u}},{\omega})$ formulation in rotational form {#subsec::the_v_omega_formulation_in_the_rotational_form}
---------------------------------------------------------
The evolution of 2D viscous incompressible flows is governed by the Navier-Stokes equations, which are most commonly expressed as a set of conservation laws for momentum and mass involving the velocity ${\vec{u}}$ and pressure $p$: $$\begin{dcases}
\frac{\partial{\vec{u}}}{\partial t} + \left({\vec{u}}\cdot\nabla\right){\vec{u}}+ \nabla p = \nu\Delta{\vec{u}}+ \vec{s}, \\
\nabla\cdot{\vec{u}}= 0,
\end{dcases} \label{eq::ns_convective_form}$$ with $\nu$ the kinematic viscosity, $\vec{s}$ the body force per unit mass, and $\Delta = \nabla\cdot\nabla$ the Laplace operator. These equations are valid on the fluid domain $\Omega$, together with suitable initial and boundary conditions. In this work we consider only periodic boundary conditions and we set $\vec{s}=0$.
The form of the Navier-Stokes equations presented in is the so called *convective form*. Its name stems from the particular form of the nonlinear term $ \left({\vec{u}}\cdot\nabla\right){\vec{u}}$, which underlines its convective nature. This form is not unique. Using well known vector calculus identities it is possible to rewrite the nonlinear term in three other forms, see for example Zang [@Zang1991], Morinishi [@Morinishi1998], and R[ø]{}nquist [@Ronquist1996].
The first alternative, called *divergence form* or *conservative form*, expresses the nonlinear term as a divergence: $$\begin{dcases}
\frac{\partial{\vec{u}}}{\partial t} + \nabla\cdot\left({\vec{u}}\otimes{\vec{u}}\right)+ \nabla p = \nu\Delta{\vec{u}}, \\
\nabla\cdot{\vec{u}}= 0.
\end{dcases} \label{eq::ns_divergence_form}$$
The second alternative is obtained as a linear combination of the convective and divergence forms, producing a skew-symmetric nonlinear term, thus its name *skew-symmetric form*: $$\begin{dcases}
\frac{\partial{\vec{u}}}{\partial t} + \frac{1}{2} \left({\vec{u}}\cdot\nabla\right){\vec{u}}+ \frac{1}{2}\nabla\cdot\left({\vec{u}}\otimes{\vec{u}}\right)+ \nabla p = \nu\Delta{\vec{u}}, \\
\nabla\cdot{\vec{u}}= 0.
\end{dcases} \label{eq::ns_skew_symmetric_form}$$
The third and final alternative we consider here, the *rotational form*, makes use of the vorticity ${\omega}:=\nabla\times{\vec{u}}$: $$\begin{dcases}
\frac{\partial{\vec{u}}}{\partial t} + \omega\times{\vec{u}}+ \nabla {\bar{p}}= \nu\Delta{\vec{u}}, \\
\nabla\cdot{\vec{u}}= 0,
\end{dcases} \label{eq::ns_rotational_form}$$ where the static pressure $p$ has been replaced by the total pressure ${\bar{p}}:= \frac{1}{2}{\vec{u}}\cdot{\vec{u}}+ p$.
A different, but equivalent, approach used to describe fluid flow problems resorts to expressing the momentum equation in terms of the vorticity ${\omega}$. By taking the curl of the momentum equation and using the kinematic definition ${\omega}:= \nabla\times{\vec{u}}$ we can obtain the flow equations based on vorticity transport: $$\begin{dcases}
\frac{\partial{\omega}}{\partial t} + \frac{1}{2}\left({\vec{u}}\cdot\nabla\right){\omega}+ \frac{1}{2}\nabla\cdot\left({\vec{u}}\,{\omega}\right) = \nu\Delta{\omega}, \\
\nabla\cdot{\vec{u}}= 0, \\
{\omega}= \nabla\times{\vec{u}}\,.
\end{dcases} \label{eq::ns_vorticity_transport}$$ This velocity-vorticity $({\vec{u}},{\omega})$ formulation of the Navier-Stokes equations is of particular interest for vortex dominated flows, see for example Gatski [@Gatski1991] for an overview and Daube [@Daube1992] and Clercx [@Clercx1997a] for applications.
All these different ways of expressing the governing equations of fluid flow are equivalent at the continuous level. As stated before, one can start with and derive all other sets of equations simply by using well known vector calculus identities and definitions. The often overlooked aspect is that each of these formulations leads to a different discretization, with its own set of properties, e.g. [@Zang1991; @Gatski1991; @Ronquist1996; @Morinishi1998]. Only in the limit of vanishing mesh size ($h \rightarrow 0$) are the discretizations expected to be equivalent. This introduces an additional degree of freedom associated to the choice of formulation that, combined with the discretization approach to use, will determine the final properties of the method.
In this work we construct a $({\vec{u}},{\omega})$ formulation by combining the rotational form with the vorticity transport equation , similar to the work of Benzi et al. [@Benzi2012] and Lee et al. [@Lee2011]: $$\begin{dcases}
\frac{\partial{\vec{u}}}{\partial t} + \omega\times{\vec{u}}+ \nabla {\bar{p}}= -\nu\nabla\times{\omega}, \\
\frac{\partial{\omega}}{\partial t} + \frac{1}{2}\left({\vec{u}}\cdot\nabla\right){\omega}+ \frac{1}{2}\nabla\cdot\left({\vec{u}}\,{\omega}\right) = \nu\Delta{\omega}, \\
\nabla\cdot{\vec{u}}= 0\,,
\end{dcases} \label{eq::ns_meevc_form}$$ where we use the vector calculus identity $\Delta{\vec{u}}= \nabla\left(\nabla\cdot{\vec{u}}\right) - \nabla\times\nabla\times{\vec{u}}$ to derive the equality $\Delta{\vec{u}}= -\nabla\times{\omega}$. An important aspect we wish to stress at this point is that although at the continuous level the kinematic definition ${\omega}:= \nabla\times{\vec{u}}$ is valid, at the discrete level it is not always guaranteed that this identity holds. In fact, in the discretization presented here this identity is satisfied only approximately. This, as will be seen, enables the construction of a mass, energy, enstrophy and vorticity conserving discretization.
Finite element discretization {#subsec::finite_element_discretization}
-----------------------------
In this work we set out to construct a finite element discretization for the Navier-Stokes equations as given by . In particular we use a mixed finite element formulation, for more details on the mixed finite elements see for example [@brezzi1991mixed]. The first step for developing this discretization is the construction of the weak form of : $$\begin{dcases}
\text{Find } {\vec{u}}\in {H(\mathrm{div},\Omega)}, {\bar{p}}\in {L^{2}(\Omega)}\text{ and } {\omega}\in {H(\mathrm{curl},\Omega)}\text{ such that:}\\
\langle\frac{\partial{\vec{u}}}{\partial t},{\vec{v}}\rangle_{\Omega} + \langle\omega\times{\vec{u}},{\vec{v}}\rangle_{\Omega} - \langle {\bar{p}},\nabla\cdot{\vec{v}}\rangle_{\Omega} = -\nu\langle\nabla\times{\omega},{\vec{v}}\rangle_{\Omega},\quad \forall {\vec{v}}\in {H(\mathrm{div},\Omega)}, \\
\langle\frac{\partial{\omega}}{\partial t},{\xi}\rangle_{\Omega} - \frac{1}{2}\langle{\omega},\nabla\cdot\left({\vec{u}}\,{\xi}\right)\rangle_{\Omega} + \frac{1}{2}\langle\nabla\cdot\left({\vec{u}}\,{\omega},\right),{\xi}\rangle_{\Omega} = \nu\langle\nabla\times{\omega},\nabla\times{\xi}\rangle_{\Omega}, \quad\forall {\xi}\in {H(\mathrm{curl},\Omega)}, \\
\langle\nabla\cdot{\vec{u}},{q}\rangle_{\Omega} = 0, \quad\forall {q}\in {L^{2}(\Omega)}\,,
\end{dcases} \label{eq::ns_meevc_weak_form_continuous}$$ where we have used integration by parts and the periodic boundary conditions to obtain the identities $\langle {\bar{p}},\nabla\cdot{\vec{v}}\rangle_{\Omega} = - \langle \nabla{\bar{p}},{\vec{v}}\rangle_{\Omega}$, $\langle{\omega},\nabla\cdot\left({\vec{u}}\,{\xi}\right)\rangle_{\Omega} = -\langle\left({\vec{u}}\cdot\nabla\right){\omega},{\xi}\rangle_{\Omega}$ and $\langle\nabla\times{\omega},\nabla\times{\xi}\rangle_{\Omega} = \langle\Delta{\omega},{\xi}\rangle_{\Omega}$. The space ${L^{2}(\Omega)}$ corresponds to square integrable functions and the spaces ${H(\mathrm{div},\Omega)}$ and ${H(\mathrm{curl},\Omega)}$ contain square integrable functions whose divergence and curl are also square integrable. The second step is the definition of conforming finite dimensional function spaces, where we will seek our discrete solutions for velocity ${\vec{u}_{h}}$, pressure ${\bar{p}_{h}}$ and vorticity ${\omega_{h}}$: $${\vec{u}_{h}}\in {U_{h}}\subset {H(\mathrm{div},\Omega)}, \quad {\bar{p}_{h}}\in {Q_{h}}\subset {L^{2}(\Omega)}\quad \mathrm{and} \quad {\omega_{h}}\in {W_{h}}\subset {H(\mathrm{curl},\Omega)}.$$ As usual, each of these finite dimensional function spaces, ${U_{h}}$, ${Q_{h}}$ and ${W_{h}}$, has an associated finite set of basis functions, ${\vec{\epsilon}^{\,U}_{i}}$, ${\epsilon^{Q}_{i}}$, ${\epsilon^{W}_{i}}$, such that $${U_{h}}= \mathrm{span}\{{\vec{\epsilon}^{\,U}_{1}}, \dots,{\vec{\epsilon}^{\,U}_{{d_{U}}}}\}, \quad {Q_{h}}= \mathrm{span}\{{\epsilon^{Q}_{1}}, \dots,{\epsilon^{Q}_{{d_{Q}}}}\}\quad\mathrm{and}\quad{W_{h}}= \mathrm{span}\{{\epsilon^{W}_{1}}, \dots,{\epsilon^{W}_{{d_{W}}}}\},$$ where ${d_{U}}$, ${d_{Q}}$ and ${d_{W}}$ denote the dimension of the discrete function spaces and therefore correspond to the number of degrees of freedom for each of the unknowns. As a consequence, the approximate solutions for velocity, pressure and vorticity can be expressed as a linear combination of these basis functions $${\vec{u}_{h}}:= \sum_{i=1}^{{d_{U}}}{u_{i}}\,{\vec{\epsilon}^{\,U}_{i}}, \quad {\bar{p}_{h}}:= \sum_{i=1}^{{d_{Q}}}{p_{i}}\,{\epsilon^{Q}_{i}} \quad\mathrm{and}\quad {\omega_{h}}:= \sum_{i=1}^{{d_{W}}}{\omega_{i}}\,{\epsilon^{W}_{i}}, \label{eq:basis_expansion}$$ with ${u_{i}}$, ${p_{i}}$ and ${\omega_{i}}$ the degrees of freedom of velocity, total pressure and vorticity, respectively. Since the Navier-Stokes equations form a time dependent set of equations, in general these coefficients will be time dependent, ${u_{i}} = {u_{i}}(t)$, ${p_{i}}={p_{i}}(t)$ and ${\omega_{i}}={\omega_{i}}(t)$.
The choice of the finite dimensional function spaces dictates the properties of the discretization. In order to have exact conservation of mass we must exactly satisfy the divergence free constraint at the discrete level. A sufficient condition that guarantees divergence-free velocities, e.g. [@arnold2010finite; @Cockburn2006; @Buffa2011], is: $$\left\{\nabla\cdot{\vec{u}_{h}}\,|\, {\vec{u}_{h}}\in{U_{h}}\right\} \subseteq {Q_{h}}\,. \label{eq:divu_subspace_q}$$ In other words, the divergence operator must map ${U_{h}}$ into ${Q_{h}}$: $${U_{h}}\stackrel{\nabla\cdot}{\longrightarrow}{Q_{h}}\,. \label{eq:div_subcomplex}$$ If we set $${U_{h}}= {\mathrm{RT}_{N}}\quad \mathrm{and} \quad {Q_{h}}={\mathrm{DG}_{N-1}}\,,$$ where ${\mathrm{RT}_{N}}$ are the Raviart-Thomas elements of degree $N$, see [@RaviartThomas1977; @kirby2012], and ${\mathrm{DG}_{N-1}}$ are the discontinuous Lagrange elements of degree $(N-1)$, see [@kirby2012], we satisfy , see for example [@arnold2010finite] for a proof. Additionally, this pair of finite elements satisfies the LBB stability condition, see for example [@RaviartThomas1977; @arnold2010finite].
What remains to define is the finite element space associated to the vorticity, ${W_{h}}$. Since we wish to exactly represent the diffusion term in the momentum equation $\langle\nabla\times{\omega},{\vec{v}_{h}}\rangle_{\Omega}$ the space ${W_{h}}$ must satisfy a relation similar to $$\left\{\nabla\times{\omega_{h}}\,|\, {\omega_{h}}\in{W_{h}}\right\} \subseteq {U_{h}}\,. \label{eq:curlw_subspace_u}$$ In other words, the curl operator must map ${W_{h}}$ into ${U_{h}}$: $${W_{h}}\stackrel{\nabla\times}{\longrightarrow}{U_{h}}\,. \label{eq:curl_subcomplex}$$ The Lagrange elements of degree $N$, denoted by ${\mathrm{CG}_{N}}$ in [@kirby2012], satisfy , see [@arnold2010finite], therefore we set $${W_{h}}= {\mathrm{CG}_{N}}\,.$$
Together, the combination of these three finite element spaces forms a Hilbert subcomplex $$0 \longrightarrow {W_{h}}\stackrel{\nabla\times}{\longrightarrow}{U_{h}}\stackrel{\nabla\cdot}{\longrightarrow}{Q_{h}}\longrightarrow 0\,,$$ that mimics the 2D Hilbert complex associated to the continuous functional spaces: $$0 \longrightarrow {H(\mathrm{curl},\Omega)}\stackrel{\nabla\times}{\longrightarrow}{H(\mathrm{div},\Omega)}\stackrel{\nabla\cdot}{\longrightarrow}{L^{2}(\Omega)}\longrightarrow 0\,.$$ The Hilbert complex is an important structure that is intimately related to the de Rham complex of differential forms. The construction of a discrete subcomplex is an important requirement to obtain stable and accurate finite element discretizations, see for example for a detailed discussion.
The finite element spatial discretization used in this work to construct an approximate solution of is then obtained by the following weak formulation $$\begin{dcases}
\text{Find } {\vec{u}_{h}}\in {\mathrm{RT}_{N}}, {\bar{p}_{h}}\in {\mathrm{DG}_{N-1}}\text{ and } {\omega_{h}}\in {\mathrm{CG}_{N}}\text{ such that:}\\
\langle\frac{\partial{\vec{u}_{h}}}{\partial t},{\vec{v}_{h}}\rangle_{\Omega} + \langle{\omega_{h}}\times{\vec{u}_{h}},{\vec{v}_{h}}\rangle_{\Omega} - \langle {\bar{p}_{h}},\nabla\cdot{\vec{v}_{h}}\rangle_{\Omega} = -\nu\langle\nabla\times{\omega_{h}},{\vec{v}_{h}}\rangle_{\Omega},\quad \forall {\vec{v}_{h}}\in {\mathrm{RT}_{N}}, \\
\langle\frac{\partial{\omega_{h}}}{\partial t},{\xi_{h}}\rangle_{\Omega} - \frac{1}{2}\langle{\omega_{h}},\nabla\cdot\left({\vec{u}_{h}}\,{\xi_{h}}\right)\rangle_{\Omega} + \frac{1}{2}\langle\nabla\cdot\left({\vec{u}_{h}}\,{\omega_{h}}\right),{\xi_{h}}\rangle_{\Omega} = \nu\langle\nabla\times{\omega_{h}},\nabla\times{\xi_{h}}\rangle_{\Omega}, \quad\forall {\xi_{h}}\in {\mathrm{CG}_{N}}, \\
\langle\nabla\cdot{\vec{u}_{h}},{q_{h}}\rangle_{\Omega} = 0, \quad\forall {q_{h}}\in {\mathrm{DG}_{N-1}}\,.
\end{dcases} \label{eq::ns_meevc_weak_form_discrete}$$
Using the expansions for ${\vec{u}_{h}}$, ${\bar{p}_{h}}$ and ${\omega_{h}}$ in , can be rewritten as $$\begin{dcases}
\text{Find } {\boldsymbol{u}}\in \mathbb{R}^{{d_{U}}}, {\boldsymbol{\bar{p}}}\in \mathbb{R}^{{d_{Q}}} \text{ and } {\boldsymbol{\omega}}\in \mathbb{R}^{{d_{W}}} \text{ such that:}\\
\sum_{i=1}^{{d_{U}}}\frac{\mathrm{d}{u_{i}}}{\mathrm{d}t}\langle{\vec{\epsilon}^{\,U}_{i}},{\vec{\epsilon}^{\,U}_{j}}\rangle_{\Omega} + \sum_{i=1}^{{d_{U}}}{u_{i}}\langle{\omega_{h}}\times{\vec{\epsilon}^{\,U}_{i}},{\vec{\epsilon}^{\,U}_{j}}\rangle_{\Omega} - \sum_{k=1}^{{d_{Q}}}{p_{k}}\langle {\epsilon^{Q}_{k}},\nabla\cdot{\vec{\epsilon}^{\,U}_{j}}\rangle_{\Omega} = -\nu\langle\nabla\times{\omega_{h}},{\vec{\epsilon}^{\,U}_{j}}\rangle_{\Omega},\quad j=1,\dots,{d_{U}}, \\
\sum_{i=1}^{{d_{W}}}\frac{\mathrm{d}{\omega_{i}}}{\mathrm{d}t}\langle{\epsilon^{W}_{i}},{\epsilon^{W}_{j}}\rangle_{\Omega} - \sum_{i=1}^{{d_{W}}}\frac{{\omega_{i}}}{2}\langle{\epsilon^{W}_{i}},\nabla\cdot\left({\vec{u}_{h}}\,{\epsilon^{W}_{j}}\right)\rangle_{\Omega} + \sum_{i=1}^{{d_{W}}}\frac{{\omega_{i}}}{2}\langle\nabla\cdot\left({\vec{u}_{h}}\,{\epsilon^{W}_{i}}\right),{\epsilon^{W}_{j}}\rangle_{\Omega} = \nu\sum_{i=1}^{{d_{W}}}{\omega_{i}}\langle\nabla\times{\epsilon^{W}_{i}},\nabla\times{\epsilon^{W}_{j}}\rangle_{\Omega}, \quad j=1,\dots,{d_{W}}, \\
\sum_{i=1}^{{d_{U}}}{u_{i}}\langle\nabla\cdot{\vec{\epsilon}^{\,U}_{i}},{\epsilon^{Q}_{j}}\rangle_{\Omega} = 0, \quad j = 1,\dots,{d_{Q}}\,,
\end{dcases} \label{eq::ns_meevc_weak_form_discrete_expansion}$$ with ${\boldsymbol{u}}:= [{u_{1}},\dots,{u_{{d_{U}}}}]^{\top}$, ${\boldsymbol{\bar{p}}}:= [{p_{1}},\dots,{p_{{d_{Q}}}}]^{\top}$ and ${\boldsymbol{\omega}}:= [{\omega_{1}},\dots,{\omega_{{d_{W}}}}]^{\top}$. Using matrix notation, can be expressed more compactly as $$\begin{dcases}
\text{Find } {\boldsymbol{u}}\in \mathbb{R}^{{d_{U}}}, {\boldsymbol{\bar{p}}}\in \mathbb{R}^{{d_{Q}}} \text{ and } {\boldsymbol{\omega}}\in \mathbb{R}^{{d_{W}}} \text{ such that:}\\
{\boldsymbol{\mathsf{M}}} \frac{\mathrm{d}{\boldsymbol{u}}}{\mathrm{d}t} + {\boldsymbol{\mathsf{R}}}\,{\boldsymbol{u}}- {\boldsymbol{\mathsf{P}}}\,{\boldsymbol{\bar{p}}}= -\nu\,\boldsymbol{l}, \\
{\boldsymbol{\mathsf{N}}}\frac{\mathrm{d}{\boldsymbol{\omega}}}{\mathrm{d}t} - \frac{1}{2}{\boldsymbol{\mathsf{W}}}\,{\boldsymbol{\omega}}+ \frac{1}{2}{\boldsymbol{\mathsf{W}}}^{\top}{\boldsymbol{\omega}}= \nu\,{\boldsymbol{\mathsf{L}}}\,{\boldsymbol{\omega}}, \\
{\boldsymbol{\mathsf{D}}}\,{\boldsymbol{u}}= 0,
\end{dcases} \label{eq::ns_meevc_weak_form_discrete_matrix_notation}$$ The coefficients of the matrices ${\boldsymbol{\mathsf{M}}}$, ${\boldsymbol{\mathsf{R}}}$ and ${\boldsymbol{\mathsf{P}}}$, and the column vector $\boldsymbol{l}$ are given by $${\mathsf{M}}_{ij} := \langle{\vec{\epsilon}^{\,U}_{j}},{\vec{\epsilon}^{\,U}_{i}}\rangle_{\Omega}, \quad {\mathsf{R}}_{ij} := \langle{\omega_{h}}\times{\vec{\epsilon}^{\,U}_{j}},{\vec{\epsilon}^{\,U}_{i}}\rangle_{\Omega}, \quad {\mathsf{P}}_{ij} := \langle {\epsilon^{Q}_{j}},\nabla\cdot{\vec{\epsilon}^{\,U}_{i}}\rangle_{\Omega}\quad\mathrm{and}\quad l_{i} := \langle\nabla\times{\omega_{h}},{\vec{\epsilon}^{\,U}_{i}}\rangle_{\Omega}. \label{eq:matrix_coefficients_1}$$ Similarly, the coefficients of the matrices ${\boldsymbol{\mathsf{N}}}$, ${\boldsymbol{\mathsf{W}}}$, ${\boldsymbol{\mathsf{L}}}$ and ${\boldsymbol{\mathsf{D}}}$ are given by $${\mathsf{N}}_{ij} := \langle{\epsilon^{W}_{j}},{\epsilon^{W}_{i}}\rangle_{\Omega}, \quad {\mathsf{W}}_{ij} := \langle{\epsilon^{W}_{j}},\nabla\cdot\left({\vec{u}_{h}}\,{\epsilon^{W}_{i}}\right)\rangle_{\Omega}, \quad {\boldsymbol{\mathsf{L}}}_{ij} := \langle\nabla\times{\epsilon^{W}_{j}},\nabla\times{\epsilon^{W}_{i}}\rangle_{\Omega} \quad \mathrm{and}\quad {\boldsymbol{\mathsf{D}}}_{ij} :=\langle\nabla\cdot{\vec{\epsilon}^{\,U}_{j}},{\epsilon^{Q}_{i}}\rangle_{\Omega}. \label{eq:matrix_coefficients_2}$$
Temporal discretization {#sec::temporal_discretization}
=======================
In this section we present the time discretization. The choice of a time discretization is essential in preserving invariants. Not all time integrators satisfy conservation of energy, even though the spatial discretization imposes conservation of kinetic energy as a function of time, see [@Sanderse2013]. In this work we choose to use a Gauss method. This time integrator is a type of collocation method based on Gauss quadrature. It is known to be an implicit Runge-Kutta method and to have optimal convergence order $2s$ for $s$ stages. Two of its most attractive properties are that (i) it conserves energy when applied to the discretization of the Navier-Stokes equations, see [@Sanderse2013], and (ii) it is time-reversible, see [@Hairer2006]. For a more detailed discussion of its properties and construction see [@Hairer2006]. Although any Gauss integrator could be used, we choose the lowest order, $s=1$, also known as the *midpoint rule*, because it enables the construction of an explicit staggered integrator in time.
When applied to the solution of a 1D ordinary differential equation of the form $$\begin{dcases}
\frac{\mathrm{d}f}{\mathrm{d}t} = g(f(t),t), \\
f(0) = f_{0},
\end{dcases}$$ the one stage Gauss integrator results in the following implicit time stepping scheme $$\frac{f^{k} - f^{k-1}}{\Delta t} = g\left(\frac{f^{k}+f^{k-1}}{2},t+\frac{\Delta t}{2}\right), \quad k=1,\dots,M, \label{eq:gauss_integrator_1D}$$ where $f^{0} = f_{0}$, $\Delta t$ is the time step and $M$ is the number of time steps. The direct application of to the discrete weak form results in $$\begin{dcases}
\text{Find } {{\vec{u}_{h}}^{\,k+1}}\in {\mathrm{RT}_{N}}, {{\bar{p}_{h}}^{\,k+1}}\in {\mathrm{DG}_{N-1}}\text{ and } {{\omega_{h}}^{\,k+1}}\in {\mathrm{CG}_{N}}\text{ such that:}\\
\langle\frac{{{\vec{u}_{h}}^{\,k+1}}- {{\vec{u}_{h}}^{\,k}}}{\Delta t},{\vec{v}_{h}}\rangle_{\Omega} + \langle{\tilde{\omega}_{h}^{\,k+\frac{1}{2}}}\times\frac{{{\vec{u}_{h}}^{\,k+1}}+{{\vec{u}_{h}}^{\,k}}}{2},{\vec{v}_{h}}\rangle_{\Omega} - \langle {\tilde{\bar{p}}_{h}^{\,k+\frac{1}{2}}},\nabla\cdot{\vec{v}_{h}}\rangle_{\Omega} = -\nu\langle\nabla\times{\tilde{\omega}_{h}^{\,k+\frac{1}{2}}},{\vec{v}_{h}}\rangle_{\Omega},\quad \forall {\vec{v}_{h}}\in {\mathrm{RT}_{N}}, \\
\langle\frac{{{\omega_{h}}^{\,k+1}}-{{\omega_{h}}^{\,k}}}{\Delta t},{\xi_{h}}\rangle_{\Omega} - \frac{1}{2}\langle\frac{{{\omega_{h}}^{\,k+1}}+{{\omega_{h}}^{\,k}}}{2},\nabla\cdot\left({\tilde{\vec{u}}_{h}^{\,k+\frac{1}{2}}}\,{\xi_{h}}\right)\rangle_{\Omega} + \frac{1}{2}\langle\nabla\cdot\left({\tilde{\vec{u}}_{h}^{\,k+\frac{1}{2}}}\,\frac{{{\omega_{h}}^{\,k+1}}+{{\omega_{h}}^{\,k}}}{2}\right),{\xi_{h}}\rangle_{\Omega} = \nu\langle\nabla\times\frac{{{\omega_{h}}^{\,k+1}}+{{\omega_{h}}^{\,k}}}{2},\nabla\times{\xi_{h}}\rangle_{\Omega}, \quad\forall {\xi_{h}}\in {\mathrm{CG}_{N}}, \\
\langle\nabla\cdot{{\vec{u}_{h}}^{\,k+1}},{q_{h}}\rangle_{\Omega} = 0, \quad\forall {q_{h}}\in {\mathrm{DG}_{N-1}}\,,
\end{dcases} \label{eq::ns_meevc_weak_form_discrete_naive_gauss}$$ where, for compactness of notation, we have set $${\tilde{\vec{u}}_{h}^{\,k+\frac{1}{2}}}:= \frac{{{\vec{u}_{h}}^{\,k+1}}+ {{\vec{u}_{h}}^{\,k}}}{2} \quad \mathrm{and}\quad {\tilde{\omega}_{h}^{\,k+\frac{1}{2}}}:= \frac{{{\omega_{h}}^{\,k+1}}+ {{\omega_{h}}^{\,k}}}{2}. \label{eq:mid_steps_compact_notation}$$
The time stepping scheme in consists of a coupled system of nonlinear equations. Therefore its solution will necessarily require an iterative procedure, which is computationally expensive. To overcome this drawback, instead of defining all the unknown physical quantities ${\vec{u}_{h}}$, ${\omega_{h}}$ and ${\bar{p}_{h}}$, at the same time instants $t^{k}$ we choose to stagger them in time. In this way it is possible to transform into two systems of quasi-linear equations. The unknown vorticity and total pressure are defined at the integer time instants ${{\omega_{h}}^{\,k}}$, ${{\bar{p}_{h}}^{\,k}}$ and the unknown velocity is defined at the intermediate time instants ${\vec{u}_{h}^{\,k+\frac{1}{2}}}$, see [Figure \[fig:time\_stepping\]]{}. Taking into account this staggered approach, can be rewritten as $$\begin{dcases}
\text{Find } {\vec{u}_{h}^{\,k+\frac{3}{2}}}\in {\mathrm{RT}_{N}}, {{\bar{p}_{h}}^{\,k+1}}\in {\mathrm{DG}_{N-1}}\text{ and } {{\omega_{h}}^{\,k+1}}\in {\mathrm{CG}_{N}}\text{ such that:}\\
\langle\frac{{\vec{u}_{h}^{\,k+\frac{3}{2}}}- {\vec{u}_{h}^{\,k+\frac{1}{2}}}}{\Delta t},{\vec{v}_{h}}\rangle_{\Omega} + \langle{{\omega_{h}}^{\,k+1}}\times\frac{{\vec{u}_{h}^{\,k+\frac{3}{2}}}+{\vec{u}_{h}^{\,k+\frac{1}{2}}}}{2},{\vec{v}_{h}}\rangle_{\Omega} - \langle {{\bar{p}_{h}}^{\,k+1}},\nabla\cdot{\vec{v}_{h}}\rangle_{\Omega} = -\nu\langle\nabla\times{{\omega_{h}}^{\,k+1}},{\vec{v}_{h}}\rangle_{\Omega},\quad \forall {\vec{v}_{h}}\in {\mathrm{RT}_{N}}, \\
\langle\frac{{{\omega_{h}}^{\,k+1}}-{{\omega_{h}}^{\,k}}}{\Delta t},{\xi_{h}}\rangle_{\Omega} - \frac{1}{2}\langle\frac{{{\omega_{h}}^{\,k+1}}+{{\omega_{h}}^{\,k}}}{2},\nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,{\xi_{h}}\right)\rangle_{\Omega} + \frac{1}{2}\langle\nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,\frac{{{\omega_{h}}^{\,k+1}}+{{\omega_{h}}^{\,k}}}{2}\right),{\xi_{h}}\rangle_{\Omega} = \nu\langle\nabla\times\frac{{{\omega_{h}}^{\,k+1}}+{{\omega_{h}}^{\,k}}}{2},\nabla\times{\xi_{h}}\rangle_{\Omega}, \quad\forall {\xi_{h}}\in {\mathrm{CG}_{N}}, \\
\langle\nabla\cdot{\vec{u}_{h}^{\,k+\frac{3}{2}}},{q_{h}}\rangle_{\Omega} = 0, \quad\forall {q_{h}}\in {\mathrm{DG}_{N-1}}\,,
\end{dcases} \label{eq::ns_meevc_weak_form_discrete_staggered_gauss}$$ where ${\vec{u}_{h}^{\,k+\frac{1}{2}}}$ and ${{\omega_{h}}^{\,k}}$ are known at the start of each time step.
Using , it is possible to rewrite in a compact matrix notation $$\begin{dcases}
\text{Find } {\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}\in \mathbb{R}^{{d_{U}}}, {{\boldsymbol{\bar{p}}}^{\,k+1}}\in \mathbb{R}^{{d_{Q}}} \text{ and } {{\boldsymbol{\omega}}^{\,k+1}}\in \mathbb{R}^{{d_{W}}} \text{ such that:}\\
{\boldsymbol{\mathsf{M}}} \frac{{\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}- {\boldsymbol{u}^{\,k+\frac{1}{2}}}}{\Delta t} + {\boldsymbol{\mathsf{R}}}^{k+1}\,\frac{{\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}+ {\boldsymbol{u}^{\,k+\frac{1}{2}}}}{2} - {\boldsymbol{\mathsf{P}}}\,{{\boldsymbol{\bar{p}}}^{\,k+1}}= -\nu\,\boldsymbol{l}^{k+1}, \\
{\boldsymbol{\mathsf{N}}}\frac{{{\boldsymbol{\omega}}^{\,k+1}}- {{\boldsymbol{\omega}}^{\,k}}}{\Delta t} - \frac{1}{2}{\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}}\,\frac{{{\boldsymbol{\omega}}^{\,k+1}}+ {{\boldsymbol{\omega}}^{\,k}}}{2} + \frac{1}{2}\left({\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}}\right)^{\top}\frac{{{\boldsymbol{\omega}}^{\,k+1}}+ {{\boldsymbol{\omega}}^{\,k}}}{2} = \nu\,{\boldsymbol{\mathsf{L}}}\,\frac{{{\boldsymbol{\omega}}^{\,k+1}}+ {{\boldsymbol{\omega}}^{\,k}}}{2}, \\
{\boldsymbol{\mathsf{D}}}\,{\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}= 0,
\end{dcases} \label{eq::ns_meevc_weak_form_discrete_staggered_gauss_matrix_notation}$$ where all matrix operators are as in and , with the exception of ${\boldsymbol{\mathsf{R}}}^{k+1}$, ${\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}}$ and $\boldsymbol{l}^{k+1}$, the coefficients of which are $${\mathsf{R}}^{k+1}_{ij} := \langle{{\omega_{h}}^{\,k+1}}\times{\vec{\epsilon}^{\,U}_{j}},{\vec{\epsilon}^{\,U}_{i}}\rangle_{\Omega}, \quad l^{k+1}_{i} := \langle\nabla\times{{\omega_{h}}^{\,k+1}},{\vec{\epsilon}^{\,U}_{i}}\rangle_{\Omega}\quad\mathrm{and}\quad {\mathsf{W}}^{k+\frac{1}{2}}_{ij} := \langle{\epsilon^{W}_{j}},\nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,{\epsilon^{W}_{i}}\right)\rangle_{\Omega}. \label{eq:matrix_coefficients_staggered}$$
To start the iteration procedure ${\vec{u}_{h}^{\,\frac{1}{2}}}$ and ${{\omega_{h}}^{\,0}}$ are required. Since only ${\vec{u}_{h}^{\,0}}$ and ${{\omega_{h}}^{\,0}}$ are known, the first time step needs to be implicit, according to . Once ${\vec{u}_{h}^{\,1}}$ is known, can be used to retrieve ${\vec{u}_{h}^{\,\frac{1}{2}}}$. The remaining time steps can then be computed explicitly with .
![Diagram of the time stepping. Left: all unknowns at the same time instant, as in . Right: staggered in time unknowns, as in .[]{data-label="fig:time_stepping"}](./figures/time_stepping/time_stepping)
Conservation properties and time reversibility {#sec::conservation_properties_and_time_reversibility}
==============================================
It is known that in the absence of external forces and viscosity the incompressible Navier-Stokes equations contain important symmetries or invariants, e.g. [@Arnold1966; @Arnold1992; @Majda2001; @Foias2001]. Conservation of mass, energy, enstrophy and total vorticity are such invariants.
A discretization for the incompressible Navier-Stokes equations has been proposed in . In the following sections we prove its conservation properties, with respect to mass, energy, enstrophy and vorticity. Additionally, we also prove time reversibility.
Mass and vorticity conservation {#sec:mass_vorticity_conservation}
-------------------------------
Mass conservation is given by the divergence of the velocity field. Since with this discretization the discrete velocity field is exactly divergence free mass is conserved.
Regarding vorticity conservation, the time evolution of vorticity is governed by the vorticity transport equation, : $$\frac{\partial {\omega}}{\partial t} +\frac{1}{2}\left({\vec{u}}\cdot\nabla\right){\omega}+ \frac{1}{2}\nabla\cdot\left({\vec{u}}{\omega}\right) = \nu \Delta{\omega}.$$ This equation, by , is discretized as $$\begin{dcases}
\text{Find } {{{\omega_{h}}^{\,k+1}}}\in {\mathrm{CG}_{N}}\text{ such that:}\\
\langle\frac{{{{\omega_{h}}^{\,k+1}}}-{{{\omega_{h}}^{\,k}}}}{\Delta t},{\xi_{h}}\rangle_{\Omega} - \frac{1}{2}\langle\frac{{{{\omega_{h}}^{\,k+1}}}+{{{\omega_{h}}^{\,k}}}}{2},\nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,{\xi_{h}}\right)\rangle_{\Omega} + \frac{1}{2}\langle\nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,\frac{{{{\omega_{h}}^{\,k+1}}}+{{{\omega_{h}}^{\,k}}}}{2}\right),{\xi_{h}}\rangle_{\Omega} = \nu\langle\nabla\times\frac{{{{\omega_{h}}^{\,k+1}}}+{{{\omega_{h}}^{\,k}}}}{2},\nabla\times{\xi_{h}}\rangle_{\Omega}, \quad\forall {\xi_{h}}\in {\mathrm{CG}_{N}},
\end{dcases} \label{eq::generic_transport_equation}$$ Conservation of vorticity ${{\omega_{h}}}$ corresponds to $$\int_{\Omega}{{{\omega_{h}}^{\,k+1}}}:= \langle{{{\omega_{h}}^{\,k+1}}},1\rangle = \langle{{{\omega_{h}}^{\,k}}},1\rangle := \int_{\Omega}{{{\omega_{h}}^{\,k}}}.$$ Therefore, the proof of conservation of vorticity is obtained from evaluating for the special case ${\xi_{h}}= 1$: $$\langle\frac{{{{\omega_{h}}^{\,k+1}}}-{{{\omega_{h}}^{\,k}}}}{\Delta t},1\rangle_{\Omega} - \frac{1}{2}\langle\frac{{{{\omega_{h}}^{\,k+1}}}+{{{\omega_{h}}^{\,k}}}}{2},\nabla\cdot{\vec{u}_{h}^{\,k+\frac{1}{2}}}\rangle_{\Omega} + \frac{1}{2}\langle\nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,\frac{{{{\omega_{h}}^{\,k+1}}}+{{{\omega_{h}}^{\,k}}}}{2}\right),1\rangle_{\Omega} = \nu\langle\nabla\times\frac{{{{\omega_{h}}^{\,k+1}}}+{{{\omega_{h}}^{\,k}}}}{2},\nabla\times1\rangle_{\Omega} . \label{eq:proof_conservation_mass_vorticity_1}$$ Since the function spaces have been chosen such that $\nabla\cdot{\vec{u}_{h}}= 0$ is satisfied exactly, see [Section \[subsec::finite\_element\_discretization\]]{}, equation becomes $$\langle{{{\omega_{h}}^{\,k+1}}},1\rangle_{\Omega} + \frac{\Delta t}{2}\langle\nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,\frac{{{{\omega_{h}}^{\,k+1}}}+{{{\omega_{h}}^{\,k}}}}{2}\right),1\rangle_{\Omega} = \langle{{{\omega_{h}}^{\,k}}},1\rangle_{\Omega} . \label{eq:proof_conservation_mass_vorticity_2}$$ The second term on the left hand side can be rewritten as $$\langle\nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,\frac{{{{\omega_{h}}^{\,k+1}}}+{{{\omega_{h}}^{\,k}}}}{2}\right),1\rangle_{\Omega} := \int_{\Omega} \nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,\frac{{{{\omega_{h}}^{\,k+1}}}+{{{\omega_{h}}^{\,k}}}}{2}\right) = \int_{\partial\Omega} \left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,\frac{{{{\omega_{h}}^{\,k+1}}}+{{{\omega_{h}}^{\,k}}}}{2}\right)\cdot\vec{n}, \label{eq:proof_conservation_mass_vorticity_3}$$ with $\vec{n}$ the outward unit normal at the boundary of $\Omega$. At the continuous level, the boundary integral in is trivially equal to zero for periodic boundary conditions, which is the case we will consider here. Note that since the domain is periodic, the nodes of the mesh at opposite sides of domain must coincide, otherwise the periodicity is lost. Since ${{\omega_{h}}}\in {\mathrm{CG}_{N}}$, it is continuous across elements. In a similar way, the normal component of ${\vec{u}_{h}}$ is continuous across elements because ${\vec{u}_{h}}\in {\mathrm{RT}_{N}}$, see for example [@RaviartThomas1977; @arnold2010finite; @kirby2012]. Therefore, in the case of periodic boundary conditions the boundary integral in will still be exactly equal to zero and becomes $$\langle{{{\omega_{h}}^{\,k+1}}},1\rangle_{\Omega} = \langle{{{\omega_{h}}^{\,k}}},1\rangle_{\Omega},\label{eq:proof_conservation_mass_vorticity_4}$$ which proves that vorticity ${\mathcal{W}_{h}^{k}}$ is conserved because $${\mathcal{W}_{h}^{k+1}}:= \langle{\omega_{h}}^{k+1},1\rangle_{\Omega} = \langle{\omega_{h}}^{k},1\rangle_{\Omega} =:{\mathcal{W}_{h}^{k}}$$ is conserved.
Kinetic energy conservation {#sec::kinetic_energy_conservation}
---------------------------
Kinetic energy conservation is one of the two *secondary conservation* properties of the numerical solver proposed in this work. Here we prove that total kinetic energy ${\mathcal{K}}$ is conserved by .
At the continuous level, kinetic energy is defined as proportional to the $L^{2}(\Omega)$ norm of the velocity velocity $${\mathcal{K}}:= \frac{1}{2}\|{\vec{u}}\|_{L^{2}(\Omega)} := \frac{1}{2}\langle{\vec{u}},{\vec{u}}\rangle,$$ is conserved in the inviscid limit $\nu = 0$, see for example [@Majda2001; @Foias2001]. This definition can be directly extended to the discrete level as $${\mathcal{K}_{h}}:= \frac{1}{2}\|{\vec{u}_{h}}\|_{L^{2}(\Omega)} := \frac{1}{2}\langle{\vec{u}_{h}},{\vec{u}_{h}}\rangle.$$ Therefore, kinetic energy conservation at the discrete level corresponds to $${\mathcal{K}_{h}^{k+\frac{3}{2}}}:= \frac{1}{2}\langle{\vec{u}_{h}^{\,k+\frac{3}{2}}},{\vec{u}_{h}^{\,k+\frac{3}{2}}}\rangle = \frac{1}{2}\langle{\vec{u}_{h}^{\,k+\frac{1}{2}}},{\vec{u}_{h}^{\,k+\frac{1}{2}}}\rangle := {\mathcal{K}_{h}^{k+\frac{1}{2}}}. \label{eq:energy_conservation_definition}$$
To prove this identity we take the first equation in (momentum equation) in the inviscid limit $\nu = 0$ and choose ${\vec{v}_{h}}= \frac{1}{2}{\vec{u}_{h}^{\,k+\frac{3}{2}}}+ \frac{1}{2}{\vec{u}_{h}^{\,k+\frac{1}{2}}}$, resulting in $$\frac{1}{2}\langle\frac{{\vec{u}_{h}^{\,k+\frac{3}{2}}}- {\vec{u}_{h}^{\,k+\frac{1}{2}}}}{\Delta t},{\vec{u}_{h}^{\,k+\frac{3}{2}}}+ {\vec{u}_{h}^{\,k+\frac{1}{2}}}\rangle_{\Omega} + \frac{1}{2}\langle{{\omega_{h}}^{\,k+1}}\times\frac{{\vec{u}_{h}^{\,k+\frac{3}{2}}}+{\vec{u}_{h}^{\,k+\frac{1}{2}}}}{2},{\vec{u}_{h}^{\,k+\frac{3}{2}}}+ {\vec{u}_{h}^{\,k+\frac{1}{2}}}\rangle_{\Omega} - \frac{1}{2}\langle {{\bar{p}_{h}}^{\,k+1}},\nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{3}{2}}}+ {\vec{u}_{h}^{\,k+\frac{1}{2}}}\right)\rangle_{\Omega} =0. \label{eq:energy_conservation_proof_1}$$ The term involving the total pressure ${\bar{p}_{h}}$ is identically zero because the velocity field is divergence free at every time step, due to the particular choice of function spaces, as discussed in [Section \[subsec::finite\_element\_discretization\]]{}. After rearranging, equation then becomes $$\frac{1}{2}\langle{\vec{u}_{h}^{\,k+\frac{3}{2}}},{\vec{u}_{h}^{\,k+\frac{3}{2}}}\rangle_{\Omega} + \frac{\Delta t}{2}\,\langle{{\omega_{h}}^{\,k+1}}\times\frac{{\vec{u}_{h}^{\,k+\frac{3}{2}}}+{\vec{u}_{h}^{\,k+\frac{1}{2}}}}{2},{\vec{u}_{h}^{\,k+\frac{3}{2}}}+ {\vec{u}_{h}^{\,k+\frac{1}{2}}}\rangle_{\Omega} = \frac{1}{2}\langle{\vec{u}_{h}^{\,k+\frac{1}{2}}},{\vec{u}_{h}^{\,k+\frac{1}{2}}}\rangle_{\Omega}. \label{eq:energy_conservation_proof_1b}$$ Because of Lemma 1.3 in [@TemamNS], becomes $$\frac{1}{2}\langle{\vec{u}_{h}^{\,k+\frac{3}{2}}},{\vec{u}_{h}^{\,k+\frac{3}{2}}}\rangle_{\Omega} = \frac{1}{2}\langle{\vec{u}_{h}^{\,k+\frac{1}{2}}},{\vec{u}_{h}^{\,k+\frac{1}{2}}}\rangle_{\Omega},$$ which proves .
Enstrophy conservation {#sec::enstrophy_conservation}
----------------------
The second *secondary conservation* property of the numerical solver presented here is enstrophy conservation. We then proceed to prove conservation of enstrophy ${\mathcal{E}}$ at the discrete level.
Similarly to kinetic energy, enstrophy is defined as proportional to the $L^{2}(\Omega)$ norm of vorticity $${\mathcal{E}}:= \frac{1}{2} \|{\omega}\|_{L^{2}(\Omega)} := \langle{\omega},{\omega}\rangle_{\Omega},$$ and is also a conserved quantity in the inviscid limit $\nu = 0$, see for example [@Majda2001; @Foias2001]. This definition can be straightforwardly extended to the discrete level as $${\mathcal{E}_{h}}:= \frac{1}{2} \|{\omega_{h}}\|_{L^{2}(\Omega)} := \langle{\omega_{h}},{\omega_{h}}\rangle_{\Omega},$$ This directly implies that conservation of enstrophy at the discrete level requires that $${\mathcal{E}_{h}^{k+1}}:= \frac{1}{2}\langle{{\omega_{h}}^{\,k+1}},{{\omega_{h}}^{\,k+1}}\rangle = \frac{1}{2}\langle{{\omega_{h}}^{\,k}},{{\omega_{h}}^{\,k}}\rangle := {\mathcal{E}_{h}^{k}}.$$
The proof of enstrophy conservation follows a similar approach to the one used for the proof of energy conservation. In this case we start with the second equation in (vorticity transport equation) in the inviscid limit $\nu = 0$ and choose ${\xi_{h}}= \frac{1}{2}{{\omega_{h}}^{\,k+1}}+ \frac{1}{2}{{\omega_{h}}^{\,k}}$, obtaining $$\frac{1}{2}\langle\frac{{{\omega_{h}}^{\,k+1}}-{{\omega_{h}}^{\,k}}}{\Delta t},{{\omega_{h}}^{\,k+1}}+{{\omega_{h}}^{\,k}}\rangle_{\Omega} - \frac{1}{4}\langle\frac{{{\omega_{h}}^{\,k+1}}+{{\omega_{h}}^{\,k}}}{2},\nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,\left({{\omega_{h}}^{\,k+1}}+{{\omega_{h}}^{\,k}}\right)\right)\rangle_{\Omega} + \frac{1}{4}\langle\nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,\frac{{{\omega_{h}}^{\,k+1}}+{{\omega_{h}}^{\,k}}}{2}\right),{{\omega_{h}}^{\,k+1}}+{{\omega_{h}}^{\,k}}\rangle_{\Omega} = 0. \label{eq:enstrophy_conservation_proof_1}$$ The last two terms on the left hand side of are equal and have opposite signs, therefore cancelling each other. Conservation of enstrophy then follows directly.
Time reversibility
------------------
It is well known that the Navier-Stokes equations in the inviscid limit $\nu = 0$ are time reversible, e.g. [@Majda2001; @Duponcheel2008]. This important property has served in the past as a benchmark for numerical flow solvers [@Duponcheel2008] and for the construction of improved LES subgrid models [@CARATI2001], for example. It is intimately related to the numerical dissipation and therefore it is desirable to satisfy it.
To prove time reversibility of the proposed scheme we follow the same approach presented by Hairer et al. in [@Hairer2006]. Our time integrator $\Phi_{\Delta t}$ is a *one-step* method since it uses only the information of the previous time step: $$\Phi_{\Delta t} ({\vec{u}_{h}^{\,k+\frac{1}{2}}},{{\omega_{h}}^{\,k}}) = ({\vec{u}_{h}^{\,k+\frac{3}{2}}},{{\omega_{h}}^{\,k+1}}).$$ A numerical one-step method $\Phi_{\Delta t}$ is *time-reversible*, if it satisfies, see Hairer et al. [@Hairer2006], $$\Phi_{\Delta t}\circ\Phi_{-\Delta t} = id \qquad \text{or equivalently} \qquad \Phi_{\Delta t} = \Phi^{-1}_{-\Delta t}\,.$$ To prove time reversibility of our method we will show that $\Phi_{\Delta t}\circ\Phi_{-\Delta t} = id$.
Start with known velocity ${\vec{u}_{h}^{\,k+\frac{1}{2}}}$ at time instant $t^{k+\frac{1}{2}}$ and known vorticity ${{\omega_{h}}^{\,k}}$ at time instant $t^{k}$. Advance both quantities one time step to obtain the velocity ${\vec{u}_{h}^{\,k+\frac{3}{2}}}$ at time instant $t^{k+\frac{3}{2}}$ and the vorticity ${{\omega_{h}}^{\,k+1}}$ at the time instant $t^{k+1}$. Reverse time and compute one time step. If the numerical method is time reversible the initial velocity and vorticity fields must be retrieved, as represented in [Figure \[fig:time\_reversibility\]]{}. First we prove the reversibility of the vorticity time step and then the reversibility of the velocity time step.
![Diagram of time reversibility. Left: forward step. Right: backward step.[]{data-label="fig:time_reversibility"}](./figures/time_reversibility/time_reversibility)
### Reversibility of vorticity time step
In the inviscid limit $\nu = 0$, the forward vorticity time step, [Figure \[fig:time\_reversibility\]]{} left, which advances the vorticity field from the time instant $t^{k}$ to the time instant $t^{k+1}$ is given by the second equation in (vorticity transport equation) $$\begin{dcases}
\text{Find } {{\omega_{h}}^{\,k+1}}\in {\mathrm{CG}_{N}}\text{ such that:}\\
\langle\frac{{{\omega_{h}}^{\,k+1}}-{{\omega_{h}}^{\,k}}}{\Delta t},{\xi_{h}}\rangle_{\Omega} - \frac{1}{2}\langle\frac{{{\omega_{h}}^{\,k+1}}+{{\omega_{h}}^{\,k}}}{2},\nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,{\xi_{h}}\right)\rangle_{\Omega} + \frac{1}{2}\langle\nabla\cdot\left({\vec{u}_{h}^{\,k+\frac{1}{2}}}\,\frac{{{\omega_{h}}^{\,k+1}}+{{\omega_{h}}^{\,k}}}{2}\right),{\xi_{h}}\rangle_{\Omega} = 0,\quad\forall{\xi_{h}}\in{\mathrm{CG}_{N}}.
\end{dcases}\label{eq:time_reversibility_vorticity_proof_1}$$ Using , this expression can be rewritten in matrix notation as an algebraic system of equations: $${\boldsymbol{\mathsf{N}}}\frac{{{\boldsymbol{\omega}}^{\,k+1}}- {{\boldsymbol{\omega}}^{\,k}}}{\Delta t} - \frac{1}{2}{\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}}\,\frac{{{\boldsymbol{\omega}}^{\,k+1}}+ {{\boldsymbol{\omega}}^{\,k}}}{2} + \frac{1}{2}\left({\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}}\right)^{\top}\frac{{{\boldsymbol{\omega}}^{\,k+1}}+ {{\boldsymbol{\omega}}^{\,k}}}{2} = 0.\label{eq:time_reversibility_vorticity_proof_1b}$$ Rearranging this expression it is possible to solve for ${{\boldsymbol{\omega}}^{\,k+1}}$ $${{\boldsymbol{\omega}}^{\,k+1}}= \left({\boldsymbol{\mathsf{N}}} - \frac{\Delta t}{4}{\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}} + \frac{\Delta t}{4}\left({\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}}\right)^{\top}\right)^{-1}\left({\boldsymbol{\mathsf{N}}} + \frac{\Delta t}{4}{\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}} - \frac{\Delta t}{4}\left({\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}}\right)^{\top}\right){{\boldsymbol{\omega}}^{\,k}}. \label{eq:time_reversibility_vorticity_proof_2}$$ If we now introduce the compact notation $${\boldsymbol{\mathsf{A}}}_{-}:= {\boldsymbol{\mathsf{N}}} - \frac{\Delta t}{4}{\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}} + \frac{\Delta t}{4}\left({\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}}\right)^{\top} \quad\mathrm{and}\quad{\boldsymbol{\mathsf{A}}}_{+}:={\boldsymbol{\mathsf{N}}} + \frac{\Delta t}{4}{\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}} - \frac{\Delta t}{4}\left({\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}}\right)^{\top},$$ equation becomes $${{\boldsymbol{\omega}}^{\,k+1}}= {\boldsymbol{\mathsf{A}}}_{-}^{-1}{\boldsymbol{\mathsf{A}}}_{+}\,{{\boldsymbol{\omega}}^{\,k}}. \label{eq:vorticity_reversibility_forward_compact}$$
The backward time step for vorticity, [Figure \[fig:time\_reversibility\]]{} right, is obtained by substituting $\Delta t$ by $-\Delta t$. Substituting this transformation in and using leads to an expression identical to $${\boldsymbol{\mathsf{N}}}\frac{{{\boldsymbol{\omega}}^{\,k}}- {{\boldsymbol{\omega}}^{\,k+1}}}{-\Delta t} - \frac{1}{2}{\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}}\,\frac{{{\boldsymbol{\omega}}^{\,k}}+ {{\boldsymbol{\omega}}^{\,k+1}}}{2} + \frac{1}{2}\left({\boldsymbol{\mathsf{W}}}^{k+\frac{1}{2}}\right)^{\top}\frac{{{\boldsymbol{\omega}}^{\,k}}+ {{\boldsymbol{\omega}}^{\,k+1}}}{2} = 0.\label{eq:time_reversibility_vorticity_proof_3}$$ This expression can be rearranged to yield a result similar to $${{\boldsymbol{\omega}}^{\,k}}= {\boldsymbol{\mathsf{A}}}_{+}^{-1}{\boldsymbol{\mathsf{A}}}_{-}\,{{\boldsymbol{\omega}}^{\,k+1}}. \label{eq:vorticity_reversibility_forward_compact_2}$$ Combining the forward time step with the backward time step results in $${{\boldsymbol{\omega}}^{\,k}}= {\boldsymbol{\mathsf{A}}}_{+}^{-1}{\boldsymbol{\mathsf{A}}}_{-}\left({\boldsymbol{\mathsf{A}}}_{-}^{-1}{\boldsymbol{\mathsf{A}}}_{+}{{\boldsymbol{\omega}}^{\,k}}\right) = {{\boldsymbol{\omega}}^{\,k}}.$$ Thus showing the reversibility of the vorticity time step.
### Reversibility of velocity time step
For the reversibility of the velocity time step consider first the forward time step, [Figure \[fig:time\_reversibility\]]{} left, in the inviscid limit $\nu = 0$. The first equation in (momentum equation) computes the evolution of the velocity field from time instant $t^{k+\frac{1}{2}}$ to the time instant $t^{k+\frac{3}{2}}$ $$\begin{dcases}
\text{Find } {\vec{u}_{h}^{\,k+\frac{3}{2}}}\in {\mathrm{RT}_{N}}, {{\bar{p}_{h}}^{\,k+1}}\in {\mathrm{DG}_{N-1}}\text{ such that:}\\
\langle\frac{{\vec{u}_{h}^{\,k+\frac{3}{2}}}- {\vec{u}_{h}^{\,k+\frac{1}{2}}}}{\Delta t},{\vec{v}_{h}}\rangle_{\Omega} + \langle{{\omega_{h}}^{\,k+1}}\times\frac{{\vec{u}_{h}^{\,k+\frac{3}{2}}}+{\vec{u}_{h}^{\,k+\frac{1}{2}}}}{2},{\vec{v}_{h}}\rangle_{\Omega} - \langle {{\bar{p}_{h}}^{\,k+1}},\nabla\cdot{\vec{v}_{h}}\rangle_{\Omega} = 0,\quad \forall {\vec{v}_{h}}\in {\mathrm{RT}_{N}}, \\
\langle\nabla\cdot{\vec{u}_{h}^{\,k+\frac{3}{2}}},{q_{h}}\rangle_{\Omega} = 0, \quad\forall {q_{h}}\in {\mathrm{DG}_{N-1}}\,.
\end{dcases} \label{eq::velocity_reversibility_proof_1}$$ Using we can write this expression as an algebraic system of equations $$\begin{dcases}
{\boldsymbol{\mathsf{M}}} \frac{{\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}- {\boldsymbol{u}^{\,k+\frac{1}{2}}}}{\Delta t} + {\boldsymbol{\mathsf{R}}}^{k+1}\,\frac{{\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}+ {\boldsymbol{u}^{\,k+\frac{1}{2}}}}{2} - {\boldsymbol{\mathsf{P}}}\,{{\boldsymbol{\bar{p}}}^{\,k+1}}=0, \\
{\boldsymbol{\mathsf{D}}}\,{\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}= 0,
\end{dcases} \label{eq::velocity_reversibility_proof_2}$$ of which ${\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}$ and ${{\boldsymbol{\bar{p}}}^{\,k+1}}$ are the solution. Once ${{\boldsymbol{\bar{p}}}^{\,k+1}}$ is known, it is possible to write an explicit expression for ${\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}$ as a function of ${\boldsymbol{u}^{\,k+\frac{1}{2}}}$ and ${{\boldsymbol{\bar{p}}}^{\,k+1}}$ $${\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}= \left({\boldsymbol{\mathsf{M}}} +\frac{\Delta t}{2} {\boldsymbol{\mathsf{R}}}^{k+1}\right)^{-1}\left({\boldsymbol{\mathsf{M}}} -\frac{\Delta t}{2} {\boldsymbol{\mathsf{R}}}^{k+1}\right) {\boldsymbol{u}^{\,k+\frac{1}{2}}}+ \Delta t \,\left({\boldsymbol{\mathsf{M}}} +\frac{\Delta t}{2} {\boldsymbol{\mathsf{R}}}^{k+1}\right)^{-1}{\boldsymbol{\mathsf{P}}}\,{{\boldsymbol{\bar{p}}}^{\,k+1}}. \label{eq:velocity_reversibility_proof_3}$$ Introducing the compact notation $${\boldsymbol{\mathsf{B}}}_{+} := {\boldsymbol{\mathsf{M}}} +\frac{\Delta t}{2} {\boldsymbol{\mathsf{R}}}^{k+1} \quad\mathrm{and}\quad {\boldsymbol{\mathsf{B}}}_{-} := {\boldsymbol{\mathsf{M}}} -\frac{\Delta t}{2} {\boldsymbol{\mathsf{R}}}^{k+1},$$ equation becomes $${\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}= {\boldsymbol{\mathsf{B}}}_{+}^{-1}{\boldsymbol{\mathsf{B}}}_{-}\,{\boldsymbol{u}^{\,k+\frac{1}{2}}}+ \Delta t \,{\boldsymbol{\mathsf{B}}}_{+}^{-1}{\boldsymbol{\mathsf{P}}}\,{{\boldsymbol{\bar{p}}}^{\,k+1}}. \label{eq:velocity_reversibility_proof_4}$$
To compute the backward time step for velocity, [Figure \[fig:time\_reversibility\]]{} right, we proceed in the same manner as for vorticity: reverse the time step by replacing $\Delta t$ by $-\Delta t$. We can now apply this transformation to and use to obtain an expression identical to $$\begin{dcases}
{\boldsymbol{\mathsf{M}}} \frac{{\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}- {\boldsymbol{u}^{\,k+\frac{1}{2}}}}{-\Delta t} + {\boldsymbol{\mathsf{R}}}^{k+1}\,\frac{{\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}+ {\boldsymbol{u}^{\,k+\frac{1}{2}}}}{2} - {\boldsymbol{\mathsf{P}}}\,{{\boldsymbol{\bar{p}}}^{\,k+1}}=0, \\
{\boldsymbol{\mathsf{D}}}\,\left({\boldsymbol{u}^{\,k+\frac{1}{2}}}\right) = 0,
\end{dcases} \label{eq::velocity_reversibility_proof_5}$$ If we assume for now that ${{\boldsymbol{\bar{p}}}^{\,k+1}}$ is known and equal to the one obtained in the forward step, it is possible to use to write an expression equivalent to but for the backward step $${\boldsymbol{u}^{\,k+\frac{1}{2}}}= {\boldsymbol{\mathsf{B}}}_{-}^{-1}{\boldsymbol{\mathsf{B}}}_{+}\,{\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}- \Delta t \,{\boldsymbol{\mathsf{B}}}_{-}^{-1}{\boldsymbol{\mathsf{P}}}\,{{\boldsymbol{\bar{p}}}^{\,k+1}}. \label{eq:velocity_reversibility_proof_6}$$ Replacing into yields $${\boldsymbol{u}^{\,k+\frac{1}{2}}}= {\boldsymbol{\mathsf{B}}}_{-}^{-1}{\boldsymbol{\mathsf{B}}}_{+}\,\left({\boldsymbol{\mathsf{B}}}_{+}^{-1}{\boldsymbol{\mathsf{B}}}_{-}\,{\boldsymbol{u}^{\,k+\frac{1}{2}}}+ \Delta t \,{\boldsymbol{\mathsf{B}}}_{+}^{-1}{\boldsymbol{\mathsf{P}}}\,{{\boldsymbol{\bar{p}}}^{\,k+1}}\right) + \Delta t \,{\boldsymbol{\mathsf{B}}}_{-}^{-1}{\boldsymbol{\mathsf{P}}}\,{{\boldsymbol{\bar{p}}}^{\,k+1}}, \label{eq:velocity_reversibility_proof_7}$$ which can be simplified to $${\boldsymbol{u}^{\,k+\frac{1}{2}}}={\boldsymbol{u}^{\,k+\frac{1}{2}}}+\Delta t\,{\boldsymbol{\mathsf{B}}}_{-}^{-1}{\boldsymbol{\mathsf{P}}}\,\left({{\boldsymbol{\bar{p}}}^{\,k+1}}- {{\boldsymbol{\bar{p}}}^{\,k+1}}\right). \label{eq:velocity_reversibility_proof_7}$$ The last term on the right will be trivially equal to zero proving reversibility of the velocity time step. This term cancels out because we assumed the total pressure computed with the backward time step is equal to the one computed in the forward step. For this reason, to finalise the proof we need to show that indeed the forward step total pressure is the solution for the pressure in the backward step. We start with a modified version of $$\begin{dcases}
{\boldsymbol{\mathsf{M}}} \frac{{\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}- {\boldsymbol{u}^{\,k+\frac{1}{2}}}}{\Delta t} + {\boldsymbol{\mathsf{R}}}^{k+1}\,\frac{{\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}+ {\boldsymbol{u}^{\,k+\frac{1}{2}}}}{2} - {\boldsymbol{\mathsf{P}}}\,{{\boldsymbol{\bar{p}}}^{\,k+1}}=0, \\
{\boldsymbol{\mathsf{D}}}\,\left({\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}+ {\boldsymbol{u}^{\,k+\frac{1}{2}}}\right) = 0.
\end{dcases} \label{eq::velocity_reversibility_proof_8}$$ This system of equations is equivalent to $\eqref{eq::velocity_reversibility_proof_2}$ because the divergence of the discrete velocity is zero, therefore ${\boldsymbol{\mathsf{D}}}\,{\vec{u}_{h}^{\,k+\frac{1}{2}}}= 0$. Rearranging gives $$\begin{dcases}
\left({\boldsymbol{\mathsf{M}}} + \frac{\Delta t}{2}{\boldsymbol{\mathsf{R}}}^{k+1}\right)\,{\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}- \left({\boldsymbol{\mathsf{M}}} - \frac{\Delta t}{2}{\boldsymbol{\mathsf{R}}}^{k+1}\right)\,{\boldsymbol{u}^{\,k+\frac{1}{2}}}- \Delta t\,{\boldsymbol{\mathsf{P}}}\,{{\boldsymbol{\bar{p}}}^{\,k+1}}=0, \\
{\boldsymbol{\mathsf{D}}}\,\left({\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}+ {\boldsymbol{u}^{\,k+\frac{1}{2}}}\right) = 0.
\end{dcases} \label{eq::velocity_reversibility_proof_8}$$ In the same way the backwards step results in the following system of equations $$\begin{dcases}
-\left({\boldsymbol{\mathsf{M}}} - \frac{\Delta t}{2}{\boldsymbol{\mathsf{R}}}^{k+1}\right)\,{\boldsymbol{u}^{\,k+\frac{1}{2}}}+ \left({\boldsymbol{\mathsf{M}}} + \frac{\Delta t}{2}{\boldsymbol{\mathsf{R}}}^{k+1}\right)\,{\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}- \Delta t\,{\boldsymbol{\mathsf{P}}}\,{{\boldsymbol{\bar{p}}}^{\,k+1}}=0, \\
{\boldsymbol{\mathsf{D}}}\,\left({\boldsymbol{u}_{h}^{\,k+\frac{3}{2}}}+ {\boldsymbol{u}^{\,k+\frac{1}{2}}}\right) = 0.
\end{dcases} \label{eq::velocity_reversibility_proof_9}$$ The system is the same as , therefore the pressure is necessarily the same in both cases. Thus proving the reversibility of the velocity time step.
Numerical test cases {#sec::numerical_test_cases}
====================
In order to verify the numerical properties of the proposed method we apply it to the solution of two well known test cases: (i) Taylor-Green vortex and (ii) inviscid shear layer roll-up. With the first test the convergence properties of the method will be assessed and with the second the time reversibility and mass, energy, enstrophy, and vorticity conservation will be verified.
Order of accuracy study: Taylor-Green vortex {#subsec::taylor_green_vortex}
--------------------------------------------
In order to verify the accuracy of the proposed numerical method we compare the numerical results against an analytical solution of the Navier-Stokes equations. A suitable analytical solution is the Taylor-Green vortex, given by: $$\begin{dcases}
u_{x} (x,y,t) = -\sin(\pi x)\cos(\pi y)\,e^{-2\pi^{2} \nu t}, \\
u_{y} (x,y,t) = \cos(\pi x)\sin(\pi y)\,e^{-2\pi^{2} \nu t}, \\
p(x,y,t) = \frac{1}{4}\,(\cos(2\pi x)+ \cos(2\pi y))\,e^{4\pi^{2} \nu t}, \\
\omega (x,y,t) = -2\pi\sin(\pi x)\sin(\pi y)\,e^{-2\pi^{2} \nu t}.
\end{dcases} \label{eq:taylor_green_exact_solution}$$ The solution is defined on the domain $\Omega = [0,2]\times[0,2]$, with periodic boundary conditions. The initial condition for both the velocity ${\vec{u}_{h}^{\,0}}$ and vorticity ${{\omega_{h}}^{\,0}}$ are given by the exact solution . The kinematic viscosity is set to $\nu = 0.01$. For this study we consider the evolution of the solution from $t=0$ to $t=1$.
The first study focusses on the time convergence. For this we have used 1024 triangular elements and polynomial basis functions of degree $p=4$ and time steps equal to $\Delta t = 1, \frac{1}{2}, \frac{1}{4}, \frac{1}{8}, \frac{1}{16}$. As can be observed in [Figure \[fig:convergence\_plots\]]{} left, this method achieves a first order convergence rate, as opposed to a second order convergence rate of an implicit formulation, see [@Sanderse2013].
Regarding the spatial convergence we have tested the convergence rate of discretizations with basis functions of different polynomial degree, $p=1,2,4$. In order not to pollute the spatial convergence rate with the temporal integration error, we have used different time steps: $\Delta t = 2.5\times 10^{-2}$ for $p=1$, $\Delta t = 1.0\times 10^{-3}$ for $p=2$ and $\Delta t = 1.0\times 10^{-4}$ for $p=4$. As can be seen in [Figure \[fig:convergence\_plots\]]{}, this method has a convergence rate of $p$-th order for basis functions of polynomial degree $p$.
![Convergence plots of velocity error, $\|{\vec{u}}-{\vec{u}_{h}}\|_{L^{2}(\Omega)}$. Left: error in velocity as function of time step size $\Delta t$. Right: error in velocity as function of mesh size, $h$, for basis functions of polynomial degree $p=1,2,4$.[]{data-label="fig:convergence_plots"}](./figures/plots/convergence_plots/dt_convergence "fig:"){width="45.00000%"} ![Convergence plots of velocity error, $\|{\vec{u}}-{\vec{u}_{h}}\|_{L^{2}(\Omega)}$. Left: error in velocity as function of time step size $\Delta t$. Right: error in velocity as function of mesh size, $h$, for basis functions of polynomial degree $p=1,2,4$.[]{data-label="fig:convergence_plots"}](./figures/plots/convergence_plots/h_convergence "fig:"){width="45.00000%"}
Time reversibility and mass, energy, enstrophy, and vorticity conservation: inviscid shear layer roll-up {#subsec::inviscid_shear_layer_roll_up}
--------------------------------------------------------------------------------------------------------
The second group of tests focusses on the conservation properties of the proposed numerical method. For this set of tests we consider inviscid flow $\nu=0$. We consider here the simulation of the roll-up of a shear layer, see e.g. [@Minion1997; @Knikker2009; @Hokpunna2010; @Sanderse2013]. This solution is particularly challenging because during the evolution large vorticity gradients develop. Several methods are known to *blow up*, e.g. [@Minion1997]. We consider on the periodic domain $\Omega = [0,2\pi]\times[0,2\pi]$ the following initial conditions for velocity ${\vec{u}}$ $$u_{x}(x,y) =
\begin{dcases}
\tanh\left(\frac{y-\frac{\pi}{2}}{\delta}\right), & y\leq\pi, \\
\tanh\left(\frac{\frac{3\pi}{2}-y}{\delta}\right), & y>\pi,
\end{dcases}
\qquad\qquad\qquad
u_{y}(x,y) = \epsilon\sin(x),$$ and vorticity ${\omega}$ $$\omega(x,y) =
\begin{dcases}
\frac{1}{\delta}\,\text{sech}^2\left(\frac{y-\frac{\pi }{2}}{\delta }\right), & y\leq\pi, \\
-\frac{1}{\delta}\,\text{sech}^2\left(\frac{\frac{3 \pi }{2}-y}{\delta }\right), & y>\pi,
\end{dcases}$$ with $\delta = \frac{\pi}{15}$ and $\epsilon=0.05$ as in [@Knikker2009; @Sanderse2013].
The small perturbation $\epsilon$ in the $y$-component of the velocity field will trigger the roll-up of the shear layer. We show the contour lines of vorticity at $t=4$ and $t=8$, [Figure \[fig:vorticity\_evolution\_comparison\_plots\_contour\_1\]]{} and [Figure \[fig:vorticity\_evolution\_comparison\_plots\_contour\_2\]]{} respectively. The plots on the left side of [Figure \[fig:vorticity\_evolution\_comparison\_plots\_contour\_1\]]{} and [Figure \[fig:vorticity\_evolution\_comparison\_plots\_contour\_2\]]{} correspond to a spatial discretization of 6400 triangular elements of polynomial degree $p=1$ and time step size $\Delta t = 0.1$. On the right we present the results for a spatial discretization of 6400 triangular elements of polynomial degree $p=4$ and time step size $\Delta t = 0.01$. An example mesh with 1600 elements is shown in [Figure \[fig:mesh\]]{}.
![Example mesh with 1600 triangular elements.[]{data-label="fig:mesh"}](./figures/plots/mesh/plots/mesh.pdf){width="40.00000%"}
Although it is possible to observe oscillations in the vorticity plots, especially for $p=1$, none of them *blows up*, as is reported in [@Minion1997]. These oscillations necessarily appear whenever the flow generates structures at a scale smaller than the resolution of the mesh. Since energy and enstrophy are conserved, there is no possibility for the numerical simulation to dissipate the small scale information. This is a feature observed in all conserving inviscid solvers without a small scale model.
![Vorticity field of inviscid shear layer roll-up at $t=4$ obtained with 6400 triangular finite elements, contour lines ($\pm 1, \pm 2, \pm 3, \pm 4, \pm 5, \pm 6$) as in [@Knikker2009]. Left: solution with basis functions of polynomial degree $p=1$ and time step $\Delta t = 0.1$. Right: solution with basis functions of polynomial degree $p=4$ and time step $\Delta t = 0.01$.[]{data-label="fig:vorticity_evolution_comparison_plots_contour_1"}](./figures/plots/shear_layer_evolution/plots/r600/contour/contour_without_w_0/pdf/shear_layer_vorticity_contour_p_1_t_4_0.pdf "fig:"){width="45.00000%"} ![Vorticity field of inviscid shear layer roll-up at $t=4$ obtained with 6400 triangular finite elements, contour lines ($\pm 1, \pm 2, \pm 3, \pm 4, \pm 5, \pm 6$) as in [@Knikker2009]. Left: solution with basis functions of polynomial degree $p=1$ and time step $\Delta t = 0.1$. Right: solution with basis functions of polynomial degree $p=4$ and time step $\Delta t = 0.01$.[]{data-label="fig:vorticity_evolution_comparison_plots_contour_1"}](./figures/plots/shear_layer_evolution/plots/r600/contour/contour_without_w_0/pdf/shear_layer_vorticity_contour_p_4_t_4_0.pdf "fig:"){width="45.00000%"}
![Vorticity field of inviscid shear layer roll-up at $t=8$ obtained with 6400 triangular finite elements, contour lines ($\pm 1, \pm 2, \pm 3, \pm 4, \pm 5, \pm 6$) as in [@Knikker2009]. Left: solution with basis functions of polynomial degree $p=1$ and time step $\Delta t = 0.1$s. Right: solution with basis functions of polynomial degree $p=4$ and time step $\Delta t = 0.01$s.[]{data-label="fig:vorticity_evolution_comparison_plots_contour_2"}](./figures/plots/shear_layer_evolution/plots/r600/contour/contour_without_w_0/pdf/shear_layer_vorticity_contour_p_1_t_8_0.pdf "fig:"){width="45.00000%"} ![Vorticity field of inviscid shear layer roll-up at $t=8$ obtained with 6400 triangular finite elements, contour lines ($\pm 1, \pm 2, \pm 3, \pm 4, \pm 5, \pm 6$) as in [@Knikker2009]. Left: solution with basis functions of polynomial degree $p=1$ and time step $\Delta t = 0.1$s. Right: solution with basis functions of polynomial degree $p=4$ and time step $\Delta t = 0.01$s.[]{data-label="fig:vorticity_evolution_comparison_plots_contour_2"}](./figures/plots/shear_layer_evolution/plots/r600/contour/contour_without_w_0/pdf/shear_layer_vorticity_contour_p_4_t_8_0.pdf "fig:"){width="45.00000%"}
In order to assess the conservation properties of the proposed method we computed the evolution of the shear-layer problem from $t=0$ to $t=16$, using different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$. We have used a coarse grid with 6400 triangular finite elements and basis functions of polynomial degree $p=1$ for the spatial discretization. In [Figure \[fig:energy\_enstrophy\_conservation\_plots\]]{} left we show the evolution of the kinetic energy error with respect to the initial kinetic energy, $\frac{{\mathcal{K}}(0)-{\mathcal{K}_{h}}(t)}{{\mathcal{K}}(0)}$. This result confirms conservation of kinetic energy, as proven in [Section \[sec::kinetic\_energy\_conservation\]]{}. A similar study was performed for enstrophy in [Figure \[fig:energy\_enstrophy\_conservation\_plots\]]{} right. The magnitude of the error observed is of machine precision, verifying conservation of enstrophy, as proven in [Section \[sec::enstrophy\_conservation\]]{}. The evolution of the error of total vorticity is shown in [Figure \[fig:vorticity\_divergence\_conservation\_plots\]]{} left. The error of vorticity shows again an error of the order of machine precision. In [Figure \[fig:vorticity\_divergence\_conservation\_plots\]]{} right we show the divergence of the velocity field. As can be seen, this discretization results in a velocity field that is divergence free.
![Simulation of shear-layer problem with 6400 triangular finite elements, basis functions of polynomial degree $p=1$, and different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$. Left: kinetic energy error with respect to initial kinetic energy, $\frac{{\mathcal{K}}(0)-{\mathcal{K}_{h}}(t)}{{\mathcal{K}}(0)}$. Right: enstrophy error with respect to initial enstrophy, $\frac{{\mathcal{E}}(0)-{\mathcal{E}_{h}}(t)}{{\mathcal{E}}(0)}$.[]{data-label="fig:energy_enstrophy_conservation_plots"}](./figures/plots/conservation_meev_plots/plots/energy_conservation_1 "fig:"){width="40.00000%"} ![Simulation of shear-layer problem with 6400 triangular finite elements, basis functions of polynomial degree $p=1$, and different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$. Left: kinetic energy error with respect to initial kinetic energy, $\frac{{\mathcal{K}}(0)-{\mathcal{K}_{h}}(t)}{{\mathcal{K}}(0)}$. Right: enstrophy error with respect to initial enstrophy, $\frac{{\mathcal{E}}(0)-{\mathcal{E}_{h}}(t)}{{\mathcal{E}}(0)}$.[]{data-label="fig:energy_enstrophy_conservation_plots"}](./figures/plots/conservation_meev_plots/plots/enstrophy_conservation_1 "fig:"){width="40.00000%"}
![Simulation of shear-layer problem with 6400 triangular finite elements, basis functions of polynomial degree $p=1$, and different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$. Left: total vorticity error with respect to initial total vorticity, ${\mathcal{W}}(0)-{\mathcal{W}_{h}}(t)$. Right: evolution of the divergence of the velocity field.[]{data-label="fig:vorticity_divergence_conservation_plots"}](./figures/plots/conservation_meev_plots/plots/vorticity_conservation_1 "fig:"){width="40.00000%"} ![Simulation of shear-layer problem with 6400 triangular finite elements, basis functions of polynomial degree $p=1$, and different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$. Left: total vorticity error with respect to initial total vorticity, ${\mathcal{W}}(0)-{\mathcal{W}_{h}}(t)$. Right: evolution of the divergence of the velocity field.[]{data-label="fig:vorticity_divergence_conservation_plots"}](./figures/plots/conservation_meev_plots/plots/mass_conservation_1 "fig:"){width="40.00000%"}
In order to assess the conservation properties of the proposed method for long time simulations we computed the evolution of the shear-layer problem from $t=0$ to $t=128$, using different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$. We have used a very coarse grid with 100 triangular finite elements and basis functions of polynomial degree $p=4$ for the spatial discretization. In [Figure \[fig:energy\_enstrophy\_conservation\_plots\_long\_time\]]{} left we show the evolution of the kinetic energy error with respect to the initial kinetic energy, $\frac{{\mathcal{K}}(0)-{\mathcal{K}_{h}}(t)}{{\mathcal{K}}(0)}$. This result confirms conservation of kinetic energy, as proven in [Section \[sec::kinetic\_energy\_conservation\]]{}. The same study was performed for enstrophy in [Figure \[fig:energy\_enstrophy\_conservation\_plots\_long\_time\]]{} right. The error observed is very small, verifying conservation of enstrophy, as proven in [Section \[sec::enstrophy\_conservation\]]{}. The evolution of error for total vorticity is shown in [Figure \[fig:vorticity\_divergence\_conservation\_plots\_long\_time\]]{} left. The error of total vorticity shows that this method is capable of conserving the total vorticity on long time simulations. In [Figure \[fig:vorticity\_divergence\_conservation\_plots\_long\_time\]]{} right we show the divergence of the velocity field. As can be seen, this discretization results in a velocity field that is divergence free. Similar results are obtained for $p=1$.
![Simulation of shear-layer problem with 100 triangular finite elements, basis functions of polynomial degree $p=4$, and different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$. Left: kinetic energy error with respect to initial kinetic energy, $\frac{{\mathcal{K}}(0)-{\mathcal{K}_{h}}(t)}{{\mathcal{K}}(0)}$. Right: enstrophy error with respect to initial enstrophy, $\frac{{\mathcal{E}}(0)-{\mathcal{E}_{h}}(t)}{{\mathcal{E}}(0)}$.[]{data-label="fig:energy_enstrophy_conservation_plots_long_time"}](./figures/plots/conservation_meev_plots/plots/long_time_simulation_p_4/energy_conservation_1 "fig:"){width="40.00000%"} ![Simulation of shear-layer problem with 100 triangular finite elements, basis functions of polynomial degree $p=4$, and different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$. Left: kinetic energy error with respect to initial kinetic energy, $\frac{{\mathcal{K}}(0)-{\mathcal{K}_{h}}(t)}{{\mathcal{K}}(0)}$. Right: enstrophy error with respect to initial enstrophy, $\frac{{\mathcal{E}}(0)-{\mathcal{E}_{h}}(t)}{{\mathcal{E}}(0)}$.[]{data-label="fig:energy_enstrophy_conservation_plots_long_time"}](./figures/plots/conservation_meev_plots/plots/long_time_simulation_p_4/enstrophy_conservation_1 "fig:"){width="40.00000%"}
![Simulation of shear-layer problem with 100 triangular finite elements, basis functions of polynomial degree $p=4$, and different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$. Left: total vorticity error with respect to initial total vorticity, ${\mathcal{W}}(0)-{\mathcal{W}_{h}}(t)$. Right: evolution of the divergence of the velocity field.[]{data-label="fig:vorticity_divergence_conservation_plots_long_time"}](./figures/plots/conservation_meev_plots/plots/long_time_simulation_p_4/vorticity_conservation_1 "fig:"){width="40.00000%"} ![Simulation of shear-layer problem with 100 triangular finite elements, basis functions of polynomial degree $p=4$, and different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$. Left: total vorticity error with respect to initial total vorticity, ${\mathcal{W}}(0)-{\mathcal{W}_{h}}(t)$. Right: evolution of the divergence of the velocity field.[]{data-label="fig:vorticity_divergence_conservation_plots_long_time"}](./figures/plots/conservation_meev_plots/plots/long_time_simulation_p_4/mass_conservation_1 "fig:"){width="40.00000%"}
The final test addresses time reversibility. In order to investigate this property of the solver, we let the flow evolve from $t=0$ to $t=8$ and then reversed the time evolution and evolved again for the same time, corresponding to the evolution from $t=8$ to $t=0$. We have performed this study with 6400 triangular finite elements of polynomial degree $p=1$ and different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$. In [Figure \[fig:time\_reversibility\_plots\]]{} we show the error between the initial vorticity field and the final vorticity field. As can be seen, the errors are in the order of $10^{-12}$, showing the reversibility of the method up to machine accuracy.
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Error between the initial vorticity field and the vorticity field computed by evolving the inviscid Navier-Stokes equations to $t=8$s and then reversing the time step until $t=0$s is reached again. The results shown correspond to a mesh of 6400 triangular finite elements, polynomial degree $p=1$, and different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$ (starting from the bottom right and going clockwise).[]{data-label="fig:time_reversibility_plots"}](./figures/plots/reversibility_plots/plots/shear_layer_reversibility_pcolor_dt_0_125.png "fig:"){width="45.00000%"}
\#2
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Error between the initial vorticity field and the vorticity field computed by evolving the inviscid Navier-Stokes equations to $t=8$s and then reversing the time step until $t=0$s is reached again. The results shown correspond to a mesh of 6400 triangular finite elements, polynomial degree $p=1$, and different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$ (starting from the bottom right and going clockwise).[]{data-label="fig:time_reversibility_plots"}](./figures/plots/reversibility_plots/plots/shear_layer_reversibility_pcolor_dt_0_250.png "fig:"){width="45.00000%"}
\#2
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Error between the initial vorticity field and the vorticity field computed by evolving the inviscid Navier-Stokes equations to $t=8$s and then reversing the time step until $t=0$s is reached again. The results shown correspond to a mesh of 6400 triangular finite elements, polynomial degree $p=1$, and different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$ (starting from the bottom right and going clockwise).[]{data-label="fig:time_reversibility_plots"}](./figures/plots/reversibility_plots/plots/shear_layer_reversibility_pcolor_dt_0_500.png "fig:"){width="45.00000%"}
\#2
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Error between the initial vorticity field and the vorticity field computed by evolving the inviscid Navier-Stokes equations to $t=8$s and then reversing the time step until $t=0$s is reached again. The results shown correspond to a mesh of 6400 triangular finite elements, polynomial degree $p=1$, and different time steps $\Delta t = 1,\frac{1}{2}, \frac{1}{4},\frac{1}{8}$ (starting from the bottom right and going clockwise).[]{data-label="fig:time_reversibility_plots"}](./figures/plots/reversibility_plots/plots/shear_layer_reversibility_pcolor_dt_1_000.png "fig:"){width="45.00000%"}
\#2
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Conclusions {#sec::conclusions}
===========
This paper presents a new arbitrary order finite element solver for the Navier-Stokes equations. The advantages of this method are: (i) high order character of the spatial discretization, (ii) geometric flexibility, allowing for unstructured triangular grids, (iii) exact preservation of conservation laws for mass, kinetic energy, enstrophy and vorticity, and (iv) time reversibility.
The construction of a numerical flow solver with these properties leads to a very robust method capable of performing very under-resolved simulations without the characteristic blow up of standard discretizations.
For a simpler Taylor-Green analytical solution, the convergence properties of the proposed method have been shown. With the more challenging shear layer roll-up test case, we have shown the robustness of the method. Conservation of kinetic energy, enstrophy and vorticity up to machine accuracy was shown for this test case, even for low order approximations.
Nevertheless, the proposed method still allows further improvement, mainly in terms of time integration. One of the advantages of the method is the fact that the time integration method avoids the solution of a fully coupled nonlinear system of equations. Instead two quasi-linear systems of equations are solved, without the need for expensive iterations for each time step. The downside of this approach is that, for now, the time integration convergence rate is limited to first order. In the future we intend to compare this approach to a fully implicit formulation using higher order Gauss integration. Another aspect which is currently being investigated is the treatment of solid boundaries.
\#1
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study three dimensional $O(N)_k$ and $U(N)_k$ Chern-Simons theories coupled to a scalar field in the fundamental representation, in the large $N$ limit. For infinite $k$ this is just the singlet sector of the $O(N)$ ($U(N)$) vector model, which is conjectured to be dual to Vasiliev’s higher spin gravity theory on $AdS_4$. For large $k$ and $N$ we obtain a parity-breaking deformation of this theory, controlled by the ’t Hooft coupling $\lambda = 4 \pi N / k$. For infinite $N$ we argue (and show explicitly at two-loop order) that the theories with finite $\lambda$ are conformally invariant, and also have an exactly marginal $(\phi^2)^3$ deformation. For large but finite $N$ and small ’t Hooft coupling $\lambda$, we show that there is still a line of fixed points parameterized by the ’t Hooft coupling $\lambda$. We show that, at infinite $N$, the interacting non-parity-invariant theory with finite $\lambda$ has the same spectrum of primary operators as the free theory, consisting of an infinite tower of conserved higher-spin currents and a scalar operator with scaling dimension $\Delta=1$; however, the correlation functions of these operators do depend on $\lambda$. Our results suggest that there should exist a family of higher spin gravity theories, parameterized by $\lambda$, and continuously connected to Vasiliev’s theory. For finite $N$ the higher spin currents are not conserved.'
author:
- |
Ofer Aharony, Guy Gur-Ari, and Ran Yacoby\
\
[*Department of Particle Physics and Astrophysics*]{}\
[*Weizmann Institute of Science, Rehovot 76100, Israel*]{}\
title: ' $d=3$ Bosonic Vector Models Coupled to Chern-Simons Gauge Theories'
---
Introduction
============
The $AdS$/CFT correspondence [@Maldacena:1997re] is an exact duality between quantum gravitational theories on space-times that include anti-de Sitter space $AdS_{d+1}$, and conformal field theories in $d$ space-time dimensions. This correspondence has many applications, and it has taught us a lot about strongly coupled field theories and about quantum gravity. However, while we know how to translate computations on one side of the duality to the other side, we do not yet have a derivation of the $AdS$/CFT correspondence, that would enable us in particular to know which quantum gravity theory is dual to a given conformal field theory, and vice versa. Finding such a derivation is complicated by the fact that in most examples, either one or both sides of the correspondence are strongly coupled. This is partly because the gravitational dual of any weakly coupled field theory must include light fields of arbitrarily high spin.
There is one example of the $AdS$/CFT correspondence in which both sides are weakly coupled in the large $N$ limit; this is the conjectured duality [@Sundborg:2000wp; @Sezgin:2002rt; @Klebanov:2002ja] between the singlet sector of the $O(N)$ vector model (namely, $N$ free real scalar fields) in three space-time dimensions, and Vasiliev’s higher-spin gravity theory on $AdS_4$ [@Fradkin:1987ks] (see [@Vasiliev:1999ba] for a review). While the gravitational side of this duality is only understood at the classical level, and it is not yet known how to give it a quantum completion, in the classical gravity limit (governed by tree-level diagrams in the bulk) this provides an example of the $AdS$/CFT correspondence in which both sides are weakly coupled. This allows many detailed tests of the correspondence to be performed in this case [@Giombi:2009wh; @Giombi:2010vg; @Giombi:2011ya], and it also suggests that this could be an ideal toy model for which a derivation of the $AdS$/CFT correspondence could be found (and perhaps then generalized to more complicated cases). Indeed, there are several suggestions in the literature [@Das:2003vw; @Koch:2010cy; @Douglas:2010rc; @Jevicki:2011ss] for how to derive the $AdS$/CFT correspondence explicitly in this example.
In this paper we study a small deformation of the duality above, on the field theory side; it should be possible to map any such deformation to the gravity side as well, and to utilize the extra structure that it provides to learn more about the explicit $AdS$/CFT mapping in this case. A simple way to obtain a theory that contains only the singlet sector of the $O(N)$ vector model is by coupling $N$ free scalar fields to an $O(N)$ gauge theory; since we do not want to add any dynamics of the gauge field, we should not have standard kinetic terms for the gauge fields, but we can view their action as the $k\to \infty$ limit of the $O(N)_k$ Chern-Simons gauge theory [@Giombi:2009wh].[^1] It is then natural to deform the theory by making the Chern-Simons level $k$ a finite integer; this theory has a ’t Hooft limit, controlled by a ’t Hooft coupling $\lambda \equiv 4 \pi N / k$, and at large $N$ this gives a continuous parity-breaking deformation of the original theory. On the field theory side one can then perform computations in perturbation theory in $\lambda$, and it should be possible to translate these into perturbative computations also on the gravity side, and to obtain a more detailed weak-weak coupling duality. We will consider both the $O(N)$ case with a real scalar field in the fundamental representation, and the $U(N)$ case with a complex scalar. These theories were previously studied perturbatively in [@Chen:1992ee; @Avdeev:1992jt].
We begin in section \[On\] by introducing our action and our methods of regularizing and renormalizing it. In section \[confSym\] we study whether the theory at small $\lambda$ is still conformally invariant. The Chern-Simons level is quantized and does not run [@Deser:1981wh; @Deser:1982vy]. One problem that may arise whenever we have interactions is that relevant operators of the form $\phi^2$ and $(\phi^2)^2$ (where $\phi$ is the scalar field) may be generated, even if they are tuned to zero at some scale. However, in our renormalization scheme these couplings do not run away if they are initially set to zero. A more serious problem is that this theory has a classically marginal $\lambda_6 (\phi^2)^3 / N^2$ coupling, which could start running once we turn on $\lambda$. However, we provide an argument (and check explicitly at two-loop order) that at infinite $N$ the beta function for this coupling vanishes. Therefore, there is a two-dimensional family of large $N$ conformal field theories, parameterized by $\lambda$ and by $\lambda_6$. For large but finite $N$ we show that a beta function for $\lambda_6$ is generated, but that (at least) for small $\lambda$ this beta function still has an IR-stable fixed point, so that there still exists a one-parameter family of conformal field theories, parameterized by $\lambda$. Note that while $\lambda$ is a discrete parameter for finite $N$, it is almost continuous when $N$ is very large.
In section \[HScurrents\] we analyze the spectrum of the large $N$ family of conformal field theories that we find, and show that it is independent of $\lambda$ (and thus identical to that of the free theory with $\lambda=0$). In particular, conserved higher-spin currents still exist for infinite $N$ and any $\lambda$, though the corresponding symmetries are broken for finite $N$. Such an appearance of an infinite number of conserved currents in an interacting theory is quite surprising, and this could lead one to suspect that the theories we discuss may be independent of $\lambda$ in the large $N$ limit. In section \[conjecture\] we show that this is not the case, by computing a correlation function in these theories at leading order in $\lambda$ (in the large $N$ limit) and showing that it depends on $\lambda$. We end in section \[summary\] with a summary of our results and a discussion of some future directions.
The $O(N)$ Model with Chern-Simons Interactions {#On}
===============================================
Consider the theory of a real scalar field $\phi$ in the fundamental representation of $O(N)$, coupled to gauge bosons $A_\mu$ with Chern-Simons interactions at level $k$ in three Euclidean dimensions (the generalization to $N$ complex scalar fields coupled to a $U(N)_k$ Chern-Simons theory is straightforward, and we will occasionally discuss below this case as well). We regulate the theory using dimensional reduction [@Siegel:1979wq] (see below), and work in Lorenz gauge (Landau gauge), $\partial^{\mu}A_\mu=~0$. The regularized action in terms of the renormalized fields and couplings is $$\begin{aligned}
S &= S_\mathrm{CS} + S_\mathrm{gh} + S_\mathrm{b}{\,,}\label{eq:action}\\
S_\mathrm{CS} &= \int \! d^dx\, \left\{
- \frac{i}{2} Z_A
\epsilon_{\mu\nu\lambda} A_\mu^a \partial_\nu A_\lambda^a
- \frac{i}{6} \mu^{\epsilon/2} g Z_g \epsilon_{\mu\nu\lambda} f^{abc}
A_\mu^a A_\nu^b A_\lambda^c
\right\}{\,,}\label{eq:actionCS}\\
S_\mathrm{gh} &= \int \! d^dx\, \left\{
- \frac{1}{2 \gamma_R} (\partial_\mu A_\mu^a)^2
+ Z_\mathrm{gh} \partial_\mu \bar{c}^a \partial^\mu c^a
+ \mu^{\epsilon/2} \tilde{Z}_g g f^{abc}
\partial_\mu \bar{c}^a A_\mu^b c^c
\right\}{\,,}\label{eq:actionGH}\\
S_\mathrm{b} &= \int \! d^dx\, \left\{
\frac{1}{2} Z_\phi (\partial_\mu \phi_i)^2
+ \mu^{\epsilon/2} Z_g' g \partial_\mu \phi_i T_{ij}^a
A_\mu^a \phi_j
- \frac{1}{4} \mu^\epsilon Z_g'' g^2 \{T^a,T^b\}_{ij}
\phi_i \phi_j A_\mu^a A_\mu^b
\right. \notag\\
&\quad \qquad \qquad \left.
+ \mu^{2\epsilon} Z_{g_6}\frac{g_6}{3!\cdot 2^3}(\phi_i\phi_i)^3
\right\} {\,,}\label{eq:actionB}\end{aligned}$$ where $d=3-\epsilon$, and $\mu$ is the renormalization scale (for additional conventions, see Appendix \[conventions\]). The coupling $g$ is related to the integer Chern-Simons level $k$ by $k = 4\pi/g^2$. When taking the ’t Hooft large $N$ limit, the couplings $\lambda = g^2 N$ and $\lambda_6 = g_6 N^2$ are held fixed, and in this limit $\lambda$ becomes a continuous parameter. Note that while parity is broken due to the Chern-Simons interaction, the theory is dual under the combined transformation of parity plus $\lambda\to-\lambda$ ($k\to -k$), and physical results must be invariant under this transformation.
Once $\lambda > 0$, in order to renormalize the theory in a generic scheme we must add also two relevant interactions that will be generated by quantum corrections: a mass term $(\phi_i \phi_i)$ and an interaction of the form $g_4 (\phi_i \phi_i)^2$. We are interested in interacting conformal fixed points of our field theory, so we will generally fine-tune our couplings so that the physical mass and $\phi^4$ couplings vanish, and then for the purposes of our computations we can just ignore these terms. In fact, in the scheme we are using (of dimensional reduction and minimal subtraction), once we fix the renormalized dimensionful couplings to zero, they remain zero so we do not even have to add them to our action.
At least in the large $N$ limit, we could also study the theory in which the coupling $g_4$ does not vanish; if it is non-zero then the theory flows to another fixed point, which at large $N$ is closely related to the original fixed point (at infinite $N$ it has the same spectrum of operators, except for the operator $\phi_i \phi_i$ whose dimension at the interacting fixed point is $\Delta=2$). For the theory with $\lambda=0$ this was discussed in the $AdS$/CFT context in [@Klebanov:2002ja; @Gubser:2002vv; @Petkou:2003zz; @Giombi:2011ya], and the same analysis holds also at finite $\lambda$. Therefore, most of our results also apply to the “critical” fixed point with a non-zero $g_4$ coupling. However, for simplicity, we will focus here on the case where the physical $g_4$ coupling is tuned to vanish.
Note that dimensional regularization of this theory is subtle, since the 3-form integration of the Chern-Simons interaction (\[eq:actionCS\]) is not well-defined for arbitrary dimension. To regulate loop integrals we first perform the tensor algebra in 3 dimensions, and then compute the resulting scalar integral in $d=3-\epsilon$ dimensions. This method, known as dimensional reduction [@Siegel:1979wq], has been shown in [@Chen:1992ee] to preserve gauge-invariance in this theory at least up to two-loop order.
Conformal Symmetry {#confSym}
==================
In this section we analyze the conditions under which the theory defined by is conformal, both for finite and for infinite $N$. The Chern-Simons level $k$ is quantized to be an integer and is therefore not renormalized, except perhaps by an integer shift at one-loop order; this has been verified explicitly in [@Chen:1992ee]. The corresponding one-loop shift in $\lambda$ is of order $1/N$ in the ’t Hooft large-$N$ limit that we study here, so we will ignore it. However, the classically-marginal $\lambda_6$ coupling may receive corrections.[^2] In order to check for conformal fixed points we need to compute its beta function $\beta_{\lambda_6}(\lambda,\lambda_6)$, and show that it vanishes. In section \[2loop\] we compute this beta function at the first non-trivial order, by computing the divergent contributions to the amputated correlator $$\begin{aligned}
\label{sixpoint}
\left< \phi^{i_1}(x_1) \cdots \phi^{i_6}(x_6) \right>_{\mathrm{amp.}} {\,.}\end{aligned}$$ We might expect that solving $\beta_{\lambda_6}(\lambda,\lambda_6)=0$ would result in a line of fixed points in the $(\lambda,\lambda_6)$ plane. For large and finite $N$ we indeed find two such lines; however, at infinite $N$ we find that $\lambda$ and $\lambda_6$ are both exactly marginal at 2-loops. In section \[allorders\] we show that this is actually true to all orders in perturbation theory, so that at infinite $N$ there is a family of conformal field theories labeled by continuous parameters $\lambda$ and $\lambda_6$. In section \[sponbreak\] we argue that there is no spontaneous breaking of the conformal symmetry in our theories.
The Beta Function $\beta_{\lambda_6}$ at Two Loops {#2loop}
--------------------------------------------------
In this section we compute $\beta_{\lambda_6}$ in momentum space using minimal subtraction. In our theory, using our dimensional reduction regularization, all 1-loop integrals are finite. Indeed, for quadratic and logarithmic divergences in three dimensions, the numerator must be an odd power of the loop momentum $q$, which must be of the form $q^\mu q^{2n}$, and then the $q$ integral vanishes by the $q\to -q$ symmetry. Linear divergences are rendered finite by dimensional regularization. In the specific case of $\beta_{\lambda_6}$, a one-loop contribution is also not allowed by the parity transformation.
Therefore, the leading contribution to this beta function arises at 2-loop order. The $\left<\phi^6\right>$ correlator (\[sixpoint\]) is superficially log-divergent, with over 50 two-loop diagrams contributing to it in the planar limit alone. However, the number of diagrams that may contribute to its divergence is greatly reduced by the following observation. Consider a diagram that includes a $\phi A^\mu \partial_\mu \phi$ vertex, with the gluon carrying a loop momentum $q$ and one of the scalar lines carrying an external momentum $p$. In the numerator we then have $\epsilon_{\mu\nu\rho} q^\rho$ from the gluon propagator (\[eq:GluonProp\]) and $(q+2p)^\mu$ from the vertex, and the leading high-energy term of order $q^2$ cancels by antisymmetry. Therefore, in such a situation the degree of divergence is reduced and the diagram is finite. As a result, the only diagrams that can contribute to $\beta_{\lambda_6}$ at 2-loop order are the following:
---------------------------------- ---------------------------------- ---------------------------------- ----------------------------------
{height="3.5cm"} {height="3.5cm"} {height="3.5cm"} {height="3.5cm"}
(A1) (A2) (A3) (A4)
{height="3.5cm"} {height="3.5cm"} {height="3.5cm"} {height="3.5cm"}
(A5) (A6) (A7) (A8)
---------------------------------- ---------------------------------- ---------------------------------- ----------------------------------
The diagrams (A1-3) include planar diagrams, while the others are suppressed by powers of $1/N$ in the ’t Hooft large $N$ limit. In order to compute the 2-loop beta function, we need in addition to the diagrams above also the anomalous dimension of the scalar field at this order. This comes from the following diagrams:[^3]
--------------------------------- --------------------------------- --------------------------------- ---------------------------------
{width="3.7cm"} {width="3.7cm"} {width="3.7cm"} {width="3.7cm"}
(B1) (B2) (B3) (B4)
--------------------------------- --------------------------------- --------------------------------- ---------------------------------
The divergent parts of all the diagrams above are listed in Appendix \[diags\]. By summing these we can determine the renormalization constants, $$\begin{aligned}
Z_{\phi} &\equiv 1 + \sum_i(Bi) = 1 - \frac{g^4\left(3N^2-23N+20\right)}{384\pi^2}\frac{1}{\epsilon}{\,,}\label{eq:Zphi}\\
g_6Z_{g_6} &\equiv g_6 + \sum_i(Ai) = g_6 + \frac{66g^8\left(N-1\right) + 4g_6^2\left(3N+22\right) - 3g^4g_6\left(N^2+19N-20\right)}{128\pi^2}\frac{1}{\epsilon}{\,.}\label{eq:Zg6}\end{aligned}$$ The bare sextic coupling $g_{6,0} = \mu^{2\epsilon} g_6 Z_{g_6} / Z_{\phi}^3$ may thus be written in the form $g_{6,0} = g_6 + b_1(g,g_6) / \epsilon + (\mathrm{other~terms})$, where $$\begin{aligned}
b_1(g,g_6) = \frac{33 g^8(N-1) - 40 g^4 g_6 (N-1) + 2 g_6^2 (3 N + 22)}{64 \pi^2}{\,.}\end{aligned}$$ The beta function for the $\lambda_6$ coupling is related to the single pole in dimensional regularization by $\beta_{g_6} = - 2 b_1 + 2 g_6 \partial_{g_6} b_1 + {1\over 2} g \partial_g b_1$ [@Weinberg:1996kr], leading to $$\begin{aligned}
\label{eq:Full2LoopBeta6}
\beta_{\lambda_6}(\lambda,\lambda_6) = \frac{33 (N-1) \lambda ^4-40 (N-1) \lambda ^2\lambda_6+2 (3 N + 22) \lambda_6^2}{32 N^2 \pi ^2} {\,.}\end{aligned}$$
In the ’t Hooft large $N$ limit we see that $\beta_{\lambda_6}=0$, so that both the $(\phi^2)^3$ coupling and the Chern-Simons interaction are marginal at this order. For the theory with only $(\phi^2)^3$ couplings it is easy to see that the large $N$ beta function vanishes to all orders, since there are no contributing diagrams; it is indeed well-known that this coupling is exactly marginal in the large $N$ limit [@Bardeen:1983rv] (see [@Moshe:2003xn] for a review). However, for finite $\lambda$ there do exist divergent planar diagrams. The vanishing of the $\lambda^4$ term in at this order is due (in our gauge choice) to a non-trivial cancelation between the diagrams (A2) and (A3). There is also a large $N$ divergence proportional to $\lambda^2 \lambda_6$ arising from (A1), that exactly cancels in the planar limit with the similar contribution from the anomalous dimension of $\phi^i$. In fact, one can show that, at large $N$, contributions to the beta function can have either zero or one $(\phi^2)^3$ vertices, and that the planar diagrams contributing with a single $(\phi^2)^3$ vertex are the same as the diagrams contributing to the anomalous dimension of $\phi^2$. Thus, the large $N$ beta function takes the form $$\beta_{\lambda_6}(\lambda,\lambda_6) = b\, \gamma_{\phi^2}(\lambda)\lambda_6 + f(\lambda) + O(1/N),$$ where $\gamma_{\phi^2}$ is the anomalous dimension of $\phi^2$ and $b$ is a constant. In the next subsection we argue that both this anomalous dimension and the beta function $\beta_{\lambda_6}$ vanish in the large $N$ limit, so that the couplings $\lambda$ and $\lambda_6$ are both exactly marginal in this limit.
At finite but large $N$ the beta function does not vanish. Without the coupling $\lambda$, the beta function is positive so the theory with $\lambda_6 > 0$ is trivial (IR-free). However, when $\lambda \neq 0$ and for large $N \geq 10$, we find from two lines of non-trivial fixed points of the two-loop beta function, $$\begin{aligned}
\lambda_{6}^\pm(\lambda) &=
\frac{\left(20 N - 20 \pm \sqrt{1852-2054 N+202 N^2} \right) \lambda^2}
{44+6 N} {\,.}\end{aligned}$$ The line $\lambda_6^+(\lambda)$ is IR-stable, while $\lambda_6^-(\lambda)$ is UV-stable – see Figure 1. Note that since $\beta_\lambda=0$, the renormalization group flow is always in the $\lambda_6$ direction.
\[fig:RG\] {width="60.00000%"}
The Large $N$ Beta Function $\beta_{\lambda_6}$ to All Orders {#allorders}
-------------------------------------------------------------
In this section we argue that $\beta_{\lambda_6}(\lambda,\lambda_6) = O(1/N)$ to all orders in perturbation theory, generalizing our explicit two-loop computation of the previous subsection. We could not find a direct argument for this, so instead we will use a trick. We focus on the $U(N)$ vector model for simplicity, but the argument can be generalized to $O(N)$ as well.
Consider the ${\ensuremath{\mathcal{N}}}=2$ supersymmetric generalization of our theory, which is the ${{\ensuremath{\mathcal{N}}}=2}$ supersymmetric Chern-Simons $U(N)$ gauge theory, coupled to a single matter chiral superfield $\Phi_i$ with components $(\phi_i,\psi_i)$ in the fundamental representation (we use $i,j = 1,\ldots,N$ to label the fundamental representation of $U(N)$).[^4] We will relate $\beta_{\lambda_6}$ in our theory (for infinite $N$) to the beta function of the $(\phi^\dagger\phi)^3$ coupling in the ${\ensuremath{\mathcal{N}}}=2$ theory. The action of the ${\ensuremath{\mathcal{N}}}=2$ theory [@Schwarz:2004yj; @Gaiotto:2007qi], after integrating out all the auxiliary fields, is $$\begin{aligned}
\label{eq:SUSicCS}
S^{{\ensuremath{\mathcal{N}}}=2}_{CS} &= -\frac{ik}{4\pi}\int \operatorname{Tr}\left[A\wedge dA + \frac{2}{3}A^3\right] + \notag\\
&\quad \int \! d^3x \, \left[ |{\ensuremath{\mathcal{D}}}_{\mu}\phi_i|^2 +i\bar{\psi}^i\gamma^{\mu}{\ensuremath{\mathcal{D}}}_{\mu}\psi_i - 2\frac{\lambda}{N}\bar{\phi}^i\phi_i\bar{\psi}^j\psi_j - \frac{\lambda}{N}\bar{\phi}^i\phi_j\bar{\psi}^j\psi_i - \frac{\lambda^2}{N^2}(\bar{\phi}^i\phi_i)^3 \right]
{\,.}\end{aligned}$$ It was shown in [@Gaiotto:2007qi] that this action is exactly conformal quantum mechanically, for all values of $k$ and $N$ (with $\lambda=4\pi N/k$). In particular, this means that the beta function of the $(\phi^\dagger\phi)^3$ coupling in this theory vanishes identically to all orders in $\lambda$ and $1/N$. The theory has a global $U(1)_f$ symmetry acting on the matter superfield as $\Phi \to e^{i\alpha}\Phi$, and in [@Gaiotto:2007qi] it was noticed that the operators ${\ensuremath{\mathcal{O}}}_1 = \bar{\phi}^i\phi_i$ and ${\ensuremath{\mathcal{O}}}_2 = \bar{\psi}^i\psi_i + \frac{4\pi}{k}(\bar{\phi}^i\phi_i)^2$ sit in the same supermultiplet as the $U(1)_f$ symmetry current.[^5] As a consequence, the dimensions of ${\ensuremath{\mathcal{O}}}_1$ and ${\ensuremath{\mathcal{O}}}_2$ are protected to be 1 and 2 respectively. The “double-trace” term in ${\ensuremath{\mathcal{O}}}_2$ does not contribute to the 2-point function $\langle{\ensuremath{\mathcal{O}}}_2(x){\ensuremath{\mathcal{O}}}_2(y)\rangle$ to leading order in $1/N$, and therefore the operator $\bar{\psi}^i\psi_i$ by itself is also protected at large $N$, with dimension $\Delta=2+O(1/N)$.
Let us begin by arguing that the anomalous dimension of ${\ensuremath{\mathcal{O}}}_1$ vanishes at large $N$ also in our non-supersymmetric theory. Consider the diagrams that contribute to $\left< {\ensuremath{\mathcal{O}}}_1 {\ensuremath{\mathcal{O}}}_1 \right>$ and $\left< {\ensuremath{\mathcal{O}}}_2 {\ensuremath{\mathcal{O}}}_2 \right>$ in the ${\ensuremath{\mathcal{N}}}=2$ theory and involve a single matter loop, with possible additional gluon lines. We will denote them collectively as
---------------------------------- ----------------------------------
{height="3.5cm"} {height="3.5cm"}
(E1) (E2)
---------------------------------- ----------------------------------
For the rest of the section we will keep gluon lines implicit in all diagrams; at large $N$ when we draw the diagrams in double-line notation these lines must sit inside the scalar/fermion loop so that the topology of each scalar/fermion loop is that of a disk. We will show below that in the large $N$ limit the sum of such diagrams at a given order in perturbation theory is finite. However, the diagrams (E1) (with gluon lines running in the loop) are precisely those that contribute to the correlator $\left< {\ensuremath{\mathcal{O}}}_1 {\ensuremath{\mathcal{O}}}_1 \right>$ in our ${\ensuremath{\mathcal{N}}}=0$ model in the large $N$ limit.[^6] Thus, it will follow that the dimension of ${\ensuremath{\mathcal{O}}}_1 = \bar{\phi}^i \phi_i$ in our non-supersymmetric vector model is $1+O(1/N)$ to all orders in planar perturbation theory.
We now prove the finiteness of (E1) and (E2) at large $N$ by induction. At zeroth order in perturbation theory, (E1) and (E2) are single 1-loop diagrams which are finite in our regularization scheme. At the next order the only diagrams contributing to the two-point functions in the ${\ensuremath{\mathcal{N}}}=2$ theory are still of the form (E1) and (E2) (with an extra gluon line), so all divergences in these diagrams must cancel (in fact, it follows from the parity transformation that these diagrams vanish). At higher orders in perturbation theory, there are more general diagrams contributing to $\left< {\ensuremath{\mathcal{O}}}_1 {\ensuremath{\mathcal{O}}}_1 \right>$ at large $N$, which have the general form:
----------------------------------
{height="3.5cm"}
(F1)
----------------------------------
Again, gluon lines running inside the loops are implicit, and all the other diagrams (not drawn in (F1)) contain tadpole matter loops causing them to vanish. We know that in the ${\ensuremath{\mathcal{N}}}=2$ theory the sum of all these (F1) diagrams, with any (odd) number of matter loops, is finite, since ${\ensuremath{\mathcal{O}}}_1$ is not renormalized. Working in momentum space, each (F1) diagram factorizes at large $N$ into a product of sub-diagrams of the form (E1) or (E2). If a given (F1) diagram has more than one matter loop, its sub-diagrams will be of a lower order in perturbation theory. The sum over such sub-diagrams is finite by the induction assumption, and therefore (F1) diagrams with more than one matter loop are finite in total. Since the sum over all (F1) diagrams is also finite, the sum over single-matter-loop diagrams — which are the (E1) diagrams at the order we are in — must be finite. This concludes the induction step for (E1); the step for (E2) is analogous.[^7] In appendix \[adim\] we verify that indeed the anomalous dimension of $\phi^2$ vanishes in the non-supersymmetric theory at two-loop order in the large $N$ limit.
The argument above can be easily generalized to diagrams of the topology (E1), which have three insertions of ${\ensuremath{\mathcal{O}}}_1$ on the scalar loop instead of two. Namely, the sum of such diagrams is also finite (in the large $N$ limit) in the non-supersymmetric theory at a given order in perturbation theory. To see this consider the correlator $\left< ({\ensuremath{\mathcal{O}}}_1)^3 \right>$ in the ${\ensuremath{\mathcal{N}}}=2$ theory, which does not contain divergences since both ${\ensuremath{\mathcal{O}}}_1$ and the $(\phi^\dagger\phi)^3$ coupling are not renormalized in the ${\ensuremath{\mathcal{N}}}=2$ theory. The diagrams contributing to this correlator again factorize into a product of matter loops, that are in general of a lower order in perturbation theory (the only difference is that the diagrams may now include both $\bar{\psi}\psi\bar{\phi}\phi$ and $(\phi^\dagger\phi)^3$ vertices). The proof then follows in a similar way.
We are now ready to show that $\beta_{\lambda_6} = O(1/N)$. In our non-supersymmetric model, at large $N$ the correlator $\left< ({\ensuremath{\mathcal{O}}}_1)^3 \right>$ receives contributions from two types of diagrams, with either zero or one $(\phi^\dagger\phi)^3$ vertices:
---------------------------------- ----------------------------------
{height="4.5cm"} {height="4.5cm"}
(G1) (G2)
---------------------------------- ----------------------------------
The $\lambda_6$ coupling contributes to $\left< ({\ensuremath{\mathcal{O}}}_1)^3 \right>$ at large $N$ through diagrams of the form (G2), some of which are non-zero (such as the leading order diagram which is explicitly drawn). Thus, if the beta function is non-zero we must have divergences in $\left< ({\ensuremath{\mathcal{O}}}_1)^3 \right>$. However, we have shown above that at every order in $\lambda$ the sum of diagrams (G1) is finite, and also the sum of diagrams (G2) is finite. Thus, the beta function must vanish at large $N$.
Spontaneous Breakdown of Conformal Symmetry {#sponbreak}
-------------------------------------------
In order to verify that our theories are conformally invariant, we should also make sure that they do not spontaneously break conformal invariance, by a vacuum expectation value for $\phi^2$. For the theory with $\lambda=0$ and $\lambda_6 \neq 0$, this was analyzed in detail in [@Bardeen:1983rv], and it was found that for $\lambda_6 < (4\pi)^2$ such a breaking does not arise. In fact, the effective potential for $\sigma = \phi^2/N$ can be computed exactly for infinite $N$, and it takes the form [@Amit:1984ri] $$V(\sigma) = \frac{N}{6} \left[ (4\pi)^2 - \lambda_6 \right] |\sigma|^3.$$ Thus, for small $\lambda_6$ the only minimum of the effective potential is at the conformal point $\phi^2=0$. We expect that turning on a small coupling $\lambda$, as we analyzed above, will lead to small changes in the coefficient of $|\sigma|^3$ in this effective potential (which can be computed in perturbation theory in $\lambda$), but at least for small $\lambda$ and small $\lambda_6$ it seems clear that there will still be a minimum of the effective potential at $\sigma=0$. Thus, at least for weak couplings and large $N$, the conformal symmetry is not spontaneously broken in the two-parameter family of conformal field theories that we discussed above. For $\lambda=0$ a spontaneous breaking of the conformal symmetry can occur when $\lambda_6 = (4\pi)^2$ exactly, and it would be interesting to investigate how this statement is modified at finite $\lambda$ (see [@Dias:2003pw; @Dias:2010it] for a study of the effective potential in the Abelian Chern-Simons-Matter theory, and [@Rabinovici:2011jj] for a similar study of the $O(N)$ vector model with a Chern-Simons term for a $U(1)$ subgroup of $O(N)$). For small values of $N$, spontaneous breaking of the conformal symmetry might happen (as in [@Ferrari:2010ex]), and it would be interesting to check if it happens in our theories.
Higher-Spin Currents {#HScurrents}
====================
The main goal of this section is to find the spectrum of primary operators of the large $N$ interacting fixed points discovered in the previous section. Let us begin by considering the free theory, taking $\lambda=\lambda_6=0$. For each positive, even spin $s$ it has a unique $O(N)$-singlet primary operator $J_s$ that saturates the unitarity bound $\Delta \ge s + d - 2$. (In the theory with a complex scalar in the fundamental representation of $U(N)$ there is such a primary for each positive spin, not just the even ones.) These are symmetric, traceless tensors that can be written schematically as[^8] $$\begin{aligned}
J_{\mu_1 \dots \mu_s} &= \frac{1}{\sqrt{N}}
\phi^i \partial_{\mu_1} \cdots \partial_{\mu_s} \phi^i + \cdots {\,.}\end{aligned}$$ For example, the first two such operators are $$\begin{aligned}
J_{\mu\nu} &= \frac{1}{\sqrt{N}} \left\{
-\frac{1}{3} \phi^i \partial_\mu \partial_\nu \phi^i + \partial_\mu \phi^i \partial_\nu \phi^i - \frac{1}{3} \delta_{\mu\nu} \partial_\rho \phi^i \partial_\rho \phi^i + \frac{1}{9} \delta_{\mu\nu} \phi^i \square \phi^i \right\}
{\,,}\label{eq:J2free}\\
J_{\mu\nu\rho\sigma} &= \frac{1}{\sqrt{N}} \left\{
\frac{3}{2} \phi^i \partial_\mu \partial_\nu \partial_\rho \partial_\sigma
\phi^i
-
42 \partial_{(\mu} \phi^i
\partial_\nu \partial_\rho \partial_{\sigma)} \phi^i
+
\frac{105}{2} \partial_{(\mu} \partial_\nu \phi^i
\partial_\rho \partial_{\sigma)} \phi^i
\right. \nonumber\\ &\quad
+
18 \delta_{(\mu\nu} \partial_\rho \partial_{\sigma)}
\partial_\chi \phi^i \partial_\chi \phi^i
-
30 \delta_{(\mu\nu|} \partial_\chi \partial_{|\rho|} \phi^i
\partial_{\chi} \partial_{|\sigma)} \phi^i
+
3 \delta_{(\mu\nu} \delta_{\rho\sigma)}
\partial_\chi \partial_\xi \phi^i
\partial_\chi \partial_\xi \phi^i
\nonumber\\ &\quad
-
\frac{9}{7} \delta_{(\mu\nu} \phi^i \partial_\rho \partial_{\sigma)}
\square \phi^i
+
18 \delta_{(\mu\nu} \partial_\rho \phi^i \partial_{\sigma)}
\square \phi^i
-
15 \delta_{(\mu\nu} \partial_\rho \partial_{\sigma)} \phi^i
\square \phi^i
\nonumber\\ &\quad \left.
+
\frac{9}{70} \delta_{(\mu\nu} \delta_{\rho\sigma)}
\phi^i \square \square \phi^i
+
\frac{3}{2} \delta_{(\mu\nu} \delta_{\rho\sigma)}
\square \phi^i \square \phi^i
-
\frac{18}{5} \delta_{(\mu\nu} \delta_{\rho\sigma)}
\partial_\chi \phi^i \partial_\chi \square \phi^i \right\} {\,,}\label{eq:J4free}\end{aligned}$$ where parentheses around indices denote an averaging over all permutations of the indices. When discussing the large $N$ limit we shall call such scalar bilinears “single-trace” operators. Since they saturate the unitarity bound, these primaries are also conserved currents, $\partial_\mu J^{\mu}_{\;\;\mu_1\dots\mu_{s-1}} = 0$, and therefore the free theory has an infinite number of conserved currents. In addition there is a scalar singlet operator $J_0 = \frac{1}{\sqrt{N}} \phi^i \phi^i$, also a primary, with dimension $\Delta=1$. In the large $N$ limit, all operators in the theory are products of these basic “single-trace” operators, or descendants of such products. Note that adding the Chern-Simons sector does not add any additional non-trivial local operators.
Let us now turn on the Chern-Simons coupling $\lambda$. As we showed in section \[allorders\], the theory is still conformally-invariant at infinite $N$. The currents of the free theory, as written above, are not gauge-invariant, but they can be made gauge-invariant by promoting derivatives to covariant derivatives and projecting onto the symmetric traceless part. The promoted currents, which will also be denoted $J_s$, are the “single-trace” primary operators of the new theory. At finite $N$ they are generally not conserved, and they also mix with “multi-trace operators”; however, as we shall now see (following a similar analysis in [@Girardello:2002pp]) they are still conserved at $N=\infty$.
In the free theory, the primary operator $J_s$ heads a short representation of the conformal group that we label $(\Delta=s+1,s)$, where $\Delta$ is the conformal dimension and $s$ the spin. The shortening condition is the conservation equation $\partial_\mu J^\mu_{\,\,\,\,\,\mu_1\dots\mu_{s-1}} = 0$. For $J_s$ to become non-conserved, there must appear on the right-hand side of this equation a non-zero operator in the representation $(s+2,s-1)$. In other words, $J_s$ must combine with another operator in this representation to form a long representation [@Girardello:2002pp], $$\begin{aligned}
\label{confreps}
{\mathrm{lim}}_{\epsilon \to 0}(s+1+\epsilon,s)_{\mathrm{long}} &\cong (s+1,s)_{\mathrm{short}} \oplus (s+2,s-1) {\,.}\end{aligned}$$ By acting with special conformal transformations on $\partial_\mu J^\mu_{\,\,\,\,\,\mu_1\dots\mu_{s-1}}$ one can show that in the limit in which $J_s$ is conserved, the $(s+2,s-1)$ operator in must be a primary of the conformal algebra [@Heidenreich:1980xi] (the coefficient of this operator in the equation for $d*J_s$ vanishes in this limit, but the special conformal generator acting on $d*J_s$ vanishes even faster). Now, a connected correlator of the form $\partial_\mu \left< J^\mu_{\,\,\,\,\mu_1\dots\mu_{s-1}} {\ensuremath{\mathcal{O}}}\right>$ can have a leading, $O(N^0)$ contribution in the large $N$ limit only when ${\ensuremath{\mathcal{O}}}$ is a “single-trace” operator. Therefore, at $N=\infty$, $J_s$ can only combine with other “single-trace” operators. Since there are no such primary operators with $(s+2,s-1)$, $J_s$ must remain conserved even when the Chern-Simons interaction is turned on. Because the representations for conserved currents are short, this also implies that the currents do not acquire an anomalous dimension at this order.
Next we consider the $O(1/\sqrt{N})$ corrections. At this order the currents with $s > 2$ can become non-conserved, but only by combining with a “double-trace” operator [@Girardello:2002pp] of the schematic form $$\begin{aligned}
\partial \cdot J_s &\sim
\frac{f(\lambda)}{\sqrt{N}} \, \epsilon \, \partial^2 J_{s-2} \, J_0 +
(\mathrm{other\ double\!-\!trace\ operators}) {\,,}\label{noncons}\end{aligned}$$ where $\epsilon$ is the Levi-Civita tensor, and the indices are implicit and can be contracted in various ways. Parity implies that the function $f(\lambda)$ must be odd. Such an equation implies that $J_s$ has an anomalous dimension of order $1/N$, times some function of $\lambda$.
From it is easy to obtain a non-renormalization theorem for the anomalous dimension of $J_0$ at large $N$ (which we derived by different methods in the previous section).[^9] By making a scale transformation of and using the fact that $\Delta_s~=~s+1$, we see that the scaling dimension of $J_0$ must be $\Delta_0=1+O(1/\sqrt{N})$, namely it does not get corrections at $N=\infty$, for any value of $\lambda$. The implicit assumption in this argument is that the coefficient $f(\lambda)$ on the right-hand side of does not vanish. This is indeed what we find for the divergence of (for example) $J_4$ at leading order in $\lambda$ by using the equations of motion, $$\begin{aligned}
\partial^\sigma J_{\mu\nu\rho\sigma} =
-\frac{i}{2} \frac{\lambda}{\sqrt{N}} &\left\{
\frac{540}{7} \,\epsilon_{\alpha\beta(\mu}
J_{\nu|\alpha}
\partial_{|\rho)} \partial_\beta J_0
+
\frac{396}{7} \,\epsilon_{\alpha\beta(\mu|}
\partial_\alpha J_{\beta|\nu} \cdot
\partial_{\rho)} J_0
-
\frac{468}{7} \,\epsilon_{\alpha\beta(\mu}
\partial_\nu J_{\rho)\alpha} \cdot
\partial_\beta J_0
\right. \notag\\ &\quad
-
\frac{108}{7} \,\epsilon_{\alpha\beta(\mu|}
\partial_\alpha \partial_{|\nu} J_{\rho)\beta} \cdot
J_0
-
\frac{108}{7} \,\delta_{(\mu\nu} \epsilon_{\rho)\alpha\beta}
J_{\alpha\chi}
\partial_\beta \partial_\chi J_0
\notag\\ &\quad
-
\frac{1989}{224} \,\delta_{(\mu\nu} \epsilon_{\rho)\alpha\beta}
\partial_\alpha J_{\beta\chi} \cdot
\partial_\chi J_0
+
36 \,\epsilon_{\alpha\beta(\mu|}
\partial_\alpha J_{|\nu\rho)} \cdot
\partial_\beta J_0
\notag\\ &\quad \left.
+
\frac{3141}{224} \,\delta_{(\mu\nu|} \epsilon_{\alpha\beta\gamma}
\partial_\alpha J_{|\rho)\beta} \cdot
\partial_\gamma J_0
+
O(\lambda^2,\lambda_6)
\right\} {\,.}\label{divJ4}\end{aligned}$$ One can verify that the right-hand side of is traceless at this order, as expected: the traced right-hand side is proportional to $$\begin{aligned}
\epsilon_{\mu\alpha\beta} \partial_{\alpha} \phi^{i} \partial_{\nu} \partial_{\beta} \phi^{i} \partial_{\nu} \phi^{j} \phi^{j}
- \epsilon_{\nu\alpha\beta} \partial_{\alpha} \phi^{i} \partial_{\mu} \partial_{\beta} \phi^{i} \partial_{\nu} \phi^{j} \phi^{j}
+ \epsilon_{\mu\alpha\beta} \partial_{\nu} \phi^{i} \partial_{\alpha} \partial_{\nu} \phi^{i} \partial_{\beta} \phi^{j} \phi^{j}
{\,,}\end{aligned}$$ and this expression can be shown to vanish by choosing a specific value for $\mu$ and using the equations of motion.
As another check of one can act on both sides with $K^\rho$, the generator of special conformal transformations (see Appendix \[conf\] for our conventions). On the left-hand side we have (when the operator is at $x=0$) $$\begin{aligned}
[K^\rho, [P^\sigma, J_{\mu\nu\rho\sigma}]] &=
[[K^\rho, P^\sigma], J_{\mu\nu\rho\sigma}]
+ [P^\sigma, [K^\rho, J_{\mu\nu\rho\sigma}]]
= 2i [\delta^{\rho\sigma} D + M^{\rho\sigma},
J_{\mu\nu\rho\sigma}]
= 0 {\,,}\end{aligned}$$ where we used the fact that $J_4$ is a primary operator, and that it is symmetric and traceless. The commutator of $K^{\rho}$ with the right-hand side of should therefore also vanish, and this can be verified directly. The calculation is straightforward, and does not require substituting the explicit expressions for $J_2$ and $J_0$. We have also explicitly verified in appendix \[adim\] that $J_0$ indeed has vanishing anomalous dimension at leading order in $1/N$, to two-loop order.
Let us summarize this section. We considered the spectrum of primaries in the 2-parameter family of conformal theories at infinite $N$, found in section \[confSym\]. We showed that the spectrum of single-trace, gauge-invariant primaries in these theories is the same as that of the free theory; namely, it consists of conserved higher-spin currents of all even positive spins in the $O(N)$ model (and all positive spins in the $U(N)$ model), plus a scalar operator of conformal dimension 1. For finite $N$, all these operators (except for the conserved energy-momentum tensor $J_2$, and (for the $U(N)$ model) the conserved $U(1)$ current $J_1$) obtain anomalous dimensions.
Correlation Functions {#conjecture}
=====================
We have seen above that for infinite $N$ the scaling dimensions in our family of fixed points are independent of $\lambda$, and the deformation of the spectrum is trivial at large $N$. One could then worry that perhaps all correlation functions are independent of $\lambda$. In this section we compute a specific correlation function of currents, $\left<J_2 J_1 J_1\right>$, and show that it does depend on $\lambda$ (already at leading order in $\lambda$).
One motivation for this computation is to obtain clues towards finding a holographic dual for the theories discussed above. The free theory with $\lambda=\lambda_6=0$ is conjectured [@Klebanov:2002ja] to be dual to Vasiliev’s higher-spin gauge theory on $AdS_4$, and our theories should be (in the classical limit) continuous deformations of this. The existence of a deformation of Vasiliev’s theory, which is dual to the $(\phi^2)^3$ deformation of the free vector model, was first mentioned in [@Elitzur:2005kz]. For that deformation the holographic picture is clear, since this is a “multi-trace” deformation that is manifested in the holographic dual as a change in boundary conditions of the scalar field dual to $\phi^2$ [@Witten:2001ua; @Berkooz:2002ug; @Elitzur:2005kz].
On the other hand, the Chern-Simons deformation by $\lambda$ should be realized on the gravity side as a continuous, parity-breaking deformation of Vasiliev’s theory. One natural conjecture could be that it is dual to one of the known parity-breaking deformations of Vasiliev’s theory, which were parameterized in [@Sezgin:2002rt] by some odd function ${\cal V}(X)$. However, as mentioned in [@Giombi:2011kc], such a deformation seems not to lead to a non-vanishing $\left<J_2 J_1 J_1\right>$ at leading order in $\lambda$ as we find below. If so, there should be some new, unknown deformation of Vasiliev’s theory that is dual to turning on $\lambda$, and it would be very interesting to discover it.
Computation of $\left<J_2 J_1 J_1\right>$ {#J2J1J1}
-----------------------------------------
Corrections to correlation functions at order $\lambda$ necessarily break parity. For simplicity, we study here the $U(N)$ case, which has a conserved current $J_1$, since the correlator $\left<J_2 J_1 J_1\right>$ is the simplest correlator of conserved currents that can exhibit a parity-breaking structure [@Giombi:2011rz]. The conserved currents $J=J_1$, $T=J_2$ of the theory of $N$ complex scalars with $U(N)$ Chern-Simons interactions are given by $$\begin{aligned}
J_\mu &= \frac{1}{\sqrt{N}}\left\{i (D_\mu \phi)^\dagger \phi - i \phi^\dagger D_\mu \phi\right\}
{\,,}\label{J1cpx}
\\
T_{\mu\nu} &= \frac{1}{\sqrt{N}} \left\{
-\frac{1}{6} ( \phi^\dagger D_\mu D_\nu \phi
+ \phi^\dagger D_\nu D_\mu \phi
+ D_\mu D_\nu \phi^\dagger \cdot \phi
+ D_\nu D_\mu \phi^\dagger \cdot \phi )
+ D_\mu \phi^\dagger D_\nu \phi
+ D_\nu \phi^\dagger D_\mu \phi
\notag \right.\\ &\quad \left.
\qquad \quad
- \frac{2}{3} \delta_{\mu\nu} D_\rho \phi^\dagger D_\rho \phi
+ \frac{1}{9} \delta_{\mu\nu} \phi^\dagger D^2 \phi
+ \frac{1}{9} \delta_{\mu\nu} D^2 \phi^\dagger \cdot \phi \right\}
{\,,}\label{emtensorcpx}\end{aligned}$$ where $D_\mu = \partial_\mu + g A_\mu^a T^a$.
With these definitions the 2-point functions of $J$ and $T$ in the free theory are fixed to be (denoting e.g. $J_{{\varepsilon}}(x) = {\varepsilon}^{\mu} J_{\mu}(x)$) $$\begin{aligned}
\langle J_{{\varepsilon}_1}(x_1) J_{{\varepsilon}_2}(x_2) \rangle &= \frac{1}{8\pi^2|x_{12}|^2}{\varepsilon}_1^{\mu}{\varepsilon}_2^{\nu}\left(\frac{\delta^{\mu\nu}}{|x_{12}|^2}-2\frac{x_{12}^{\mu}x_{12}^{\nu}}{|x_{12}|^4}\right) {\,,}\\
\langle T_{{\varepsilon}_1{\varepsilon}_1}(x_1) T_{{\varepsilon}_2{\varepsilon}_2}(x_2) \rangle &= \frac{1}{3\pi^2|x_{12}|^2}\left[{\varepsilon}_1^{\mu}{\varepsilon}_2^{\nu}\left(\frac{\delta^{\mu\nu}}{|x_{12}|^2}-2\frac{x_{12}^{\mu}x_{12}^{\nu}}{|x_{12}|^4}\right)\right] ^2 {\,.}\end{aligned}$$
We now compute the correlator $\left< T_{{\varepsilon}_1 {\varepsilon}_2}(x_1) J_{{\varepsilon}_3}(x_2) J_{{\varepsilon}_4}(x_3) \right>$ in $x$-space at order $\lambda$. It has a unique parity-violating tensor structure, and to compute its coefficient it will prove useful (as in [@Giombi:2011kc]) to choose all polarizations equal and null, ${\varepsilon}_i = {\varepsilon}$, ${\varepsilon}^2 = 0$, and to take the limit $x_2 \to x_1$. With these choices, the parity-violating tensor structure has the form [@Giombi:2011rz] $$\begin{aligned}
\frac{1}{|x_{12}||x_{23}||x_{13}|} \left(
Q_1^2 S_1 + 2 P_2^2 S_3 + 2 P_3^2 S_2 \right) \to
-\frac{4 \epsilon_{\mu\nu\rho} x_{13}^\mu x_{12}^\nu {\varepsilon}^\rho
({\varepsilon}\cdot x_{12})^2 ({\varepsilon}\cdot x_{13}) }
{ |x_{12}|^6 |x_{13}|^6 } {\,.}\label{J2J1J1struct}\end{aligned}$$ In the limit $x_2 \to x_1$ it diverges as $|x_{12}|^{-3}$, and we shall use this fact to discard subleading terms in $|x_{12}|$.
There are 3 diagrams, up to permutations of the current insertions, contributing to the correlator at order $\lambda$:
--------------------------------- --------------------------------- ---------------------------------
{width="5.2cm"} {width="5.2cm"} {width="5.2cm"}
(D1) (D2) (D3)
--------------------------------- --------------------------------- ---------------------------------
Diagrams of the type (D1) vanish because all polarizations are equal: from the gluon propagator we have $\epsilon_{\mu\nu\rho} \varepsilon^\mu \varepsilon^\nu = 0$. The computation of (D2), (D3) is not completely straightforward; it involves the repeated application of several techniques, as we will demonstrate by computing one of the (D3) diagrams in detail. The results for the other diagrams are listed in Appendix \[diags\].
In this theory the scalar and gluon propagators are given by $$\begin{aligned}
I_{xy} &= \frac{1}{4 \pi}\frac{1}{|x-y|} {{\,,}\quad}I_{\mu\nu;xy} = -\frac{i}{4\pi}\frac{\epsilon_{\mu\nu\rho}(x-y)^{\rho}}{|x-y|^3}{\,.}\end{aligned}$$
To regularize the diagrams we change the loop variables to be $d$ dimensional: $d^3{\omega}\to d^d{\omega}$. While this is not a gauge-invariant regulator, we found that each of the diagrams is finite and thus independent of $d$. We have also checked that using a different regulator gives the same results.
The diagram (D3), with the gluon line stretched between the two scalar propagators connected to the energy-momentum tensor, is given by $$\begin{aligned}
-2 \frac{\lambda}{\sqrt{N}}
\int \! d^d{\omega}_1 \, d^d{\omega}_2 \, I_{\alpha\beta;{\omega}_1{\omega}_2}
\left[ I_{{\omega}_2 x_1} T^0_{{\epsilon}{\epsilon}}(x_1) I_{x_1{\omega}_1}
{\overleftrightarrow}{{\partial}}_{\!\!{\omega}_1,\alpha} I_{{\omega}_1 x_2}
({\varepsilon}\cdot{\overleftrightarrow}{{\partial}}_{\!\!x_2}) I_{x_2 x_3}
({\varepsilon}\cdot{\overleftrightarrow}{{\partial}}_{\!\!x_3}) I_{x_3{\omega}_2}
{\overleftrightarrow}{{\partial}}_{\!\!{\omega}_2,\beta} \right]
{\,.}\label{eq:D3Perm1}\end{aligned}$$ Here ${\overleftrightarrow}{{\partial}}\equiv \overrightarrow{{\partial}}-\overleftarrow{{\partial}}$, and $$\begin{aligned}
\phi^{i\dag}(x_1)T^0_{{\varepsilon}{\varepsilon}}(x_1)\phi^i(x_1)\equiv
\phi^{i\dag}(x_1)\left[-\frac{1}{3}({\varepsilon}\cdot{\overrightarrow}{{\partial}_{x_1}})^2-\frac{1}{3}({\varepsilon}\cdot{\overleftarrow}{{\partial}_{x_1}})^2+
2({\varepsilon}\cdot{\overleftarrow}{{\partial}_{x_1}})({\varepsilon}\cdot{\overrightarrow}{{\partial}_{x_1}})\right]\phi^{i}(x_1)\end{aligned}$$ is the energy-momentum tensor at leading order. It is understood that the right-most derivative $\overrightarrow{{\partial}}$ in acts on the left-most propagator inside the brackets.
Let us try and take as many derivatives as possible out of the integral. We are limited by the fact that there are two propagators involving $x_1$, and the combination $T^0_{{\varepsilon}{\varepsilon}}$ is not a total derivative acting on them. To proceed let us first split the point $x_1$ into two points $x_1,x_1'$, each connected to a different scalar line; eventually we will take $x_1' \to x_1$. This procedure does not spoil gauge invariance, since at this order in $\lambda$ it is equivalent to stretching a Wilson line between the separated points. The result can be written as $$\begin{aligned}
\frac{2i \lambda}{(4\pi)^6\sqrt{N}}\frac{1}{|x_{23}|}
\left. ({\varepsilon}\cdot{\overleftrightarrow}{{\partial}_{x_2}})
({\varepsilon}\cdot{\overleftrightarrow}{{\partial}_{x_3}})
\left({\partial}_{x_2,\alpha}-{\partial}_{x_1,\alpha}\right) \cal{I}_\alpha \, \right|_{x_1'\rightarrow x_1} {\,,}\label{D3simp}\end{aligned}$$ where $x_{ij} \equiv x_i - x_j$, and $$\begin{aligned}
\cal{I}_\alpha &= 2 \int \! d^d{\omega}_1 \, d^d{\omega}_2 \,
\frac{{\epsilon}_{\alpha\beta\gamma}{\omega}_{12}^{\gamma}({\omega}_2-x_3)^{\beta}}{|{\omega}_{12}|^3|{\omega}_1-x_1||{\omega}_2-x_1'||{\omega}_1-x_2||{\omega}_2-x_3|^3} \times \notag\\ &\quad \qquad
\times \left[ 2\frac{({\varepsilon}\cdot({\omega}_1-x_1))({\varepsilon}\cdot({\omega}_2-x_1'))}{|{\omega}_1-x_1|^2|{\omega}_2-x_1'|^2} -\frac{({\varepsilon}\cdot({\omega}_1-x_1))^2}{|{\omega}_1-x_1|^4} - \frac{({\varepsilon}\cdot({\omega}_2-x_1'))^2}{|{\omega}_2-x_1'|^4} \right]
\label{eq:A10} {\,.}\end{aligned}$$ To arrive at this form, we used the relation $({\omega}_2-x_3)^{\beta} / |{\omega}_2-x_3|^3 =-{\partial}_{{\omega}_2}^{\beta} |{\omega}_2-x_3|^{-1}$, and integrated by parts with respect to ${\omega}_2$. Note that we have chosen to take out a single $x_1$ derivative, while acting with the rest explicitly.
Next, note that ${\partial}_{x_1',\alpha} \cal{I}_\alpha = 0$, as can be seen by rewriting ${\partial}_{1'}^{\alpha}$ as ${\partial}_{{\omega}_2}^{\alpha}$ and integrating by parts. This means that we can take $x_1'\rightarrow x_1$ before acting with the outer derivatives in , since the $x_1$ derivative there acts in the $\alpha$ direction. In addition, let us shift ${\omega}_{1,2} \to {\omega}_{1,2} + x_1$. With these changes, the integral simplifies to $$\begin{aligned}
\cal{I}_\alpha &= 2\int \! d^d{\omega}_1 \, d^d{\omega}_2 \, \frac{{\epsilon}_{\alpha\beta\gamma} {\omega}_{12}^{\gamma}({\omega}_2+x_{13})^{\beta}}{|{\omega}_{12}|^3|{\omega}_1||{\omega}_2||{\omega}_1+x_{12}||{\omega}_2+x_{13}|^3}
\left[
2\frac{({\varepsilon}\cdot{\omega}_1)({\varepsilon}\cdot{\omega}_2)}{|{\omega}_1|^2|{\omega}_2|^2}
- \frac{({\varepsilon}\cdot{\omega}_2)^2}{|{\omega}_2|^4}
- \frac{({\varepsilon}\cdot{\omega}_1)^2}{|{\omega}_1|^4}
\right]{\,.}\label{eq:A1}\end{aligned}$$
Let us now consider the limit $x_2 \to x_1$, in which the integral diverges as $|x_{12}|^{-1}$. In this limit, the full diagram diverges as $|x_{12}|^{-3}$, and therefore contributes to the parity-violating tensor structure . We first compute the last term in the brackets in . Using the fact that ${\omega}_{12}^\gamma/|{\omega}_{12}|^3 = - \partial_{{\omega}_2,\gamma} |{\omega}_{12}|^{-1}$ and integrating by parts we rewrite this term as $$\begin{aligned}
\cal{I}_\alpha^{\text{last~term}} &= -2\int \! d^d{\omega}_1 \, d^3{\omega}_2 \, \frac{ ({\varepsilon}\cdot{\omega}_1)^2{\epsilon}_{\alpha\beta\gamma}{\omega}_2^{\gamma}x_{13}^{\beta} }{ |{\omega}_{12}||{\omega}_1|^5|{\omega}_2|^3|{\omega}_1+x_{12}||{\omega}_2+x_{13}|^3} \nonumber\\
&= -\frac{1}{\pi}\int \! d^d{\omega}_1 \, d^5{\omega}_2 \, \frac{ ({\varepsilon}\cdot{\omega}_1)^2{\epsilon}_{\alpha\beta\gamma}{\omega}_1^{\gamma}x_{13}^{\beta} }{ |{\omega}_{12}|^3|{\omega}_1|^5|{\omega}_2|^3|{\omega}_1+x_{12}||{\omega}_2+x_{13}|^3} {\,.}\end{aligned}$$ The second equality can be verified by introducing Feynman parameters and performing the dimensional integration on both sides. The integral over ${\omega}_2$ can now be carried out [@Boos:1987bg], and we find $$\begin{aligned}
\cal{I}_\alpha^{\text{last~term}} = -\frac{8\pi}{|x_{13}|}\int \! d^3{\omega}_1 \, \frac{({\varepsilon}\cdot{\omega}_1)^2{\epsilon}_{\alpha\beta\gamma}{\omega}_1^{\gamma}x_{13}^{\beta}}{|{\omega}_1|^6|{\omega}_1+x_{12}||{\omega}_1+x_{13}|}\frac{1}{|{\omega}_1|+|x_{13}|+|{\omega}_1+x_{13}|} {\,.}\end{aligned}$$ As mentioned above, this integral diverges as $|x_{12}|^{-1}$ in the limit $x_2 \to x_1$, and the divergence comes from the region $|{\omega}_1| \ll 1$. As we approach the limit, most of the contribution to the integral will therefore come from this region. We can therefore expand around ${\omega}_1 = 0$, keeping only the leading term; the remaining terms will give sub-leading corrections in $|x_{12}|$. We thus arrive at a straightforward integral, $$\begin{aligned}
\cal{I}_\alpha^{\text{last~term}} = -4\pi\frac{x_{13}^{\beta}}{|x_{13}|^3}\int \! d^3{\omega}_1 \, \frac{({\varepsilon}\cdot{\omega}_1)^2{\epsilon}_{\alpha\beta\gamma}{\omega}_1^{\gamma}}{|{\omega}_1|^6|{\omega}_1+x_{12}|}
+ O((x_{12})^0) {\,.}\label{Alast}\end{aligned}$$
The other two terms in have an $|x_{12}|^{-1}$ divergence in the limit $x_2 \to x_1$ coming from the region $|{\omega}_1|,|{\omega}_2| \ll 1$, and we can similarly take the leading order in the expansion around ${\omega}_{1,2} = 0$. The resulting integral is again straightforward to evaluate, $$\begin{aligned}
\cal{I}_\alpha^{\text{terms~1,2}} &= 2\frac{x_{13}^{\beta}}{|x_{13}|^3}
\int \! d^d{\omega}_1 \, d^d{\omega}_2 \,
\frac{{\epsilon}_{\alpha\beta\gamma}{\omega}_{12}^{\gamma}}{|{\omega}_{12}|^3|{\omega}_1||{\omega}_2||{\omega}_1+x_{12}|}
\left[
2\frac{({\varepsilon}\cdot{\omega}_1)({\varepsilon}\cdot{\omega}_2)}{|{\omega}_1|^2|{\omega}_2|^2}
- \frac{({\varepsilon}\cdot{\omega}_2)^2}{|{\omega}_2|^4} \right]
+ O((x_{12})^0) {\,.}\label{eq:A12}\end{aligned}$$
Combining the results of and and plugging into , the contribution of the specific (D3) diagram that we computed to the parity-violating tensor structure is $$\begin{aligned}
\label{eq:D3Perm1Res}
\frac{i}{24\pi^4}\frac{\lambda}{\sqrt{N}}\frac{({\varepsilon}\cdot x_{12})^2({\varepsilon}\cdot x_{13}){\epsilon}_{\alpha\beta\gamma}x_{12}^{\alpha}x_{13}^{\beta}{\varepsilon}^{\gamma}}{|x_{13}|^6|x_{12}|^6}
{\,.}\end{aligned}$$
By applying similar techniques one can compute the other (D2), (D3) diagrams and their permutations; the results are listed in Appendix \[diags\]. Summing these contributions, we find the following non-zero result at order $\lambda$, $$\begin{aligned}
\left. \langle T_{{\varepsilon}{\varepsilon}}(x_1)J_{{\varepsilon}}(x_2)J_{{\varepsilon}}(x_3)\rangle \right|_{x_2 \to x_1} = \frac{i}{24\pi^4}\frac{\lambda}{\sqrt{N}}\frac{({\varepsilon}\cdot x_{12})^2({\varepsilon}\cdot x_{13}){\epsilon}_{\alpha\beta\gamma}x_{12}^{\alpha}x_{13}^{\beta}{\varepsilon}^{\gamma}}{|x_{13}|^6|x_{12}|^6}.\end{aligned}$$ Using the known tensor structure , for general coordinates and polarizations this implies $$\begin{aligned}
\langle T_{{\varepsilon}_1{\varepsilon}_1}(x_1)J_{{\varepsilon}_2}(x_2)J_{{\varepsilon}_3}(x_3)\rangle &= \frac{i}{96\pi^4}\frac{\lambda}{\sqrt{N}}\frac{1}{|x_{12}||x_{23}||x_{13}|} \left(
Q_1^2 S_1 + 2 P_2^2 S_3 + 2 P_3^2 S_2 \right)
+ o(\lambda^2) {\,.}\end{aligned}$$
Summary and Future Directions {#summary}
=============================
In this paper we studied the three dimensional $O(N)$ ($U(N)$) vector model coupled to a Chern-Simons theory at level $k$, in the limit of large $N,k$ with a fixed ratio $\lambda = 4 \pi N / k$. We found that for infinite $N$ this theory has two exactly marginal deformations, corresponding to $\lambda$ and to a $(\phi^2)^3$ coupling, while for finite large $N$ we showed that there is (at least for small $\lambda$) a single IR-stable fixed point for every $\lambda$. For infinite $N$ we showed that none of the operators of the theory have anomalous dimensions, so that the infinite tower of conserved currents of the theory with $\lambda=0$ remains also for finite $\lambda$ (and finite $\lambda_6$). We showed explicitly that some of the correlation functions of the infinite $N$ theory do depend on $\lambda$.
The fact that at infinite $N$ we find an infinite tower of conserved currents even in the interacting theory at finite $\lambda$ is quite surprising, and suggests that this theory may have some interesting integrable structure. In this paper we only performed explicit computations at low orders in perturbation theory. However, the existence of an infinite number of conserved currents may be useful towards performing exact computations as a function of $\lambda$ in these theories. When our scalar fields are replaced by fermion fields, many such exact computations can indeed be performed [@Giombi:2011kc]. In this case there is a choice of gauge for which only rainbow diagrams contribute, simplifying the resummation of all planar diagrams. For scalar fields we have not yet been able to find similar simplifications.
Vector models of the type we analyze here exhibit large $N$ phase transitions at temperatures of order $\sqrt{N}$ [@Shenker:2011zf; @Giombi:2011kc]. It would be interesting to generalize these transitions to our finite $\lambda$ theories.
It would also be interesting to understand the holographic duals of the theories with finite $\lambda$ that we discussed here, which should be continuous deformations of Vasiliev’s higher-spin theories. Unlike standard marginal deformations, here we are not deforming by the integral of a gauge-invariant local operator, so it is not obvious how to identify this deformation. Perhaps the attempted derivations of the $AdS$/CFT correspondence for $\lambda=0$ [@Das:2003vw; @Koch:2010cy; @Douglas:2010rc; @Jevicki:2011ss] can be generalized to finite $\lambda$, by replacing the scalar bilinear operators $\phi_i(x) \phi_i(y)$ appearing in these derivations by a gauge-invariant bilinear (in which the two scalars are connected by an open Wilson line); if so then this should provide clues towards the construction of this holographic dual.
It would also be interesting to understand finite $N$ corrections to our theories on the gravity side, though this may require a quantum completion of Vasiliev’s higher-spin theory that is not yet known. Since on the field theory side our theories are vector models, it seems that they should not correspond to closed string theories, but to open string theories coupled to a trivial (topological) closed string background. For instance, since the closed string duals of the $O(N)_k$ and $U(N)_k$ Chern-Simons theories are known topological string theories [@Gopakumar:1998ki; @Sinha:2000ap], one could imagine that adding fundamental matter fields to these theories (as we have done) should correspond to adding non-topological D-branes to these topological string theories.
There are many possible generalizations of our computations. The generalization to the case of $l$ vectors of scalar fields is straightforward, and all the operators we discuss just become $l\times l$ matrices (the description of this on the gravity side is straightforward). The anomalous dimensions of all these operators still vanish in the large $N$ limit, so in particular we have many massless “gravitons” in this case, as expected for a theory involving $l$ D-branes. The generalization to fermionic fields instead of scalars will be discussed in [@Giombi:2011kc]. One can also consider an ${{\ensuremath{\mathcal{N}}}=1}$ supersymmetric generalization of our theories, whose field content includes both a scalar and a fermionic field, with specific interactions between them. The gravity dual for this case was discussed in [@Leigh:2003gk; @Sezgin:2003pt], and it would be interesting to generalize our discussion of the theory with finite $\lambda$ to this case. It would also be interesting to find the gravity dual for the ${{\ensuremath{\mathcal{N}}}=2}$ generalizations of our theories, that we briefly discussed in §\[allorders\].
We hope that further study of these theories will shed more light on the structure of the $AdS$/CFT correspondence in the case where it gives a weak-weak coupling duality, and hopefully also in general.
Acknowledgments {#s:acks .unnumbered}
---------------
We would like to thank R. Gopakumar and S. Minwalla for interesting discussions that initiated this project, and S. Giombi and S. Minwalla for many useful discussions and for notifying us of the results of [@Giombi:2011kc]. We also thank D. Jafferis, Z. Komargodski, E. Rabinovici, M. Smolkin, S. Wadia and S. Yankielowicz for useful discussions. This work was supported in part by the Israel–U.S. Binational Science Foundation, by a research center supported by the Israel Science Foundation (grant number 1468/06), by the German-Israeli Foundation (GIF) for Scientific Research and Development, and by the Minerva foundation with funding from the Federal German Ministry for Education and Research.
Conventions
===========
Starting with the action , let us separate it to the physical coupling part plus counterterms, $\delta Z_x = Z_x - 1$, $\delta\alpha = \frac{1}{2\gamma_R} - \frac{1}{2\alpha}$, so that $$\begin{aligned}
S &=
S_\mathrm{CS}^\mathrm{phys.}
+ S_\mathrm{gh}^\mathrm{phys.}
+ S_\mathrm{b}^\mathrm{phys.}
+ S_\mathrm{CS}^\mathrm{c.t.}
+ S_\mathrm{gh}^\mathrm{c.t.}
+ S_\mathrm{b}^\mathrm{c.t.} \, ,\\\nonumber
\\
S_\mathrm{CS}^\mathrm{phys.} &= \int \! d^dx\, \left\{
- \frac{i}{2}
\epsilon_{\mu\nu\lambda} A_\mu^a \partial_\nu A_\lambda^a
- \frac{i}{6} \mu^{\epsilon/2} g \epsilon_{\mu\nu\lambda} f^{abc}
A_\mu^a A_\nu^b A_\lambda^c
\right\} {\,,}\\
S_\mathrm{gh}^\mathrm{phys.} &= \int \! d^dx\, \left\{
- \frac{1}{2 \alpha} (\partial_\mu A_\mu^a)^2
+ \partial_\mu \bar{c}^a \partial^\mu c^a
+ \mu^{\epsilon/2} g f^{abc}
\partial_\mu \bar{c}^a A_\mu^b c^c
\right\} {\,,}\\
S_\mathrm{b}^\mathrm{phys.} &= \int \! d^dx\, \left\{
\frac{1}{2} (\partial_\mu \phi_i)^2
+ \mu^{\epsilon/2} g \partial_\mu \phi_i T_{ij}^a
A_\mu^a \phi_j
- \frac{1}{4} \mu^\epsilon g^2 \{T^a,T^b\}_{ij}
\phi_i \phi_j A_\mu^a A_\mu^b
\right. \notag\\ &\quad \left. \qquad \qquad
+ \mu^{2\epsilon} \frac{g_6}{3!\cdot 2^3}(\phi_i\phi_i)^3
\right\} {\,,}\\\nonumber
\\
S_\mathrm{CS}^\mathrm{c.t.} &= \int \! d^dx\, \left\{
- \frac{i}{2} \delta Z_A
\epsilon_{\mu\nu\lambda} A_\mu^a \partial_\nu A_\lambda^a
- \frac{i}{6} \mu^{\epsilon/2} g \delta Z_g \epsilon_{\mu\nu\lambda} f^{abc}
A_\mu^a A_\nu^b A_\lambda^c
\right\} {\,,}\\
S_\mathrm{gh}^\mathrm{c.t.} &= \int \! d^dx\, \left\{
- \delta\alpha (\partial_\mu A_\mu^a)^2
+ \delta Z_\mathrm{gh} \partial_\mu \bar{c}^a \partial^\mu c^a
+ \mu^{\epsilon/2} \delta\tilde{Z}_g g f^{abc}
\partial_\mu \bar{c}^a A_\mu^b c^c
\right\} {\,,}\\
S_\mathrm{b}^\mathrm{c.t.} &= \int \! d^dx\, \left\{
\frac{1}{2} \delta Z_\phi (\partial_\mu \phi_i)^2
+ \mu^{\epsilon/2} \delta Z_g' g \partial_\mu \phi_i T_{ij}^a
A_\mu^a \phi_j
- \frac{1}{4} \mu^\epsilon \delta Z_g'' g^2 \{T^a,T^b\}_{ij}
\phi_i \phi_j A_\mu^a A_\mu^b \right. \nonumber\\
&\quad \qquad \qquad
\left.+ \mu^{2\epsilon} \delta Z_{g_6}\frac{g_6}{3!\cdot 2^3}(\phi_i\phi_i)^3
\right\}
{\,.}\end{aligned}$$ We use Landau gauge, $\alpha\to 0$, in which the gluon propagator is $$\begin{aligned}
\label{eq:GluonProp}
- \delta_{ab} \epsilon_{\mu\nu\lambda} \frac{p^\lambda}{p^2} {\,.}\end{aligned}$$ The $O(N)$ generators in the fundamental are taken to be real and anti-symmetric, $(T^a)^{\dag}=(T^a)^T=-T^a$. They satisfy $$\begin{gathered}
\operatorname{Tr}\left(T^aT^b\right)=\delta^{ab}C_1 \,,\quad
f^{acd}f^{bcd}=\delta^{ab}C_2 \,,\quad
T^a_{ij}T^a_{kl}=I_{ij,kl}C_3 \,,\quad \notag\\
f^{abc} T^b_{ik} T^c_{kj} = \frac{1}{2} C_2 T^a_{ij} \, ,\quad
f^{abc} = \operatorname{Tr}\left( T^a [T^b,T^c] \right)
{\,,}\label{eq:TrRel}\end{gathered}$$ where $$\begin{aligned}
\label{eq:TrRelOn}
C_1=C_3=1 {{\,,}\quad}I_{ij,kl}=\frac{1}{2}\left(\delta_{il}\delta_{kj}-\delta_{ik}\delta_{jl}\right) {{\,,}\quad}C_2=2-N {\,.}\end{aligned}$$
We will also be interested in the case of a complex scalar field in the fundamental representation of $U(N)$, again coupled to gauge fields with a Chern-Simons interaction. In this case the scalar action is $$\begin{aligned}
S_\mathrm{b} &= \int \! d^dx \,
\left\{
Z_\phi |\cal{D}_{\mu} \phi_i|^2
+ \mu^{2\epsilon} Z_{g_6} \frac{g_{6}}{3!}
(\phi^\dagger \phi)^3
\right\} {\,,}\label{eq:cSb}\end{aligned}$$ and the generators of $U(N)$ in the fundamental representation satisfy , with $$\begin{aligned}
\label{eq:TrRelUn}
C_1=C_3=1 {{\,,}\quad}I_{ij,kl}=\delta_{il}\delta_{kj}
{{\,,}\quad}C_2=-2N {\,.}\end{aligned}$$ The $SU(N)$ case is identical at large $N$, differing by an extra term in $I_{ij,kl}$. The counterterms for the complex and real theories are related by $$\begin{aligned}
\delta Z^{SU(N)}_{\phi}&= 4\delta Z^{O(N)}_{\phi} {{\,,}\quad}\delta Z^{SU(N)}_{g_6} = 4\delta Z^{O(N)}_{g_6} {\,.}\label{eq:cScalarRen}\end{aligned}$$
2-Loop Diagram Results {#diags}
======================
The following are the diverging parts of the diagrams of the $O(N)$ Chern-Simons-matter theory appearing in sections \[2loop\] and \[J2J1J1\], and in appendix \[adim\].
$$\begin{aligned}
(\mathrm{A}1) &= -g^4g_6{
\left(
\delta_{i_1i_2}\delta_{i_3i_4}\delta_{i_5i_6} + 14 \text{\ perms.}
\right)
}\left(\frac{3}{2}N^2+\frac{21}{2}N-12\right)\frac{1}{64\pi^2}\frac{1}{\epsilon}{\,,}\\
(\mathrm{A}2) &= g^8{
\left(
\delta_{i_1i_2}\delta_{i_3i_4}\delta_{i_5i_6} + 14 \text{\ perms.}
\right)
}\left(N^2+N-2\right)\frac{3}{64\pi^2}\frac{1}{\epsilon}{\,,}\\
(\mathrm{A}3) &= -g^8{
\left(
\delta_{i_1i_2}\delta_{i_3i_4}\delta_{i_5i_6} + 14 \text{\ perms.}
\right)
}\left(N^2-3N+2\right)\frac{3}{64\pi^2}\frac{1}{\epsilon}{\,,}\\
(\mathrm{A}4) &= -g^4g_6{
\left(
\delta_{i_1i_2}\delta_{i_3i_4}\delta_{i_5i_6} + 14 \text{\ perms.}
\right)
}\left(N-1\right)\frac{9}{32\pi^2}\frac{1}{\epsilon}{\,,}\\
(\mathrm{A}5) &= g^8{
\left(
\delta_{i_1i_2}\delta_{i_3i_4}\delta_{i_5i_6} + 14 \text{\ perms.}
\right)
}\left(N-1\right)\frac{3}{64\pi^2}\frac{1}{\epsilon}{\,,}\\
(\mathrm{A}6) &= 0{\,,}\\
(\mathrm{A}7) &= g^8{
\left(
\delta_{i_1i_2}\delta_{i_3i_4}\delta_{i_5i_6} + 14 \text{\ perms.}
\right)
}\left(N-1\right)\frac{9}{32\pi^2}\frac{1}{\epsilon}{\,,}\\
(\mathrm{A}8) &= g_6^2{
\left(
\delta_{i_1i_2}\delta_{i_3i_4}\delta_{i_5i_6} + 14 \text{\ perms.}
\right)
}\left(3N+22\right)\frac{1}{32\pi^2}\frac{1}{\epsilon}{\,.}\end{aligned}$$
$$\begin{aligned}
(\mathrm{B}1) &= -g^4\delta_{ij}p^2\left(N^2-3N+2\right)\frac{1}{96\pi^2}\frac{1}{\epsilon}{\,,}\\
(\mathrm{B}2) &= g^4\delta_{ij}p^2\left(N^2-N\right)\frac{1}{384\pi^2}\frac{1}{\epsilon}{\,,}\\
(\mathrm{B}3) &= g^4\delta_{ij}p^2\left(N-1\right)\frac{1}{48\pi^2}\frac{1}{\epsilon}{\,,}\\
(\mathrm{B}4) &= g^4\delta_{ij}p^2\left(N-1\right)\frac{1}{96\pi^2}\frac{1}{\epsilon}{\,.}\end{aligned}$$
$$\begin{aligned}
(\mathrm{C}1) &= g^4\delta_{i_1i_2}\left(\frac{3}{2}N^2+\frac{21}{2}N-12\right)\frac{1}{96\pi^2}\frac{1}{\epsilon} {\,,}\\
(\mathrm{C}2) &= g^4\delta_{i_1i_2}\left(N-1\right)\frac{3}{16\pi^2}\frac{1}{\epsilon} {\,.}\end{aligned}$$
Let us denote the diagrams (D2),(D3) of section \[J2J1J1\], including permutations, as
----------------------------------- ----------------------------------- -----------------------------------
{height="4.5cm"} {height="4.5cm"} {height="4.5cm"}
(D21) (D22) (D23)
{height="4.5cm"} {height="4.5cm"} {height="4.5cm"}
(D31) (D32) (D33)
----------------------------------- ----------------------------------- -----------------------------------
Their contributions to the parity-violating tensor structure at order $\lambda$, for null polarizations in the limit $x_2 \to x_1$, are given by $$\begin{aligned}
\frac{\lambda}{\sqrt{N}}\frac{({\varepsilon}\cdot x_{12})^2({\varepsilon}\cdot x_{13}){\epsilon}_{\alpha\beta\gamma}x_{12}^{\alpha}x_{13}^{\beta}{\varepsilon}^{\gamma}}{|x_{13}|^6|x_{12}|^6}\end{aligned}$$ times the following factors, $$\begin{aligned}
(\mathrm{D}21) &\to \frac{i}{12\pi^4} {\,,}&
(\mathrm{D}22) &\to 0 {\,,}&
(\mathrm{D}23) &\to 0 {\,,}\\
(\mathrm{D}31) &\to \frac{i}{24\pi^4} {\,,}&
(\mathrm{D}32) &\to \frac{-i}{12\pi^4} {\,,}&
(\mathrm{D}33) &\to 0 {\,.}\end{aligned}$$
Conformal Transformations {#conf}
=========================
The conformal algebra in Euclidean space is $$\begin{aligned}
[M_{\mu\nu},P_\rho] &=
-i(\delta_{\mu\rho}P_\nu - \delta_{\nu\rho}P_\mu) {\,,}&
[M_{\mu\nu},K_\rho] &=
-i(\delta_{\mu\rho}K_\nu - \delta_{\nu\rho}K_\mu) {\,,}\nonumber\\
[D,P_\mu] &= -iP_\mu {\,,}&
[D,K_\mu] &= iK_\mu {\,,}\\
[D,M_{\mu\nu}] &= 0 {\,,}&
[K_\mu,P_\nu] &= 2i(\delta_{\mu\nu}D + M_{\mu\nu}) \nonumber{\,,}\end{aligned}$$ $$\begin{aligned}
[M_{\mu\nu},M_{\rho\sigma}] &=
-i \delta_{\mu\rho} M_{\nu\sigma}
+i \delta_{\nu\rho} M_{\mu\sigma}
+i \delta_{\mu\sigma} M_{\nu\rho}
-i \delta_{\nu\sigma} M_{\mu\rho} {\,.}\end{aligned}$$ The action of $D$ on a local primary operator ${\ensuremath{\mathcal{O}}}(x)$ with dimension $\Delta$ is $$\begin{aligned}
[D,{\ensuremath{\mathcal{O}}}(0)] &= -i \Delta {\ensuremath{\mathcal{O}}}(0) \,.\end{aligned}$$ The Lorentz generators in the vector representation are $$\begin{aligned}
({\tilde M}_{\mu\nu})_{\alpha\beta} &=
i ( \delta_{\mu\alpha} \delta_{\nu\beta}
- \delta_{\mu\beta} \delta_{\nu\alpha} )
{\,,}\end{aligned}$$ and their action on a tensor operator $J_{\rho_1 \cdots \rho_n}$ is $$\begin{aligned}
[M_{\mu\nu}, J_{\rho_1 \cdots \rho_n}] &=
-({\tilde M}_{\mu\nu})_{\rho_1 \alpha} J_{\alpha \rho_2 \cdots \rho_n}
- \cdots
-({\tilde M}_{\mu\nu})_{\rho_n \alpha} J_{\rho_1 \rho_2 \cdots \alpha}
{\,.}\end{aligned}$$
Anomalous Dimension of $\phi^i\phi^i$ {#adim}
=====================================
In this appendix we verify explicitly that $J_0 = \phi^i \phi^i / \sqrt{N}$ does not receive an anomalous dimension at two loops and infinite $N$, in accordance with the general results of sections \[allorders\] and \[HScurrents\]. To compute the anomalous dimension of $J_0$ we consider the correlator $$\begin{aligned}
\langle \phi^2(x)\phi^{i_1}(x_1)\phi^{i_2}(x_2)\rangle_{\mathrm{amp.}}
\label{eq:ADcorr}\end{aligned}$$ in momentum space. The following two diagrams contribute to the divergence:
--------------------------------- ---------------------------------
{width="3.0cm"} {width="3.0cm"}
(C1) (C2)
--------------------------------- ---------------------------------
The divergent parts of these diagrams appear in Appendix \[diags\]. Only (C1) contains a planar diagram, and its contribution at large $N$ to the correlator with amputated $\phi$ legs is $$\begin{aligned}
\lambda^2 \delta_{i_1i_2} \frac{1}{64\pi^2} \frac{1}{\epsilon}{\,.}\label{adcon}\end{aligned}$$ The bare amputated correlator is related to the amputated correlator of the physical theory by $$\begin{gathered}
\left< \phi^2 \phi^{i_1} \phi^{i_2} \right>_{\text{phys.}} =
\frac{Z_\phi}{Z_{\phi^2}}
\left< \phi^2 \phi^{i_1} \phi^{i_2} \right>_{\text{bare}} {\,,}\label{corrrel}\end{gathered}$$ where $J_0^{\text{bare}} = Z_{\phi^2} J_0^{\text{phys.}}$, and $Z_{\phi}=1-\lambda^2\frac{1}{128} \frac{1}{\epsilon}+O(1/N)$ (see ). This should not have any divergence. Using , and noting that the amputated correlator equals ($2 \delta_{i_1 i_2}$) at tree-level, the divergence of at the large $N$ limit, can be seen to be $$\begin{aligned}
-2 \delta_{i_1i_2} \delta Z_{\phi^2} + O(\lambda^3) {\,.}\end{aligned}$$ Therefore, to leading order in $1/N$, $\delta Z_{\phi^2} = O(\lambda^3)$ and $\phi^2$ has no anomalous dimension.
For the theory of a complex scalar field in the fundamental representation of $U(N)$, using the relations and the fact that the tree-level correlator equals 1, we also reach the conclusion that the operator $|\phi|^2$ has no anomalous dimension.
[99]{}
J. M. Maldacena, “The Large N limit of superconformal field theories and supergravity,” Adv. Theor. Math. Phys. [**2** ]{} (1998) 231-252. \[[hep-th/9711200](http://arxiv.org/abs/hep-th/9711200)\].
B. Sundborg, “Stringy gravity, interacting tensionless strings and massless higher spins,” Nucl. Phys. Proc. Suppl. [**102** ]{} (2001) 113-119. \[[hep-th/0103247](http://arxiv.org/abs/hep-th/0103247)\].
E. Sezgin, P. Sundell, “Massless higher spins and holography,” Nucl. Phys. [**B644** ]{} (2002) 303-370. \[[hep-th/0205131](http://arxiv.org/abs/hep-th/0205131)\].
I. R. Klebanov, A. M. Polyakov, “AdS dual of the critical O(N) vector model,” Phys. Lett. [**B550** ]{} (2002) 213-219. \[[hep-th/0210114](http://arxiv.org/abs/hep-th/0210114)\].
E. S. Fradkin, M. A. Vasiliev, “On the Gravitational Interaction of Massless Higher Spin Fields,” Phys. Lett. [**B189** ]{} (1987) 89-95.
M. A. Vasiliev, “Higher spin gauge theories: Star product and AdS space,” In \*Shifman, M.A. (ed.): The many faces of the superworld\* 533-610. \[[hep-th/9910096](http://arxiv.org/abs/hep-th/9910096)\].
S. Giombi, X. Yin, “Higher Spin Gauge Theory and Holography: The Three-Point Functions,” JHEP [**1009** ]{} (2010) 115. \[[arXiv:0912.3462](http://arxiv.org/abs/0912.3462) \[hep-th\]\].
S. Giombi, X. Yin, “Higher Spins in AdS and Twistorial Holography,” JHEP [**1104** ]{} (2011) 086. \[[arXiv:1004.3736](http://arxiv.org/abs/1004.3736) \[hep-th\]\].
S. Giombi, X. Yin, “On Higher Spin Gauge Theory and the Critical O(N) Model,” \[[arXiv:1105.4011](http://arxiv.org/abs/1105.4011) \[hep-th\]\].
S. R. Das, A. Jevicki, “Large N collective fields and holography,” Phys. Rev. [**D68** ]{} (2003) 044011. \[[hep-th/0304093](http://arxiv.org/abs/hep-th/0304093)\].
R. d. M. Koch, A. Jevicki, K. Jin, J. P. Rodrigues, “$AdS_4/CFT_3$ Construction from Collective Fields,” Phys. Rev. [**D83** ]{} (2011) 025006. \[[arXiv:1008.0633](http://arxiv.org/abs/1008.0633) \[hep-th\]\].
M. R. Douglas, L. Mazzucato, S. S. Razamat, “Holographic dual of free field theory,” Phys. Rev. [**D83** ]{} (2011) 071701. \[[arXiv:1011.4926](http://arxiv.org/abs/1011.4926) \[hep-th\]\].
A. Jevicki, K. Jin, Q. Ye, “Collective Dipole Model of AdS/CFT and Higher Spin Gravity,” \[[arXiv:1106.3983](http://arxiv.org/abs/1106.3983) \[hep-th\]\].
W. Chen, G. W. Semenoff, Y. -S. Wu, “Two loop analysis of nonAbelian Chern-Simons theory,” Phys. Rev. [**D46** ]{} (1992) 5521-5539. \[[hep-th/9209005](http://arxiv.org/abs/hep-th/9209005)\].
L. V. Avdeev, D. I. Kazakov, I. N. Kondrashuk, “Renormalizations in supersymmetric and nonsupersymmetric nonAbelian Chern-Simons field theories with matter,” Nucl. Phys. [**B391** ]{} (1993) 333-357.
S. Deser, R. Jackiw, S. Templeton, “Topologically Massive Gauge Theories,” Annals Phys. [**140**]{}, 372-411 (1982).
S. Deser, R. Jackiw, S. Templeton, “Three-Dimensional Massive Gauge Theories,” Phys. Rev. Lett. [**48**]{}, 975-978 (1982).
W. Siegel, “Supersymmetric Dimensional Regularization via Dimensional Reduction,” Phys. Lett. [**B84** ]{} (1979) 193.
S. S. Gubser, I. R. Klebanov, “A Universal result on central charges in the presence of double trace deformations,” Nucl. Phys. [**B656** ]{} (2003) 23-36. \[[hep-th/0212138](http://arxiv.org/abs/hep-th/0212138)\].
A. C. Petkou, “Evaluating the AdS dual of the critical O(N) vector model,” JHEP [**0303** ]{} (2003) 049. \[[hep-th/0302063](http://arxiv.org/abs/hep-th/0302063)\].
V. S. Alves, M. Gomes, S. L. V. Pinheiro, A. J. da Silva, “A Renormalization group study of the (phi\* phi)\*\*3 model coupled to a Chern-Simons field,” Phys. Rev. [**D61** ]{} (2000) 065003. \[[hep-th/0001221](http://arxiv.org/abs/hep-th/0001221)\].
L. C. de Albuquerque, M. Gomes, A. J. da Silva, “Renormalization group study of Chern-Simons field coupled to scalar matter in a modified BPHZ subtraction scheme,” Phys. Rev. [**D62** ]{} (2000) 085005. \[[hep-th/0005258](http://arxiv.org/abs/hep-th/0005258)\].
S. Weinberg, “The quantum theory of fields. Vol. 2: Modern applications,” Cambridge, UK: Univ. Pr. (1996) 489 p.
W. A. Bardeen, M. Moshe, M. Bander, “Spontaneous Breaking of Scale Invariance and the Ultraviolet Fixed Point in O($n$) Symmetric $(phi^{6}$ in Three-Dimensions) Theory,” Phys. Rev. Lett. [**52** ]{} (1984) 1188.
M. Moshe, J. Zinn-Justin, “Quantum field theory in the large N limit: A Review,” Phys. Rept. [**385** ]{} (2003) 69-228. \[[hep-th/0306133](http://arxiv.org/abs/hep-th/0306133)\].
J. H. Schwarz, “Superconformal Chern-Simons theories,” JHEP [**0411** ]{} (2004) 078. \[[hep-th/0411077](http://arxiv.org/abs/hep-th/0411077)\].
D. Gaiotto, X. Yin, “Notes on superconformal Chern-Simons-Matter theories,” JHEP [**0708** ]{} (2007) 056. \[[arXiv:0704.3740](http://arxiv.org/abs/0704.3740) \[hep-th\]\].
D. J. Amit, E. Rabinovici, “BREAKING OF SCALE INVARIANCE IN phi\*\*6 THEORY: TRICRITICALITY AND CRITICAL END POINTS,” Nucl. Phys. [**B257**]{}, 371 (1985).
A. G. Dias, M. Gomes, A. J. da Silva, “Dynamical breakdown of symmetry in (2+1) dimensional model containing the Chern-Simons field,” Phys. Rev. [**D69** ]{} (2004) 065011. \[[hep-th/0305043](http://arxiv.org/abs/hep-th/0305043)\].
A. G. Dias, A. F. Ferrari, “Renormalization Group and Conformal Symmetry Breaking in the Chern-Simons Theory Coupled to Matter,” Phys. Rev. [**D82** ]{} (2010) 085006. \[[arXiv:1006.5672](http://arxiv.org/abs/1006.5672) \[hep-th\]\].
E. Rabinovici, M. Smolkin, “On the dynamical generation of the Maxwell term and scale invariance,” JHEP [**1107**]{}, 040 (2011). \[[arXiv:1102.5035](http://arXiv.org/abs/arXiv:1102.5035) \[hep-th\]\].
A. F. Ferrari, E. A. Gallegos, M. Gomes, A. C. Lehum, J. R. Nascimento, A. Y. .Petrov, A. J. da Silva, “Coleman-Weinberg mechanism in a three-dimensional supersymmetric Chern-Simons-Matter model,” Phys. Rev. [**D82**]{}, 025002 (2010). \[[arXiv:1004.0982](http://arXiv.org/abs/arXiv:1004.0982) \[hep-th\]\].
L. Girardello, M. Porrati, A. Zaffaroni, “3-D interacting CFTs and generalized Higgs phenomenon in higher spin theories on AdS,” Phys. Lett. [**B561** ]{} (2003) 289-293. \[[hep-th/0212181](http://arxiv.org/abs/hep-th/0212181)\].
W. Heidenreich, “Tensor Products Of Positive Energy Representations Of So(3,2) And So(4,2),” J. Math. Phys. [**22** ]{} (1981) 1566.
S. Elitzur, A. Giveon, M. Porrati, E. Rabinovici, “Multitrace deformations of vector and adjoint theories and their holographic duals,” JHEP [**0602** ]{} (2006) 006. \[[hep-th/0511061](http://arxiv.org/abs/hep-th/0511061)\].
E. Witten, “Multitrace operators, boundary conditions, and AdS / CFT correspondence,” \[[hep-th/0112258](http://arxiv.org/abs/hep-th/0112258)\].
M. Berkooz, A. Sever, A. Shomer, “’Double trace’ deformations, boundary conditions and space-time singularities,” JHEP [**0205** ]{} (2002) 034. \[[hep-th/0112264](http://arxiv.org/abs/hep-th/0112264)\].
S. Giombi, S. Minwalla, S. Prakash, S. P. Trivedi, S. R. Wadia, X. Yin, “Chern-Simons Theory with Vector Fermion Matter,” \[[arXiv:1110.4386](http://arXiv.org/abs/arXiv:1110.4386) \[hep-th\]\].
S. Giombi, S. Prakash, X. Yin, “A Note on CFT Correlators in Three Dimensions,” \[[arXiv:1104.4317](http://arxiv.org/abs/1104.4317) \[hep-th\]\].
E. E. Boos, A. I. Davydychev, “A Method Of The Evaluation Of The Vertex Type Feynman Integrals,” Moscow Univ. Phys. Bull. [**42N3** ]{} (1987) 6-10.
S. H. Shenker, X. Yin, “Vector Models in the Singlet Sector at Finite Temperature,” \[[arXiv:1109.3519](http://arxiv.org/abs/1109.3519) \[hep-th\]\].
R. Gopakumar, C. Vafa, “On the gauge theory / geometry correspondence,” Adv. Theor. Math. Phys. [**3** ]{} (1999) 1415-1443. \[[hep-th/9811131](http://arxiv.org/abs/hep-th/9811131)\].
S. Sinha, C. Vafa, “SO and Sp Chern-Simons at large N,” \[[hep-th/0012136](http://arxiv.org/abs/hep-th/0012136)\].
R. G. Leigh, A. C. Petkou, “Holography of the N=1 higher spin theory on AdS(4),” JHEP [**0306** ]{} (2003) 011. \[[hep-th/0304217](http://arxiv.org/abs/hep-th/0304217)\].
E. Sezgin, P. Sundell, “Holography in 4D (super) higher spin theories and a test via cubic scalar couplings,” JHEP [**0507** ]{} (2005) 044. \[[hep-th/0305040](http://arxiv.org/abs/hep-th/0305040)\].
[^1]: The Chern-Simons action is required to make the operator $F_{\mu \nu}(x)$ trivial, to ensure that we do not add any additional local operators to the theory beyond the singlets of the vector model.
[^2]: For the Abelian Chern-Simons theory coupled to a scalar field, such corrections were studied in [@Alves:1999hw; @deAlbuquerque:2000ec].
[^3]: These diagrams were already computed in [@Chen:1992ee].
[^4]: This theory has a “parity anomaly” which means that $k$ must be a half-integer rather than an integer, but this will not affect our large $N$ discussion here.
[^5]: The $U(1)$ flavor symmetry is in fact part of the gauged $U(N)$ symmetry. Nevertheless, in the large $N$ limit that we are interested in, we can obtain the same results by gauging only an $SU(N)$ group, and then the $U(1)$ is a global symmetry.
[^6]: There are also diagrams that include $(\phi^\dagger\phi)^3$ vertices, but they have tadpole matter loops, and all such loops vanish in our regularization scheme.
[^7]: Note that we are using here the fact that we only have marginal couplings. In a theory with relevant operators like $(\phi^2)^2$, anomalous dimensions can arise even from finite diagrams, but this is not true in our case.
[^8]: We use a normalization in which the 2-point functions remain finite in the large $N$ limit.
[^9]: We thank S. Minwalla for discussions on this issue.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'H. Jönsson'
- 'N. Ryde'
- 'M. Schultheis'
- 'M. Zoccali'
bibliography:
- '/Users/henrik/Documents/Bibliografi/papers.bib'
- '/Users/henrik/Documents/Bibliografi/kurucz.bib'
- '/Users/henrik/Documents/Bibliografi/sulphur.bib'
- '/Users/henrik/Documents/Bibliografi/GESreferencesv5all.bib'
date: 'Submitted 2016; accepted 2016'
subtitle: 'II. O, Mg, Ca, and Ti in the bulge sample'
title: 'Abundances of disk and bulge giants from hi-res optical spectra[^1]'
---
Introduction {#sec:introduction}
============
The Galactic bulge holds a significant part of the stars of our Galaxy, but its history and evolution is still unknown. From cosmological $\Lambda$CDM-models it is expected that the bulge was formed via mergers of smaller dwarf galaxies, but recently it has been repeatedly shown that a major part of the bulge is dynamically formed from the inner disk, for example the fact that it shows cylindrical rotation [@2007ApJ...658L..29R; @2012AJ....143...57K], and that it has two red clumps [@2010ApJ...724.1491M]. On the other hand, the two red clumps might not be visible for lower-metallicity stars [@2012ApJ...756...22N; @2013MNRAS.432.2092N], and old RR Lyrae stars trace a component that is less elongated and is rotating slower [@2013ApJ...776L..19D; @2016ApJ...821L..25K], suggesting that there possibly is an old spheroid-bulge co-existing with the dynamically formed bar.
@2012MNRAS.421..333S predict that a classical merger-formed spheroidal bulge, if present, would have been spun up to bar-kinematics and therefore impossible/hard to find using kinematics alone, but the possible two stellar populations need to be distinguished by chemical differences. This is something that has been attempted several times during recent years: for example measured O, Na, Mg, and Al in 53 bulge giants, 35 of which overlap with our sample. determined O, Na, Mg, Al, Si, Ca, and Ti in 25 bulge giants and when comparing to 55 similarly analyzed local giants (thin disk, thick disk, and halo), they find that the bulge has had a chemical evolution similar to the local thick disk. More recently used micro-lensing to observe 58 bulge dwarf and sub-giant stars, finding a wide age distribution with several young stars, a broad MDF, possibly including several components, and, when comparing to a similar sample of Solar neighborhood stars, they conclude that the ‘knee’ in the \[alpha/Fe\] vs. \[Fe/H\] plots is shifted to $\sim0.1$ dex higher metallicities in the bulge, suggesting a faster chemical enrichment. Even more recently, analyzed slightly lower resolution spectra (the high-resolution GIBS-sample using FLAMES GIRAFFE with $R\sim 22500$) of 400 bulge giants in four bulge fields, finding a knee in their \[Mg/Fe\] vs \[Fe/H\] plot around \[Fe/H\]$\simeq-0.44$ dex, approximately 0.1 dex lower than . However, they lack a large enough, similarly observed and analyzed, solar neighborhood sample of giants to compare their bulge results to. They are therefore not able to conclude whether the difference in their sample compared to that of the microlensed dwarfs is due to systematic differences between the studies, or due to the very differently sized samples. @2014AJ....148...67J determine O, Na, Mg, Al, Si, Ca, Cr, Fe, Co, Ni, and Cu in 156 giants, using FLAMES GIRAFFE $R\sim22500$ spectra, in two bulge fields in common with ours (B3 and BL), finding a higher knee for the bulge giants as compared to literature samples of local dwarf stars . The fact that they are using two different types of stars, giants in the bulge and dwarfs in the local disk, analyzed differently, might introduce systematic differences that could account for the different position of the knee. In spite of the efforts put into these and many more works, still no consensus on the absolute abundance trends of the bulge and its evolution is reached. This is mainly because observing stars in the bulge is hard: it is situated relatively far away and covered behind dust in the disk. To handle these problems, one could do one or more of the following: go down in spectral resolution, thereby sacrificing abundance precision, possibly only enabling determination of the general metallicity (for example the low-resolution GIBS-sample, ), observe in the infrared, where determining the stellar parameters still is a problem (for example @2012ApJ...746...59R), use microlensing events, thereby not being able to select your targets and their positions , and/or use long integration times (as is done in for example and this work).
This paper (hereafter, Paper II) is the second in a series determining abundances of bulge giants from optical high-resolution spectra ($R\sim47000$). Because of the long integration times needed for observing stars in the bulge at this high resolution not many such observations have been attempted. The spectra that were first used in are therefore an unique dataset that has been analyzed in several subsequent articles: , , , and . re-determine the stellar parameters as derived in the original article of , for a small subset of the stars, and show that their all-spectroscopic approach in some cases is giving significantly different results, possibly influencing some of the abundance determinations and conclusions in , , and . In order to eliminate systematic differences and ensure a homogeneous, differential comparison, we attempt to re-determine these stellar parameters, add eleven similarly observed spectra in a new field even closer to the Galactic center, and determine the alpha abundances oxygen, magnesium, calcium, and titanium. Thereby we will re-determine the oxygen abundances of 35 stars from and magnesium abundances the same 35 stars from , opening up for an interesting comparison between the results. In @2016arXiv161105462J (Paper I) we presented a similarly analyzed local disk sample of 291 similar giants, their stellar parameters and the abundances of oxygen, magnesium, calcium, and titanium. There we found that our stellar parameters were accurate and precise with a low dispersion compared to benchmark values based on fundamentally determined stellar parameters, such as effective temperatures from angular diameter measurements [@2003AJ....126.2502M], and surface gravities from asteroseismic measurements . Furthermore, we found that the derived abundance trends show similar scatter as the trends of other Solar neighborhood works using dwarfs . In this paper we will determine the same abundances for a bulge sample of 46 giants of similar type as the previously published Solar neighborhood sample, enabling a differential comparison of abundances in the bulge and in the local disk.
Observations {#sec:observations}
============
We have observed 46 K-giants in the Galactic bulge using the spectrometer FLAMES/UVES mounted on VLT. The basic data of our stars are listed in Table \[tab:basicdata\] and the Figure \[fig:bulge\_fields\] shows the location of our five fields (SW, B3, BW, B6, and BL) in comparison to the COBE/DIRBE outline of the Galactic bulge [@1994ApJ...425L..81W], the locations of the microlensed bulge dwarfs of , and the high-resolution sample of the GIBS survey . As can be seen in Figure \[fig:bulge\_fields\], the new field in the Sagittarius Window (SW, $(l,b)=(1.25,-2.65)$) is closer to the Galactic centre than the other previously analyzed fields. Furthermore, it can be seen that it is situated in a region where the optical extinction is lower than the surroundings. To go even closer to the Galactic centre, infrared observations are needed due to the optical extinction being to high ; the corresponding infrared extinction in the bulge is is essentially zero outside of the plane (for $b<-1.5$ or $b>1.5$).
---------- -------------- -------------- --------
Star$^a$ RA (J2000) Dec (J2000) $V$
(h:m:s) (d:am:as)
SW-09 17:59:04.533 -29:10:36.53 16.153
SW-15 17:59:04.753 -29:12:14.77 16.326
SW-17 17:59:08.138 -29:11:20.10 16.388
SW-18 17:59:06.455 -29:10:30.53 16.410
SW-27 17:59:04.457 -29:10:20.67 16.484
SW-28 17:59:07.005 -29:13:11.35 16.485
SW-33 17:59:03.331 -29:10:25.60 16.549
SW-34 17:58:54.418 -29:11:19.82 16.559
SW-43 17:59:04.059 -29:13:30.26 16.606
SW-71 17:58:58.257 -29:12:56.97 16.892
SW-76 17:58:54.192 -29:12:09.31 16.943
B3-b1 18:08:15.840 -25:42:09.83 18.653
B3-b5 18:09:00.527 -25:48:06.78 18.345
B3-b7 18:09:16.540 -25:49:26.08 18.659
B3-b8 18:08:24.602 -25:48:44.39 18.915
B3-f1 18:08:16.176 -25:43:19.18 18.514
B3-f2 18:09:15.609 -25:57:32.75 18.924
B3-f3 18:08:49.628 -25:40:36.93 18.575
B3-f4 18:08:44.293 -26:00:25.05 18.650
B3-f7 18:09:23.694 -25:50:38.19 18.304
B3-f8 18:08:12.632 -25:50:04.45 18.490
BW-b1 18:03:34.710 -29:54:33.80 19.222
BW-b2 18:04:23.950 -30:05:57.80 18.484
BW-b5 18:04:13.270 -29:58:17.80 18.832
BW-b6 18:03:51.840 -30:06:27.90 18.410
BW-b8 18:04:02.870 -30:01:29.20 18.146
BW-f1 18:03:37.140 -29:54:22.30 17.994
BW-f5 18:04:39.620 -29:55:19.80 17.481
BW-f6 18:03:36.890 -30:07:04.30 18.387
BW-f7 18:04:43.920 -30:03:15.20 18.106
B6-b1 18:09:50.480 -31:40:51.61 17.995
B6-b3 18:10:19.060 -31:40:28.19 17.705
B6-b4 18:10:07.770 -31:52:41.36 17.842
B6-b5 18:10:37.380 -31:40:29.14 17.711
B6-b6 18:09:49.100 -31:50:07.66 17.793
B6-b8 18:09:55.950 -31:45:46.33 18.397
B6-f1 18:10:04.460 -31:41:45.31 17.901
B6-f3 18:10:17.720 -31:41:55.20 17.255
B6-f5 18:10:41.510 -31:40:11.88 17.632
B6-f7 18:10:52.300 -31:46:42.18 17.841
B6-f8 18:09:56.840 -31:43:22.56 17.263
BL-1 18:34:58.510 -34:33:15.24 16.905
BL-3 18:35:27.510 -34:31:59.36 16.884
BL-4 18:35:21.110 -34:44:48.22 16.451
BL-5 18:36:01.010 -34:31:47.91 16.911
BL-7 18:35:57.260 -34:38:04.61 16.579
---------- -------------- -------------- --------
: Basic data for the observed bulge giants.
\[tab:basicdata\]
{width="170mm"}
The spectra of the 35 stars in the B3, BW, B6, and BL fields analyzed here, are the same that have been analyzed for O in , Na, Mg, Al in , Mn , and Ba, La, Ce, Nd, Eu in . These observations were carried out May-Aug 2003-2004. Since the fiber-array FLAMES was used in combination with UVES, seven stars could be observed in each pointing. Four spectra have been excluded from the analysis because the determined stellar parameters are outside of the parameter space tested in Paper I: for one star, B3-b3, we derive a large $\log g$ of 3.23, making it possible to be a foreground disk star, for two stars, B3-b4 and B6-f2, we need to use atmospheric turbulence parameters outside of the ranges $1.0<v_{\mathrm{mic}}<2.0$ and $1.0<v_{\mathrm{mac}}<8.0$ that is making the determination of the surface gravity uncertain, and for one star, BW-f4, we derive a \[Fe/H\]$=-1.55$, that is lower than any of the stars analyzed in Paper I, and we are not certain how well our stellar parameter determination works in this regime.
The 11 new stars in Sagittarius Window were observed in the same way using the same telescope, instrument, and setting in service mode during Aug 2011 (ESO program 085.B-0552(A)).
The total integration time in each setting was 5-12 hours depending on extinction. The achieved S/N is listed in Table \[tab:stellarparams\]. The resolving power of the spectra is 47000 and the spectra cover the region 5800 Å to 6800 Å.
Analysis and results {#sec:analysis}
====================
The spectra were analyzed using the exact same techniques and spectral lines used to analyze the Solar neighborhood stars in Paper I. In short, the software `Spectroscopy Made Easy`, SME was used to, via $\chi^2$-minimization of a synthetic spectrum and the observed spectrum, determine the stellar parameters as well as the abundances. In the analysis we used spherical symmetric, \[$\alpha$/Fe\]-enhanced, LTE MARCS-models. Furthermore, NLTE-corrections were used for the iron-lines [@2012MNRAS.427...50L].
Reference sample
----------------
The main point of Paper I was to analyze a Solar neighborhood reference sample of giant stars similar to the bulge stars analyzed here. The HR-diagram of this Solar neighborhood reference sample and the Solar neighborhood dwarf stars of are shown in the leftmost panel of Figure \[fig:hr-dia\].
{width="180mm"}
The bulge sample
----------------
In order to enable a strictly differential comparison to the reference sample, we have re-determined the stellar parameters as well as the abundances also for the previously published B3-BW-B6-BL-stars using the exact same purely spectroscopic analysis as is used for the Solar neighborhood reference sample of Paper I. The resulting stellar parameters are plotted as a HR-diagram in the middle panel of Figure \[fig:hr-dia\]. As a comparison, the parameters for the same stars as determined in are shown in the rightmost panel. As can be seen, the largest differences between the two sets of stellar parameters are seen in the surface gravities, where our results are spread out along the red giant branch, while this is not shown in the previous stellar parameters of . Furthermore, our results, in contrast to the older parameters, are sorted in \[Fe/H\] as expected in Figure \[fig:hr-dia\]. This, together with the careful evaluation of the method used (in Paper I), gives us confidence in our determined stellar parameters.
The determined parameters and abundances are listed in Table \[tab:stellarparams\]. Furthermore, the determined abundances are shown in Figure \[fig:all\_abundances\] together with the microlensed bulge dwarfs of , the abundances of @2014AJ....148...67J, and the GIBS survey .
------- ------------------ ---------- ---------- ---------------------- ----- ----------------------------- ------------------------------ ------------------------------ ------------------------------
Star $T_\textrm{eff}$ $\log g$ \[Fe/H\] $\xi_\textrm{micro}$ S/N $\log \epsilon(\mathrm{O})$ $\log \epsilon(\mathrm{Mg})$ $\log \epsilon(\mathrm{Ca})$ $\log \epsilon(\mathrm{Ti})$
\[K\] (cgs) \[kms$^{-1}$\]
SW-09 4095 1.79 -0.18 1.32 16 8.47 7.59 6.16 5.06
SW-15 4741 1.96 -1.01 1.62 15 ... 7.13 5.65 ...
SW-17 4245 2.09 0.21 1.44 11 8.93 7.84 6.61 5.21
SW-18 4212 1.67 -0.16 1.49 14 ... 7.71 6.25 4.86
SW-27 4423 2.34 0.08 1.60 13 8.76 7.93 6.43 5.12
SW-28 4254 2.36 -0.17 1.44 16 8.78 7.84 6.07 4.94
SW-33 4580 2.72 0.13 1.39 14 8.97 7.80 6.36 5.01
SW-34 4468 1.75 -0.48 1.63 12 ... 7.75 6.06 4.68
SW-43 4892 2.34 -0.80 1.84 16 8.41 7.34 5.76 ...
SW-71 4344 2.66 0.36 1.31 14 9.11 ... 6.64 5.32
SW-76 4427 2.45 0.10 2.00 12 9.02 7.76 6.63 5.22
B3-b1 4414 1.35 -0.92 1.41 21 8.22 7.38 5.92 4.28
B3-b5 4425 2.70 0.22 1.43 43 8.87 7.99 6.50 5.17
B3-b7 4303 2.36 0.05 1.58 38 8.80 7.77 6.42 5.07
B3-b8 4287 1.79 -0.70 1.46 65 8.47 7.27 5.88 4.43
B3-f1 4485 2.25 -0.18 1.88 31 8.74 7.81 6.31 5.02
B3-f2 4207 1.64 -0.69 1.74 22 ... 7.55 5.96 4.74
B3-f3 4637 2.96 0.21 1.89 31 8.98 8.00 6.49 5.14
B3-f4 4319 2.60 -0.15 1.50 11 8.77 ... 6.20 5.03
B3-f7 4517 2.93 0.14 1.55 24 ... 7.89 6.44 5.18
B3-f8 4436 2.88 0.21 1.54 63 8.79 7.96 6.50 5.23
BW-b1 4042 2.39 0.43 1.43 29 ... 8.07 6.58 5.47
BW-b2 4367 2.39 0.15 1.68 20 ... 7.99 6.50 5.17
BW-b5 3939 1.68 0.22 1.31 43 ... 7.82 6.47 5.22
BW-b6 4262 1.98 -0.35 1.44 23 8.60 7.72 6.31 4.97
BW-b8 4424 2.54 0.27 1.52 44 ... 7.99 6.52 5.14
BW-f1 4359 2.51 0.25 1.93 37 8.96 8.15 6.59 5.26
BW-f5 4818 2.89 -0.54 1.29 39 ... 7.37 6.08 4.65
BW-f6 4117 1.43 -0.46 1.69 33 8.55 7.73 6.09 4.60
BW-f7 4592 2.96 0.53 1.50 15 9.10 8.05 6.71 5.69
B6-b1 4372 2.59 0.22 1.57 51 ... 7.87 6.50 5.13
B6-b3 4468 2.48 0.02 1.67 59 8.91 7.82 6.33 5.06
B6-b4 4215 1.38 -0.65 1.68 41 8.43 7.38 5.87 4.43
B6-b5 4340 2.02 -0.51 1.34 54 8.49 7.66 6.09 4.71
B6-b6 4396 2.37 0.16 1.77 44 8.86 7.95 6.49 5.16
B6-b8 4021 1.90 0.03 1.45 55 8.68 7.71 6.36 5.10
B6-f1 4149 2.01 0.07 1.65 77 8.84 7.83 6.37 5.06
B6-f3 4565 2.60 -0.38 1.28 82 8.63 7.57 6.14 4.80
B6-f5 4345 2.32 -0.36 1.41 32 ... 7.68 6.15 4.88
B6-f7 4250 2.10 -0.34 1.65 29 ... 7.69 6.16 4.91
B6-f8 4470 2.78 0.10 1.30 81 8.89 7.81 6.45 5.10
BL-1 4370 2.19 -0.22 1.50 38 ... 7.57 6.25 4.87
BL-3 4555 2.48 -0.12 1.53 57 8.74 7.73 6.26 4.88
BL-4 4476 2.94 0.24 1.41 36 8.93 8.00 6.63 5.20
BL-5 4425 2.65 0.25 1.68 58 8.91 8.05 6.60 5.25
BL-7 4776 2.52 -0.53 1.53 60 ... 7.49 6.07 4.68
------- ------------------ ---------- ---------- ---------------------- ----- ----------------------------- ------------------------------ ------------------------------ ------------------------------
\[tab:stellarparams\]
{width="170mm"}
Discussion {#sec:discussion}
==========
We are not able to see any trends in metallicity nor abundances across the different fields, why we in the following handle the entire sample as a bulge sample. To see possible trends more stars in every field are needed.
It is hard to estimate the age of giant stars, but from the isochrones in Figure \[fig:hr-dia\], one can see a slight splitting up with respect of age for the giants with highest gravities (close to $\log g=3$). This might possibly be visible in a slight split/spread of the solar metallicity (orange in the plot) stars of the solar neighborhood sample, while the same effect is not clearly visible in the bulge sample, which is expected since the bulge stars are predominately old [e.g., @2008ApJ...684.1110C], . From Figure \[fig:hr-dia\] it is also obvious that the bulge stars generally are more metal-rich than the giants found in the Solar neighborhood.
Comparison to other studies
---------------------------
From Figure \[fig:all\_abundances\], one can see that the trends of our stars and the microlensed dwarfs are quite similar, and the scatter seem to be rather similar, with our stars possibly showing marginally higher scatter. Our \[O/Fe\] vs. \[Fe/H\] trend is much less scattered and less steep than that of @2014AJ....148...67J. The differences are likely due to the large uncertainties inherent in determining the oxygen abundance from the 6300 Å \[\]-line in the relatively low resolution spectra of @2014AJ....148...67J. For example, avoid determining the oxygen abundance all-together from the exact same data due to these uncertainties. When it comes to our \[Ca/Fe\] vs. \[Fe/H\] trend, it is much tighter and less alpha enhanced than the corresponding trends of @2014AJ....148...67J and . These differences may be attributed to our stellar parameters being more accurate due to our larger wavelength coverage, higher resolution, and thorough tests of our method in Paper I. On the other hand, the \[Mg/Fe\] vs. \[Fe/H\] trends of @2014AJ....148...67J and are less alpha enhanced and tighter than ours. The reason for our data showing a larger scatter is not clear: all three works use the same three lines around 6318-19 Å and our data has higher resolution suggesting that our data, at least theoretically, should be of higher quality. However, these three lines have several difficulties: first of all, they have uncertain $gf$-values. @2014AJ....148...67J and use astrophysical values and we use the (very similar) results of @2016arXiv161107000P. Secondly, the lines are affected by an autoionizing -line producing a very wide depression of the spectrum. @2014AJ....148...67J, like us, solve this problem by setting a local pseudo-continuum around the lines, while both model the autoionizing -line using their determined Ca-abundance, *and* place a local continuum to get rid of possible residual mis-matches between the observed and synthetic spectra. Thirdly, the lines are in a region affected by telluric lines. @2014AJ....148...67J remove these by division by an observed ‘telluric’ spectrum, while we and simply avoid using the Mg-lines visibly affected by telluric contamination. To conclude, a possible explanation for our more scattered \[Mg/Fe\] vs. \[Fe/H\] trend might be that the (necessary) lower S/N of our higher resolution data makes the continuum-placement more difficult, and our tendency to derive higher \[Mg/Fe\] might be due to the lower S/N making it harder to identify and avoid telluric lines, implying that we would derive too high magnesium abundances in the cases where we possibly fail to identify a telluric line.
Several previous studies, but not all, have found different trends in \[O/Fe\] and \[Mg/Fe\] in the bulge, which is often attributed to a higher degree massive stars in the bulge compared to the Solar neighborhood. One example is , who used 35 of the same spectra as we do, opening up for an interesting comparison. Therefore, we have plotted the oxygen abundances of (these are the very same abundances also presented in ), and the magnesium abundances of in Figure \[fig:sn-bulge\_abundances\]. Our re-analysis show a similar, but slightly less scattered, oxygen-trend as , while our magnesium-trend is lower in \[Mg/Fe\] for \[Fe/H\]$>-0.5$, showing a rather thick-disk-like trend at odds with what is found in . We believe that most of these differences can be attributed to our new all-spectroscopic stellar parameters, but also to the different handling of the autoionizing -line affecting the derived magnesium abundances: model this line to get rid of its influence in spite of its uncertain spectroscopic data, while we avoid synthesizing it and instead place a local pseudo-continuum around the three -lines (for some example of their modeling of this line, see Figures 3 and 5 in ).
There are several ongoing large spectroscopic projects surveying the entire Galaxy, and including the bulge. For example the APOGEE survey [@2011AJ....142...72E] has observed a wealth of stars (over 150 000), with several fields towards the bulge. APOGEE has the advantage of observing in the H-band, reducing the problem with extinction of light due to dust, but the rather small diameter of the telescope used, means that the stars that actually are *in* the bulge are the most luminous giants, that are the hardest to analyze. As of yet there has not been any APOGEE-paper on the bulge, but only on the very special stellar population of the absolute galactic centre .
The Gaia ESO-survey has some fields in the bulge, but has sofar not published any comparison between the alpha elemental trends of the local disk and the bulge, but only an investigation on the metallicity and kinematic trends of the bulge .
Comparison to the Solar neighborhood sample
-------------------------------------------
As has been mentioned several times before, the abundance trends found in the bulge must be compared to similarly determined trends in the disk: most importantly the type of stars and the spectral lines used in the analysis should be the same to minimize systematic differences. Ideally also the quality of the spectra should be the same - the resolution and S/N - but this is harder to obtain: it is impossible to obtain the same S/N for the faint bulge giants as for the bright nearby disk giants, and for this difference in magnitude, possibly the same telescope/spectrometer cannot be used in both cases. In our case we used FIES [@2014AN....335...41T] at NOT and data retrieved from the NARVAL and ESPaDOnS spectral archive PolarBase [@2014PASP..126..469P] to collect spectra for the Solar neighborhood sample of Paper I, while we used UVES/FLAMES at VLT for the bulge sample. The FIES and PolarBase spectra have a resolving power of 67000 and 65000, respectively, and high S/N (typically around 100), while the UVES/FLAMES spectra have $R \sim 47000$ and much lower S/N (see Table \[tab:stellarparams\]). The effect of this difference in spectral quality is expected to manifest itself as more scatter in the bulge trends.
In Figure \[fig:sn-bulge\_abundances\], where we compare the abundance trends from our Solar neighborhood sample of Paper I to that of the bulge stars of this article, it is obvious that the abundance trends in the bulge are indeed not as tight as the trends from the Solar neighborhood. Since the type of stars analyzed, and the lines used are the same, this larger spread can only be attributed to lower S/N in the bulge-observations. Looking at Table \[tab:stellarparams\], and comparing to Figure 2 in Paper I where we investigate the impact of S/N on the stellar parameters and the abundances, all SW-stars are expected to have an uncertainty in the \[X/Fe\] abundance ratio of around 0.2 dex (standard deviation) stemming from the S/N alone. The B3-BW-B6-BL-stars generally have higher S/N and are expected to show lower uncertainties due to the S/N, in general around 0.1 dex (standard deviation) for the \[X/Fe\] abundance ratios. As mentioned earlier, the scatter in the \[Mg/Fe\]-trend for the bulge stars seems higher than for the other elements, and is possibly slightly enhanced compared to the thick disk, which is not seen for the other elements. This strongly suggests that the larger scatter in \[Mg/Fe\] for the bulge stars is linked to both the low S/N making it hard to place the continuum and identify telluric lines in the spectrum.
{width="170mm"}
Comparing the abundance trends of Figure \[fig:sn-bulge\_abundances\], we find the bulge trends to generally follow that of the local thick disk, but possibly tracing the upper envelope in the case of magnesium, calcium and titanium, while the bulge oxygen trend more seem to follow the lower envelope of the local thick disk (or upper envelope of the local thin disk). These lower oxygen abundances could potentially be explained by the lower S/N of the bulge stars: from Figure 2 in Paper I, an asymmetry for the lowest S/N is seen in the oxygen abundance, suggesting that lower oxygen abundances are derived for lower S/N. In general, the oxygen abundance is expected to be more sensitive to lower quality of the spectra, since it is based on a single line, in contrast to the magnesium, calcium, and titanium abundances. On the other hand, the oxygen and calcium abundance trends are the tightest of the four, suggesting that the determined surface gravity is precise: the oxygen and calcium abundances are mainly dependent on the surface gravity, as is shown in Table 3 in Paper I for oxygen, and is evident in the case of calcium since it is used to constrain the surface gravity.
In spite of the slight differences of oxygen and magnesium when comparing our bulge sample to our local disk sample in Figure \[fig:sn-bulge\_abundances\], we cannot see any evidence for \[O/Fe\] and \[Mg/Fe\] showing different trends in our bulge data, see Figure \[fig:omg\]. The \[O/Mg\] in our local disk sample is around zero for all metallicities, while the \[O/Mg\] in our bulge data is negative, but still constant. This difference might be on account of our oxygen abundances in the bulge possibly being systematically too low on account of the lower S/N in the bulge spectra. Furthermore, a slightly higher magnesium abundances in the bulge trend can possibly be attributed to difficulties in avoiding telluric lines in the more noisy bulge spectra as mentioned earlier.
{width="170mm"}
From Figure \[fig:sn-bulge\_abundances\], we see no clear evidence for different positions of the knees of the bulge population and the thick disk population, thereby corroborating , but the conclusion is weak. To make a firmer statement, we would ideally need a larger bulge sample extending to lower metallicities, and more thick disk stars in our local sample.
To resolve the question about the possible higher knee in the alpha elemental abundance plots for the bulge as compared to the local thick disk, one would need two decent sized stellar samples: one from the bulge and one from the local disk. This can be reached in several ways, some of which are listed below:
- The investigation of has a large and representative sample of 714 local dwarf stars, while the bulge sample is much smaller with 58 microlensed dwarfs. Ideally, the bulge sample should be enlarged, but with the unpredictability of the microlensing events, this is not readily done. Also, the type of stars in the two samples are not exactly the same with the local sample being F and G dwarf stars, while the bulge sample contains several slightly cooler subgiant stars, see Figure \[fig:hr-dia\].
- The investigation of @2014AJ....148...67J and , on the other hand, both have large bulge sample of hundreds of giant stars, with a very tight and un-scattered \[Mg/Fe\]-trend, see Figure \[fig:all\_abundances\]. However, they both have a small similarly observed and analyzed local disk sample of giants to contrast their bulge-trend to. Furthermore, their other $\alpha$-elements show larger scatter than our trends.
- Our investigation has a local disk sample of 291 giants, but would benefit from having more thick-disk stars. The bulge-sample consists of 46 giants of very similar types to the local sample, see Figure \[fig:hr-dia\].
The strategy used here, observing K-giants in the optical with high resolution spectroscopy, has the upside that it is easy to find and observe suitable, very similar local disk stars and also telescopes/instruments with which to carry out these observations. The downside is the long integration times needed for the bulge observations. However, FLAMES/UVES offers the ability to observe seven stars simultaneously, resulting in about one hour telescope-time per star, similar to the amount of time spent per microlensed dwarf star in .
For the future, a similar methodology as presented here but performed in the near-IR H and/or K bands would be rewarding. This is indeed possible now with the new cross-dispersed high-resolution, near-IR spectrometers recently available or planned for [see e.g. @2014SPIE.9147E..1DP; @2014SPIE.9147E..1EO], but to do so, a serious effort in exploring usable and reliable spectral features in the near-IR needs to be adressed.
Conclusions {#sec:conclusion}
===========
We have determined the abundances of oxygen, magnesium, calcium, and titanium in a sample of 46 bulge K-giants, 35 of which have been analyzed for oxygen and magnesium in previous works , and compare the abundances to those of 291 similarly analyzed K-giants in the solar neighborhood.
To conclude, our re-analysis of the bulge oxygen abundances from and the magnesium abundances from , result in similar oxygen trends, while we do not see the high \[Mg/Fe\]-values for the highest \[Fe/H\]-stars. Thereby we contradict and their claim that the oxygen and magnesium trends are very different in the bulge.
Furthermore, the question of a possible shift in position of the knee in the \[$\alpha$/Fe\]-plot in the bulge as compared to the local disk is not unambiguously answered.
This research has been partly supported by the Royal Physiographic Society in Lund, Stiftelsen Walter Gyllenbergs fond. N.R. acknowledges support from the Swedish Research Council, VR (project number 2014-5640). M.Z. acknowledges support by the Ministry of Economy, Development, and Tourism’s Millennium Science Initiative through grant IC120009, awarded to The Millennium Institute of Astrophysics (MAS), by Fondecyt Regular 1150345 and by the BASAL-CATA Center for Astrophysics and Associated Technologies PFB-06. This publication made use of the SIMBAD database, operated at CDS, Strasbourg, France, NASA’s Astrophysics Data System, and the VALD database, operated at Uppsala University, the Institute of Astronomy RAS in Moscow, and the University of Vienna.
[^1]: Based on observations collected at the European Southern Observatory, Chile (ESO programs 71.B-0617(A), 073.B-0074(A), and 085.B-0552(A))
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Our multi-wavelength analysis of GRB060124 shows the unusual behaviour of the decaying lightcurve as well as supporting the recently proposed phenomenon of long-lasting central engine activity. The prompt X-ray emission displays uncommonly well resolved flaring behaviour, with spectral evolution – indicative of central engine activity – which allows us to estimate the energy injection time for each flare. The otherwise smooth X-ray & optical afterglows demonstrate achromatic breaks at about 1 day which differ significantly from the usual jet break in the blastwave model of afterglows.'
author:
- 'P.A. Curran[^1], D.A. Kann, P. Ferrero, E. Rol, R.A.M.J. Wijers'
title: 'The prompt emission & peculiar break of GRB060124'
---
\[1999/12/01 v1.4c Il Nuovo Cimento\]
Introduction
============
On 2006 January 24 the bright gamma-ray burst, GRB060124, triggered instruments on multiple satellites, including Swift and Konus-Wind. Swift immediately slewed to the burst and started observing in X-rays, $\sim350$s before the main period of activity. The discovery of the bright optical counterpart only 1 hour after the burst allowed for well-sampled follow-up. The brightness of the burst permitted observations in X-rays for nearly a month despite a redshift of $z = 2.3$.
Prompt Emission
===============
We analysed the previously observed high-energy flaring from the Swift BAT & XRT instruments and Konus-Wind[@ref:kann],[@ref:romano]. This exhibits spectral evolution, supporting the theory of a central engine which is active for extended periods. Our joint spectral fit of the data (0.2keV – 1160keV) gives an average spectral index of $\beta= 0.5$ after a break at 1.2keV and in combination with negligible Hydrogen column density above that of the Galactic value. This is in contrast to previous interpretations of the spectrum[@ref:romano].
We also fit the XRT data to the curvature effect model[@ref:curve], and estimate the times at which the central engine ejects the energy required to power the two main flares. Our results are consistent with similar fits to other GRBs[@ref:liang], in that the ejection time of flares is during the rising period of the flare, though ours is a most striking example.
Peculiar Break
==============
![The X-ray and optical lightcurves of GRB060124.[]{data-label="lc"}](curranpa_fig1-col.eps){width="90mm"}
The fading optical & X-ray lightcurves ([@ref:kann]; Fig. \[lc\]) display simultaneous achromatic breaks. Before the break the temporal decay indices, and spectral decay indices, are consistent with the blastwave model of GRB afterglows in a homogeneous circumburst medium, with a power law index of the electron energy distribution, $p \approx 2$. The achromatic nature of the break points to a change in the dynamics of the blastwave.
After the break there is a deviation from the behaviour expected from a laterally spreading jet. As observed in the majority of bursts, the temporal decay equals $p$, in both optical & X-rays. Our analyses show shallower and unequal decay in both wavelength regimes. A possible explanation for this peculiar behaviour is that the jet is spreading at lower velocities than usually observed.
[0]{} in preparation (2006)
[^1]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The paper is devoted to the prospects of using the laser radiation interaction with plasmas in the laboratory relativistic astrophysics context. We discuss the dimensionless parameters characterizing the processes in the laser and astrophysical plasmas and emphisize a similarity between the laser and astrophisical plasmas in the ultrarelativistic energy limit. In particular, we address basic mechanisms of the charged particle acceleration, the collisionless shock wave and magnetic reconnection and vortex dynamics properties relevant to the problem of ultrarelativistic particle acceleration.'
author:
- |
S. V. Bulanov$^{1,2}$, T. Zh. Esirkepov$^1$, D. Habs$^{3,4}$, F. Pegoraro$^5$, T. Tajima$^{1,3,4}$\
$^1$\
$^2$\
$^3$\
$^4$[\
$^5$[Physics Dept. and CNISM, University of Pisa, Largo Pontecorvo, 3, 56127 Pisa, Italy]{} ]{}
title: 'Relativistic Laser-Matter Interaction and Relativistic Laboratory Astrophysics'
---
Introduction {#intro}
============
High-power laser facilities have made unprecedented progress in recent years and the nearest future their radiation may reach intensities of 10$^{24}$W/cm$^{2}$ and higher [@ELIILE]. As a result of laser technology progress the laser-matter interaction entered regimes of interest for astrophysics. Typically in the course of laser irradiation of targets shock waves are generated; the target compression is accompanied by the Rayleigh-Taylor (RT) and Richtmayer-Meshkov (RM) instability development; collimated plasma jets are observed; the matter equation of state (EOS) acquires new properties under extreme pressure, density and temperature conditions; the laser plasma emits high energy charged particle beams and high- and low-frequency powerful electromagnetic radiation. Gathering of these facts principal for both space and laboratory physics has initiated works in the so-called laboratory astrophysics [@LabAstr] with the aim to model the processes of key importance for the space objects under laboratory conditions. Concerning the laser facilities, the present day laser systems can be subdivided into two categories. The first category includes lasers with a relatively long pulse of pico- and nanosecond duration and generally low repetition rate. These high energy and power laser facilities have been mainly developed for purposes of inertial confinement fusion with the laser pulse and target parameters corresponding to the collisional hydrodynamics phenomena [@Lindl]. In context of laboratory astrophysics they are used for experiments on shock waves, including the radiative shocks and RT&RM instability, the jet formation, and the EOS studies. The second category includes table top size lasers, whose pulse duration is of the order of a few tens of a femto-second with high repetition rate [@S-Mou]. Due to ultra short pulse duration and high contrast, these relatively moderate energy lasers can produce extremely high power and relativistically high intensity electromagnetic pulses. However, the role of both kinds of laser systems is complementary for the development of experimental facilities for the purposes of relativistic laboratory astrophysics.
Generic questions for astrophysics such as whether we are living in the Universe or in the Multiverse [@Weinberg], related discussions of the inflation era in the Multiverse evolution [@Linde] and probing our world’s dimensions are related to quantum gravitation physics and deal with the observational cosmology, in particular with an analysis of the cosmic black body radiation, the nuclear synthesis and Type I supernovae radiation (see [@Khlopov]), are yet out of the energy range accessible with present day lasers. The quantum gravitation energy scale is given by the Planck energy, $\sqrt{\hbar c^{3}/G}\approx 10^{19}$GeV, which corresponds to the mass $\sqrt{\hbar c/G}\approx 10^{-5}$g and the length $\sqrt{\hbar
G/c^{3}}\approx 10^{-33}\ $cm. In quantum field theory the unification energy scale corresponds to $10^{16}$GeV [@Perkins; @Wilczek]. These energy frontiers are yet well above of nowadays laser pulse energies. Fortunately, new physics such as the Higgs boson detection and exploration of the physics beyond the Standard Model is anticipated to be met at a substantially lower energy level in the range of several TeV in the experiments planned with the Large Hadron Collider (LHC), as summarized in Ref. [@Wilczek]. If relativistic laser plasmas can provide the charged particle acceleration up to the TeV energy level, laser accelerators will make a considerable impact to high energy physics, to finding answers on black hole and brane production under the terrestrial conditions [@MBH], to test causality [@TM] and to study the quark-gluon plasmas [@JR].
We may see that the main field of studies of astrophysical phenomena with high power lasers lies in the electrodynamics of continuous media in the relativistic regime [@MTB-06]. Since matter irradiated by ultrastrong electromagnetic waves (EMW) is ionized during a time interval comparable with the wave period and becomes a plasma and under astrophysical conditions approximately 95% of barionic part of matter is in the plasma state, the object of our studies is the relativistic laser and astrophysical plasma.
If we address to the problems of contemporary relativistic astrophysics, first of all questions on the mechanisms of the cosmic ray acceleration and on the properties of strong EMW interaction with relativistic plasmas attract our attention [@ACR]. In space plasmas basic mechanisms of charged particle acceleration are connected with the reconnection of magnetic field lines, which is accompanied by the strong and regular electric field generation (it occurs in the planet magnetosheres, in binary stellar systems, in accretion disks, in the magnetar magnetospheres, etc.) and with collisionless shock waves, at the fronts of which the charged particle acceleration occurs (this happens in interplanetary space, during supernova explosions, in colliding galaxies, etc.) [@ACR; @AH].
The laser accelerator development relies upon the fact that under the terrestrial laboratory conditions presently one of the most powerful sources of coherent electromagnetic radiation is provided by lasers [@MTB-06]. Wakefield accelerators, [@T-D] and [@Ch-D], presently provide the most advanced schemes for electron acceleration and they may be suggested to be good candidates for the charged particle acceleration in space [ACR,CTT]{}. One of the efficient mechanisms of ion acceleration in laser plasmas utilizes the radiation pressure of electromagnetic waves interacting with plasmas (see Refs. [@RPDA] and [@BEKT] ). Radiation pressure is a very effective mechanism of momentum transfer to charged particles. This mechanism was introduced long ago [@leb] and physical conditions of interest range from stellar structures and radiation generated winds (see e.g. Refs. [@miln]), to the formation of photon bubbles in very hot stars and accretion disks [Arons]{}, to particle acceleration in the laboratory [@RPDA; @VEK; @TER], see in addition Refs. [@lif; @ma], and in high energy astrophysical environments [@astr1].
Utilization of the plasma nonlinear properties for the electromagnetic wave intensification can result in much higher intensity and power. In this case a fundamental role is played by relativistic mirrors, which are thin electron sheets induced by the laser radiation moving with a speed close to the speed of light in vacuum, as proposed in Ref. [@BET-03]. We note the fruitfulness of the relativistic mirror concept for solving a wide range of problems in modern theoretical physics. Relativistic mirrors are important elements in the theory of the dynamical Casimir effect [@Casimir], with regard to the Unruh radiation [@Unruh] and other nonlinear vacuum phenomena [@Rozanov; @four-wave; @Narozhny-Fedotov; @MSh-06; @SHHK-06]. Relativistic mirrors made by wake waves may lead to an electromagnetic wave intensification resulting in an increase of pulse power up to the level when the electric field of the wave reaches the Schwinger limit [@SCHWIN] when electron-positron pairs are created from the vacuum and the vacuum refractive index becomes nonlinearly dependent on the electromagnetic field strength. In quantum field theory particle creation from the vacuum attracts a great attention, because it provides a typical example of non perturbative processes [@QED]. Nonlinear QED vacuum properties can in future be probed with such strong and powerful electromagnetic pulses.
If we trace a relationship between astrophysics and laser physics, we can see a number of publications devoted to the laboratory modeling of astrophysical processes [@LabAstr]. As known there has been an interest in modeling space physics with laboratory experiments for many years. The first modeling of processes fundamental for space physics in terrestrial laboratories has been done by Kristian Birkeland, who more than 100 years ago conducted first experiments on studying the auroral regions in the earth magnetosphere [@BIRK]. Lateron progress has been achieved in the laboratory modeling of various processes [@SPEXP; @FORT], including the magnetic field reconnection [@REC], collisionless shock waves [@SHW], which provide mechanisms for charged particle acceleration under various astrophysical conditions (see Ref. [@ACR]).
In the present paper we address plasma processes relevant to space physics, which occur in the relativistic and collisionless regimes.
Dimensionless Parameters that Characterize the Interaction Regimes of High Intensity Electromagnetic Waves with Matter
======================================================================================================================
Principle of Qualitative Scaling
--------------------------------
Laboratory experiments for studying astrophysical phenomena are of two types [@FH]. The first type of experiments can be referred to as *configuration modeling*, which is aiming at simulating the actual configuration of a system, e.g. the whole Earth’s magnetosphere (for example see Ref. [@ZABM], where the results of the laser-plasma experiments on the simulation of the global impact of the coronal mass ejections onto the Earth’s magnetosphere are presented). The second type of experiments corresponds to *process simulation*, i. e. they are aiming at studying the properties of physical processes relevant to astrophysical phenomena [@LabAstr]. There are a number of nonlinear plasma physical processes that require their clarification.
Physical systems obey scaling laws, which can also be presented as similarity rules. In the theory of similarity and modeling the key role is played by dimensionless parameters that characterize the phenomena under consideration [@Sedov]. The principle requirement of the laboratory modeling is the equality of the key dimensionless parameters in the modeled processes. In cases of modeling astrophysical phenomena where this equality can hardly be respected, instead the* principle of limited similarity* (PLS) or *principle of qualitative scaling* has been formulated in Refs. [@PodSag] and [@FH]. According to the PLS those dimensionless parameters, which are relevant in a certain context and which are much larger or smaller than unity under astrophysical conditions must retain this property (i.e. be much larger or smaller than unity) in the laboratory experiments modeling the astrophysical process. Below we present the key dimensionless parameters that characterize the high intensity electromagnetic wave (EMW) interaction with matter (see also Refs. [NEW,RELPEG]{}).
Parameters of Strong EMW Propagating in Plasmas
-----------------------------------------------
The intensity of an electromagnetic wave pulse is defined by its electric field amplitude through the expression: $I=cE_{0}^{2}/4\pi $, which is related to the Poynting vector $$P=\frac{c}{4\pi }\left[ {E\times B}\right] .$$ The power of the EMW is equal to the integral over its transverse cross-section $S$, $$P=\frac{{c}}{4\pi }\oint\limits_{S}{\left( {\left[ {E\times B}\right] \cdot n%
}\right) }dS=I\,S.$$ The time integral of the power gives the pulse energy, $\mathcal{E}=\mathcal{%
P}\tau _{p}$, where $\tau _{p}$, is the pulse duration. Other important parameters are the pulse frequency, $\omega _{0}$, which is related to its wavelength, $\lambda _{0}=2\pi c/\omega _{0}$, and the pulse polarization.
The first of the dimensionless parameters which characterizes the EMW packet is the ratio of the pulse length, $l_{p}=c\tau _{p}$, to the radiation wavelength, $\lambda _{0}$. We shall denote this ratio as $%
N_{p}=l_{p}/\lambda _{0}$. It is equal to the number of wavelengths per pulse, and is Lorentz invariant.
If the EMW intensity is relatively low, irradiated matter is not ionized. We notice that the typical energy of a photon in the laser parameter range with wavelengths in the micron range is of the order of one electron-volt and is substantially smaller than the binding energy of an electron inside an atom, $\hbar \omega _{0}\ll W_{b}$, i.e. it is smaller than the atomic ionization potential. In this case the characteristic dimensionless parameter of the interaction is the ratio between the amplitude of the electric field in the laser pulse, $E=\sqrt{4\pi I/c}$, and the atomic electric field, $E_{a}$. The latter is equal to the electric field of the proton at a distance of a Bohr radius, $a_{B}=\hbar ^{2}/m_{e}e^{2}\simeq 5.3\times 10^{-9}cm$, i.e. $%
E_{a}=e/a_{B}^{2}=m_{e}^{2}e^{5}/\hbar ^{4}$. The electron binding energy is $W_{b}=\hbar ^{2}/2a_{B}^{2}m_{e}$ and it corresponds to the frequency $%
\omega _{a}=W_{b}/\hbar $. The above condition, $\hbar \omega _{0}\ll W_{b}$, is equivalent to the inequality $\hbar \omega _{0}/W_{b}=\omega
_{0}/\omega _{a}\ll 1$. The dimensionless parameter $$\frac{E_{0}}{E_{a}}=\frac{\hbar ^{4}E_{0}}{m_{e}^{2}e^{5}}$$ becomes equal to unity for a laser radiation intensity equal to $%
m_{e}^{4}e^{10}c/4\pi \hbar ^{8}\simeq 10^{16}$ W/cm$^{2}$. For small but finite values of this parameter, i.e. in the limit $I<10^{16}$ W/cm$^{2}$, the atom is not ionized, unless the multiphoton processes come into play, the EMW–matter interaction can be described within the framework of perturbation theory. When the parameter $E_{0}/E_{a}$ approaches unity, the potential inside the atom changes its form and the so-called tunnel ionization becomes possible. The tunnel ionization probability is given by the Keldysh formula [@Keldysh] $$w=\omega _{a}\exp \left[ {-\frac{2W_{b}}{\hbar \omega _{0}}f(\gamma _{K})}%
\right] , \label{eq:tunion}$$ where the function $f(\gamma _{K})\approx 2\gamma _{K}/3$ for $\gamma
_{K}\ll 1$ and $f(\gamma _{K})\approx \ln 2\gamma _{K}-1/2$ for $\gamma
_{K}\gg 1$. The adiabatic parameter $\gamma _{K}$ is defined as $$\gamma _{K}=\omega _{0}\frac{\sqrt{2m_{e}W_{b}}}{eE_{0}}=\sqrt{\frac{2\hbar
\omega _{a}}{a_{0}^{2}m_{e}c^{2}}}.$$ Here introduced is the EMW dimensionless amplitude,
$$a_{0}=\frac{eE_{0}}{m_{e}\omega _{0}c}. \label{eq:dimla}$$
In the limit $\gamma _{K}\ll 1$, i.e. for a relatively strong electromagnetic wave, Eq. (\[eq:tunion\]) corresponds to the ionization probability by a constant electric field, $$w=2\omega _{a}\frac{E_{a}}{E_{0}}\exp \left( {-}\frac{{2E_{a}}}{{3E_{0}}}%
\right) .$$ For intensities larger than $m_{e}^{4}e^{10}c/4\pi \hbar ^{8}\simeq 10^{16}$ W/cm$^{2}$ the deformation of the potential inside the atoms caused by the laser pulse field becomes so strong that the electron energy level becomes larger than the maximum value of the potential. As a result, the electron appears as if in a free state and leaves the atom. Due to the periodicity of the electric field, there is a probability that the electron will return after a half of the wave period. Recollisions with the ions lead to the generation of high order harmonics [@Corkum]. However, for a very strong electromagnetic wave the effects of the wave magnetic field decrease this probability. In this case the matter becomes ionized in one optical period and plasma processes start to play a key role.
Under the action of the electromagnetic wave the plasma electrons oscillate at the wave frequency. In the limit $v<<c$ their quiver velocity is approximately equal to $v_{E}=eE_{0}/m_{e}\omega _{0}$. In the non-relativistic limit, when $v_{E}/c\ll 1$ or $a_{0}\ll 1$, the electron quiver amplitude is smaller than the laser wavelength, $\lambda _{0}$. Under the action of the electromagnetic wave, given by the vector potential $%
A_{\bot }(x-ct)$, the electrons oscillate at the wave frequency. From the equations of the motion we obtain that the transverse component of the generalized momentum $p_{\bot }-eA_{\bot }(x-ct)/c$ is constant. The particle energy and the longitudinal momentum component are related as [LLTF]{} $$\sqrt{m_{e}^{2}c^{4}+p_{\bot }^{2}+p_{\parallel }^{2}}-p_{||}c=h.$$ In the reference frame where the particle was at rest before interaction with the laser pulse, the particle kinetic energy $$K=m_{e}c^{2}\left( {\sqrt{1+(p/m_{e}c)^{2}}-1}\right)$$ and momentum $\mathbf{p=(}p_{||},p_{\bot }\mathbf{)}$ are given by expressions $K=m_{e}c^{2}\left| {a_{\bot }}\right| ^{2}/2$, $p_{\bot
}=m_{e}c\,a_{\bot }$, $p_{||}=m_{e}c\left| {a_{\bot }}\right| ^{2}/2$. Here $%
a_{\bot }=eA_{\bot }(x-ct)/m_{e}c^{2}$. For $|a_{\bot }|>1$ the particle acquires a relativistic energy, and the longitudinal component of its momentum is larger than the transverse component. Fig. \[fig:01\] shows a typical trajectory of the charged particle in the electromagnetic wave.
![Projectories of the charged particle trajectory, when it interacts with the elliptically polarized EMW.[]{data-label="fig:01"}](fig-01.jpg){width="9cm"}
The EMW behavior in a plasma differs from its behavior in vacuum, and depends on the electron density. In a plasma with a density $n$, a displacement of the electrons with respect to the ions generates the electric field. Its ratio to the laser electric field is $E/E_{0}=4\pi
ne^{2}/m_{e}\omega _{0}^{2}=(\omega _{pe}/\omega _{0})^{2}=n/n{{{{_{cr}}}}}$, where $\omega _{pe}=\sqrt{4\pi ne^{2}/m_{e}}$ is the Langmuir frequency and $n_{cr}={m_{e}\omega _{0}^{2}{{/{4\pi e^{2}}}}}$ is the critical density. The dimensionless parameter $$\frac{\omega _{pe}}{\omega _{0}}=\sqrt{\frac{n}{n_{cr}}}=\sqrt{\frac{4\pi
ne^{2}}{m_{e}\omega _{0}^{2}}}$$ is a measure of the plasma collective response to a periodic electromagnetic field.
When an EMW propagates through a plasma, its group velocity, $v_{g}=\partial
\omega /\partial k$, and phase velocity, $v_{ph}=\omega /k$, are not equal to each other and are related as $v_{g}v_{ph}=c^{2}$. While in vacuum the dispersion equation for the frequency, $\omega $, and wave vector,$k$, takes the form $\omega ^{2}=k^{2}c^{2}$, in a plasma it becomes $\omega
^{2}=k^{2}c^{2}+\omega _{pe}^{2}$. This dispersion equation can be rewritten as $k=\sqrt{\omega ^{2}-\omega _{pe}^{2}}/c$, which shows that an EMW with a frequency below the Langmuir frequency cannot propagate through the plasma and that the electromagnetic field evanescence length in a high density plasma is of the order of the collisionless skin depth, $d_{e}=c/\omega
_{pe} $, i.e. an overdense plasma with the electron density higher than the critical density is not transparent.
The collective response of the plasma, in addition to the transverse electromagnetic mode, exhibits longitudinal plasma oscillations, i.e., Langmuir waves. The electric field in a Langmuir wave oscillates with frequency $\omega =\omega _{pe}$. The group velocity of the Langmuir waves vanishes, $v_{g}=\partial \omega _{pe}/\partial k=0$, and their phase velocity is determined by the wave number.
Relativistic effects change the dispersion equation due to the dependence of the Langmuir frequency on the wave amplitude. As found in Ref. [@A-P], the frequency of a longitudinal wave depends on its amplitude $%
a_{L}=eE/m_{e}\omega _{pe}c$ as $\omega \approx \omega _{pe}(1-3a_{L}^{2}/4)$ for $a_{L}\ll 1$ and as $\omega \approx \omega _{pe}/\sqrt{8a_{L}}$ in the case $a_{L}\gg 1$.
For a circularly polarized electromagnetic wave the dispersion equation takes the form: $$\omega ^{2}=k^{2}c^{2}+\frac{\omega _{pe}^{2}}{\sqrt{1+a_{0}^{2}}}.$$ We see that the effective critical density increases as the EMW amplitude grows, i.e., the plasma is more transparent to high intensity electromagnetic radiation.
Large amplitude, finite length pulses of electromagnetic and Langmuir waves do not propagate independently since they are coupled by nonlinear processes. The Langmuir wave that is generated by an ultra short laser pulse, being left behind in the plasma and thus called the wake wave. It is of special interest since the structure of the electric field of this wake wave is favorable for charged particle acceleration. In a low density plasma the phase velocity of the wake wave can be very close to the speed of light in vacuum. In analogy to linear accelerators that use electric fields in the radio-frequency range in Ref. [@T-D] it was proposed to use the wake field for charged particle acceleration.
The dimensionless amplitude, Eq. (\[eq:dimla\]), is equal to the electron quiver momentum normalized to $m_{e}c$. For a pulse with an intensity corresponding to $a_{0}>1$, relativistic effects must be taken into account. The intensity of a linearly polarized electromagnetic wave can be written via $a_{0}$ as $$I_{L}=\frac{\pi }{2}\frac{a_{0}^{2}}{\lambda _{0}^{2}}\frac{{m_{e}c^{3}}}{%
r_{e}}\approx 1.37\times 10^{18}\times a_{0}^{2}\times \left( {\frac{1\mu m}{%
\lambda _{0}}}\right) ^{2}\frac{W}{cm^{2}}.$$
If the wave is focused into a one wavelength spot, this intensity corresponds to the power $\mathcal{P}=a_{0}^{2}\times 43\;$GW. At present laser intensities have reached a level above $10^{22}$W/cm$^{2}$ [@YAN].
When the electron energy approaches $3m_{e}c^{2}$, electron-positron pairs are generated during electron-nuclei collisions, [@QED]. The cross section of this process is given by $$\sigma _{\pm }=\frac{28}{27\pi }r_{e}^{2}(\alpha Z)^{2}\left[ {\ln \left(
\sqrt{1+\left( \frac{p}{m_{e}c}\right) ^{2}}\right) }\right] .$$
Here $Ze$ is the nucleus electric charge and $\alpha =e^{2}/\hbar c=1/137$ is the fine-structure constant. Positron generation in a plasma has been discussed in a number of publications (e.g. see [@BKZS]) and was observed in the terawatt laser plasma interaction experiments [@POS]. We note a discussion of the pion and muon production in electron-positron and gamma plasmas [@IKUZ].
Interaction of EMW with Plasmas in the Radiation-Dominated Regime
-----------------------------------------------------------------
The dimensionless parameters characterizing the electromagnetic emission by an electron are the ratio between the classical electron radius and the electromagnetic wavelength, $r_{e}/\lambda _{0}=e^{2}\omega _{0}/2\pi
m_{e}c^{3}$, and the ratio between the photon energy and the electron rest mass energy, $\hbar \omega _{0}/m_{e}c^{2}$.
When an electron moves under the action of the electric and magnetic field of a wave, it emits electromagnetic radiation. The intensity of this radiation is given by the formula $W=(2e^{2}/3m_{e}^{2}c^{3})(dp_{\mu
}/d\tau )^{2}$, where $p_{\mu }$ is the particle 4-momentum and $\tau $ is its proper time. When an ultrarelativistic charged particle moves along a circular trajectory in a circularly polarized electromagnetic wave, the radiation intensity is $W=\left( {4\pi r_{e}/3\lambda _{0}}\right) \omega
_{0}\,m_{e}c^{2}a_{0}^{4}$. We see that the relative role of the radiation damping force is determined by the dimensionless parameter$\varepsilon
_{rad} $, which is equal to $$\varepsilon _{rad}=\frac{4\pi r_{e}}{3\lambda _{0}}.$$By comparing the energy radiated by the particle per unit of time with the maximum energy gain in the electromagnetic wave $\partial _{t}\mathcal{E}%
=\omega _{0}m_{e}c^{2}a_{0}$, we obtain that the radiation effects become dominant at $a_{0}\geq a_{rad}=\varepsilon _{rad}^{-1/3}$ , i. e. in the limit $I>10^{23}W/cm^{2}$ for 1 $\mu $m wavelength laser [@BEKT; @RAD]. In the limit of a relatively low amplitude laser pulse, $a_{0}\ll a_{rad}$, the momentum of an electron moving in a circularly polarized electromagnetic wave in a plasma scales with the laser pulse amplitude as $p=m_{e}ca_{0}$, while in the limit $a_{0}\gg a_{rad}$, it scales as $p=m_{e}c(a_{0}/%
\varepsilon _{rad})^{1/4}$.
Quantum effects become important, when the energy of the photon generated by Compton scattering is of the order of the electron energy, i.e. $\hbar
\omega _{m}\approx \mathcal{E}_{e}$. An electron with energy $\mathcal{E}%
_{e}=\gamma m_{e}c^{2}$ rotates with frequency $\omega _{0}$ in a circularly polarized wave propagating in a plasma and emits photons with frequency $%
\omega _{m}=\gamma ^{3}\omega _{0}$. We obtain that quantum effects come into play when $\gamma \geq \gamma _{Q}=\sqrt{m_{e}c^{2}/\hbar \omega _{0}}$. For an electron interacting with one-micron laser light we find $\gamma
_{Q}\approx 600$. From the previous analysis we obtain that the quantum limit on the electron gamma factor corresponds to $$a_{Q}=\frac{2e^{2}m_{e}c}{3\hbar ^{2}\omega _{0}}.$$ The energy flux reemitted by the electron is equal to $e(E\cdot
v)=\varepsilon _{rad}\omega _{0}\gamma ^{2}p_{\bot }^{2}/m_{e}$. The total scattering cross section defined as the ratio of the reemitted energy to the Poynting vector $P=cE_{0}^{2}/4\pi $, is given by $$\sigma =\sigma _{T}\frac{\gamma ^{2}}{1+\varepsilon _{rad}^{2}\gamma ^{6}},$$ where the Thomson scattering cross section is $\sigma _{T}=8\pi
r_{e}^{2}/3=6.65\times 10^{-25}cm^{2}$. We see that, as the wave amplitude increases in the range $1\ll a_{0}\ll a_{rad}$, the scattering cross section increases according to the law $\sigma =\sigma _{T}(1+a_{0}^{2})$ and reaches its maximum $\sigma =\sigma _{T}a_{rad}^{2}$ at $a_{0}\approx
a_{rad} $; for $a_{0}\gg a_{rad}$, it decreases according to the law $\sigma
=\sigma _{T}a_{rad}^{3}/a_{0}$. In Fig. \[fig:02\] we show the scattering cross section dependence on the EMW amplitude and wavelength.
![Scattering cross section dependence on the EMW amplitude and wavelength.[]{data-label="fig:02"}](fig-02.jpg){width="9cm"}
In the radiation-dominated regime of the EMW interaction with charged particles, i.e. at $a_{0}>a_{rad}$, the emitted gamma quanta can produce secondary electron-positron pairs, which in turn emit gamma ray photons, producing an avalanche of $\gamma $ rays and electron-positron pairs [Bell]{}.
Probing Nonlinear Vacuum
------------------------
When the amplitude of the electromagnetic wave approaches the critical electric field of quantum electrodynamics (also called the “Schwinger field”), vacuum becomes polarized and electron-positron pairs are created in vacuum [@QED; @DGi]. On a distance equal to the Compton length, $%
\lambda _{C}=\hbar /m_{e}c$, the work of the critical field on an electron is equal to the electron rest mass energy, $m_{e}c^{2}$, i.e. $%
eE_{QED}\lambda _{C}=m_{e}c^{2}$. The dimensionless parameter
$$\frac{E}{E_{QED}}=\frac{e\hbar E}{m_{e}^{2}c^{3}}$$
becomes equal to unity for an electromagnetic wave intensity of the order of
$$I=\frac{c}{r_{e}\lambda _{C}^{2}}\frac{m_{e}c^{2}}{4\pi }\approx 4.7\times
10^{29}\frac{W}{cm^{2}}.$$
For such ultrahigh intensities the effects of nonlinear quantum electrodynamics play a key role: an electromagnetic wave excites virtual electron-positron pairs. An observable manifestation of this process could be detection of light birefringence during the propagation of an electromagnetic wave in a strong electric or magnetic field in vacuum. The cross section for the photon-photon interaction in the limit $\hbar \omega
\ll m_{e}c^{2}$ is given by $$\sigma _{\gamma \gamma \to \gamma \gamma }=\frac{973}{10125}\frac{\alpha ^{2}%
}{\pi ^{2}}r_{e}^{2}\left( {\frac{\hbar \omega }{m_{e}c^{2}}}\right) ^{6},$$ where $\hbar \omega $ is the photon energy (see [@QED]). This cross section reaches its maximum, $\sigma _{\max }\approx 10^{-20}cm^{2}$, for $%
\hbar \omega \approx m_{e}c^{2}$, i.e. for the interactions of photons in the gamma range. Also attention is focused on the process of electron-positron pair creation in vacuum by an electromagnetic wave. For an electric field small compared to $E_{QED}$, this process is sub-barrier, similarly to the tunnel ionization of atom discussed above \[see Eq. ([eq:tunion]{})\]. The probability of electron-positron pair creation per unit volume and per unit time is exponentially small and is given by $$w=\left( \frac{\alpha c}{\pi ^{2}\lambda _{C}^{4}}\right) \left( \frac{E}{%
E_{QED}}\right) ^{2}exp\left( -\pi \frac{E_{QED}}{E}\right) . \label{eq:w}$$ Here $\lambda _{C}=\hbar /m_{e}c$ is the Compton length and $\alpha
=e^{2}/\hbar c=1/137$ is the fine structure constant.
We may formally estimate the number of electron-positron pairs produced by a 10 fs long laser pulse in a volume $V=\lambda ^{3}=10^{-12}$cm$^{3}$ as $%
N_{\pm }=wV\tau _{p}$. It is easy to show that $N_{\pm }$ is equal to one pair for a laser intensity equal to $I=10^{26}$W/cm$^{2}$ (a more detailed description of this process can be found in [@NAR] and in Refs. [QED,VSP,NBMP]{}). Obviously, this latter number is overestimated because the minimum needed energy is by many orders of magnitude larger than the total energy of the laser pulse. At intensities of the order $I=10^{30}$W/cm$^{2}$ Eq. (\[eq:w\]) is not applicable and a depletion of the laser pulse must be taken into account. The electromagnetic pulse depletion due to its energy conversion into electron-positron pairs has been studied in Ref. [@BFP].
![Various regimes of relativistically strong EMW interaction with plasmas.[]{data-label="fig:03"}](fig-03.jpg){width="9cm"}
The nonlinear dependence of the vacuum susceptibilities on the electromagnetic-field amplitude results in the finite value of the Kerr constant of vacuum. It can be found to be $$K_{K}=\frac{7\alpha }{90\pi }\frac{\lambda _{C}^{3}}{m_{e}c\lambda _{0}}$$ The Kerr constant in vacuum for $\lambda _{0}=1\mu m$ is of the order of $%
10^{27}cm^{2}/erg$, which is a factor $10^{20}$ smaller than for water. As shown in Ref. [@Rozanov], in a QED nonlinear vacuum two counterpropagating electromagnetic waves mutually focus each other. A nonlinear modification of the refraction index in vacuum within the framework of the Heisenberg-Euler approximation is characterized by the critical value of the electromagnetic wave power $$\mathcal{P}_{QED}=45\pi ^{2}\frac{cE_{QED}^{2}\lambda _{0}^{2}}{4\pi \alpha }.$$ When the electromagnetic wave power exceeds this value, the cross modulation nonlinear effects affect the wave propagation. We see that the critical power, $\mathcal{P}_{cr}$, depends only on the laser pulse wavelength, $%
\lambda _{0}$, and on fundamental constants. It is easy to show that for $%
\lambda _{0}=1\mu $m the critical power $\mathcal{P}_{cr}=cE^{2}w^{2}/4$, where $w$ is the laser beam waist. For the mutual self-focusing $\mathcal{P}%
_{cr}=2.5\times 10^{24}$W can be found to be for $\lambda _{0}=1\mu m$.
Nonlinear modifications of the vacuum refraction index lead to the vacuum birefringence [@Rozanov], to the four-wave interaction [@four-wave], to the high order harmonic generation [@Narozhny-Fedotov], and to the laser-photon splitting and merging [@DIP] (see also review articles [@MTB-06], [@MSh-06; @SHHK-06]). According to Ref. [@Unruh] the Unruh radiation intensity of the electron moving in the field of a strong electromagnetic wave becomes comparable with the nonlinear Thomson scattering intensity under the condition $4\pi a_{0}\hbar
\omega _{0}/m_{e}c^{2}\approx 1$. The multi-photon Compton scattering during the collision of counter-propagating laser beams and ultrarelativistic electron bunches leading to the gamma quanta generation $$e^{-}+n\hbar \omega _{0}\rightarrow \hbar \omega _{\gamma },$$ with their subsequent interaction with the laser light accompanied by the electron -positron pair creation in vacuum via the Breit-Wheeler process $$\hbar \omega _{\gamma }+n\hbar \omega _{0}\rightarrow e^{-}+e^{+}$$ has been investigated in Ref. [@Burke]. In Ref. [@ErL] the cross section of the Breit-Wheeler process on the laser pulse intensity has been investigated.
Various regimes of the relativistically strong EMW interaction with plasmas are illustrated in Fig. \[fig:03\].
EMW Parameters under Space Plasma Conditions
--------------------------------------------
![Pulsar magnetosphere. The inset: The Crab pulsar [Hester]{}.[]{data-label="fig:04"}](fig-04.jpg){width="9cm"}
In one of the first works on the charged particle acceleration by strong electromagnetic waves in astrophysical plasmas, pulsars [@BG] have been considered as sources of ultraintense radiation [@GO]. Pulsars are considered to be oblique rotators with non-parallel rotation and magnetic dipole axes, as illustrated in Fig. \[fig:03\]. The power of magneto-dipole radiation is given by the expression $\ $$$W=\frac{2\mu ^{2}\sin ^{2}\chi \Omega _{P}^{4}}{3c^{3}},$$ where $\mu $ is the magnetic momentum, $\chi $ is an angle between the rotational and magnetic dipole axes, and $\Omega _{P}$ is the pulsar rotation frequency. Even for parallel magnetic and angular moments, i.e. for $\chi =0$, the expression $W=(2/3)\mu ^{2}\Omega _{P}^{4}/c^{3}$ gives the pulsar electromagnetic energy losses, as it follows from the theoretical model of the pulsar magnetosphere [@GR]. The magnetic moment is related to the pulsar magnetic field and radius as $\mu \approx Br_{P}^{3}$. For typical values of $r_{P}=10^{6}$cm and $B=10^{12}\ $G we obtain $\mu
=10^{30}\ $G cm$^{3}$. The electromagnetic wave intensity at the distance $%
r $ is equal to $I=W/4\pi r^{2}$. At the wave zone boundary, $r=c/\Omega
_{P} $, the dimensionless amplitude of the electric field is $$a_{P}=\frac{e\mu \Omega _{P}^{2}}{m_{e}c^{4}}.$$ For the Crab pulsar with the rotation frequency $\Omega _{P}=200\ $s$^{-1}$ we find $a_{P}=2\times 10^{10}$.
According to Ref. [@BEKT; @RAD] in the limit of high radiation intensity the effects of the radiation damping should be incorporated into the theory of the electromagnetic wave interaction with plasmas. A dimensionless parameter, $$\varepsilon _{rad}=\frac{2e^{2}\Omega _{P}}{3m_{e}c^{3}},$$ gives a value of the wave amplitude, $a_{rad}=\varepsilon _{rad}^{-1/3}$, above which the radiation damping cannot be neglected. For $\Omega _{P}=200\
$s$^{-1}$ this yields $a_{rad}=10^{7}$, which is substantially less than above found value of $a_{P}=2\times 10^{10}$.
In the case of laser - plasma interaction for a typical laser wavelength of $%
1\mu $m the dimensionless amplitude $a_{rad}$ corresponds to an intensity of the order of $10^{23}$W/cm$^{2}$ which can be achieved by tight focusing of the PW power laser beams onto the one-lambda size focus spot. We see that the laser plasmas can be used for modeling the radiation damping effects, which are important for relativistic astrophysics.
Acceleration of Charged Particles in the EMW Interaction with Plasmas {#LACP}
=====================================================================
General requirements for the laser accelerator parameters are principally the same as for standard accelerators of charged particles [ACC-EM]{}, i. e. they should have a reasonable acceleration scale length, a high enough efficiency and the required maximal energy, a high quality, emittance and luminosity of charged particle beams. In the 1940-s Enrico Fermi paid attention to the high energy limit of $\approx 1\ $PeV$=10^{15}\ $eV for accelerated particles, which could be reached under terrestrial conditions, when the accelerator size is limited by the equator circumference. These limitations resulted in the 1950-ties in the proposal to use collective electric fields excited in a plasma (collective methods of acceleration) in order to accelerate charged particles [@VEK].
Electron Accelerator {#EA}
--------------------
Wakefield acceleration has been proposed in Ref. [@CTT] for the generation of ultra high energy cosmic rays. Below we describe the wake field acceleration mechanism using as an example the LWFA scheme.
![Schematic view of the standard linear accelerator of charged particles, a) and the LWFA, b).[]{data-label="fig:05"}](fig-05.jpg){width="9cm"}
Under the condition of minimum laser energy the one stage LWFA accelerator scaling is described as it follows [@T-D]. The electric field in a plasma has the form of a wave propagating with a phase velocity, $v_{ph,W}$. A gamma factor corresponding to the wave phase velocity is given by the expression $\gamma _{ph,W}=(1-v_{ph,W}^{2}/c^{2})^{-1/2}$. A condition of the wake wave synchronization with the driver laser pulse yields $%
v_{ph,W}=v_{g,las}$, where $v_{g,las}\approx c(1-\omega _{pe}^{2}/2\omega
_{0}^{2})$ is the laser pulse group velocity. The wavelength of the weakly nonlinear wake wave is $\lambda _{p}=\lambda _{0}\gamma _{ph,W}$. Assuming the electrostatic potential in the wake is equal to $m_{e}c^{2}/e$, we obtain for the fast electron gamma factor $\gamma _{e}=2\gamma _{ph,W}^{2}$. The acceleration length is given by $l_{acc}=\lambda _{p}\gamma _{ph,W}^{2}$, i.e. $l_{acc}=\lambda _{0}\gamma _{ph,W}^{3}$. This gives a relationship between the acceleration length and the fast electron energy: $%
l_{acc}=\lambda _{0}\gamma _{e}^{3/2}$. For $\lambda _{0}=1\ \mu $m and $%
\gamma _{e}=10^{6}$ we obtain $l_{acc}\approx 1$km [@KanTeV].
![2D PIC simulations show that the electrons pushed away by the ponderomotive pressure of the laser pulse form the “bow wave” [NEW]{}. The electron density distribution (a) clearly shows the ’swallow-tail’ formation during the wake wave breaking in the first period of the wave behind the laser pulse. The wakefield (the x-component of the electric field) is excited by the laser pulse in an underdense plasma (b). Inset: The bow wave formed by colliding galaxies in the Bullet Cluster [@DClowe].[]{data-label="fig:06"}](fig-06.jpg){width="9cm"}
In the opposite limit, when the laser transverse width $r_{las}\le \lambda
_{p}$, we need to take into account the formation of an electron density cavity moving with the group velocity of the laser pulse (see Fig.[fig:06]{}, where the wake wave left behind the ultra short laser pulse in the underdense plasma is shown). The cavity’s transverse size is determined by the laser pulse width and its length is of the order of the Langmuir wave wavelength. In this limit, the wavelength depends on the amplitude of the Langmuir wave, which in turn depends on the laser pulse intensity. For a given laser pulse width the electrostatic potential in the cavity is of the order of $\phi \approx \pi n_{0}er_{las}^{2}$, and the group velocity of a narrow laser pulse is determined by its width, i.e. $\gamma _{ph,W}\approx
r_{las}/\lambda _{0}$. As a result we find the electron energy scaling [NEW]{}: $$\gamma _{e}=\frac{r_{las}^{4}}{\lambda _{0}^{2}\lambda _{p}^{2}}.$$ It does not depend on the laser pulse amplitude provided $a_{0}>e\phi
/m_{e}c^{2}$. The laser energy depletion length in this limit is given by $$l_{dep}=a_{0}l_{las}\left( \frac{\lambda _{p}}{\lambda _{0}}\right) ^{2},$$ i. e. it is by a factor $a_{0}$ greater than in the 1D case.
Considering the laser electron accelerator for the applications in the high energy physics, we find that its parameters should satisfy several conditions in addition to the requirement on the maximum particle energy. Parameters of fundamental importance such as the luminosity characterize the number of reactions produced by the particles in colliding beams of a collider. The luminosity is given by the expression $$\mathcal{L}=f\frac{N_{1}N_{2}}{4\pi \sigma _{y}\sigma _{z}},$$ where $N_{1}$ and $N_{2}$ are the numbers of particles in each of the beams, $\sigma _{y}$ and $\sigma _{z}$ are the transverse size of the beam in the $%
y $ and $z$ directions, and $f$ is the frequency of the beam collisions. A product of the luminosity and the reaction cross section gives the reaction rate. We see that the luminosity can be increased by increasing the particle number in a bunch, $N_{j}$, and/or by increasing the repetition rate, $f$, or by decreasing the transverse size of the bunch, $\sigma _{i}$, by focusing the particle beam into the minimum size focal spot. The focal spot size depends on the beam emittance, which is defined as the surface occupied by the bunch in the phase plane ($(y,p_{y})$ or $(z,p_{z})$). A calculation under the assumption of a round transverse shape of the beam ($\sigma
_{y}=\sigma _{z}=r$) results in the expression given by the integral $$\varepsilon _{\bot }=\frac{1}{\pi }\int drdr^{\prime },$$ where $r$ is the transverse size of the bunch and $r^{\prime }=dr/dx=dr/cdt$ [@ACC-EM].
The transverse dynamics of the electron in the field of the wake wave is described by the equation (see for example Ref. [@B-PoP]) $$\frac{d}{dt}\left( \gamma _{||}\frac{dr}{dt}\right) +\omega _{pe}^{2}r=0,$$ where the electron gamma factor depends on time as $\gamma _{||}(t)=\gamma
_{e}(1-t^{2}/t_{acc}^{2})$ with $\gamma _{e}=\gamma _{ph,W}^{2}$ and $\gamma
_{ph,W}=\omega _{0}/\omega _{pe}$. In the limit $\gamma _{||}\gg 1$ the electron transverse oscillations are described by the dependence of the radial displacement on time: $$r(t)=r_{inj}\left( \frac{\gamma _{inj}}{\gamma _{||}(t)}\right) ^{1/4}\cos %
\left[ \int_{t_{inj}}^{t}\omega _{b}(t^{\prime })dt^{\prime }\right] ,$$ where $\omega _{b}(t)=\omega _{pe}\gamma _{||}(t)^{2}$ is the betatron oscillation frequency and $r_{inj}$ and $\gamma _{inj}\approx \gamma _{ph,W}$ are the radial coordinate and the electron energy at the injection time, $%
t_{inj}$ normalized on $m_{e}c^{2}$. Calculating the transverse emittance, we find $\varepsilon _{\perp }=\pi \kappa ^{2}(\omega _{pe}/\omega _{0})^{3}$ mm mrad, with $\kappa =r_{inj}/\lambda _{p}$. The normalized emittance, $%
\varepsilon _{N}=\varepsilon _{\perp }\gamma _{e}$, is equal to $\varepsilon
_{\bot }=\pi \kappa ^{2}(\omega _{pe}/\omega _{0})$ mm mrad.
The electron motion in the electric field of the wake plasma wave is characterized by the structure of the phase plane ($p_{x},X=x-v_{ph}t$). A calculation of the energy spectrum of fast electrons is done in Refs. [B-PoP,QME]{}). It uses the property of electrons injected at the breaking point to move along the separatrix. The electrons, whose trajectories lie on the separatrix, where they are uniformly distributed, near the top of the separatrix have an electron momentum dependence on the coordinate $$p_{x}=p_{m}(1-X^{2}\omega _{pe}^{2}/c^{2}a_{0})=p_{m}(1-t^{2}/t_{acc}^{2}).$$ The distribution function of the electrons at the target has the form $$f(t,\mathcal{E})=(n_{b}\omega _{pe}/\sqrt{2}ca_{0})\delta (\mathcal{E}-%
\mathcal{E}_{m}(1-t^{2}/t_{acc}^{2})).$$ Here $\delta (z)$ is the Dirac delta function and we have assumed that the electrons are ultrarelativistic with $\mathcal{E}=p_{x}c$ and $\mathcal{E}%
_{m}=p_{m}c$ . In order to find the energy spectrum of the electrons on the target, we must integrate the function on time in the limits between $%
-t_{acc}$ and $t_{acc}$. We obtain
$$\frac{d\mathcal{N}(\mathcal{E})}{d\mathcal{E}}=\frac{n_{b}\omega _{pe}}{%
\sqrt{2}ca_{0}}\int\limits_{-t_{acc}}^{t_{acc}}\delta \left( \mathcal{E}-%
\mathcal{E}_{m}\left( 1-\frac{t^{2}}{t_{acc}^{2}}\right) \right) dt=\frac{%
n_{b}\omega _{pe}}{2\sqrt{2}ca_{0}\sqrt{\mathcal{E}_{m}(\mathcal{E}_{m}-%
\mathcal{E})}},$$
i.e. the particle spectrum has a typical form $\propto 1/\sqrt{\mathcal{E}%
_{m}-\mathcal{E}}$ near maximum energy.
Using the above given relationships and estimating a maximum number of particles in a bunch as $N\approx \kappa ^{2}n_{e}\lambda _{p}^{3}$, we obtain the luminosity to be equal to $$\mathcal{L}=10^{34}\left( \frac{f}{10KHz}\right) \left( \frac{\kappa }{0.1}%
\frac{\lambda _{0}}{r_{inj}}\right) ^{2}\left( \frac{\gamma _{e}}{10^{6}}%
\right) ^{3/2}\ \frac{1}{\mathrm{cm}^{2}\mathrm{s}}.$$ Here we assume a round transverse shape of the bunch with $r\approx
r_{inj}(\gamma _{inj}/\gamma _{e})^{1/4}$. Utilization of flat bunches with $%
\sigma _{y}\gg \sigma _{z}$ allows and to achieve larger luminosity [kando-07]{}. In addition, in the case of flat beams the space charge effects and beamsstahlung can be weakened. We notice that the radiation damping effects on the LWFA operation have been considered in Ref. [@MSSEL].
Ion Accelerator {#IA}
---------------
The mechanism of laser acceleration of ions (protons and other ions) is determined by the electric field set up by the space charge separation of hot or energetic electrons and the ions. The exact mechanisms entering into the energy transfer from the fast electron to the ion energy depends on the specific conditions of the laser-target interaction (see review articles [@MTB-06] and [@BFB]). The proton generation is a direct consequence of the electron acceleration.
The typical energy spectrum of laser accelerated particles observed both in experiments and in computer simulations can be approximated by a quasi-thermal distribution with a cut-off at a maximum energy. On the other hand, the applications require high quality proton beams, i.e. beams with sufficiently small energy spread $\Delta \mathcal{E}_{i}/\mathcal{E}_{i}\ll
1_{i}$. For example, for hadron therapy it is highly desirable to have a proton beam with $\Delta \mathcal{E}_{i}/\mathcal{E}_{i}\leq 2\%$ in order to provide the conditions for a high irradiation dose being delivered to the tumor, while sparing neighboring tissues. In Ref. [@BKh] it was shown that such a required beam of laser accelerated ions can be obtained using a double layer target. Extensive computer simulations of this target were performed in Ref. [@Es-2002] and the results of experimental studies of this ion acceleration mechanism are presented in Ref. [@Nat].
### Ion acceleration during plasma expansion into vacuum
Ion acceleration during the collisionless plasma expansion into vacuum appears to be one of the most obvious mechanisms of the ion acceleration [@GP]. In particular, it has been considered as one of the possible acceleration mechanisms in space plasmas [@ACR]. When the electrons that have been heated and tend to expand overtake the ions in a relatively small volume, the electric neutrality of the plasma breaks and the generated electric field induces the ion motion. Although a velocity of the bulk ion and electron motion is of the order of the ion acoustic speed, $v_{s}=\sqrt{%
T_{e}/m_{p}}$, a small fraction at the plasma front gains energy efficiently. Here $T_{e}$ is the electron temperature and $m_{p}$ is the ion (proton) mass. Under the most favorable conditions the ions achieve a kinetic energy which corresponds to an ion velocity of the order of the electron thermal velocity, i.e. the maximum ion energy can be of the order of $m_{p}\mathcal{E}_{e}/m_{e}$. We notice here that for the electron distributions with the energy cut-off this conclusion requires careful analysis (see Refs. [@M-T; @BEKTF])
In the limit when the electron energy is relativistic, in order to analyze the ion motion one should use the equations of relativistic hydrodynamics, $%
\nu _{\alpha} T_{\mu} ^{\nu} = 0$ with $T_{\mu}^{\nu} $ being the energy-momentum tensor, $$\partial _{\mu }\left( {nu^{\mu }}\right) =0,$$ $$\mathcal{W}u^{\mu }\partial _{\mu }u^{\nu }=-\left( {\delta ^{\mu \nu
}-u^{\mu }u^{\nu }}\right) \partial _{\mu }P.$$ Here $u^{\mu} $ is the four-dimensional velocity vector, $P$ is the pressure, n is the density in the proper frame of reference, $\mathcal{W} =
P + \varepsilon $ is the enthalpy with $\varepsilon $ being the internal energy density.
The self-similar plasma motion depending on the variable $\chi =x/t$ is described by a system of ordinary differential equations $$\frac{u\chi -1}{(u-\chi )(1-u^{2})}u^{\prime }-(\ln n)^{\prime }=0,$$ $$\mathcal{W}\frac{u-\chi }{1-u^{2}}u^{\prime }-(u\chi -1)P^{\prime }=0,$$ with a prime denoting a differentiation with respect to $\chi $ and $u=v/c$. Here we use the relativistic equation of state of an ideal gas, $$\mathcal{W}=c^{2}\frac{K_{3}\left( {m_{e}c^{2}/T_{e}}\right) }{K_{2}\left( {%
m_{e}c^{2}/T_{e}}\right) },$$ $$P=nT_{e},$$ where $K_{n}(x)$ are the modified Bessel functions. In the case of $T_{e}=$constant we find $$u=\frac{c_{s}+c\eta }{c+c_{s}},$$ $$n=n_{0}\left( \frac{c_{s}-c\eta }{c+c_{s}}\right) ^{c/c_{s}}$$ with $c_{s}$ being the relativistic speed of sound, $$c_{s}=\sqrt{\frac{T_{e}}{m_{e}}\frac{K_{3}(m_{e}c^{2}/T_{e})}{%
K_{2}(m_{e}c^{2}/T_{e})}}.$$ In the ultra-relativistic limit the energy spectrum of fast ions has a power-law form, $$\frac{d\mathcal{N}_{p}(\mathcal{E}_{i})}{d\mathcal{E}_{i}}\propto \mathcal{E}%
_{i}^{-2c^{2}/c_{s}^{2}}.$$
### Radiation pressure dominated regime of the ion acceleration
A regime of ion acceleration that exhibits very favorable properties has been identified in Ref. [@RPDA]. Among the wide variety of ion acceleration mechanisms realized in the laser-plasma interaction, the radiation pressure dominated ion acceleration (RPDA) has the highest efficiency. In the RPDA ion accelerator the laser pulse radiation pressure pushes forward the irradiated region of a thin foil as a whole. In the relativistic limit, when the electrons and ions move together with the same velocity due to a smallness of the electron to ion mass ratio, the ion kinetic energy is by a factor $m_{i}/m_{e}$ times higher than the electron energy. In this case the laser pulse interacts with an accelerated foil like with a relativistic co-propagating mirror. The electromagnetic radiation reflected back by the relativistic mirror has almost negligible energy compared to the energy in the incident laser pulse, i.e. the laser energy is almost completely transformed into the energy of fast ions. In Fig. [fig:07]{} we show results of 3D PIC simulations of this ion acceleration regime. In the course of the interaction with a thin overdense plasma slab the multi-petawatt laser pulse forms a cocoon confining the EMW energy, thus increasing the coupling of the electromagnetic wave with the target (see frame a) and the 2D inset). The ions accelerated beyond the GeV energy level have a quasi-monoenergetic spectrum (Fig. \[fig:07\]b). We notice that a combination of the RPDA mechanism with the use of double layer targets can substantially increase the ion acceleration efficiency as demonstrated in Ref. [@lif].
![ Results of 3D PIC simulations of the PPDA ion acceleration regime. a) The electromagnetic pulse forms a cocoon confining the EMW energy. The right inset shows a cocoon seen in the plasma density and an EMW distribution obtained with the 2D PIC simulation. In the left inset we see a cocoon formed by the Black Widow pulsar (Pic. NASA). b) Quasi-monoenergetic ion spectrum.[]{data-label="fig:07"}](fig-07.jpg){width="9cm"}
The equations of the irradiated foil motion can be cast into the form [PB-07]{}: $$\frac{dp_{i}}{dt}=\mathcal{P}d\sigma _{i}\mathbf{\;}, \label{eq1}$$ where $p_{i}$ is a momentum of the foil element, $d\sigma _{i}$ is a vector normal to the foil, the index $i=1,2,3$, and $\mathcal{P}$ is the relativistically invariant pressure. In the frame of reference co-moving with the foil the radiation pressure is equal to $\mathcal{P}=E_{M}^{2}/2\pi
$, with $E_{M}$ being the EMW amplitude. In the laboratory frame of reference we have $E_{0}^{2}=E_{M}^{2}\left( \omega _{0}/\omega _{M}\right)
^{2}$, where $\omega _{0}$ and $\omega _{M}$ are the wave frequency in the laboratory and boosted frames. They are related to each other as $\omega
_{M}/\omega _{0}=\sqrt{\left( 1-\beta \right) /\left( 1+\beta \right) }$. Introducing the Lagrange variables $\eta $ and $\xi $, related to the Euler coordinates as $x=x(\eta ,\xi ,t)$, $y=y(\eta ,\xi ,t)$, $z=z(\eta,\xi,t)$, we find that the vector normal to the foil surface element is given by $%
d\sigma _{i}=\varepsilon _{ijk}dx_{j}dx_{k}$. Here $dx_{j}$ are the vectors directed along the $i$-axes, $\varepsilon _{ijk}$ is the fully antisymmetric unity tensor, and a summation over repeated indices is assumed. Using these relationships we can find the equations of foil motion $$\frac{\partial p_{i}}{\partial t}=\frac{\mathcal{P}}{\nu _{0}}\varepsilon
_{ijk}\frac{\partial x_{j}}{\partial \eta }\frac{\partial x_{k}}{\partial
\xi },$$ $$\frac{\partial x_{i}}{\partial t}=c\frac{p_{i}}{\sqrt{%
m_{p}^{2}c^{2}+p_{k}p_{k}}}.$$ Here $\nu _{0}=n_{0}l_{0}$ is the initial surface density, $%
p_{i}=(p_{x},p_{y},p_{z})$ is the momentum, $x_{i}=(x,y,z)$ is the foil element coordinate and, the index, $m_{p}$ is the ion mass. In the nonrelativistic limit for constant pressure $\mathcal{P}$, this system is reduced to the equations obtained in Ref. [@Ott].
When a planar foil is irradiated by a normally incident EM pulse, the ions achieve the energy $$\gamma _{i}=1+\frac{2w^{2}}{1+2w},$$ where $w$ is the normalized fluence, $$w=\int_{-\infty }^{t-x/c}\frac{E_{0}^{2}(\psi )}{2\pi n_{0}lm_{i}c}d\psi .$$ In the limit $w\gg 1$ the resulting ion energy is equal to the ratio of the laser pulse energy, $\mathcal{E}_{las}$ , to the total number of accelerated ions, $N_{tot}$, i.e. $\gamma _{i}\approx \mathcal{E}%
_{las}/m_{i}c^{2}N_{tot} $. As an example, we consider a solid density foil, $n_{0}=10^{24}$cm$^{-3}$, of 1$\ \mu $m thickness irradiated by a laser pulse with a transverse size of 100$\ \mu $m. For the laser pulse energy of the order of 200 kJ we find that the accelerated ion energy is equal to 1$\
$TeV with a total ion number of 10$^{12}$. The ion acceleration length in this case is approximately equal to $l_{acc}\approx 0.5~$km.
In order to achieve high values of the ion bunch luminosity it is highly desirable to decrease the transverse bunch size. This can be achieved by modulating the density inside the foil, e.g. by a properly modulated laser pulse. The analysis of the linearized equations of the foil motion demonstrates the exponential growth of the modulations $$x_{i}^{(1)}(\alpha ,\beta ,t)\propto \exp \left[ {\left( {\frac{t}{\tau _{RT}%
}}\right) ^{1/3}-iq\eta -ir}\xi \right] , \label{eq34}$$ where $$\tau _{RT}=\omega _{0}^{-1}\frac{(2\pi )^{3/2}R_{0}^{1/2}}{6\left( {%
q^{2}+r^{2}}\right) ^{3/2}\lambda _{0}^{2}}. \label{eq37}$$ and [@PB-07] $$R_{0}=\frac{E_{0}^{2}}{2\pi n_{0}l\omega _{0}^{2}}.$$ This opens a way for focusing the accelerated ions onto a narrow spot with lower limit given by the foil thickness. Using these results we can estimate the RPDA accelerated ion bunch luminosity as $$\mathcal{L}=10^{35}\left( \frac{f}{10KHz}\right) \left( \frac{N_{tot}}{%
10^{12}}\right) ^{2}\left( \frac{10^{-4}cm}{\sigma _{\bot }}\right) ^{2}\
\frac{1}{\mathrm{cm}^{2}\mathrm{s}}.$$
A first indication of the RPDA - regime has been obtained in the experiments [@Kar], when a thin foil target has been irradiated by a laser pulse with an intensity approaching 10$^{20}$W/cm$^{2}$.
Mini-black-holes on Earth {#MBHE}
=========================
We may see that when LWFA and RPDA accelerators will reach 100 GeV and TeV particle energies, which corresponds to the energy range of interest for high energy physics, laser accelerators may be considered as a source of ultrarelativistic particle beams with parameters comparable to those, which are produced by standard accelerators.
As an example for the problems in the field of high energy physics and astrophysics which may be explored with laser accelerators of charged particles, we note the mini-black-hole detection. In the general relativity theory black holes play a fundamental role. The Einstein equation [LLTF,WMTW]{}, $$R_{\mu \nu }-\frac{1}{2}g_{\mu \nu }R=-\frac{8\pi }{m_{p}^{2}}T_{\mu \nu },$$ where $m_{p}^{2}=1/G$ is the square of the Planck mass, and $T_{\mu \nu }$ is the energy-momentum tensor (the units $\hbar =c=1$ are used), has a Schwarzschild solution for the interval: $$ds^{2}=g_{\mu \nu }dx^{\mu }dx^{\nu }=-\eta (r)dt^{2}+dr^{2}/\eta
(r)+r^{2}d\Omega ^{2}$$ with $$\eta (r)=1-(2/m_{p}^{2})M/r.$$ Here $M$ and $\Omega $ are the object mass and the surface element in the 3D space. The metric given by this interval has a singularity at $r$ equal to the Schwarzschild radius, $R_{BH}=2M/m_{p}^{2}$. For an object with a mass of the order of the solar mass the black hole radius is equal to 2 km. As it was noted above, a black hole with the size of about the Planckian length, 10$^{-33}$cm, has the mass $m_{p}=10^{-5}$g, which corresponds to an energy approximately equal to 10$^{19}$GeV.
The situation may change, if our world’s dimension is higher than 3. In accordance with modern quantum field theory [@ArH], our world may have higher dimensions $(d+3)$. The additional dimensions are compactified in a sufficiently small scale, $R_{comp}$. Gravitational interaction is present in the whole space due to its universal character. At the small scale for $%
r\ll R_{comp}$, the gravitational potential of the field produced by an object with mass $M$ behaves as $\phi (r)=M_{f}^{d+2}M/r^{(1+d)}$. A constant $M_{f} $ characterizes the gravitational interaction in the small scale limit. In the limit of large scale compared with $R_{comp}$, i.e. for $%
r\gg R_{comp}$ we have the expression $\phi
(r)=M/m_{p}^{2}r=M/rR_{comp}^{d}M_{f}^{d+2}$. It yields a relationship between $m_{p}$, $R_{comp}$, and $M_{f}$ being $%
m_{p}^{2}=M_{f}^{d+2}R_{comp} $. The solution of the Einstein equation in space with d+3 dimension gives for the interval the formula $$ds^{2}=-\eta (r)dt^{2}+dr^{2}/\eta (r)+r^{2}d\Omega _{d+3}^{2},$$ where $\Omega _{d+3}$ is the surface element and the metric element is equal to $$\eta (r)=1-(R_{BH}/r)^{d+1}.$$ This results in the expression for the black hole radius $$R_{BH}^{d+1}=(2/(d+1))M_{f}^{-(d+1)}M/M_{f}.$$
For the constant $M_{f}$ of the order of 1 TeV the black hole radius is equal to $R_{BH}\approx 10^{-4}\ $fm (here 1 fm=10$^{-13\ }$cm). A probability of a black hole creation is proportional to the cross section of this process [@MBH] $$\sigma _{BH}=\pi R_{BH}^{2}.$$ In the collision of 7 TeV proton bunches with a luminosity of the order of the LHC luminosity, $\mathcal{L}=10^{34}$cm$^{-2}$s$^{-1}$, it is expected that approximately 10$^{9}$ mini-black-holes may be detected per year. The created mini-black-holes are to be detected by their emission of electromagnetic radiation and of elementary particles according to the Hawking mechanism. At the end of its evolution the black hole is thought be strings.
The TeV range laser accelerator of charged particles can generate 10$^{6}$ mini-black-hole per year, when its repetition rate is 1 Hz.
Flying Mirror Concept of the Electromagnetic Wave Intensification {#FMC}
=================================================================
An electromagnetic wave reflected off a moving mirror undergoes frequency multiplication and corresponding increase in the electric field magnitude. The multiplication factor $(1+\beta _{M})/(1-\beta _{M})$ is approximately proportional to the square of the Lorentz factor of the mirror, $\gamma
_{M}=1/\sqrt{1-\beta _{M}^{2}}$, making this effect an attractive basis for a source of powerful high-frequency radiation. Several ways have been suggested to extremely high intensity (see articles, Refs. [BET-03,ARUT,NAU]{}, [@MTB-06; @PPR] and literature quoted in). A specular reflection by a sufficiently dense relativistic electron cloud as suggested in Refs. [@LAND]. The reflection at the moving ionization fronts was studied in Refs. [@IONS].
Here we consider the flying mirrorconcept[@BET-03]. It uses a fact that at optimal conditions, the dense shells formed in the electron density in a strongly nonlinear plasma wake, generated by a short laser pulse, reflect a portion of a counter-propagating laser pulse. In the wake wave generated by the ultrashort laser pulse electron density modulations take the form of a paraballoid moving with the phase velocity close to the speed of light in vacuum [@PAR]. At the wave breaking threshold the electron density in the nonlinear wake wave tends towards infinity. The formation of peaked electron density maxima breaks the geometric optics approximation and provides conditions for the reflection of a substantially high number of photons of the counterpropagating laser pulse. As a result of the electromagnetic wave reflection from such a “relativistic flying mirror”, the reflected pulse is compressed in the longitudinal direction, which is a consequence of frequency upshifting. The paraboloidal form of the mirrors leads to a reflected wave focusing into the spot with the size determined by the shortened wavelength of the reflected radiation (see Fig. \[fig:08\]). This mechanism allows to generate extremely short, femto-, atto-, zepto-second duration pulses of coherent electromagnetic radiation with extremely high intensity, which pave the way for studying such nonlinear quantum electrodynamics effects as the electron-positron pair creation and nonlinear refraction in vacuum.
![Flying Mirror Concept. a) The reflection of EMW at the relativistic mirror results in a frequency upshifting and compression of the wave. b) Paraboloidal modulations of the electron density in the plasma wake wave. c) The electric field pattern of the laser pulse driver and of the reflected EMW. Inset: The reflected electromagnetic pulse frequency is upshifted, it is focused and its intensity increases.[]{data-label="fig:08"}](fig-08.jpg){width="9cm"}
The key parameter in the problem of Flying Relativistic Mirror (FRM) is the wake wave gamma factor, $\gamma _{ph,W}$. According to the special theory of relativity [@STR], the frequency of the electromagnetic wave reflected from FRM increases by a factor approximately equal to $4\gamma _{ph,W}^{2}$. A number of back reflected photons is proportional to $\gamma _{ph,W}^{-3}$ (for details see Ref. [@PAN]), which results in the reflected light intensification [@BET-03] $$I_{r}/I_{0}\approx \gamma _{ph,W}^{3}(S/\lambda _{0})^{2}, \label{eq:intFM}$$ where $S$ is the transverse size of the laser pulse incident on the FRM. The reflected pulse power increases as $\mathcal{P}_{r}=\mathcal{P}_{0}\gamma
_{ph,W}$.
Using the expression for the reflected pulse intensity (\[eq:intFM\]), we obtain that the interaction of two laser pulses with energies 10 kJ and 30 J, respectively, counterpropagating in a plasma with a density $\approx
10^{18}$cm$^{-3}$ can result in a light intensification of up to $\approx
10^{28}$W cm$^{-2}$. This corresponds to the generation of an electric field with a value close to the nonlinear quantum electrodynamics (QED) limit, $E_{QED}=m_{e}^{2}c^{3}/e\hbar $, when electron-positron pairs can be created in vacuum. This QED electric field is also called the “Schwinger field”.
Experiments utilizing the electromagnetic pulse intensified with the FRM technique may allow studying regimes of super-Schwinger fields, when $%
E>E_{QED}$. This may be possible because the light reflected by the parabaloidal FRM is focused into a focus spot moving with a relativistic velocity and is well collimated within an angle $\approx 1/\gamma _{ph,W}$ [@BET-03]. The wave localization within the narrow angle corresponds to the fact that the wave properties are close to the plane wave properties to the extent of the smallness of the parameter $1/\gamma _{ph,W}$. In this case the second Poincare invariant of the electromagnetic field, $%
B^{2}-E^{2} $, has a value of the order of $E^{2}/\gamma _{ph,W}^{2}$. Therefore the electric field amplitude in the reflected electromagnetic wave can exceed the Schwinger limit by $\gamma _{ph,W}$ times. We note that a tightly focused electromagnetic wave cannot have an amplitude above $E_{QED}$, due to the electron-positron pair creation [@NAR] when $E\rightarrow
E_{QED}$ leads to the depletion of the electromagnetic wave [@BFP].
As it was shown above, the critical power for mutual focusing of two counterpropagating EMW is equal to $\mathcal{P}_{cr}=2.5\times 10^{24}$W, which is beyond the reach of existing and planned lasers. Fortunately, if we take into account that the radiation reflected by the FRM has a shortened wavelength $\lambda _{r}=\lambda _{0}/4\gamma _{ph,W}^{2}$ and that its power is increased by a factor $\gamma _{ph,W}$, we may find that for $%
\gamma _{ph,W}=30$, i.e. for a plasma density $\approx 3\times 10^{17}$cm$%
^{-3}$, nonlinear vacuum properties can be seen for laser light the incident on the FRM with a power of about 10 PW. This makes the FRM concept attractive for the purpose of studying nonlinear quantum electrodynamics effects.
Within the framework of the Flying Mirror concept, it has been demonstrated [@BET-03] that the wavelength of the laser pulse, which has been reflected and focused at the wake plasma wave, becomes shorter by a factor $%
4\gamma _{ph}^{2}$ and its power increases by a factor $2\gamma _{ph}$. From this it follows that nonlinear QED vacuum polarization effects are expected to be observable for 50 PW 1-$\mu $m lasers.
A demonstration of the Flying Mirror concept has been accomplished in the experiments of Ref. [@KP]. Two beams of terawatt laser radiation interacted with an underdense plasma slab. The first laser pulse excited the nonlinear wake wave in a plasma with parameters required for the wave breaking, which has been seen in the quasi-mono-energetic electron generation and in the stimulated Raman scattering. The second counter-crossing laser pulse has been partially reflected from the relativistic mirrors formed by the wake plasma wave. We detected the electromagnetic pulses with a duration of femtoseconds and wavelengths from 7 nm to 15 nm. These results demonstrate the feasibility of constructing sources of coherent X-ray radiation with the parameters that are tunable in a broad range.
Reconnection of Magnetic Field Lines & Vortex Patterns
======================================================
The term magnetic field line reconnection refers to a broad range of problems that are of interest for space and laboratory plasmas. The results of theoretical and experimental studies of magnetic reconnection have been reviewed in many papers and monographs [@ACR; @SYROV; @BISKAMP]. As it concerns relativistic laser plasmas, earlier a conclusion was made in Ref. [@ASK] about the important role of the generation of magnetic fields by fast electron currents and their reconnection in the relativistic laser-matter interaction regime. The experiments conducted in Refs. [@Nil] revealed magnetic reconection phenomena in laser plasmas, when two high power laser beams irradiated a thin foil target.
Processes of reconnection are accompanied by an ultra fast magnetic energy release, which is transformed into different forms, such as internal plasma energy, radiation and fast particles.
Dimensionless parameters describing the relative roles of nonlinear, dissipative and Hall effects {#S.DIM}
-------------------------------------------------------------------------------------------------
Inside current sheets, which are basic entities in the reconnection process, as well as in the vicinity of shock wave fronts, the effects of dissipation and of nonlinearity play a crucial role being comparable in magnitude. Together with the Hall effect, which leads to the appearance of small scale structures, these effects violate the freezing of magnetic field in plasma motion.
Magnetic field is frozen in a plasma in the limit of a large Lundquist number $S\rightarrow \infty $. It obeys the equation: $$\partial _{t}\mathbf{B}=\mathbf{v}\times \mathbf{B},$$ which corresponds to the conservation of magnetic field flux through a contour moving with the plasma. The dimensionless parameter $S$ is equal to the ratio of two characteristic time scales: the magnetic diffusion time $%
\tau _{\sigma }=l^{2}/\nu _{m}$ and a typical time $\tau _{A}=l/v_{A}$ that it takes for an Alfvèn wave to propagate along the distance $l$; $S=\tau
_{\sigma }/\tau _{A}$. Here the magnetic diffusivity is $\nu _{m}=c^{2}/4\pi
\sigma $, where $\sigma $ is the electric conductivity of the plasma, and $%
v_{A}=|\mathbf{B}|/\sqrt{4\pi \rho }$ is the Alfvèn wave velocity.
In the vicinity of the zero point the scale of the field nonuniformity $l$ equals to the distance $r$ from the zero point. The magnetic field and hence the Alfvèn velocity are proportional to $r$: $|\mathbf{B}|=hr$, $%
v_{A}=hr/\sqrt{4\pi \rho }\equiv \Omega _{A}r$. Here $h$ is a typical value of the gradient of the magnetic field.
The measure of the significance of nonlinear effects is given by the ratio $%
\delta B/hl$ between the magnitude of the magnetic field perturbation $%
\delta B$ and the background magnetic field $B=hl$. This ratio depends on the distance from the null point, due to both the nonuniformity of the background magnetic field $B$ and to the change of the MHD wave amplitude in the course of its propagation.
If the plasma is pinched by a the quasi-cylindrical electric current $I$ with a radius of the order of $r$, the value of the magnetic field at its boundary is approximately $\delta B=2I/cr$. The dimensionless ratio $\delta
B/hl$ is equal to one for $r\approx r_{m}=\sqrt{I/hc}$. If the electric current has the form of a quasi-one-dimensional slab pinch, and if the pinching occurs in the direction of its small size, the characteristic value of the magnetic field perturbation is constant: $\delta B=B_{\Vert }$ and the ratio $\varepsilon $ becomes of order unity at the distance $%
r_{A}=B_{\Vert }/h$. In the approximation of small amplitude perturbations, these two types of pinching correspond to the effects of the propagation of magnetoacoustic and of Alfvèn waves, respectively. The magnetoacoustic waves focus towards the null line, while the energy of the Alfvèn waves accumulate near the magnetic field separatrices. The values $r_{m}$ and $%
r_{A}$ determine the size of the region, where the magnetoacoustic wave and, respectively, the Alfvèn one become nonlinear.
The dimensionless parameters $$(r_{m}/r_{\sigma })^{2}=I\Omega _{A}/c\nu _{m}\equiv L_{m}, \label{eq:4.3}$$ $$(r_{A}/r_{\sigma })^{2}=B_{\Vert }^{2}\Omega _{A}/h^{2}\nu _{m}\equiv L_{A}
\label{eq:4.4}$$ determine the relative role of the dissipation and of the nonlinearity effect in the course of the current sheet formation due to finite amplitude perturbations of the magnetoacoustic and of the Alfvèn wave type, respectively.
Now we discuss the relationship between the dimensionless parameter $L_{m}$ and the current sheet parameters obtained in the framework of the Sweet – Parker model [@SWEET; @PARKER]. In this model it is supposed the current sheet has a width $b$ and a thickness $a$ with $b\gg a$. The plasma flows into the current sheet with a velocity $v_{\mathrm{in}}\approx \nu _{m}/a$ and exits through its narrow edges with a velocity $v_{\mathrm{out}}$, which is of the order of the Alfvèn wave velocity, $v_{A}\approx \Omega _{A}b$. From mass conservation we obtain $v_{\mathrm{in}}b=av_{\mathrm{out}}$. From this it follows that the thickness of the current sheet is equal to $$a=\sqrt{\nu _{m}/\Omega _{A}},$$ i.e. of order $r_{\sigma }$. Estimating the current sheet width as $r_{m}$, we find that the ratio of its width to its thickness is $$b/a=\sqrt{I\Omega _{A}/hc\nu _{m}}\equiv \sqrt{L_{m}}.$$ Thus, the condition for the formation of a wide current sheet with $b\gg a$ is equivalent to the requirement $L_{m}\gg 1$. Similarly, current sheets are formed in the vicinities of the magnetic field separatrices when $L_{A}\gg 1$.
Considering the case when the Hall effect, i.e. the electron inertia, plays a dominant role in the reconnection process, we define the dimensionless parameter which measures the role of the Hall effect as $\tilde{\alpha}%
=\alpha h/\Omega _{A}l\equiv c/\omega _{pi}l$. When the length $r_{H}$, at which the Hall effect starts to be important, is larger than the current sheet thickness, $r_{H}/a=\alpha h/\sqrt{\nu _{m}\Omega _{A}}>1$, the effects of dispersion lead to the formation of small scale structures. In the limit $r_{H}/b=\alpha h\sqrt{ch/I}/\Omega _{A}\gg 1$ the pattern of the plasma flow is completely determined by the Hall effect. Similar to the way used to define the parameters $L_{m}$ and $L_{A}$, we define the dimensionless parameter $$L_{H}=(b/r_{H})^{2}=E^{2}c^{2}\Omega _{a}/h^{4}\alpha ^{2}\nu _{m}\equiv
E^{2}\omega _{pi}^{2}/h^{2}\nu _{m}\Omega _{a}.$$ When $L_{H}\gg 1$, nonlinear effects are much stronger than the Hall effect.
Current Sheet
-------------
In a simple 2D configuration the current sheet is formed in the magnetic field described by a complex function $B(x,y)=B_{x}-iB_{y}=h\zeta $ of a complex variable $\zeta =x+iy$. The magnetic field vanishes at the coordinate origin. The magnetic field lines lie on the surfaces of constant vector potential, $A(x,y)=\mathrm{Re}\{h\zeta ^{2}/2\}$. They are hyperbolas as we can see in Fig. \[fig:09\] a). This is a typical behaviour of the magnetic field lines in the vicinity of null lines (they are the so called X-lines) in magnetic configurations.
Under finite time perturbations the magnetic X-line evolves to the magnetic configurations of the form $B=h(\zeta -b)^{1/2}$, which desribes the magnetic field created by thin current sheet between two points $\pm b$ [SYROV]{}. The magnetic field lines lie on the constant surfaces of $$A(x,y)=\frac{h}{2}\mathrm{Re}\left\{ \zeta \sqrt{\zeta ^{2}-b^{2}}-\mathrm{%
Log}\left[ \zeta +\sqrt{\zeta ^{2}-b^{2}}\right] \right\} .$$ They are shown in Fig. \[fig:09\] b. The width of the current layer $b$ is determined by the total electric current $I$ inside, and by the magnetic field gradient, $h$. It is equal to
$$b=\sqrt{4I/hc.}$$
In the strongly nonlinear stage of the magnetic field and plasma evolution a quite complex pattern in the MHD flow in the nonadiabatic region near the critical point can be formed, with shock waves and current sheets. In Fig. \[fig:10\] we show the results of the dissipative magnetohydrodynamics simulations of the current sheet formation near the X-line.
Magnetic Reconnection in Collisionless Plasmas
----------------------------------------------
When the Hall effect is dominant, i.e. the electron inertia determines the relationship between the electric field and the electric current density carried by the electron component, the magnetic field evolution is described by the equation (see [@EMHD; @BPS-92]) $$\partial _{t}(\mathbf{B}-\Delta \mathbf{B})=\nabla \times \left[ \left(
\nabla \times \mathbf{B}\right) \times (\mathbf{B}-\Delta \mathbf{B})\right]
, \label{eq:EMHD}$$ which corresponds to the condition of generalized vorticity, $\mathbf{\Omega
}=\mathbf{B}-\Delta \mathbf{B}$, be frozen into the electron component motion with the velocity $\mathbf{v}_{e}=c\nabla \times \mathbf{B/}4\pi
n_{0}e$. Here the space scale is chosen to be equal to the collisionless electron skin-depth, $d_{e}=c/\omega _{pe}$, and the time unit is $\omega
_{Be}^{-1}=m_{e}c/eB$. The range of frequencies described by the EMHD equations is given by $\omega _{Bi}<\omega <\omega _{Be}$.
In the linear approximation Eq. (\[eq:EMHD\]) describes the propagation of whistler waves, for which the relationship between the wave frequency and the wave vector, is $\omega =|\mathbf{k}|(\mathbf{k}\cdot \mathbf{B_{0}}%
)/(1+k^{2})$. From this relationship it follows that in a weakly inhomogeneous magnetic field the critical points are the points and lines where $|\mathbf{B_{0}}|=0$ or/and $(\mathbf{k}\cdot \mathbf{B_{0}})=0$.
The electron inertia effects make the reversed magnetic field configuration unstable against tearing modes [@FKR; @LPV], which result in magnetic field line reconnection. The slab equilibrium configuration with a magnetic field given by $\mathbf{B_{0}}=B_{0z}\mathbf{e_{z}}+B_{0x}(y/L)\mathbf{e_{y}}
$, where $B_{0x}(y/L)$ is the function that gives the current sheet magnetic field, is unstable with respect to perturbations of the form $f(y)\exp
(\gamma t+ikx)$ with $kL<1$. For this configuration one has $(\mathbf{k}%
\cdot \mathbf{B_{0}})=0$ at the surface $y=0$. The growth rate of the tearing mode instability is [@BPS-92; @BASOVA; @FRUSTRA] $\gamma \approx
(1-kL)^{2}\Delta ^{\prime 2}/kL^{2}$.
In Fig. \[fig:11\] the results of a numerical solution of Eq. ([eq:EMHD]{}) in a 2D geometry with magnetic field $\mathbf{B}(x,y,t)=\left(
\nabla \times a\right) \times \mathbf{e_{\bot }}+b\mathbf{e_{\Vert }}$ are shown. The unperturbed configuration is chosen to be a current sheet, infinite in the $x$-direction, that separates two regions with opposite magnetic field. Both the line pattern of generalized vorticity, $\Omega
=a-\Delta a$, and of the magnetic field show the formation of quasi–one–dimensional singular distributions in the electric current density and in the distribution of the generalized vorticity. The magnetic field topology changes, as is seen from Fig. \[fig:11\].
Charged Particle Acceleration
-----------------------------
A fully developed tearing mode results in a current sheet break up into parts separated by a distance $2a$, as it is illustrated in Fig. \[fig:12\]a (see. Ref. [@SYROV] and literature quoted therein). Under the magnetic field line tension the plasma is thrown out. The model magnetic field describing this configuration is given by the complex variable function $%
B(\zeta )=B_{0}\varsigma /\sqrt{a^{2}-\varsigma ^{2}}$. The magnetic field lines lie on the surfaces of constant vector potential, $$A(x,y,t)=\mathrm{Re}\left\{ B_{0}\sqrt{a^{2}(t)-\varsigma ^{2}}\right\} .$$ Due to a dependence of the function $a$ on time the electric field parallel to the $z$ axis arrises. It is given by $$E(x,y,t)=-\frac{1}{c}\partial _{t}A=-\frac{1}{c}\frac{B_{0}a(t)\dot{a}(t)}{%
\sqrt{a^{2}(t)-\varsigma ^{2}}}.$$ In the vicinity of the null line we have a quadrupole structure of the magnetic $B(\zeta )\approx B_{0}\varsigma /a$ field and a locally homogeneous electric field, $E\approx \dot{a}B_{0}/c$.
The magnetic field reconnection, the study of which has been started by Dungey [@Dungey], on its initial stage had had as a main goal to explain the generation of suprathermal particles during solar flares and substorms in the earth’s magnetosphere. Despite the simplycity of the formulation of the problem, it is quite far from a complete solution. Even in the test particle approximation, which describes the particle motion in the given magnetic and electric fields, the solution of this problem meets serious difficulties [@ACR; @RCACC]. The reason of that is due to the fact that in the vicinity of critical points of magnetic configurations the standard approximations adopted to describe the plasma dynamics are no longer valid. In such regions the drift approximation, i.e., the assumption that the adiabatic invariants are constant, can no longer be applied. On the other hand, the particle spends only a finite time interval in the nonadiabatic region, since there its motion is unstable. After a finite time interval it gets out of the nonadiabatic region, and gets into the drift region as it is seen in Fig. \[fig:12\]b. Matching the solution described by the particle trajectories in different regions, we can describe the particle motion and hence the acceleration near critical points of the magnetic configurations.
Under the conditions of space plasmas, the radiation losses during the charged particle acceleration in the magnetic reconnection processes are caused by the backward Compton scattering and by synchrotron losses. A characteristic time of the synchrotron losses for the electron with energy $%
\mathcal{E}$ is given by the expression $$\tau _{B}=\frac{3m_{e}^{4}c^{7}}{2e^{4}B^{2}\mathcal{E}}.$$ As it was shown in Ref. [@BKORF], during solar flares this effect limits the ultrarelativistic electron energy to a value of about several tens of GeV.
Electron Vortices in Collisionless Plasmas
------------------------------------------
The vortical fluid motion is well known to be widely present under the earth’s and space conditions. In laser plasmas, when ultra short and high intensity EMW pulse propagates in the collisionless plasmas, it accelerates a copious number of relativistic electrons. The electric current of fast electrons produces quasistatic magnetic field, whose evolution results in the formation of electron vortex structures. They naturally take a form of the vortex rows [@HMV], as it is shown in the LHS inset to Fig.[fig:13]{}. A strong magnetic field in the relativistic laser plasma has been detected experimentaly [@BBfield].
The interacting vortices can be described within the framework of a two-dimensional theoretical model. By taking $\mathbf{B}$ to be along the $z$-axis ($\mathbf{B=}B\mathbf{{e}_{z}}$), and assuming all the quantities to depend on the $x,y-$coordinates, we obtain from vector equations ([eq:EMHD]{}) one equation $$\partial _{t}(\Delta B-B)+\{B,(\Delta B-B)\}=0 \label{eq:8.2D}$$ for a scalar function $B(x,y,t)$. Hear $$\{f,g\}=\partial _{x}f\ \partial _{y}g-\partial _{x}g\ \partial _{y}f$$ are the Poisson brackets. Equation (\[eq:8.2D\]) is known as the Charney equation [@Batch] or the Hasegawa-Mima (HM) [@HasMima] equation in the limit of zero drift velocity. In this case linear perturbations with the dispersion equation $\omega =|\mathbf{k}|(\mathbf{k}\cdot \mathbf{B_{0}}%
)/(1+k^{2})$ correspond to the Rossby waves, the drift waves or to the whistler waves, respectively.
Equation (\[eq:8.2D\]) has a discrete vortex solution, for which the generalized vorticity is localized at the points $\mathbf{x=x}^{\alpha }$: $$\Omega =\Delta B-B=\sum_{\alpha }\kappa _{\alpha }\delta (\mathbf{x-x}%
^{\alpha }(t)).$$ Solving this equation we find that the magnetic field is a superposition of the magnetic fields created at isolated vortices localized at the coordinates $\mathbf{x}^{\alpha }(t)$: $\,B=\sum_{\alpha }B^{\alpha }$ with $%
~$$$B^{\alpha }(\mathbf{x},\mathbf{x}^{\alpha }(t))=-\frac{\kappa _{\alpha }}{{%
2\pi }}K_{0}(|\mathbf{x}-\mathbf{x}^{\alpha }(t)|)\mathbf{.}
\label{eq:8.BVrt}$$ Here and below $K_{n}(\xi )$ are modified Bessel functions.
The curves $\mathbf{x}^{\alpha }(t)$ are determined by the characteristics of equation (\[eq:8.2D\]). The characteristic equations have the Hamiltonian form $$\kappa _{\alpha }\dot{x}_{i}^{\alpha }=J_{ij}\frac{\partial \mathcal{H}}{%
\partial x_{j}^{\alpha }}=-\frac{1}{2\pi }J_{ij}\sum_{\beta \neq \alpha
}\kappa _{\alpha }\kappa _{\beta }\frac{(x_{i}^{\alpha }-x_{j}^{\beta })}{%
l_{\alpha \beta }^{2}}, \label{eq:8.HpointHM}$$ where $J_{ij}$ is the antisymmetric unit matrix. The Hamiltonian is given by $$\mathcal{H}=-\sum_{\alpha <\beta }\kappa _{\alpha }\kappa _{\beta
}K_{0}(l_{\alpha \beta })/2\pi .$$
In the case of the Euler hydrodynamics, a point vortex is described by $%
(\kappa _{\alpha }/2\pi )\ln \mathbf{|x}-\mathbf{x}^{\alpha }(t)|$, instead of the expression (\[eq:8.BVrt\]) which involves the Bessel function $%
K_{0}\,(\mathbf{|x}-\mathbf{x}^{\alpha }(t)|\mathbf{)}$. The later results in the shiealding of the interaction between vortices at large distances. A typical scale length of the problem under consideration in the case of the EMHD vortex systems, is equal to the collisionless electron skin-depth, $%
d_{e}=c/\omega _{pe}$.
Considering the problem of the stability of an infinite vortex chain we assume that all vortices have the same absolute intensity and take. In the initial equilibrium the vortices have coordinates (Fig. (\[fig:13\]))
In the case of an antisymmetrical vortex row with $\sigma =1/2$, we expect a more complicated behavior of the perturbations, compared to that of the symmetrical configuration. As noted in Lamb’s monograph [@Lamb], in standard hydrodynamics the antisymmetrical von Karman’s vortex row is stable for $q/s\approx 0.281$, where $s$ and $q$ give a distance between the vortices in the unperturbed vortex row along the $x$ and $y$ coordinates. A dependence of the instability growth rate on $s$ and $q$ for the vortex row described within the framework of the Euler hydrodynamics approximation is shown in Fig. \[fig:13\]a.
By direct inspection of the row instability described by the Hasegawa-Mima equations we can see that for large distance between neighbouring vortices the antisymmetric vortex row is stable for $$3s^{2}/4>q>s/2.$$ (see Fig. \[fig:13\]b).
A Role of the Weibel Instability in the Quasistatic and Turbulent Magnetic Field Generation
-------------------------------------------------------------------------------------------
The quasistatic magnetic field generation in relativistic laser plasmas occurs due to the fast electron beam interaction with the background plasma. It can be understood in terms of the Weibel instability [@Weibel] or in the generic case, in terms of the electromagnetic filamentation instability. When the fast electron beam propagates in the plasma, its electric current is compensated by the current carried by the plasma electrons. A repulsion of the oppositely directed electric currents results in the electron beam filamentation and in the generation of a strong magnetic field [@Cali]. An electromagnetic filamentation instability leads to the generation of a quasistatic magnetic field and is associated with many small-scale current filaments [@Honda]. Each filament consists of a direct and of a return electric current which repel each other. This produces a strong electric field, which accelerates the ions in the radial direction. In the long term evolution, the successive coalescence of the small-scale current filaments forms a large scale magnetic structure. This process is accompanied by the reconnection of the magnetic field lines, by the formation of current sheets, and by strong ion acceleration inside these sheets [@Sakai1].
The filamentation phenomena are of great interest for the explanation of the quasi-static magnetic field origin in laser plasmas irradiated by relativistically strong EMW [@BAsk]. Counterstreaming electric current configurations naturally appear in space at the fronts of colliding electron-positron and electron-ion plasma clouds [@Kazimura] as in the cases of the Galactic Gamma Ray Bursts and in shock waves in supernova remnants. The filamentation instability generates the magnetic field required by the theory of the synchrotron afterglow in GRB [@Med1]. The Weibel instability has been invoked as a mechanism of the primordial magnetic field generation by colliding electron clouds in cosmological plasmas [@Sakai2].
The filamentation instability developing in the vicinity of shock wave fronts together with other types of instabilities [@TKrll] plays the role of the source of strong electromagnetic turbulence invoked in the theoretical models of the Fermi acceleration of cosmic rays [Med2,Takabe]{}. A realization of the Fermi acceleration mechanism of Type A at the shock wave front is discussed below in Section \[Sec:SW\].
Relativistic Rotator
====================
In Ref. [@GG] the antenna mechanism of the pulsar radiation emission has been proposed. According to this mechanism in the pulsar magnetoshere, which is a rotating magnetic dipole, the magnetic dipole interaction with a plasma at the magnetosphere periphery induces strong modulations of the electron density, an electron density lump. The phase velocity of the electron lump can be arbitrarily close to the speed of light in vacuum. It is directed along a circle as illustrated in Fig. \[fig:GG\]. As a result of the curvilinear acceleration, the electron lump emits a radiation, whose properties are similar to the synchrotron radiation [@VLG].
In the context of Relativistic Laboratory Astrophysics it is remarkable that the relativistic rotating dipole can naturally be formed in the laser plasma. Laser-plasma interactions provide an opportunity to reproduce nonlinear electrodynamics effects under astrophysical conditions in the laboratory. In Ref. [@Afterglow] it is demonstrated that high-power coherent synchrotron-like radiation can be generated by the relativistic charge density wave rotating self-consistently inside an electromagnetic-dipole solitary wave, dwelling in a laser plasma. The relativistically strong laser pulse can generate relativistic EM subcycle solitary waves in a plasma [@EMSoliton], as it was indicated by particle-in-cell (PIC) simulations. An analytical description of solitons of this type was developed in Refs. [@Kozlov]. Figure \[fig:15\] presents the structure of electric and magnetic fields inside the soliton [3Dsoliton]{}. The soliton ressembles an oscillating or rotating electric dipole. The toroidal magnetic field, shown in Fig. \[fig:15\], indicates that, besides the strong electrostatic field, the soliton also has an electromagnetic field. The electrostatic and electromagnetic components in the soliton are of the same order of magnitude.
The 3D solitons emit high-frequency EM radiation, whose frequency is much higher than the Langmuir frequency [@Afterglow]. This radiation is emanated from the electron density hump rotating in the wall of the soliton cavity, similar to coherent synchrotron-like emission. This radiation has the characteristics of a well pronounced outgoing spiral EM wave, Fig. [fig:16]{} a). The emission of the spiral wave correlates to the rotation of the electron density hump in the cavity wall, and it leads to the spiral modulations of the electron density (see Fig. \[fig:16\] b)). The density hump gyrates in a circle, and the period of revolution is exactly equal to the soliton period. The polarization of the spiral wave corresponds to the well known synchrotron radiation [@VLG] and the density hump emission is coherent.
The results of 3D PIC simulations, presented in Ref. [@Afterglow], distinctly demonstrated relativistic rotating dipoles excited by the circularly polarized laser pulse in an underdense plasma. The dipoles are associated with the relativistic electromagnetic solitons.
Shock Waves {#Sec:SW}
===========
Phenomena taking place at shock-wave fronts play a key role in various astrophysical conditions. The characteristic dimensionless parameters that determine the shock wave propagagation are the magnetic Mach number, $%
M_{A}=v_{SW}/v_{A}$, equal to the ratio of the shock wave front velocity, $%
v_{SW}$, to the Alfven velocity, $v_{A}$, the ratio of the gas pressure to the magnetic pressure, $8\pi nT/B^{2}$, and $\theta $, the angle between the normal to the front and the magnetic field.
Shock Waves in Supernova Remnants
---------------------------------
The origin of cosmic rays (CR) is one of the most interesting problems in astroparticle physics [@ACR; @CRLEC]. The observation of ultrahigh-energy cosmic rays indicates that cosmic rays exist beyond 10$^{20}$ eV and certainly beyond 10$^{19}$ eV energies greater than the GZK cutoff [@GZK] for the extragalactic sources due to the pionization loss of protons that decay by collision with cosmic microwave background photons. The galactic CR spectra in the energy range above a few GeV and below $\approx 10^{7}$GeV are power-laws with the total cosmic ray spectrum being $$I_{CR}=1.8\times \mathcal{E}^{-\kappa }\frac{\mathrm{particles}}{\mathrm{cm}%
^{2}\mathrm{s\ st\ GeV}} \label{eq:CRspectr}$$in the energy range from a few GeV to 100 TeV with $\kappa \approx 2.7$. Around 10$^{15}$ eV (the knee), the slope steepens from $\kappa \approx 2.7$ to $\kappa \approx 3$. The energies 10$^{18}$ eV correspond to the ultra high energy cosmic rays (UHECR), which sources are associated with active galactic nuclei (AGNs) [@AUGER].
For the most advanced theoretical models of galactic cosmic ray acceleration with the energy below 10$^{17}$ eV the shock waves formed in the supernova explosions are most important. This process is related to the nature of collisionless shock waves [@TKrll].
During explosions of type II supernovae an energy $\mathcal{E}_{tot}$ of the order of $10^{51}$erg is released. The frequency of supernova explosions is about 1/30 per year. Estimates [@ACR] show that approximately 2% of the energy of a supernova should be transferred into the cosmic ray energy.
In the initial stage of the evolution of a supernova envelope a system of shocks is formed (\[fig:17\]). The matter ejected from a star is decelerated and compressed in the inner shock wave. Through the circumstellar gas a second shock wave propagates. The matter ejected from a star is separated from the circumstellar gas by a contact discontinuity, which is unstable with respect to a Rayleigh-Taylor instability. The RT instability leads to the relatively long scale modulations of the gas density inside the supernova shells.
When the mass of the swept interstellar gas becomes larger than the mass ejected from the star, the propagation of outer shock in Fig. (\[fig:17\]) is described by the Sedov-Taylor self-similar solution. The radius of the shock, $R_{SW}$, as a function of time is related to the energy, $\mathcal{E}%
_{SN}$, released in the explosion and to the gas density $\rho _{0}$ by the relation $$R_{SW}(t)=1.51\left( \frac{\mathcal{E}_{SN}}{\rho _{0}}\right) ^{1/5}t^{2/5}=%
\frac{5}{2}v_{SW}t.$$The shock wave velocity, $v_{SW}(t)\approx t^{-3/5}$, decreases with time. At a later time when the radiation losses become important, the law of the supernova envelope expansion changes. The asymptotic time dependence of the SN envelope radius is given by $R_{SW}(t)\approx t^{2/7}$ (see Ref. [ACR]{} and references therein).
Collisionless Shock Waves
-------------------------
If the shock wave has a relatively small amplitude, $M_{A}<M_{1}\approx 1.5$ (the precise value depends on $\beta $ and $\theta $), then the front profile is laminar in structure and it is determined by a joint action of the dispersion and dissipation on the nonlinear waves propagation. These effects are described in the framework of the Korteweg-de Veries-Burgers equation: $$\partial _{t}u+u\partial _{x}u-\nu \partial _{xx}u+\beta \partial _{xxx}u=0.
\label{eq:10kdvB}$$
The stationary wave propagating with constant velocity is described by a solution, which shows the change of the amplitude of the wave from zero far ahead of the shock wave front, to $u_{1}=2v_{sw}$ far behind the shock wave front.
The decay of the oscillation amplitude, with the coefficient equal $\nu
/\beta $, results in the decrease of the amplitude of solitons as it is shown in Fig. \[fig:18\].
If dissipation effects are more important than the effects of dispersion, $%
\nu /\beta \gg 1$, there are no oscillations at the shock wave front. More precisely, the decay should be large enough, $\nu \gg v_{cr},$ with $$\nu _{cr}=\sqrt{4\beta u_{1}}.$$ In this case the wave has a monotonous structure.
In the case of $\nu /\beta \ll 1,$ the dispersion effects are dominant and there are many well seen solitons near the front. For $\beta >0$ the oscillations are localized behind the front (Fig. \[fig:12.4\]a), while for $\beta <0$ they are ahead of the front (Fig.\[fig:12.4\]b).
For example, in the case of the magnetoacoustic shock waves in a plasma, propagating almost perpendicularly to the direction of the magnetic field, the dispersion coefficient, $\beta \approx v_{a}c^{2}/2\omega _{pe}^{2}$, is positive. This means that the oscillations are localized behind the front of the magnetoacoustic shock wave propagating perpendicularly to the magnetic field. When the direction of the magnetoacoustic wave propagation is almost parallel to the direction of the magnetic field, the coefficient $\beta
\approx -v_{a}c^{2}/2\omega _{pi}^{2}$ is negative with $\omega _{pi}=\sqrt{%
4\pi ne^{2}/m_{i}}$. The oscillations at the front of the magnetoacoustic shock wave, propagating quasi-parallely with respect to the magnetic field, are localized ahead of the front.
Dissipation, which determines the distance of the oscillation decay, can be due to anomalous resistance and viscosity arising from an excitation of the plasma instability, i.e. the Weibel instability of counterpenetrating plasmas. If the amplitude of the shock wave is large, $M_{A}>3$, a high level of turbulent fluctuations of electric and magnetic fields are excited ahead and behind the wave front.
In the laser-plasma physics context, the observation of collisionless shocks was reported by several authors [@SHW], aiming to reproduce astrophysical phenomena in small scale laboratories. However, in general when the shocks were observed with optical probing techniques, the front structure could hardly be resolved. In Ref. [@LRom] the propagation in a rarefied plasma ($n_{e}<10^{15}cm^{-3}$) of collisionless shock waves being excited following the interaction of a long ($L=470ps$) and intense ($%
I=10^{15}Wcm^{-2}$) laser pulse with solid targets, has been investigated via proton probing techniques [@PrIm]. The shocks’ structures and related electric field distributions were reconstructed with high spatial and temporal resolution. The experimental results are described within the framework of the nonlinear wave description based on the Korteweg–de Vries–Burgers equation (\[eq:10kdvB\]).
Diffusive Acceleration of Charged Particles at the Shock Wave Front
-------------------------------------------------------------------
The charged particle interaction with fluctuations of the electric and magnetic field in a turbulent plasma may result in particle scattering and diffusion. When the shock wave propagates in a turbulent medium, an average velocity of electromagnetic fluctuations is different in the regions ahead and behind the shock front. Efficiently the particle appears to move between semitransparent (due to diffusion) walls with a decreasing distance between them. A model transport equation describing the particle convection, diffusion and acceleration has a form [@ACR] $$\partial _{t}\mathrm{\,}f+\mathrm{div}(\mathbf{u}\mathrm{\,}f-D\mathrm{\,}%
\nabla f)=$$ $$\frac{1}{p^{2}}\partial _{p}\left[ p^{2}\left( \frac{p}{3}\mathrm{div\ }%
\mathbf{u}-K(p)\right) \mathrm{\,}f\right] , \label{eq:CRtrEq}$$ where $f(p,x,t)$ is the fast particle distribution function, $p$, $x$ and $t$ the particle momentum, coordinate and time, $v_{SW}$ being the speed of the shock wave propagation, and $D$ is the diffusion coefficient. A term in the right hand side describes regular acceleration or deceleration of the charged particles: $$\frac{dp}{dt}=-K(p)-\frac{1}{3}p\ \mathrm{div\,}\mathbf{u}.$$ The function $K(p)$ corresponds to the Compton and synchrotron losses important for the cosmic ray electron component: $$K(p)=-\beta _{B}cp^{2}$$ with $$\beta _{B}=8\times 10^{-25}\left( \frac{B^{2}}{8\pi }+w_{ph}\right) \frac{1}{%
eV\ s}$$ An average change of the particle momentum proportional to $p\mathrm{\,div\,}%
\mathbf{u}$ occurs due to the particle bouncing between converging, $\mathrm{%
div\,}\mathbf{u}<0$ , or diverging, $\mathrm{div\,}\mathbf{u}<0$, scattering centres.
The average particle bouncing between two reflecting plates with distance $L
$ as a function of time provides a simple example of a dynamic system with conservation of the longitudinal adiabatic invariant, $J_{||}=pL$ [LibLich,SVB-SNS]{}. The phase plane shown in the inset to Fig. \[fig:19\], illustrates the Fermi acceleration mechanism of the first type (type A according to Ref. [@Fermi]). By virtue of the longitudinal adiabatic invariant conservation, for decreasing distance between the plates, $dL/dt<0$, the particle momentum grows, i.e. the particle acquires energy.
The velocity distribution in the vicinity of the front of an infinitely thin shock wave, propagating from left to right, has the form: $u(X)=u_{1}$ in the region $X>0$, and $u(X)=u_{2}$ for $X<0$. Here $X=x-v_{SW}t$. The velocities ahead the shock front and behind it are related to each other as $$u_{2}=u_{1}\frac{\kappa +1}{\kappa -1}.$$ Here, $\kappa $ is the polytropic index. For an infinitely thin shock wave front the divergence of the velocity is equal to $$\mathrm{div\ }\mathbf{u=(}u_{1}-u_{2})\delta (X).$$
Substituting this expression into Eq. (\[eq:CRtrEq\]), we obtain that the charged particle acceleration at the fronts of collisionless shock waves propagating in a turbulent plasma is described by the equation (see Ref. [@ACR] and references therein) $$\partial _{X}(u(X)f-D\partial _{X}f)+\frac{1}{p^{2}}\partial
_{p}(p^{2}K(p)f)=$$$$-2\frac{u_{2}}{3(\kappa +1)}\delta (X)\frac{1}{p^{2}}\partial _{p}(p^{3}f).$$In the limit, when the energy losses are negligibly small, this equation has a solution, which gives a power law dependence of the distribution function, $f\propto p^{-k}$ with the index value $k=3u_{2}/(u_{2}-u_{1})$. For $\kappa
=5/3$ the index equals $k=4$, i.e. $f\propto p^{-4}$, or the energy spectrum $d\mathcal{N}_{CR}(\mathcal{E})/d\mathcal{E}\propto \mathcal{E}^{-3}$ is close to the power law index observed in the galactic cosmic ray energy spectrum (see Eq. (\[eq:CRspectr\])).
For the cosmic ray electron component in the high energy limit, at the energy when we cannot neglect the Compton and synchrotron losses, there is a cut off in the spectrum [@SVBVAD]. For typical parameters in supernova remnants, $D=10^{25}$cm$^{2}$s$^{-1}$, $B=10^{-4}$G, $u_{1}=10^{8}$ cm s$%
^{-1}$, the radiation losses limit the energy of ultrarelativistic electrons by values of the order of 10 TeV.
Under the conditions of typical timescale of the laser plasmas the synchrotron losses of ultrarelativistic electrons interacting with the self-generated magnetic field is of the order of $$\tau _{B}=5\left( \frac{10^{3}}{\gamma _{e}}\right) \left( \frac{10^{9}G}{B}%
\right) ^{2}\ fs.$$
Conclusions
===========
Finally, we note that the development of superintense lasers with parameters in the ELI range will provide the necessary conditions for experimental physics where it will become possible to study ultrarelativistic energy of accelerated charged particles, super high intensity EMW and the relativistic plasma dynamics. A fundamental property of the plasma to create nonlinear coherent structures, such as relativistic solitons and vortices, collisionless shock waves and high energy particle beams, and to provide the conditions for relativistic regimes of the magnetic field line reconnection, makes the area of relativistic laser plasmas attractive for modeling of processes of key importance for relativistic astrophysics.
Acknowledgement {#acknowledgement .unnumbered}
===============
We appreciate discussions and comments from V. S. Beskin, M. Borghesi, P. Chen, R. Diehl, A. Ya. Faenov, M. Kando, Y. Kato, T. Kawachi, J. K. Koga, K. Kondo, G. Korn, G. Mourou, N. B. Narozhny, T. A. Pikuz, A. S. Pirozhkov, N. N. Rosanov, V. I. Telnov, A. G. Zhidkov. This work was partially supported by the Ministry of Education, Science, Sports and Culture of Japan, Grant-in-Aid for Creative Scientific Research (A), 202244065, 2008.
The authors acknowledge the support by the European Commission under contract ELI pp 212105 in the framework of the program FP7 Infrastructures-2007-1.
[999]{} M. Shriber, Science **310** (2005) 1610; M. Dunne, Nature Phys. **2** (2006) 2; J. Shambaret, *et al.,* Quantum Electronics and Laser Sci. Conf. QETS’07 (2007) 2.
B. Remington, D. Arnett, P. Drake, and H. Takabe, Science **248** (1999) 1488; P. Chen, AAPPS Bull. **13** (2003) 3 \[ArXiv:astro-ph/03003350\]; N. C. Woolsey, C. Courtois, and R. O. Dendy, Plasma Phys. Control. Fusion **46** (2004) B397; B. Remington, R. Drake, D. Ryutov, Rev. Mod. Phys. **78** (2006) 755.
J. D. Lindl, *et al.,* Phys. Plasmas **11** (2004) 339.
A. D. Strickland and G. Mourou, Opt. Commun. **56** (1985) 212.
S. Weinberg, arXiv:hep-th/0511037 v1.
A. Linde, *Particle Physics and Inflationary Cosmology* (Harwood, Chur, Switzerland, 1990).
M. Yu. Khlopov and S. G. Rubin, *Cosmological Pattern of Microphysics in the Inflationary Universe* (Kluwer, New York, 2004).
D. H. Perkins, *Particle Astrophysics* (Oxford University Press, Oxford, 2003).
F. Wilczek, International J. Mod. Phys. A, **23** (2008) 1791.
N. Kaloper and J. Terning, arxiv.org/abs/physics/0703062v2; M. Bleicher, arxiv.org/abs/physics/0703062v2.
I. Ya. Aref’eva and I. V. Volovich, arxiv.org/abs/0710.2696v2.
J. Rafelski, International J. Mod. Phys. A, **16** (2007) 813.
G. Mourou, T. Tajima, and S. V. Bulanov, Rev. Mod. Phys. **78** (2006) 309.
V. S. Berezinskii, S. V. Bulanov, V. L. Ginzburg, V. A. Dogiel, V. S. Ptuskin, *Astrophysics of cosmic rays* (North Holland Publ. Co. Elsevier Sci. Publ. Amsterdam, 1990).
F. A. Aharonian, *et al.*, Phys. Rev. D **66** (2002) 023005; M. V. Medvedev, Phys. Rev. E **67** (2003) 045401 (R).
T. Tajima and J.M. Dawson, Phys. Rev. Lett. **43** (1979) 267.
P. Chen, J. M. Dawson, R. Huff, and T. Katsouleas, Phys. Rev. Lett. **54** (1985) 693.
P. Chen, T. Tajima, and Y. Takahashi, Phys. Rev. Lett. **89** (2002) 161101; M. Hoshino, Astrophys. J. **672** (2008) 940; J. T. Frederiksen, Astrophys. J. **680** (2008) L5; P. Chen, [*et al.,*]{} Plasma Phys. Control. Fusion [**51**]{} (2009) 024012.
T. Zh. Esirkepov, *et al.,* Phys. Rev. Lett. **92** (2004) 175003.
S. V. Bulanov, T. Zh. Esirkepov, J. Koga, and T. Tajima, Plasma Phys. Rep. **30** (2004) 196.
P. N. Lebedev, Ann. der Physik, **6** (1901) 433; A. S. Eddington, Mon. Not. Roy. Astr. Soc. **85** (1925) 408.
E. A. Milne, Mon. Not. Roy. Astr. Soc. **86** (1926) 459; S. Chandrasekhar, Mon. Not. Roy. Astr. Soc. **94** (1934) 522; N. J. Shaviv, [Astrophys. J.]{} **532** (2000) L137.
J. Arons, [Astrophys. J.]{} **388** (1992) 561; C. F. Gammie, Mon. Not. Roy. Astr. Soc. **297** (1998) 929; M. C. Begelman, [Astrophys. J.]{} **551** (2001) 897.
V. I. Veksler, Atomic Energy **2** (1957) 427.
S. V. Bulanov, *et al.,* Nucl. Instrum. Methods Phys. Res. A **540** (2005) 25.
S. S. Bulanov, *et al.*, [Phys. Rev.]{} E** 78**, (2008) 026412.
T. V. Liseykina, M. Borghesi, A. Macchi, and S. Tuveri, Plasma Phys. Control. Fusion **50** (2008) 124033; A. P. L. Robinson, *et al.*, [New J. Phys.]{} **10** (2008) 013021.
P. Goldreich, [Phys. Scripta]{} **17** (1978) 225; T. Piran, [Astrophys. J.]{} **257** (1982) L23.
S. V. Bulanov, T. Zh. Esirkepov, and T. Tajima, Phys. Rev. Lett. **91** (2003) 085001.
M. Bordag, U. Mohideen, and V.M. Mostepanenko, Phys. Rep. **353** (2001) 1; A. M. Fedotov, Yu. E. Lozovik, N. B. Narozhny, and A. N. Petrosyan, Phys. Rev. A **74** (2006) 013806.
W. G. Unruh, Phys. Rev. D **14** (1976) 870; P. Chen and T. Tajima, Phys. Rev. Lett. **83** (1999) 256; R. Schutzhold, G. Schaller, and D. Habs, Phys. Rev. Lett. **97** (2006) 121302; L. Crispino, *et al.*, Rev. Mod. Phys. **80** (2008) 787.
N. N. Rosanov, JETP **76** (1993) 991; N. N. Rosanov, JETP **86** (1998) 284.
J. J. Klein and B. P. Nigam, Phys. Rev. **135**, B1279 (1964); J. Lundin, *et al.,* Phys.Rev. A **74** (2006) 043821; D. Tommasini, *et al.*, arXiv:0802.0101v1 \[physics.optics\].
A. M. Fedotov and N. B. Narozhny, Physics Letters A **362** (2007) 1.
A. Di Piazza, A. I. Milstein, and C. H. Keitel Phys. Rev. A **76** (2007) 032103; A. Di Piazza, K. Z. Hatsagortsyan, and C. H. Keitel, Phys.Rev. A **78** (2008) 062109.
M. Marklund and P. Shukla, Rev. Mod. Phys. **78** (2006) 591.
Y. I. Salamin, S. X. Hu, K. Z. Hatsagortsyan, and C. H. Keitel, Phys. Reports **427** (2006) 41.
J. Schwinger, Phys. Rev. **82** (1951) 664.
V. B. Beresteskii, E. M. Lifshitz, and L. P. Pitaevskii, *Quantum Electrodynamics* (Pergamon, New York, 1982).
K. Birkeland, *The Norwegian Aurora Polaris Expedition, 1902-1903* (Aschenhoug & Co. Christiania, 1908).
H. Alfven and C.-G. Falthammar, *Cosmic Electrodynamics* (Oxford, 1963); T. Tajima and K. Shibata *Plasma Astrophysics* (Addison-Wesley, Reading, 1997).
V. E. Fortov, Physics - Uspekhi **50** (2007) 347.
P. J. Baum and A. Bratenahl, Adv. Electronics and Electron Phys. **54** (1980) 1; S. V. Bulanov, I. Ya. Butov, Yu. S. Gvaladze, A. M. Zaborov, A. N. Kuzyutin, M. A. Ol’shanetskij, R. G. Salukvadze, B. D. Tsurtsumija, Sov. J. Plasma Phys. **12**, 180 (1986); S. V. Bulanov and A. G. Frank, Sov. J. Plasma Phys. **18** (1992) 1535; R. L. Stenzel, *et al.*, Phys. Plasmas **9** (2002) 1925; J. Egedal, *et al.*, Phys. Rev. Lett. **90** (2003) 135003; A. G. Frank, *et al.*, Phys. Plasmas **12** (2005) 052316; M. Yamada, *et al.,* Phys. Plasmas **14** (2007) 058102.
D. W. Koopman and D. A. Tidman, Phys. Rev. Lett. **18** (1967) 533; A. R. Bell *et al.*, Phys. Rev. A **38** (1988) 1363; N. C. Woolsey, *et al.*, Phys. Plasmas **8** (2001) 2439.
Yu. P. Zakharov, *et al.*, Journal of Physics: Conf. Series **112** (2008) 042011.
L. I. Sedov, *Similarity and dimensional methods in mechanics* (New York, Academic Press, 1959); G. I. Barenblatt, *Scaling, Self-similarity, and Intemediate Asymptotics: Dimensional Analysis and Intermediate Asymptotics* (Cambridge, Cambridge University Press, 1996).
I. M. Podgorny and R. Z. Sagdeev, Sov. Phys. Uspekhi **98** (1970) 445.
C.-G. Falthammar, Space Sci. Rev. **15** (1974) 803.
S. V. Bulanov, Plasma Phys. Control. Fusion **48** (2006) B29.
F. Pegoraro, T. Zh. Esirkepov, and S. V. Bulanov, Physics Letters A **347** (2005) 133.
L. V. Keldysh, Sov. Phys. JETP **20** (1965) 1307.
P. Corkum, Phys. Rev. Lett. **71** (1993) 1994.
L. D. Landau and E. M. Lifshitz, *The Classical Theory of Fields* (Pergamon Press, Oxford, 1980).
A. I. Akhiezer and R. V. Polovin, Sov Phys. JETP **30** (1956) 915.
S.-W. Bahk, *et al.,* Opt. Lett. **29** (2004) 2837; V. Yanovsky, *et al.,* Opt. Express **16** (2008) 2109.
G. S. Bisnovaty-Kogan, Ya. B. Zel’dovich, and R. A. Syunyaev, Sov. Astron. J. 15 (1971) 17; S. Wilks,* et al.,* Astrophys. Space. Sci. **298** (2005) 347.
I. Kuznetsova, D. Habs, and J. Rafelski, Phys. Rev. D. **78** (2008) 014027.
C. Gahn, *et al.,* Appl. Phys. Lett. **77** (2000) 2662.
Ya. B. Zel’dovich and A. F. Illarionov, Sov. Phys. JETP **34** (1972) 467; A. D. Steiger and C. H. Woods, Phys. Rev. D **5** (1972) 2912; Ya. B. Zel’dovich, Sov. Phys. Uspekhi **18** (1975) 79; C H Keitel, *et al.,* J. Phys. B: At. Mol. Opt. Phys. **31** (1998) L75; A. G. Zhidkov, *et al.,* Phys. Rev. Lett. **88** (2002) 185002; J. K. Koga, T. Zh. Esirkepov, and S. V. Bulanov, Phys. Plasmas **12** (2005) 093106; J. Koga, T. Z. Esirkepov, and S.V. Bulanov, J. Plasma Phys. **72** (2006) 1315; V. I. Berezhiani, S. Mahajan, and Z. Yoshida, Phys. Rev. E. **78** (2008) 066403; A. Di Piazza, Lett. Math. Phys. **83** (2008) 305.
A. R. Bell and J. G. Kirk, Phys. Rev. Lett. **101** (2008) 200403.
W. Dittrich and H. Gies, *Probing the Quantum Vacuum: Perturbative Effective Action Approach in Quantum Electrodynamics and Its Applications* (Springer, Berlin, 2000).
V. S. Popov, Phys. At. Nuclei **68**, 686 (2005).
N. B. Narozhnyi, S. S. Bulanov, V. D. Mur, and V. S. Popov, JETP Lett. **80** (2004) 382.
S. S. Bulanov, A. M. Fedotov, and F. Pegoraro, Phys. Rev. **71** (2005) 016404; R. Ruffini, *et al.,* Phys. Lett. A **371** (2007) 399.
N. B. Narozhny, *et al.,* Phys. Lett. A **330** (2004) 1.
C. Bamber, *et al.,* Phys. Rev. D **60** (1999) 092004.
E. Loetstedt, *et al.,* Phys. Rev. Lett. **101** (2008) 203001.
V. S. Beskin, A. V. Gurevich, and Ya. N. Istomin, *Physics of the Pulsar Magnetosphere* (Cambridge Univ. Press, Cambridge, 1993).
A. Gruzinov, Phys. Rev. Lett. **94** (2005) 021101.
J. Gunn and J. Ostriker, Phys. Rev. Lett. **22** (1969) 728.
J. Hester, Ann. Rev. Astron. Astrophys. **46** (2008) 127.
D. Clowe, *et al.,* Astrophys. J. **648** (2006) L109.
S. Humphries, Jr., *Charged Particle Beams* (Wiley, New York, 1990).
T. Zh. Esirkepov, Y. Kato, and S. V. Bulanov, Phys. Rev. Lett. **101** (2008).
M. Kando, *et al.,* in: 1st International Symposium Laser-Driven Relativistic Plasmas Applied for Science, Industry, and Medicine// eds. S. V. Bulanov, H. Daido. AIP Conf. Proc. **1024** (2008) 197.
S. V. Bulanov, *et al.*, Phys. Plasmas **12** (2005) 073103.
S. V. Bulanov and T. Tajima, J. Particle Accelerator Society of Japan **2** (2005) 35; T. Zh. Esirkepov, *et al.,* Phys. Rev. Lett. **96** (2006) 014803.
M. Kando, *et al.*, JETP **105** (2007) 916.
P. Michel, *et al.*, Phys. Rev. E **74** (2006) 026501; I. Y. Dodin and N. J. Fisch, Phys. Plasmas **15** (2008) 103105.
M. Borghesi, *et al.*, Fus. Sci. Technology, **49** (2006) 412; M. Borghesi, *et al.*, Plasma Phys. Control. Fusion, **50** (2008) 124040.
S. V. Bulanov and V. S. Khoroshkov, Plasma Phys. Rep. **28** (2002) 453.
T. Zh. Esirkepov, *et al.*, Phys. Rev. Lett. **89** (2002) 175003; S. V. Bulanov, *et al.*, Plasma Phys. Rep. **28** (2002) 975; A.P.L Robinson and P. Gibbon, Phys. Rev. E **75** (2007) 015401(R); T. Morita, [*et al.,*]{} Phys. Rev. Lett. **100** (2008) 145001.
C. Schwoerer, *et al.*, Nature **439** (2006) 445.
A. V. Gurevich, *et al.*, Sov. Phys. JETP **22** (1966) 449; E. G. Gamalii and R. Dragila, Kratk. Soobshch. Fiz. AN SSSR **2** (1988) 16; S. V. Bulanov, L. M. Kovrizhnykh, and A. S. Sakharov, Phys. Rep. **186** (1990) 1; P. Mora, Phys. Rev. Lett. **90** (2003) 185002.
Y. Kishimoto, K. Mima, T. Watanabe, and K. Nishikawa, Phys. Fluids **26** (1983) 2308; F. Mako and T. Tajima, Phys. Fluids **27** (1984) 1815.
S. V. Bulanov, *et al.,* Plasma Phys. Rep. **30** (2004) 21.
F. Pegoraro and S.V. Bulanov, Phys. Rev. Lett. **99** (2007) 065002.
E. Ott, Phys. Rev. Lett. **29** (1972) 1429; W. Manheimer, *et al.,* Phys. Fluids **27** (1984) 2164; F. Pegoraro, *et al.,* Phys. Rev. E **64** (2001) 016415.
S. Kar, *et al.,* Phys. Rev. Lett. **100** (2008) 225004; M. Borghesi, *et al.,* Plasma Phys. Control. Fusion **50** (2008) 124040.
C. W. Misner, K. Thorn, J. A. Wheeler, *Gravitation* (Freeman, San Francisco, 1973); S. Weinberg, *Gravitation and Cosmology: Principles and Applications of the General Theory of Relativity* (Wiley, New York, 1973)
N. Arkani-Hamed, S. Dimopoulos, and G. Dvali, Phys. Lett. B **429** (1998) 263.
F. R. Arutyunian and V. A. Tumanian, Phys. Lett. **4** (1963) 176; Y. Li*, et. al.*, Phys. Rev. ST Accel. Beams **5** (2002) 044701.
N. M. Naumova, J. A. Nees, I. V. Sokolov, B. Hou, and G. A. Mourou, Phys. Rev. Lett. **92** (2004) 063902; N. M. Naumova, J. Nees, and G. Mourou, Phys. Plasmas **12** (2005) 056707; A. Isanin, et al., Phys. Lett. A **337** (2005) 107; S. S. Bulanov, T. Z. Esirkepov, F. F. Kamenets, and F. Pegoraro, Phys. Rev. E **73** (2006) 036408; N. M. Naumova*, et al.,* New J. Phys. **10** (2008) 025022; V. V. Kulagin*, et al.,* Phys. Plasmas **14** (2007) 113101; N. N. Rosanov, JETP Lett. **88** (2008) 577; D. Habs, M. Hegelish, *et al.,* Appl Phys B **93** (2008) 349354; H.- C. Wu and J. Meyer-ter-Vehn, arXiv:0812.0703v1 \[physics.plasm-ph\] 3 Dec 2008; J. Meyer-ter-Vehn and H. - C. Wu, arXiv:0812.0710v1 \[physics.plasm-ph\] 3 Dec 2008; T. Zh. Esirkepov, et al., arXiv:0812.0401v1 \[physics.plasm-ph\] 2 Dec 2008.
K. Landecker, Phys. Rev. **86** (1952) 852; V. L. Granatstein et al., Phys. Rev. A **14** (1976) 1194; J. A. Pasour, V. L. Granatstein, and R. K. Parker, ibid. **16** (1977) 2441.
V. I. Semenova, Sov. Radiophys. Quantum Electron. **10** (1967) 599; W. B. Mori, Phys. Rev. A **44** (1991) 5118; R. L. Savage, Jr.*, et al.,* Phys. Rev. Lett. **68** (1992) 946.
A. Einstein, Ann. Phys. (Leipzig) **17** (1905) 891; W. Pauli, *Theory of Relativity* (Pergamon, New York, 1958).
A. V. Panchenko, *et al.*, Phys. Rev. E **78** (2008) 056402.
S. V. Bulanov and A. S. Sakharov, JETP Lett. **54** (1991) 203; S. V. Bulanov, F. Pegoraro, and A. M. Pukhov, Phys. Rev. Lett. **74** (1995) 710; N. H. Matlis, *et al.*, Nature Phys. **2** (2006) 749.
M. Kando, *et al.,* Phys. Rev. Lett. **99** (2007) 135001; A.S. Pirozhkov, *et al.,* Phys. Plasmas **14** (2007) 080904; M. Kando, *et al.,* :0705.0872v1 \[physics.plasm-ph\] 2007.
W. Heisenberg and H. Euler, Z. Phys. **98** (1936) 714.
S. V. Bulanov and P. V. Sasorov, Soviet Astronomy J. **19** (1976) 464; S. V. Bulanov and F. Cap, Soviet Astronomy J. **32** (1988) 436; J. Buchner and L. M. Zelenyi, J. Geophys. Res. **94** (1989) 11821; D. L.Vainshtein , L. M. Zelenyi , and A. I*.*Neishtadt, Plasma Physics Reports, **21** (1995) 484; J. F. Drake, M. Swisdak, H. Che and M. A. Shay, Nature **443** (2006) 553; S. Zenitani and M. Hoshino, Astrophys. J. **670** (2007) 702 \[arXiv:0708.1000v2\]; Y. Lyubarsky and M. Liverts, Astrophys. J. **682** (2008)1436.
S. I. Syrovatskii, Ann. Rev. Astron. Astrophys. **19** (1981) 163; 1; S. I. Syrovatskii, S. V. Bulanov, V. A. Dogiel, in: [*Scientific Reviews E, Astrophysics and Space Physics Reviews*]{}, vol. 2 (North Holl. Publ. Co, Amsterdam, 1983) p. 385
D. Biskamp,* Magnetic Reconnection in Plasmas* (Cambridge Univ. Press, Cambridge, 2000).
G. A. Askar’yan, *et al.,* Comm. Plasma Phys. Controlled Fusion **17** (1995) 35.
P. M. Nilson,* et al.,* Phys. Rev. Lett. **97** (2006) 255001; C. K. Li, *et al.,* Phys. Rev. Lett. **99** (2007) 055001; P. M. Nilson, *et al.,* Phys. Plasmas **15** (2008) 092701.
P. A. Sweet, in *Electromagnetic Phenomena in Cosmic Physics./ Ed. by B. Lehnert.* (Cambridge University Press, 1958) p. 122.
E. Parker, *Cosmical Magnetic Fields: their origin and their activity* (Oxford University Press, New York, 1979).
S. V. Bulanov, *et al.,* Phys. Lett. A **203** (1995) 219.
S. V. Bulanov, *et al.,* Sov. Astronomy. Lett. **11** (1985) 159.
S. V. Bulanov, *et al.,* in: *Reviews of Plasma Physics*, vol. 22 // edited by V. D. Shafranov, (Kluwer Academic / Plenum Publishers, New York, 2001), p. 227.
S. V. Bulanov, M. Lontano, T. Zh. Esirkepov, F. Pegoraro, and A. M. Pukhov, Phys. Rev. Lett. **76** (1996) 3562.
M. Borghesi, *et al.,* Phys. Rev Lett. **80** (1998) 5137; M. Tatarakis, *, et al.,* Nature (London) **415** (2002) 280..
G. K. Batchelor, *Introduction to Fluid Mechanics* (Cambridge University Press, Cambridge, 1973).
A. Hasegawa and K. Mima, Phys. Rev. Lett. **39** (1977) 205.
H. Lamb, *Hydrodynamics* (Cambridge University Press, Cambridge, 1932).
A. S. Kingsep, K. V. Chukbar, and V. V. Yan’kov, in: *Reviews of Plasma Physics*, vol. 16 // Ed. by B. B. Kadomtsev (Consultants Bureau, New York, 1990) p. 243.
S. V. Bulanov, F. Pegoraro, and A. S. Sakharov, Phys. Fluids B **4** (1992) 2499; K. Avinash, et al., Phys. Plasmas **5** (1998) 2946; P. A. Cassak*, et al.,* Phys. Rev. Lett. **98** (2007) 215001.
H. P. Furth, I. Killen, and M. N. Rosenbluth, Phys. Fluids **6** (1963) 459; B. Coppi, G. Laval, and R. Pellat, Phys. Rev. Lett. **16** (1966) 1207.
S. V. Basova,* et al.,* Sov. J. Plasma Phys. **17** (1991) 615.
A. Fruchtman and H. R. Strauss, Phys. Fluids B **5** (1993) 1408; M. Hosseinpour, [*et al.,*]{} Phys. Plasmas [**16**]{} (2009) 012104.
J. W. Dungey, Phil. Mag. Ser. 7. **44** (1953) 725.
E. S. Weibel, Phys. Rev. Lett **2** (1959) 83.
V. Yu. Bychenkov, *et al.,* Sov. Phys. JETP **71** (1990) 709; G. A. Askar’yan, *et al.,* JETP Lett. **60** (1994) 251.
B. M. Gaensler, *et al.,* Astrophys. J. **616** (2004) 383.
F. Califano, F. Pegoraro, and S. V. Bulanov, Phys. Rev. E **56** (1997) 963.
M. Honda, J. Meyer-ter-Vehn, and A. M. Pukhov, Phys. Plasmas **7** (2000) 1302.
J. I. Sakai, *et al.,* Phys. Plasmas **9** (2002) 2959.
Y. Kazimura, J.-I. Sakai, T. Neubert, and S. V. Bulanov, Astrophys. J. **498** (1998) L183.
M. V. Medvedev and A. Loeb, Astrophys. J. **526** (1999) 697.
J. I. Sakai, R. Schlickeiser, and P. Shukla, Phys. Lett. A **330** (2004) 384.
D. A. Tidman and N. A. Krall, *Shock Waves in Collisionless Plasmas* (Wiley-Interscience, New York, 1971); T. M. O’Neil and F. V. Coroniti, Rev. Mod. Phys. **71** (1999) 404.
L. O. Silva, *et al.,* Astrophys. J. **596** (2003) L121.
H. Takabe, *et al.,* Plasma Phys. Control. Fusion **50** (2008) 124057.
V. L. Ginzburg and V. V. Zheleznyakov, Ann. Rev. Astron. Astrophys. **13** (1975) 511.
V. L. Ginzburg, *Applications of Electrodynamics in Theoretical Physics and Astrophysics* (Gordon and Breach, New York, 1989).
T. Zh. Esirkepov, *et al.,* Phys. Rev. Lett. **92** (2004) 255001.
S. V. Bulanov, *et al.,* Phys. Fluids B **4** (1992) 1935; S. V. Bulanov*, et al.,* Phys. Rev. Lett. **82** (1999) 3440.
V. A. Kozlov, A. G. Litvak, and E. V. Suvorov, Sov. Phys. JETP **49** (1979) 75; P. K. Kaw, A. Sen, and T. Katsouleas, Phys. Rev. Lett. **68** (1992) 3172; T. Zh. Esirkepov,* et al.,* JETP Lett. **68** (1998) 36; D. Farina and S. V. Bulanov, Phys. Rev. Lett. **86** (2001) 5289.
T. Esirkepov, *et al.,* Phys. Rev. Lett.** 89** (2002) 275002.
N. M. Naumova, *et al.,* Phys. Rev. Lett. **87** (2001) 185004.
M. Borghesi, *et al.,* Phys. Rev. Lett. **88** (2002) 135002.
M. Kachelries, arXiv:0801.4376v1 \[astro-ph\] 29 Jan 2008.
K. Greisen, Phys. Rev. Lett. **16** (1966) 748; G. T. Zatsepin and V. A. Kuz’min, Sov. Phys. JETP. Lett. **4** (1966) 78.
J. Abraham (Pier Auger Collaboration) Science **318** (2007) 938.
L. Romagnani, *et al.,* Phys. Rev. Lett. **101** (2008) 025004.
M. Borghesi, *et al.,* Phys. Plasmas **9** (2002) 2214.
G. F. Krymskii, Sov. Phys.–Dokl. **23** (1977); W. I. Axford, E. Leer and G. Skadron, Proc. 15th Int. Cosmic Ray Conf. (Plovdiv) **11** (1977) 132; A. R. Bell, Mon. Not. Roy. Astr. Soc. **182** (1978) 147; R. D. Blandford and J. P. Ostriker, Astrophys. J. **221** (1978) 29; S. V. Bulanov and I. V. Sokolov, Soviet Astronomy J. **28** (1984) 515; M. A. Malkov and L. O. Drury, Rep. Prog. Phys. **64** (2001) 429.
D. Burrows, *et al.,* Astrophys. J. **543** (2000) L149.
V. S. Imshennik and D. K. Nadezhin, Usp. Fiz. Nauk, **158** (1988) 561; V. S. Imshennik and D. K. Nadezhin, in: [ *Sov. Sci. Rev. E Astrophysics and Space Physics Reviews*]{}, vol. 8 (North Holl. Publ. Co, Amsterdam, 1983) p. 1.
A. Lichtenberg and M. Liberman, *Regular and stochastic dynamics* (Wiley, New York, 1984).
S. V. Bulanov, *Introduction to Nonlinear Physics* (Scuola Normale Superiore, Pisa, 2000).
E. Fermi, Phys. Rev. **75** (1949) 1169; Astrophys. J. **119** (1954) 1.
S. V. Bulanov and V. A. Dogiel, Sov. Astron. J. Lett. **5** (1979) 521.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We establish the future stability of nonlinear perturbations of a class of homogeneous solutions to the relativistic Euler equations with a linear equation of state $p=K\rho$ on exponentially expanding FLRW spacetimes for the equation of state parameter values $1/3 < K < 1/2$.'
address: |
School of Mathematics\
9 Rainforest Walk\
Monash University, VIC 3800\
Australia
author:
- 'Todd A. Oliynyk'
bibliography:
- 'Kgtot\_v4.bib'
title: 'Future global stability for relativistic perfect fluids with linear equations of state $p=K\rho$ where $1/3<K<1/2$'
---
Introduction\[intro\]
=====================
Relativistic perfect fluids on a prescribed spacetime $(M,\gt)$ are governed by the relativistic Euler equations given by[^1] $$\nablat_i \Tt^{ij}=0 \label{relEulA}$$ where $$\Tt^{ij} = (\rho+p)\vt^i \vt^j + p \gt^{ij}$$ is the stress energy tensor, $\rho$ is the fluid proper energy density, $p$ is the fluid pressure, and $\vt^{i}$ is the fluid four-velocity normalized by $\gt_{ij}\vt^i \vb^j=-1$. In this article, we will be interested in analyzing the relativistic Euler equations on exponentially expanding Friedmann-Lemaître-Robertson-Walker (FLRW) spacetime of the form $(M,\gt)$ where $$M = (0,1]\times \Tbb^3$$ and[^2] $$\label{FLRW}
\gt = \frac{1}{t^2} g$$ with $$\label{conformal}
g = -dt\otimes dt + \delta_{IJ}dx^I \otimes dx^J.$$ It is important to note that, due to our conventions, the future is located in the direction of *decreasing* $t$ and future timelike infinity is located at $t=0$. Consequently, we require that $$\vt^0 < 0$$ in order to ensure that the four-velocity is future directed.
The future stability of nonlinear perturbations of homogeneous solutions to the relativistic Euler equations with a linear equation of state $$p = K \rho$$ on exponentially expanding FLRW spacetimes has been well studied for the parameter range $$0\leq K\leq \frac{1}{3}.$$ The first such stability result was, building on the earlier stability results for the Einstein-scalar field system [@Ringstrom:2008], established[^3] for the parameter values $0<K<1/3$ in the articles [@RodnianskiSpeck:2013; @Speck:2012]. Stability results for the end points $K=1/3$ and $K=0$ were established later in [@LubbeKroon:2013] and [@HadzicSpeck:2015], respectively. See also [@Friedrich:2017; @LiuOliynyk:2018b; @LiuOliynyk:2018a; @Oliynyk:CMP_2016] for different proofs and perspectives, the articles [@LeFlochWei:2015; @LiuWei:2019] for related stability results for fluids with nonlinear equations of state, and the articles [@FOW:2020; @Ringstrom:2009; @Speck:2013; @Wei:2018] for stability results on other expanding FLRW spacetimes (e.g. power law expansion). The importance of all of these works is that they demonstrate spacetime expansion can suppress shock formation in fluids, which was first discovered in the Newtonian cosmological setting with $K=0$ by [@BrauerRendallReula:1994]. This should be compared to the work of [@Christodoulou:2007] where it is established that arbitrary small perturbations of a class of homogeneous solutions to the relativistic Euler equations, for relatively general equations of state, on Minkowski spacetime, which is a FLRW spacetime with spatial manifold $\Rbb^3$ and no expansion, form shocks in finite time.
For linear equations of states, the parameter $K$ determines the square of the sound speed, and consequently, it is natural to assume[^4] that $K$ satisfies $$\label{KrangeC}
0\leq K \leq 1$$ so that the propagation speed for the fluid is less than or equal to the speed of light. When the sound speed is equal to the speed of light, that is $K=1$, it is well known that the irrotational relativistic Euler equations coincide, under a change of variables, with the linear wave equation. In this case, the future global existence of solutions on exponentially expanding FLRW spacetimes can be inferred from standard existence results for linear wave equations; we note also that similar results can be established for fluids with rotation. This leaves us to consider the parameter range $$\label{KrangeA}
\frac{1}{3}<K<1,$$ which we will assume holds for the remainder of the article.
The asymptotic behavior of relativistic fluids on exponentially expanding FLRW spacetimes with a linear equation of state for $K$ satisfying was investigated in the article [@Rendall:2004] by Rendall using formal expansions. In that article, Rendall observed that the formal expansions can become inconsistent for $K$ in the range if the leading order term in the expansion of the four-velocity vanishes somewhere. In that case, he speculated that inconsistent behavior in the expansions could be due to inhomogeneous features developing in the fluid density that would lead to the density contrast blowing up. This possibility for instability in solutions to the relativistic Euler equations for the parameter range was also commented on by Speck in [@Speck:2013 §1.2.3]. There, Speck presents a heuristic analysis that suggest uninhibited growth should set in for solutions of the relativistic Euler equations for the parameter values . These speculations leave the existence of future global solutions to the relativistic Euler equations in doubt for $K$ satisfying .
In this article, we rule out, under a small initial data hypothesis, the possibility of any pathologies developing in finite time for $K$ satisfying $$\frac{1}{3}<K < \frac{1}{2}$$ by establishing, for these parameter values, the future stability of nonlinear perturbations of a class of homogeneous solutions, see , to the relativistic Euler equations on exponentially expanding FLRW spacetimes. For a precise statement of our stability result, see Theorem \[mainthm\], which is the main result of this article. This, of course, leaves open the possibility of finite time blow-up for $K$ satisfying $1/2 < K < 1$. As a first step towards understanding the behavior of solutions in this regime, we establish in Theorem \[symthm\] the future stability of $\Tbb^2$-symmetric nonlinear perturbations of the same class of homogenous solutions for the full parameter range $1/3 < K < 1$. Here, the stability proof relies heavy on the $\Tbb^2$ symmetry that allows us to reduce the relativistic Euler equations to an essentially regular $1+1$ dimensional problem. It is unclear at the moment if one should expect that this result will still hold for $K$ satisfying $1/2 \leq K < 1$ if the $\Tbb^2$-symmetry assumption is removed. We plan to revisit this interesting question in a separate article.
The proof of our main stability result, Theorem \[mainthm\], is based on the Fuschsian method for establishing the global existence of solutions to systems of hyperbolic equations that was first employed in [@Oliynyk:CMP_2016] and further developed in the articles [@BOOS:2019; @FOW:2020; @LiuOliynyk:2018b; @LiuOliynyk:2018a]. This method relies on transforming the global existence problem for a given hyperbolic system into an existence problem for a Fuchsian symmetric hyperbolic equation of the form $$\Asc^0(t,\Wsc)\del{t}\Wsc+ \Asc^i(t,\Wsc)\del{i}\Wsc = \frac{1}{t}\Af(t,\Wsc)\Pbb \Wsc + \Fsc(t,\Wsc)$$ on a finite time interval $(0,T_0]$. Once in this form, the existence of solutions on the time interval $(0,T_0]$ can be deduced, under a suitable smallness assumption on the initial data specified at time $t=T_0$, from general existence theorems for such Fuchsian systems that have been established in the articles [@BOOS:2019; @FOW:2020; @LiuOliynyk:2018b; @LiuOliynyk:2018a].
In this article, we transform the relativistic Euler equations into a suitable Fuchsian form in a number of steps. We start in Section \[relEulsec\] with a formulation, see , of the relativistic Euler equations that was first employed in [@Oliynyk:CMP_2015]. We then modify this system by introducing a new density variable defined by , which results in the system . In Section \[Ftrans\], we proceed by decomposing the conformal three-velocity $v_I$ into its length determined by the variables $u$, $w_1$ and into a normalized vector determined by the variables $w_2$, $w_3$; see - for the relevant formulas. Here, $u$ only depends on $t$ and is used to parameterize a class of homogeneous solutions of relativistic Euler equations. After some straightforward, but lengthy calculations, we obtain two equivalent versions of Euler equations, now expressed in terms of the new variables $u$, $w_1$, $w_2$ and $w_3$, given by and . We then use the second version , see Section \[Homsec\], to identify the ODE satisfied by $u$ that determines homogeneous solutions of the relativistic Euler equations. The existence of solutions to this ODE is established in Proposition \[Homprop\]. The transformation of the relativistic Euler equations into a suitable Fuchsian form is then completed in Section \[Fform\], see , while the coefficients of are analyzed in Section \[coeff\] in order to verify that this system satisfies the required properties in order to apply the existence theory from [@BOOS:2019]. This existence theory is then applied in the proof of Theorem \[mainthm\], which is given in Section \[stability\], to establish the future stability of nonlinear perturbations of the homogeneous solutions to the relativistic Euler equations from Proposition \[Homprop\]. Finally, in Section \[symsec\], we establish the future stability of $\Tbb^2$-symmetric nonlinear perturbations of the same class of homogenous solutions for the full parameter range $1/3 < K < 1$ by using the $\Tbb^2$ symmetry to reduce the relativistic Euler equations to an essentially regular $1+1$ dimensional problem. The precise statement of the stability result in this setting is given in Theorem \[symthm\].
A symmetric hyperbolic formulation of the Relativistic Euler equations\[relEulsec\]
===================================================================================
The first step in transforming the relativistic Euler equations into a suitable Fuchsian form is to find a symmetric hyperbolic formulation of the relativistic Euler equations. Here, we start with the symmetric hyperbolic formulation derived in [@Oliynyk:CMP_2015 §2.2], see also [@Oliynyk:CMP_2016 §2.2]. This involves introducing the *conformal fluid four-velocity* $v_i$ and the $\zeta$ according to $$\label{cov1}
v_i= \frac{1}{t} g_{ij}\vt^{j} \AND \rho = t^{3(1+K)}\rho_c e^{(1+K)\zeta}, \quad \rho_c\in \Rbb_{>0}.$$ Using these variables, the computations carried out in [@Oliynyk:CMP_2015 §2.2] show that relativistic Euler equations can be cast into the following symmetric hyperbolic form: $$\label{relEulB}
B^k \del{k}V = \frac{1}{t}\Bc \pi V$$ where $$\begin{aligned}
V &= (\zeta, v_J )^{\tr} , \label{Vdef}\\
v_0 & = \sqrt{|v|^2 +1} , \qquad |v|^2 = \delta^{IJ}v_I v_J, \label{v0def}\\
v^i & = \delta^{iJ}v_J - \delta^{i}_0 v_0, \label{viupdef}\\
\Bc &= \frac{-1}{v^0}\begin{pmatrix} 1 & 0 \\ 0 & \frac{1-3K}{v_0}\delta^{JI} \end{pmatrix}, \label{Bcdef}\\
\pi &= \begin{pmatrix} 0 & 0 \\ 0 & \delta_{I}^J \end{pmatrix}, \\
L^k_I &= \delta^k_J - \frac{v_J}{v_0} \delta^k_0, \\
M_{IJ} &= \delta_{IJ} - \frac{1}{(v_0)^2}v_I v_J, \label{Mdef}\\
B^0 &= \begin{pmatrix} K & \frac{K}{v^0} L^0_M \delta^{MJ} \\ \frac{K}{v^0} \delta^{LI} L^0_L & \delta^{LI} M_{LM} \delta^{MJ} \end{pmatrix}
\intertext{and}
B^K &= \frac{1}{v^0}\begin{pmatrix} Kv^K & K L^K_M \delta^{MJ} \\ K \delta^{LI} L^K_L & \delta^{LI} M_{LM} \delta^{MJ} v^K \end{pmatrix}.
\label{BKdef}\end{aligned}$$
Defining a new modified density variable $\zetat$ by $$\label{zetatdef}
\zetat = \zeta + \ln(v_0),$$ we obtain from differentiating the relation $$\label{dV2dVt}
\del{k} V = Q\del{k}\Vt$$ where $$\label{Vtdef}
\Vt = (\zetat, v_J)^{\tr}$$ and $$Q= \begin{pmatrix} 1 & -\frac{1}{v_0^2}v^J \\
0 & \delta^J_I \end{pmatrix}.$$ From - and , we see, after multiplying on the left by $Q^{\tr}$, that the relativistic Euler equations can be expressed in terms of the new variables as $$\label{relEulC}
\Bt^k \del{k}\Vt = \frac{1}{t}\Bc \pi \Vt$$ where $$\begin{aligned}
\Bt^0 &:= Q^{\tr}B^0 Q = \begin{pmatrix} K & 0 \\ 0 & \delta^{IL}M_{LM}\delta^{MJ}-\frac{K}{v_0^4}v^I v^J\end{pmatrix}
\label{Bt0def}
\intertext{and}
\Bt^K &:= Q^{\tr}B^K Q = -\frac{1}{v_0}\begin{pmatrix} K v^K & -\frac{K}{v_0^2}v^J v^K + K\delta^{KJ} \\
-\frac{K}{v_0^2}v^I v^K + K\delta^{KI}& \bigl(\delta^{IL}M_{LM}\delta^{MJ}+\frac{K}{v_0^4}v^I v^J\bigr)v^K -\frac{K}{v_0^2}
(v^I \delta^{JK}+v^J \delta^{IK}) \end{pmatrix}. \label{BtIdef}\end{aligned}$$
Transformation to Fuchsian form\[Ftrans\]
=========================================
We proceed with the transformation of the relativistic Euler equations into a suitable Fuchsian form by defining a second change of variables via $$\begin{aligned}
v_1 &= \frac{t^{-\mu } e^{u(t)+w_1}}{\sqrt{t^{2 \mu
}
\left((w_2-w_3)^2+(w_2+w_3)^2\right)+1}}, \label{cov2a} \\
v_2 &= \frac{(w_2+w_3) e^{u(t)+w_1}}{\sqrt{t^{2
\mu }
\left((w_2-w_3)^2+(w_2+w_3)^2\right)+1}} \label{cov2b}
\intertext{and}
v_3 &= \frac{(w_2-w_3) e^{u(t)+w_1}}{\sqrt{t^{2
\mu }
\left((w_2-w_3)^2+(w_2+w_3)^2\right)+1}}, \label{cov2c}\end{aligned}$$ where $u(t)$ is a time dependent function and $\mu\in \Rbb$ is a constant both of which will be fixed below. Using -, we find from differentiating that $$\begin{aligned}
\del{t} \Vt &= P \del{t}W + Z \label{dVt2dWa}
\intertext{and}
\del{I}\Vt &= P \del{I} W \label{dVt2dWb}\end{aligned}$$ where $$\begin{gathered}
W = (\zetat, w_1, w_2, w_3 )^{\tr} , \label{Wdef}\\
Z = \begin{pmatrix}
0 \\
\frac{t^{-\mu -1} e^{\wbr_1} \left(t u'(t)
\phi-\mu
\left(4 t^{2 \mu }
\left(w_2^2+w_3^2\right)+1\right)\right)}{\phi^{3/2}} \\
\frac{(w_2+w_3) e^{\wbr_1} \left(t
u'(t) \phi-2 \mu
t^{2 \mu }
\left(w_2^2+w_3^2\right)\right)}{t \phi^{3/2}} \\
\frac{(w_2-w_3) e^{\wbr_1} \left(t
u'(t) \phi -2 \mu
t^{2 \mu }
\left(w_2^2+w_3^2\right)\right)}{t \phi^{3/2}}
\end{pmatrix}, \label{Zdef} \\
P=\begin{pmatrix}
1 & 0 & 0 & 0 \\
0 & \frac{t^{-\mu } e^{\wbr_1}}{\sqrt{\phi}} &
-\frac{2 w_2 t^{\mu }
e^{\wbr_1}}{\phi^{3/2}} &
-\frac{2 w_3 t^{\mu }
e^{\wbr_1}}{\phi^{3/2}} \\
0 & \frac{(w_2+w_3)
e^{\wbr_1}}{\sqrt{\phi}} &
-\frac{e^{\wbr_1}\eta_3}{\phi^{3/2}} &
\frac{e^{\wbr_1} \eta_2}{\phi^{3/2}} \\
0 & \frac{(w_2-w_3)
e^{\wbr_1}}{\sqrt{\phi}} &
\frac{e^{\wbr_1}\xi_3}{\phi^{3/2}} &
-\frac{e^{\wbr_1} \xi_2}{\phi^{3/2}}
\end{pmatrix} \label{Pdef}\end{gathered}$$ and we have set $$\begin{aligned}
\wbr_1 &= u+w_1, \label{wbr1def} \\
\phi &= 2 t^{2\mu } \left(w_2^2+w_3^2\right)+1, \label{phidef} \\
\eta_\Lambda &= \left(2 w_\Lambda t^{2 \mu }
(w_2-w_3)+(-1)^\Lambda 1\right), \quad \Lambda =2,3, \label{etadef}
\intertext{and}
\xi_\Lambda &= \left(2 w_\Lambda t^{2 \mu }
(w_2+w_3)+1\right), \quad \Lambda =2,3. \label{xidef}\end{aligned}$$ By multiplying on the left by $P^{\tr}$, we see, with the help of -, that $W$ satisfies $$\label{relEulD}
A^0 \del{t}W + A^I \del{I}W = \frac{1}{t}P^{\tr}\bigl(\Bc \pi \Vt-t \Bt^0 Z)$$ where $$\label{Aidef}
A^i = P^{\tr} \Bt^i P.$$
Next, setting $$\label{Acdef}
\Ac^I = (A^0)^{-1}A^I$$ and $$\label{Fcdef}
\Fc = \frac{1}{t}(A^0)^{-1} P^{\tr}\bigl(\Bc \pi \Vt-t \Bt^0 Z),$$ we deduce from that $W$ satisfies $$\label{relEulE}
\del{t}W + \Ac^I \del{I}W = \Fc.$$ Moreover, straightforward, but lengthy, calculations using -, , -, -, - and - yield the following explicit formulas for the matrices $A^0$, $\Ac^I$ and the source term $\Fc$: $$\label{A0rep}
A^0 = \begin{pmatrix}
K & 0 & 0 & 0 \\
0 & \frac{t^{2 \mu } e^{2 \wbr_1}-(K-1) e^{4
\wbr_1}}{\psi^2} & 0 & 0 \\
0 & 0 & \frac{2 e^{2 \wbr_1} \left(2 w_3^2
t^{2 \mu }+1\right)}{\phi^2} &
-\frac{4 w_2 w_3 t^{2 \mu } e^{2
\wbr_1}}{\phi^2} \\
0 & 0 & -\frac{4 w_2 w_3 t^{2 \mu } e^{2
\wbr_1}}{\phi^2} &
\frac{2 e^{2 \wbr_1} \left(2 w_2^2 t^{2
\mu }+1\right)}{\phi^2} \\
\end{pmatrix},$$ $$\label{Ac1rep}
\Ac^1 = \frac{1}{\sqrt{\frac{t^{2\mu}}{e^{2\wt_1}}+1}}\begin{pmatrix}
-\frac{1}{\sqrt{\phi}} &
-\frac{t^{2 \mu }}{\psi \sqrt{\phi}} &
\frac{2 t^{2\mu } w_2}{\phi^{3/2}} &
\frac{t^{2\mu} w_3}{\phi^{3/2}}
\\
-\frac{K t^{2 \mu } e^{-2
\wbr_1} \psi}{\sqrt{\phi}\chi} &
\frac{(2 K-1)
t^{2 \mu }+(K-1) e^{2
\wbr_1}}{\sqrt{\phi} \chi} & -\frac{2 K
t^{2\mu } \psi w_2}{\phi^{3/2}\chi} & -\frac{2 K t^{2\mu }
\psi w_3}{\phi^{3/2}\chi} \\
K t^{2\mu }w_2 e^{-2 \wbr_1}
\sqrt{\phi} & -\frac{K
t^{2\mu } w_2 \sqrt{\phi}}{\psi} &
-\frac{1}{\sqrt{\phi}} & 0
\\
K t^{2\mu }w_3 e^{-2 \wbr_1} \sqrt{\phi} &
-\frac{K t^{2\mu } w3 \sqrt{\phi}}{ \psi} & 0 &
- \frac{1}{\sqrt{\phi}} \\
\end{pmatrix},$$ $$\label{Ac2rep}
\Ac^2 = \frac{1}{\sqrt{\frac{t^{2\mu}}{e^{2\wt_1}}+1}}\begin{pmatrix}
-\frac{t^{\mu } (w_3+w_2)}{\sqrt{\phi}} &
-\frac{t^{3 \mu } (w_3+w_2)}{\psi \sqrt{\phi}} &
\frac{t^{\mu } \eta_3}{\phi^{3/2}} &
-\frac{t^\mu \eta_2}{\phi^{3/2}}
\\
-\frac{K t^{3 \mu } (w_2+w_3) e^{-2
\wbr_1} \psi}{\sqrt{\phi}\chi} &
\frac{t^{\mu } (w_2+w_3) \left((2 K-1)
t^{2 \mu }+(K-1) e^{2
\wbr_1}\right)}{\sqrt{\phi} \chi} & -\frac{K
t^{\mu } \psi \eta_3}{\phi^{3/2}\chi} & \frac{K t^{\mu }
\psi\eta_2}{\phi^{3/2}\chi} \\
-\frac{1}{2} K t^{\mu } e^{-2 \wbr_1}
\sqrt{\phi} & \frac{K
t^{\mu } \sqrt{\phi}}{2 \psi} &
-\frac{t^{\mu } (w_3+w_2)}{\sqrt{\phi}} & 0
\\
-\frac{1}{2} K t^{\mu } e^{-2 \wbr_1} \sqrt{\phi} &
\frac{K t^{\mu } \sqrt{\phi}}{2 \psi} & 0 &
-\frac{t^{\mu } (w_3+w_2)}{\sqrt{\phi}} \\
\end{pmatrix},$$ $$\label{Ac3rep}
\Ac^3 = \frac{1}{\sqrt{\frac{t^{2\mu}}{e^{2\wt_1}}+1}}\begin{pmatrix}
\frac{t^{\mu } (w_3-w_2)}{\sqrt{\phi}} &
\frac{t^{3 \mu } (w_3-w_2)}{\psi \sqrt{\phi}} &
-\frac{t^{\mu } \xi_3}{\phi^{3/2}} &
\frac{t^\mu \xi_2}{\phi^{3/2}}
\\
-\frac{K t^{3 \mu } (w_2-w_3) e^{-2
\wbr_1} \psi}{\sqrt{\phi}\chi} &
\frac{t^{\mu } (w_2-w_3) \left((2 K-1)
t^{2 \mu }+(K-1) e^{2
\wbr_1}\right)}{\sqrt{\phi} \chi} & \frac{K
t^{\mu } \psi \xi_3}{\phi^{3/2}\chi} & -\frac{K t^{\mu }
\psi\xi_2}{\phi^{3/2}\chi} \\
-\frac{1}{2} K t^{\mu } e^{-2 \wbr_1}
\sqrt{\phi} & \frac{K
t^{\mu } \sqrt{\phi}}{2 \psi} &
\frac{t^{\mu } (w_3-w_2)}{\sqrt{\phi}} & 0
\\
\frac{1}{2} K t^{\mu } e^{-2 \wbr_1} \sqrt{\phi} &
-\frac{K t^{\mu } \sqrt{\phi}}{2 \psi} & 0 &
\frac{t^{\mu } (w_3-w_2)}{\sqrt{\phi}} \\
\end{pmatrix}$$ and $$\label{Fcrep}
\Fc = t^{2\mu -1}\Gc + \Fc_0,$$ where $$\begin{gathered}
\psi = t^{2 \mu }+e^{2 \wbr_1}, \label{psidef}\\
\chi = t^{2\mu }-(K-1) e^{2 \wbr_1}, \label{chidef}\end{gathered}$$ $$\label{Gcdef}
\Gc =\begin{pmatrix}
0 \\
-\frac{K (3 K-1) \left(e^{2 w_1}-1\right) e^{2
u}}{\left((K-1) e^{2 u}-t^{2 \mu }\right)
\left((K-1) e^{2 \wbr_1}-t^{2 \mu
}\right)} \\
0 \\
0
\end{pmatrix},$$ $$\label{Fc0def}
\Fc_0 = -\frac{\mu}{t}\Pi W
+ \frac{1}{t}
\begin{pmatrix}
0 \\
\frac{-(\mu -3 K+1) t^{2 \mu }+t u'\!(t) \left(t^{2
\mu }-(K-1) e^{2 u(t)}\right)+(-\mu +(\mu
+3) K-1) e^{2 u(t)}}{ \left((K-1) e^{2 u(t)}-t^{2
\mu }\right)} \\
0 \\
0
\end{pmatrix}$$ and $$\label{Pidef}
\Pi = \begin{pmatrix} 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{pmatrix}.$$ For later use, we also define $$\label{Piperpdef}
\Pi^\perp = \id - \Pi,$$ and observe that $\Pi$ and $\Pi^\perp$ satisfy the relations $$\Pi^2 = \Pi, \quad (\Pi^\perp)^2 = \Pi^\perp, \quad \Pi\Pi^\perp =\Pi^\perp \Pi = 0 \AND \Pi+\Pi^\perp = \id. \label{Pirel}$$
Homogeneous solutions\[Homsec\]
-------------------------------
To proceed, we need to identify the homogeneous solutions that we will show are stable to the future under nonlinear perturbations. We locate these solutions by noting from that $$\Gc|_{W=0} = 0.$$ From this, and , it is then clear that the trivial solution $W=0$ will solve provided that $\mu$ and $u(t)$ are chosen to satisfy $$\label{mufixA}
-\mu +(\mu
+3) K-1 = 0$$ and $$\label{HomeqA}
u'\!(t) =\frac{(\mu -3 K+1) t^{2 \mu-1}}{t^{2
\mu }-(K-1) e^{2 u(t)}},$$ respectively. Solving for $\mu$ yields $$\label{mufixB}
\mu = \frac{3K-1}{1-K},$$ which we observe by satisfies $$\label{mufixC}
\mu > 0.$$ Moreover using , we note that can be expressed as $$\label{HomeqB}
u'\!(t) =\frac{K\mu t^{2 \mu-1}}{t^{2
\mu }+(1-K) e^{2 u(t)}}.$$ The following proposition guarantees the existence of solutions to that exist for all $t\in (0,1]$.
\[Homprop\] Suppose $1/3<K<1$, $\mu = (3K-1)/(1-K)$, and $u_0 \in \Rbb$. Then there exists a unique solution $u \in C^\infty((0,1]) \cap C^0([0,1])$ to the initial value problem $$\begin{aligned}
u'\!(t) &=\frac{K\mu t^{2 \mu-1}}{t^{2
\mu }+(1-K) e^{2 u(t)}},\quad 0<t\leq 1, \label{HomeqB.1} \\
u(0) &= u_0, \label{HomeqB.2}\end{aligned}$$ that satisfies $$|u(t)-u(0)| \lesssim t^{2\mu} \AND |u'\!(t)| \lesssim t^{2\mu-1} \label{Hombounds}$$ for all $t\in (0,1]$. Moreover, for each $\rho_c\in \Rbb>0$, the solution $u$ determines a homogenous solution of the relativistic Euler equations given by $$\label{Homsol}
\bigl(\rho,\vt^i) = \biggl( \frac{\rho_c t^{\frac{2(1+K)}{1-K}}}{(t^{2\mu}+ e^{2u})^{\frac{1+K}{2}}},
-t^{1-\mu}\sqrt{e^{2u}+t^{2 \mu} }, t^{1-\mu }e^{u},0,0\biggr) .$$
By standard local existence theorems for ODEs, we know there exists a $T \in [0,1)$ and a unique solution $u \in C^\infty((T,0])$ to the initial value problem - that can be continued to smaller times as long as $u(t)$ stays bounded. Noting that is separable, it can be integrated to yield the implicit solution $${\textstyle \frac{K}{2}} \ln \bigl(t^{2 \mu }+e^{2
u(t)}\bigr)- u(t)=c$$ where the constant $c$ is uniquely determined by the initial condition $u_0$ and the constants $K,\mu$. Solving for $t^{2\mu}$ shows that $$e^{\frac{2(c+u(t))}{K}}-e^{2 u(t)} = t^{2\mu}.$$ Since $\mu>0$, this implies the inequality $$0\leq e^{\frac{2(c+u(t))}{K}}-e^{2 u(t)} \leq 1, \quad 0\leq T<t \leq 1,$$ from which we deduce the lower bound $$\label{Homprop1}
e^{2 u(t)} \leq e^{\frac{2(c+u(t))}{K}}\quad \Longrightarrow \quad 2 u(t) \leq \frac{2(c+u(t))}{K} \quad \Longrightarrow\quad -\frac{c}{1-K}\leq u(t), \quad 0\leq T<t \leq 1.$$ On the other hand, since the right hand side of the ODE is positive, $u(t)$ must be increasing, and hence, it is bounded above by $$\label{Homprop2}
u(t) \leq u_0 \quad 0\leq T<t \leq 1.$$ Thus $u(t)$ is bounded above and below, and so we conclude via the continuation principle for ODEs that the solution $u(t)$ must exist for all $t\in (0,1]$, that is, $T=0$.
Next, integrating in time, we see, with the help of the lower and upper bounds - and the triangle inequality, that $u(t)$ satisfies the estimate $$\label{Homprop3}
|u(t_2)-u(t_1)| \lesssim \int^{t_2}_{t_1} \tau^{2\mu-1}\, d\tau \lesssim t_2^{2\mu} - t_1^{2\mu}, \quad 0<t_1< t_2 \leq 1.$$ From this, we conclude that the limit $\lim_{t\searrow 0} u(t)$ exists and $u(t)$ extends to a uniformly continuous function on $[0,1]$. Setting $t_2=t$ and sending $t_1\searrow 0$ in gives $$|u(t)-u(0)| \lesssim t^{2\mu}, \quad 0<t\leq 1.$$ We further note that the inequality $$|u'\!(t)| \lesssim t^{2\mu-1}, \quad 0<t\leq 1,$$ follows directly from the bounds - and the ODE . To complete the proof, we observe, by construction, that the trivial solution $W=0$ to determines via , , - and a homogeneous solution to the relativistic Euler equations given by .
Fuchsian form\[Fform\]
----------------------
To complete the transformation of the relativistic Euler equations into Fuchsian form, we let $u(t)$ denote one of the homogeneous solutions to the IVP - from Proposition \[Homprop\]. Then by construction, $\Fc_0=0$, and so, reduces, see and , to $$\label{relEulF}
\del{t}W + \Ac^I \del{I}W =-\frac{\mu}{t}\Pi W + t^{2\mu-1}\Gc.$$ Applying the projection operator $\Pi$ to this equation and noting the $\Pi \Gc = 0$ by and , we get that $$\del{t}(\Pi W) + \Pi\Ac^I \del{I}W =-\frac{\mu}{t}\Pi W,$$ which we observe can equivalently written as $$\label{relEulG}
\del{t}(t^{\mu}\Pi W) + t^{\mu}\Pi\Ac^I \del{I}W = 0.$$ Next, applying $\Pi^\perp$, see , to shows, with the help of , that $$\del{t}(\Pi^\perp W) + \Pi^\perp\Ac^I \del{I}W = t^{2\mu-1}\Pi^\perp\Gc.$$ Adding this equation to gives $$\label{relEulH}
\del{t}\Wb+t^\mu\Pi\Ac^I \del{I}W+\Pi^\perp\Ac^I \del{I}W = t^{2\mu-1}\Pi^\perp\Gc$$ where we have set $$\label{Wbdef}
\Wb := \Pi^\perp W+t^{\mu}\Pi W = (\zetat,w_1,t^{\mu}w_2,t^{\mu}w_3)^{\tr}.$$
We then differentiate spatially to get $$\del{t}\del{J}W + \Ac^I \del{I}\del{J}W + \del{J}\Ac^I \del{I}W = -\frac{\mu}{t}\Pi \del{J}W + t^{2\mu-1}\del{J}\Gc.$$ Setting $$\label{WbJdef}
\Wb\!_J := t^\mu \del{J}W = (t^\mu\del{J}\zetat,t^\mu \del{J}w_1,t^{\mu}\del{J}w_2,t^{\mu} \del{J} w_3)^{\tr},$$ we can write this as $$\del{t}\Wb\!_J + \Ac^I \del{I}\Wb\!_J + \del{J}\Ac^I \Wb_I = \frac{\mu}{t}\Pi^\perp \Wb\!_J + t^{3\mu-1}\del{J}\Gc.$$ Multiplying on the left by $A^0$ and recalling the definitions , we find that $\Wb\!_J$ satisfies $$\label{relEulI}
A^0\del{t}\Wb\!_J + A^I \del{I}\Wb\!_J = \frac{\mu}{t}A^0\Pi^\perp \Wb\!_J + t^{3\mu-1}A^0\del{J}\Gc- A^0\del{J}\Ac^I \Wb_I.$$ Additionally, using the definition , we observe that can be written as $$\label{relEulJ}
\del{t}\Wb=-\Pi\Ac^I \Wb_I- t^{-\mu}\Pi^\perp\Ac^I \Pi \Wb_I + t^{2\mu-1}\Pi^\perp\Gc- t^{-\mu}\Pi^\perp\Ac^I \Pi^\perp \Wb_I.$$
Finally, combining and yields the system $$\label{relEulK}
\Asc^0\del{t}\Wsc + \Asc^I \del{I}\Wsc = \frac{\mu}{t}\Asc^0\Pbb \Wsc + \Fsc_0 +\Fsc_1$$ where $$\begin{aligned}
\Wsc &= \begin{pmatrix} \Wb \\ \Wb\!_J \end{pmatrix}, \label{Wscdef} \\
\Asc^0 &= \begin{pmatrix} \id & 0 \\ 0 & A^0 \end{pmatrix}, \label{Asc0def} \\
\Asc^I &= \begin{pmatrix} 0 & 0 \\ 0 & A^I \end{pmatrix}, \label{AscIdef} \\
\Pbb &= \begin{pmatrix} 0 & 0 \\ 0 & \Pi^\perp \end{pmatrix}, \label{Pbbdef} \\
\Fsc_0 &=\begin{pmatrix} -\Pi\Ac^I \Wb_I- t^{-\mu}\Pi^\perp\Ac^I \Pi \Wb_I + t^{2\mu-1}\Pi^\perp\Gc \\
t^{3\mu-1}A^0\del{J}\Gc- A^0\del{J}\Ac^I \Wb_I\end{pmatrix} \label{Fsc0def}
\intertext{and}
\Fsc_1 & =\begin{pmatrix} - t^{-\mu}\Pi^\perp\Ac^I \Pi^\perp \Wb_I \\ 0\end{pmatrix}. \label{Fsc1def}\end{aligned}$$ The point of this system, as will be established in the proof of Theorem \[mainthm\], is that it is now of a suitable Fuchsian form to which we can apply the existence theory from [@BOOS:2019]. This will allow us to establish the future stability of nonlinear perturbations of the homogeneous solutions to relativistic Euler equations that are defined by .
Coefficient properties\[coeff\]
-------------------------------
We now turn to showing that the coefficients of the system satisfy the required properties needed to apply the existence theory from [@BOOS:2019] in the proof of Theorem \[mainthm\]. To begin, we define $$\label{bvarsdef}
\tb = t^{2\mu}, \AND \wb_\Lambda = t^\mu w_\Lambda, \quad \Lambda=2,3,$$ and observe from -, and that the matrix $A^0$ can be treated as a map depending on the variables , that is, $$\label{A0smooth}
A^0 = A^0(\tb,\wbr_1,\wb_2,\wb_3),$$ where for each $R>0$ there exists constants $r,\omega >0$ such that $A^0$ is smooth on the domain defined by $$\label{smoothdom}
(\tb,\wbr_1,\wb_2,\wb_3) \in (-r,2) \times (-R,R) \times (-R,R)\times (-R,R),$$ and satisfies $$\label{A0lb}
A^0(\tb,\wbr_1,0,0) \geq \omega \id$$ for all $(\tb,\wbr_1)\in (-\rho,2)\times (-R,R)$. Differentiating $A^0$ with respect to $t$ then shows, with the help of , and , that $$\begin{aligned}
\del{t}A^0 &= DA^0(\tb,\wbr_1,\wb_2,\wb_3) \begin{pmatrix} 2\mu t^{2\mu-1} \\ u'(t)+\del{t}w_1\\ \del{t}\wb_2\\ \del{t}\wb_3 \end{pmatrix}
\notag \\
&= DA^0(\tb,\wbr_1,\wb_2,\wb_3) \left(\begin{pmatrix} 2\mu t^{2\mu-1} \\ u'(t)\\ 0 \\ 0 \end{pmatrix}
+ \Pc_1\del{t}\Wb \right) \label{dtA0}\end{aligned}$$ where $$\Pc_1 = \begin{pmatrix} 0 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{pmatrix},$$ $\del{t}\Wb$ can be computed using the equation of motion , and $u'(t)$ is bounded by .
Next, setting $$\label{wh1def}
\wh_1 = t^\mu e^{-2 \wbr_1},$$ it follows from -, -, - and that we can express the matrices $\Ac^I$ as $$\label{AcIsmooth}
\Ac^I = \Ac^I_1(\wh_1,\wb_2,\wb_3)+ t^\mu \Ac^I_2(\tb,\wbr_1,\wb_2,\wb_3)+ t^{2\mu} \Ac^I_3(\tb,\wbr_1,\wb_2,\wb_3)$$ where the $\Ac^I_2$, $\At^I_3$ are smooth on the domain and the $\Ac^I_1$ are smooth on the domain defined by $$(\wh_1,\wb_2,\wb_3) \in (-R,R)\times (-R,R)\times (-R,R).$$ It is also not difficult to verify from - that the $\Ac^I_1$ satisfy $$\label{PiperpAcIPi}
\Pi^\perp \Ac^I_1 \Pi = 0.$$ Differentiating the matrices $\Ac^I$ spatially, we get from , , , and that $$\begin{aligned}
\del{J}\Ac^I &= D\Ac^I_1(\wh_1,\wb_2,\wb_3)\begin{pmatrix}-2 e^{-2 \wbr_1} t^\mu \del{J}w_1 \\ t^\mu \del{J}w_2 \\
t^\mu \del{J}w_2
\end{pmatrix}\notag \\
&+ t^\mu D\Ac^I_2(\tb,\wbr_1,\wb_2,\wb_3)\begin{pmatrix} 0 \\ \del{J}w_1\\ t^\mu \del{J}w_2
\\ t^\mu\del{J} w_3 \end{pmatrix} +t^{2\mu} D\Ac^I_3(\tb,\wbr_1,\wb_2,\wb_3)\begin{pmatrix} 0 \\ \del{J}w_1\\ t^\mu \del{J}w_2
\\ t^\mu\del{J} w_3 \end{pmatrix} \notag \\
& = \Bigl(D\Ac^I_1(\wh_1,\wb_2,\wb_3)\Pc_2 + D\Ac^I_2(\tb,\wbr_1,\wb_2,\wb_3)\Pc_3+t^\mu D\Ac^I_2(\tb,\wbr_1,\wb_2,\wb_3)\Pc_3\Bigr) \Wb\!_J, \label{dJAcI}\end{aligned}$$ where $$\Pc_2 = \begin{pmatrix} 0 & -2 e^{-2 \wbr_1} & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{pmatrix} \AND
\Pc_3 = \begin{pmatrix} 0 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & t^\mu & 0 \\ 0 & 0 & 0 & t^\mu \end{pmatrix}.$$
We further observe from and - that $A^0$ satisfies $$[\Pi^\perp,A^0] = 0$$ and $$\Pi^\perp A^0 \Pi = \Pi A^0 \Pi^\perp = 0.$$ From the definitions and , it is then clear that $\Asc^0$ satisfies $$\label{Asc0Pbbcom}
[\Pbb,\Asc^0] = 0$$ and $$\label{PbbAsc0Pbbperp}
\Pbb^\perp \Asc^0 \Pbb = \Pbb \Asc^0 \Pbb^\perp = 0,$$ where $$\label{Pbbperpdef}
\Pbb^\perp = \id -\Pbb.$$ Additionally, it follows immediately from - that $\Pbb$ satisfies $$\label{Pbbrel}
\Pbb^2 = \Pbb, \quad \Pbb^{\tr} = \Pbb, \quad \del{t}\Pbb = 0 \AND \del{I} \Pbb = 0,$$ while the symmetry of the matrices $\Asc^i$, that is, $$\label{Ascisym}
(\Asc^i)^{\tr} = \Asc^i$$ is an immediate consequence of the definitions -, , and -.
\[dtA0rem\] From the definitions , , , , , , and , the evolution equation , the estimates for $u(t)$ and $u'(t)$, the derivative formula , and the smoothness properties and of the matrices $A^0$ and $\Ac^I$, respectively, and the identity , it is not difficult to verify, for $\mu$ satisfying $0\leq \mu \leq 1/2$ and $R>0$ small enough, that there exists constants $\theta, \beta>0$ such that $\del{t}A^0$ is bounded by $$|\del{t}A^0| \leq \theta t^{2\mu-1} + t^{\mu-1}\beta |\Pbb \Wsc|$$ for all $(t,\Wsc)\in [0,1]\times B_R(\Rbb^{16})$. Furthermore, from the formulas -, it is also clear that $$\Pbb\Fsc_1=0$$ and there exists a constant $\mu>0$ such that $\Pbb^\perp\Fsc_1$ is bounded by $$|\Pbb^\perp\Fsc_1| \leq t^{\mu-1}\mu |\Pbb \Wsc|$$ for all $(t,\Wsc)\in [0,1]\times B_R(\Rbb^{16})$, while $\Fsc_0$ is bounded by $$|\Fsc_0| \lesssim t^{2\mu-1}|\Wsc|$$ for all $(t,\Wsc)\in [0,1]\times B_R(\Rbb^{16})$.
By similar considerations, we see for $\mu$ satisfying $1/2<\mu <1$, that $\del{t}A^0$, $\Pbb^\perp \Fsc_1$ and $\Fsc_0$ are bounded by $$\begin{aligned}
|\del{t}A^0| &\leq \theta t^{1-2\mu} + t^{-\mu}\beta |\Pbb \Wsc|, \\
|\Pbb^\perp\Fsc_1| &\leq t^{-\mu}\mu |\Pbb \Wsc|
\intertext{and}
|\Fsc_0| &\lesssim t^{1-2\mu}|\Wsc|,\end{aligned}$$ respectively, for all $(t,\Wsc)\in [0,1]\times B_R(\Rbb^{16})$.
Future stability\[stability\]
=============================
We are now ready to establish the future stability of nonlinear perturbations of the homogeneous solutions of relativistic Euler equations.
\[mainthm\] Suppose $k\in\Zbb_{>3/2+3}$, $1/3<K < 1/2$, $\mu = (3K-1)/(1-K)$, $\sigma > 0$, $u_0\in \Rbb$, $u \in C^\infty((0,1])\times C^0([0,1])$ is the unique solution to the IVP -, and $\zetat_0, w^0_J \in H^{k+1}(\Tbb^3)$. Then for $\delta>0$ small enough, there exists a unique solution $$W=(\zetat,w_J)^{\tr} \in C^0\bigl((0,1], H^{k+1}(\Tbb^3,\Rbb^4)\bigr)\cap C^1\bigl((0,1],H^{k}(\Tbb^3,\Rbb^4)\bigr)$$ to (see and ) the initial value problem $$\begin{aligned}
A^0 \del{t}W + A^I \del{I}W &= A^0 \Fc && \text{in $(0,1]\times \Tbb^3$,} \label{relEulL1} \\
W &= (\zetat_0, w^0_J)^{\tr} && \text{in $\{1\}\times \Tbb^3$,} \label{relEulL2}\end{aligned}$$ provided that $$\biggl(\norm{\zetat_0}_{H^{k+1}}^2+\sum_{J=1}^3\norm{w^0_J}_{H^{k+1}}^2\biggr)^{\frac{1}{2}}\leq \delta.$$ Moreover,
(i) $W=(\zetat,w_J)^{\tr}$ satisfies the energy estimate $$\Ec(t) + \int_t^1 \tau^{2\mu-1}\bigl(\norm{D\zetat(\tau)}_{H^k}^2+\norm{Dw_1(\tau)}_{H^k}^2\bigr)\,d\tau \lesssim \norm{\zetat_0}_{H^{k+1}}^2+\sum_{J=1}^3\norm{w^0_J}_{H^{k+1}}^2$$ for all $t\in (0,1]$ where[^5] $$\Ec(t)=\norm{\zetat(t)}_{H^k}^2+\norm{w_1(t)}_{H^k}^2+t^{2\mu}\Bigl(\norm{D\zetat(t)}_{H^k}^2+\norm{Dw_1(t)}_{H^k}^2+\norm{w_2(t)}_{H^{k+1}}^2+\norm{w_3(t)}_{H^{k+1}}^2\Bigr),$$
(ii) there exists functions $\zetat_*, w_1^* \in H^{k-1}(\Tbb^3)$ and $\wb_2^*,\wb_3^* \in H^{k}(\Tbb^3)$ such that the estimate $$\begin{aligned}
\bar{\Ec}(t) \lesssim
t^{\mu-\sigma}\end{aligned}$$ holds for all $t\in (0,1]$ where $$\bar{\Ec}(t)=\norm{\zetat(t) - \zetat_*}_{H^{k-1}}+\norm{w_1(t) - w_1^*}_{H^{k-1}}
+\norm{t^\mu w_2(t) - \wb_2^*}_{H^{k}}+\norm{t^\mu w_3(t) - \wb_3^*}_{H^{k}},$$
(iii) and $u$ and $W=(\zetat,w_J)^{\tr}$ determine a solution of the relativistic Euler equations on the spacetime region $M=(0,1]\times \Tbb^3$ via the formulas $$\begin{aligned}
\rho &= \frac{\rho_c t^{\frac{2(1+K)}{1-K}} e^{(1+K)\zetat}}{(t^{2\mu}+ e^{2(u+w_1)})^{\frac{1+K}{2}}}, \label{relEulsol.1}\\
\vt^0 &= -t^{1-\mu}\sqrt{e^{2 (u+w_1)}+t^{2 \mu} },\label{relEulsol.2}\\
\vt^1 &=t^{1-\mu }\biggl( \frac{e^{u+w_1}}{\sqrt{ (t^{\mu}w_2-t^{\mu}w_3)^2+(t^{\mu}w_2+t^{\mu}w_3)^2+1}} \biggr), \label{relEulsol.3} \\
\vt^2 &= t^{1-\mu }\biggl( \frac{(t^{\mu}w_2+t^{\mu}w_3) e^{u+w_1}}{\sqrt{ (t^{\mu}w_2-t^{\mu}w_3)^2+(t^{\mu}w_2+t^{\mu}w_3)^2+1}}\biggr)\label{relEulsol.4}
\intertext{and}
\vt^3 &= t^{1-\mu }\biggl( \frac{(t^{\mu}w_2-t^{\mu}w_3) e^{u+w_1}}{\sqrt{ (t^{\mu}w_2-t^{\mu}w_3)^2+(t^{\mu}w_2+t^{\mu}w_3)^2+1}}\biggr). \label{relEulsol.5}\end{aligned}$$
By - and , we know that the matrices $A^i$ are symmetric. Furthermore, from the analysis carried out in Section \[coeff\], we know that the maps $A^i$ and $\Fc$ depend smoothly on the variables $(t,\zeta,w_J)$ for $t\in (0,1]$ and $(\zeta,w_J)$ in an open neighborhood of zero, and that the matrix $A^0$ is positive definite. This shows that the system is symmetric hyperbolic. Since $k\in\Zbb_{>3/2+3}$ and $$W_0 :=(\zetat_0, w^0_J)^{\tr}\in H^{k+1}(\Tbb^3,\Rbb^4))$$ by assumption, we can appeal to standard local-in-time existence and uniqueness theorems and the continuation principle for symmetric hyperbolic systems, see Propositions 1.4, 1.5 and 2.1 from [@TaylorIII:1996 Ch. 16], to conclude that there exists a maximal time $T_* \in [0,1)$ such that the IVP - admits a unique solution $$W=(\zetat,w_J)^{\tr} \in C^0((T_*,1], H^{k+1}(\Tbb^3,\Rbb^4))\cap C^1((T_*,1],H^{k}(\Tbb^3,\Rbb^4).$$ We also know from the computations carried out in Section \[Ftrans\] that $$\label{Wscvar}
\Wsc = (\Wb,\Wb\!_J),$$ where $\Wb$ and $\Wb\!_J$ are determined from the solution $W$ via the formulas and , respectively, will solve the IVP $$\begin{aligned}
\Asc^0\del{t}\Wsc + \Asc^I \del{I}\Wsc &= \frac{\mu}{t}\Asc^0\Pbb \Wsc + \Fsc_0 + \Fsc_1 &&
\text{in $(T_*,1]\times \Tbb^3$,}
\label{relEulM1} \\
\Wsc &= \Wsc_0 := (W_0,\del{J}W_0)^{\tr} &&\text{in $\{1\}\times \Tbb^3$.} \label{relEulM2}\end{aligned}$$ We further observe that if the initial data $W_0$ is chosen to satisfy $\norm{W_0}^{H^{k+1}} \leq \delta$, then $$\norm{\Wsc_0}_{H^k} \lesssim \norm{W_0}_{H^{k+1}} \leq \delta.$$
On the other hand, we can view as an equation for the variables $\Wsc=(\Wb,\Wb_J)$, with $\Wb=(w_1,\wb_2,\wb_3)$ and $\Wb=(w_{1,J},\wb_{1,J},\wb_{1,J})$, where the maps $A^0$, $A^I=A^0\Ac^I$ and $\Fsc_0$, $\Fsc_1$ depend on the variables $(t,\Wb)$ and $\Wsc$ respectively; see Section \[coeff\] above. Then from **(i)** the smoothness properties , and the identity satisfied by the matrices $A^0$ and $\Ac^I$, **(ii)** the derivative formulas and , **(iii)** the variable definitions , , , , and , **(iv)** the properties of the homogeneous solution $u(t)$ as given by Proposition \[Homprop\], **(v)** the assumption $1/3<K < 1/2$, which, in particular, implies that $0<\mu <1$, and **(vi)** the properties - of the matrices $\Asc^i$ and $\Pbb$, it is not difficult to verify using the definitions - that, for $R>0$ chosen small enough, there exists, see also Remark \[dtA0rem\], constants $\theta,\gamma_1=\gammat_1,\gamma_2=\gamma_2,\mu_2>0$ and $\beta_0,\beta_2,\beta_4,\beta_6>0$, where the $\beta_q$ can be chosen as small as we like by shrinking $R>0$ if necessary, such that system satisfies all the assumptions from Section 3.4 of [@BOOS:2019] for following choice of constants: $$\begin{gathered}
p=\begin{cases}2\mu & \text{if $0<\mu\leq 1/2$} \\
2(1-\mu) & \text{if $1/2\mu < 1$} \end{cases}, \\
\kappa=\kappat=\mu
\intertext{and}
\beta_1=\beta_3=\beta_5=\beta_7=\mu_1=\mu_3= \alpha=0.\end{gathered}$$ As discussed in [@BOOS:2019 §3.4], this implies that under the time transformation[^6] $t \mapsto t^p$, the transformed version of will satisfy all of the assumptions from Section 3.1 of [@BOOS:2019]. Moreover, since the $\Asc^I$ have a regular limit as $t\searrow 0$ (equivalently as $\tau\searrow 0$), the constants $\btt$ and $\tilde{\btt}$ from Theorem 3.8 of [@BOOS:2019] will satisfy $\btt=\tilde{\btt}=0$, and consequently, the constant[^7] $\mathfrak{z}$ that is involved in determining the decay is given by $$\mathfrak{z}= \kappa - \Half \gammat_1(\beta_1+(k-1)\tilde{\btt}) = \mu.$$ We can therefore apply Theorem 3.8 from [@BOOS:2019] to the time transformed version of as described in [@BOOS:2019 Section 3.4] to deduce, for $\delta>0$ chosen small enough and the initial data satisfying $\norm{\Wsc(0)}_{H^k}< \delta$, the existence of a unique solution $$\Wsc^* \in C^0\bigl((0,1],H^k(\Tbb^3,\Rbb^{16})\bigr)\cap L^\infty\bigl((0,1],H^k(\Tbb^3,\Rbb^{16}))\bigr)\cap
C^1\bigl((0,1],H^{k-1}(\Tbb^3,\Rbb^{16})\bigr)$$ to the IVP - with the following properties:
(1) The limit $\lim_{t\searrow 0} \Pbb^\perp \Wsc^*$, denoted $\Pbb^\perp \Wsc^*(0)$, exists in $H^{k-1}(\Tbb^3,\Rbb^{16})$.
(2) The solution satisfies the energy estimate $$\label{eestA}
\norm{\Wsc^*(t)}_{H^k}^2 + \int_{t}^1 \frac{1}{\tau} \norm{\Pbb \Wsc^*(\tau)}_{H^k}^2\, d\tau \lesssim \norm{\Wsc_0}_{H^k}^2$$ for all $t\in (0,1]$.
(3) The solution decays as $t\searrow 0$ according to $$\begin{gathered}
\label{decayA1}
\norm{\Pbb \Wsc^*(t)}_{H^{k-1}} \lesssim
t^{\mu-\sigma}
\intertext{and}
\label{decayA2}
\norm{\Pbb^\perp \Wsc^*(t) - \Pbb^\perp \Wsc^*(0)}_{H^{k-1}} \lesssim
t^{\mu-\sigma} \end{gathered}$$ for all $t\in (0,1]$.
By uniqueness, the two solutions $\Wsc$ and $\Wsc^*$ to the IVP - must coincide on their common domain of definition, and so, we have $$\Wsc(t)=\Wsc^*(t), \quad T_*<t \leq 1.$$ But this implies via , the energy estimate , and Sobolev’s inequality [@TaylorIII:1996 Ch. 13, Prop 2.4] that $$\norm{\Wb(t)}_{W^{1,\infty}} \lesssim \norm{\Wb(t)}_{H^k} \leq \norm{\Wsc(t)}_{H^{k-1}} \lesssim \norm{\Wsc_0},
\quad T^*<t\leq 1.$$ By choosing the initial data $\Wsc_0$ so that $\norm{\Wsc_0}_{H^k}$ is sufficiently small, we can then ensure that $$\norm{\Wb(t)}_{W^{1,\infty}} \leq \frac{R}{2}
\quad T^*<t\leq 1,$$ where $R>0$ is as defined in Section \[coeff\], which, in particular, is enough to guarantee that the coefficients $A^i$ and $\Fc$ of remain well-defined. By the continuation principle and the maximality of $T_*$, we conclude that $T_*=0$, and hence that $$\Wsc(t)=\Wsc^*(t), \quad 0< t\leq 1.$$ From this, the definitions , , , , and , and the energy estimate , it is then straightforward to verify $$\Ec(t) + \int_t^1 \tau^{2\mu-1}\bigl(\norm{D\zetat(\tau)}_{H^k}^2+\norm{Dw_1(\tau)}_{H^k}^2\bigr)\,d\tau \lesssim \norm{W_0}_{H^k}^2,
\quad 0<t\leq 1,$$ where $$\Ec(t)=\norm{\zetat(t)}_{H^k}^2+\norm{w_1(t)}_{H^k}^2+t^{2\mu}\Bigl(\norm{D\zetat(t)}_{H^k}^2+\norm{Dw_1(t)}_{H^k}^2+\norm{w_2(t)}_{H^{k+1}}^2+\norm{w_3(t)}_{H^{k+1}}^2\Bigr).$$ Furthermore, from the decay estimate and the definition , we obtain the existence of functions $\zetat_*, w_1^* \in H^{k-1}(\Tbb^3)$ and $\wb_2^*,\wb_3^* \in H^{k}(\Tbb^3)$ such that the estimate $$\begin{aligned}
\bar{\Ec}(t) \lesssim
t^{\mu-\sigma}\end{aligned}$$ holds for all $t\in (0,1]$, where $$\bar{\Ec}(t)=\norm{\zetat(t) - \zetat_*}_{H^{k-1}}+\norm{w_1(t) - w_1^*}_{H^{k-1}}
+\norm{t^\mu w_2(t) - \wb_2^*}_{H^{k}}+\norm{t^\mu w_3(t) - \wb_3^*}_{H^{k}}.$$ To complete the proof, we recall from , , and -, that $u$ and $W=(\zetat,w_J)^{\tr}$ determine a solution of the relativistic Euler equations on the spacetime region $M=(0,1]\times \Tbb^3$ via the formulas -.
$\Tbb^2$-symmetric future stability\[symsec\]
=============================================
In this section, we focus on solutions of the relativistic Euler equations that are independent of the coordinates $(x^2,x^3)\in \Tbb^2$, or in other words, admit a $\Tbb^2$-symmetry. To find such solutions, we set $$\label{symvars}
\zetat=\ztt(t,x^1), \quad w_1 = \wtt(t,x^1) \AND w_2=w_3=0,$$ and observe, with the help of the , -, , and -, that this ansatz leads to a consistent reduction of to a symmetric hyperbolic equations for the variables $(\ztt,\wtt)$ in $1+1$ dimensions given by $$\label{symEulA}
\Att^0 \del{t}\Wtt+ \Att^1 \del{1}\Wtt = \Ftt$$ where $$\begin{gathered}
\Wtt = (\ztt,\wtt)^{\tr}, \label{Wttdef} \\
\Att^0 =\begin{pmatrix}
K & 0 \\
0 & \frac{t^{2 \mu } e^{2 (u+\wtt)}+(1-K) e^{4
(u+\wtt)}}{(t^{2 \mu }+e^{2
(u+\wtt)})^2}
\end{pmatrix} , \label{Att0def}\\
\Att^1 =\frac{1}{\sqrt{e^{2 (u+\wtt)}+t^{2
\mu } }}\begin{pmatrix}
-K e^{u+\wtt} & -\frac{K
t^{2\mu } e^{u+\wtt}}{t^{2 \mu }+e^{2
(u+\wtt)}} \\
-\frac{K t^{2\mu } e^{u+\wtt}}{t^{2 \mu
}+e^{2 (u+\wtt)}} & \frac{
(2 K-1) t^{2 \mu } e^{3
(u+\wtt)}+(K-1) e^{5
(u+\wtt)}}{ (t^{2 \mu }+e^{2
(u+\wtt)})^2}
\end{pmatrix} ,\label{Att1def}\\
\Ftt =\begin{pmatrix}
0 \\
-\frac{ t^{2\mu -1 }K (3 K-1) (e^{2 \wtt}-1) e^{4
u+2 \wtt}}{(t^{2 \mu }-(K-1) e^{2
u}) (t^{2 \mu }+e^{2
(u+\wtt)})^2}
\end{pmatrix}
,\label{Fttdef} \end{gathered}$$ and in deriving this equation, we have assumed, as above, that $u=u(t)$ solves the IVP -.
The system is almost regular in that $\Att^0(t,\wtt)$, $\Att^1(t,\wtt)$ and $\Ftt(t,\wtt)$ are smooth in $(t,\wtt)$ for $(t,\wtt)\in (0,1]\times \Rbb$ and $\Att^0$ and $\Att^1$ are, for any $R>0$, uniformly bounded for $(t,\wtt)\in (0,1]\times [-R,R]$ by virtue of the assumption $1/3<K<1$, which implies that $\mu>0$. The slight difficulty in establishing existence is that $\del{t}A^0(t,\wtt)$ and $\Ftt(t,\wtt)$ are not bounded as $t\searrow 0$ for all $K\in (1/3,1)$. However, the worst that these coefficients can diverge is like $t^{2\mu -1}$, which is always integrable since $\mu>0$. As we shall see in the proof of the following theorem, this integrability allows us to modify standard local-in-time existence results in a straightforward fashion to establish the existence of solutions of on $(0,1]\times \Tbb^1$ under a suitable small initial data assumption.
\[symthm\] Suppose $k\in \Zbb_{>1/2+1}$, $1/3<K<1$, $\mu=(3K-1)/(1-K)$, $u_0\in \Rbb$ and $u\in C^\infty((0,1])\times C^0([0,1])$ is the unique solution to the IVP -, and $\ztt_0,\wtt_0 \in H^k(\Tbb^1)$. Then for $\delta>0$ small enough, there exists a unique solution $$\Wtt=(\ztt,\wtt)^{\tr} \in C^0\bigl((0,1],H^k(\Tbb^1,\Rbb^2)\bigr)\times C^1\bigl((0,1],H^{k-1}(\Tbb^1,\Rbb^2)\bigr)$$ of the IVP $$\begin{aligned}
\Att^0 \del{t}\Wtt+ \Att^1 \del{1}\Wtt &= \Ftt && \text{in $(0,1]\times \Tbb^1$,}\label{globalS1}\\
\Wtt &= (\ztt_0,\wtt_0)^{\tr} && \text{in $\{1\}\times \Tbb^1,$} \label{globalS2} \end{aligned}$$ provided that $$\bigl(\norm{\ztt}^2_{H^k}+\norm{\wtt}^2_{H^k} \bigr)^{\frac{1}{2}} \leq \delta.$$ Moreover,
(i) the solution and its time derivative are bounded by $$\norm{\Wtt}_{H^k} \lesssim 1 \AND \norm{\del{t}\Wtt}_{H^{k-1}} \lesssim 1+t^{2\mu-1}$$ respectively, for all $t\in (0,1]$,
(ii) there exists functions $\ztt_*,\wtt_*\in H^{k-1}(\Tbb^1)$ such that $$\norm{\ztt(t)-\ztt_*}_{H^{k-1}}+\norm{\wtt(t)-\wtt_*}_{H^{k-1}} \lesssim t+ t^{2\mu}$$ for all $t\in (0,1]$,
(iii) and $u$ and $\Wtt=(\ztt,\wtt)^{\tr}$ determine a solution of the relativistic Euler equations on the spacetime region $M=(0,1]\times \Tbb^3$ via the formulas $$\begin{aligned}
\rho &= \frac{\rho_c t^{\frac{2(1+K)}{1-K}} e^{(1+K)\ztt}}{(t^{2\mu}+ e^{2(u+\wtt)})^{\frac{1+K}{2}}}, \\ \vt^0 &= -t^{1-\mu}\sqrt{e^{2 (u+\wtt)}+t^{2 \mu} }, \\ \vt^1 &=t^{1-\mu }e^{u+\wtt} \intertext{and}
\vt^2 &=\vt^3 = 0. $$
Since $K\in (1/3,1)$, the inequality $\mu>0$ holds, and so, fixing $R>0$, we observe from that there exists a constant $\gamma>0$ such that $$\label{globalP2}
\frac{1}{\gamma}\id \leq \Att^0(t,\wtt) \leq \gamma\id$$ for all $(t,\wtt)\in (0,1]\times [-R,R]$. From this inequality and the smooth dependence, see -, of $\Att^0$, $\Att^1$ and $\Ftt$ on $(t,\wtt)$ for $(t,\wtt)\in (0,1]\times \Rbb$, it follows that the system is symmetric hyperbolic. Consequently, fixing $k \in \Zbb_{> 1/2+1}$ and choosing initial data $\ztt_0,\wtt_0\in H^{k}(\Tbb^1)$ satisfying $$\label{globalP3}
\norm{\Wtt(1)}_{H^k} = \sqrt{\norm{\ztt_0}_{H^k}^2 + \norm{\wtt_0}_{H^k}^2} \leq \delta$$ for some $\delta>0$, we know from standard local-in-time existence and uniqueness theorems and the continuation principle for symmetric hyperbolic systems, see Propositions 1.4, 1.5 and 2.1 from [@TaylorIII:1996 Ch. 16], that there exists a unique solution $$\Wtt=(\ztt, \wtt )^{\tr} \in C^0\bigl((T^*,1],H^{k}(\Tbb^1)\bigr)\cap C^1 \bigl((T^*,1],H^{k-1}(\Tbb^1)\bigr)$$ to satisfying the initial condition for some time $T^*\in [0,1)$, which we can take to be maximal.
Next, applying $\Att^0\del{1}^\ell (\Att^0)^{-1}$ to gives $$\label{globalP7}
\Att^0 \del{t}\del{1}^\ell \Wtt+ \Att^1 \del{1} \del{1}^\ell \Wtt= \Ftt_\ell, \quad \ell =0,1,\dots,k,$$ where $$\label{globalP9}
\Ftt_\ell =-\Att^0[\del{1}^\ell,(\Att^0)^{-1}\Att^1]\del{1}\Wtt + \Att^0\del{1}^\ell\bigl((\Att^0)^{-1}\Ftt\bigr).$$ Employing a standard $L^2$ energy estimate, we obtain the energy inequality $$\label{globalP10}
-\del{t}\nnorm{\del{1}^\ell \Wtt}_0^2 \leq \norm{\textrm{Div}\Att}_{L^\infty}\norm{\del{1}^\ell \Wtt}_{L^2}^2 +
2\norm{\del{1}^\ell \Wtt}_{L^2}\norm{\Ftt_\ell }_{L^2}, \quad \ell =0,1,\dots,k,$$ from , where $$\textrm{Div}\Att = \del{t}\Att^0 + \del{1}\Att^1$$ and $$\nnorm{(\cdot)}^2_0 = \ip{(\cdot)}{A^0(\cdot)}$$ is the energy norm.
To proceed, we define the higher energy norms $$\nnorm{\Wtt}_k^2= \sum_{\ell=0}^k\nnorm{\del{1}^\ell \Wtt}_0^2,$$ and observe via that the equivalence of norms $$\label{globalP11}
\frac{1}{\sqrt{\gamma}}\norm{\Wtt}_{H^k}\leq \nnorm{\Wtt}_k \leq \sqrt{\gamma}\norm{\Wtt}_{H^k}$$ holds. Using this equivalence, we obtain, after summing over $\ell$ from $0$ to $k$, the differential energy estimate $$\label{globalP12}
-\del{t}\nnorm{\Wtt}_k^2 \lesssim \norm{\textrm{Div}\Att}_{L^\infty}\nnorm{\Wtt}_{k}^2 +
\nnorm{\Wtt}_{k}\left(\sum_{\ell=0}^k\norm{\Ftt_\ell }_{L^2}\right).$$ Since $k>1/2+1$, we have by Sobolev’s inequality [@TaylorIII:1996 Ch. 13, Prop 2.4] that $$\label{globalP20}
\norm{\Wtt}_{L^\infty}+ \norm{\del{1} \Wtt}_{L^\infty} \leq C_{\text{Sob}} \norm{\Wtt}_{H^k}$$ for some constant $C_{\text{Sob}}>0$ independent of the solution $\Wtt$. We then set $\Rc = \frac{R}{\sqrt{\gamma}C_{\text{Sob}}}$ so that $$\label{globalP21}
\nnorm{\Wtt}_{k} < \Rc \quad \Longrightarrow \quad \norm{\Wtt}_{L^\infty} < R$$ by and . We also choose $\delta$, see above, so that $0< \delta < \frac{\Rc}{4\sqrt{\gamma}}$ in order to guarantee that $\nnorm{\Wtt(1)}< \frac{\Rc}{4}$, and we let $T_* \in (T^*,1),$ be the first time such that $\norm{\Wtt(T_*)}_{H^k}= \frac{\Rc}{2}$ or if that time does not exist, then we set $T^*=T_*$, the maximal time of existence. In either case, we have that $$\label{globalP22}
\nnorm{\Wtt(t)}_{k} < \frac{\Rc}{2}, \quad 0<T^*\leq T_* < t \leq 1.$$
From the formulas - and the bounds obeyed by $u(t)$, it is then clear that there exists a constant $C_\ell>0$, $\ell\in \Zbb_{\geq 0}$, such that $\Att^0$, $\Att^1$ and $\Ftt$ are bounded by $$\begin{gathered}
|\del{\wtt}^\ell \Att^0(t,\wtt)| + |\del{\wtt}^\ell \Att^1(t,\wtt)| \leq C_\ell , \quad
|\del{\wtt}^\ell\del{t} \Att^0(t,\wtt)| \leq C_\ell (1+t^{2\mu-1}) \label{globalP22a}
\intertext{and}
\bigl|\del{\wtt}^\ell \Ftt(t,\wtt)\bigr| \leq C_\ell(1+t^{2\mu-1})|\wtt| \label{globalP22c}\end{gathered}$$ for all $(t,\wtt)\in (0,1]\times [-R,R]$. These bounds in conjunction with the Moser and commutator estimates, see Propositions 3.7 and 3.9 from [@TaylorIII:1996 Ch. 13], and the inequality imply that $$\begin{gathered}
\norm{\textrm{Div}\Att}_{L^\infty} \leq (1+t^{2\mu-1})C(\norm{\Wtt}_{H^k}) \AND
\left(\sum_{\ell=0}^k\norm{\Ftt_\ell }_{L^2}\right) \leq (1+ t^{2\mu-1})C(\norm{\Wtt}_{H^k})\norm{\Wtt}_{H^k}.\end{gathered}$$ With the help of these inequalities and , we see that implies the energy estimate $$-\del{t}\nnorm{\Wtt}_k \leq (1+ t^{2\mu-1})C(\nnorm{\Wtt}_{k})\nnorm{\Wtt}_{k}, \quad 0<T^*\leq T_* < t \leq 1.$$ By Gronwall’s inequality, we obtain the bound $$\label{globalP24}
\nnorm{\Wtt(t)}_k\leq e^{C(\Rc)\int_{t}^1 1+\tau^{2\mu-1}\, d\tau}\nnorm{\Wtt(1)}_k, \quad 0<T^*\leq T_* < t \leq 1,$$ where in deriving this we have used . But $$\int_{t}^1 1+\tau^{2\mu-1}\, d\tau\lesssim 1, \quad 0<t\leq 1,$$ since $\mu>0$, and consequently, we have by , and that $$\nnorm{\Wtt(t)}_k\leq C(\Rc)\delta, \quad 0<T^*\leq T_* < t \leq 1.$$ By shrinking $\delta>0$ more if necessary, it follows that $$\label{globalP25}
\nnorm{\Wtt(t)}_k\leq \frac{\Rc}{2}, \quad 0<T^*\leq T_* < t \leq 1.$$ We therefore conclude by the continuation principle and the definition of $T_*$ that $T_*=T^*=0$, which establishes the existence of a unique solution $$\Wtt=(\ztt, \wtt )^{\tr} \in C^0\bigl((0,1],H^{k}(\Tbb^1.\Rbb^2)\bigr)\cap C^1 \bigl((0,1],H^{k-1}(\Tbb^1,\Rbb^2)\bigr)$$ to the initial value problem -.
Next, by integrating $\del{t}\Wtt$ in time, we get $$\label{globalP26}
\Wtt(t_2)-\Wtt(t_1) = \int_{t_1}^{t_2} \del{t}\Wtt(\tau)\,d\tau, \quad 0<t_1<t_2\leq 1.$$ Using to write $\del{t}\Wtt$ as $$\del{t}\Wtt= (\Att^0)^{-1}[\Att^1 \del{1}\Wtt +\Ftt],$$ it is not difficult to verify from the bounds , , , -, and , where $T_*=T^*=0$, and the Moser estimates that $$\norm{\del{t}\Wtt}_{H^{k-1}} \lesssim 1+t^{\mu-1}.$$ From this estimate and the triangle inequality, we see, after applying the $H^{k-1}$ norm to , that $$\label{globalP27}
\norm{\Wtt(t_2)-\Wtt(t_1)}_{H^{k-1}} \leq \int_{t_1}^{t_2} \norm{\del{t}\Wtt(\tau)}_{H^{k-1}}\,d\tau \lesssim |t_2-t_1| +|t_2^{2\mu}-t_1^{2\mu}|, \quad 0<t_1<t_2\leq 1.$$ From this inequality, we conclude that the limit $\lim_{t\searrow 0}\Wtt(t)$, denoted $(\ztt_*,\wtt_*)$, exists in $H^{k-1}(\Tbb^1,\Rbb^2)$. Furthermore, sending $t_1\searrow 0$ in shows that $$\norm{\ztt(t)-\ztt_*}_{H^{k-1}}+\norm{\wtt(t)-\wtt_*} \lesssim \norm{\Wtt(t)-\Wtt(0)}_{H^{k-1}} \lesssim t+t^{2\mu}, \quad 0\leq t \leq 1.$$ To complete the proof, we observe, by construction, that $\Wtt=(\ztt,\wtt)^{\tr}$ will determine a solution of the relativistic Euler equations on the spacetime region $M=(0,1]\times\Tbb^3$ by replacing $(\zetat,w_1,w_2,w_3)$ in the formulas - with .
*Acknowledgements:* This work was partially supported by the Australian Research Council grant DP170100630.
[^1]: Our indexing conventions are as follows: lower case Latin letters, e.g. $i,j,k$, will index spacetime coordinate indices that run from $0$ to $3$ while upper case Latin letters, e.g. $I,J,K$, will index spatial coordinate indices that run from $1$ to $3$.
[^2]: By introducing the change of coordinate $\tilde{t}=-\ln(t)$, the metric can be brought into the more recognizable form $$\gt = -d\tilde{t}\otimes d\tilde{t} + e^{2\tilde{t}}\delta_{ij}dx^I \otimes dx^J,$$ where now $\tilde{t} \in [0,\infty)$.
[^3]: In these articles, stability was established in the more difficult case where the fluid is coupled to the Einstein equation. However, the techniques used there also work in the simpler setting considered in this article where coupling to gravity is neglected.
[^4]: While this restriction on the sound speed is often taken for granted and implicitly assumed, it is strictly speaking not necessary; see [@Geroch:2010] for an extended discussion.
[^5]: Here, the norm $\norm{Df}_{H^k}$ is defined by $\norm{Df}^2_{H^k}= \sum_{J=1}^3 \norm{\del{J}f}^2_{H^k}$.
[^6]: By our conventions, the time variable $t$ is assumed to be positive as opposed to [@BOOS:2019] where it is taken to be negative. This causes no difficulties since one can change between these two conventions by using the simple time transformation $t\rightarrow -t$.
[^7]: In the article [@BOOS:2019], this constant is denoted by $\zeta$, but since we are already using $\zeta$ to denote the modified fluid density, we will refer to this parameter as $\mathfrak{z}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In recent years significant efforts have been made to design and fabricate functional nanomaterials for biomedical applications based on the control of light matter interaction at the nanometer scale. Among many other artificial materials, hyperbolic dispersion metamaterials allow to access unprecedented physical effects and mechanisms due to the extreme anisotropy of their optical constants. The unbound isofrequency surface of hyperbolic metamaterials (HMMs) enable the possibility to support a virtually infinite density of states and ultra-high confinement of electromagnetic fields, allowing perfect absorption of light and extreme sensing properties. Optical sensor technology based on plasmonic metamaterials offers significant opportunities in the field of clinical diagnostics, particularly for the detection of low-molecular-weight biomolecules in highly diluted solutions. In this context, we present a computational effort to engineer a biosensing platform based on hyperbolic metamaterials, supporting highly confined bulk plasmon modes integrated with out-of-plane chiral metasurfaces. The role of the helicoidal chiral metasurface is manifold: i) as a diffractive element to increase the momentum of the incoming light to excite the plasmon sensing modes with linearly and circularly polarized light; ii) as out-of-plane extended sensing surface to capture target analytes away from the substrate thereby the diffusion limit; iii) as a plamonic chiral nanostructure with enhanced sensing performance over circularly polarized reflectance light.'
author:
- Giovanna Palermo
- 'Giuseppe E. Lio'
- Marco Esposito
- Loredana Ricciardi
- Mariachiara Manoccio
- Vittorianna Tasco
- Adriana Passaseo
- Antonio De Luca
- Giuseppe Strangi
bibliography:
- 'Helix\_sensing.bib'
- 'Helix\_sensing.bib'
title: Biomolecular sensing at the interface between chiral metasurfaces and hyperbolic metamaterials
---
In the fight against cancer it is essential to diagnose the disease at an early stage. To date, one of the most widely used methods for the diagnosis and staging of the disease is represented by the tissue biopsy. However, this clinical exam is invasive and it is performed when the site of the tumor has already been identified. In contrast, liquid biopsy has proved to be a valid alternative to detect and monitor cancer biomarkers. [@joosse2013biologic; @palmirotta2018liquid; @cohen2018detection; @gorgannezhad2018circulating; @mattox2019applications; @lim2019liquid] For this reason, researchers are focused to develop extremely sensitive platforms, able to determine specific nucleic acids and proteins in circulating tumor cells (CTCs), in body fluids as blood, urine, sweat and tears.
To the present, several investigation techniques have been developed to identify for example the presence of proteins in the blood by means of an electrochemical analysis,[@leca2005biosensors] or in saliva by means of electrochemical immunoassay biosensors. [@corrie2015blood]
Optical sensor technology based on plasmonic metamaterials offers significant opportunities in the field of clinical diagnosis, particularly for the detection of low-molecular weight biomolecules in highly diluted solutions.[@sreekanth2016extreme; @baqir2018nanoscale; @sreekanth2016biosensing]
Hyperbolic metamaterials (HMMs) represent a novel class of promising artificial plasmonic materials used in the field of bio-sensing, supporting highly confined bulk plasmon polaritons (BPPs) in addition to surface plasmon polaritons (SPPs).[@poddubny2013hyperbolic; @ferrari2015hyperbolic; @shekhar2014strong; @shekhar2014hyperbolic]
It has been experimentally demonstrated that it is possible to excite both surface and bulk plasmon polaritons in HMM by means of a hypergrating, consisting of 1D and 2D metallic diffraction gratings, such as stripes, disk or holes gratings.[@sreekanth2013experimental] However, further sensing properties could be reached by engineering the hypergrating features towards the realization of a high sensitivity and high accuracy biosensing platform.
Here, we report, for the first time, a numerical simulation of a new biosensing platform based on HMMs integrated with an out-of-plane 3D chiral metasurface hypergrating (CMH), where the fundamental building block is represented by a helical shaped nanostructure. In particular, the implementation of 3D metallic grating could enable the excitation of the high-k modes of the underlying HMM and reduce the diffusion limit by the detection of molecules that do not necessarily flows close to the surface. Moreover, the strong intrinsic chirality belonging to plasmonic nanohelices induces circular polarization dependent BPP mode excitation, resulting in a giant circular dichroism value (CD), boosting the sensing properties of the system.
We demonstrate that our advanced biosensing platform exhibits extremely high sensing properties since it takes advantages of the elevated sensitivity of the HMM bulk modes (high Q factor) and the significant accuracy of the circular dichroism (CD) signal of the CMH that exhibits more spectral features to track the environment changes.
Moreover, considering the increased out-of-plane surface of the 3D nanohelices, able to that interact with analyte targets, we have introduced the mass sensitivity as analytical figure of merit, that describes the surface coverage-refractive index variation relationship. We have calculated that the minimum quantity of binding analyte that can leads to an appreciable variation in terms of refractive index unit (RIU) results equal to 0.04 pg/mm$^2$.
Similarly, we carried out a study of the analyte diffusion calculating the sensitivity changes as a function of the relative surface coverage, at different distance away from the HMM surface, noting that our system allows to detect analytes over the diffusion limit of planar hypergrating.
Finally, we quantified the sensitivity and Figure of merit (FOM) of the sensor tracking the spectral shift of the CD features as a function of the glycerol/water concentration ratio leading to a remarkable FOM value of about 20.0 RIU$^{-1}$.
These results indicate that our innovative sensing platform offers unique advantages for high sensitivity detection of analytes in many biological applications.
Results and discussion
======================
The sketch reported in Figure \[fig:1\]a shows the schematic diagram of a type II HMM configuration consisting of alternating dielectric and metallic thin films. In particular, to obtain a hyperbolic dispersion at optical frequencies (with dielectric permittivity tensor component $\varepsilon_{//} < 0$ and $\varepsilon_{\perp} > 0$, above 418 nm), we designed a HMM composed by indium tin oxide (ITO - 20 nm) and silver (Ag - 20 nm) layers using Effective Medium Theory (EMT)[@choy2015effective] (see Supporting Information).
Figure \[fig:1\]b shows the numerical reflectance and transmittance curves of the HMM, provided by COMSOL Multiphysics simulation.[@multiphysics1998introduction] The reflectance and transmittance are calculated by considering a TM-wave (p-polarization) and an incident angle $\theta_i =50^\circ$. The minimum value at about 358 nm in the reflectance curve (and the related maximum in the transmittance one) is referred to as the Ferrel-Berreman mode for silver nanometric layers.[@caligiuri2016dielectric]
![(a) Schematic diagram of the ITO/Ag HMM and (b) corresponding calculated Reflectance and Transmittance spectra. (c) Sketch of one lateral unit cell of the out-of-plane chiral structure composed of a right-handend Au helix on a glass substrate, the relevant structure parameters are illustrated. (d) Calculated Reflectance and Transmittance spectra for the chiral metasurface. The reflectance and transmittance spectra are calculated for TM wave, angle of incidence $\theta_i =50^\circ$.[]{data-label="fig:1"}](Fig1.png){width="0.7\columnwidth"}
In order to excite both surface and bulk plasmon modes inside the HMM, we designed an out-of-plane CMH as shown in Figure \[fig:1\]c. The geometrical parameters of the helix array are wire radius ($r$), helix radius ($R_h$), pitch height or axial pitch ($p$), the number of pitches ($N_p$) and the lattice constant ($a$).[@gansel2010gold] Currently CMH with these geometrical features could be fabricated by focused ion beam induced deposition (FIBID),[@esposito2014three] or by the shadow-growth or glancing-angle deposition techniques. [@mark2013hybrid; @hawkeye2007glancing; @robbie1996chiral; @zhao2003designing]
The simulations were conducted considering an infinite array of helices both on a glass substrate and onto the HMM structures. This has been possible by simulating a single unit cell of the chiral metasurface composed of a right-handed Au helix arranged in a square lattice. In the first analysis distilled water (n = 1.3330) was considered as the surrounding medium for the chiral metasurface. The reflectance and transmittance curves are calculated for a TM wave and same angle of incidence ($\theta_i =50^\circ$) with respect to the helix axis. The resulting spectra, shown in Figure \[fig:1\]d, are determined by the interplay of internal resonances and their mutual coupling that strongly depends on the structure parameters of the CMH.
It is important to note that the calculated reflectance for the chiral metasurface on glass substrate, does not exceed the 40% in all the considered spectral range (400-1600 nm). In order to study the optical behavior in terms of reflectance and transmittance of the chiral metasurface on the HMM (CMH-HMM), we combined the two above mentioned structures into a single 3D geometry built in COMSOL. The obtained 3D system is composed by a parallelepiped 8 times higher than the lattice constant $a$ - Figure \[fig:2\]a. The choice of the box height is fundamental for the optimal propagation of the lightwave inside the system preventing any diffraction or boundary problems. The numerical geometry of the model is sketched in Figure \[fig:2\]a, with, starting from the top, the superstrate containing the Au helix (distilled water as surrounding medium), the multi-layer system (grey stack of ITO/Ag) and the substrate (glass).
![(a) Unit cell of the CMH-HMM simulated geometry. COMSOL permits to simulate sources or detectors of e.m. radiation by creating Ports: in our case, there is a port on the top (P$_{in}$), from where the radiation propagates and a port on the bottom (P$_{out}$), that behaves as a detector. (b) Calculated Reflectance and Transmittance spectra of an Au helix array on HMM - water as surrounding medium - for TM, angle of incidence $\theta_i$=50$^\circ$.[]{data-label="fig:2"}](Fig2.png){width="0.8\columnwidth"}
Here, an electromagnetic plane wave impinges on the combined system from the superstrate, with a specific incident angle ($\theta_i$) and polarization (s-polarization (TE) or p-polarization (TM)). Light interacting with the structure is then collected through both its substrate (typically glass) and superstrate, allowing to compute the optical quantities of interest (transmittance (T) and reflectance (R)).
This is done by solving the frequency-domain partial differential equation (PDE) that governs the $\mathbf{E}$ and $\mathbf{H}$ fields associated with the electromagnetic wave propagating through the structure:
$$\nabla \times\mu_r^{-1}(\nabla \times\mathbf{E})-\omega^2\varepsilon_0\mu_0(\varepsilon_r-i\sigma/\omega\varepsilon_0)\mathbf{E}=0
\label{eq1}$$
In the condition of the electric conductivity $\sigma=0$ and non-magnetic materials ($\mu_r=1$), the previous equation reduces to:
$$\nabla \times(\nabla \times\mathbf{E})-k_0\textsuperscript{2}\varepsilon_r\mathbf{E}=0
\label{eq2}$$
Here $k_0$ is the incident wavevector in vacuum or air ($k_0 = 2\pi/\lambda$), while $\varepsilon_r$ represents the material dielectric permittivity. By providing as an input the values of real and imaginary parts of the refractive index of any considered material, the software retrieves the corresponding dielectric permittivity ($\varepsilon_r$) and numerically solves Eq. \[eq2\] to obtain the $\mathbf{E}$ field distribution. Scattering parameters (S-parameter) for transmittance ($S_{21}$) and reflectance ($S_{11}$) are computed (See Supporting information).
In order to simulate the behavior of TM and TE light in a 3D environment, it is necessary to properly write the components of the electromagnetic fields with respect to the incidence plane ($xz$) and set the polarizations accordingly: $\mathbf{H}= (0,1,0)$ for TM and $\mathbf{E}=(0,1,0)$ for TE.[@lio2019comprehensive]
As we can see in Figure \[fig:2\]b, the calculated transmittance for the CMH-HMM turns out to be zero in all the considered vis-NIR regions, while the reflectance spectrum, calculated for a TM and $\theta_i = 50^\circ$, is totally modified respect to the reflectance of the HMM and the Au helix array considered separately (see Figures 1b-1d). In particular, we can distinguish three reflectance minima at 635 nm (mode A), 710 nm (mode B) and 890 nm (mode C).
It is important to note that these reflectance dips are closely related to the geometrical and material properties of the designed CMH. In fact, considering an out-of-plane 3D pillars hypergrating characterized by the same sizes of the wire radius ($r$), or the helix radius ($R_h$) it does not leads to comparable signals usable for sensing (see Supplementary materials).
According to the grating coupling technique of surface plasmon excitation,[@sambles1991optical; @sreekanth2013experimental] the surface plasmon modes can be excited when the wavevector of the grating diffraction orders are greater than that of the incident light. Under this condition, diffraction orders are no longer propagating waves, but evanescent field and the enhanced wavevector results to be responsible for the coupling of the incident light to the surface plasmon modes according to the coupling condition:[@barnes2004surface]
$$k^2_{SPP}=n^2_0 k^2_0 sin^2 \theta \pm 2 n_0 m k_g k_0 sin \theta cos \phi + (m k_g)^2
\label{eq3}$$
where $n_0$ is the refractive index of the incident medium, $k_0$ is the vacuum wavevector, $m$ is the grating diffraction order and $\theta$ and $\phi$ are the incident grazing and azimuthal angle, respectively. $k_g = 2\pi/ \Lambda$ is the grating wavevector, with $\Lambda$ the grating period, in our case corresponding to the the lattice constant $a$. When $\phi =0$, Eq. \[eq3\] results to be: $k_{SPP}=n_0 k_0 sin \theta \pm m k_g$. By using this equation we calculated the corresponding $k_x$ for our three modes: $k_{xA}$,$k_{xB}$ and $k_{xC}$.
By considering the dispersion relations for the SPP and BPP modes given by:[@avrutsky2007highly] $$k_{SPP}= k_0 \sqrt{\varepsilon_d \varepsilon_m \over \varepsilon_d + \varepsilon_m}
\label{eq4}$$
$$k_{BPP_N}=k_0 \sqrt{\varepsilon_d - \frac{\lambda^2 N}{\pi^2 t_d t_m } \frac{\varepsilon_d }{\varepsilon_m}}
\label{eq5}$$
we can plot the SPP and BPP dispersion curves (Figure \[fig:3\]a) supported by the structure. In the above equations, $t_d$, $\varepsilon_d$ and $t_m$, $\varepsilon_m$ are the thickness and dielectric permittivity of the dielectric and metal, in our case ITO and Ag, respectively, while $N$ represents the order mode ($N$=1,2,3...). In this plot it is possible to identify three points corresponding to the intersection between the $k_x$ component of modes A, B and C and the BPP modes of the structure, Figure \[fig:3\]a.
Since both surface and bulk plasmon mode excitation depends on the incident angle, we calculated the reflectance of the CMH-HMM for different angle of incidence (Figure \[fig:3\]b). We observed that there is a blue shift in the reflectance minima when the incident angle is increased. This is attributed to the variation in modal indices with incident angles.[@fan2006all]
![(a) Dispersion relations of SPP and BPP modes and (b) calculated reflection spectra for three different incident angles ($\theta_i$) for the CMH-HMM.[]{data-label="fig:3"}](Fig3.png){width="0.7\columnwidth"}
The reported data clearly show that the CMH can excite the BPP modes of the underneath hyperbolic metamaterial. These modes are expected to undergo a frequency shift as a function of the analytes concentration binding at the metasurface, defining the sensitivity of the sensing platform. To this end we use the experimental data of Koohyar et al.[@koohyar2012study] who tabulated the refractive index (RI) variations related to different molar fractions of aqueous solutions of methanol, ethanol, ethylene glycol, 1-propanol and 1, 2, 3-propantriol. In particular, we consider the data corresponding to the aqueous solution of 1, 2, 3-propantriol characterized by an ultra low molecular weight ($C_3H_8O_3 \approx$ 60 Da).
Figure \[fig:4\]a shows the calculated reflection spectra of the CMH-HMM sensor in measuring 1, 2, 3-propantriol solutions with different concentrations. The shift of the resonances as a function of the different percentage of propantiol in the aqueous solution allows us to determine the sensitivity of the sensor. The reflectance minima (mode dips) shift towards longer wavelengths with the increase of RI, in the range $0 <\Delta_{RI} <0.0680$, this last case corresponding to a molar fraction of $C_3H_8O_3$ of about 17% in the aqueous solution; the resonance wavelength shift ($\Delta\lambda$) shows a linear behaviour with the RI change ($\Delta_{RI}$), as reported in Figure \[fig:4\]b.
![(a) Performance evaluation of the CMH-HMM sensor: calculated reflectance spectra with different mole fractions of 1, 2, 3-propantriol in distilled water. (b) Resonance wavelength shift for the mode A, mode B and the third mode C as a function of the refractive index variation and corresponding linear fitting. (c)X-Z maps of the electric field intensity through the CMH-HMM structure in logarithmic scale for the three modes.[]{data-label="fig:4"}](Fig4.png){width="0.9\columnwidth"}
The spectral sensitivity of the sensor is defined as $S= \Delta n/\Delta \lambda$, where $\Delta n$ represents the change of the refractive index solution flowing on top of the sensor and $\Delta\lambda$ the shift of the sensor resonance in $nm$.
In particular, $\Delta \lambda_A$=20 nm and $S_A$=3.3 10$^{-3}$ RIU/nm for the first mode A (635 nm), $\Delta \lambda_B$=26 nm and $S_B$=2.5 10$^{-3}$ RIU/nm for the second mode B (710 nm) and $\Delta \lambda_C$=37 nm and $S_C$=1.8 10$^{-3}$ RIU/nm for the third mode C (890 nm). The corresponding limit of detection (LOD), corresponding to the minimum RI variation that can be distinguished (shift of 1 nm) results to be equal to 0.0015. The difference in sensitivity for the three modes is related to their different $\Delta \lambda$. In particular, mode C is the one that shows the best result in terms of sensitivity respect to the other two modes, this is due to the highly confined field distributions on the superstrate at hyperbolic dispersion (Figure \[fig:4\]c) and to the fact that the transverse decay of the field in the HMM strongly varies from one mode to another.
Another important aspect related to the 3D hypergrating is represented by the significantly increase of the out-of-plane sensing surface, where the specific binding of the analytes occurs via functional groups immobilized in this specific case on a helical surface. At the same time, a chiral structure can modify the fluid dynamics around it, inducing an increase of the probability of a specific binding.
![(a) Sketch of the simulated geometry with different percentage of the surface coverage and (b) corresponding reflectance spectra. (c) Resonance wavelength shift for the mode A, B and C as a function of the variation of the surface coverage. []{data-label="fig:5"}](Fig5.png){width="0.7\columnwidth"}
Clearly, the wavelength shift of BPP modes is strongly related to the quantity of molecules that bind selectively on the surface of the helices. To demonstrate this behavior numerically, we have considered different percentages of surface coverage of the helicoidal structures, starting from the HMM surface (Figure \[fig:5\]a), where the maximum variation of the considered refractive index ($\Delta_{RI}$=0.0680) affects only particular regions around the helix. To this end, we have considered the two limiting cases: the 0% of surface coverage (s.c.) and the 100%, in which the whole surface of the helix is totally covered by molecules; in this last case we obtain the maximum shift of 31 nm for mode C, 21 nm for mode B and 15 nm for mode A. Then we considered four intermediate cases of partial s.c.: 25%, 33%, 50%, and 75% (Figure \[fig:5\]b). As we can see in Figure \[fig:5\]c the resonance wavelength shift of the helix/HMM modes shows a linear behaviour until 50% of surface coverage, once this value is exceeded the slope of the curve decreases, by showing that the sensitivity is decreasing too.
At the same time, it is possible to find the minimum percentage value, that in our case results to be equal to 16%, corresponding to the minimum surface coverage, necessary to have an appreciable shift of the modes. This leads to calculate the mass sensitivity, a key parameter used to describe the performance of a biosensor, that is strongly related to the surface coverage; for the considered system the mass sensitivity results equal to 0.04 $pg/mm^2$ (see Supplementary materials).
![(a)Sketch of the geometry with the surface coverage (s.c.) of 20% at two different heights. (b) Corresponding simulated reflectance spectra and (c) a magnification of them in proximity of the mode A, B and C. []{data-label="fig:6"}](Fig6.png){width="0.6\columnwidth"}
The last analysis is dedicated to investigate how the sensitivity changes as function of the relative surface coverage at different heights away from the HMM surface. In particular, it is possible to demonstrate that the local change of refractive index in a small disk surrounding the helix produces appreciable shifts even when the binding is confined exclusively in the upper region of the helix, which represents the maximum distance from the HMM. For this reason a small disk, with $n$=1.401 and corresponding to an adsorbed surface of 20%, is positioned at different distance ($D$) respect to the surface of the HMM (Figure \[fig:6\]a). By plotting the resonance shift of the three modes A, B and C as a function of the distance $D$ normalized to the helix height ($h$) it is possible to see that the proposed biosensing platform is able to appreciate a shift of the considered modes even in the worst case ($D/h$=1) of about 1.4 nm for the mode A, 3.6 nm for the mode B and 3.0 nm for the mode C, as reported in Figure 6b. It is interesting to note that the resonance shift decreases exponentially by moving away from the surface of the HMM. The corresponding sensitivity results to be 48.6 10$^{-3}$RIU/nm for the mode A, 18.9 10$^{-3}$RIU/nm for the mode B and 22.6 10$^{-3}$RIU/nm for the mode C. These results demonstrate the strong sensing power of our innovative system to detect target analytes also away from the surface of the HMM exploiting the increased surface/volume ratio of the 3D CMH exposed to the analyte.
On the other hand, considering the intrinsic chirality of the nanohelices, CMH could excite new BPP modes of the HMM by coupling with their circular polarization-dependent plasmon modes. Indeed, as we can see in Figure \[fig:7\]a, for an angle of incidence of $\theta_i$=75$^\circ$, different reflectance dips are obtained for left-handed circular polarized (LCP) and right-handed circular polarized (RCP) light. In particular, the reflectance dips obtained for LCP light, indicated in the figure as BPP$_3$ to BPP$_6$, are strongly related to a coupling between the plasmonic modes of the gold helix array and the HMM underneath. This can be verified through the intersection of the dispersion curves of the HMM with the wavelength corresponding to the reflectance dips that are expressed in terms of energy versus momentum and are indicated in the Figure \[fig:7\]b with the blue triangles. For LCP light, at these wavelengths, respectively: 597 nm, 713nm, 796 nm and 873 nm the electric field intensity |E| results to be confined between the helix antenna surface and the HMM layers - Figure \[fig:7\]c. A similar analysis could be carried out for RCP reflectance dips.
![(a)Reflectance curves of the CMH-HMM for LCP and RCP light, angle of incidence $\theta_i$=75$^\circ$; in the inset a sketch of the simulated unit cell probed with circular polarized light. (b) Intersection of the dispersion curves of the HMM with BPP modes excited by circular polarized light. (c) X-Z maps of the electric field intensity through the CMH-HMM structure in logarithmic scale for LCP light and $\theta_i$=75$^\circ$. (d) Reflectance circular dichroism (RCD) as a function of refractive index variations.[]{data-label="fig:7"}](Fig7.png){width="0.8\columnwidth"}
From the analysis of the chiroptical response of the CMH we can calculate the CD signal, typically riches of spectral signatures, such as bipolar peaks and crossing points, increasing the sensitivity and accuracy of the monitoring refractive index changes due to the analyte absorption. This sensing modality offers strong optical contrast even in the presence of highly achiral absorbing media, increasing the signal-to-noise CD measurements of a chiral analyte, an important consideration for use in complex biological media with limited transmission.[@jeong2016dispersion] For this purpose, the resulting reflectance curves obtained for LCP and RCP light, at $\theta_i$=75$^\circ$ are used to calculate the reflectance circular dichroism (RCD) spectrum, which characterizes the reflectance difference between LCP and RCP light, leading to a RCD amplitude (RCD=R$_{LCP}$-R$_{RCP}$). As expected, the RCD spectra for the CMH-HMM exhibit multiple features: different maxima, minima and crossing points (Figure \[fig:7\]d). We have studied the RCD, as a function of the refractive indices variation between 1.333 and 1.401 (glycerol - water mixtures varying from 0 to 17.5% concentration - Figure \[fig:7\]d) to quantify the performance in terms of sensitivity and FOM of our sensor.
We can distinguish in the range 700-900 nm a minimum ($\lambda_m$), a crossing point ($\lambda_0$) and a maximum ($\lambda_M$) respectively at 797 nm, 816 nm and 852 nm which significantly shift and modify their intensity as the refractive index changes. These signals show a chiral plasmon sensitivity S$_{CP}$=$\Delta n/\Delta\lambda$ of 2.42 10$^{-3}$RIU/nm for $\lambda_m$, 2.12 10$^{-3}$RIU/nm for $\lambda_0$ and 2.19 10$^{-3}$RIU/nm for $\lambda_M$, respectively. After extracting the classical FWHM for the $\lambda_m$ and $\lambda_M$ modes and the FWHM in the |RCD| spectrum (see Supplementary materials) for $\lambda_0$, we have calculated the FOM,[@mayer2011localized] defined in our case as $FOM$=$(S \cdot FWHM)^{-1}$, that resulted to be, respectively, equal to 18.72, 15.18 and 15.19, very high values if compared to previously reported results.
Conclusions
===========
In this paper, we have numerically demonstrated how an out-of-plane chiral metasurface hypergrating can be used to excite the plasmonic modes of an underlying hyperbolic metamaterial with linearly and circularly polarized light. The electromagnetic modes resulting from the coupling between the CMH-HMM depend strongly on the material and geometrical parameters of the CMH-HMM. In our numerical study, we have demonstrated how the HMM BPPs modes can be used for the detection of ultra low molecular weight molecules ($\approx$ 60 Da) with extremely low concentrations (mass sensitivity of 0.04 $ pg / mm ^ 2 $) on the whole surface of the helices array. The CMH represents a new versatile way to couple circularly polarized light with the HMM, by exploiting the features of the reflectance circular dichroism (RCD) spectra, rich of signals that can be used to increase the processed data for the detection of analytes by the proposed biosensing platform.
This research has been supported by the *“AIM: Attraction and International Mobility”* - PON R&I 2014-2020 Calabria and “Progetto Tecnopolo per la Medicina di precisione, Deliberazione della Giunta Regionale n. 2117” Puglia. The authors thank the Area della Ricerca di Roma 2, Tor Vergata, for the access to the ICT Services (ARToV-CNR) for the use of the COMSOL Multiphysics Platform, Origin Lab and Matlab, and the Infrastructure BeyondNano (PONa3-00362) of CNR-Nanotec for the access to research instruments.
Contents:
- Electric permittivity of the HMM calculated by means of the effective medium theory
- Equations for Scattering parameters $S_{11}$ and $S_{21}$
- Meshed geometry
- Unchiral hypergratings
- Calculation of the mass sensitivity
- Detailed plots for the FOM calculation
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[ For example, the low-energy theory of the two-dimensional (2D) toric code model [(i.e. the deconfined phase of $\mathbb{Z}_2$ gauge theory)]{} is a $U(1)\times U(1)$ Chern-Simons theory in which gauge charges (i.e., $e$ and $m$ particles) are deconfined and the gauge fields are gapped, while the confined phase is topologically trivial. In this paper, we point out a new route to [constructing exotic 3D gapped fermionic]{} phases [in a confining phase of a gauge theory]{}. [Starting from a parton construction with strongly fluctuating compact $U(1)\times U(1)$ gauge fields, we construct gapped phases of interacting fermions]{} by condensing two linearly independent bosonic composite particles [consisting]{} of partons and $U(1)\times U(1)$ magnetic monopoles. [This]{} can be regarded as a 3D generalization of the 2D Bais-Slingerland condensation mechanism. Charge fractionalization [results]{} from a Debye-Hückel-like screening cloud formed by the condensed composite particles. [Within our general framework, we explore two aspects of symmetry-enriched]{} 3D Abelian topological phases. First, [we construct a new fermionic state of matter with time-reversal symmetry and $\Theta\neq \pi$, the fractional topological insulator]{}. Second, we generalize the notion of *anyonic symmetry* of 2D Abelian topological phases to the *charge-loop excitation symmetry* ($\mathsf{Charles}$) [of]{} 3D Abelian topological phases. [We show that line twist defects, which realize $\mathsf{Charles}$ transformations, exhibit non-Abelian fusion properties.]{}]{}'
author:
- Peng Ye
- 'Taylor L. Hughes'
- Joseph Maciejko
- Eduardo Fradkin
title: '[Composite Particle Theory of Three-dimensional Gapped Fermionic Phases: Fractional Topological Insulators and Charge-Loop Excitation Symmetry ]{}'
---
Introduction
============
Background and overview: Parton construction and gauge confinement
------------------------------------------------------------------
In models of non-interacting fermions, several topological phases of matter have been found, such as integer quantum Hall states (IQH), Chern insulators, and topological insulators (TI)[@iqh; @haldane88; @TI1; @TI2; @TI3; @TI4; @TI5; @TI6; @TIexp]. Owing to the non-interacting nature of the problem, tremendous progress has been made in both theory and experiment. In the presence of weak interactions these phases can also be analytically understood. If interactions are strong enough such that a perturbative analysis is no longer meaningful, one usually faces a problem in strongly-correlated electron physics. While exact solutions are possible in a few specific models, one often constructs approximate effective descriptions of such systems. One such approach is the parton construction approach, also known as the projective construction, or slave-particle approach. It has been widely applied in studies of strongly correlated electron systems such as high-temperature superconductors and fractional quantum Hall states (FQH).[@BZA8773; @BA8880; @AM8874; @KL8842; @SHF8868; @AZH8845; @DFM8826; @WWZcsp; @Wsrvb; @LN9221; @MF9400; @WLsu2; @WenRMP; @weng; @ye10; @ye11a; @ye_zhang; @ye_ma; @ye_wang; @jain_parton; @wen_parton; @Wen99; @Barkeshli2010] Recently, it has also been applied [@LL1263; @YW12; @YW13a; @Wangsenthil2015; @Ye14b] to bosonic symmetry-protected topological phases (SPT) as well.[@1DSPT; @Chenlong; @Chen_science; @Chen10]
Generally speaking, in the parton construction we start with a lattice action $S$ that describes a strongly-correlated electronic system. In this paper, the electron operator $c$ is meant to represent a generic Grassmann variable that is the only dynamical variable in $S$. We further write the electron operator $c$ in terms of several parton operators $f^i$. The Hilbert space for $S$ is equivalently replaced by a projected Hilbert space formed by partons and gauge fields. In practice, there are many different kinds of parton constructions. We focus on one of them, where all partons are fermionic such that an odd number of partons is required to form an electron. Mathematically, the electron operator is formally fractionalized as $c=f^1f^2\cdots f^{2n+1}$. The electron operator $c$ is a singlet of the $SU(2n+1)$ gauge group. The largest totally commuting subgroup, or maximal torus, is given by the compact Abelian $(U(1))^{2n}$ gauge group which acts with gauge transformations: $f^1\rightarrow f^1e^{i\theta_1},\cdots, f^i\rightarrow f^ie^{i\theta_{i}-i\theta_{i-1}},\cdots, f^{2n+1}\rightarrow f^{2n+1}e^{-i\theta_{2n}}$, where $\{\theta_i\}$ ($i=1,2,\cdots,2n$) are arbitrary functions of the lattice sites and a continuous time variable. As such, by applying the ’t Hooft gauge projection,[@Gauge_Confinement_tHooft3] the lattice action deep in the confined phase is reformulated to describe a system of interacting partons and $2n$ dynamical compact abelian gauge fields $\{a^{(i)}_\mu\}$.
It should be noted that the gauge-field coupling constants $g_i$ at the lattice scale should be treated as being very strong since the usual lattice kinetic terms (with coefficient $1/g^2_i$) for compact gauge fields are not present in $S$. From here one can usually proceed further by assuming a mean-field theory of partons where the effects of gauge fluctuations are assumed negligible. As such, a very important feature—the compactness of gauge fields—is totally ignored. A standard perturbative analysis can be applied in order to quantitatively recover the effect of gauge fluctuations at leading order. In some cases, this assumption is legitimate. A typical example is a 2D system where fermionic partons occupy energy bands with non-zero Chern number at the mean-field level.[@YW12] In this case, a Chern-Simons term is generated and a topological mass gap[@csgap] for the gauge fields is produced, which suppresses instanton tunneling. However, there is no reason to rule out the possibility that $g$ at low energies flows to strong coupling, such that the compactness of the gauge fields plays a fundamental role in reshaping the nature of the emergent ground states. In such cases, mean-field theories of partons fail to describe the physical states formed by electrons, even at a qualitative level.
Despite the strong coupling nature of the problem, the leading order effect of gauge fluctuations can still be perturbatively treated by considering the Bose-Einstein condensation of composite particles. In the regime of strong gauge coupling, condensed composites contain magnetic monopoles of the internal compact gauge fields (and possible electric gauge charge as well). Historically, this line of thinking was developed in the context of strongly coupled gauge theories. For example, condensed monopole phases are relevant for studies of (3+1)D compact quantum electrodynamics, the Georgi-Glashow models, and supersymmetric Yang-Mills theory.[@mandelstam; @mt78; @Gauge_Confinement_Polyakov_1; @Gauge_Confinement_Polyakov_2; @Gauge_Confinement_Polyakov_3; @Gauge_Confinement_tHooft1; @Gauge_Confinement_tHooft2; @Gauge_Confinement_susskind; @Gauge_Confinement_Fradkin; @Gauge_Confinement_tHooft3; @Gauge_Confinement_proceedings; @susy] For an Abelian gauge theory with a compact $U(1)$ gauge group, the monopole creation operator has been constructed explicitly and gains a nonzero vacuum expectation value as shown in Ref. . Recently, it was further suggested that the behavior of the non-trivial line operators may be used to make the proper distinction between confinement phases of strongly coupled gauge theories.[@seiberg_order_2013] Note that since electric (i.e., gauge) charge excitations are linearly confined during this condensation process, the monopole condensation phase is also called the confinement phase.
Unfortunately, the usual monopole condensation scenario, [@mandelstam; @mt78; @Gauge_Confinement_Polyakov_1; @Gauge_Confinement_Polyakov_2; @Gauge_Confinement_Polyakov_3; @Gauge_Confinement_tHooft1; @Gauge_Confinement_tHooft2; @Gauge_Confinement_susskind; @Gauge_Confinement_Fradkin; @Gauge_Confinement_tHooft3; @Gauge_Confinement_proceedings; @susy] when applied to the 3D parton construction, will simply confine all partons back into electrons, [resulting in]{} a non-fractionalized trivial insulator. In this sense, the parton construction with gauge confinement [driven]{} by the usual monopole condensation [does not seem to be a good pathway to reach topological states]{}. To save the parton construction approach, we should look for new scenarios of gauge confinement. More precisely, can we have a new kind of condensation that confines partons while still [leading to]{} a fractionalized insulator? How can we imagine the existence of fractionalized excitations when partons are confined? If these questions can be solved, a new systematic treatment of 3D fermionic fractionalized phases will be established. [This is the main goal of this paper, and that this is possible can be gleaned from the success of the 2D Bais-Slingerland condensation mechanism]{}.[@bais2009] Indeed, we show that there are new pathways to fractionalization in 3D, now in the [*confining regime*]{} of the gauge theory, provided that confinement occurs as the result of condensation of a class of composites made of fermionic partons and monopoles (from different sectors of the gauge group). We will see that this new form of [*oblique confinement*]{}[@Gauge_Confinement_tHooft3] leads to unexpected phases of matter, particularly states with fractionalized $\Theta$ angles and yet compatible with time-reversal invariance. [While oblique confinement as a pathway to topological phases has been considered before in the context of bosonic phases of matter,[@vonkeyserlingk2015] here we focus on topological phases of interacting fermions.]{} In addition, we also study several applications. For example, how can we impose symmetry in such a parton construction with gauge confinement? The latter leads to the notion of symmetry-enriched topological phases (SET) in the parton construction approach.
[Summary of main results]{}
----------------------------
[***(1)** Composite particle theory of fermionic phases*]{}. In this article, we will consider the condensation of “composites” that not only carry magnetic charges but also contain fermionic partons that are charged under different internal gauge fields \[i.e., $U(1)\times U(1)$ strongly fluctuating gauge fields in our concrete example $c=f^1f^2f^3$\] and under the external electromagnetic (EM) field $A_\mu$. One caveat is that, despite the mixture of partons and magnetic monopoles, those condensed composites are *not* dyons. More concretely, they carry either electric charge or magnetic charge in a given gauge group, not both. This fact allows us to make reliable statements and calculations from a *local* theory. All excitations can be organized as a set of *charge-loop composites*, and, as a whole, form a *charge-loop-lattice* in which each lattice site corresponds to a deconfined excitation. Especially, partons are confined as usual but some composites constituted by partons and magnetic monopoles of $U(1)\times U(1)$ gauge fields may be deconfined and carry fractionalized EM electric charge. Many universal physical properties can be easily determined from the charge-loop-lattice, such as the braiding statistics between point-like excitations and loop excitations, the self-statistics of point-like excitations, the EM charge of the excitation, and the bulk axion $\Theta$ angle.[@witten1; @Qi2008] We will refer to this approach to constructing 3D fermionic gapped phases as a *composite particle theory*.
In this theory, charge fractionalization is achieved via a Debye-H[ü]{}ckel-type charge-screening cloud formed by the composite condensates. This is analogous to the charge-screening phenomenon in the composite fermion theory of the FQH effect.[@composite_Jain_1; @composite_Jain_2; @composite_Fradkin; @composite_HLR; @composite_sm1; @cmposite_H_2; @composite_sm2] In fact, we prove that the Debye-H[ü]{}ckel-type screening is the unique source of charge fractionalization. In principle, all physical quantities of the resulting phases can be expressed as functions of a set of parameters that characterize composite particle theory. This line of thinking is also analogous to the composite fermion theory of FQH states where the filling fraction is unified in a sequence of discrete numbers, each of which corresponds to a specific ansatz in the composite fermion construction. From this perspective, the composite particle theory may be regarded as an attempt to find a 3D analog of the composite fermion theory of FQH states[, with the caveat that we are considering confined phases while in the composite fermion theory, all gauge fields are deconfined. ]{}
[***(2)** Fractional topological insulators*]{}. Based on the composite particle theory, we will study two symmetry-enriched properties. The first property is the bulk axion angle $\Theta$ in the presence of time-reversal symmetry.[@witten1; @Qi2008] When $\Theta$ is non-vanishing an externally inserted EM monopole with integral magnetic charge $M$ will induce an electric polarization charge $\frac{\Theta}{2\pi}M$, a phenomenon known as the Witten effect.[@witten1; @Qi2008; @franz] For free-fermion topological insulators, the $\Theta$ angle is $\pi \text{ mod } 2\pi$ with $M\in\Z$.[@Qi2008] However, it was theoretically proposed that $\Theta$ could be different from $\pi$ if strong interactions and correlations are taken into account,[@maciejkoFTI; @maciejko_model; @maciejko2015; @swingle2011; @swingle_fti_2012] leading to the notion of 3D fractional topological insulators (FTI) with deconfined gauge fields. The periodicity of $\Theta$ should also be properly modified so as to preserve time-reversal symmetry.
[In Ref. , [FTI]{}s were obtained via parton constructions where the internal gauge fields are in the Coulomb phase (photons are gapless). Therefore, a gapless channel, despite of its electric neutrality, can in principle adiabatically connect the FTIs to a fractionalized state with vanishing axion angle. In Ref. , bosonic FTIs were obtained via parton constructions where the mean-field Hamiltonian of partons explicitly breaks SU(2) gauge group down to $\Z_2$ discrete gauge group. As a result, the unbroken discrete gauge group leads to bulk topological order and deconfined fractionalized excitations. The gauge fluctuations in both Ref. and Ref. are perturbatively weak. In the present work, we explore the possibility of realizing FTIs via the condensation of composites (introduced above) when gauge fluctuations are sufficiently strong and gauge confinement occurs. ]{}
[ In Ref. and Ref. , each parton is assumed to carry a fractional EM electric charge such that a fractionalized $\Theta$ angle should be expected (by simply noting that the coefficient of $F\wedge F$ has unit of $e^2$). This is not the case in our work. We show that even if partons carry integral EM electric charge (i.e., both $f^1$ and $f^2$ carry $+1$ electric charge and $f^3$ carries $-1$ eletric charge, see Sec. \[sec:bec\] for more details), a fractionalized $\Theta$ and gapped bulk can also be achieved as long as a proper composite condensation is considered and partons occupy non-trivial topological insulator bands. This feature is unique in the parton construction with gauge confinement.]{}
[ In the [FTI]{} state constructed in this work (Sec. \[sec\_FTI\], \[sec\_FTI\_plus\]), we show that the EM electric charge of deconfined excitations is fractionalized at $1/3$. This is consistent to the claim by Swingle, *et. al.*[@swingle2011] that an FTI necessarily has a fractionalized bulk. Indeed, the fractionalization nature of the FTI state in the present work can be traced back to the presence of $\Z_2\times \Z_6$ topological order (see Sec. \[sec\_FTI\_plus\] for more details). The latter arises from the deconfined discrete subgroup of the confined $SU(3)$ gauge group.]{}
[***(3)** Charge-loop excitation symmetry and extrinsic twist defects*]{}. Noting that the set of all excitations forms a charge-loop-lattice, the second symmetry-enriched property is the concept of “charge-loop excitation symmetry”, abbreviated as $\mathsf{Charles}$ (see Definition \[dfn\_charlos\]). $\mathsf{Charles}$ can be viewed as a hidden symmetry of (3+1)D topological quantum field theories. Meanwhile, $\mathsf{Charles}$ has a geometric interpretation as a point-group symmetry of the charge-loop-lattice that preserves physical properties of the excitations. The study of $\mathsf{Charles}$ is motivated by the theory of anyonic symmetry [@Teo2015; @Teo2014; @ran; @barkeshli_wen; @Teo2013; @Barkeshli2014; @You2012; @genon_1; @genon_2; @genon_3; @genon_4; @Bombin2010] and its relation to extrinsic twist defects of 2D Abelian topological phases. We expect that 3D Abelian topological phases where charge-loop composite excitations are allowed may host even more exotic physics if extrinsic defects are considered.
Physically, extrinsic defects (which may come in the form of vortices or disclinations, for example) are semi-classical objects that are externally imposed into a 2D topological phase. An extrinsic *twist* defect is one which may be associated with an element of an anyonic symmetry group that acts to permute the set of anyons. The inclusion of such defects enriches the tensor category theory of the Abelian parent topological phase. Indeed, this line of thinking has attracted a lot of attention since extrinsic twist defects can bind non-Abelian objects even though all of excitations of the parent topological phase without defects are Abelian.[@genon_1; @genon_2; @Bombin2010; @genon_3; @genon_4; @You2012] A typical example is found in some lattice systems exhibiting $\mathbb{Z}_N$ topological order that contain, for example, the $\Z_N$ charge and flux anyons $e$ and $m$.[@Bombin2010; @You2012] In these cases the anyonic symmetry is intertwined with a lattice translation symmetry such that a dislocation defect acts to exchange the $e$ particle-type with the $m$ particle-type when they orbit around the defect. This implies that the defect harbors a rich internal (non-Abelian) structure so that it can convert between the anyon types. In the present work we propose $\mathsf{Charles}$ as a 3D version of anyonic symmetry in 2D. In analogy to 2D, each extrinsic defect in 3D is also associated with a $\mathsf{Charles}$ group element. We also study defect species and some defect fusion properties (see Fig. \[figure\_defect\_composite\]).
The remainder of the paper is organized as follows. Sec. \[sec:gauge\_strc\] is devoted to a general discussion of the parton construction and composite condensation. In Sec. \[sec:theta\], [FTI]{}s are constructed from a composite condensation phase where all of the partons occupy topological insulator bands. A concrete example with time-reversal symmetry and fractional $\Theta=\frac{\pi}{9}\text{ mod }\frac{2}{9}\pi$ is shown (see Fig. \[figure\_theta\], Sec. \[sec:theta\]). As a comparison, we also show a parton construction in the Coulomb (gapless photon) phase using a perturbative approach, which leads to $\Theta=\pi\text{ mod }2\pi$ and two gapless neutral modes in the bulk. In Sec. \[sec:charlos\_defect\_symmetry\], the charge-loop excitation symmetry ($\mathsf{Charles}$) of the charge-loop excitations, and its relation to 3D extrinsic defects, is studied. Sec. \[sec:conclusion\_direction\] is devoted to the conclusion and future directions. Many key notations, mathematical formulae, and terminologies are introduced in Sec. \[sec:gauge\_strc\], which provides the preliminaries for the subsequent parts. Several technical details can be found in the Appendices. In Appendix \[appendix:notation\], several notations and abbreviations are collected.
Composite particle theory in three dimensions: A general discussion {#sec:gauge_strc}
===================================================================
Compact $U(1)\times U(1)\times U(1)$ gauge symmetry of composites {#sec:bec}
-----------------------------------------------------------------
![(Color online) Parton construction of electron operators in the present work. The wavy lines denote interactions mediated by gauge bosons. $A_\mu$ is the external non-dynamical EM (electromagnetic) field, serving as a probe of the electromagnetic response of the system. $a_\mu$ and $b_\mu$ are two dynamical, compact $U(1)$ gauge fields, belonging to the $U(1)_a$ and $U(1)_b$ gauge groups respectively. The partons $f^1$ and $f^2$ carry $1$ and $-1$ gauge charges of the $U(1)_a$ gauge group respectively. The partons $f^3$ and $f^2$ carry $1$ and $-1$ gauge charges of the $U(1)_b$ gauge group respectively. The EM electric charges carried by the partons $f^1$, $f^2$, $f^3$ are $e$, $e$, $-e$, respectively.[]{data-label="figure_newelectron"}](figure_newelectron.pdf){width="8.5cm"}
In the simplest fermionic parton construction, the electron operator is decomposed into three fermionic partons: $c=f^1f^2f^3$ (Fig. \[figure\_newelectron\]), where both $f^1$ and $f^2$ carry unit EM charge $e$ while $f^3$ carries $-e$. As a result, the electron carries $e$. In a 3D mean-field theory where the partons are deconfined, we consider that all fermionic partons have a gapped spectrum and form either a trivial band insulator ($\theta=0$) or a strong topological insulator ($\theta=\pi$) simultaneously, where $\theta$ denotes the axion angle of the partons [@witten1; @Qi2008]. To avoid confusion, we will use the capital letter $\Theta$ to denote the axion angle of the electron, which will be calculated in detail in Sec. \[sec:theta\]. The internal gauge group is $SU(3)$ whose maximal torus (maximal commuting subgroup) $U(1)_a\times U(1)_b$ is sufficient to capture the confinement phase properties due to the ’t Hooft gauge projection [@Gauge_Confinement_tHooft3]. Here, both $U(1)$ factors are compact gauge groups that support magnetic monopoles [@Gauge_Confinement_tHooft3]. $U(1)_a$ corresponds to the gauge field $a_\mu$ that glues $f^1$ and $f^2$ together, while $U(1)_b$ corresponds to the gauge field $b_\mu$ that glues $f^2$ and $f^3$ together (Fig. \[figure\_newelectron\]). Adding the EM gauge group with gauge field $A_\mu$, the total gauge group is given by $U(1)_a\times U(1)_b\times U(1)_{\rm EM}$. It should be noted that $A_\mu$ is a non-dynamical (i.e., background) gauge field, which is useful for diagnosing the EM linear response properties of the resulting phases. For the same reason, we consider monopole configurations of $A_\mu$ (with magnetic charge $M$) as externally imposed background configurations.
Alternatively, one may also define the following three gauge fields: $$\begin{aligned}
A^{f1}_\mu=a_\mu+A_\mu , A^{f2}_\mu=-a_\mu-b_\mu+A_\mu , A^{f3}_\mu=b_\mu-A_\mu\,,\nonumber
\end{aligned}$$ where $A^{fi}$ is the gauge field that only couples to $f^i$ ($i=1,2,3$). The relation between the two sets of gauge fields can be expressed in matrix form: $$\begin{aligned}
\left(\begin{matrix}
A_\mu\\
a_\mu\\
b_\mu\end{matrix}\right)=\left(\begin{matrix}
1&1&1\\
0&-1&-1\\
1&1&2
\end{matrix}\right)
\left(\begin{matrix}
A_\mu^{f1}\\
A_\mu^{f2}\\
A_\mu^{f3}
\end{matrix}\right)\,,\label{transtion_matrix_gauge}\end{aligned}$$ where the matrix is integer-valued and invertible, i.e., belongs to the $\mathbb{GL}(3,\Z)$ group. We now turn to the description of generic composite particles, which are labeled by a set of electric charges and magnetic charges. We use $N_{a,b}$ to denote the electric charge of the $U(1)_{a,b}$ gauge group and $N_m^{a,b}$ to denote the magnetic charge of that same gauge group. We use $N_A$ and $M$ to denote the bare electric and magnetic charges in the EM gauge group, and $N^{fi}$ and $N^{fi}_m$ to denote the electric and magnetic charges of the $U(1)_{fi}$ gauge groups.
Due to Eq. (\[transtion\_matrix\_gauge\]), the magnetic charges transform as: $$\begin{aligned}
&N^{f1}_m=N^a_m+ M,
N^{f2}_m=-N^a_m-N^b_m+M, \nonumber\\
&N^{f3}_m=N^b_m-M,\nonumber\end{aligned}$$ and electric charges transform as $$\begin{aligned}
&N_A=N^{f1}+N^{f2}-N^{f3},
N_a=N^{f1}-N^{f2}, \nonumber\\
&N_b=N^{f3}-N^{f2}.\nonumber
\end{aligned}$$ For convenience, we can easily derive the following useful formulae: $$\begin{aligned}
&M=N^{f1}_m+N^{f2}_m+N^{f3}_m,
N^a_m=-N^{f2}_m-N^{f3}_m,\nonumber\\
&N^b_m=N^{f1}_m+N^{f2}_m+2N^{f3}_m.\nonumber\end{aligned}$$ To summarize, a composite particle can be uniquely labeled by six numbers (three electric charges and three magnetic charges). The above relations can be recast in matrix form, $$\begin{aligned}
&\left(\begin{matrix}
N_A\\
N_a\\
N_b
\end{matrix}\right)=\left(\begin{matrix}
1&1&-1\\
1&-1&0\\
0&-1&1\\
\end{matrix}\right)
\left(\begin{matrix}
N^{f1}\\
N^{f2}\\
N^{f3}
\end{matrix}\right)\,,\label{eqn:transition_matrix}
\\
&\left(\begin{matrix}
M\\
N_m^a\\
N_m^b
\end{matrix}\right)=\left(\begin{matrix}
1&1&1\\
0&-1&-1\\
1&1&2
\end{matrix}\right)
\left(\begin{matrix}
N^{f1}_m\\
N^{f2}_m\\
N^{f3}_m
\end{matrix}\right)\,,\label{eqn:transition_matrix1}\end{aligned}$$ where the two matrices belong to the $\mathbb{GL}(3,\Z)$ group. All magnetic charges take values in an integral domain, i.e., $M$, $N_m^a$, $N_m^b$, $N_m^{fi}\in\Z$, where $i=1,2,3$. However, we will soon see that this integral domain will be potentially restricted to a smaller domain if we only consider the deconfined *excitations* in the presence of a composite condensate. We will introduce the notion of excitations in Sec. \[section\_of\_screen\].
By the bare electric charge, we mean that $N_A$ is a naive count of the EM electric charge. In Sec. \[section\_of\_screen\], it will be shown that composite condensates will partially screen the charge, leading to a *net* EM electric charge $Q$ to be defined in Eq. (\[NE1\]). The electric charges $N^{fi}$ ($i=1,2,3$) are related to the number of attached fermions via the Witten effect formula: $$\begin{aligned}
&N^{fi}=n^{fi}+\frac{\theta}{2\pi}N^{fi}_m\,,\,\text{with }n^{fi}\in\Z\,.\label{Nnf1}\end{aligned}$$ The integer $n^{fi}$ counts the total number of fermions $f^i$ in the composite, and $\theta$ is determined by the $\mathbb{Z}_2$ index of a 3D time-reversal invariant topological insulator. If $\theta=0$, the partons occupy a trivial band structure; if $\theta=\pi$, the partons occupy a non-trivial topological insulator band structure. The defining domains of $N^{f1,f2,f3},N_a,N_b,N_A$ can be either integer or potentially half-integer, depending on $\theta$.
![(Color online) Schematic representation of composite particle condensation. Before condensation, the system is an electromagnetic plasma of composites in the $U(1)_\textrm{EM}\times U(1)_a\times U(1)_b$ gauge group. There are many composite particles (denoted by solid circles) including $\varphi_1$ and $\varphi_2$. There are also two gapless photons (denoted by $a$ and $b$ in the figure), indicating that the phase before condensation is a gapless Coulomb phase for both internal dynamical gauge fields. After condensing $\varphi_1$ and $\varphi_2$, the system enters a gapped phase in the absence of photons. All composites (denoted by black solid circles on the left) that have nonzero mutual statistics with both $\varphi_1$ and $\varphi_2$ are confined. Otherwise, those composites that have trivial (zero) mutual statistics with both condensates survive as excitations (denoted by blue solid circles) of the gapped phase. The red loops on the right represent loop excitations due to the two condensates.[]{data-label="figure_condensation"}](figure_condensation.pdf){width="8.5cm"}
A local field theoretic description of condensed composites
-----------------------------------------------------------
Since there are two internal gauge fields with strong gauge fluctuations, we can consider two linearly independent Bose condensates denoted by $\varphi_1$ and $\varphi_2$, as shown in Fig. \[figure\_condensation\]. Both condensates should contain magnetic monopoles of the internal gauge fields but be neutral under both the $U(1)_a$ and $U(1)_b$ gauge groups, i.e., $N_a=0\,,N_b=0$. Since the EM gauge field is treated as a background gauge field for the purpose of the EM response, the condensates should not carry $M$. Otherwise, the EM gauge field must be strongly fluctuating, which is not our working assumption. In summary, the electric and magnetic charges of $\varphi_1$ and $\varphi_2$ can be completely determined by six parameters $(q,u,v,q',u',v')$ in Table \[table:em\]. Since the condensates are not dyonic in each gauge group the order parameters $\langle \varphi_1\rangle$ and $\langle \varphi_2\rangle$ are local to each other and can be described by an effective local quantum field theory. More concretely, we may start with a phenomenological Ginzburg-Landau-type action in 4D Euclidean spacetime: $$\begin{aligned}
\! \! \! \!\! S_{\rm GL}\!=\!\int \!d^4x \sum^2_I\!\left(|\hat{D}_\mu\varphi_I |^2+\mu^2 |\varphi_I|^2\! +\! \lambda |\varphi_I|^4\right)+ \!S_M,\label{equation_free_energy}\end{aligned}$$ where the Ginzburg-Landau parameter $\lambda$ is positive. $\hat{D}_\mu$ is the covariant derivative defined by: $$\begin{aligned}
\hat{D}_\mu\equiv \partial_\mu+i N_A A_\mu+i N_m^a\tilde{a}_\mu+i N_m^b \tilde{b}_\mu.\end{aligned}$$ Here, $N_A,N^a_m,N^b_m$ are two sets of electric/magnetic charges of $\varphi_1$ and $\varphi_2$, which can be found in Table \[table:em\]. The one-form gauge fields $\tilde{a}_\mu$ and $\tilde{b}_\mu$ serve as the magnetic dual of the gauge fields $a_\mu$ and $b_\mu$, respectively. For example, $\tilde{a}_\mu$ is introduced such that its gauge charge is carried by magnetic monopoles of the $U(1)_a$ gauge group. Meanwhile, the magnetic flux of $\tilde{a}_\mu$ gives the electric field $\mathbf{E}^a$, namely, $\mathbf{E}^a=\nabla\times \mathbf{\tilde{a}}$. $\tilde{b}_\mu$ can be understood analogously to $\tilde{a}_\mu$. $S_M$ includes the Maxwell terms: $S_M=\int d^4x(\frac{1}{4}\tilde{f}^a_{\mu\nu}\tilde{f}^a_{\mu\nu}+\frac{1}{4}\tilde{f}^b_{\mu\nu}\tilde{f}^b_{\mu\nu})\,.$ In the condensed phase where the mass parameter $\mu^2<0$, nonzero expectation values $\langle\varphi_I\rangle\neq0$ develop. Here, $\tilde{f}^{a,b}_{\mu\nu}$ are the field strength tensors of $\tilde{a}_\mu$ and $\tilde{b}_\mu$. One advantage to using dual gauge fields is that the problem of strong gauge fluctuations ($g_a\gg 1\,,g_b\gg 1$) of the $a_\mu$ and $b_\mu$ gauge fields is transformed into the problem of weak gauge fluctuations of the dual gauge fields $\tilde{a}_\mu$ and $\tilde{b}_\mu$ by noting that the coupling constants between magnetic charges and dual gauge fields is the inverse of the original coupling constants, i.e., $1/g_{a,b}$.
It is noteworthy that the six numbers $(u,v,u',v',q,q')$ describing the condensates are not completely free since the following three conditions should be satisfied:
1. $\varphi_1$ and $\varphi_2$ are bosonic;\[condition\_phi2\]
2. $\varphi_1$ and $\varphi_2$ are allowed to condense simultaneously;\[condition\_phi3\]
3. $\varphi_1$ and $\varphi_2$ are linearly independent;\[condition\_phi1\]
such that the composite condensates $\varphi_1$ and $\varphi_2$ are physically viable. In more detail, according to the domains of definition of every charge (e.g., all magnetic charges are integer-valued, all $n^{fi}$ are integer-valued), we can deduce the domains of the six numbers (see Table \[table:em\]): $$\begin{aligned}
&u\,,v\,,u'\,,v'\,,q-\frac{\theta}{2\pi}u \,, q+\frac{\theta}{2\pi}(u+v) \,,q-\frac{\theta}{2\pi}v \in\Z,\label{constr_1plus}\\
& q'-\frac{\theta}{2\pi}u' \,,q'+\frac{\theta}{2\pi}(u'+v')\, ,q'-\frac{\theta}{2\pi}v' \in\Z\,.\label{constr_1}\end{aligned}$$ Since only bosonic particles can undergo Bose condensation, one should carefully check the self-statistics of $\varphi_1$ and $\varphi_2$. Furthermore, the mutual statistics between $\varphi_1$ and $\varphi_2$ must be zero so that they are allowed to condense simultaneously.
Let us first consider the latter. The trivial mutual statistics between two composites (with and without prime) is given by the following equation: $$\begin{aligned}
\sum^3_{i}(N^{fi}_m{n^{fi}}'-{N^{fi}_m}'n^{fi})=0\label{equation:mutual_stat}\end{aligned}$$ or equivalently: $\sum^3_{i}(N^{fi}_m{N^{fi}}'-{N^{fi}_m}'N^{fi})=0\,.$ If this equation is satisfied, then the two composites can condense simultaneously. Furthermore, condensation of one of the composites will lead to deconfined particles (an excitation spectrum) having electric and magnetic charges are determined by this equation. If the equation is not satisfied, the condensation of one of the composites will confine the other [@cardy1; @cardy2]. Inserting the electric and magnetic charges of $\varphi_1,\varphi_2$ into the equation, it turns out that $\varphi_1$ and $\varphi_2$ always satisfy the condition of trivial mutual statistics.
Next, we need to further check the self-statistics of $\varphi_1$ and $\varphi_2$. For a generic composite, the self-statistics phase $e^{i\pi \Gamma}$ is determined by the following integer: $$\begin{aligned}
\Gamma=\sum_i^3 (N_m^{fi}n^{fi}+n^{fi})\,,\label{eq:quantum_stat}\end{aligned}$$ where the second term $n^{fi}$ counts the number of fermionic partons inside the composite. The first term $N_m^{fi}n^{fi}$ arises from the angular momentum of the relative motion between the electric charge and magnetic charge. Note that the polarization charge “$\frac{\theta}{2\pi}N^{fi}_{m}$” due to the Witten effect in Eq. (\[Nnf1\]) does not enter the statistics. A field theoretic understanding of this phenomenon can be found in Ref. . For later convenience, we may express Eq. (\[eq:quantum\_stat\]) in terms of the $U(1)_{\rm EM}\times U(1)_a\times U(1)_b$ gauge groups: $$\begin{aligned}
\Gamma=&N^a_m(n^{f1}-n^{f2})+N^b_m (n^{f3}-n^{f2})\nonumber\\
&+(M+1)(n^{f1}+n^{f2}-n^{f3})\,,\label{eq:stat123}\end{aligned}$$ where we have added even integers during the derivation as only the value of $\Gamma$ mod $2$ is meaningful. If $\Gamma$ is even, the composite is bosonic; otherwise, it is fermionic. After inserting the values of the electric and magnetic charges of $\varphi_1$ and $\varphi_2$ into $\Gamma$, we may obtain the $\Gamma$ formulae of both $\varphi_1$ and $\varphi_2$ (denoted as $\Gamma(\varphi_1),\Gamma(\varphi_2)$) as functions of $u,v,u',v',q,q'$ (see Appendix \[appendix\_bosonic\]). The requirement that both $\varphi_1$ and $\varphi_2$ are bosonic leads to the following constraints on $u,v,u',v',q,q'$: $$\begin{aligned}
&\Gamma(\varphi_1)\in\Z_{\rm even}\,,\,~~\Gamma(\varphi_2)\in\Z_{\rm even}\,.\label{constr_3}\end{aligned}$$ So far, we have deduced several constraints on the six numbers: Eqs. (\[constr\_1plus\],\[constr\_1\],\[constr\_3\]), but there is one more constraint, i.e., Eq. (\[constr\_5\]), which enforces that $\varphi_1$ and $\varphi_2$ are linearly independent. It is possible that one of the composites consists of several copies of the other composite, in which case there is actually only one condensate. To avoid this situation, the following condition should be strictly imposed: $$\begin{aligned}
uv'-u'v\neq 0\,.\label{constr_5}
\end{aligned}$$ A physical understanding of this condition will be presented in Sec. \[sec:loop\_loop\]. For convenience, we introduce the following notation: $$\begin{aligned}
&K=\bpm u & v \\ u'& v'\epm,~~ \mathbf{N}_m=\bpm N_m^a\\N_m^b\epm,~~ \mathbf{q}=\bpm q\\q'\epm,\label{define_KNq1}
\\
&\mathbf{\Phi}_e=\bpm \Phi^a_e\\ \Phi^b_e\epm, ~~\mathbf{N}_e=\bpm N_a\\N_b\epm.
\label{define_KNq}\end{aligned}$$ Then, the matrix $K$ is invertible, namely, its determinant should be nonzero, as given by Eq. (\[constr\_5\]). In summary, the conditions Eqs. (\[constr\_1plus\],\[constr\_1\],\[constr\_3\],\[constr\_5\]) should be imposed on the six integers $u,v,q;u',v',q'$ such that the two condensates satisfy conditions: (\[condition\_phi2\],\[condition\_phi3\],\[condition\_phi1\]).
Generalized flux quantization and loop excitations {#sec:loop_loop}
--------------------------------------------------
In order to gain a better physical understanding of the condition (\[constr\_5\]), we need to carefully study the “generalized flux quantization” induced by the two condensates $\varphi_1$ and $\varphi_2$ whose electric and magnetic charges are listed in Table \[table:em\]. In a usual type-II superconductor, we know that the EM magnetic flux denoted by $\Phi^A_M$ is screened and quantized according to $2\Phi^A_M/2\pi=\Phi^A_M/\pi\in\Z$ since the Cooper pair condensate carries $2e$ EM electric charge. In our case, the two condensates $\varphi_1$ and $\varphi_2$ carry not only EM electric charges but also magnetic charges of the $a$ and $b$ gauge groups as shown in Table \[table:em\]. As a result, we have the following generalized flux quantization conditions: $$\begin{aligned}
& q\Phi^A_M+u\Phi^a_e+v\Phi^b_e=2\pi\ell,\label{EM1}\\
&q'{\Phi^A_M}+u'{\Phi^a_e}+v'{\Phi^b_e}=2\pi\ell',\!\label{EM1+}\end{aligned}$$ where $\Phi^A_M$ is the EM magnetic flux piercing a spatial loop $\mathcal{S}^1$. $\Phi^a_e$ and $\Phi^b_e$ are the $a$- and $b$-electric fluxes piercing $\mathcal{S}^1$, respectively. Here, instead of magnetic fluxes, electric fluxes of the $U(1)_a\times U(1)_b$ gauge group are involved since the condensates carry magnetic charges rather than electric charges of the $U(1)_a\times U(1)_b$ gauge group. $\ell,\ell'\in\Z$ label the winding numbers of the mapping $\mathcal{S}^1\rightarrow U(1)$ of the condensate order parameters $\varphi_1,\varphi_2$.
In contrast with fluxes of the internal gauge groups, arbitrary values of $\Phi^A_M$ are allowed to be inserted. In other words, $A_\mu$ itself is not higgsed, and the EM electric charge of the electrons is a well-defined quantum number. This implies that the two condensates must provide a new charge screening mechanism such that the net EM electric charge of each condensate is zero, although both condensates carry a nonzero bare EM electric charge ($N_A=q,q'$). This screening effect can lead to fractionalization of the charge of excitations, even in the absence of an external EM magnetic charge. We will postpone a discussion of this issue until Sec. \[section\_of\_screen\].
Since $A_\mu$ is an external non-dynamical field, we may temporarily turn it off in Eqs. (\[EM1\],\[EM1+\]) to find: $$\begin{aligned}
& u\Phi^a_e+v\Phi^b_e=2\pi \ell\,, \,u'{\Phi^a_e}+v'{\Phi^b_e}=2\pi \ell'\,.\label{phie1}\end{aligned}$$ A generic solution of Eq. (\[phie1\]) is given by: $$\begin{aligned}
&\Phi^a_e=2\pi\frac{\ell v'-\ell'v}{\mathsf{Det}K}\,,\,\Phi^b_e=2\pi\frac{\ell'u-\ell u'}{\mathsf{Det}K}\,.\label{phi_ae}
\end{aligned}$$ Here, $u,v,u',v'\in\Z$ satisfy the condition (\[constr\_5\]). By noting that $\ell v'-\ell'v$ is divisible by the greatest common divisor ${\mathsf{GCD}}(v,v')$ and $\ell'u-\ell u'$ is divisible by the greatest common divisor $\mathsf{GCD}(u,u')$, one can use [Bézout’s lemma]{} (see Appendix \[appendix\_proof\_theorem\_flux\]) to obtain the minimal quantized electric fluxes: $$\begin{aligned}
\!\!\!\!(\Phi^a_e)_{\rm min}\!=\!2\pi\!\bigg|\frac{ \mathsf{GCD}(v,v')}{ \mathsf{Det}K}\bigg|, (\Phi^b_e)_{\rm min}\!=\!2\pi\!\bigg|\frac{ \mathsf{GCD}(u,u')}{ \mathsf{Det}K}\bigg|.\!\label{discreteflux1}
\end{aligned}$$ Since $| uv'-u'v|$ is divisible by both $\mathsf{GCD}(v,v')$ and $\mathsf{GCD}(u,u')$, we have the following two useful inequalities: $$\begin{aligned}
\!\!\!\!\!|uv'-u'v|\!\geq\!|\mathsf{GCD}(u,u')|\,, |uv'-u'v|\!\geq\!|\mathsf{GCD}(v,v')|.\label{eq:ineq}\end{aligned}$$
Based on Bézout’s lemma, we can easily prove the following theorem. The proof is shown in Appendix \[appendix\_proof\_theorem\_flux\]:
$(\Phi^a_e)_{\rm min}=2\pi$ and $(\Phi^b_e)_{\rm min}=2\pi$ if and only if $|uv'-u'v|=1$.\[theorem\_flux\]
The above theorem leads to the following criterion for loop/flux excitations:
\[Criterion for loop excitations\] If $|\mathsf{Det}K|=1$, the bulk has no deconfined discrete gauge fluxes (therefore, no detectable loop excitations); If $|\mathsf{Det}K|>1$, the bulk has deconfined discrete gauge fluxes (therefore, detectable loop excitations with minimal flux strength smaller than $2\pi$.)\[crt\_loop\_exc\]
The solutions $(\Phi_e^a, \Phi_e^b)$ in Eq. (\[phi\_ae\]) can be recast in the following form: $$\begin{aligned}
\mathbf{\Phi}_e=2\pi K^{-1}\mathbf{L} \,,\label{eq:phie_L}\end{aligned}$$ where the integer vector $\mathbf{L}=(\ell,\ell')^T$. Thus, we may define a 2D loop-lattice generated by a dimensionless integer vector $\mathbf{L}$:
\[[Loop-lattice]{}\] A loop-lattice is a 2D square lattice where each site corresponds to a loop excitation labeled by $\mathbf{L}=(\ell,\ell')^T$. The corresponding electric flux strength $\mathbf{\Phi}_e$ of each site is determined by Eq. (\[eq:phie\_L\]).\[dfn\_loops\]
Point-particle excitations and charge fractionalization {#section_of_screen}
-------------------------------------------------------
In addition to loop excitations, we also have point-particle excitations:
\[[Excitation]{} and [Charge lattice]{}\] Excitations are defined as deconfined particles that have trivial mutual statistics with both condensates. All excitations form a 4D charge lattice which is a sublattice of the original 6D lattice. Unless otherwise specified, excitations always refer to point-particle excitations.\[dfn\_excitation\]
By definition, all excitations have trivial mutual statistics with respect to the condensates. In other words, Eq. (\[equation:mutual\_stat\]) holds between any excitation and $\varphi_1$, and also holds between any excitation and $\varphi_2$. By explicitly using the parameters of $\varphi_1$ and $\varphi_2$ in Table \[table:em\], the electric and magnetic charges of excitations are constrained by the following two equations: $$\begin{aligned}
&qM=uN_a+vN_b\,, ~~q'M=u'N_a+v'N_b\,.\label{excitation_1}
\end{aligned}$$ Therefore, a generic particle that has six independent charges $(N_A,N_a,N_b,M,N_m^a,N_m^b)$ is now completely determined by four of them $(N_A,M,N_m^a,N_m^b)$ if the particle is a deconfined excitation in the condensed phase. Keeping Eq. (\[constr\_5\]) in mind, $N_a$ and $N_b$ are fully determined by $M$: $ N_a=\frac{qv'-q'v}{\mathsf{Det}K}M\,,\, N_b=\frac{q'u-qu'}{\mathsf{Det}K}M\,$ which can be written as $$\begin{aligned}
\mathbf{N}_e=MK^{-1}\mathbf{q}\label{NabM}
\end{aligned}$$ by using the notation in Eqs. (\[define\_KNq1\],\[define\_KNq\]).
![(Color online) Schematic representation of the charge screening mechanism. Consider a composite particle carrying $N_A$ units of the EM ($A_\mu$) electric charge (i.e., $U(1)$ symmetry charge), $N_m^a$ units of magnetic charge of the $a_\mu$ field, and $N_m^b$ units of magnetic charge of the $b_\mu$ gauge field. Due to the condensates, $N_A$ is partially screened such that the net EM electric charge $Q$ is given by Eq. (\[NE1\]), which is different from $N_A$. In (a) the physics of Aharonov-Bohm effect in Eq. (\[ABPhase\]) is illustrated. An excitation (denoted by the blue ball) adiabatically moves along a closed trajectory and feels the EM magnetic flux $\Phi^A_M$, the electric flux $\Phi^a_e$ of the $a_\mu$ gauge field, and the electric flux of the $b_\mu$ gauge field. In (b), condensed particles $\varphi_1$ and $\varphi_2$ form a Debye-H[ü]{}ckel-like charge cloud around an excitation, providing the screening charge $Q_{\rm Debye}$ in Eq. (\[NE\_screen\]).[]{data-label="figure_flux"}](figure_flux.pdf){width="8cm"}
As mentioned in Sec. \[sec:gauge\_strc\], the EM electric charge of a particle, $N_A$, is called the “bare” charge, which suggests that it will be partially screened due to the condensates. In order to clearly see the screening, we turn on the external EM field $A_\mu$ to probe the EM response and consider a spatial loop $C$. The total Aharonov-Bohm phase accumulated by an adiabatically moving test particle is given by (see Fig. \[figure\_flux\]): $$\begin{aligned}
&\text{Aharonov-Bohm phase}\nonumber\\
=
&\exp\{iN_A \Phi^A_M+i M \Phi^A_E+iN^a\Phi^a_m+iN^b\Phi^b_m\nonumber\\
&+iN^a_m \Phi^a_e+iN^b_m \Phi^b_e\}\,,\label{ABPhase}\end{aligned}$$ where $\Phi^A_E$ is EM electric flux piercing $C$, $\Phi_m^a$ and $\Phi_m^b$ are the $a$- and $b$-magnetic fluxes respectively. However, Eqs. (\[EM1\],\[EM1+\]) indicate that $\Phi^a_e$ and $\Phi^b_e$ depend linearly on the external EM magnetic flux $\Phi^A_M$. Solving Eqs. (\[EM1\],\[EM1+\]) leads to: $$\begin{aligned}
&\Phi^a_e=2\pi\frac{kv'-k'v}{\mathsf{Det}K}-{ \Phi^A_M\frac{qv'-q'v}{\mathsf{Det}K}}\,,\label{screening1}\\
&\Phi^b_e=2\pi\frac{k'u-ku'}{\mathsf{Det}K}-{ \Phi^A_M\frac{q'u-qu'}{\mathsf{Det}K}}\,.\label{screening2}
\end{aligned}$$ The terms that depend linearly on $\Phi_M^A$ correct the saddle point solutions in Eq. (\[phi\_ae\]).
Taking Eqs. (\[screening1\],\[screening2\]) into account, the contribution to the Aharonov-Bohm phase due to the external EM gauge field can be isolated. Eq. (\[ABPhase\]) can be recast into $e^{iQ \Phi^A_M+\cdots}$, where $\cdots$ denotes the remaining terms that do not contain the factor $\Phi^A_M$, and, $Q$ is the *net EM electric charge*: $$\begin{aligned}
Q=N_A-Q_{\rm Debye}\,, \label{NE1}
\end{aligned}$$ where $$\begin{aligned}
\!\!\!\! Q_{\rm Debye}\!=\!{ N_m^a\frac{qv'-q'v}{\mathsf{Det}K}\!+\!N_m^b\frac{q'u-qu'}{\mathsf{Det}K}}\!=\!\mathbf{N}_m^T K^{-1}\!\mathbf{q}\,.\!\!\!\label{NE_screen}
\end{aligned}$$ denotes the charge carried by the Debye-H[ü]{}ckel-like screening cloud (see Fig. \[figure\_flux\]). The matrix $K$, vector $\mathbf{N}_m,$ and the vector $\mathbf{q}$ are defined in Eq. (\[define\_KNq\]). Concerning the screening charge, there are two interesting limits. First, the total EM electric charge of each condensate is zero, i.e., $Q=0$ for both condensates $\varphi_1$ and $\varphi_2$ (see Table \[table:em\]). For example, we have $Q_{\rm Debye}=q$ for $\varphi_1$, which completely screens its bare EM electric charge $N_A=q$. Second, let us consider an intrinsic excitation whose bare EM electric charge vanishes, $N_A=0$. Its net EM electric charge $Q$ is nonzero and completely given by that of the Debye screening cloud: $
Q=-Q_{\rm Debye}= - \mathbf{N}_m^T K^{-1}\!\mathbf{q}
$.
In fact, charge fractionalization in Abelian FQH states can also be understood via the above Aharonov-Bohm thought experiment. As an example, let us derive the fractionalization of charge in the $\nu=1/3$ Laughlin state. The effective field theory is described by the following Lagrangian: $$\begin{aligned}
\mathcal{L}=\frac{3}{4\pi}a_\mu\partial_\nu a_\lambda \epsilon^{\mu\nu\lambda}+\frac{1}{2\pi}A_\mu \partial_\nu a_\lambda\epsilon^{\mu\nu\lambda}\,.
\end{aligned}$$ Here the gauge field $a_\mu$ is a dual description of the electron current $J_\mu$: $J_\mu=\frac{1}{2\pi}\partial_\nu a_\lambda \epsilon^{\mu\nu\lambda}$. The second term in the Lagrangian $\mathcal{L}$ means that each electron carries one unit of electric charge. Excitations in FQH states are labeled by gauge charges of the $a_\mu$ gauge group, since they minimally couple to $a_\mu$. In this sense, let us consider the Aharonov-Bohm experiment for an excitation that carries one unit of gauge charge of $a_\mu$. The Aharonov-Bohm phase is given by $e^{i\Phi_a}$ where $\Phi_a$ is the $a_\mu$ magnetic flux felt by the excitation. In the hydrodynamical field theory $\mathcal{L}$, $\frac{1}{2\pi}\Phi_a$ corresponds to the electron density. By studying the equation of motion of $a_\mu$ in $\mathcal{L}$, we obtain: $\Phi_A=3\Phi_a$. Physically, this identity means that each electron effectively corresponds to three units of magnetic flux of the background EM field, which is nothing but the definition of filling fraction $\nu=\frac{1}{3}$. Thus, the Aharonov-Bohm phase accumulated by the excitation is identical to: $e^{i\Phi_a}=e^{i\frac{1}{3}\Phi_A}$. The coefficient $\frac{1}{3}$ indicates that the excitation in the presence of $A_\mu$ behaves as an electrically charged particle with $\frac{1}{3}$ charge.
Let us come back to our 3D system. We introduce the following two equivalent criteria for charge fractionalization:
\[Criterion for charge fractionalization\] Charge fractionalization exists if excitations with zero $M$ and fractionalized $Q$ exist. \[crt\_charge\_1\]
and,
\[Criterion for charge fractionalization\] Equivalently, charge fractionalization exists if the EM magnetic charge $M$ of excitations is quantized in units of an integer $w>1$, i.e., $M=0,\pm w, \pm 2w,\cdots$. \[crt\_charge\_2\]
In Appendix \[appendix\_dirac\], the equivalence of the above two criteria is explained by using the well-known Dirac-Zwanziger-Schwinger quantization condition. The requirement of $M=0$ in Criterion \[crt\_charge\_1\] can be understood as follows. Typically, in the presence of $M$, excitations can potentially carry a fractionalized $Q$ due to the Witten effect. However, this does not mean our 3D quantum system is fractionalized. The topological insulator (TI) is a typical example. If a single EM monopole ($M=1$) is inserted into the bulk, there is a half-charge cloud surrounding the monopole [@Qi2008; @franz]. However, since the TI can be realized in a non-interacting band insulator we do not consider it to be fractionalized. In order to highlight the set of excitations with $M=0$, we introduce the notion of intrinsic excitations and intrinsic charge lattice:
\[[Intrinsic excitations]{} and [intrinsic charge lattice]{}\] Intrinsic excitations are excitations with zero EM magnetic charge, i.e., $M=0$; the intrinsic charge lattice is a special 3D charge lattice with zero EM magnetic charge, i.e., $M=0$.\[dfn\_intrinsic\_exc\]
One can verify that $Q_{\rm Debye}$ in Eq. (\[NE1\]) is the *unique* source of charge fractionalization. In other words, $N_A$ in Eq. (\[NE1\]) is always integer-valued when $M=0$ (see Appendix \[appendix\_na\_integer\] for details); charge fractionalization exists if and only if $Q_{\rm Debye}$ is fractional when $M=0$.
Finally, we show that the Debye charge cloud $Q_{\rm Debye}$ in Eq. (\[NE\_screen\]) can also be understood in a more formal way, i.e., from a topological $BF$ field theory. Without loss of generality, we consider the London limit (i.e., deep in the confined phase) such that the amplitude fluctuations of $|\varphi_I|$ are negligible. In this limit, we may dualize $S_{\rm GL}$ in Eq. (\[equation\_free\_energy\]) into a two-component topological $BF$ field theory [@horowitz89]: $$\begin{aligned}
S=&i\int \frac{1}{2\pi} \mathscr{B}^T \wedge K d \mathscr{A}+i\int \frac{1}{2\pi} \mathbf{q}^T \mathscr{B}\wedge d A+\mathcal{S}_{ex},\label{BFaction}\end{aligned}$$ where we define the two-component vectors $\mathscr{B}=\bpm \mathcal{B},\mathcal{B}'\epm^T$ and $\mathscr{A}=\bpm \tilde{a},\tilde{b}\epm^T$, and use a differential form notation. Here, $\mathcal{B}$ and $\mathcal{B}'$ are two Kalb-Ramond $2$-form gauge fields introduced as a result of the particle-vortex line duality transformation in (3+1)D [@savit]. Physically, they are related to the supercurrents of the condensates $\varphi_1$ and $\varphi_2$, respectively, via: $J^{\varphi_1}=\frac{1}{2\pi} \star d\mathcal{B}\,,\,{J^{\varphi_2}}=\frac{1}{2\pi} \star d \mathcal{B}'\,,
$ where $\star$ is the usual Hodge-dual operation. Since the energy gap in the bulk of the topological $BF$ field theory is effectively infinite, the term $\mathcal{S}_{ex}$ is added by hand in order to take into account the point-like excitations labeled by $\mathbf{N}_m=(N_m^a,N_m^b)^T,$ and the loop excitations labeled by the integer vector $\mathbf{L}=(\ell,\ell')^T$ (see Definition \[dfn\_loops\]): $$\begin{aligned}
\mathcal{S}_{ex}=i \int \mathbf{N}_m^T \mathscr{A} \wedge \star j+i \int\mathbf{L}^T \mathscr{B} \wedge \star \Sigma\,,\label{BFex}\end{aligned}$$ where the vector $j$ denotes the composite excitation current, and the tensor $\Sigma$ denotes the loop excitation current. Integrating out the dynamical fields $\mathscr{A}$ and $\mathscr{B}$ yields an effective theory for $j$ and $\Sigma$ in the presence of the external EM field $A_\mu$: $$\begin{aligned}
\!\!\!\!S_{\rm eff}\!= \!i\mathbf{N}_m^T K^{-1}\!\mathbf{q} \!\!\int \! j\!\wedge \star A +\!i 2\pi\mathbf{N}_m^TK^{-1}\mathbf{L} \! \!\int \!\Sigma \wedge d^{-1}\,j.\!\! \label{eqn:effective}\end{aligned}$$ It is remarkable that the first term in the effective action (\[eqn:effective\]) is nothing but the Debye screening charge cloud $Q_{\rm Debye}$ defined in Eq. (\[NE\_screen\]). Thus, $Q_{\rm Debye}$ is a topological property of an excitation. The second term represents the long-range Aharonov-Bohm statistical interaction between fluxes and particles. The operator $d^{-1}$ is a formal notation defined as the operator inverse of $d$, whose exact form can be understood in momentum space by Fourier transformation. The coefficient $\mathbf{N}^T_m K^{-1} \mathbf{L}$ gives rise to the charge-loop braiding statistics $\vartheta^{cl}$ between composite particles with quantum number $\mathbf{N}_m$ and loop excitations with electric fluxes $\mathbf{\Phi}_e=2\pi(K)^{-1}\mathbf{L}$ due to Eq. (\[eq:phie\_L\]): $$\begin{aligned}
\vartheta^{cl}=2\pi\mathbf{N}^T_m K^{-1} \mathbf{L}=\mathbf{N}^T_m \mathbf{\Phi}_e \,.\label{eq:cltheta}\end{aligned}$$
Now that we have carefully developed a theory that describes topological phases in the presence of $U(1)$ composite condensates we will use the results to construct fractionalized 3D topological insulators with time-reversal symmetry.
Fractional topological insulators {#sec:theta}
=================================
In this section, we will study 3D topological phases of matter with non-vanishing axion angle $\Theta$. The presence of nontrivial values of $\Theta$ lead to several observable phenomena including a surface quantum Hall effect, and the celebrated Witten effect: a magnetic monopole will bind a electric charge. For free-fermion time-reversal invariant topological insulators, the angle is $\pi \text{ mod } 2\pi$ [@Qi2008]. In fractionalized states where strong interactions and correlations are taken into account, in principle, the axion angle can be fractional (i.e., $\Theta/\pi$ is not integral) while time-reversal invariance is still maintained [@maciejkoFTI; @maciejko_model; @maciejko2015; @swingle2011; @YW13a]. Such topological phases are called “fractional topological insulators” (FTI). In Ref. , FTIs were obtained via parton constructions where the partons themselves carry fractional EM electric charges, and the internal gauge fields are in the Coulomb phase where gauge fluctuations are weak and the photon(s) are gapless. In a different non-fractionalized state where $U(1)\times U(1)\rtimes Z_2$ symmetry is considered [@bti6], the $\Theta$ term may signal a mutual Witten effect where a monopole of one U(1) gauge group induces an electric charge of another $U(1)$ gauge group.
In the following, we will explore [FTI]{}s via the parton construction introduced in Sec. \[sec:gauge\_strc\]. However in Sec. \[sec\_FTI\_Coulomb\], we shall first study charge lattices for which all partons occupy topological insulator bands and the internal gauge fields are in the Coulomb phase. The resulting state is a non-fractional topological insulator (i.e., $\Theta=\pi$) and there are two massless gauge bosons in the bulk. In order to obtain a gapped bulk and a fractional $\Theta$, in Sec. \[sec\_FTI\] we again assume that all partons occupy topological insulator bands and then we condense certain composites (see $\varphi_1$ and $\varphi_2$ of [FTI]{} in Table \[table:em\]). We focus on a concrete example and show that the resulting state is a [FTI]{} with time-reversal symmetry and $\Theta=\frac{\pi}{9}$ (c.f. Eq. (\[eqn:new\_Theta\_reduced\])). As a side result, in Appendix \[appendix\_Gamma\_trivial\_1\], we show that the ansätze in which all partons are in a topologically trivial band structure always gives a topologically trivial state with $\Theta=0$ regardless of the condensate structure.
Topological insulators in the Coulomb phase {#sec_FTI_Coulomb}
-------------------------------------------
In the following, we consider partons occupying non-trivial 3D topological insulator bands (i.e., $\theta=\pi$). Previously, it was shown that partons with $\theta=\pi$ can potentially support [a]{} fractional $\Theta$ angle if [the]{} Coulomb phase is considered, and a special parton representation of an electron is used [@maciejkoFTI]. In the Coulomb phase, the dynamical gauge fields $a_\mu$ and $b_\mu$ are weakly fluctuating and non-compact; hence, the standard perturbative analysis is applicable. Integrating [out]{} the partons [to quadratic order in the gauge fields [@Qi2008], we obtain the following effective action $S_{\rm eff}$:]{} $$\begin{aligned}
\!S_{\rm eff}\!=\!& \!\int\! d^4x\frac{\theta}{32\pi^2}\left(g_a f^a_{\mu\nu}\!+\!eG_{\mu\nu}\right) \! \!\left(g_af^a_{\lambda\rho}\!+\!eG_{\lambda\rho}\right)\!\ep
+ \! \!\int \!d^4x\frac{\theta}{32\pi^2}\!\!\left(-g_af^a_{\mu\nu}\!-\!g_bf^b_{\mu\nu}+eG_{\mu\nu}\right)\nonumber\\
&\times \! \!
\left(-g_af^a_{\lambda\rho}\!-\!g_bf^b_{\lambda\rho}\!+\!eG_{\lambda\rho}\right)\!\ep+\!\!\int d^4x\frac{\theta}{32\pi^2}\left(g_bf^b_{\mu\nu}-eG_{\mu\nu}\right) \! \!\left(g_bf^b_{\lambda\rho}-eG_{\lambda\rho}\right)\ep+S_\text{Maxwell}\,,
\end{aligned}$$
where $\theta=\pi$. The quantities $f^a_{\mu\nu}=\partial_\mu a_\nu-\partial_\nu a_\mu$ and $f^b_{\mu\nu}=\partial_\mu b_\nu-\partial_\nu b_\mu$ are field strength tensors of $a_\mu$ and $b_\mu$ respectively. Both $a_\mu$ and $b_\mu$ are smooth variables and do not support monopole configurations. $G_{\mu\nu}$ is defined as: $G_{\mu\nu}=F_{\mu\nu}-\frac{2\pi}{e}S_{\mu\nu}$, where $F_{\mu\nu}=\partial_\mu A_\nu-\partial_\nu A_\mu$, $A_\mu$ is smooth external EM field, and the tensor $S_{\mu\nu}$ forms the EM monopole current via: $M_\mu=\frac{1}{2}\ep \partial_\nu S_{\lambda\rho}\,
$. The constant $e^2$ denotes the fine structure constant of the EM field $A_\mu$. The coupling constants $g_{a,b}$ of the $a_\mu$ and $b_\mu$ gauge fields are written explicitly and $0<g_a,g_b\ll 1$ in the Coulomb phase. $S_\text{Maxwell}$ includes all non-topological terms (Maxwell-type) of $a_\mu,b_\mu$ and $A_\mu$. Since both $a_\mu$ and $b_\mu$ are smooth variables, all terms of the form $f^a\wedge f^a$, $f^a\wedge f^b$, and $f^b\wedge f^b$ are total-derivative terms that can be neglected in the bulk effective field theory. The term $-\frac{4\theta \,e \,g_b}{32\pi^2}f^b_{\mu\nu} G_{\lambda\rho}\ep=\frac{\theta g_b}{\pi}M_\mu b_\mu$ implies that $M_\mu$ carries integer gauge charge of the $U(1)_b$ gauge group by noting that $\frac{\theta}{\pi}=1$. As such, after integrating $a_\mu,b_\mu$ $S_{\rm eff}$ reduces to: $$\begin{aligned}
S_{\rm eff}=&\frac{\Theta e^2}{32\pi^2}\int d^4x G_{\mu\nu}G_{\lambda\rho}\ep+\cdots\,,\label{Effective_Action_Coulomb}\end{aligned}$$ where $\Theta=3\pi$. The terms represented by $\cdots$ include the long-range Coulomb interactions between the monopole currents $M_\mu$ mediated by the $b_\mu$-photons, and other non-topological terms. Since the periodicity of $\Theta$ is still $2\pi$ in the absence of charge fractionalization, $\Theta$ reduces to $\pi$ by a $2\pi$ periodic shift. In summary, the resulting state shows a $\Theta$ angle that is the same as a free-fermion topological insulator. The bulk admits two gapless, electrically neutral excitations, i.e., photons of the $U(1)_a$ and $U(1)_b$ gauge fields.
Fractional topological insulators in the composite condensation phase {#sec_FTI}
---------------------------------------------------------------------
The charge [lattice]{} in Sec. \[sec\_FTI\_Coulomb\] was obtained from the assumptions that (i) [partons]{} occupy $\theta=\pi$ topological insulator bands, and (ii) the internal gauge fields are in the Coulomb phase. However, the [resulting]{} phase supports a non-fractional $\Theta=\pi$ angle and the bulk spectrum is gapless. In the following we consider composite condensation phases as discussed in Sec. \[sec:gauge\_strc\] instead of the Coulomb phase. When the partons are in topological insulator bands the [resulting]{} phase can support fractionalized $\Theta$ angles and a *fully* gapped bulk.
[Let]{} us start with the scenario that all partons occupy topological insulator bands (i.e., $\theta=\pi$) and then consider composite condensations. One can prove that parameters $u,v,u',v',q,q'$ must be even: $$\begin{aligned}
u,v,u',v',q,q'\in\Z_{\rm even}\label{uvuvqq_even}\,{,}\end{aligned}$$ in order to satisfy the set of constraints given by Eqs. (\[constr\_1\],\[constr\_3\]). [We obtain the following relations]{} via Eqs. (\[eqn:transition\_matrix\],\[eqn:transition\_matrix1\],\[Nnf1\]) ($\theta=\pi$): $$\begin{aligned}
n^{f1}-n^{f2}=&N_a-N^a_m-\frac{1}{2}N^b_m\,,\label{N_a_n1n2}\\
n^{f3}-n^{f2}=&N_b-N^b_m-\frac{1}{2}N^a_m+M\,,\label{N_b_n1n2}\\
n^{f1}+n^{f2}-n^{f3}=&N_A+N^b_m-\frac{3}{2}M\,.\label{NA_is_integer}\end{aligned}$$ In order to see whether or not there is charge fractionalization (Definition \[crt\_charge\_1\]), we may check the value of $Q$ defined in Eq. (\[NE1\]) when $M=0$. Then, Eq. (\[NA\_is\_integer\]) indicates that $N_A$ is always integer-valued when $M=0$ by noting that $n^{fi}$ and $N_m^b$ are integer-valued.
Thus, we should further check whether or not $Q_{\rm Debye}$ defined in Eq. is fractional when $M=0$. In principle, one may deduce $\Theta$ as a function of the parameters $(u,v,u',v',q,q',\theta)$. However, such a generic discussion is technically intricate and not illuminating. Instead, we will proceed further with a concrete example as a proof of principle (see Table \[table:em\]): $
u=2,v=2,u'=4,v'=10,q=2,q'=2\,$. In terms of the matrix notation defined in Eq. (\[define\_KNq1\]), we have: $$\begin{aligned}
K=\bpm 2 & 2 \\ 4& 10\epm, ~~~~\mathbf{q}=\bpm 2\\2\epm\,.\label{define_KNq12345}\end{aligned}$$ From Table \[table:em\], we see that $\varphi_1$ is a bosonic bound state of two $a_\mu$ magnetic monopoles, two $b_\mu$ magnetic monopoles, one $f^1$ parton, four $f^2$ partons, and one $f^3$ parton. $\varphi_2$ is a bosonic bound state of four $a_\mu$ magnetic monopoles, ten $b_\mu$ magnetic monopoles, nine $f^2$ partons, and three hole-like $f^3$ partons.
By using Eq. (\[NabM\]), it is straightforward to work out the relation between $N_{a,b}$ and $M$: $$\begin{aligned}
N_a=\frac{4}{3}M\,,~~~~~ N_b=-\frac{1}{3}M\,{,}\label{example_NaNb}\end{aligned}$$ which must be satisfied for all excitations. Plugging Eq. (\[example\_NaNb\]) into Eqs. (\[N\_a\_n1n2\],\[N\_b\_n1n2\]), we end up with $$\begin{aligned}
n^{f1}-n^{f2}=&\frac{4}{3}M-N^a_m-\frac{1}{2}N^b_m\,,\label{N_a_n1n2111}\\
n^{f3}-n^{f2}=&\frac{2}{3}M-N^b_m-\frac{1}{2}N^a_m\,.\label{N_b_n1n2111}\end{aligned}$$ It is obvious that the single parton $f^i$ whose charges are shown in Table \[table:em\] is confined since Eqs. (\[N\_a\_n1n2111\],\[N\_b\_n1n2111\]) are not satisfied simultaneously.
By noting that $n^{fi}, N_m^a,N_m^b,M$ are integer-valued, Eqs. (\[N\_a\_n1n2111\],\[N\_b\_n1n2111\]) require that: $$\begin{aligned}
&N_m^a, N_m^b\in\Z_{\rm even} \,;~~~\,M=0,\pm3,\pm6,\pm9,\cdots\,.\label{double_m_charge}\end{aligned}$$ Therefore, the quantization of $M$ is modified compared to the usual quantization $M=0,\pm1,\pm2{,\ldots}$ found in the vacuum and a non-fractionalized TI. A direct consequence is that an $M=1$ particle is not allowed to pass through [the]{} [FTI]{}, which is illustrated in Fig. \[figure\_ti\].
![(Color online) Throwing three external EM magnetic monopoles (denoted by blue balls) in vacuum into topological materials TI and [FTI]{}. Only [the]{} EM magnetic monopole with [$3k$, $k\in\Z$]{} magnetic charge shown in (\[double\_m\_charge\]) can penetrate the [FTI]{} boundary in our example. An EM magnetic monopole with $M=1,2$ will be completely reflected on the [FTI]{} boundary, as shown by the leftwards arrows. The shadow of each ball pictorially denotes the polarization charge cloud induced by the Witten effect.[]{data-label="figure_ti"}](figure_ti.pdf){width="8.6cm"}
We may use $n^{f1}$, $n^{f2}$, $n^{f3}$, and $M$ to uniquely label all excitations. Solving Eqs. (\[N\_a\_n1n2111\],\[N\_b\_n1n2111\]) gives rise to: $$\begin{aligned}
&N_m^a=\frac{4}{3}M-\frac{4}{3}n^{f1}+\frac{2}{3}n^{f2}+\frac{2}{3}n^{f3}\,,\label{equation_NMA}\\
&N_m^b=\frac{2}{3}n^{f1}+\frac{2}{3}n^{f2}-\frac{4}{3}n^{f3}\,.\label{equation_NMB}\end{aligned}$$ By using the above two equations, $N_A$ in Eq. (\[NA\_is\_integer\]) and $Q_{\rm Debye}$ in Eq. (\[NE\_screen\]) can be expressed as: $$\begin{aligned}
&N_A=\frac{1}{3}n^{f1}+\frac{1}{3}n^{f2}+\frac{1}{3}n^{f3}+\frac{3}{2}M\,,\\
&Q_{\rm Debye}=-2n^{f1}+\frac{2}{3}n^{f2}+\frac{4}{3}n^{f3}+\frac{16}{9}M\,.\end{aligned}$$ The net EM electric charge $Q$ is defined as $N_A-Q_{\rm Debye}$ and thus is given by: $$\begin{aligned}
Q=\frac{7}{3}n^{f1}-\frac{1}{3}n^{f2}-n^{f3}-\frac{5}{18}M\,.\label{Q_expression}\end{aligned}$$ Thus, the quantization of $Q$ is given by: (see Appendix \[appendix\_gamma\_stat4321\_new\]) $$\begin{aligned}
&Q=0,\pm\frac{1}{3},\pm\frac{2}{3},\pm1\,,\cdots \,\,\,\text{ when } \frac{M}{3}=0,\pm2,\cdots\,,\label{Q_fractional_M0}\\
&Q=0,\pm\frac{1}{6},\pm\frac{3}{6},\pm\frac{5}{6}\,,\cdots \,\,\, \text{ when } \frac{M}{3}=\pm1,\pm3,\cdots.\label{Q_fractional_M1}\end{aligned}$$ Eq. (\[Q\_fractional\_M0\]) indicates that the intrinsic excitations of [the]{} [FTI]{} (Definition \[dfn\_intrinsic\_exc\]) carry $1/3$ quantized EM electric charge. In other words, the [FTI]{} bulk supports charge fractionalization (Criterion \[crt\_charge\_1\]). Due to the quantization of $M$ in Eq. (\[double\_m\_charge\]), Criterion \[crt\_charge\_2\] is automatically satisfied.
Eq. (\[Q\_fractional\_M1\]) indicates that the 2D $(M,Q)$ lattice is tilted by an angle $\frac{5}{18}M$. More precisely, an axion angle $\Theta$ can be defined as: $\Theta=-\frac{5}{9}\pi$ by identifying $- \frac{5}{18}M=\frac{\Theta}{2\pi}M$. This $M$-dependent EM electric charge is a known consequence of the Witten effect [@witten1; @Qi2008; @franz].
[The]{} self-statistics of excitations (i.e., either fermionic or bosonic) can also be derived as a function of $(n^{f1},n^{f2},n^{f3},M)$. For this purpose, let us start with $\Gamma$ defined in Eq. (\[eq:stat123\]) and take Eq. (\[double\_m\_charge\]) into account. Therefore, the first two terms of Eq. (\[eq:stat123\]) are even and can be removed giving: $$\begin{aligned}
\Gamma=&(M+1)(n^{f1}+n^{f2}+n^{f3})\,,\label{eq:stat4321}\end{aligned}$$ where $-n^{f3}$ is also changed to $n^{f3}$ leaving the even-odd property of $\Gamma$ unaltered. In analogy to a TI, time-reversal symmetry should also be [maintained]{}. From the point of view [of the charge lattice]{}, time-reversal symmetry is a reflection [symmetry]{} $M\rightarrow -M$ [that keeps the]{} net EM electric charge and self-statistics invariant[:]{} $Q\rightarrow Q$[,]{} $\Gamma\rightarrow \Gamma+\text{even integer}$. [One possible definition of time-reversal symmetry that satisfies these properties is as follows,]{} $$\begin{aligned}
& \mathcal{T}n^{f1} \mathcal{T}^{-1}= n^{f1}-\frac{1}{3}M\,,~ \mathcal{T}n^{f2} \mathcal{T}^{-1} =n^{f2}-\frac{2}{3}M\,, \\
& \mathcal{T}n^{f3} \mathcal{T}^{-1}= n^{f3}\,, ~~\mathcal{T}M \mathcal{T}^{-1}=-M\,,
\end{aligned}$$ where $\mathcal{T}$ denotes [the]{} time-reversal operator. It can be [verified]{} that $Q$ is invariant and $\Gamma$ is only changed by [an]{} even integer, thus leaving its even-odd property unaltered. [Using]{} the above transformations, we may also derive the transformations below: $$\begin{aligned}
\mathcal{T}N^a_m \mathcal{T}^{-1}=N_m^a-\frac{8}{3}M\,,~ \mathcal{T}N^b_m \mathcal{T}^{-1}=N_m^b-\frac{2}{3}M\,.
\end{aligned}$$ The shifted amounts $-\frac{8}{3}M$ and $-\frac{2}{3}M$ are even integers, which guarantees the transformed $N^{a,b}_m$ are still even as required by Eq. (\[double\_m\_charge\]). [A time-reversed excitation is still an excitation, in the sense that the]{} transformed electric and magnetic charges also satisfy all equations that are satisfied by the excitation before [time reversal]{}. Geometrically, [this]{} means that after the above transformations, the new particle is still on the 4D charge lattice. Furthermore, $Q$ and [$\Gamma$]{} are unchanged. From this geometric point of view, the time-reversal symmetry defined above effectively acts like a subgroup of [the point group]{} of the 4D charge lattice.
$n^{f1}$ $n^{f2}$ $n^{f3}$ $M$ $N_A$ $Q_{\rm Debye}$ $Q$ $N_m^a$ $N_m^b$ $\Gamma$
------------------------ ---------- ---------- ---------- ----- ---------------- ----------------- ---------------- --------- --------- ----------
Elementary charge $1$ $0$ $2$ $0$ $1$ $\frac{2}{3}$ $\frac{1}{3}$ $0$ $-2$ F
Elementary charge $2$ $9$ $1$ $0$ $4$ $\frac{10}{3}$ $\frac{2}{3}$ $4$ $6$ B
Electron $1$ $1$ $1$ $0$ $1$ $0$ $1$ $0$ $0$ F
Elementary EM monopole $0$ $0$ $0$ $3$ $\frac{9}{2}$ $\frac{16}{3}$ $-\frac{5}{6}$ $4$ $0$ B
An example with $M=3$ $1$ $1$ $1$ $3$ $\frac{11}{2}$ $\frac{16}{3}$ $\frac{1}{6}$ $4$ $0$ B
An example with $M=3$ $2$ $1$ $0$ $3$ $\frac{11}{2}$ $2$ $\frac{7}{2}$ $2$ $2$ B
An example with $M=6$ $1$ $1$ $1$ $6$ $10$ $\frac{32}{3}$ $-\frac{2}{3}$ $8$ $0$ F
An example with $M=6$ $2$ $1$ $0$ $6$ $10$ $\frac{22}{3}$ $\frac{8}{3}$ $6$ $2$ F
: Examples of excitations in our [FTI]{}. The electric and magnetic charges are explicitly shown. “F” is short for “fermionic” where $\Gamma$ is odd. “B” is short for “bosonic” where $\Gamma$ is even. We call an elementary charge an intrinsic excitation (Definition \[dfn\_intrinsic\_exc\]) that [carries]{} $Q=1/3$ or $Q=2/3$ EM electric charge, in analogy to the fractionalized charge excitation[s]{} in [the]{} $\nu=1/3$ [FQH]{} state. [The]{} two elementary charges in the Table are just two concrete examples, and there are many other excitations that carry $Q=1/2,2/3$ and $M=0$. [The]{} elementary EM monopole is an excitation that carries the minimal nonzero EM magnetic charge $M=3$ and do[es]{} not contain any partons (i.e., $n^{fi}=0$, $\forall i=1,2,3$). [A]{} nonzero $M$ can be externally added into the bulk in order to probe the EM response (see Definition \[dfn\_intrinsic\_exc\]). [The]{} minimal quantum of charge fractionalization is $\frac{1}{3}$ determined by the [intrinsic excitations]{}, i.e., Eq. (\[Q\_fractional\_M0\]) rather than Eq. (\[Q\_fractional\_M1\]).[]{data-label="table:FTI"}
An important result is that the [FTI]{} with $\Theta=-\frac{5}{9}\pi$ is actually topologically equivalent to the [FTI]{} with $\Theta=\frac{1}{9}\pi$ by [a]{} periodic shift. The minimal choice of $\Theta$ for our [FTI]{} phase is given by: $$\begin{aligned}
\Theta=\frac{1}{9}\pi \text{ mod }\frac{2}{9}\pi\,.\label{eqn:new_Theta_reduced}\end{aligned}$$ [To understand]{} this result, let us revisit the self-statistics $\Gamma$ in Eq. (\[eq:stat4321\]). In fact, $\Gamma$ can be reformulated as a unique function of $M$ and $Q$: (see Appendix \[appendix\_gamma\_stat4321\_new\] for details) $$\begin{aligned}
\Gamma=&3(M+1)(Q-\frac{\Theta}{2\pi}M)\,,\label{eq:stat4321_new}\end{aligned}$$ where $\Theta$ is given by Eq. (\[eqn:new\_Theta\_reduced\]). We manifestly see that the even-odd property of $\Gamma$ is unaltered by the *minimal* shift $\Theta\rightarrow \Theta+\frac{2}{9}\pi$. $\Gamma$ is pictorially illustrated in Fig. \[figure\_theta\](b) and we can see that geometrically, $\Theta$ describes how tilted the charge lattice is with respect to its initial orientation (Fig. \[figure\_theta\](a)). To illustrate, the red dashed line in Fig. \[figure\_theta\](b) can be either more or less tilted with respect to the vertical axis via a shear deformation. The charge lattice \[Fig. \[figure\_theta\](b)\] with a nonzero $\Theta$ can be obtained through such a shear deformation from the non-tilted charge lattice \[Fig. \[figure\_theta\](a)\]. Since $\Theta=\frac{1}{9}\pi$, the charge lattice shown in Fig. \[figure\_theta\](b) is time-reversal invariant ($Q\rightarrow Q, M\rightarrow -M, \Gamma\rightarrow \Gamma+\text{even integer}$), which can be viewed as a reflection symmetry about $Q$-axis.
It is obvious that the entire charge lattice \[Fig. \[figure\_theta\](b)\] as well as the self-statistics distribution is unaltered if we further increase $\Theta$ by $\frac{2}{9}\pi$ (i.e. increase $\tan\alpha$ by $1/9$). For this reason, $\Theta$ is well defined only mod $\frac{2}{9}\pi$ as shown in Eq. (\[eqn:new\_Theta\_reduced\]). For example, the bosons on the site $(\frac{1}{6},3)$ are shifted to the bosons on the site $(\frac{1}{2},3)$; the fermions on the site $(0, 6)$ are shifted to the fermions on the site $(\frac{2}{3},6)$. Furthermore, since the charge lattice is actually 4-dimensional (Definition \[dfn\_excitation\]), each lattice site of Fig. \[figure\_theta\]-(b) actually corresponds to many excitations that are different from each other by $N^a_m,N^b_m$ as shown in Fig. \[figure\_theta\]-(c) where $Q=\frac{1}{6},M=3$ is illustrated. The lattice sites in Fig. \[figure\_theta\]-(c) follow a simple relation: $({N^b_m-N^a_m-1})/{3}\in\Z$ where $N^a_m,N^b_m$ are even. (see Appendix \[appendix\_gamma\_stat4321\_new\] for details.)
![(Color online) Self-statistics distribution on the 4D charge lattice. (a) Self-statistics distribution as a function of $M$ and $Q$ by turning off $\Theta$ in Eq. (\[eq:stat4321\_new\]). (b) Self-statistics distribution as a function of $M$ and $Q$ as shown in Eq. (\[eq:stat4321\_new\]). The allowed values of $M$ and $Q$ are determined by Eqs. (\[Q\_fractional\_M0\],\[Q\_fractional\_M1\]). Geometrically, $\Theta=2\pi\tan\alpha$, where $\tan\alpha=\frac{1/6}{3}=1/18$. Thus, $\Theta$ angle can be viewed as a consequence of a shear deformation from (a) to (b). During the shear deformation, the area of “Dirac unit cell” (denoted by the shaded area) is invariant. Since the charge lattice is 4D (Definition \[dfn\_excitation\]), each site in (b) on the $(M-Q)$ parameter space corresponds to more than one excitation. An example is shown in (c), where $N^a_m,N_m^b$ are used to [label]{} excitations that have the same $Q$ and $M$: $Q=\frac{1}{6},M=3$.[]{data-label="figure_theta"}](figure_theta.pdf){width="8.5cm"}
Experimentally, one may understand the physics of $\Theta$ via the surface quantum Hall effect on a surface with broken time-reversal symmetry. For example, by placing a ferromagnetic thin film on top of the surface of [a]{} [FTI]{}, we may observe [a]{} Hall effect with Hall conductance $\sigma_H=\frac{\Theta}{2\pi}\frac{e^2}{h}$ [@Qi2008]: $$\begin{aligned}
\sigma_H=\left(\frac{1}{18}+\frac{n}{9}\right)\frac{e^2}{h}\,,n\in\Z\,,\end{aligned}$$ where $h$ is the usual Planck constant. It should be kept in mind that, although the minimal nonzero $\sigma_H$ is $\frac{1}{18}$, the corresponding charge induced by the Hall response is not $\frac{1}{18}$ since the quantization of $M$ in Eq. (\[double\_m\_charge\]) is modified from its non-fractionalized value. More precisely, an external EM monopole with $M$ magnetic charge can be viewed as $2\pi M$ EM magnetic fluxes threading the surface [@bti5]. By using [the]{} Laughlin argument, the surface will generate $\frac{1}{18}M$ response charge once the EM monopole penetrates the surface. Since the minimal nonzero $M$ is $3$ due to Eq. (\[double\_m\_charge\]), the minimal surface response charge is $\frac{1}{18}\times 3=\frac{1}{6}$ rather than $1/18$. Further, the $\frac{1}{6}$ charge will be attached onto the EM magnetic monopole that moves into the [FTI]{} bulk, which renders the Witten effect [@witten1; @Qi2008; @franz]. This phenomenon is nontrivial in a sense that the $\frac{1}{6}$ charge cannot be formed by the bulk intrinsic excitations (Definition \[dfn\_intrinsic\_exc\]) whose $Q$ is quantized at $\frac{1}{3}$ due to Eq. (\[Q\_fractional\_M0\]).
[An]{} [FTI]{} can be viewed as a symmetry-enriched topological phase [(SET)]{} which [is]{} characterized by both bulk topological order (TO) data and a symmetry action. In our case, the latter is encoded by the structure of the charge lattice in Fig. \[figure\_theta\]. The former is given by [the set of]{} all intrinsic excitations (Definition \[dfn\_intrinsic\_exc\], i.e., all sites along the $Q$-axis in Fig. \[figure\_theta\]) and also loop excitations (Definition \[dfn\_loops\]). With these preliminaries, we may discuss the consequence of stacking operations in the context of topological order [@string8; @wen_stacking]. [Stacking operations]{}, denoted as $\boxtimes$, form a monoid that does not contain inverse elements. It is known that stacking two TIs leads to [the topologically trivial vacuum state]{}: $\text{TI}\boxtimes \text{TI}=\text{Vacuum}
$. Let us stack a 3D TI and a 3D [FTI]{} together. The [resulting]{} phase is a TO: $\text{FTI}\boxtimes \text{TI}=\text{TO}
$. In other words, the stacking operation removes the nontrivial Witten effect of the [FTI]{}, rendering a [state with]{} pure topological order. This can be understood [in]{} two steps. First, since the bulk intrinsic excitations of a TI only contain electron excitations, the above stacking operation indeed does not change the TO of the [FTI]{}. Second, in the stacked phase, the net EM electric charge $Q$ is given by $Q=\left(\frac{\pi}{18}M+\frac{n}{3}\right)+\left(\frac{\pi}{2}M+n'\right)
$ [with $n,n'\in\Z$]{}, where the first term is given by [the]{} Witten effect of [the]{} [FTI]{} while the second term is given by [the]{} Witten effect of [the]{} TI. Since the stacked phase [is]{} formed by putting [the]{} [FTI]{} and TI in [the]{} same [spatial]{} 3D region, the quantization of $M$ in Eq. (\[double\_m\_charge\]) still holds in the stacked phase. As a result, [the electric charge]{} $Q$ [in]{} the stacked phase is given by: $Q=\frac{5}{9}M+\frac{n}{3}+n'$, where the $M$-induced charge $\frac{5}{9}M$ is quantized [to]{} $5/3$. This charge can be [completely]{} screened by $\frac{n}{3}+n'$, e.g.[,]{} $n=-2\,,\,n'=-1$. Thus, the charge lattice of the stacked phase is not tilted, meaning that $\Theta=0$.
In summary, the stacking of a [FTI]{} and a TI leads to a [phase with]{} pure topological order where [the]{} Witten effect is absent. We may also consider stacking [two]{} [FTI]{}s: $
\text{FTI}\boxtimes \text{FTI}=\text{TO}\boxtimes \text{TO}$, which means that the stacked phase is a purely topologically ordered phase where the charge lattice is not tilted and the topological order is given by $\text{TO}\boxtimes \text{TO}$. Surely, this is just an example while [it]{} is possible that other examples of [FTI]{} may [produce different phases when]{} stacked together with TI or [with]{} themselves. [S]{}tacking operations in SET phases [generally]{} change TO to a new topological order denoted as “$\widetilde{\text{TO}}$”. For example, stacking two [FTI]{}s here gives rise to $\widetilde{\text{TO}}=\text{TO}\boxtimes \text{TO}$. In order to see if the [resulting]{} phase is a new SET or not, one should further consider symmetry-respecting condensations that change $\widetilde{\text{TO}}$ back to TO. [In]{} this way, we may make progress toward the classification of SETs. As it is beyond the scope of the present work, we will leave this issue to further studies.
The above calculation is based on concrete numerical inputs (\[define\_KNq12345\]). As [mentioned previously]{}, one may in principle generically deduce $\Theta$ as a function of the parameters $(u,v,u',v',q,q',\theta)$ that fully determine the two permissible composite condensations and the [entire]{} bulk spectrum. In 2D, we know that some FQH states can be unified into Jain’s sequence [@composite_Jain_1; @composite_Jain_2] such that they can be understood in the composite-fermion theory with different microscopic designs of the composite particles. Our 3D composite particle theory is similar to this 2D scenario: the $\Theta$ angle, and other properties of composite condensation phases, are also determined by the different designs of composite condensations. Therefore, all phases constructed from composite condensations can be thought to form a sequence. We expect more studies in the future along this line of thinking will be helpful in uncovering the physics of 3D Abelian topological phases of strongly interacting fermion systems.
Deconfined descrete gauge subgroup $\Z_2\times \Z_6$: Abelian topological order in the bulk {#sec_FTI_plus}
-------------------------------------------------------------------------------------------
The FTI state obtained in Sec. \[sec\_FTI\] supports fractionalized intrinsic excitations as indicated by Eq. (\[Q\_fractional\_M0\]) and the texts around it. Ref. ever pointed out that FTIs necessarily has a fractionalized bulk. Therefore, our construction is consistent to the claim. Usually, a fractionalized gapped bulk can be understood as the presence of a topological order of some form. To see more clearly the exact form of the topological order of our FTI, let us start with the $K$ matrix in Eq. (\[define\_KNq12345\]). By using two independent unimodular matrices (i.e., $\Omega$ and $W$ that will be discussed in details in Sec. \[sec:charlos\_defect\_symmetry\]), we may diagonalize $K$: $$\begin{aligned}
& \Omega K W^T=\bpm 2 & 0 \\ 0& 6\epm, \\
&\text{where }\Omega=\bpm 1&0\\-2&1 \epm\,,\,W=\bpm 1&0\\-1&1 \epm\,.\end{aligned}$$ In the new basis, it is clear that the bosonic sectors of the ground state are described by deconfined $\Z_2\times\Z_6$ gauge group. In other words, the maximal torus $U(1)\times U(1)$ of the $SU(3)$ gauge group of the parton construction is confined *except* the $\Z_2\times\Z_6$ gauge subgroup. In Ref. , the discrete gauge group $\Z_2$ arises since the choice of parton mean-field Hamiltonian explicitly breaks the original pseudospin $SU(2)$ gauge group down to $\Z_2$ subgroup. However, in our FTI state, the discrete gauge subgroup arises from the the deconfined subgroup of a confined non-Abelian gauge group, physically due to the condensation of composites that contain magnetic monopoles.
Charge-loop excitation symmetry and its relation to extrinsic twist defects {#sec:charlos_defect_symmetry}
===========================================================================
In Sec. \[sec:theta\], we have explored the axion angle of the charge lattice with composite condensation. In this section, we will explore the charge-loop excitation symmetry based on the composite particle theory introduced in Sec. \[sec:gauge\_strc\].
The topological $BF$ field theory (\[BFaction\]), which is derived from the two permissible composite condensates, only captures the statistical interaction between particles that carry $N^a_m,N^b_m$ magnetic charges and loops that carry $\Phi_e^a,\Phi^b_e$ electric fluxes. Specifically, several important properties of composites, such as the self-statistics $\Gamma$ in Eq. (\[eq:quantum\_stat\]), and the net EM electric charge $Q$ in Eq. (\[NE1\]), are not encoded in Eq. (\[BFaction\]). However, the topological $BF$ field theory reproduces $Q_{\rm Debye}$, an important part of $Q$. In this section, we further study the topological $BF$ field theory and show that it serves as a useful platform to study “*Char*ge-*Lo*op *E*xcitation *S*ymmetry” (abbreviated as “$\mathsf{Charles}$”, see Definition \[dfn\_charlos\]) that can be viewed as a 3D generalization of “anyonic symmetry” [@Teo2015] (or “topological symmetry” in Ref. and references therein) in 2D Abelian topological phases. We expect that 3D Abelian topological phases where loop excitations are allowed may host even more exotic physics if extrinsic twist defects are imposed, and anticipate that 3D charge-loop excitation symmetry will be a useful tool in future studies of such extrinsic defects.
Definition of $\mathsf{Charles}$ {#sec:charlos}
--------------------------------
In 2D topological phases, each point-like extrinsic twist defect is associated with an element of an anyonic symmetry group $G$. The anyonic symmetry group is a finite group that acts to permute a subset of anyons in the parent TO phase while preserving all of the topological properties (topological spin, statistics) of the anyons (and sometimes their symmetry properties as well, e.g., their EM charge). For example, the permutation of $e$ and $m$ particles in the 2D Wen-plaquette model (with $\Z_2$ TO)[@Wen2003] is a typical anyonic symmetry transformation. For this particular model this transformation can be realized by extrinsically imposing a lattice dislocation which enacts the permutation of $e$ and $m$ when an anyon passes through a 1D branch cut that terminates at the extrinsic point defect [@Kitaev2006; @Bombin2010; @You2012; @Teo2015]. Interestingly, this quasiparticle permutation mechanism endows the dislocation with an attached non-Abelian object at the defect core, which opens up a possible new platform for topological quantum computation. Mathematically speaking, the incorporation of extrinsic defects into 2D Abelian topological phases described by a category theory $\mathcal{C}$, promotes $\mathcal{C}$ to a $G$-crossed tensor category theory $\mathcal{C}^{\times}_{G}$ [@Barkeshli2014; @Teo2015].
Let us briefly recall some properties of anyonic symmetry in 2D Abelian topological phases. As mentioned, these phases are described using Abelian Chern-Simons theory using the data in a symmetric, integer $K$-matrix. There is an important class of unimodular, integer transformations $W$ satisfying $WKW^{T}=K$ that act as the automorphisms of $K$ (or the automorphisms of the integer lattice, and dual/quasiparticle lattice, determined by $K$). These transformations relabel the different anyonic excitations, but most of them preserve the anyon type, and just attach local quasiparticles (e.g., attaching extra electrons). These trivial transformations are called the inner automorphisms ${\mathsf{Inner}(K)}$ and they form a normal subgroup of the full set of automorphisms $\mathsf{Auto}(K).$ The non-trivial anyonic relabeling symmetries are hence given by the group $G\equiv\mathsf{Outer}(K)=\frac{\mathsf{Auto}(K)}{\mathsf{Inner}(K)}.$ This captures the conventional anyonic symmetries that act as point-group operations on the quasiparticle lattice, although it leaves out possible non-symmorphic lattice operations or symmetries of stably-equivalent $K$-matrices [@Teo2015; @Teo2014; @cano2013]. We will not consider these more complicated possibilities for anyonic symmetries any further and leave their 3D generalization to future work.
In order to generalize this discussion of anyonic symmetry and extrinsic defects to 3D, let us revisit some basic facts of excitations in our 3D fermionic gapped phase formed by two permissible composite condensates. The 2D vectors $\mathbf{L}$ form a 2D loop-lattice in Definition \[dfn\_loops\]. The 2D vectors $\mathbf{N}_m$ form a 2D lattice which is a sublattice of the 4D charge lattice in Definition \[dfn\_excitation\]. As a whole, we may define a 6D *charge-loop-lattice* (N.B., this is not the same 6D lattice mentioned earlier).
\[[Charge-loop-lattice]{}\] The charge-loop-lattice is a 6D lattice whose sites are given by the 6D lattice vector $\vec{V}=(N_A,\mathbf{N}^T_m,M,\mathbf{L}^T)=(N_A,N^a_m,N^b_m,M,\ell,\ell')$. Each site corresponds to a *charge-loop composite*.\[dfn\_cloop\]
In order to avoid confusions in terminology, the word “composite,” if used by itself, always denotes a point-like particle, unless otherwise specified. The symmetry group $\mathsf{Charles}$ is then defined as below:
\[[Charge-Loop Excitation Symmetry]{} ($\mathsf{Charles}$)\] The charge-loop excitation symmetry group is a subset of the point group of the 6D charge-loop-lattice and corresponds to the following quotient group: $$\begin{aligned}
\mathsf{Charles}=\frac{\mathsf{Auto}(K)}{\mathsf{Inner}(K)}\,,\label{auto_inner}\end{aligned}$$ where $\mathsf{Auto}(K)$ is the group of generalized automorphisms of $K$. $\mathsf{Inner}(K)$ is the group of generalized inner automorphisms of $K$, which is a subgroup of $\mathsf{Auto}(K)$. Group elements of $\mathsf{Auto}(K)$ have the matrix representation $\mathscr{G}=W\oplus \Omega$, where the two independent rank-two unimodular matrices $W$ and $\Omega$ satisfy the following two conditions: $$\begin{aligned}
&(i).~~~~ \Omega K W^T=K\,,\label{eqn:auto}\\
&(ii).~~~~\Gamma(\cdots, {\mathbf{N}_m},\cdots)=\Gamma (\cdots,W^{-1}\mathbf{N}_m,\cdots)\,.\label{eqn:auto1}\end{aligned}$$ Here, $\Gamma$ is the self-statistics of composites, which is a function of lattice sites labeled by the 4D coordinates ($N_A,\mathbf{N}^T_m,M$). In addition to conditions (i) and (ii), the group elements in $\mathsf{Inner}(K)$ have the property that $W^{-1}\mathbf{N}_m-\mathbf{N}_m=K^T\,(n_1,n_2)^T$ and $\Omega^{-1}\mathbf{L}-\mathbf{L}=K(n_3,n_4)^T$, where $n_1,\cdots,n_4$ are integers. $n_1$ and $n_2$ are functions of $\mathbf{N}_m,W$; $n_3$ and $n_4$ are functions of $\mathbf{L},\Omega$. \[dfn\_charlos\]
Just like the 2D anyonic symmetry group, the definition of $\mathsf{Charles}$ also involves the definitions of $\mathsf{Auto}(K)$ and $\mathsf{Inner}(K)$. One can prove that $\mathsf{Auto}(K)$ satisfies the usual group axioms (identity element, inverse element, closure, associativity) and that $\mathsf{Inner}(K)$ is a normal subgroup of $\mathsf{Auto}(K)$, such that $\mathsf{Charles}$ forms a group. Details of this proof can be found in Appendix \[appendix\_proof\_group\].
Physically, group elements $\mathscr{G}=W\oplus\Omega$ in $\mathsf{Charles}$ correspond to point group transformations: $({\mathbf{N}_m})_{\text{new}} = W^{-1} \mathbf{N}_m, \,({\mathbf{L}})_{\text{new}} = \Omega^{-1}\mathbf{L}$. Conditions (i) and (ii) guarantee that the transformed charge-loop-lattice is identical to the original one, which means that $\mathsf{Charles}$ keeps not only the lattice geometry invariant, but also leaves all topological properties of particle excitations and loop excitations (denoted by lattice sites) unaffected. Those topological properties include the self-statistics of particle excitations $\Gamma$, the charge-loop braiding statistics $\vartheta^{cl}$, and the Debye screening $Q_{\rm Debye}$. However, there is a redundancy corresponding to $\mathsf{Inner}(K)$ that should be removed. $\mathsf{Inner}(K)$ includes all trivial transformations whose point-group effects are equivalent to effectively shifting both $\mathbf{N}_m$ and $\mathbf{L}$ by undetectable amounts (i.e., $\vartheta^{cl}=0 \text{ mod } 2\pi$ in Eq. (\[eq:cltheta\]); see also Sec. \[sec:loop\_loop\]), and thereby must be modded out from $\mathsf{Auto}(K)$ if we only want to keep non-trivial transformations. Again, the transformations in $\mathsf{Inner}(K)$ can be interpreted as changing the excitations by a trivial, topologically-undetectable charge or flux.
In contrast to the 2D definition of “anyonic symmetry” where condition (i) (where the simpler structure only allows for $W=\Omega$) is enough to guarantee the invariance of the self-statistics of anyons, one now needs condition (ii) in order to guarantee that the self-statistics of excitations on the 4D charge lattice remains invariant under $\mathsf{Charles}$ transformations. The main reason for this is that the self-statistics of an excitation (Definition \[dfn\_excitation\]) cannot be captured by the topological $BF$ field theory. *Whether or not $W$ satisfies condition (ii) relies on the specifics of the parton decomposition, and in the following subsections, we will assume condition (ii) is satisfied.*
General theory of $\mathsf{Charles}$ and its tensor-network-type representation {#sec:general_charlos}
-------------------------------------------------------------------------------
It should be noted that $W$, $\Omega$, and $K$ in Definition \[dfn\_charlos\] can be naturally generalized to arbitrary rank if a physical realization using a scenario having any number of permissible composite condensates in Sec. \[sec:bec\] can be achieved. For example, one can consider a single composite condensate or three linearly independent condensates with entirely different parton constructions, which leads to a number $K\in\Z$ or a rank-three $K$ matrix respectively.
Before proceeding further, we introduce a simplified notation that will be useful for subsequent discussions. The notation in the two-component $BF$ action (\[BFaction\]), such as $\tilde{a}$ and $\tilde{b}$, comes from the specific physical realization described in Sec. \[sec:bec\]. It is however inconvenient for the purpose of generalizing $\mathsf{Charles}$. Thus, in the current section (Sec. \[sec:general\_charlos\]), we temporarily use a new notation for the gauge fields: $b=(b^1,b^2,\cdots)$ and $a=(a^1,a^2,\cdots)$ where $b$ is a set of $2$-form Kalb-Ramond $U(1)$ gauge fields while $a$ is a set of $1$-form $U(1)$ gauge fields. As a result, the topological $BF$ term is expressed as: $$\begin{aligned}
\frac{iK^{IJ}}{2\pi}\int b^{I}\wedge da^J=\frac{i}{2\pi}\int b^T\wedge K da
\end{aligned}$$ with a square matrix $K$ of rank $N$. The excitation terms in Eq. (\[BFex\]) are rewritten as: $$\begin{aligned}
\mathcal{S}_{ex}=i \int \mathbf{t}^T a\wedge \star j+i \int\mathbf{L}^T b \wedge \star \Sigma\,,
\end{aligned}$$ where $\mathbf{t}=(t_1,t_2,\cdots)$ is an integer vector replacing the notation $\mathbf{N}_m$. Then, the charge-loop-lattice is formed by an $N$-dimensional charge lattice labeled by vectors $\mathbf{t}$ and an $N$-dimensional loop-lattice labeled by vectors $\mathbf{L}$. Group elements of $\mathsf{Charles}$ are still denoted as “$\mathscr{G}=W\oplus \Omega$” with the transformations: $(\mathbf{t})_{\text{new}} = W^{-1}\mathbf{t}$ and $(\mathbf{L})_{\text{new}} = \Omega^{-1}\mathbf{L}$.
Let us consider some examples. In Table \[table:charlos\], all possible $\mathsf{Charles}$ groups are listed for a $1\times 1$ matrix $K\in\Z$. From the table, we see that $\mathbb{Z}_2$ gauge theory in (3+1)D $(K=2)$ only has trivial $\mathsf{Charles}$, which is surprisingly different from a deconfined $\mathbb{Z}_2$ gauge theory in (2+1)D (e.g., as appears in the Wen-plaquette model), where the $e\leftrightarrow m$ exchange process is an anyonic symmetry transformation. Nontrivial $\mathsf{Charles}$ arises for $\mathbb{Z}_K$ gauge theory in (3+1)D *only* when $|K|\geq 3$. For example, for $\mathbb{Z}_3$ gauge theory, the nontrivial element of $\mathsf{Charles}$ is $-\mathbb{I}\oplus-\mathbb{I}$ which means that $W=\Omega=-\mathbb{I}$ (here $\mathbb{I}$ reduces to the natural number “1”). Under the transformation of this group element, there is an exchange symmetry between a particle with one unit of gauge charge and a particle with two units of gauge charge since the latter is trivially equivalent to a particle with gauge charge $-1$. There is also an exchange symmetry between a loop with magnetic flux $2\pi/3$ and a loop with magnetic flux $4\pi/3$ ($=-\frac{2\pi}{3}+2\pi$). These two exchange processes must occur *simultaneously*.
$K$ $\mathsf{Charles}$
------------- -------------------------------------------------------------
$K=\pm 1$ $\{\mathscr{G}_{\mathbb{I}}\}$
$K=\pm 2$ $\{\mathscr{G}_{\mathbb{I}}\}$
$|K|\geq 3$ $\{\mathscr{G}_{\mathbb{I}},-\mathbb{I}\oplus-\mathbb{I}\}$
: Examples of $\mathsf{Charles}$ (Sec. \[sec:general\_charlos\]) when the matrix $K$ reduces to an integer. A generic group element is denoted by $\mathscr{G}=W\oplus \Omega$. $\mathscr{G}_{\mathbb{I}}$ denotes the identity element: $\mathscr{G}_{\mathbb{I}}=\mathbb{I}\oplus \mathbb{I}$. []{data-label="table:charlos"}
A simple example of a rank-2 $K$ matrix is $K=2\sigma_x$. If we do not worry about $\mathsf{Charles}$ for a moment, a diagonalization can be achieved by using $W=\sigma_x, \Omega=\mathbb{I}_{2\times 2}$. In the new basis, we end up with two copies of the level-2 topological $BF$ field theory, thereby obtaining a $\mathbb{Z}_2\times\mathbb{Z}_2$ discrete gauge theory (i.e., $\mathbb{Z}_2\times\mathbb{Z}_2$ topological order). Due to Definition \[dfn\_charlos\], such a basis change is clearly not a group element of $\mathsf{Charles}$, but it reveals that the gauge structure is $\mathbb{Z}_2\times\mathbb{Z}_2$ rather than $\mathbb{Z}_4$. It is important to distinguish these possibilities since those two gauge structures produce the same ground state degeneracy ($\mathsf{GSD}$) on a 3-torus [@bf1; @bf2; @bf3; @horowitz89]. For this example, a typical group element of $\mathsf{Charles}$ is: $\mathscr{G}= \sigma_x\oplus \sigma_x$, which satisfies condition (i) in Eq. (\[eqn:auto\]). Physically, $\Omega$ exchanges a particle labeled by $\mathbf{t}=(0,1)^T$ and a particle labeled by $\mathbf{t}=(1,0)^T$. At the same time, $\Omega$ exchanges a loop labeled by $\mathbf{L}=(0,1)^T$ and a loop labeled by $\mathbf{L}=(1,0)^T$.
For convenience, condition (i) in Eq. (\[eqn:auto\]) can be visually represented by a tensor network-type graph as shown in Fig. \[figure\_tensor\](a) of Appendix \[appendix\_proof\_group\]. It indicates that $K$ is a fixed point tensor (here, a matrix) that is invariant under $\mathsf{Charles}$ renormalization-group-like transformations. The bond dimension is given by the rank of $K$. This graphical representation allows us to straightforwardly generalize the notion of $\mathsf{Charles}$ to more general Abelian topological quantum field theories (TQFTs) in (3+1)D that include more exotic topological terms. For instance, let us consider a TQFT with the action: $$\begin{aligned}
S=\frac{i}{2\pi}\int b^T\wedge K da+ i\int \Lambda^{IJK}a^I \wedge a^J \wedge da^K\,,\end{aligned}$$ where $\Lambda^{IJK}$ is a real tensor with three legs as shown in Fig. \[figure\_tensor\](b) of Appendix \[appendix\_proof\_group\]. By itself, and at a classical level, the second term in this action corresponds to a topological invariant for the mutual linkage of three electromagnetic flux loops [@bod]. At a quantum level, the action $S$ was also proposed as a continuum field theory description of Dijkgraaf-Witten lattice gauge theory [@dwitten; @corbodism3]. It can also be derived by gauging the global on-site symmetry group $G=\mathbb{Z}_{N_1}\times \mathbb{Z}_{N_2}\times \cdots$ of the TQFT action of a 3D SPT phase with $N_1\times N_2\times \cdots=|\text{det}K|$, where the quantization of $\Lambda^{IJK}$ is determined by the number of topologically distinct ways to impose $G$ in SPT phases [@YeGu2015]. The relation to 3-loop statistics [@wang_levin1; @levin_talk; @wang_levin2; @lin_levin] is being investigated [@YeGu2015; @wang_levin1; @levin_talk; @3loop_ryu]. It is believed that the coefficient $\Lambda^{IJK}$ encodes the information of $3$-loop statistics that classifies topologically distinct twisted discrete Abelian gauge field theories in (3+1)D. In analogy to the topological $BF$ field theory, the tensor $\Lambda^{IJK}$ must also be transformed accordingly under the charge-loop-lattice point group transformations. In order to keep the important 3-loop statistics data [@wang_levin1; @levin_talk] invariant, the generalized $\mathsf{Charles}$ should incorporate the following new condition: $$\begin{aligned}
\sum_{I'J'K'}W^{I'I}W^{J'J}W^{K'K}\Lambda^{I'J'K'}=\Lambda^{IJK}\label{eqn:auto33}\end{aligned}$$ in addition to those conditions in Definition \[dfn\_charlos\]. Likewise, we can also use a tensor-network-type graph \[Fig. \[figure\_tensor\](b) of Appendix \[appendix\_proof\_group\]\], to graphically represent Eq. (\[eqn:auto33\]), where the bond dimension is no less than two.
Finally, we can also consider a TQFT with the action: $$\begin{aligned}
S=\frac{i}{2\pi}\!\int \!b^T\!\wedge K da+ i\!\!\int \!\Xi^{IJKL}a^I \wedge a^J \wedge a^K\wedge a^L,
\end{aligned}$$ where the coefficient $\Xi^{IJKL}$ is a tensor with four legs. Likewise, the quantized values of $\Xi$ encode the information of the four-loop braiding process [@wang_levin2] and can provide topological invariants for classifying twisted discrete gauge field theories in (3+1)D. In order to keep $\Xi$ invariant under point-group transformations of the charge-loop-lattice, the following relation should be obeyed: $$\begin{aligned}
\sum_{I'J'K'L'}\!W^{I'I}W^{J'J}W^{K'K}W^{L'L}\Xi^{I'J'K'L'}=\Xi^{IJKL}.\label{eqn:auto44}\end{aligned}$$ A tensor-network-type representation is shown in Fig. \[figure\_tensor\](c) of Appendix \[appendix\_proof\_group\], where the bond dimension is no less than four.
We may also consider the scenario that $\mathsf{Charles}$ transformations can be performed *locally* so that $\mathsf{Charles}$ becomes dynamically *gauged*. In this case, the extrinsic twist defects become well-defined, deconfined excitations of a new topological phase. The resulting phases have been thoroughly studied in 2D and are non-Abelian topological phases called twist liquids [@Teo2015; @Teo2014; @ran; @barkeshli_wen; @Teo2013; @Barkeshli2014; @Bombin2010; @genon_1; @genon_2; @genon_3; @genon_4; @You2012]. As a result, $W$ and $\Omega$ become space-time dependent. The difference between the next-nearest lattice sites is compensated by locally twisting matter fields. The tensor-network graph representations of the various symmetry transformations in Fig. \[figure\_tensor\] of Appendix \[appendix\_proof\_group\] are suggestive that such a tensor-network analysis may be a useful tool for future studies of 3D twist liquids.
![(Color online) Extrinsic twist defects in 3D. (a) line defect; (b) point defect. The two cubic boxes denote the 3D bulk of an underlying quantum many-body system. The shaded plane in (a) denotes a 2D branch cut/plane ending at the line defect, while the dashed line in (b) denotes a 1D branch cut ending at the point defect. A line defect can act on both composite particles denoted by a black dot, and loops denoted by a red circle. Once a point-like excitation and a loop excitation move around a line defect, the defect performs the $\mathsf{Charles}$ symmetry transformation $\Omega$ and $W$ on the point-like excitation and the loop excitation, respectively. In (b), the loop moves around the point defect such that the branch line intersects at the loop’s spatial trajectory (a torus) only once. A $\mathsf{Charles}$ transformation induced by a point defect can only be $\mathscr{G}=W\oplus\Omega=\mathbb{I}\oplus\Omega$, which acts only on the loop excitations. However, due to Eq. (\[eqn:auto\]), the only candidate for $\Omega$ is $\mathbb{I}$. This means that point defects can only behave like the identity element $\mathscr{G}_{\mathbb{I}}$ of $\mathsf{Charles}$. Therefore, in 3D, we only consider line defects.[]{data-label="figure_defect"}](figure_defect.pdf){width="8.5cm"}
Theory of $\mathsf{Charles}$-defects: twist defect species and fusion {#sec:defect_fusion}
---------------------------------------------------------------------
{width="17.5cm"}
In the following, we study extrinsic twist defects associated with $\mathsf{Charles}$ group elements in analogy to extrinsic twist defects in 2D Abelian topological phases with anyonic symmetry [@Teo2015; @Teo2014; @Teo2013; @Barkeshli2014; @genon_1; @genon_2; @Bombin2010; @You2012]. More specifically, we explore two issues: (i) the universal labeling of a defect in 3D, and (ii) the fusion properties of defect-charge-loop composites. Recently remarkable progress in the study of various aspects of string/loop excitations in (3+1)D topological phases of matter, such as their description using lattice models and field theories, their associated ground-state degeneracy ($\mathsf{GSD}$), and their braiding and fusion properties has been made[@string1; @string1.5; @wang_levin1; @string2; @string3; @string4; @ran; @string5; @string6; @string7; @string8; @string9; @string10]. As will be seen below, the consideration of extrinsic defects within the framework of charge-loop excitation symmetry introduces a new aspect to the physics of loop excitations in (3+1)D topological phases.
We begin by reviewing the physics of twist defects and defect-anyon composites in 2D Abelian topological phases. From a topological point of view, externally imposed defects in such phases are a set of special, potentially non-Abelian objects. For Abelian groups, each group element of the anyonic symmetry group $G$ corresponds to a bare defect and there are $\mathsf{ord}(G)$ distinct bare defects, where $\mathsf{ord}(G)$ is the order of $G$. Generically, the defects are labeled by the conjugacy classes of $G,$ but since we only deal with Abelian groups here we will not often make this distinction. By a “bare” defect, we really mean that the defect is externally imposed alone in the bulk. In general, defects can be bound to anyons of the parent Abelian topological phase, thereby forming a *defect-anyon composite* which is, by definition, not bare. When given a group element of $G$, the total number of topologically distinct defects (also known as defect species) includes bare ones and composite ones, and is not always the same as the number of topologically distinct anyons. In other words, two defect-anyon composites might be topologically equivalent to each other if there do not exist gauge-invariant Wilson measurements that can distinguish them. Indeed, there is a consistency equation for determining the equivalence classes of defect types [@Teo2014; @Teo2015], $$\begin{aligned}
\mathcal{D}^0_\mathscr{G} \times qp =\mathcal{D}^0_\mathscr{G}\times
(qp+(\mathbb{I}-\mathscr{G})qp_1) \,,\label{equation_fusion_2d}\end{aligned}$$ which is diagrammatically shown in Fig. \[figure\_defect\_composite\](a). Here, $qp$ denotes a quasiparticle (i.e., an anyon) that is provided by the parent 2D Abelian topological phase. $\mathcal{D}^0_\mathscr{G}$ denotes a bare defect labeled by a group element $\mathscr{G}$ of the anyonic symmetry group $G$. The particle $qp_1$ is any anyon provided by the parent 2D Abelian topological phase. The composite object $ \mathcal{D}^0_{\mathscr{G}}\times qp$ denotes the fusion between $\mathcal{D}^0_{\mathscr{G}}$ and $qp$ that forms a defect-anyon composite. Specifically, Eq. (\[equation\_fusion\_2d\]) determines when this defect-anyon composite is topologically identical to a defect-anyon composite that is formed by the fusion between the same bare defect and a different anyon given by $qp+(\mathbb{I}-\mathscr{G})qp_1$. The symbol “$+$” should be regarded as the addition of quasiparticle vectors in the $K$-matrix Chern-Simons theory. The physical reason of this equivalence is really due to the nontrivial internal structure of a defect-anyon composite. More specifically, the anyon $qp$ that is trapped at the defect can emit anyon $qp_1$ which moves around the defect once. As a result, anyon $qp_1$ is changed to anyon $\mathscr{G}qp_1$ that is finally absorbed by the defect. Such a process occurs inside the defect-anyon composite and cannot change the defect species [@Teo2014]. Therefore, the process provides an equivalence between two defect-anyon composites.
A typical example in (2+1)D is $K=2\sigma_x$ Chern-Simons theory that describes $\Z_2$ topological order. Its anyonic symmetry group is given by $\{\mathbb{I},\sigma_x\}$. The nontrivial group element $\sigma_x$ interchanges the anyon $e$, labeled by the quasiparticle vector $(1 \text{ mod }2,0\text{ mod }2)^T$, and the anyon $m$, labeled by the quasiparticle vector $(0\text{ mod }2,1\text{ mod }2)^T$. A defect labeled by this group element can in principle be realized by externally imposing a dislocation in the Wen-plaquette model as mentioned previously. For convenience, the identity quasiparticle (vacuum) $vac$ is labeled by $(0\text{ mod }2,0\text{ mod }2)^T,$ and the fermion quasiparticle $\psi$ is labeled by $(1\text{ mod }2,1\text{ mod }2)^T$. Thus, the only non-trivial, bare defect is given by $\mathcal{D}^0_{\sigma_x}$. Next we need to deduce equivalence classes of defect-anyon composites. Taking into account Eq. (\[equation\_fusion\_2d\]) and $\mathbb{I}-\sigma_x=\left(\begin{smallmatrix}1&-1\\-1&1\end{smallmatrix}\right)$, we have: $$\begin{aligned}
&\mathcal{D}^0_{\sigma_x}\times qp=\mathcal{D}^0_{\sigma_x}\times (qp+\epsilon)\,,
\forall qp\in\{vac,e,m,\epsilon\}.\nonumber
\end{aligned}$$ Here, the symbol “$+$” denotes the usual addition of quasiparticle vectors of $qp$ and $\epsilon$. Therefore, there are two equivalence classes: $\mathcal{D}^0_{\sigma_x}\times e=\mathcal{D}^0_{\sigma_x}\times m$ and $\mathcal{D}^0_{\sigma_x}\times \psi=\mathcal{D}^0_{\sigma_x}$. In other words, there are two topologically distinct defects: one is bare, given by a bare defect $\mathcal{D}^0_{\sigma_x}$; the other one is a defect-anyon composite, denoted by $\mathcal{D}^1_{\sigma_x}=\mathcal{D}^0_{\sigma_x}\times e$. The fusion rules of these two defects are given by: $$\begin{aligned}
\mathcal{D}^0_{\sigma_x}\times\mathcal{D}^0_{\sigma_x}=\mathcal{D}^1_{\sigma_x}\times\mathcal{D}^1_{\sigma_x}=vac+\psi, \mathcal{D}^0_{\sigma_x}\times\mathcal{D}^1_{\sigma_x}=e+m\,,\nonumber\end{aligned}$$ where “$+$” here denotes the collection of different fusion channels into quasiparticles of simple-type.
Now that we have reviewed the lower dimensional case, let us move on to 3D. Simply from a dimensionality point of view, there are two types of extrinsic defects in 3D: line defects and point defects. The latter also appear in 2D and serve as end points on which 1D branch cuts (i.e., the dashed line in Fig. \[figure\_defect\](b)) terminate. The former are really loop-like. In Fig. \[figure\_defect\](a), the line defect is drawn as a finite line that ends at the top and bottom boundaries where a periodic boundary condition is implicitly imposed. A 2D branch “brane” (i.e., the shaded plane in Fig. \[figure\_defect\](a)) is attached to each line defect.
From Fig. \[figure\_defect\], we see that line defects can perform generic $\mathsf{Charles}$ operations where both point particles and loops are transformed. In contrast, point defects can only perform $\mathsf{Charles}$ operations on loops, meaning that $\mathscr{G}=W\oplus \Omega=\mathbb{I}\oplus \Omega$ for point defects. However, due to Eq. (\[eqn:auto\]), the only candidate for $\Omega$ is $\mathbb{I}$. This means that point defects can only behave like the identity element $\mathscr{G}_{\mathbb{I}}=\mathbb{I}\oplus\mathbb{I}$ of $\mathsf{Charles}$. Therefore, in 3D, we only consider line defects since point defects *cannot* perform nontrivial $\mathsf{Charles}$ operations.
In a manner similar to 2D, a *defect-charge-loop composite* is allowed, where the term “defect” corresponds to a line defect, “charge” corresponds to a point-like excitation, and “loop” corresponds to a loop excitation. Since loops are always transformed to loops by $\Omega,$ and particles are always transformed to particles by $W$, we may study defect-charge-composites and defect-loop-composites separately. In order to determine defect species for a given $\mathsf{Charles}$ group element $\mathscr{G}$, we need to study the equivalence classes of the above two kinds of defect composites. For defect-charge-composites, the following equation determines the equivalence classes: $$\begin{aligned}
\mathcal{D}^0_{\mathscr{G}}\times qp=\mathcal{D}^0_{\mathscr{G}}\times (qp+(\mathbb{I}-W)qp_1)\,,\label{equation_fusion_3d_charge}\end{aligned}$$ which is diagrammatically shown in Fig. \[figure\_defect\_composite\](b). $qp$ denotes point-like particle excitations. $\mathcal{D}^0_{\mathscr{G}}$ denotes the bare line defect that is labeled by a $\mathsf{Charles}$ group element (or conjugacy class for a non-Abelian group) $\mathscr{G}=W\oplus \Omega$. $ qp_1$ is any particle excitation provided by the parent 3D Abelian topological phase. Eq. (\[equation\_fusion\_3d\_charge\]) means that the defect-charge composite $\mathcal{D}^0_{\mathscr{G}}\times qp$ is topologically equivalent to the defect-charge composite that is formed by the fusion between the same bare line defect and a different particle excitation given by $(qp+(\mathbb{I}-W)qp_1)$. Likewise, we have a similar equation for defect-loop composites: $$\begin{aligned}
\mathcal{D}^0_{\mathscr{G}}\times loop=\mathcal{D}^0_{\mathscr{G}}\times (loop+(\mathbb{I}-\Omega)loop_1)\,,\label{equation_fusion_3d_loop}\end{aligned}$$ which is diagrammatically shown in Fig. \[figure\_defect\_composite\](c). One can also unify Eqs. (\[equation\_fusion\_3d\_charge\],\[equation\_fusion\_3d\_loop\]) by considering charge-loop composites. We will show this in the following example.
Let us take $K=3$ in Table \[table:charlos\] as an example. There is only one nontrivial group element given by $\mathscr{G}=-\mathbb{I}\oplus-\mathbb{I}$. For convenience, we label the three topologically distinct particle excitations as $t_0,t_1,t_2$ and the three distinct loop excitations as $l_0,l_1,l_2$. Using numerical labels, we have: $t_0=0\text{ mod }3, t_1=1\text{ mod }3, t_2=2\text{ mod }3$, and $l_0=0\text{ mod }3, l_1=1\text{ mod }3, l_2=2\text{ mod }3$. We can consider the set of 2D vectors $\mathbf{V}_{ij}=(t_i,l_j)^T$ where $i,j=0,1,2$ and hence, there are $3^2=9$ vectors that label the 9 topologically distinct charge-loop composites: $$\begin{aligned}
\!\!\!\!\{\mathbf{V}_{ij}\}\!=\!\!\bpm0 \\ 0\epm\!\!, \bpm1 \\ 0\epm\!\!, \bpm2 \\ 0\epm\!\!, \bpm0 \\1\epm\!\!, \bpm1\\1 \epm\!\!, \bpm2 \\ 1\epm\!\!, \bpm0 \\ 2\epm\!\!,\bpm1 \\ 2\epm\!\!, \bpm2\\ 2\epm.\nonumber\end{aligned}$$ As a result, Eqs. (\[equation\_fusion\_3d\_charge\],\[equation\_fusion\_3d\_loop\]) can be unified as: $$\begin{aligned}
\mathcal{D}^0_{\mathscr{G}}\times \mathbf{V}=\mathcal{D}^0_{\mathscr{G}}\times (\mathbf{V}+(\mathscr{G}_{\mathbb{I}}-\mathscr{G})\mathbf{V}')\,,\label{equation_fusion_3d_unified}\end{aligned}$$ where $\mathbf{V},\mathbf{V}'\in\{\mathbf{V}_{ij}\}$. By noting that $\mathscr{G}_{\mathbb{I}}-\mathscr{G}=\left(\begin{smallmatrix}1&0\\ 0&1\end{smallmatrix}\right)-\left(\begin{smallmatrix}-1&0\\ 0&-1\end{smallmatrix}\right)=\left(\begin{smallmatrix}2&0\\ 0&2\end{smallmatrix}\right)$, the above relation reduces to: $$\begin{aligned}
\mathcal{D}^0_{\mathscr{G}}\times \mathbf{V}=\mathcal{D}^0_{\mathscr{G}}\times (\mathbf{V}+2\mathbf{V}').\end{aligned}$$ As a result, all defect-charge-loop composites are topologically equivalent to the bare defect: $ \mathcal{D}^0_{\mathscr{G}}=\mathcal{D}^0_{\mathscr{G}}\times \mathbf{V}_{ij}\,,$ where $\{\mathbf{V}_{ij}\}$ denotes the 9 vectors ($i,j=0,1,2$). The resulting fusion rules are given by: $$\begin{aligned}
\mathbf{V}_{ij}\times \mathbf{V}_{i'j'}&=\mathbf{V}_{(i+i')\text{mod}3,(j+j')\text{mod}3}\,,\\
\mathcal{D}^0_{\mathscr{G}}\times \mathcal{D}^0_{\mathscr{G}}&=\sum_{ij} \mathbf{V}_{ij}\,,
\end{aligned}$$ from which we see that there are multiple fusion channels when two defects are fused together. It indicates that the externally imposed line defect $\mathcal{D}^0_{\mathscr{G}}$ is of non-Abelian nature.
Conclusions {#sec:conclusion_direction}
===========
In this work, a composite particle theory for 3D fermionic gapped phases was formulated based on a specific parton construction of electrons. Composite particles are bound states of partons and magnetic monopoles for a set of internal gauge fields and the external electromagnetic field $A_\mu$. The resulting fully-gapped phases were constructed by condensing two composite particles. All excitations including point-like and string-like excitations as a whole form a charge-loop-lattice. Each site of the charge-loop-lattice corresponds to a deconfined excitation of the condensed phase. A general mechanism for charge fractionalization in 3D was studied in detail. Based on the general framework of composite particle theory, we further explored two important properties of 3D Abelian topological phases. First, we studied phases with non-vanishing axion $\Theta$ angle which is characteristic of the tilted charge lattice. It was found that time-reversal invariant fractional topological insulators with $\Theta\neq \pi$ can be constructed from composite particle theory. Second, we generalized the notion of anyonic symmetry of 2D Abelian topological phases to a charge-loop excitation permutation symmetry ($\mathsf{Charles}$) group in 3D Abelian topological phases. We also investigated the relation between $\mathsf{Charles}$ group elements and line twist defects in (3+1)D Abelian topological phases.
There are several interesting directions for future studies. *First*, it is interesting to propose a systematic theory of the symmetric surface states of fractional topological insulators based on the composite particle theory. The 2D surface may exhibit quantum phenomena that are even more exotic than the surface topological order recently found on the surface of interacting topological insulators and interacting bosonic topological insulators [@bti1; @bti2; @bti3; @bti4; @bti5; @bti7; @bti8; @sto1; @sto2; @sto3; @sto4; @sto5]. For the [FTI]{} bulk lattice model construction and the phase diagram of confinement-deconfinement, the idea in Ref. may be helpful. *Second*, one may consider the composite particle theory by assuming that partons form topological superconductor ansätze, which may lead to interacting topological superconductors with fractional gravito-electromagnetism and a fractional version of the gravitational Witten effect [@grav1; @grav2]. *Third*, as discussed in Sec. \[sec:general\_charlos\], the tensor-network-type graphs may be helpful for understanding 3D analogs of the twist liquid, i.e., the topological phases obtained by gauging $\mathsf{Charles}$. *Fourth*, it is interesting to think if there are simple 3D lattice models that can demonstrate the physics of extrinsic defects and $\mathsf{Charles}$, in analogy to the 2D case where there are lattice models like the Wen-plaquette model. In addition, some group elements of $\mathsf{Charles}$ may break $U(1)$ charge symmetry. A line defect associated with such a group element might be realized in a $U(1)$-symmetric 3D lattice model as an extrinsic defect coated with a superconducting region. *Fifth*, in analogy to 2D anyonic symmetry where $G$-crossed tensor category theory [@Teo2015; @Barkeshli2014] was proposed, a generic mathematical framework is also needed for 3D extrinsic defects. *Sixth*, it would be useful to have a microscopic theory of 3D line twist defects in terms of a cutting and gluing procedure where the twist defects are formed by tuning/twisting allowed tunneling terms between the two sides of a gapless cut [@genon_1; @genon_2; @genon_3; @dhl_cut_glue].
Acknowledgement {#acknowledgement .unnumbered}
===============
We would like to thank K. Shiozaki, Y. Qi, and S. Ryu for helpful discussions. P.Y. would like to thank Z.-Y. Weng and X.-G. Wen for beneficial collaborations and insightful discussions on parton constructions, and also acknowledges S.-T. Yau’s hospitality at the Center of Mathematical Sciences and Applications at Harvard University where the work was done in part. This work was supported in part by the NSF through grant DMR 1408713 at the University of Illinois.(P.Y. & E.F.) T.L.H. is supported by the US National Science Foundation under grant DMR 1351895-CAR. J.M. was supported by NSERC grant \#RGPIN-2014-4608, the CRC Program, CIFAR, and the University of Alberta.
Summary of notations, abbreviations, and definitions {#appendix:notation}
====================================================
In this Appendix, several notations, abbreviations, definitions, and criteria are collected for the reader’s convenience.
*1. Mathematical notations:*
$u,v,u',v',q,q'$: a set of parameters that label the two condensed composites as shown in Table. \[table:em\].
$Q$: the net EM electric charge carried by a composite.
$Q_{\rm Debye}$: the screening charge cloud around a composite. It is induced by the two composite condensates $\varphi_1$ and $\varphi_2$.
$N_A$: the bare EM electric charge carried by a composite. It is related to $Q$ via Eqs. (\[NE1\],\[NE\_screen\]).
$M$: the EM magnetic charge carried by a composite.
$M_\mu$: the 4-current of EM magnetic monopoles, introduced in Sec. \[sec\_FTI\_Coulomb\].
$N_{a,b}$: gauge charges in $U(1)_a$ and $U(1)_b$ gauge groups. An integer vector $\mathbf{N}_e$ is formed via Eq. (\[define\_KNq\]).
$N_m^{a,b}$: magnetic charges in $U(1)_a$ and $U(1)_b$ gauge groups. An integer vector $\mathbf{N}_m$ is formed via Eq. (\[define\_KNq1\]).
$\Gamma$: self-statistics of a composite. $\Gamma$ is even (odd) if the composite is bosonic (fermionic), see Eq. (\[eq:stat123\]).
$\theta$: $\theta=0$ if all partons ($f^1,f^2,f^3$) form trivial band insulators. $\theta=\pi$ if all partons form topological insulators.
$\vartheta^{cl}$: the mutual statistics between a point-like particle excitation and a loop excitation, see Eq. (\[eq:cltheta\]).
$\Theta$: the axion angle of the electron states (i.e., the resulting fermionic gapped phase constructed via the composite particle theory).
$g_{a,b}$ dimensionless gauge coupling constants of $U(1)_{a,b}$ gauge groups.
$\mathcal{D}^0_\mathscr{G}$: a bare line defect associated with $\mathsf{Charles}$ group element $\mathscr{G}$.
*2. Abbreviation:*
$\mathsf{Charles}$: **char**ge-**l**oop **e**xcitation **s**ymmetry.
EM: electromagnetic (specific to the usual background electromagnetic field $A_\mu$).
FQH: fractional quantum Hall effect.
[FTI]{}: fractional topological insulator.
$\mathsf{GCD}$: greatest common divisor.
GSD: ground state degeneracy.
IQH: integer quantum Hall effect.
SET: symmetry-enriched topological phase.
SPT: symmetry-protected topological phase.
TI: free-fermion topological insulator.
TO: topological order
TQFT: topological quantum field theory.
*3. Definitions:*
Loop-lattice: Definition \[dfn\_loops\] on Page .
Excitation and charge lattice: Definition \[dfn\_excitation\] on Page .
Intrinsic excitation and intrinsic charge lattice: Definition \[dfn\_intrinsic\_exc\] on Page .
Charge-loop-lattice: Definition \[dfn\_cloop\] on Page .
Charge-loop excitation symmetry: Definition \[dfn\_charlos\] on Page .
*4. Others*
Criterion \[crt\_loop\_exc\] for loop excitations on Page
Criterion \[crt\_charge\_1\] for charge fractionalization on Page
Criterion \[crt\_charge\_2\] for charge fractionalization on Page
Technical details in Sec. \[sec:gauge\_strc\]
=============================================
Details of Eq. (\[constr\_3\]) {#appendix_bosonic}
------------------------------
By inserting the data of $\varphi_1$ and $\varphi_2$ in Table \[table:em\] into Eq. (\[eq:quantum\_stat\]), one may obtain: $$\begin{aligned}
\Gamma(\varphi_1)
= 3q-\frac{\theta}{2\pi}[u(u+1)+v(v+1)+(u+v)(u+v-1)]\,,\end{aligned}$$ and $$\begin{aligned}
\Gamma(\varphi_1)
=3q'-\frac{\theta}{2\pi}[u'(u'+1)+v'(v'+1)+(u'+v')(u'+v'-1)].
\end{aligned}$$
Proof of Theorem \[theorem\_flux\] {#appendix_proof_theorem_flux}
----------------------------------
We present Bézout’s lemma as a preliminary:
Bézout’s lemma [@bezout79]
: Let $a$ and $b$ be nonzero integers and let $d$ be their greatest common divisor ($\mathsf{GCD}$). Then there exist integers $x$ and $y$ such that $ax+by=d$. In addition, $d$ is the smallest positive integer that can be written as $ax + by$; every integer of the form $ax + by$ is a multiple of $d$.
Let us now prove Theorem 1.
[Sufficiency:]{} When $|uv'-u'v|=1$, according to Eq. (\[eq:ineq\]), we straightforwardly obtain $|\mathsf{GCD}(u,u')|=1,|\mathsf{GCD}(v,v')|=1$. Then, the equalities in Eq. (\[eq:ineq\]) hold. Therefore, $(\Phi^a_e)_{\rm min}=2\pi, (\Phi^b_e)_{\rm min}=2\pi$ in Eq. (\[discreteflux1\]).
[Necessity:]{} We start with the equalities in Eq. (\[eq:ineq\]), i.e., $|uv'-u'v|=|\mathsf{GCD}(u,u')|=|\mathsf{GCD}(v,v')|$. If $|uv'-u'v|\neq1$, meaning that $|\mathsf{GCD}(u,u')|=|\mathsf{GCD}(v,v')|\neq 1$. Therefore, $u,u'$ are not co-prime; $v,v'$ are not co-prime. Then, we consider: $$\begin{aligned}
1=\bigg|\frac{u}{\mathsf{GCD}(u,u')}v'-\frac{u'}{\mathsf{GCD}(u,u')}v\bigg|\,,\label{in1}\end{aligned}$$ where $\frac{u}{\mathsf{GCD}(u,u')}, \frac{u'}{\mathsf{GCD}(u,u')}$ are co-prime by definition, i.e., $$\begin{aligned}
\bigg|\mathsf{GCD}(\frac{u}{\mathsf{GCD}(u,u')}, \frac{u'}{\mathsf{GCD}(u,u')})\bigg|=1\,.\nonumber
\end{aligned}$$ Then, according to Eq. (\[eq:ineq\]), we can also have the following inequalities if we just replace $u$ and $u'$ in Eq. (\[eq:ineq\]) by $\frac{u}{\mathsf{GCD}(u,u')}$ and $\frac{u'}{\mathsf{GCD}(u,u')}$, respectively: $$\begin{aligned}
\bigg|\frac{u}{\mathsf{GCD}(u,u')}v'-\frac{u'}{\mathsf{GCD}(u,u')}v\bigg|\geq |\mathsf{GCD}(v,v')|\,.\label{in2}\end{aligned}$$ Due to Eqs. (\[in1\],\[in2\]), we obtain $|\mathsf{GCD}(v,v')|=1$. This is contradictory to our starting point $|\mathsf{GCD}(v,v')|\neq 1$. Therefore, the only possibility is $|uv'-u'v|=|\mathsf{Det}K|=1$.
Equivalence between Criterion \[crt\_charge\_1\] and Criterion \[crt\_charge\_2\] {#appendix_dirac}
---------------------------------------------------------------------------------
Consider two excitations in a $U(1)_{\rm EM}$-symmetric system. Let one carry zero EM magnetic charge $M=0$ and minimal non-vanishing EM electric charge $Q=\frac{1}{w}$ with $w\in\Z$. Let the other excitation carry a minimal nonzero EM magnetic charge $w'$ and an EM electric charge, say, $y$. $y$ can be either integer or non-integer. Due to the Dirac-Zwanziger-Schwinger quantization condition [@dirac1; @dirac2; @dirac2.5; @dirac3; @dirac3.5], the magnetic and electric charges of the above two excitations satisfy: $$\begin{aligned}
\left(\frac{1}{w}\times w'-0\times y\right)=0,\pm1,\pm2,\cdots\,.\end{aligned}$$ Therefore, the minimal choice of $w'$ is $w'=w$, indicating that the change of quantization of the EM magnetic charge $M$ is accompanied with a change of the charge quantization. Once $w>1$, $w'$ is also larger than one. In this sense, the two criteria are equivalent.
Debye-H[ü]{}ckel charge cloud $Q_{\rm Debye}$ is the unique source of charge fractionalization {#appendix_na_integer}
----------------------------------------------------------------------------------------------
In this Appendix, we prove that the Debye-H[ü]{}ckel charge cloud $Q_{\rm Debye}$ defined in Eq. (\[NE\_screen\]) is the *unique* source of charge fractionalization. In other words, $N_A$ is always integer-valued when $M=0$. By definition in Eqs. (\[eqn:transition\_matrix\],\[eqn:transition\_matrix1\],\[Nnf1\]), $N_A$ is given by: $$\begin{aligned}
N_A=&N^{f1}+N^{f2}-N^{f3}\nonumber\\
=&(n^{f1}+n^{f2}-n^{f3})+\frac{3\theta }{2\pi}M-\frac{2\theta}{2\pi}N^b_m\,,\end{aligned}$$ where $n^{fi}$ are integer-valued. $N^b_m$ is integer-valued, and $\theta=0,\pi$. Therefore, $-\frac{2\theta}{2\pi}N^b_m$ is always integer-valued. As a result, $N_A$ is integer-valued when $M=0$.
Technical details in Sec. \[sec:theta\]
=======================================
Partons occupying trivial bands {#appendix_Gamma_trivial_1}
-------------------------------
We assume that all partons $f^{i}$ (pure gauge charge carriers) form three trivial band insulators ($\theta=0$). According to Eqs. (\[constr\_1plus\],\[constr\_1\],\[constr\_3\]), we have: $$\begin{aligned}
u,v,u',v'\in\Z\,;~~q,q'\in\Z_{\rm even} \,.\label{theta_0_q_even}
\end{aligned}$$ Since $\theta=0$, we have: $$\begin{aligned}
N^{fi}=n^{fi}\in\Z
\end{aligned}$$ according to Eq. (\[Nnf1\]). Due to the definitions in Sec. \[sec:gauge\_strc\], we end up with: $$\begin{aligned}
N_a,N_b\in\Z\,.
\end{aligned}$$ Thus, in the mean-field ansätze with $\theta=0$, all magnetic charges and electric charges of composites are integer-valued. However, $Q_{\rm Debye}$ and $Q$ may be fractional, depending on the condensate parameters.
According to Definition \[dfn\_excitation\], excitations are a subset of generic composites and satisfy the two equations in Eq. (\[excitation\_1\]). Thus, only a 4D sublattice embedded in the 6D lattice survives, i.e., the charge lattice in Definition \[dfn\_excitation\]. Since $N_a(=N^{f1}-N^{f2})$ and $N_b(=N^{f3}-N^{f2})$ are fully determined by $M$ via Eq. (\[NabM\]), we may use the labels $(N^{f2},M,N^a_m,N_m^b).$ These four linearly independent integer numbers are “4D coordinates” of the 4D lattice that label excitations. Then, the bare EM electric charge $N_A$ is expressed as: $$\begin{aligned}
N_A=N^{f1}+N^{f2}-N^{f3}=(r-s)M+N^{f2}\,.\end{aligned}$$ The net EM electric charge $Q$ is given by: $$\begin{aligned}
Q=&N_A-Q_{\rm Debye}\nonumber\\
=&N^{f1}+N^{f2}-N^{f3}-Q_{\rm Debye}\nonumber\\
=&(r-s)M + N^{f2}-rN_m^a-sN_m^b\,,\label{q_trivial}\end{aligned}$$ where $N^{f1}-N^{f2}=rM$, $N^{f3}-N^{f2}=sM$ with $r$ and $s$: $$\begin{aligned}
r=\frac{qv'-q'v}{\mathsf{Det}K}\,,~s=\frac{q'u-qu'}{\mathsf{Det}K}\,.
\end{aligned}$$ We note that $r$ and $s$ can be either integer or non-integral rational numbers. However, $rM$ and $sM$ must be integer-valued in order to ensure the $N^{fi}$ are integer-valued. Thus, the quantization of $M$ should be altered properly if $r$ and $s$ are non-integral rational numbers. In summary, we can define the following domains: $$\begin{aligned}
N^{f2}\in\Z,N^a_m\in\Z,N^b_m\in\Z\,, \frac{M}{w}\in\Z\,,\end{aligned}$$ where $w$ is a positive minimal integer such that both $rM\in\Z$ and $sM\in\Z$ are satisfied. In Eq. (\[q\_trivial\]), the $M$-dependent charge $(r-s)M$ is integer-valued: $$\begin{aligned}
(r-s)M\in\Z\,.\label{theta_trivial_response}
\end{aligned}$$ Therefore, the minimal quantized value of $Q$ is sufficiently determined by $rN_m^a$ and $sN^b_m$ by noting that the latter two terms can be potentially fractionalized depending on $r$ and $s$. In the language of the EM response theory, $M$-dependent charge means that the EM magnetic current minimally couples to the EM gauge field $A_\mu$. In other words, the bulk supports an EM response action with $\Theta$ term. If we define $\frac{\Theta}{2\pi}M=(r-s)M$, then $Q=\frac{\Theta}{2\pi}M+N^{f2}-rN_m^a-sN^b_m$ with $\Theta=2\pi(r-s)$. However, due to Eq. (\[theta\_trivial\_response\]), this nonzero $\Theta$ gives rise to an integer charge cloud surrounding EM magnetic monopoles. This additional charge cloud does not render a new quantization of $Q$ different from the quantization when $M=0$. In other words, the EM charge lattice ($M-Q$ plane) is just a square lattice that is not tilted. The allowed values of $Q$ when $M=0$ are completely the same as when $M\neq 0$. In this sense, the resulting state with $\Theta=2\pi(r-s)$ is equivalent to a trivial state with $\Theta=0$. By comparison, a typical example with nontrivial $\Theta$ angle has a $Q$ quantization shown in Eqs. (\[Q\_fractional\_M0\],\[Q\_fractional\_M1\]) of Sec. \[sec\_FTI\] where the quantization of $Q$ manifestly depends on $M$.
Derivation of Eqs. (\[Q\_fractional\_M0\],\[Q\_fractional\_M1\]), Eq. (\[eq:stat4321\_new\]), and the site distribution in Fig. \[figure\_theta\](c) {#appendix_gamma_stat4321_new}
----------------------------------------------------------------------------------------------------------------------------------------------------
Since $M$ is quantized in multiples of $3$ as indicated in Eq. (\[double\_m\_charge\]), one may introduce an integer $k$ such that $M=3k$. Meanwhile, Eq. (\[Q\_expression\]) indicates that $Q$ is generically quantized in multiples of $1/6$. Thus, we can introduce an integer $k_0$ such that $Q=\frac{k_0}{6}$. Then, Eq. (\[Q\_expression\]) is formulated as: $$\begin{aligned}
k_0+5k=2(7n^{f1}-n^{f2}-3n^{f3})\,,\end{aligned}$$ where the r.h.s. is always even. $5k$ has the same even-odd property as $k$. As a result, $k_0$ and $k$ must be simultaneously either odd or even, which leads to Eqs. (\[Q\_fractional\_M0\],\[Q\_fractional\_M1\]).
Then, we start with $\Gamma$ in Eq. (\[eq:stat4321\]) and derive its equivalent expression (\[eq:stat4321\_new\]). Due to Eqs. (\[double\_m\_charge\],\[Q\_fractional\_M1\]), we introduce four integer numbers $k_0,k_1,k_2,k_3$ via $$\begin{aligned}
M=3k\,, \,N_m^a=2k_1\,, \,N_m^b=2k_2\,, \,Q=\frac{k_0}{6}
\end{aligned}$$ so as to simplify the analysis below. Then, solving Eqs. (\[equation\_NMA\],\[equation\_NMB\],\[Q\_expression\]) leads to: $$\begin{aligned}
&n^{f1}=-\frac{5}{2}k+\frac{k_0}{6}+\frac{5}{3}k_1-\frac{2}{3}k_2\,,\\
&n^{f2}=-\frac{13}{2}k+\frac{k_0}{6}+\frac{11}{3}k_1+\frac{1}{3}k_2\,,\\
&n^{f3}=-\frac{9}{2}k+\frac{k_0}{6}+\frac{8}{3}k_1-\frac{5}{3}k_2\,.
\end{aligned}$$ Therefore, $\Gamma$ in Eq. (\[eq:stat4321\_new\]) can be reformulated as: $$\begin{aligned}
\Gamma=(M+1)(-14k+\frac{1}{2}k+\frac{k_0}{2}+8k_1-2k_2)\,.\end{aligned}$$ Since $(M+1)(-14k+8k_1-2k_2)$ is always an even integer, we may remove it and end up with: $$\begin{aligned}
\Gamma=&(M+1)(\frac{1}{2}k+\frac{k_0}{2})\,
=3(M+1)(Q+\frac{1}{18}M)\end{aligned}$$ which can be rewritten as: $$\begin{aligned}
\Gamma=&3(M+1)(Q-\frac{\Theta}{2\pi}M)\end{aligned}$$ with $\Theta=-\frac{1}{9}\pi$. One can check that $\Gamma$ is invariant under the shift $\Theta\rightarrow \Theta+\frac{2}{9}\pi$ since the additional term $-3(M+1)\frac{1}{9}M=-k(3k+1)$ is always an even integer which leaves the even / odd property of $\Gamma$ unaltered. From this point of view, we say that two $\Theta$’s are topologically equivalent if their difference is given by multiples of $\frac{2}{9}\pi$. In conclusion, $$\begin{aligned}
\Theta=\frac{1}{9}\pi \text{ mod }\frac{2}{9}\pi\,.\end{aligned}$$ As a result, $-\frac{1}{9}\pi$, $\frac{1}{9}\pi$, $-\frac{5}{9}\pi$, etc. describe the same [FTI]{} states. The periodicity $\frac{2}{9}\pi$ is the minimal one in the sense that any shift smaller than $\frac{2}{9}\pi$ does not keep the even-odd property of $\Gamma$ invariant. In other words, the charge lattice with “tilt angle” $\Theta=\frac{1}{9}\pi$ is always different from a lattice with $\Theta=0$. This periodicity check is very important since it is possible that a nonzero $\Theta$ might be entirely removed by a periodic shift. If this happens, the resulting bulk state is actually a trivial state. Next, we calculate the lattice sites in Fig. \[figure\_theta\](b). Since $Q=\frac{1}{6}$ and $M=3$, we have: $k_0=1,k=1$: $$\begin{aligned}
&n^{f1}=(-2+2k_1-k_2)+\frac{-1-k_1+k_2}{3}\,,\\
&n^{f2}=(-6+4k_1)+\frac{-1-k_1+k_2}{3}\,,\\
&n^{f3}=(-4+3k_1-2k_2)+\frac{-1-k_1+k_2}{3}\,.
\end{aligned}$$ Therefore, $-1-k_1+k_2$ should be quantized in multiples of 3 such that the $n^{fi}$’s are integer-valued. By noting that $N^a_m=2k_1,N^b_m=2k_2$, we end up with Fig. \[figure\_theta\](c) where $k_0=1,k=0$ are assumed.
Technical details in Sec. \[sec:charlos\_defect\_symmetry\] {#appendix_proof_group}
===========================================================
$\mathsf{Charles}$ is a group
-----------------------------
is to prove that $\mathsf{Auto}(K)$ is a group. In other words, the elements satisfy the four group axioms. (identity element, inverse element, closure, associativity).
Identity element.— The identity element is $\mathscr{G}_{\mathbb{I}}=W\oplus \Omega=\mathbb{I}\oplus\mathbb{I}$ where $\mathbb{I}$ is a rank-2 identity matrix. For every element $\mathscr{G}$ in $\mathsf{Auto}(K)$, the equation $\mathscr{G}\cdot \mathscr{G}_{\mathbb{I}}=\mathscr{G}_{\mathbb{I}} \cdot\mathscr{G}=\mathscr{G} $ holds. Here the symbol $\cdot$ denotes matrix multiplication. We will also omit it unless otherwise specified.
Associativity.— Associativity is guaranteed by matrix multiplication rules.
Inverse element.— The inverse element of $\mathscr{G}$ is given by: $\mathscr{G}^{-1}=W^{-1}\oplus \Omega^{-1}$. One may check that $(W^{-1}\oplus \Omega^{-1})\cdot (W\oplus \Omega)=\mathbb{I}\oplus\mathbb{I}=\mathscr{G}_{\mathbb{I}}$ and $ (W\oplus \Omega)\cdot (W^{-1}\oplus \Omega^{-1})=\mathscr{G}_{\mathbb{I}}$, which means that: $\mathscr{G}^{-1}\cdot\mathscr{G}=\mathscr{G}\cdot\mathscr{G}^{-1}=\mathscr{G}_{\mathbb{I}}$.
Closure.— Suppose $\mathscr{G}'=W'\oplus \Omega'\in \mathsf{Auto}(K)$. Thus, $W,\Omega, W',\Omega'$ matrices satisfy conditions (i) and (ii) in Definition \[dfn\_charlos\]. Then, by definition, $\mathscr{G}'\cdot \mathscr{G}=(W' W)\oplus (\Omega' \Omega)$. Both $W' W$ and $\Omega' \Omega$ are still rank-2 unimodular matrices. Furthermore, $$\begin{aligned}
(\Omega'\Omega)K (W' W)^T= \Omega'(\Omega K W^T) {W'}^T= \Omega'K{W'}^T=K\,.\nonumber\end{aligned}$$ $\therefore$ condition (i) is satisfied. And, $$\begin{aligned}
&\Gamma (\cdots,(W'W)^{-1}\mathbf{N}_m,\cdots)\nonumber\\
=&\Gamma (\cdots,W^{-1}{W'}^{-1}\mathbf{N}_m,\cdots)\nonumber\\
=&\Gamma (\cdots,{W'}^{-1}\mathbf{N}_m,\cdots)\nonumber\\
=&\Gamma (\cdots,\mathbf{N}_m,\cdots)\,.\nonumber
\end{aligned}$$ $\therefore$ condition (ii) is also satisfied.
is to verify that $\mathsf{Inner}(K)$ is a subgroup of $\mathsf{Auto}(K)$. First, it is a subset of $\mathsf{Auto}(K)$, i.e., $\mathsf{Inner}(K)\subset\mathsf{Auto}(K)$ since not only do $W$ and $\Omega$ satisfy condition (i) and (ii), but also satisfy $$\begin{aligned}
W^{-1}\mathbf{N}_m-\mathbf{N}_m=K^T\,(n_1,n_2)^T\end{aligned}$$ and $$\begin{aligned}
\Omega^{-1}\mathbf{L}-\mathbf{L}=K(n_3,n_4)^T\,.\end{aligned}$$ Here $n_1,\cdots,n_4$ are integers. $n_1,n_2$ depend on $\mathbf{N}_m,W$; $n_3,n_4$ depend on $\mathbf{L},\Omega$.
$\mathscr{G}_{\mathbb{I}}\in \mathsf{Inner}(K)$ since one can obtain $\mathscr{G}_{\mathbb{I}}\mathbf{L}-\mathbf{L}=0$ and $\mathscr{G}_{\mathbb{I}}\mathbf{N}_m-\mathbf{N}_m=0$ by choosing $n_1=n_2=n_3=n_4=0$ for all $\mathbf{L}$’s and $\mathbf{N}_m$’s. Then, elements of $\mathsf{Inner}(K)$ have associativity arising from the standard matrix multiplication.
For the existence of the inverse element, we need to prove that $W^{-1}\oplus\Omega^{-1}\in\mathsf{Inner}(K)$. By definition, for $W\oplus \Omega\in\mathsf{Inner}(K)$, the operation of $W$ is $$\begin{aligned}
W^{-1}\mathbf{N}_m-\mathbf{N}_m=K^T \mathbf{J}_{W,\mathbf{N}_m}\,,\nonumber
\end{aligned}$$ where the notation $\mathbf{J}_{W,\mathbf{N}_m}$ denotes the integer vector $(n_1,n_2)^T$ and the subscript $W,\mathbf{N}_m$ means that the integer vector is a function of $W$ and $\mathbf{N}_m$. Likewise, we also have: $$\begin{aligned}
W^{-1} (W\mathbf{N}_m)-W\mathbf{N}_m=K^T \mathbf{J}_{W,W\mathbf{N}_m}.\nonumber
\end{aligned}$$ As a result, $W\mathbf{N}_m-\mathbf{N}_m=-K^T \mathbf{J}_{W,W\mathbf{N}_m}$. Since $-\mathbf{J}_{W,W\mathbf{N}_m}$ is an integer vector, we obtain the operation of $W^{-1}$: $$\begin{aligned}
(W^{-1})^{-1}\mathbf{N}_m-\mathbf{N}_m=K^T (-\mathbf{J}_{W,W\mathbf{N}_m})\nonumber
\end{aligned}$$ which exactly satisfies the defining property of $\mathsf{Inner}(K)$. Likewise, we can also prove that $\Omega^{-1}$ is an operation in $\mathsf{Inner}(K)$. Therefore, the inverse element $\mathscr{G}^{-1}=W^{-1}\oplus \Omega^{-1}\in\mathsf{Inner}(K)$.
For the closure property, we need to prove that $\mathscr{G}'\cdot\mathscr{G}\in \mathsf{Inner}(K)$ if $\mathscr{G}'\in \mathsf{Inner}(K)$ and $\mathscr{G}\in \mathsf{Inner}(K)$. For this purpose, let us calculate: $$\begin{aligned}
&(W'W)^{-1}\mathbf{N}_m-\mathbf{N}_m \nonumber\\
=&W^{-1}{W'}^{-1}\mathbf{N}_m-\mathbf{N}_m=W^{-1}\left( \mathbf{N}_m+K^T\mathbf{J}_{W',\mathbf{N}_m}\right)\!-\!\mathbf{N}_m\nonumber\\
=&\!\left(W^{-1} \mathbf{N}_m\!-\!\mathbf{N}_m\right)\!+W^{-1}K^T\mathbf{J}_{W',\mathbf{N}_m}\nonumber\\
=&K^T\! \mathbf{J}_{W,\mathbf{N}_m}\! +\!K^T\Omega^T\mathbf{J}_{W',\mathbf{N}_m}=K^T\!\left(\mathbf{J}_{W,\mathbf{N}_m}\!+\Omega^T\mathbf{J}_{W',\mathbf{N}_m}\!\right)\,,\nonumber\end{aligned}$$ where we have used $W^{-1}K^T=K^T\Omega^T$ that is an equivalent expression of condition (i). Since $\mathbf{J}_{W,\mathbf{N}_m}+\Omega^T\mathbf{J}_{W',\mathbf{N}_m}$ is an integer vector (by noting that $\Omega$ is unimodular), we conclude that $W'W$ satisfies the defining property of $\mathsf{Inner}(K)$. So does $\Omega\Omega'$. Therefore, $\mathscr{G}'\cdot\mathscr{G}\in \mathsf{Inner}(K)$.
is to prove that $\mathsf{Inner}(K)$ is normal. In other words, we need to verify that $W'W{W'}^{-1}\in\mathsf{Inner}(K)$ and $\Omega'\Omega{\Omega'}^{-1}\in\mathsf{Inner}(K)$ for $\forall W,\Omega \in\mathsf{Inner}(K)$ and $\forall W',\Omega'\in\mathsf{Auto}(K)$. For this purpose, let us calculate: $$\begin{aligned}
&(W'W{W'}^{-1})^{-1}\mathbf{N}_m-\mathbf{N}_m=W'(W^{-1}{W'}^{-1}\mathbf{N}_m)-\mathbf{N}_m\nonumber\\
=&W'( {W'}^{-1}\mathbf{N}_m+K^T\mathbf{J}_{W,{W'}^{-1}\mathbf{N}_m})-\mathbf{N}_m\nonumber\\
=&W'K^T\mathbf{J}_{W,{W'}^{-1}\mathbf{N}_m}.\nonumber\end{aligned}$$ Since $W'\in\mathsf{Auto}(K)$, condition (i) in Eq. (\[eqn:auto\]) leads to: $\Omega' K{W'}^{T}=K,$ and thereby ${W'}K^T{\Omega'}^{T}=K^T$. Therefore, ${W'}K^T=K^T({\Omega'}^{T})^{-1}$. $\therefore$ $$\begin{aligned}
\!\!\!\!(W'W{W'}^{-1})^{-1}\mathbf{N}_m-\mathbf{N}_m\!=\!K({\Omega'}^{T})^{-1} \mathbf{J}_{W,{W'}^{-1}\mathbf{N}_m}.\nonumber\end{aligned}$$
Since $({\Omega'}^{T})^{-1}$ is obviously a unimodular matrix, it implies that $({\Omega'}^{T})^{-1} \mathbf{J}_{W,{W'}^{-1}\mathbf{N}_m}$ is an integer vector. Thus, by the definition of $\mathsf{Inner}(K)$, $W'W{W'}^{-1}\in\mathsf{Inner}(K)$. Likewise, we also have $\Omega'\Omega{\Omega'}^{-1}\in\mathsf{Inner}(K)$. $\therefore$ we conclude that $\mathsf{Inner}(K)$ is a normal subgroup of $\mathsf{Auto}(K)$.
Then, according to the definition of $\mathsf{Charles}$, the elements of $\mathsf{Charles}$ form a quotient group of $\mathsf{Auto}(K)$ by $\mathsf{Inner}(K)$. It can be non-Abelian since $\mathscr{G}\cdot \mathscr{G}'\neq \mathscr{G}'\cdot \mathscr{G}$ may hold for some elements.
Graphical representations of $\mathsf{Charles}$ transformations
---------------------------------------------------------------
Eqs. (\[eqn:auto\],\[eqn:auto33\],\[eqn:auto44\]) are graphically represented in Fig. \[figure\_tensor\] where a tensor-network-type graph is introduced.
![(Color online) Tensor-network-type graphical representations of $\mathsf{Charles}$ transformations. (a) represents Eq. (\[eqn:auto\]) where $K$ is a fixed-point matrix. (b) represents Eq. (\[eqn:auto33\]) where $\Lambda$ is a fixed-point tensor with a bond dimension no less than two. (c) represents Eq. (\[eqn:auto44\]) where $\Xi$ is a fixed-point tensor with a bond dimension no less than four.[]{data-label="figure_tensor"}](figure_tensor.pdf){width="8.5cm"}
[139]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, ()](\doibase 10.1103/PhysRevLett.45.494) [****, ()](\doibase 10.1103/PhysRevLett.61.2015) [****, ()](\doibase 10.1103/PhysRevB.75.121306) [****, ()](\doibase 10.1103/PhysRevLett.98.106803) [****, ()](\doibase 10.1103/PhysRevB.79.195322) [****, ()](\doibase 10.1103/RevModPhys.82.3045) [****, ()](http://dx.doi.org/10.1038/nature08916) [****, ()](\doibase 10.1103/RevModPhys.83.1057) [****, ()](\doibase
10.1038/nphys1274) [****, ()](\doibase 10.1016/0038-1098(87)90642-9) [****, ()](\doibase 10.1103/PhysRevB.37.580) [****, ()](\doibase 10.1103/PhysRevB.37.3774) [****, ()](\doibase 10.1103/PhysRevB.38.5142) [****, ()](\doibase 10.1143/JPSJ.57.2768) [****, ()](\doibase
10.1103/PhysRevB.38.745) [****, ()](\doibase 10.1103/PhysRevB.38.2926) [****, ()](\doibase
10.1103/PhysRevB.39.11413) [****, ()](\doibase 10.1103/PhysRevB.44.2664) [****, ()](\doibase 10.1103/PhysRevB.46.5621) [****, ()](\doibase 10.1103/PhysRevB.49.5200) [****, ()](\doibase 10.1103/PhysRevLett.76.503) [****, ()](\doibase 10.1103/RevModPhys.78.17) [****, ()](\doibase 10.1103/PhysRevB.55.3894) [****, ()](\doibase
10.1103/PhysRevLett.106.147002) [****, ()](\doibase
10.1016/j.nuclphysb.2011.09.019) [****, ()](\doibase 10.1103/PhysRevB.85.205142) [****, ()](\doibase 10.1088/1367-2630/16/8/083039) [****, ()](\doibase 10.1016/j.nuclphysb.2013.05.023) [****, ()](\doibase 10.1103/PhysRevB.40.8079) [****, ()](\doibase 10.1142/S0217984991000058) [****, ()](\doibase 10.1103/PhysRevB.60.8827) [****, ()](\doibase 10.1103/PhysRevB.81.155302) [****, ()](\doibase 10.1103/PhysRevB.89.184417) [****, ()](\doibase 10.1103/PhysRevB.87.195128) [****, ()](\doibase 10.1103/PhysRevB.89.045127) [****, ()](\doibase 10.1103/PhysRevB.91.195131) [****, ()](\doibase
10.1103/PhysRevB.90.235146) [****, ()](\doibase 10.1103/PhysRevB.80.155131) [****, ()](\doibase 10.1103/PhysRevB.87.155114) [****, ()](\doibase 10.1126/science.1227224) [****, ()](\doibase 10.1103/PhysRevB.82.155138) [****, ()](\doibase 10.1016/0550-3213(81)90442-9) [****, ()](\doibase 10.1016/0003-4916(82)90164-6) [****, ()](\doibase 10.1016/0370-1573(80)90083-6) [****, ()](\doibase 10.1016/0003-4916(78)90252-X) [****, ()](\doibase 10.1016/0370-2693(75)90162-8) [****, ()](\doibase 10.1016/0550-3213(77)90086-4) [****, ()](\doibase 10.1016/0370-2693(78)90737-2) [****, ()](\doibase 10.1016/0550-3213(78)90153-0) [****, ()](\doibase 10.1016/0550-3213(79)90595-9) [****, ()](\doibase 10.1103/PhysRevD.20.2610) [****, ()](\doibase 10.1103/PhysRevD.17.2637) @noop [**]{} (, ) [****, ()](\doibase 10.1016/0550-3213(94)90214-3) [****, ()](\doibase 10.1007/JHEP08(2013)115) [****, ()](\doibase 10.1103/PhysRevB.79.045316) [****, ()](\doibase 10.1103/PhysRevB.91.045134) [****, ()](\doibase 10.1016/0370-2693(79)90838-4) [****, ()](\doibase 10.1103/PhysRevB.78.195424) [****, ()](\doibase 10.1103/PhysRevLett.63.199) @noop [**]{} (, ) [****, ()](\doibase 10.1103/PhysRevB.44.5246) [****, ()](\doibase 10.1103/PhysRevB.47.7312) [****, ()](\doibase 10.1103/PhysRevLett.79.4437) [****, ()](\doibase 10.1016/j.physe.2003.09.022), [****, ()](\doibase 10.1103/RevModPhys.75.1101) [****, ()](\doibase 10.1103/PhysRevB.82.035105) [****, ()](\doibase
10.1103/PhysRevLett.105.246809) [****, ()](\doibase
10.1103/PhysRevB.86.235128) [****, ()](http://dx.doi.org/10.1038/nphys3311) [****, ()](\doibase 10.1103/PhysRevB.83.195139) [****, ()](\doibase 10.1103/PhysRevB.86.245111) [****, ()](\doibase 10.1016/j.aop.2015.05.012) [****, ()](\doibase 10.1103/PhysRevB.90.235149) [****, ()](\doibase 10.1103/PhysRevB.88.035141) [****, ()](\doibase 10.1103/PhysRevB.81.045323) [****, ()](\doibase 10.1103/PhysRevB.90.155111) [“,” ](http://arxiv.org/abs/1410.4540) (), , [****, ()](\doibase 10.1103/PhysRevB.86.161107) [****, ()](\doibase 10.1103/PhysRevX.2.031013) [****, ()](\doibase 10.1103/PhysRevB.87.045130) [****, ()](\doibase 10.1103/PhysRevX.4.041035) [****, ()](\doibase 10.1103/PhysRevLett.114.026401) [****, ()](\doibase 10.1103/PhysRevLett.105.030403) [****, ()](\doibase 10.1016/0550-3213(82)90463-1) [****, ()](\doibase 10.1016/0550-3213(82)90464-3) [****, ()](\doibase 10.1142/S0217732389000046) [****, ()](\doibase 10.1007/BF01218410) [****, ()](\doibase 10.1103/RevModPhys.52.453) [****, ()](\doibase 10.1103/PhysRevB.88.235109) [****, ()](\doibase 10.1103/PhysRevB.88.035131) [“,” ](http://arxiv.org/abs/1405.5858) (), , [ (), 10.1093/nsr/nwv077](\doibase 10.1093/nsr/nwv077), [****, ()](\doibase 10.1103/PhysRevLett.90.016803) [****, ()](\doibase 10.1016/j.aop.2005.10.005) [****, ()](\doibase
10.1103/PhysRevB.89.115116) [****, ()](\doibase 10.1016/0003-4916(91)90240-9) [****, ()](\doibase 10.1016/0550-3213(94)00503-7) [****, ()](\doibase 10.1016/S0550-3213(98)00586-0) [****, ()](\doibase 10.1103/PhysRevLett.92.030406) [****, ()](\doibase 10.1007/BF02096988) [“,” ](http://arxiv.org/abs/1404.3230) (), , [****, ()](\doibase 10.1103/PhysRevB.93.205157) [****, ()](\doibase 10.1103/PhysRevLett.113.080403) [“,” ](http://pcts.princeton.edu/pcts/Symmetry2014Program-Web.pdf) (), [****, ()](\doibase 10.1103/PhysRevB.91.165119) [****, ()](\doibase 10.1103/PhysRevB.92.035115) [“,” ](http://arxiv.org/abs/1509.04266) (), , [****, ()](\doibase 10.1103/PhysRevB.71.045110) [****, ()](\doibase 10.1103/RevModPhys.77.871) [****, ()](\doibase 10.1103/PhysRevB.72.035307) [****, ()](\doibase 10.1103/PhysRevB.91.075114) [****, ()](\doibase 10.1103/PhysRevB.91.035134) [****, ()](\doibase 10.1103/PhysRevX.4.031048) [****, ()](\doibase 10.1103/PhysRevB.92.045101) [****, ()](\doibase 10.1007/JHEP02(2015)172) [ ()](http://arxiv.org/abs/1601.00981), [“,” ](http://arxiv.org/abs/1602.05951) (), , [****, ()](\doibase 10.1103/PhysRevX.3.011016) [****, ()](\doibase 10.1103/PhysRevX.5.021029) [****, ()](\doibase 10.1103/PhysRevB.87.235122) [****, ()](\doibase 10.1103/PhysRevB.87.174412) [****, ()](\doibase 10.1103/PhysRevX.4.041049) [“,” ](http://arxiv.org/abs/1404.6659) (), , [****, ()](\doibase 10.1103/PhysRevB.89.165132) [****, ()](\doibase 10.1126/science.1243326) [****, ()](\doibase 10.1088/1742-5468/2013/09/P09016) [****, ()](\doibase 10.1103/PhysRevB.92.125111) [****, ()](\doibase 10.1103/PhysRevB.88.115137) [****, ()](\doibase 10.1103/PhysRevB.91.045134) [****, ()](\doibase 10.1103/PhysRevB.85.045104) [****, ()](\doibase 10.1103/PhysRevB.84.014527) [****, ()](\doibase 10.1103/PhysRevLett.99.196805) @noop [**]{} (, , ) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRev.176.1480) [****, ()](\doibase 10.1103/PhysRev.176.1489) [****, ()](\doibase 10.1103/PhysRev.144.1087) [****, ()](\doibase 10.1103/PhysRev.173.1536)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'As part of the European Supernova Collaboration we obtained extensive photometry and spectroscopy of the type Ia SN 2002dj covering epochs from 11 days before to nearly two years after maximum. Detailed optical and near-infrared observations show that this object belongs to the class of the high-velocity gradient events as indicated by Si, S and Ca lines. The light curve shape and velocity evolution of SN 2002dj appear to be nearly identical to SN 2002bo. The only significant difference is observed in the optical to near-IR colours and a reduced spectral emission beyond 6500 Å. For high-velocity gradient Type Ia supernovae, we tentatively identify a faster rise to maximum, a more pronounced inflection in the V and R light curves after maximum and a brighter, slower declining late-time B light curve as common photometric properties of this class of objects. They also seem to be characterized by a different colour and colour evolution with respect to “normal” SNe Ia. The usual light curve shape parameters do not distinguish these events. Stronger, more blueshifted absorption features of intermediate-mass elements and lower temperatures are the most prominent spectroscopic features of Type Ia supernovae displaying high velocity gradients. It appears that these events burn more intermediate-mass elements in the outer layers. Possible connections to the metallicity of the progenitor star are explored.'
author:
- |
G. Pignata,$^{1,2}$[^1] S. Benetti,$^3$ P. A. Mazzali,$^{4,5}$ R. Kotak,$^{6}$ F. Patat,$^7$ P. Meikle,$^{8}$ M. Stehle,$^4$ B. Leibundgut,$^{7}$ N. B. Suntzeff,$^{9}$ L. M. Buson,$^3$ E. Cappellaro,$^3$ A. Clocchiatti,$^{2}$ M. Hamuy,$^{1}$ J. Maza,$^{1}$ J. Mendez,$^{10}$ P. Ruiz-Lapuente,$^{10}$ M. Salvo,$^{11}$ B. P. Schmidt,$^{11}$ M. Turatto$^{3}$ and W. Hillebrandt$^{4}$\
$^1$ Departamento de Astronomía, Universidad de Chile, Casilla 36-D, Santiago, Chile\
$^2$ Departamento de Astronomía y Astrofísica, Pontificia Universidad Católica de Chile, Casilla 306, Santiago 22, Chile\
$^3$ INAF Osservatorio Astronomico di Padova, Vicolo dell Osservatorio 5, I-35122 Padova, Italy\
$^4$ Max-Planck-Institut für Astrophysik, Karl-Schwarzschild-Str. 1, D-85741 Garching bei München, Germany\
$^5$ Osservatorio Astronomico di Trieste, Via Tiepolo 11, I-34131 Trieste, Italy\
$^6$ Astrophysics Research Centre, School of Mathematics and Physics, Queen’s University Belfast, Belfast BT7 1NN, United Kingdom\
$^7$ European Southern Observatory, Karl-Schwarzschild-Str. 2, D-85748 Garching bei München, Germany\
$^8$ Blackett Laboratory, Imperial College London, Prince Consort Road, London SW7 2BW, United Kingdom\
$^9$ Department of Physics, Texas A&M University, College Station, TX 77843-4242\
$^{10}$ Department of Astronomy, University of Barcelona, Marti i Franques 1, E-08028 Barcelona, Spain\
$^{11}$ Research School of Astronomy and Astrophysics, Australian National University, Cotter Road, Weston Creek, ACT 2611, Australia\
date: 'Accepted ...... Received .......; in original form .......'
title: 'Optical and IR observations of SN 2002dj: some possible common properties of fast expanding SNe Ia '
---
\[firstpage\]
supernovae: general - supernovae: individual: SN 2002dj
Introduction
============
In the last decade type Ia Supernovae (SNe Ia) have been extensively used for cosmology yielding evidence that we live in an accelerating Universe (Riess et al. 1998; Perlmutter et al. 1999). Ongoing surveys such as ESSENCE [@Miknaitis07; @Wood-Vasey07] and SNLS [@Astier06] use the relation between the shape of the light curve and its peak luminosity [@Phillips93; @Phillips99; @Hamuy96; @Riess96; @Goldhaber01; @Guy05; @Prieto06] to constrain the equation-of-state parameter for dark energy. An important caveat is that those relations assume that SNe Ia are a one parameter family.
However, in recent years there has been growing evidence for the observational diversity among SN Ia and this is of prime interest in their application as distance indicators. Subtle but unequivocal differences between events with large wavelength coverage and dense temporal sampling are evident in light curve shapes, colour evolutions, luminosities, evolution of spectral lines and expansion velocities derived from the line shifts. The search for accurate correlations between photometric and spectroscopic properties could improve the luminosity calibration and help to shed light on the explosion mechanisms and progenitor system. Recently, @Benetti05 have identified three classes of SNe Ia based on their spectroscopic features. A similar classification is also reported in @Branch06. Interestingly, two of these classes are nearly indistinguishable in some of their photometric parameters (e.g. $\Delta m_{15}$, Phillips et al. 1993); yet they show clear spectroscopic differences. High velocity gradient (HVG) SNe are characterized by a fast decrease in their expansion velocity over time as measured from the minimum of the Si II (6355 Å) absorption line. On the other hand in the low velocity gradient (LVG) SNe group, which represents the majority of SN Ia, the evolution of the Si II (6355 Å) velocity is smooth, the ejecta expansion is slower than in HVG SNe. These spectroscopic differences may or may not affect the relations used to calibrate the luminosities of SNe Ia. Subtle differences in colour may exist among the different classes and lead to significant bias in reddening estimates. The existence of spectroscopic families suggests possible differences in the progenitor channels and/or explosion mechanisms. @Branch93 for example reported that SNe characterized by high expansion velocities, tend to explode in late type galaxies, pointing to progenitors arising from a young population.
In this paper we present the observations of the HVG SN 2002dj carried out by the European Supernova Collaboration (ESC). SN 2002dj ($\alpha =
13^h 13^m 00^s.34$, $\delta = -19^\circ 31' 08''.7$, J2000) was discovered in NGC 5018 on June 12.2 UT [IAUC 7918, @Hutchings02] and classified by ESC members as a Type Ia event on June 14.15 UT [IAUC 7919, @Riello02]. The layout of the manuscript is as follows. Observations and data reduction are presented in Sect. 2. We describe the reddening estimate in Sect. 3. In Sect. 4 we analyze the SN 2002dj optical and IR photometry comparing its properties with those of its kinematical twin, SN 2002bo. We also investigate whether they are representative of the HVG SNe group. Sect. 5 contains the determination of the SN 2002dj absolute luminosity and a characterization of the properties of its host galaxy. The optical and IR spectroscopy is analyzed in Sect. 6 and Sect. 7, respectively. The expansion velocities are discussed in Sect. 8 and models of the early and late spectra in Sect 9. We conclude in Sect. 10 by examining possible physical conditions in the HGV SNe ejecta that could explain part of their spectroscopic and photometric behavior.
Observations and Data Reduction
===============================
Instrument settings
-------------------
SN 2002dj was observed with seven different instruments in the optical and two facilities in the near infrared. Below we list the instruments and their main characteristics:
- 0.9m telescope located at Cerro Tololo Inter-American Observatory (CTIO) equipped with a CCD camera (2048$\times$2048, pixel size = 0.40 arcsec pixel$^{-1}$) and UBVRI Johnsons, Kron-Cousin standard filters.
- 1.0m telescope located at CTIO equipped with A Novel Double-Imaging CAMera (ANDICAM; Hawaii HgCdTe 1024$\times$1024, pixel size = 0.137 arcsec pixel$^{-1}$) and JHKs standard filters.
- European Southern Observatory (ESO) New Technology Telescope (NTT) located at La Silla Observatory and equipped with the ESO Multi Mode Instrument (EMMI) in RILD mode (2$\times$MIT/LL CCD 2048$\times$4096, pixel size = 0.167 arcsec pixel$^{-1}$) and BVRI standard filters with ESO identification numbers 605, 606, 608 and 610, respectively.
- ESO-NTT equipped with the Son of ISAAC camera (SofI; Hawaii HgCdTe 1024$\times$1024, pixel size = 0.29 arcsec pixel$^{-1}$) and JHKs standard filters.
- Danish 1.54m telescope located at La Silla Observatory equipped with the Danish Faint Object Spectrograph and Camera (DFOSC; MAT/EEV CCD 44-82 2048$\times$2048, pixel size = 0.39 arcsec pixel$^{-1}$) and UBVRI standard filters with ESO identification numbers 632, 450, 451, 452 and 425, respectively.
- ESO-Kueyen Very Large Telescope (VLT) located at Paranal Observatory equipped with the FOcal Reducer and low dispersion Spectrograph (FORS1; 2$\times$E2V 2048$\times$4096 pixel size = 0.2 arcsec pixel$^{-1}$) and BVRI standard filters with ESO identification numbers +34, +35, +36 and +37, respectively.
- Nordic Optical Telescope (NOT) located at Roque de los Muchachos Observatory equipped with the Andalucía Faint Object Spectrograph and Camera (ALFOSC; E2V 2048$\times$2048, pixel size = 0.19 arcsec pixel$^{-1}$) and UBVR standard filters with NOT identification numbers 7, 74, 75 and 76, respectively, and an interference i filter with number 12.
- Isaac Newton Group (ING) Jacobus Kapteyn Telescope (JKT) located at Roque de los Muchachos Observatory equipped with a CCD camera (SITe2 2048$\times$2048, pixel size = 0.33 arcsec pixel$^{-1}$) and UBVRI standard filters with ING identification numbers 3, 27, 30, 37 and 44, respectively.
- ING-Isaac Newton Telescope (INT) located at Roque de los Muchachos Observatory equipped with the Wide Field Camera (WFC; 4 thinned EEV 2048$\times$4096, pixel size = 0.33 arcsec pixel$^{-1}$) and UBVRi(sloan) standard filters with ING identification numbers 204, 191, 192, 193 and 215, respectively.
Optical and infrared photometry
-------------------------------
id U B V R I
---- ------------------ ------------------ ------------------ ------------------ ------------------
1 $-$ 17.91 $\pm$ 0.02 17.09 $\pm$ 0.03 16.61 $\pm$ 0.03 16.18 $\pm$ 0.04
2 15.79 $\pm$ 0.04 15.08 $\pm$ 0.01 14.09 $\pm$ 0.02 13.53 $\pm$ 0.02 13.02 $\pm$ 0.01
3 $-$ 18.11 $\pm$ 0.02 17.33 $\pm$ 0.04 16.88 $\pm$ 0.03 16.43 $\pm$ 0.02
4 $-$ 18.15 $\pm$ 0.04 17.51 $\pm$ 0.03 17.13 $\pm$ 0.01 16.76 $\pm$ 0.06
5 16.44 $\pm$ 0.05 16.45 $\pm$ 0.02 15.84 $\pm$ 0.01 15.48 $\pm$ 0.01 15.11 $\pm$ 0.01
6 $-$ 18.59 $\pm$ 0.02 17.52 $\pm$ 0.05 16.90 $\pm$ 0.01 16.31 $\pm$ 0.01
7 15.05 $\pm$ 0.03 15.08 $\pm$ 0.01 14.62 $\pm$ 0.02 14.33 $\pm$ 0.01 14.04 $\pm$ 0.03
8 18.01 $\pm$ 0.08 17.60 $\pm$ 0.01 16.57 $\pm$ 0.01 15.96 $\pm$ 0.02 15.38 $\pm$ 0.01
9 $-$ 19.25 $\pm$ 0.06 18.68 $\pm$ 0.04 18.32 $\pm$ 0.05 17.85 $\pm$ 0.08
10 17.31 $\pm$ 0.07 17.28 $\pm$ 0.03 16.62 $\pm$ 0.01 16.22 $\pm$ 0.04 15.85 $\pm$ 0.03
11 16.87 $\pm$ 0.06 16.76 $\pm$ 0.03 15.98 $\pm$ 0.03 15.50 $\pm$ 0.02 15.03 $\pm$ 0.01
\[tab2.1\]
UT date M.J.D. Phase$^a$ U B V R I Instr.
---------------- --------- ----------- ------------------ ------------------ ------------------ ------------------ ------------------ -----------
13/06/2002$^p$ 52439.1 -10.9 16.07 $\pm$ 0.04 16.06 $\pm$ 0.02 15.85 $\pm$ 0.02 15.66 $\pm$ 0.02 15.90 $\pm$ 0.03 CTIO 0.9m
14/06/2002 52439.9 -10.1 $-$ 15.64 $\pm$ 0.02 15.49 $\pm$ 0.02 15.27 $\pm$ 0.02 15.46 $\pm$ 0.02 EMMI
27/06/2002$^p$ 52452.9 2.9 14.30 $\pm$ 0.03 14.36 $\pm$ 0.02 14.14 $\pm$ 0.02 14.10 $\pm$ 0.02 14.46 $\pm$ 0.02 WFC
28/06/2002 52453.9 3.9 $-$ 14.41 $\pm$ 0.02 14.16 $\pm$ 0.02 14.12 $\pm$ 0.02 14.48 $\pm$ 0.02 JKT
29/06/2002 52455.0 5.0 14.41 $\pm$ 0.04 14.46 $\pm$ 0.02 14.20 $\pm$ 0.02 14.15 $\pm$ 0.03 14.51 $\pm$ 0.03 JKT
30/06/2002 52456.0 6.0 $-$ $-$ $-$ 14.24 $\pm$ 0.02 $-$ JKT
02/07/2002 52458.0 8.0 $-$ 14.66 $\pm$ 0.09 14.31 $\pm$ 0.02 14.31 $\pm$ 0.02 14.61 $\pm$ 0.02 JKT
03/07/2002 52458.1 8.1 14.86 $\pm$ 0.04 14.67 $\pm$ 0.02 14.27 $\pm$ 0.02 14.30 $\pm$ 0.02 14.66 $\pm$ 0.05 DFOSC
04/07/2002 52460.0 10.0 15.02 $\pm$ 0.04 14.81 $\pm$ 0.02 14.36 $\pm$ 0.02 14.46 $\pm$ 0.02 14.85 $\pm$ 0.05 DFOSC
07/07/2002$^p$ 52463.0 13.0 15.42 $\pm$ 0.04 15.13 $\pm$ 0.02 14.56 $\pm$ 0.02 14.61 $\pm$ 0.03 14.94 $\pm$ 0.04 DFOSC
09/07/2002 52465.0 15.0 15.75 $\pm$ 0.04 15.39 $\pm$ 0.02 14.73 $\pm$ 0.02 14.71 $\pm$ 0.02 14.97 $\pm$ 0.02 DFOSC
11/07/2002 52467.0 17.0 16.06 $\pm$ 0.04 15.61 $\pm$ 0.02 14.85 $\pm$ 0.02 14.76 $\pm$ 0.02 14.95 $\pm$ 0.02 DFOSC
17/07/2002$^p$ 52473.0 23.0 $-$ $-$ 15.12 $\pm$ 0.02 14.86 $\pm$ 0.02 14.81 $\pm$ 0.02 CTIO 0.9m
16/07/2002 52473.1 23.1 16.75 $\pm$ 0.04 16.29 $\pm$ 0.02 15.13 $\pm$ 0.02 14.83 $\pm$ 0.02 14.89 $\pm$ 0.02 DFOSC
19/07/2002 52476.0 26.0 16.99 $\pm$ 0.04 16.54 $\pm$ 0.02 15.40 $\pm$ 0.02 15.05 $\pm$ 0.02 14.85 $\pm$ 0.02 DFOSC
24/07/2002 52479.9 29.9 $-$ 16.89 $\pm$ 0.02 15.58 $\pm$ 0.02 15.14 $\pm$ 0.02 14.83 $\pm$ 0.02 ALFOSC
25/07/2002 52481.0 31.0 17.44 $\pm$ 0.05 16.91 $\pm$ 0.02 15.59 $\pm$ 0.02 15.18 $\pm$ 0.02 14.88 $\pm$ 0.02 DFOSC
29/07/2002 52484.9 34.9 17.59 $\pm$ 0.15 $-$ $-$ $-$ $-$ ALFOSC
08/08/2002$^p$ 52495.0 45.0 $-$ 17.41 $\pm$ 0.04 16.25 $\pm$ 0.02 15.95 $\pm$ 0.03 15.77 $\pm$ 0.06 CTIO 0.9m
31/08/2002 52518.0 68.0 18.43 $\pm$ 0.09 $-$ 16.83 $\pm$ 0.02 $-$ $-$ DFOSC
02/09/2002 52520.0 70.0 $-$ $-$ 16.87 $\pm$ 0.02 $-$ $-$ DFOSC
06/09/2002 52524.0 74.0 $-$ 17.75 $\pm$ 0.19 $-$ 17.06 $\pm$ 0.06 17.01 $\pm$ 0.48 DFOSC
25/03/2003$^p$ 52724.3 274.3 $-$ 21.00 $\pm$ 0.04 20.71 $\pm$ 0.03 21.39 $\pm$ 0.06 20.78 $\pm$ 0.09 FORS1
22/05/2004$^p$ 53147.2 697.2 $-$ $>$24.7$^b$ $>$23.4$^b$ $-$ $-$ FORS1
\
$^a$ Counted since the time of the $B$ maximum brightness M.J.D.=52450 $\pm$ 0.7\
$^b$ Upper limit\
$^p$ Photometric night\
\[tab2.2\]
id J H Ks
------ ------------------ ------------------ ------------------
2 12.23 $\pm$ 0.03 11.75 $\pm$ 0.03 11.63 $\pm$ 0.03
2 IR 14.80 $\pm$ 0.03 14.18 $\pm$ 0.03 13.95 $\pm$ 0.04
7 13.63 $\pm$ 0.04 13.42 $\pm$ 0.04 13.38 $\pm$ 0.04
11 14.32 $\pm$ 0.04 13.89 $\pm$ 0.04 13.76 $\pm$ 0.04
: Magnitudes for the local IR photometric sequence in the field of SN 2002dj (Fig. \[fig2.1\] right panel). The data were obtained during three photometric nights with SofI.
\[tab2.3\]
For the optical photometric observations basic data reduction (bias and flat-field correction) was performed using standard routines in IRAF[^2]. In Table 1 are reported the magnitudes of the local photometric sequence identified in Fig. \[fig2.1\], that were calibrated against Landolt standard stars [@Landolt92] in the seven photometric nights marked in Table 2. SN photometry was performed using the PSF fitting technique. Only in the $B$ and $V$ FORS1 images obtained on March 25, 2003, was it possible to remove the galaxy. For the other instruments the lack of a suitable image without the SN has prevented us from using the template subtraction technique. We note that the FORS1 $B$ and $V$ band frames on which the SN is still relatively bright yield magnitudes in agreement to the ones derived from the template subtracted images. This gives us confidence that the PSF photometry performed around maximum, when the SN signal-to-noise ratio is high, was not significantly contaminated by the host galaxy background.
The near-IR data reductions were also performed using standard IRAF routines. Dark and flat-field corrections were applied to the scientific frames. For each night a sky background image was created by taking the median of the dithered science frames. For the ANDICAM images this approach lead to an imperfect removal of the galaxy light. Nevertheless, the good agreement between the SofI and ANDICAM photometry gives us confidence that the small galaxy residuals did not significantly bias the ANDICAM magnitudes. The sky background image was then subtracted from each single frame. For SofI, an illumination correction was also applied to all images. Four stars close to the SN position were calibrated in the $JHKs$ bands using standard stars from @Persson98 during three photometric nights marked in Table \[tab2.4\]. The resulting magnitudes are listed in Table \[tab2.3\]. They agree well with those reported in the 2MASS catalog with $J_{seq}-J_{2MASS}$=0.01 $\pm$ 0.01, $H_{seq}-H_{2MASS}$=0.00 $\pm$ 0.01 and $Ks_{seq}-Ks_{2MASS}=0.01 \pm
0.01$. The final SN calibration in all bands was performed using the S-correction technique (see Appendix for details). Thanks to the very good spectroscopic coverage we could compute the S-terms using only the spectra of SN 2002dj for the $B$, $V$, $R$, $I$ bands, while for the $U$, $J$, $H$, $Ks$, the SN 2002dj spectra were complemented with those of SN 2005cf [@Garavini07] and SN 2002bo [@Benetti04].
UT date M.J.D. Phase$^a$ J H Ks Instr.
---------------- --------- ----------- ------------------ ------------------ ------------------ ---------
13/06/2002 52439.0 -11.0 15.63 $\pm$ 0.06 15.64 $\pm$ 0.09 15.65 $\pm$ 0.13 ANDICAM
14/06/2002 52440.0 -10.0 15.27 $\pm$ 0.04 15.31 $\pm$ 0.06 15.25 $\pm$ 0.09 ANDICAM
15/06/2002 52441.1 -8.9 14.97 $\pm$ 0.03 15.13 $\pm$ 0.03 15.11 $\pm$ 0.03 SoFi
19/06/2002$^p$ 52444.1 -5.9 14.63 $\pm$ 0.04 14.84 $\pm$ 0.03 14.71 $\pm$ 0.03 SoFi
20/06/2002 52446.0 -4.0 14.58 $\pm$ 0.11 14.80 $\pm$ 0.07 14.66 $\pm$ 0.08 ANDICAM
24/06/2002 52450.0 -0.0 14.60 $\pm$ 0.03 14.84 $\pm$ 0.06 14.52 $\pm$ 0.10 ANDICAM
24/06/2002$^p$ 52450.0 -0.0 14.62 $\pm$ 0.03 14.83 $\pm$ 0.03 14.54 $\pm$ 0.03 SoFi
27/06/2002 52453.1 3.1 14.86 $\pm$ 0.05 14.90 $\pm$ 0.04 14.68 $\pm$ 0.09 ANDICAM
30/06/2002 52456.0 6.0 15.30 $\pm$ 0.13 15.09 $\pm$ 0.10 14.88 $\pm$ 0.20 ANDICAM
08/07/2002 52464.0 14.0 16.30 $\pm$ 0.09 15.05 $\pm$ 0.05 15.04 $\pm$ 0.09 ANDICAM
11/07/2002 52467.0 17.0 16.26 $\pm$ 0.05 14.97 $\pm$ 0.03 14.90 $\pm$ 0.06 ANDICAM
11/07/2002$^p$ 52467.0 17.0 16.20 $\pm$ 0.03 14.95 $\pm$ 0.03 14.83 $\pm$ 0.03 SoFi
14/07/2002 52470.0 20.0 16.08 $\pm$ 0.13 14.81 $\pm$ 0.05 14.75 $\pm$ 0.08 ANDICAM
17/07/2002 52473.0 23.0 15.86 $\pm$ 0.04 14.75 $\pm$ 0.03 14.65 $\pm$ 0.05 ANDICAM
25/07/2002 52481.0 31.0 15.57 $\pm$ 0.07 14.79 $\pm$ 0.05 14.81 $\pm$ 0.09 ANDICAM
28/07/2002 52484.0 34.0 15.63 $\pm$ 0.05 15.06 $\pm$ 0.04 15.12 $\pm$ 0.07 ANDICAM
31/07/2002 52487.0 37.0 15.89 $\pm$ 0.03 15.23 $\pm$ 0.07 15.43 $\pm$ 0.29 ANDICAM
07/08/2002 52494.0 44.0 16.58 $\pm$ 0.14 15.61 $\pm$ 0.09 15.65 $\pm$ 0.35 ANDICAM
10/08/2002 52497.0 47.0 16.68 $\pm$ 0.10 15.73 $\pm$ 0.06 15.93 $\pm$ 0.16 ANDICAM
13/08/2002 52500.0 50.0 16.89 $\pm$ 0.12 15.85 $\pm$ 0.06 16.04 $\pm$ 0.15 ANDICAM
30/08/2002 52517.0 67.0 18.20 $\pm$ 0.38 16.66 $\pm$ 0.03 16.84 $\pm$ 0.03 SoFi
$^a$ Counted since the time of the $B$ maximum brightness M.J.D.=52450 $\pm$ 0.7\
$^p$ Photometric night\
Optical and infrared spectroscopy
---------------------------------
Optical spectra were reduced using the [*ctioslit*]{} package in IRAF. Optimal extraction was usually obtained by weighting the signal according to the intensity profile along the slit. Sky subtraction was carried out by fitting a low order polynomial on either side of the SN spectrum and the wavelength solution was determined from arc spectra. The wavelength calibration was checked against bright night-sky emission lines. Flux calibration was performed by means of spectrophotometric standard stars [@Hamuy92; @Hamuy94] and checked against the photometry. When discrepancies occurred, the flux of the spectrum was scaled to match the photometry. In nights with good observing conditions the agreement with the photometry was found to be within 10%.
The near-IR spectra were reduced using standard procedures in the FIGARO 4 environment [@Shortridge95]. Wavelength calibration was carried out using Xe arc spectra and the accuracy of the solution was checked using OH sky lines. The flux calibration was secured with respect to near-IR solar-analog standard stars observed close in time and airmass to the SN to minimize variations in the atmospheric absorptions [@Maiolino96]. The SN spectrum was divided by the standard stars spectrum to remove the strong telluric IR features. The resulting spectrum was then multiplied by the solar spectrum to eliminate the intrinsic features introduced by the solar-type star. Like in the optical, the flux calibration was checked against the photometry and, if necessary, scaled to match the latter.
Interstellar extinction
=======================
Presence of dust in NGC 5018 has been claimed in several studies. The map produced by @Fort86 shows a dust lane embedded in the galaxy. Subsequently, @Carollo94 and @Goudfrooij94a detected patchy dust in the galaxy core. None of those maps shows evidence of dust at the location of SN 2002dj. Our deep images taken with the VLT add details to the @Carollo94 colour maps, but again do not show the presence of dust at, or close to, the SN position. SN spectra can be used to measure the equivalent width ($EW$) of the Na I D lines as a proxy to quantify absorption. The Galactic colour excess we obtain using one of the two linear relations reported in @Turatto03 ($E(B-V)=0.16 \times EW$ yielding $E(B-V)_{Galactic}=0.11 \pm 0.03$ from $EW=0.72 \pm 0.19$ Å) is in very good agreement with that reported by @Schlegel98 ($E(B-V)_{Galactic}=0.096$) for the SN 2002dj line of sight. No absorption line of interstellar Na I D has been detected from the host galaxy. Since SN 2002dj was observed both in the optical and in the near IR, we can apply several methods to estimate its extinction from the observed colour. The results are reported in Table \[tab3.4\]. We note that all methods involving $B-V$ point to an $E(B-V)$ greater than that due to Milky Way absorption alone, indicating a small but not negligible dust extinction towards SN 2002dj in the host galaxy, while the methods based on $V-IR$ colours provide a negligible colour excess for the host galaxy. A similar apparent inconsistency occurred in the case of SN 2002bo, where the reddening derived from $B-V$ colour seems to be larger than that derived through detailed synthetic spectra modeling [@Stehle05] and, as in the case of SN 2002dj, it is significantly larger than that derived from the $V-IR$ colours. The peculiar colours of SN 2002bo were already noted by @Benetti04 and @Krisciunas04b.
With the aim to check whether HVG SNe could in general have intrinsically different colours which could bias their reddening estimate, we examined the other HVG SNe reported in the literature for which multi-epoch spectra show a persistent high expansion velocity (see Table \[tab5.3\])[^3]. Unfortunately, all objects except SN 2002bf and SN 2004dt, show signs of absorption in their host galaxies (i.e. Na I D lines). The latter two SNe are also not useful for this purpose because the photometric follow-up for SN 2002bf started $\sim$ at +6 days and definitive photometry is not yet available for SN 2004dt. Hence, for all objects in Table \[tab5.3\] the intrinsic colour is rather uncertain. To investigate the possible peculiar colours for HVG SNe we compared the the scatter between the colour excess derived using the $B-V$ and $V-I$ colours among LVG SNe with the systematic difference we found for our sample of HVG SNe.
To determine the difference in extinction measure for LVG SNe we used Table 2 of @Phillips99 and obtained $E(B-V)_{B-V}$-$E(B-V)_{V-I}=-0.01 \pm 0.08$. This value was derived by selecting SNe with $0.95 < \Delta m_{15} < 1.60$ in order to avoid colour peculiarity not accounted by the @Phillips99 relation [@Garnavich04]. For HVG SNe in addition to SN 2002bo and SN 2002dj, only in the case of SN 1997bp, [*BVRI*]{} photometry [@Altavilla04; @Jha06] was suitable to compute the colours at maximum. For these three objects we find $E(B-V)_{B-V}$-$E(B-V)_{V-I}=0.17 \pm
0.03$. The systematic difference in reddening depending on the reference color in HVG SNe is hence nearly $2\sigma$ the scatter observed in LVG SNe hinting at intrinsic colour differences. Anyway, the colour evolution among HVG SNe shows significant variations (see Fig. \[fig4.5\]) making it difficult to establish an overall difference with LVG SNe.
\
Method $E(B-V)$ Reference
------------------- ----------------- ---------------- --
$B_{max}-V_{max}$ 0.22 $\pm$ 0.06 @Phillips99
$(B-V)_{max}$ 0.24 $\pm$ 0.06 @Altavilla04
$V_{max}-I_{max}$ 0.10 $\pm$ 0.06 @Phillips99
$(B-V)_{CMAGIC}$ 0.32 $\pm$ 0.06 @Wang03
$(B-V)_{Jha}$ 0.29 $\pm$ 0.06 @Jha06
$(B-V)_{tail}$ 0.27 $\pm$ 0.14 @Phillips99
$(B-V)_{Wang}$ 0.21 $\pm$ 0.04 @Wang05
$(B-V)_{V-H}$ 0.05 $\pm$ 0.04 @Krisciunas04a
$(B-V)_{V-K}$ 0.10 $\pm$ 0.06 @Krisciunas04a
: SN 2002dj total (Galactic + host galaxy) reddening from different methods.
\[tab3.4\]
Optical and Infrared photometry
===============================
{width="175mm"}
![Early phase $UBVRI$ light curves of SN 2002dj. The symbols, SNe and vertical shifts are the same as reported in Fig. \[fig4.1\]. []{data-label="fig4.4"}](sn2002dj_lce_opt.ps){width="84mm"}
SN 2002dj was observed in $UBVRI$ from $-$11 to +697 days and from $-$11 to +68 days in $JHKs$. The corresponding light curves are shown in Figs. \[fig4.1\], \[fig4.4\], \[fig4.2\], \[fig4.3\], while colour curves are displayed in Figs. \[fig4.5\] and \[fig4.6\].
Light curves
------------
In Fig. \[fig4.4\] the $BVI$ and, to a lesser degree $R$, HVG SNe (filled symbols; SN 2002dj, SN 2002bo and SN 2006X) show a faster rise to maximum if compared to LVG SNe (empty symbols; represented by SN 2001el, SN 2003du and SN 2005cf). On the other hand we note that SN 2006X and SN 1997bp peak later in $VRI$ ($\sim$+2 days) than the LVG average population (see histograms in @Contardo02). Therefore the fast rise in $VRI$ of the HVG objects could be induced by a peculiar behaviour of the $B$ light curve that historically is used to set the phase of the SN Ia. Comparing the light curves of SN 2006X (SN 1997bp does not have observations close to maximum) with those of SN 2005cf and leaving the $B$ maximum time as a free parameter in the fit, we found a phase shift of $-$0.9 days for the $V$ and $R$ bands and $-$0.5 for the $I$. The latter results reduce the possibility that the early time behaviour of HVG SNe could simply be a phase mis-calibration. This is also confirmed by the IR observation of SN 2002dj, SN 2002bo and SN 2006X which show that those SNe reach the peak at the same epoch as the @Krisciunas04a templates (Fig. \[fig4.2\]). For the $U$ band the situation is more complex. The data are more scattered and systematic errors due to standard filter mis-matching could play an important role. Therefore a photometrically more homogeneous data set is necessary to confirm the trends observed in the other filters.
Another feature visible in Fig. \[fig4.1\] is that the HVG SN 1997bp, SN 2002bf[^4] and SN 2006X show more pronounced inflections around +25 days in $V$ and $R$. In the case of SN 2002bf the large photometric errors associated with the corresponding points demand caution, while for SN 2006X, at least part of this effect could be due to its high extinction shifting the effective wavelength of the passband to redder [@EliasRosa06; @Wang07a]. The absorption suffered by SN 1997bp is low and the photometric errors small, hence this feature must be intrinsic. Finally, we note that SN 1997bp, SN 2002bf and SN 2006X are the highest velocity SNe Ia in our sample, with velocities about 5000 km s$^{-1}$ higher than those of SN 2002dj and SN 2002bo (Fig. \[fig8.1\]). This suggest that photometric features could be correlated with the kinematics of the ejecta.
Since, as previously mentioned, at early epochs the light curves of HVG SNe are different from those of LVG SNe, we decided to constrain the epoch of the SN 2002dj $B$ maximum brightness, hereafter T($B_{max}$), using only the $B$ light curve of SN 2002bo. In fact, it is not viable to determine the maximum brightness directly from the SN 2002dj light curve because it happens to fall in a gap of 12 days with no photometric observations. Having constrained T($B_{max}$), (M.J.D.=$52450.0 \pm 0.7$) we used the light curves of SNe with similar behavior around maximum to estimate the SN 2002dj peak magnitude in $UBVRI$ (Table \[tab5.1\]). The values of the light curve parameters $\Delta m_{15}$ and stretch factor $s$ [@Perlmutter97a] are reported in Table \[tab5.2\]. The stretch factor was obtained by fitting the SN 2002dj $B$ light curve to the @Leibundgut88 template complemented by @Goldhaber01. We note that the values of $\Delta m_{15}$ and $s$, satisfy the relations reported by @Perlmutter97b and @Altavilla04 reasonably well.
As in the optical, in the near infrared bands both SN 2002dj and SN 2002bo show a faster rise to maximum brightness than SN 2001el and the @Krisciunas04a “mid-range” decliner IR templates. SN 2006X follows the templates closely, although very early time observations are missing.
![$JHK$ light curves of SN 2002dj, different colours refer to different instruments. For comparison the light curves of SN 2002bo [@Krisciunas04b], SN 2006X [@Wang07a], SN 1998bu [@Hernandez00], SN 2001el [@Krisciunas03], SN 2005cf [@Pastorello07] and the mid range decliner templates from @Krisciunas04a (solid line) are displayed. The light curves of different SNe were vertically shifted in order to match them at maximum.[]{data-label="fig4.2"}](sn2002dj_lc_ir.ps){width="84mm"}
At epochs $>$+35 days in $I$, $R$ and $V$ HVG SNe are perhaps slightly brighter, but the points show large spread, covering possible subtle systematic differences. In $B$ HVG SNe are clearly brighter than LVG SNe. Interestingly, as in the case of the inflections occurring around +25 days, the brightest objects among HVG SNe are those displaying the highest expansion velocities. @Wang07a already pointed out that the decline rates $\gamma$ between +40 and +100 days (Pskovskii 1984) of the HVG SN 1984A, SN 2002bo and SN 2006X are smaller than for “normal” SNe Ia (i.e. $\gamma$ = 1.40 $\pm$ 0.10 mag 100 d$^{-1}$), being 1.14 $\pm$ 0.06, 1.17 $\pm$ 0.10 and 0.92 $\pm$ 0.05, respectively. Fitting the three $B$ observations of SN 2002bf at phase $>$ +40.0 days we obtained $\gamma$ = 0.90 $\pm$ 0.60. The large error is due to the noisy photometry. For SN 2002dj, considering the only two $B$ observations in the phase interval, we derived $\gamma$ =1.17 $\pm$ 0.13. Also for SN 1997bp we have only two points from which yield $\gamma$ =0.83 $\pm$ 0.18. Therefore, although with limited statistical strength, we confirm the finding of @Wang07a. In $U$ the fastest expanding SN 1997bp and SN 2006X seem to follow the $B$ band trend, while SN 2002dj and SN 2002bo are closer to the evolution of SN 2005cf and SN 1998bu.
![Same as Fig. \[fig4.1\] showing the late phase $BVRI$ light curves of SN 2002dj.[]{data-label="fig4.3"}](sn2002dj_lcl_opt.ps){width="84mm"}
At late phases SN 2002dj, although an HVG SN resembles the behavior of LVG SNe. At the latest epoch for which we obtained measurable data (+274 days), SN 2002dj appears in fact in all bands close to the light curve of SN 2001el [@Stritzinger07 see Fig. \[fig4.3\]]. SN 1997bp is instead $\sim$0.5 mag and $\sim$0.3 brighter in $B$ and $V$, respectively. Its late time decline rate is similar to the other SNe, which argues against a possible contribution of a light echo. With the aim to further investigate (or exclude) the presence of a light echo, $B$ and $V$ imaging of SN 2002dj was performed with FORS1 on April 22, 2004, i.e. 697.2 days after $B$ maximum light. Nothing is visible at the position of the SN and through re-detection of artificial stars, we estimated 3$\sigma$ upper limits of $B$ $\sim$24.7 and $V$$\sim$23.4.
Colour curves
-------------
![De-reddened $(B-V)_0$, $(V-R)_0$ and $(V-I)_0$ colour curves of SN 2002dj. []{data-label="fig4.5"}](sn2002dj_cc_bvri_opt.ps){width="84mm"}
The de-reddened colour curves of SN 2002dj are compared in Fig. \[fig4.5\] with those of SN 1997bp, SN 2002bo, SN 2002bf, SN 2001el, SN 2003du and SN 2005cf. For HVG SNe the reddening was removed using the values of $E(B-V)$ reported in Table \[tab5.3\] and assuming $R_V$=3.1. Because of their large reddening, SN 2002bo and SN 2006X are not considered in the HVG [*vs.*]{} LVG comparison.
In the $(B-V)_0$ colour, at least until +30 days, HVG SNe form a redder sequence than LVG SNe. Afterward the two groups mix due to the HVG slower decline rate in the $B$ band. Note that the latter poses problems if one wishes to apply the Lira relation [@Phillips99] to estimate the reddening of HVG [see also @Wang07a]. Therefore, both maximum light and tail $B-V$ colours could bias the reddening estimation of these objects. In $(V-R)_0$ there is not a clear separation between HVG and LVG SNe, while in $(V-I)_0$ HVG SNe seem to form a bluer sequence with respect to LVG SNe, but the significance of the differences is reduced by the large calibration uncertainties of the $I$ band [see @Pignata04].
The $(V-R)_0$ and $(V-I)_0$ colours of SN 2002bo are clearly bluer then those of SN 2002dj. This reflects the lack of flux in the red part of the SN 2002bo spectrum noticeable in the comparison with SN 2002dj (see section 6.3).
The $(V-J)_0$, $(V-H)_0$ and $(V-K)_0$ colours curves of SN 2002dj (Fig. \[fig4.6\]) follow the @Krisciunas04a “mid range” templates reinforcing the idea of negligible host galaxy reddening. As in the case of the $(V-R)_0$ and $(V-I)_0$ colours, SN 2002bo is bluer at all epochs. The difference to SN 2002dj increases toward redder bands, suggesting that the lack of flux starting around 6500 Å continues monotonically towards longer wavelengths.
![De-reddened $(V-J)_0$, $(V-H)_0$ and $(V-K)_0$ colour curves of SN 2002dj. []{data-label="fig4.6"}](sn2002dj_cc_vmir.ps){width="84mm"}
Absolute luminosity
===================
The Virgo infall model of @Kraan-Korteweg86 yields a distance to relative Virgo of 2.52 for NGC 5018. Assuming a Virgo distance of 15.3 Mpc [@Freedman01] we obtain 38.55 Mpc ($\mu=32.93$) for the host galaxy of SN 2002dj. The Nearby Galaxies Catalog [@Tully88] reports a larger distance (40.9 Mpc, $\mu$=33.1), while @Hilker96 using the globular cluster luminosity function derived a distance modulus ranging between 32.8 and 34.0. For a consistent comparison to SN 2002bo we decided to assume 38.55 Mpc as the best distance estimate for SN 2002dj. With this distance and the reddening estimate we derived the corresponding absolute magnitudes in $UBVRIJHKs$. The values are reported in Table \[tab5.1\], along with those obtained from the $\Delta m_{15}$ absolute magnitude relations of @Prieto06 for $BVRI$ and those of @Krisciunas04 for the $JHKs$. The absolute magnitudes in all the bands are in agreement within 2$\sigma$ with the values predicted by the previously mentioned relations, although SN 2002dj appears slightly underluminous.
\
Filter m(obs)$^a$ M(cor)$^b$ M(ave)$^c$
-------- ------------------ ------------------- ------------------- --
$B$ 14.30 $\pm$ 0.04 -19.03 $\pm$ 0.23 -19.34 $\pm$ 0.02
$V$ 14.15 $\pm$ 0.04 -19.08 $\pm$ 0.19 -19.26 $\pm$ 0.02
$R$ 14.10 $\pm$ 0.05 -19.06 $\pm$ 0.17 -19.26 $\pm$ 0.02
$I$ 14.35 $\pm$ 0.06 -18.72 $\pm$ 0.14 -19.00 $\pm$ 0.02
$J$ 14.56 $\pm$ 0.02 -18.46 $\pm$ 0.11 -18.57 $\pm$ 0.03
$H$ 14.80 $\pm$ 0.03 -18.18 $\pm$ 0.11 -18.24 $\pm$ 0.04
$Ks$ 14.51 $\pm$ 0.05 -18.45 $\pm$ 0.11 -18.42 $\pm$ 0.04
: Peak magnitudes of SN 2002dj
\
$^a$ Apparent magnitude\
$^b$ Absolute magnitude corrected for reddening\
$^c$ Average absolute magnitude for SNe Ia\
\[tab5.1\]
Using the $UBVRIJHKs$ observations of SN 2002dj and adding the UV contribution determined by @Suntzeff96, we constructed a $uvoir$ light curve that is compared in Fig. \[fig5.1\] with that of SN 2002bo and a model generated using a Monte Carlo light curve code [@Cappellaro97]. The curve of SN 2002bo published by @Benetti04 has been rescaled to the new distance estimate ($\mu$ = 31.76 $\pm$ 0.07) reported by @EliasRosa08. Within the errors the absolute luminosities of the two SNe are comparable. The model provides a good fit to the brightness and width near the peak, yielding a $^{56}$Ni mass of 0.45 $\pm$ 0.04 $M_{\odot}$ for SN 2002dj, but after +20 days it overestimates the observed flux. This is probably due to sudden changes in the opacity at advanced stages which our model does not properly take into account.
![Comparison between the $uvoir$ light curves of SN 2002dj and SN 2002bo. The error bars account only for the uncertainties in reddening and photometry. The best fit model of SN 2002dj ($^{56}$Ni=0.45$M_{\odot}$) is also shown. []{data-label="fig5.1"}](2002dj_bol.ps){width="84mm"}
An intriguing host galaxy
-------------------------
\
------------------------------------ -----------------------------
Host galaxy NGC 5018
Galaxy type E3
RA (2000) 13$^h$13$^m$01$^s$.7
Dec (2000) $-$19$^{\circ}$31’12”.8
Heliocentric
Recession velocity \[km s$^{-1}$\] 2816 $\pm$ 1 $\dag$
Distance modulus 32.93 $\pm$ 0.15
$E(B-V)$ $0.096 \pm 0.05$
Date of $B$ max (MJD) 52450.0 $\pm$ 0.7
Offset from the nucleus $8''.9$ West, $2''.8$ South
$\Delta m_{15}(B)$ 1.08 $\pm$ 0.05
stretch factor in $B$ 0.97 $\pm$ 0.02
------------------------------------ -----------------------------
: Main parameters of SN 2002dj and its host galaxy.
\
$\dag$ [@Rothberg06] \[tab5.2\]
\
Type $\Delta m_{15}(B)$
----------- -------------- -------- -------------------- ----------------
SN 1981B NGC 4536 SAB 1.11 $-$
SN 1983G NGC 4753 I0 1.37 $-$
SN 1984A NGC 4419 SB 1.21 $-$
SN 1989A NGC 3687 SAB 1.06 $-$
SN 1997bp NGC 4680 Sp 1.00 0.18(0.04)$^a$
SN 2002bf CGCG 266-031 SB(r)b $-$ 0.08(0.04)$^b$
SN 2002bo NGC 3190 Sap 1.17 0.38(0.10)$^c$
SN 2002dj NGC 5018 E3 1.08 0.096(0.05)
SN 2004dt NGC 799 SB(s) 1.21 $-$
SN 2006X NGC 4321 SABbc 1.31 1.41(0.04)$^d$
: Host galaxies of HVG SNe considered in the paper
\
$^a$ average value between $E(B-V)_{B-V}$ and $E(B-V)_{V-I}$\
$^b$ From @Leonard05\
$^c$ From @Stehle05\
$^d$ From @Wang07a \[tab5.3\]
NGC 5018 is the dominant giant elliptical of a small group. It is peculiar in several respects. First, @Schweizer90 classified it as one of their best candidates for a recent major merger. Second, although it is morphologically classified as gE, its nuclear optical spectrum distinguishes itself by having the weakest measurement of the absorption line index Mg$_2$ (0.218) for its velocity dispersion among over 400 gEs surveyed by @Davies87. Its UV (IUE) spectrum lacks the prominent UV-upturn shortward of 2000Å, which is typical of old, metal-rich spheroids [@Bertola93]. Nevertheless, through a spectral index study, @Leonardi00 found indications for the presence of a relatively young stellar population ($\sim$2.8 Gyr) in the central region of NGC 5018 with nearly solar metallicity. A similar age ($\sim$3Gyr) is proposed by @Buson04 through a UV spectroscopy study. @Kim88 detected a H I gas bridge connecting NGC 5018 with the nearby spiral NGC 5022, indicating an ongoing flow toward the giant elliptical. Furthermore, possible young ($\sim 10^8$ years) globular clusters have been claimed by @Hilker96. Finally the H$_{\alpha}$+\[N II\] maps reported in @Goudfrooij94b reveal the presence of an extended emission distributed as a strongly warped disk covering the SN 2002dj position that the authors suggest to be associated with star forming regions.\
@Branch93 reported that SNe Ia characterized by high expansion velocities, tend to explode in late type galaxies and suggested that those objects could have younger progenitors than “normal” SNe Ia. The HVG SNe analyzed in this paper confirm this (Table \[tab5.3\]). @Hamuy00 pointed out that slow declining events preferentially occur in late type galaxies, while fast decliners occur in all type of galaxies. A similar result was obtained by @Sullivan06 for a sample of high-z SNe Ia. These observations seem at odds with the fact that SN 2002dj was hosted by an elliptical galaxy. However, given the peculiarities of NGC5018 it is possible that SN 2002dj could be associate with a relatively young stellar population.
Optical spectra
===============
Spectra of SN 2002dj have been obtained at 20 epochs spanning phases between $-10.9$ and +274 days (Table \[tab6.1\]), following the rapid evolution during the early epochs in detail and sampling the late phases more sparsely. The early discovery of SN 2002dj allows us to compare its spectral evolution (Fig. \[fig6.1\]) with that of SN 2002bo and other SNe, especially at early epochs.
---------- --------- ------- ------------ ---- --------
UT date M.J.D. Instr.
02/06/13 52439.1 -10.9 3300-9300 9 DFOSC
02/06/14 52440.1 -9.9 3700-9800 10 EMMI
02/06/15 52441.0 -9.0 3300-9300 9 DFOSC
02/06/15 52441.1 -8.9 9400-16500 21 SofI
02/06/16 52442.0 -8.0 3300-9300 9 DFOSC
02/06/18 52444.0 -6.0 3300-9300 9 DFOSC
02/06/19 52444.1 -5.9 9400-25400 21 SofI
02/06/20 52446.0 -4.0 3200-9200 9 DFOSC
02/06/21 52447.0 -3.0 3300-9300 9 DFOSC
02/06/24 52450.0 0.0 9400-16500 21 SofI
02/07/03 52458.7 8.7 3400-9300 9 DFOSC
02/07/04 52460.0 10.0 3300-10300 9 DFOSC
02/07/07 52463.0 13.0 3200-9100 9 DFOSC
02/07/11 52467.0 17.0 3300-10300 9 DFOSC
02/07/11 52467.0 17.0 9400-25200 21 SofI
02/07/16 52473.1 23.1 3300-9100 9 DFOSC
02/07/25 52481.0 31.0 3700-10300 9 DFOSC
02/09/02 52520.0 70.0 3400-9100 9 DFOSC
03/02/01 52672.3 222.3 3400-7500 14 EFOSC
03/03/25 52724.3 274.3 3600-8600 12 FORS1
---------- --------- ------- ------------ ---- --------
: Optical and IR spectroscopic observations of SN 2002dj.
$^a$Counted since the time of the $B$ maximum brightness M.J.D.=$52450 \pm 0.7$\
\[tab6.1\]
![Optical spectral evolution of SN 2002dj. With the exception of the last three spectra, which for presentation purposes have been multiplied by a constant, the other spectra of the sequence have been only shifted vertically. The $\earth$ symbol shows the position of the main telluric features. The spectra are labeled with the epoch in days past $B$ maximum.[]{data-label="fig6.1"}](2002dj_evol.ps){width="84mm"}
Premaximum phase
----------------
In Fig. \[fig6.4\] we show the first spectrum of SN 2002dj obtained about 11 days before $B$ maximum together with coeval spectra of SN 2002bo, SN 2003du, SN 2005cf and SN 1994D. Already at a first glance it is evident how the minima of the Ca II H&K, S II and Si II absorption features are more blueshifted in the spectra of SN 2002dj and SN 2002bo than in LVG SN 2003du, SN 2005cf and SN 1994D. The Si II lines are also deeper in HVG SNe. The blueshift is due to the higher expansion velocities of the HVG objects (cf. Fig. \[fig8.1\]). The blending of high velocity detached features [@Branch04] with the photospheric line component could also play an important role in blueshifting the absorption minimum of Ca II H&K and Si II $\lambda$6355 Å in HVG SNe as it is for the Ca II IR triplet (Mazzali et al. 2005, Tanaka et al. 2008). For example, the minimum of Si II $\lambda$6355 Å in SN 2002dj and SN 2002bo is at the same position of the high velocity component identified by @Garavini07 in the spectrum of SN 2005cf. The feature around 4400 Å marked in Fig. \[fig6.4\] is attributed to Si III $\lambda$$\lambda$4553,4568 Å. The strength of this line correlates with temperature and is clearly detected in LVG SNe. The spectrum of SN 2002dj only has a hint Si III, and it is completely absent in SN 2002bo. As shown in Fig. \[fig6.10\], the low temperature of HVG SNe at early epochs is confirmed by their high $R$(Si II). If well traced by $R$(Si II) the temperature should rapidly rise in HVG SNe and quickly reach that of LVG SNe. Indeed at $-$4 days (Fig. \[fig6.5\]) the Si III is visible with similar strength in all SNe but SN 2002bo. The latter SN also lacks the weak absorption at $\sim$5550 Å also attributed to Si III ($\lambda$5740 Å) that is present in all other SNe. It is interesting to note how the minimum of the Si III line is at nearly the same wavelength for all SNe, indicating that the line is formed at roughly the same position in velocity space. The Ca II H&K doublet (blended with Si II $\lambda$3858 Å) appears to be double-dipped in LVG SN 1994D and SN 2003du, while it is a single feature in SN 2002dj and SN 2002bo. @Hatano00 suggested that this difference is related to higher expansion velocities that lead to a greater degree of line blending. The Ca II line blending could also be enhanced by the presence of high velocity features. On the other hand @Lentz00 argued that the double-dip disappears when Si II is weaker than Ca II. Given that in all fast expanding SNe the Si II lines are stronger than in the slow expanding SNe, line blending seems a most consistent explanation.
![Spectrum of SN 2002dj taken at $-$10.9 days. The coeval spectra of SN 2002bo [@Benetti04], SN1994D [@Patat96], SN 2003du [@Stanishev07] and SN 2005cf [@Garavini07] are shown for comparison. The spectra have been corrected for reddening and redshift.[]{data-label="fig6.4"}](2002dj_first.ps){width="84mm"}
![Same as Fig. \[fig6.4\], but for a phase of $-$4 days. The bibliographic sources for the spectra of SN 2002bo, SN 1994D, SN 2003du and SN 2005cf are the same as those in Fig. \[fig6.4\].[]{data-label="fig6.5"}](2002dj_second.ps){width="84mm"}
Postmaximum phase
-----------------
A week after maximum the Si II $\lambda$6355 Å line in the HVG SN 2002dj and SN 1997bp remains blueshifted with respect to the LVG SN 1994D, SN 2003du and SN 2005cf (Fig. \[fig6.6\]). The Si II $\lambda$6355 Å feature is still deeper in SN 2002dj and SN 1997bp than in LVG SNe, while, contrary to the pre-maximum spectra, the S II line $\lambda$5640 Å is weaker. SN 2002dj and SN 1997bp also show less substructure in the region between 4700 and 5100 Å with respect to SN 1994D, SN 2003du and SN 2005cf. Finally we note that the absorption minimum of the feature around 5700 Å at this epoch attributed to a blend of Na I D and Si II $\lambda$5972 Å [@Garavini07] has about the same position, but in SN 2002dj and especially in SN 1997bp it is weaker than in the other SNe. By one month after maximum the spectra are dominated by iron peak elements (Fig. \[fig6.7\]). The absorption minimum of the lines are nearly at the same position in all selected SNe. The early fast expansion of SN 2002dj, SN 2002bo and SN 1997bp is now only reflected in the broadness of the Ca II IR triplet. The Na I D line is still weaker than in SN 2003du and SN 2005cf and is now slightly redshifted. The line at $\sim$5300 Å identified by @Branch05 as due to Cr II is barely visible in the faster expanding SNe while it remains visible in the slower ones. Finally, around two months after maximum brightness the features due to Fe II are slightly redshifted or at the same position in SN 2002dj and SN 2002bo with respect to SN 2003du and SN 2005cf (Fig. \[fig6.8\]). Also the Na I D line is slightly redshifted and its intensity is comparable in all SNe. Interestingly, in the region between 4700 and 5100 Å SN 2002dj and SN 2002bo still show less substructures. At these epochs line blending due to the high velocity should not be severe in HVG SNe, so the lack of lines points to a temperature or abundance effect.
![Same as Fig. \[fig6.4\] for a phase of +9 days. The bibliographic sources for the spectra are the same as those in Fig. \[fig6.4\] and Fig. \[fig6.5\], while the spectrum of SN 1997bp comes from the Asiago archive (unpublished).[]{data-label="fig6.6"}](2002dj_third.ps){width="84mm"}
![Same as Fig. \[fig6.4\] for phase +31 days. The bibliographic sources for the spectra are the same as those in Fig. \[fig6.5\].[]{data-label="fig6.7"}](2002dj_fourth.ps){width="84mm"}
![Same as Fig. \[fig6.4\], but for phase +70 days. The bibliographic sources for the spectra are the same as those in Fig. \[fig6.4\].[]{data-label="fig6.8"}](2002dj_fiveth.ps){width="84mm"}
![Evolution of $R$(SiII) for a sample of HVG SNe (blue filled symbols) and LVG SNe (empty green symbols). []{data-label="fig6.10"}](RSiII_evol.ps){width="84mm"}
SN 2002dj versus SN 2002bo, a close comparison
----------------------------------------------
We compare the spectra of SN 2002dj with those of SN 2002bo at four different epochs (Fig. \[fig6.11\]). In the wavelength interval covered by the $B$ and $V$ filters the similarity between the two objects is evident, but around 6500 Å SN 2002bo shows less flux than SN 2002dj. Since the spectra of both SNe were calibrated in different epochs with different instruments, we can rule out an instrumental mis-calibration. Alternatively the difference could be attributed to an incorrect estimate of the reddening of the two objects. However, in this case we would expect a gradual deviation of the two spectra, but the discrepancy looks more like a break possibly due to a lack of opacity in the photosphere of SN 2002bo. For the comparison we chose to obtain the best fit between 4000 and 6000 Å. If instead we scale the spectra of SN 2002bo in order to get the best match with the redder part of those of SN 2002dj, we end up with a flux excess in SN 2002bo at blue wavelength. While this is possible, the comparison of the SN 2002bo colours with those of other LVG and HVG SNe and the results of the spectral modeling reported by @Stehle05 point to a flux deficit in the red part instead of an excess in the blue.
![Comparison of the reddening and redshift corrected spectra of SN 2002dj (solid lines) and SN 2002bo (dotted lines). The epochs labeled in the plot are for SN 2002dj; the SN 2002bo spectra are coeval within 1.0 day. The SN 2002bo spectra were scaled in order to match those of SN 2002dj between 4000 and 6000 Å. []{data-label="fig6.11"}](sn2002dj_comp.ps){width="84mm"}
IR spectra
==========
![Infrared spectral evolution of SN 2002dj. The spectra are corrected for reddening and redshift. For graphic exigence the spectra have been multiplied by a suitable factor and vertically shifted.[]{data-label="fig7.1"}](2002dj_ir_spec_evol.ps){width="84mm"}
Although the sample of published IR spectra of SNe Ia is growing rapidly, their number is still small compared with those obtained at optical wavelengths. In this context the spectral evolution of SN 2002dj in the IR from day -8.9 to +17 presented in Fig. \[fig7.1\] constitutes an important contribution to the global database.
![IR spectrum of SN 2002dj taken at -8.9 days. The coeval spectra of SN 2002bo [@Benetti04], SN1994D [@Meikle96], SN 1999ee [@Hamuy02] and SN 2003du [@Stanishev07] are shown for comparison. The spectra have been corrected for reddening and redshift.[]{data-label="fig7.2"}](2002dj_ir_first.ps){width="84mm"}
Fig. \[fig7.2\] compares the first SN 2002dj IR spectrum with those of SN 2002bo, SN 1994D and SN 1999ee at similar epochs. The IR spectrum at this phase is nearly featureless with the remarkable exception of an absorption at $\sim$10500 Å. The identification of this line remains open. It was attributed by @Meikle96 to He I $\lambda$10830 Å or Mg II $\lambda$10926 Å based on an Local Thermodynamic Equilibrium (LTE) treatment. Using a NLTE approach, @Mazzali98a discussed the conditions under which the He line could form, and found that its time evolution was at odds with the observed one. They therefore favoured either Mg or Si. A Mg II identification was also claimed by @Wheeler98.\
SN 1999ee shows an emission line at $\sim$12300 Å which @Rudy02 identified as due to Fe III in the spectra of SN 2000cx. The presence of this feature in those SNe is due to their higher photospheric temperature. As expected, it is absent in the spectra of both SN 2002dj and SN 2002bo.
![Same as Fig. \[fig7.2\] for phase -6 days for SN 2002dj, SN1994D and SN 2003du. The bibliographic sources for the spectra are the same as those in Fig. \[fig7.2\], while the spectrum of SN 2003cg was published in @EliasRosa06.[]{data-label="fig7.3"}](2002dj_ir_second.ps){width="84mm"}
The -5.9 day spectrum of SN 2002dj shows a feature around 12100 Å (Fig. \[fig7.3\]) which was not visible in the earliest spectrum (cf. Fig. \[fig7.1\]). @Marion03 identified this feature as a blend of several Ca II lines between 12430 Å and 12830 Å. Like for the Mg II line, the blueshift of the Ca II minimum is larger in SN 2002dj than in all other SNe. Finally at $\sim$20800 Å a weak emission is visible in the spectra of SN 2002dj, SN 2003du and SN 2003cg attributed to Si III by @Wheeler98.
Following @Rudy02 we fitted the two earliest IR spectra of SN 2002dj to estimate the black body temperature. We used only sections of the spectrum free of telluric features or absorption lines. Assuming $E(B-V)$=0.096 for SN 2002dj we obtain T=11900 K at -8.5 days compared with T=12900 K and T=23500 K for SN 1994D and SN 1999ee reported in Fig. \[fig7.2\]. For the latter two SNe we assumed $E(B-V)$=0.06 $\pm$ 0.02 [@Patat96] and $E(B-V)$=0.30 $\pm$ 0.04 [@Stritzinger02], respectively. The high temperature of SN 1999ee ($\Delta m_{15}(B)=0.96 \pm 0.02$) is consistent with the presence of the Fe III emission feature in its spectrum. It is also similar to the 25000 K temperature of SN 2000cx [@Rudy02] taken at -8.0 days, which shows a stronger Fe III emission than SN 1999ee. This gives us confidence that the technique provides reliable results, when applied to early time spectra. The sequence of increasing temperature from SN 2002dj to SN 1999ee confirms the results obtained from $R$(Si II) (Fig. \[fig6.10\]). Indeed, for the SN 2002dj at -5.9 days we found a much higher temperature (T=20500 K). This confirms its fast rise as traced by the $R$(Si II) and the appearance of a Si III line in the $K$ band in the spectrum of SN 2002dj.
@Rudy02 proposed that the black body IR spectrum fit can be used to derive also an estimate of absorption. We note, however, that the resulting $\chi^2$ surface for temperature and colour excess is nearly flat providing only weak constraints. At +17 days we see that several lines populate the IR spectrum of SN 2002dj and those of other SNe (Fig. \[fig7.4\]). In particular, a broad P-Cygni feature attributed to Fe II [@Marion03] is observed at $\sim$12350 Å. In SN 2002dj the blueshift of its minimum is roughly the same as in SN 2003cg and SN 2002er, while SN 1999ee displays higher velocities. Finally, @Marion03 suggested that the minimum around $\sim$15000 Å defines the transition from partial to complete silicon burning. The fact that the velocities inferred by the position of this minimum is the same in SN 2002dj, SN 2003cg and SN 2002er indicate that these SNe have an iron core of similar size.
![Same as Fig. \[fig7.2\] for phase +17.0 days. The bibliographic sources for the spectra are the same as those in Fig. \[fig7.2\] and Fig. \[fig7.3\]. The spectrum of SN 2002er that was published by @Kotak05.[]{data-label="fig7.4"}](2002dj_ir_fourth.ps){width="84mm"}
Expansion velocities
====================
In Fig. \[fig8.1\] we show the expansion velocities measured from the blueshift of the minima of the Si II ($\lambda$6355 Å), S II ($\lambda$5460 Å) and Ca II (H&K). It is evident that at early epochs the HVG SN 1984A, SN 1997bp, SN 2002bo and SN 2002dj are faster than LVG SNe in all those lines. The two groups are separated most strongly in Si II velocity. Nevertheless, as noted by @Jeffery90 strong features like Si II ($\lambda$6355 Å) reliably trace the photospheric velocity only at early epochs, when the mass above the photosphere is small. For HVG SNe like SN 2002dj and SN 2002bo, where Si II is particularly strong and where high velocity features could play an important role [@Mazzali05], this is unlikely even at very early times. In SN 2002bo for example the photospheric evolution computed by @Stehle05 even at -13 days has a velocity $\sim$2000 km s$^{-1}$ lower than that measured from the minimum of the Si II line (cf. Fig. \[fig8.1\] upper panel). The relatively weak S II (5460 Å) seems to be a better tracer of the photospheric velocity. In HVG SNe this feature is at higher velocities than in LVG SNe, but due to its steep velocity decline it approaches the LVG group already before maximum. If we take S II as a proxy for the photospheric velocity, its evolution may indicate that at very early epochs the photosphere in HVG SNe is located at larger radii than in LVG SNe, but moving inward more quickly, it reaches a comparable position already before maximum. For Ca II (H&K), some LVG SNe closely resemble the velocity evolution of HVG SNe, making this feature unsuitable to separate the two groups.
In the case of Mg II (10926 Å), the expansion velocity measured from the absorption minimum at -8.5 days is $\sim$15000 km s$^{-1}$ for SN 2002dj and SN 2002bo and significantly slower ($\sim$11000 km s$^{-1}$) for SN 1994D. The feature appears deeper in SN 1994D than in SN 2002dj (Fig. \[fig8.2\]). We do not observe a change in the line velocity from the first to the second epoch in both SNe and infer that even earlier than one week before maximum the photosphere is below the magnesium layer. Since Mg II is produced in the outer layers by oxygen burning [@Wheeler98], this means that only IR spectra observed earlier than this may tell us about the boundary between explosive carbon and oxygen burning. Nevertheless, we note in Fig. \[fig8.2\] that the blue wing of the Mg II line extends up to $\sim$19000 km s$^{-1}$ in SN 2002dj and only up to $\sim$15000 km s$^{-1}$ in SN 1994D. Hence, the burning front must have propagated further out in SN 2002dj than in SN 1994D.
![Velocity evolution of Si II (top panel), S II (middle panel) and Ca II (bottom panel) in SN 2002dj, SN 2002bo, SN 1997bp, SN 1984A, SN 2002bf, SN 1998bu, SN 2003du, SN 2005cf, SN 1994D and SN 2003cg. The solid lines represent the evolution of the photospheric velocity of SN 2002bo computed by @Stehle05, while the dashed lines show the velocities predicted by @Lentz00 model for the case of $\times$10 (top line at epoch 0), $\times$3 (middle line) and $\times$1/3 (bottom line) C+O solar metallicity.[]{data-label="fig8.1"}](CaII_SII_SiII_ev.ps){width="84mm"}
![Comparison between the Mg II $\lambda$10926 feature of SN 2002dj (solid line) and SN 1994D (dashed line) at $\sim -$8.5 days (left panel) and -$6.0$ days (right).[]{data-label="fig8.2"}](MgII_vel.ps){width="84mm"}
Spectral Modelling
==================
Early phase
-----------
We modeled the earliest spectrum of SN 2002dj obtained 10.9 days before B maximum using the Montecarlo code described in Mazzali & Lucy (1993), Lucy (1999) and Mazzali (2000), modified to include abundance stratification as described in @Stehle05. While a good fit to the overall spectrum can be obtained for the input parameters reported in the caption of Fig. \[fig9.1\], it is not possible with the density structure of W7 [@Nomoto84] to reproduce the high velocity features. One modification that we adopted here to overcome this is to increase the mass at high velocity. In particular, we assume the presence of 0.04$M_{\sun}$ of material at velocities above 18000 km s$^{-1}$. This may be caused by interaction of the outer ejecta with circumstellar material. @Tanaka08 discuss this and other possibilities in detail. Let us note that the model suffers only to a minor degree from the problem of overfitting the flux to the red of the Si II ($\lambda$6355 Å) line suggesting high line opacity in all regions. This can not be the case for SN 2002bo where the observed spectrum is noticeably below the model [@Stehle05].
Finally, we note that the spectral modelling seems to exclude the presence of C II in the earliest spectrum of SN 2002dj. The presence of this high velocity feature was found in the -9 days spectrum of SN 1998aq [@Branch03] and in the spectra of SN 2006gz from -14 to -10 days [@Hicken07]. High velocity C II ($\lambda$6578 Å) in the -14 days spectrum of SN 1990N was also claimed by @Fisher97, although @Mazzali01 showed that a high-velocity component of Si II ($\lambda$6355 Å) was another viable identification. Moreover, the presence of small quantities of carbon has been shown to improve the fit of the very early spectra of SN 1994D [@Branch05], SN 2001el and SN 2003du [@Tanaka08]. The latter SN Ia all are LVG SNe, therefore the absence of C II in the -10.9 days spectrum of SN 2002dj and in the -12.9 days spectrum of SN 2002bo [@Stehle05] could indicate that the carbon burning has penetrated further out in these explosions [@Stehle05].
![Comparison between the first spectrum of SN 2002dj taken at -10.9 days and a synthetic spectrum. The the input parameters for the latter are $E(B-V)$=0.096, v$_{ph}$=11600 km s$^{-1}$, log(L)=42.33.[]{data-label="fig9.1"}](model_2002dj.ps){width="60mm"}
Nebular phase
-------------
In order to estimate the errors on the $^{56}$Ni mass estimate obtained by modelling nebular spectra, we generated a set of synthetic spectra computed using two different codes and two different assumptions of the $^{56}$Ni mass distribution. The six synthetic spectra together with the two observed spectra of SN 2002dj are presented in Fig. \[fig9.2\]. Both codes assume line-formation in non-local thermodynamic equilibrium (non-LTE) conditions. The code described in @Bowers97 assumes a uniform density, homologously expanding sphere containing iron, cobalt and sulphur, with relative abundances modulated by the radioactive decay. In computing the SED only singly and doubly-ionized species are considered because they are predicted to dominate the spectrum. The code described in @Mazzali01 can use either a homologously expanding nebula of finite extent, uniform density and composition, as in @Bowers97, or stratification in density and abundance [@Mazzali07]. In this case, the density profile is taken from the explosion model W7 [@Nomoto84]. $\gamma$-rays and positrons are emitted at various depths according to the distribution of $^{56}$Ni and their propagation and deposition is followed using a Monte Carlo scheme similar to that discussed by @Cappellaro97 for their light-curve models. Gas heating and cooling as well as line emissivity are computed in non-LTE in each radial shell, and the line profiles are computed assuming each shell to contribute to a truncated parabola, with an inner truncation point corresponding in velocity to the inner boundary of the shell considered. The sum of these truncated parabolas is the emerging spectrum. Constant $\gamma$-ray ($\kappa_{\gamma} =
0.027$cm$^2$g$^{-1}$) and positron opacities ($\kappa_{e^+} =
7$cm$^2$g$^{-1}$) are assumed in both cases.
The synthetic spectra yield slightly different results for the $^{56}$Ni mass, reflecting the different spectral calibrations and code assumptions. The @Bowers97 model gives a $^{56}$Ni mass of 0.59 M$_{\sun}$ and 0.55 M$_{\sun}$ for the +222d and +274d spectra, respectively, while the one-zone model from @Mazzali01 yields 0.55 M$_{\sun}$ for both epochs. With the @Mazzali07 stratified model we obtain 0.45 M$_{\sun}$ for both spectra. Results obtained with the same code on the two different epochs are consistent, indicating the robustness of the fit. On the other hand, both one-zone models yield a $^{56}$Ni mass slightly larger than the stratified model. @Ruiz-Lapuente95 found similar results with their stratified nebular model as opposed to the one-zone model [@Ruiz-Lapuente92]. The main difference between the models is that in the one-zone codes the $^{56}$Ni density is uniform, while in the stratified code it is higher near the centre, which leads to a higher $\gamma$-ray deposition at late times and reduces the $^{56}$Ni mass necessary to produce a given luminosity. The $^{56}$Ni mass obtained with the stratified model is in excellent agreement with the one obtained from the $uvoir$ light curve, suggesting that this assumption provides a good description of the physical conditions in the nebula.
![Nebular spectra of SN 2002dj taken at +222 days (top panel) and +274 (bottom panel) compared with various synthetic spectra. []{data-label="fig9.2"}](2002dj_late_model.ps){width="84mm"}
Discussion and conclusion
=========================
We have presented optical and IR photometric and spectroscopic observations of the HVG Type Ia SN 2002dj. The ejecta kinematics of this object are very similar to that of SN 2002bo. The direct comparison of the line shapes and strengths of the two objects reveals an even more impressive similarity. The one obvious distinction is a peculiar flux deficit redward of 6500 Å in SN 2002bo.
The light curves are nearly identical with the notable exception of the $B$ band at epochs later than +40 days. However, the $V-IR$ colours are remarkably different. SN 2002bo is much bluer confirming the lack of flux redward of $\sim$6500 Å. Finally, also the bolometric absolute luminosity around maximum of the two objects turns out to be very similar. The IR contribution is in in fact very small at those epochs [@Suntzeff96].
Starting from the photometric similarities of SN 2002dj and SN 2002bo, we studied the characteristics of other HVG SNe. Although the number of objects for which reliable multi-band photometry covering both very early and late light curves phases is limited, the HVG SNe analyzed in this paper seem to share the following photometric characteristics if compared with LVG SNe:
1. A general fast rise to maximum brightness in all filters
2. More pronounced inflections in the $V$ and $R$ bands around +25 days
3. A brighter, more slowly declining $B$ light curve after +40 days
4. Possibly different colours
To complete the picture these properties should be combined with parameters one can derive from the spectral evolution:
1. Fast expanding ejecta
2. Strong absorption due to intermediate mass elements (IME) at early phases
3. Absence of carbon in the very early spectra
4. Early time low photospheric temperature rapidly increasing toward maximum brightness
5. Nebular spectra with kinematics not so different or even slower than that of LVG SNe.
@Benetti04 suggested that the high expansion velocities of IME are due to the fact that in HVG SNe the burning front extends further out into the outermost layers than in LVG SNe. This hypothesis is supported by the extension of the blue wing of the Mg II line at velocities close to 19000 km sec$^{-1}$ observable in the early IR spectra of SN 2002dj and by the absence of carbon in the -10.9 spectrum of SN 2002dj and -12.9 spectrum of SN 2002bo. In this scenario the density of IME is enhanced in the layers where usually unburned carbon and oxygen are present. The IME density enhancement in the outermost layers would induce stronger absorptions at early epochs and increase the opacity in these regions, moving the photosphere to larger radii. The blending between high velocity features and the photospheric component could also contribute to the absorption enhancement [@Tanaka08]. Since the photosphere is located at high velocity shells, this could justify both the fast rise to maximum observed in the light curve and, because the heating produced by the $\gamma-$rays is less effective, the lower temperature in the line forming region. Due to their high velocity the density in those shells decreases rapidly and hence the photosphere quickly recedes into the ejecta. Indeed, around maximum, the light curves of SN 2002dj, SN 2002bo and SN 2006X approached those of LVG SNe. Also the temperature is nearly identical as traced by the $R$(Si II) evolution and by the appearance of the Si III lines, the absorption minimum of which is at the same position in all SNe. The initially low and then rapidly increasing temperature of SN 2002dj was confirmed by fitting a black body to its two earliest IR spectra.
We note that in HVG SNe more pronounced inflections in the $V$ and $R$ light curves around +25 days seem to correlate with expansion velocities. The fastest expanding SN 1997bp, SN 2002bf and SN 2006X display a stronger inflection than in SN 2002bo and SN 2002dj, where the effect is barely visible. We must stress that in the case of SN 2006X, due to the large reddening, the intensity of this feature has to be taken with caution. Also the brighter $B$ light curve after +40 days seems to correlate with the expansion velocity. @Wang07b suggested that SN 2006X declined more slowly after +40 days and could be due to a light echo produced by the circumstellar dust. In the case of SN 2002dj there is no clear evidence of dust obscuring its light. but it is also the HVG SNe with the $B$ light curve closer to “normal” SNe Ia. Nevertheless, SN 1997bp and especially SN 2002bf with low absorption closely resemble the behavior of SN 2006X having also comparable expansion velocities. Moreover, part of the reddening estimated for those SNe could be biased due to their peculiar colour and colours evolution.
Regarding the colours of HVG SNe we note that @Dominguez01 using a set of Delayed Detonation models (DD) have found that increasing the metallicity of the progenitor white dwarf, the $B-V$ colour of the resulting SN Ia becomes redder. Unfortunately, this study does not extend the results to other bands and is limited to a metallicity Z=0.02. However, @Timmes03 have proposed that the effect of the progenitor metallicity is significantly enhanced at solar metallicity and above. In particular Fig. 9 of Dominguez et al. (2001) shows that passing from Z=0.001 to Z=0.02 the flux in $B$ is depressed, while that in $V$ is enhanced. If this is maintained at higher metallicity it could be the source of the red $B-V$ and blue $V-I$ colours we found in SN 2002bo, SN 1997bp and especially SN 2002dj where the effect of reddening should be negligible. Furthermore, Timmes et al. (2003) have pointed out that in metal rich progenitors due to the more efficient electron capture, more $^{54}$Fe and $^{58}$Ni are produced at the expense of $^{56}$Ni. @Mazzali06 used this finding to explore the effect on spectra and light curves by varying the fraction of $^{54}$Fe and $^{58}$Ni produced in burning to nuclear statistical equilibrium and concluded that the colours of SN Ia with a larger fraction of these elements (i.e. the more metal rich ones) have redder $B-V$ colour. Moreover, indications of larger $^{54}$Fe and $^{58}$Ni and lower $^{56}$Ni production for the HVG SN 1981B, SN 2002bo and SN 2002dj with respect the similar $\Delta m_{15}(B)$ LVG SN 1990N, SN 2003du and SN 2003cg have been found by @Mazzali07. @Lentz00 pointed out that increasing the abundance of all elements heavier than oxygen in the unburned C+O layer of the pure deflagration model W7 [@Nomoto84] leads to a larger blueshift of SN Ia absorption lines. This is due to the fact that the higher opacity provided by the large amount of metals confines the line forming region to the outer (faster) part of the ejecta. Although, this supports the hypothesis of metal rich progenitors for HVG SNe, we favour the configuration with an extended burning front to explain the HVG SNe observables.
To further complicate the picture we must mention that @Hoeflich98 found that SN Ia with metal rich progenitors should have a slower rise time to maximum than metal poor ones clearly contrary to what we found of SN 2002dj, SN 2002bo and SN 2006X. Given the very small sample of HVG SNe, the uncertainties afflicting their colours and the model shortcomings we briefly recapped here, our suggestion of metal rich progenitor for HVG SNe are tentative and provide possible avenues to explore when more data on HVG SNe and more sophisticated models will be available.
Fast expanding SNe are rare both at low and high redshift [@Blondin06; @Balland07; @Hook05]. The systematic error they could introduce into cosmological studies due to their peculiar colours appears to be negligible. Detailed analysis will become possible when a larger sample of well-observed rapidly evolving SNe Ia will be available.
Acknowledgments {#acknowledgments .unnumbered}
===============
G.P acknowledges support by the Proyecto FONDECYT 3070034. This work was also supported by the European Community’s Human Potential Programme under contract HPRN-CT-2002-00303, “The Physics of Type Ia Supernovae”. It is partially based on observations made with ESO Telescopes under programme ID 169.D-0670. This work is also based on observations performed at the Jacobus Kapteyn Telescope (JKT) and the Isaac Newton Telescope of the Isaac Newton Group at La Palma, Spain, the Nordic Optical Telescope at La Palma, Spain, and the 0.9m, 1.0m telescopes at Cerro Tololo Interamerican Observatory. A.C., M.H. and J.M. acknowledge support from Centro de Astrofisica FONDAP 15010003, and Núcleo Milenio P06-045-F funded by Programa Bicentenario de Ciencia y Tecnología from CONICYT and Programa Iniciativa Científica Milenio from MIDEPLAN. A.C. also acknowledges support from Proyecto FONDECYT 1051061. M.H. further acknowledges support from Proyecto FONDECYT 1060808. SB, EC and MT are supported by the Italian Ministry of Education via the PRIN 2006 n.2006022731 002. This work has made use of the NASA/IPAC Extragalatic database (NED) which is operated by the JET Propulsion Laboratory, California Institute of Technology, under contract with the US National Aeronautic and Space Administration.
[99]{}
Altavilla G. et al., 2004, MNRAS, 349, 1344
Astier et al., 2006, A&A, 447, 31
Baldwin J. A., Stone R. P. S., 1984 MNRAS, 206, 241
Balland C. et al., 2007, A&A, 464, 827
S. et al., 2004, MNRAS, 348, 261
S. et al., 2005, ApJ, 623, 1011
F., [Burstein]{} D., [Buson]{} L. M., 1993, ApJ, 403, 573
M. S., 1990, PASP, 102, 1181
S. et al. 2006, AJ, 131, 1648
Bowers E. J. C., Meikle, W. P. S., Geballe T. R., Walton N. A., Pinto P. A., Dhillon V. S., Howell S. B., Harrop-Allin M. K., 1997, MNRAS, 290, 663
D., [van den Bergh]{} S., 1993, AJ, 105, 2231
Branch D., 2003, AJ, 126, 1489
Branch D., 2004, ApJ, 606, 413
Branch D., Baron E., Hall N., Melakayil M., Parrent J., 2005, PASP, 117, 545
Branch D. et al., 2006, PASP, 118, 560
Buson L. M., Bertola, F., Bressan A., Burstein D., Cappellari M., 2004, A&A, 423, 965
Cappellaro E., Mazzali P. A., Benetti S., Danzinger I. J., Turatto M., Della Valle M., Patat F., 1997, A&A, 328, 203
, C. M., [Danziger]{}, I. J., 1994, MNRAS, 270, 743
G., [Leibundgut]{} B., [Vacca]{} W. D., 2000, A&A, 359, 876
R. L., [Burstein]{} D., [Dressler]{} A., [Faber]{} S. M., [Lynden-Bell]{} D., [Terlevich]{} R. J., [Wegner]{}, G., 1987, ApJS, 64, 581
Dominguez I, Höflich P., Straniero O., 2001, ApJ, 557, 279
Elias Rosa N. et al., 2006, MNRAS, 369, 1880
Elias Rosa N. et al., MNRAS, 384, 107
Fisher A., Branch D., Nugent P., Baron E., 1997, ApJ, 481, L89
B. P., [Prieur]{} J.-L., [Carter]{} D., [Meatheringham]{} S. J., [Vigroux]{}, L., 1986, ApJ, 306, 110
Freedman W. L. et al., 2001, ApJ, 553, 47
Garavini G. et al., 2007, A&A, 471, 527
Garnavich M. P. et al., 2004, AJ, 613, 1120
G. et al., 2001, ApJ, 558, 359
P., [Hansen]{} L., [Jorgensen]{}, H. E., [Norgaard-Nielsen]{} H. U., de Jong T., van den Hoek L. B., 1994a A&AS, 104, 179
P., [Hansen]{} L., [Jorgensen]{}, H. E., [Norgaard-Nielsen]{} H. U., 1994b A&AS, 105, 341
Guy J., Astier P., Nobili S., Regnault N., Pain R., 2005, A&A, 443, 781
M., [Walker]{} A. R., [Suntzeff]{} N. B., [Gigoux]{} P., [Heathcote]{} S. R., [Phillips]{} M. M., [1992]{}, PASP, 104, 533
M., [Suntzeff]{} N. B., [Heathcote]{} S. R., [Walker]{} A. R., [Gigoux]{} P., [Phillips]{} M. M., [1994]{}, PASP, 106, 566
M., [Phillips]{} M. M., [Suntzeff]{} N. B., [Schommer]{} R. A., [Maza]{} J., [Smith]{} R. C., [Lira]{} P., [Aviles]{} R., 1996, AJ, 112, 2438
M., Trager S. C., Pinto P. A., [Phillips]{} M. M., [Schommer]{} R. A., Ivanov V., [Suntzeff]{} N. B., 2000, AJ, 120, 1479
M. et al., 2002, AJ, 124, 417
K., [Branch]{} D., [Lentz]{} E. J., [Baron]{} E., [Filippenko]{} A. V., [Garnavich]{} P. M., 2000, ApJL, 543, L49
Hernandez M. et al., 2000, MNRAS, 319, 223
Hicken M., Garnavich P. M., Prieto J. L., Blondin S., DePoy D. L., Kirshner R. P., Parrent J., 2007, ApJ, 669, L17
M., [Kissler-Patig]{} M., 1996, A&A, 314, 357
Hoeflich P., Wheeler J. C., Thielemann F. K., 1998, ApJ, 495, 617
Hook I. M. et al., 2005, AJ, 130, 2788
Hutchings D., [Li]{} W. D.2002 IAUC, 7918, 1
Jeffery D. J., Branch D., 1990, in J.C. Wheeler, T. Piran, S. Weinberg eds. Proc. Jerusalem Winter School for Theoretical Physics. World Scientific Publishing Co. p. 149
S. et al., 1999, ApJS, 125, 73
S. et al., 2006, AJ, 131, 527
Kim, D. W., Jura M., Guhathakurta P., Knapp G. R., van Gorkom J. H., 1988, ApJ, 330, 684
D. L., 1985, RGO/La Palma technical note N$^{\circ}$ 31.
Kotak R. et al., 2005, A&A, 436, 1021
Kraan-Korteweg R. C., 1986, A&AS, 66, 255
K. et al., [2003]{}, AJ, 125, 166
K., Phillips M. M., Suntzeff N. B., 2004, AJ, 602, L81
K. et al., [2004a]{}, AJ, 127, 1664
K. et al., 2004b, AJ, 128, 3034
A. U., 1992, AJ, 104, 340
B., 1988, Ph.D. thesis, Univ. Basel
E. J., [Baron]{} E., [Branch]{} D., [Hauschildt]{} P. H., [Nugent]{} P. E., 2000, Apj, 530, 966
Leonard D. C., Li W., Filippenko A. V., Foley R. J., Chornock R., 2005, ApJ, 632, 450
Leonardi A. J., Worthey G., 2000, AJ, 534, 650
Lucy L. B., 1999, A&A,345, 211
Maiolino R., Rieke G. H., Rieke M. J., 1996, AJ, 111, 537
G. H., [H[" o]{}flich]{} P., [Vacca]{} W. D., [Wheeler]{} J. C., 2003, ApJ, 591, 316
Mazzali P. A., Lucy L. B., 1993, A&A, 279, 477
Mazzali P. A., Lucy L. B., 1998, MNRAS 295, 428
Mazzali P., 2000, A&A, 363, 705
Mazzali P. A., Nomoto K., Patat F., Maeda K., 2001, ApJ, 559, 1047
Mazzali P. et al., 2005, ApJ, 623, L37
Mazzali P. A., Podsiadlowski F., 2006, MNRAS, 369, L19
Mazzali, P. A., Röpke F. K., Benetti S., Hillebrandt W., 2007, Science, 315, 825
W. P. S. et al., 1996, MNRAS, 281, 263
Miknaitis G. et al., 2007, ApJ, 666, 674
S., [Goobar]{} A., [Knop]{} R., [Nugent]{} P., 2003, A&A, 404, 901
Nomoto K., Thielemann F.-K., Yokoi K., 1984, ApJ, 1984, 286, 644
P., [Phillips]{} M., [Baron]{} E., [Branch]{} D., [Hauschildt]{} P., 1995, ApJ, 455, L147
Pastorello A. et al., 2007, MNRAS, 376, 1301
F., [Benetti]{} S., [Cappellaro]{} E., [Danziger]{} I. J., [della Valle]{} M., [Mazzali]{} P. A., [Turatto]{} M., 1996, MNRAS, 278, 111
S. et al., 1997a, in Thermonuclear Supernova, ed. P. Ruiz-Lapuente, R. Canal, & J. Isern (Dordrecht:Kluwer), p. 749
S. et al, 1997b, ApJ, 483, 565
S. E., [Murphy]{} D. C., [Krzeminski]{} W., [Roth]{} M., [Rieke]{} M. J., 1998, AJ, 116, 2475
M. M., 1993, ApJ, 413, 105
M. M., [Lira]{} P., [Suntzeff]{} N. B., [Schommer]{} R. A., [Hamuy]{} M., [Maza]{} J., 1999, AJ, 118, 1766
Pignata G. et al., 2004, MNRAS, 355, 178
Pskovskii, Yu. P., 1984, Soviet Astronomy, 28, 658
Prieto J. L., Rest A., Suntzeff B. N., 2006, AJ, 647, 512.
A. G., Press W. H., Kirshner R. P., 1996, ApJ, 473, 88 R. P.
A. G. et al., 1998, AJ, 116, 1009
M., [Benetti]{} S., [Altavilla]{} G., [Pastorello]{} A., [Turatto]{} M., [Cappellaro]{} E., [O’Toole]{} S., 2002 IAUC, 7919, 1
Rothberg B., Joseph R. D., 2006, AJ, 131, 185
Rudy R. J., [Lynch]{} D. K., [Mazuk]{} S., [Venturini]{} C. C., [Puetter]{} R. C., [H[ö]{}flich]{}, P., 2002, ApJ, 565, 413
Ruiz-Lapuente P. & Lucy L. B., 1992, ApJ, 400, 127
Ruiz-Lapuente P., Kirshner R. P., Phillips M. M., Challis P. M., Schmidt B. P., Filippenko A. V., Wheeler, J. C., 1995, ApJ, 439, 60
Schlegel D. J., [Finkbeiner]{} D. P., [Davis]{} M., 1998, ApJ, 500, 525
F., [Seitzer]{} P., [Faber]{} S. M., [Burstein]{} D., [alle Ore]{} C. M., [Gonzalez]{} J. J., 1990, ApJL, 364, L33-L36
Shortridge K., Meatheringham S. J., Carter B. D., Ashley M. C. B., 1995, Astron. Soc. of Australia. Proceeding v.12:2, p. 244
Stanishev V. el al., 2007, A&A, 469, 645
, M., [Mazzali]{}, P. A., [Benetti]{}, S., [Hillebrandt]{}, W., 2005, MNRAS, 360, 1231
Stone R. P. S., Baldwin J. A., 1983 MNRAS, 204, 347
M. et al., 2002, AJ, 124, 2100.
Stritzinger M., Suntzeff N. B., Hamuy M., Challis P., Demarco R., Germany L., Soderberg A. M., 2005, PASP, 117, 810
M. & Sollerman J., 2007, A&A, 470, L1
Sullivan M. et al., 2006, ApJ, 648, 868
Suntzeff N. B., 1996, in R. McCray & Zhenru Wang eds. Proc. IAU Colloq. Supernovae and supernova remnants. Cambridge University Press, Cambridge, p. 41
, N. B. et al., 1999, AJ, 117, 1175
Tanaka .M et al., 2008, ApJ, 677, 448
Timmes F. X., Brown E. F., Truran J. W., 2003, ApJ, 590, L83
Tully R. B., 1988, Nearby Galaxies Catalog, Cambrige University Press.
Turatto M., Benetti S., Cappellaro E., 2003, in the proceedings to the ESO/MPA/MPE Workshop (an ESO Astrophysics Symposium) From Twilight to Highlight: The Physics of Supernovae, eds. B. Leibundgut and W. Hillebrandt Springer-Verlag, Heidelberg, p. 200
L., [Goldhaber]{} G., [Aldering]{} G., [Perlmutter]{} S., 2003, ApJ, 590, 944
X., [Wang]{} L., Xu Z., Yu-Qing L., Zongwei Li, 2005, AJ, 620, L87
L., Baade D., Hoeflich P., Wheeler J. C., Kawabata K., Khokhlov A., Nomoto K., Patat F., 2006, AJ, 653, 490
Wang X. et al., 2008a, ApJ, 675, 626
Wang X., Li W., Filippenko A. V.; Foley R. J., Smith N., Wang L., 2008, ApJ, 677, 1060
J. C., [Hoeflich]{} P.,[Harkness]{} R. P., [Spyromilio]{} J., 1998, ApJ, 496, 908
Wood-Vasey W. M. et al., 2007, ApJ, 666, 694
Woosley S. E., Kasen D., Blinnikov S., Sorokina E., 2007, ApJ, 662, 487
S-correction
============
To avoid the well known limitations of the colour equation to calibrate SN, we used the following calibration path to transform the SN instrumental magnitude into a standard photometric system. To simplify the explanation of the various steps, we will discuss as an example the calibration of the $B$ band and we will use only one of the local standard stars. First one needs to transform the standard magnitude of the sequence star in the natural photometric system of the instrument using the following formula:
$$B_{nat}^*=B^*-\gamma^B_{B-V}(B^*-V^*)
\label{equa1}$$
where $B^*$ and $V^*$ are the standard magnitude of the local standard star, while $\gamma^B_{B-V}$ is the colour term. Then one computes the instrument zero point in the natural photometric system $ZP(B)_{nat}$=$B_{nat}^*-b^*$, where $b^*$ is the local standard star instrumental magnitude.
Finally we calibrated the magnitude of the SN using the equation:
$$B^{SN}=ZP(B)_{nat}+b^{SN}+\delta_B
\label{equa2}$$
where $b^{SN}$ is the instrumental magnitude of the SN. We did not correct for atmospheric absorption because we assume that the SN and local standard stars are at the same airmass and hence the absorption is already included in $ZP(B)_{nat}$. We define $\delta_B$ as the difference between the synthetic magnitude of the SN computed using the standard and the natural system passbands.
$$\delta_B=B_{sy}^{SN}-b_{sy}^{SN}
\label{equa3}$$
with:
$$B_{sy}^{SN}=-2.5log\frac{\int_{-\infty}^{+\infty} N(\lambda)S_{st}^B(\lambda)d \lambda}{\int_{-\infty}^{+\infty}S_{st}^B(\lambda)d \lambda} + ZP(B)_{sy}
\label{equa4}$$
$$b_{sy}^{SN}=-2.5log\frac{\int_{-\infty}^{+\infty} N(\lambda)S_{nat}^B(\lambda)d \lambda}{\int_{-\infty}^{+\infty}S_{nat}^B(\lambda)d \lambda} + ZP(b)_{sy}
\label{equa5}$$
where $N(\lambda)$ is the photon number distribution of the SN, $S_{nat}^B(\lambda)$ and $S_{st}^B(\lambda)$ are the instrumental and standard $B$ passbands. In the equations \[equa4\] and \[equa5\], $N(\lambda)$ is used instead of $F(\lambda)$ because CCD and IR detector are photon counting devices. The @Bessell90 passbands were also reported in photon unit dividing the original ones for the wavelength, while the @Persson98 bands are already in photon units. Since in the spectra which are used to compute the synthetic magnitudes the flux is expressed in energy units, $N(\lambda)$ becomes:
$$N(\lambda)=\frac{F(\lambda)\lambda}{hc}
\label{equa9}$$
The zero points $ZP(B)_{sy}$ and $ZP(b)_{sy}$ were computed using a subset of spectrophotometric standard stars for which accurate photometry is available. In particular, for $ZP(B)_{sy}$ the standard magnitude was used directly, while for $ZP(b)_{sy}$ the magnitudes were previously reported in the instrument natural photometric system using equation (\[equa1\]). Putting together equations \[equa1\], \[equa2\] and \[equa3\] we obtain:
$$B^{SN}=B^*+b^{SN}-b^*-\gamma^B_{B-V}(B^*-V^*)+B_{sy}^{SN}-b_{sy}^{SN}
\label{equa8}$$
![Comparison of the different instrumental $UBVRI$ transmission curves normalized to the peak transmission with the standard Johnson Cousins functions [@Bessell90].[]{data-label="figA.1"}](bande_opt_2002dj.ps){width="84mm"}
![Comparison of the different instrumental $JHKs$ transmission curves normalized to the peak transmission with the standard Persson functions [@Persson98].[]{data-label="figA.2"}](bande_ir_2002dj.ps){width="84mm"}
----------- ------------------ -------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------------- -----------------
CTIO 0.9m 0.062(0.035)$^a$ 0.098(0.013) $-$0.092(0.007) $-$0.094(0.007) 0.018(0.001) 0.016(0.003) 0.010(0.001) $-$0.013(0.009) 0.004(0.001) 0.004(0.004)
ALFOSC 0.093(0.006)$^a$ 0.076(0.045) 0.032(0.003) 0.057(0.016) $-$0.046(0.002) $-$0.052(0.007) $-$0.065(0.002) $-$0.072(0.020) $-$0.041(0.002) $-$0.065(0.029)
DFOSC 0.058(0.006)$^a$ 0.029(0.016) 0.068(0.003) 0.089(0.004) 0.013(0.002) 0.013(0.007) 0.026(0.001) 0.009(0.015) $-$0.048(0.001) $-$0.060(0.036)
JKT 0.072(0.050) 0.026(0.019) 0.066(0.006) 0.085(0.032) 0.030(0.001) 0.028(0.035) 0.003(0.001) $-$0.039(0.023) 0.056(0.004) 0.018(0.030)
WFC 0.077(0.021) 0.059(0.032) 0.031(0.006) $-$0.010(0.001) 0.007(0.029) 0.013(0.001) 0.000(0.014) $-$0.197(0.004) $-$0.210(0.004)
FORS1 $-$0.083(0.007) $-$0.083(0.009) 0.040(0.002) 0.033(0.010) 0.080(0.002) 0.066(0.010) $-$0.034(0.004) $-$0.031(0.005)
EMMI 0.067(0.006) $-$0.045(0.005) 0.017(0.001) 0.056(0.023) 0.067(0.001) 0.044(0.009) $-$0.047(0.001)
----------- ------------------ -------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------------- -----------------
$^a$ Modified band.\
$^b$ Colour term computed using synthetic photometry.\
$^c$ Colour term computed using photometric standard.\
--------- ---------- -------------- ---------- ----------------- ------- -----------------
SoFi $-$0.020 $-$0.007 $-$0.022 0.006 0.006 0.023
ANDICAM 0.035 0.028(0.005) $-$0.009 $-$0.010(0.005) 0.000 $-$0.003(0.005)
--------- ---------- -------------- ---------- ----------------- ------- -----------------
date M.J.D. Phase$^a$ U B V R I Instr.
------------ --------- ----------- ------------------ ------------------ ------------------ ------------------ ------------------ -----------
13/06/2002 52439.0 -11.0 16.01 $\pm$ 0.03 16.02 $\pm$ 0.04 15.85 $\pm$ 0.03 15.64 $\pm$ 0.06 15.96 $\pm$ 0.03 CTIO 0.9m
14/06/2002 52439.9 -10.1 $-$ 15.68 $\pm$ 0.05 15.51 $\pm$ 0.02 15.25 $\pm$ 0.03 15.43 $\pm$ 0.03 EMMI
27/06/2002 52452.9 2.9 14.30 $\pm$ 0.04 14.35 $\pm$ 0.03 14.17 $\pm$ 0.02 14.09 $\pm$ 0.03 14.43 $\pm$ 0.05 WFC
28/06/2002 52453.9 3.9 $-$ 14.43 $\pm$ 0.03 14.17 $\pm$ 0.01 14.11 $\pm$ 0.03 14.47 $\pm$ 0.03 JKT
29/06/2002 52455.0 5.0 14.40 $\pm$ 0.04 14.47 $\pm$ 0.02 14.19 $\pm$ 0.02 14.14 $\pm$ 0.01 14.50 $\pm$ 0.04 JKT
30/06/2002 52456.0 6.0 $-$ $-$ $-$ 14.26 $\pm$ 0.03 $-$ JKT
02/07/2002 52458.0 8.0 $-$ 14.68 $\pm$ 0.02 14.31 $\pm$ 0.02 14.31 $\pm$ 0.03 14.64 $\pm$ 0.03 JKT
03/07/2002 52458.1 8.1 14.85 $\pm$ 0.05 14.68 $\pm$ 0.05 14.32 $\pm$ 0.03 14.29 $\pm$ 0.03 14.73 $\pm$ 0.05 DFOSC
04/07/2002 52460.0 10.0 15.01 $\pm$ 0.02 14.83 $\pm$ 0.04 14.42 $\pm$ 0.01 14.46 $\pm$ 0.04 14.96 $\pm$ 0.04 DFOSC
07/07/2002 52463.0 13.0 15.42 $\pm$ 0.03 15.16 $\pm$ 0.02 14.62 $\pm$ 0.03 14.61 $\pm$ 0.04 15.05 $\pm$ 0.04 DFOSC
09/07/2002 52465.0 15.0 15.75 $\pm$ 0.03 15.44 $\pm$ 0.04 14.79 $\pm$ 0.02 14.70 $\pm$ 0.03 15.07 $\pm$ 0.04 DFOSC
11/07/2002 52467.0 17.0 16.05 $\pm$ 0.01 15.61 $\pm$ 0.04 14.88 $\pm$ 0.04 14.74 $\pm$ 0.05 15.04 $\pm$ 0.05 DFOSC
16/07/2002 52473.0 23.0 $-$ $-$ 15.14 $\pm$ 0.03 14.84 $\pm$ 0.03 14.83 $\pm$ 0.02 CTIO 0.9m
16/07/2002 52473.1 23.1 16.75 $\pm$ 0.04 16.33 $\pm$ 0.04 15.17 $\pm$ 0.03 14.79 $\pm$ 0.05 14.99 $\pm$ 0.02 DFOSC
19/07/2002 52476.0 26.0 16.97 $\pm$ 0.05 16.57 $\pm$ 0.04 15.42 $\pm$ 0.03 15.03 $\pm$ 0.03 14.84 $\pm$ 0.04 DFOSC
24/07/2002 52479.9 29.9 $-$ 16.92 $\pm$ 0.05 15.54 $\pm$ 0.05 15.12 $\pm$ 0.03 14.85 $\pm$ 0.01 ALFOSC
25/07/2002 52481.0 31.0 17.41 $\pm$ 0.03 16.97 $\pm$ 0.02 15.63 $\pm$ 0.03 15.20 $\pm$ 0.03 14.90 $\pm$ 0.04 DFOSC
29/07/2002 52484.9 34.9 17.55 $\pm$ 0.17 $-$ $-$ $-$ $-$ ALFOSC
08/08/2002 52495.0 45.0 $-$ 17.33 $\pm$ 0.04 16.22 $\pm$ 0.04 15.91 $\pm$ 0.03 15.74 $\pm$ 0.04 CTIO 0.9m
31/08/2002 52518.0 68.0 18.38 $\pm$ 0.12 $-$ 16.89 $\pm$ 0.04 $-$ $-$ DFOSC
02/09/2002 52520.0 70.0 $-$ $-$ 16.91 $\pm$ 0.05 $-$ $-$ DFOSC
06/09/2002 52524.0 74.0 $-$ 17.76 $\pm$ 0.19 $-$ 17.05 $\pm$ 0.16 16.88 $\pm$ 0.48 DFOSC
25/03/2003 52724.3 274.3 $-$ 21.00 $\pm$ 0.04 20.78 $\pm$ 0.03 21.30 $\pm$ 0.08 20.74 $\pm$ 0.10 FORS1
$^a$ Relative to the time of the $B$ maximum brightness M.J.D.=52450 $\pm$ 0.7\
date M.J.D. Phase$^a$ J H K Instr.
------------ --------- ----------- ------------------ ------------------ ------------------ ---------
13/06/2002 52439.0 -11.0 15.64 $\pm$ 0.05 15.64 $\pm$ 0.08 15.66 $\pm$ 0.10 ANDICAM
14/06/2002 52440.0 -10.0 15.27 $\pm$ 0.03 15.31 $\pm$ 0.05 15.25 $\pm$ 0.08 ANDICAM
15/06/2002 52441.1 -8.9 14.98 $\pm$ 0.03 15.13 $\pm$ 0.03 15.09 $\pm$ 0.03 SoFi
19/06/2002 52444.1 -5.9 14.63 $\pm$ 0.03 14.85 $\pm$ 0.03 14.70 $\pm$ 0.03 SoFi
20/06/2002 52446.0 -4.0 14.56 $\pm$ 0.09 14.82 $\pm$ 0.06 14.64 $\pm$ 0.06 ANDICAM
24/06/2002 52450.0 -0.0 14.60 $\pm$ 0.03 14.85 $\pm$ 0.05 14.50 $\pm$ 0.09 ANDICAM
24/06/2002 52450.0 -0.0 14.62 $\pm$ 0.03 14.83 $\pm$ 0.03 14.54 $\pm$ 0.03 SoFi
27/06/2002 52453.1 3.1 14.88 $\pm$ 0.04 14.93 $\pm$ 0.04 14.65 $\pm$ 0.07 ANDICAM
30/06/2002 52456.0 6.0 15.33 $\pm$ 0.12 15.12 $\pm$ 0.08 14.85 $\pm$ 0.16 ANDICAM
08/07/2002 52464.0 14.0 16.38 $\pm$ 0.07 15.09 $\pm$ 0.04 15.03 $\pm$ 0.08 ANDICAM
11/07/2002 52467.0 17.0 16.36 $\pm$ 0.05 15.01 $\pm$ 0.03 14.90 $\pm$ 0.05 ANDICAM
11/07/2002 52467.0 17.0 16.15 $\pm$ 0.03 14.90 $\pm$ 0.03 14.87 $\pm$ 0.03 SoFi
14/07/2002 52470.0 20.0 16.19 $\pm$ 0.10 14.85 $\pm$ 0.04 14.76 $\pm$ 0.06 ANDICAM
17/07/2002 52473.0 23.0 15.97 $\pm$ 0.04 14.79 $\pm$ 0.03 14.66 $\pm$ 0.04 ANDICAM
25/07/2002 52481.0 31.0 15.70 $\pm$ 0.05 14.81 $\pm$ 0.04 14.83 $\pm$ 0.08 ANDICAM
28/07/2002 52484.0 34.0 15.75 $\pm$ 0.04 15.07 $\pm$ 0.04 15.13 $\pm$ 0.05 ANDICAM
31/07/2002 52487.0 37.0 16.01 $\pm$ 0.03 15.25 $\pm$ 0.06 15.45 $\pm$ 0.28 ANDICAM
07/08/2002 52494.0 44.0 16.70 $\pm$ 0.09 15.62 $\pm$ 0.07 15.67 $\pm$ 0.35 ANDICAM
10/08/2002 52497.0 47.0 16.79 $\pm$ 0.09 15.74 $\pm$ 0.05 15.94 $\pm$ 0.12 ANDICAM
13/08/2002 52500.0 50.0 16.99 $\pm$ 0.11 15.86 $\pm$ 0.05 16.05 $\pm$ 0.12 ANDICAM
30/08/2002 52517.0 67.0 18.07 $\pm$ 0.38 16.61 $\pm$ 0.03 16.87 $\pm$ 0.03 SoFi
$^a$ Relative to the time of the $B$ maximum brightness M.J.D.=52450 $\pm$ 0.7\
Band construction
-----------------
We define the transmission function of a give passband $S(\lambda)$ as: $S(\lambda)=F(\lambda)\cdot QE(\lambda)
\cdot Ac(\lambda) \cdot Al(\lambda) \cdot M(\lambda) \cdot N_{ref} \cdot
L(\lambda)$\
where $F(\lambda)$ is the filter transmission function, $QE(\lambda)$ is the detector quantum efficiency, $Ac(\lambda)$ is the continuum atmospheric transmission profile, $Al(\lambda)$ is the line atmospheric transmission profile, $M(\lambda)$ is the mirror reflectivity function, $N_{ref}$ the number of reflections and $L(\lambda)$ is the lens throughput. We downloaded $F(\lambda)$ and $QE(\lambda)$ from the instrument web sites. For $M(\lambda)$ we used a standard aluminium reflectivity curve, while we did not find information about the lens transmission for any of the instruments. In the construction of the atmosphere model the weight of the terms $Ac(\lambda)$ and $ Al(\lambda)$ change considerably going from $U$ to $Ks$ bands. In the optical the atmosphere transmission is mainly defined by a continuum function, while in the IR in most of the instruments the passband cut-offs are defined by deep atmospheric absorption bands. The latter make the use of a fixed instrumental passband not a good approximation of the real transmission function, because it depends on the atmospheric conditions. To a much lesser degree this is also a problem for the optical $R$ and $I$ filters, where some telluric features fall in the middle of the passbands. These telluric lines do not change linearly with airmass, therefore their variation is not properly accounted by the absorption coefficients. For La Palma we obtained $Ac(\lambda)$ from @atm_lapalma, while for La Silla, Paranal and CTIO we have used the CTIO transmission curve in the IRAF reduction package (Stone & Baldwin 1983, Baldwin & Stone 1984). For the IR sky spectrum of La Silla and CTIO we downloaded $Ac(\lambda)
\cdot Al(\lambda)$ from the 2MASS web side[^5]. The instrumental filter bands are compared with the standard ones in Fig. \[figA.1\] and Fig. \[figA.2\].\
As usual (see e.g. Pignata et al. 2004), we checked how well the constructed instrumental passbands match the real ones and compared the colour terms derived from standard stars (@Landolt92 for the optical and @Persson98 for the IR) with those we computed by integrating the flux of spectrophotometric standards of @Stritzinger05 for the optical and Sirius, Vega and the Sun for the IR. We decided to modify the instrumental passband only in cases where the difference between the two colour terms reported in Table \[tabA.1\] and Table \[tabA.2\] were larger than three times the error associated to both “synthetic” and “photometric” colour terms. The lack of a suitable number of spectrophotometric IR standards with an establish magnitude make the colour term comparison less effective than at optical wavelengths.
[^1]: E-mail: [email protected]
[^2]: IRAF is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc, under contract to the National Science Foundation.
[^3]: Although SN 2002er was classified as HVG SNe [@Benetti05] it was not included in the table because of its lower expansion velocities. This SN could indeed represents a transition object between HVG SNe and LVG SNe [@Tanaka08].
[^4]: The light curves of SN 2002bf have been K-corrected to the host galaxy rest frame velocity of SN 2002dj
[^5]:
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We say $(W, \{\phi_1, \cdots, \phi_t\})$ is a polarizable dynamical system of several morphisms if $\phi_i$ are endomorphisms on a projective variety $W$ such that $\bigotimes \phi_i^*L$ is linearly equivalent to $L^{\bigotimes q}$ for some ample line bundle $L$ on $W$ and for some $q>t$. If $q$ is a rational number, then we have the equidistribution of small points of given dynamical system because of Yuan’s work [@Y]. As its application, we can build a polarizable dynamical system of an automorphism and its inverse on $K3$ surface and show its periodic points are equidistributed.'
address: 'Department of Mathematics, University of Illinois at Chicago, Chicago IL 60607, US'
author:
- Chong Gyu Lee
title: |
Equidistribution of periodic points of\
some automorphisms on $K3$ surface
---
Introduction
============
The study of algebraic dynamics blooms after Northcott proved the arithmetic property of dynamical system of a morphism on projective space. Szpiro, Ullmo & Zhang [@SUZ] started one direction of algebraic dynamics, the equidistribution of small points. After various research of Bilu [@B] on some variety with group structure and of Baker & Rumely [@BR], Chambert-loir [@Ch] and Favre & Rivera-Letelier [@FR] on the equidistribution of dynamical system on dimension $1$, Yuan [@Y] proved the general equidistribution theorem: let $\phi$ be a polarizable endomorphisms. Then, we have an ample line bundle $\mathcal{L}$ with semipositive dynamical metric $||\cdot||_\phi$ defined by Zhang [@Z1], then we have the equidistribution of the small points with respect to the height function corresponding to $\overline{\mathcal{L}}= (\mathcal{L},||\cdot||_\phi)$.
For the dynamical equidistribution, “polarizable” condition is very important because it guarantees that we can define a sequence of metric defined on the same line bundle. If $\phi$ is not polarizable, then, metrics ${\phi^k}^*||\cdot||^{\frac{1}{q^k}}$ may be defined on different line bundles for each $k$ so that “convergence” of give sequence of metrics doesn’t make sense.
Still, we have hope because of Kawaguchi’s idea. He [@K0] suggested the polarizable dynamical system of several morphisms:
Let $W$ be a projective variety, let $L$ be an ample line bundle and let $M=\{ \phi_1, \cdots, \phi_t:W \rightarrow W\}$ be a finite set of morphisms. We say that a dynamical system of several morphism $(W, M)$ is *polarizable* if $$\bigotimes_{i=1}^t \phi_i^*L \sim L^{\otimes q}$$ for some rational number $q>t$.
His idea makes a way to study the dynamics of some automorphisms. In general, an automorphisms on projective variety is not polarizable in general. However, we have a good counter part, the inverse map. The existence of the inverse map makes a dynamical system of a automorphism better; all preperiodic points of an automorphisms $\sigma$ is actually periodic, and $\sigma$ and $\sigma^{-1}$ shares that same periodic point. Thus, we can consider a dynamical system of an automorphism $\sigma$ is considered as a dynamical system of $\{\sigma, \sigma^{-1}\}$. Furthermore, a dynamical system of several morphisms $(W, M=\{\phi_1, \cdots \phi_t\})$ actually works with the monoid generated by $M$. If $M$ consists of an automorphism $\sigma$ and its inverse, the monoid $\mathcal{M}$ generated by $M$ is exactly $\{ \sigma^k ~|~ k \in \mathbb{Z}\}$ and hence $\mathcal{M}$-preperiodic points is essentially ‘$\sigma$-preperiodic or $\sigma^{-1}$-prepriodic points. In Section 2, we have examples of polarizable dynamical systems of an automorphism on $K3$ surfaces.
The main purpose of this paper is to confirm that we have the dynamical equidistribution for dynamical systems of several morphisms and to apply this result on some automorphisms on $K3$ surface. In section 3, we will combine Kawaguchi’s and Yuan’s results to prove the equidistribution of small points:
Let $W$ be a projective variety of dimension $n$ over a number field $K$, let $\mathcal{L}$ be an ample line bundle and let $M= \{ \phi_1, \cdots, \phi_t\}$ be a finite set of endomorphisms on $W$. Suppose that $(W,M)$ is a polarizable with some integer $q>t$ and $\{ x_m\}$ be a generic and small sequence. Then, the a sequence of probability measure on the Galois orbit of $x_m$ weakly converges to the dynamical measure at every place $v$: $$\dfrac{1}{\deg x_m} \sum_{y\in \Gamma_{x_m}} \delta_y \rightarrow \mu_{M,v}$$ where $\Gamma_{x_m}$ is the Galois orbit or $x_m$ and $\mu_{M,v} = \dfrac{c_1(\mathcal{L})^n_v}{\deg_L W}$ is the dynamical probability measure of the dynamical system $(W,M)$ on the analytic space $W_{K_v}^{an}$.
In Section 5, we will show that we can find a generic and small sequence of periodic points. Thus, we can find some properties of the set pr periodic points of some automorphisms on $K3$ surfaces.
Let $W$ be a projective variety defined over a number field $K$, let $M = \{\sigma, \sigma^{-1}\}$ be an automorphism and its inverse on $W$. Suppose that $(W,M)$ is polarizable with some integer $q>2$. Then, ${\operatorname{Per}}(\sigma)$ is Zariski dense.
*Acknowledgements*. The author would like to thank Xinyi Yuan for helpful discussions and comments for paper, thank Joseph H. Silverman for suggesting ideal for proof of Theorem \[dense\]. Also thanks to Shu Kawaguchi and Jordan Ellenberg for useful comments.
Polarizable dynamical systems of automorphisms on $K3$ surfaces
===============================================================
We have lots of interesting examples of the polarizable dynamical system of several morphisms on $K3$ surface.
$K3$ surface with two involutions, I
------------------------------------
The space of $K3$ surfaces is a $19$-dimensional object up to isomorphism. And, a family of $K3$ surface in ${\mathbb{P}}^2\times {\mathbb{P}}^2$ defined by an intersection of hypersurfaces of bidegree $(1,1)$ and $(2,2)$ is $18$-parameter family of isomorphism classes of nonsingular surfaces. For details of such $K3$ surfaces, refer [@S2 §7.4].
\[K3I\] Let $S ={\mathbb{P}}^2\times {\mathbb{P}}^2$ be a $K3$ surface defined by an intersection of hypersurfaces of bidegree $(1,1)$ and $(2,2)$ with two involutions $\iota_1, \iota_2$, let $\pi_i$ be the projection map onto $i$-th component and let $L_i = \pi_i^*{\mathcal{O}}_{{\mathbb{P}}^2}(1)$. Then, we have $$\iota_i^*L_i = L_i, \quad \iota_i^*L_j = L_i^{\otimes 4} \otimes L_j^{\otimes-1}$$ and hence $$\iota_1^*L \otimes \iota_2^*L = L^{\otimes 4}$$ where $L =L_1 \otimes L_2$ is an ample line bundle. Therefore, $(S, \{ \iota_1, \iota_2 \})$ is a polarizable dynamical system.
\[K3IB\] Let $S$ be the $K3$ surface defined on Example \[K3I\]. Define $\sigma_1 = \iota_2 \circ \iota_1$ and $\sigma_2 = \iota_1 \circ \iota_2 = \sigma_1^{-1}$. Then, $(S, \{\sigma_1, \sigma_2\})$ is a polarizable dynamical system: we have $$\sigma_i^*L_i = L_i^{\otimes -1} \otimes L_j^{\otimes 4}, \quad \sigma_i^*L_j = L_i^{\otimes -4} \otimes L_j^{\otimes 15}$$ and $$\sigma_1^*L \otimes \sigma_2^*L = L^{\otimes 14}.$$
$K3$ surface with two involutions, II
-------------------------------------
There is another way of defining other automorphisms on $K3$ surface in ${\mathbb{P}}^2 \times {\mathbb{P}}^2$. With same method, we can calculate the number of parameters of defining equations $$\sum_{0\geq i \geq j \geq 2}\sum_{0\geq k \geq 2} A_{ijk} x_{i} x_j y_k \quad \sum_{0\geq l \geq 2}\sum_{0\geq m \geq n \geq 2}
B_{mmn} x_{l} y_m y_n$$ is $(18-1)+(18-1)$ and the dimension of ${\operatorname{PGL}}_3$, the isometry group of each ${\mathbb{P}}^2$, is 8. Hence the dimension of the family of $K3$ such surfaces is $18$ again.
\[K3II\] Let $S ={\mathbb{P}}^2 \times {\mathbb{P}}^2$ be a $K3$ surface generated by intersecting two hypersurfaces of bidegree $(1,2)$ and $(2,1)$ with two involutions $\iota_1, \iota_2$. Let $\pi_i$ be the projection map onto $i$-th component and $L_i = \pi_i^*{\mathcal{O}}_{{\mathbb{P}}^2}(1)$. Then, we have $$\iota_i^*L_i = L_i, \quad \iota_i^*L_j = L_i^{\otimes -1} \otimes L_j^{\otimes 5}.$$ Since $L =L_1 \otimes L_2$ is ample, we get $$\iota_1^*L \otimes \iota_2^*L = L^{\otimes 5}$$ and hence get a polarizable dynamical system.
\[K3IIB\] Let $S$ be the $K3$ surface defined on Example \[K3II\]. Define $\sigma_1 = \iota_2 \circ \iota_1$ and $\sigma_2 = \iota_1 \circ \iota_2=\sigma_1^{-1}$. Then, $(S, \sigma_1, \sigma_2)$ is a rational polarizable dynamical system: we have $$\sigma_i^*L_i = L_i^{\otimes -1} \otimes L_j^{\otimes 5}, \quad \sigma_i^*L_j = L_i^{\otimes -5} \otimes L_j^{\otimes 24}$$ and $$\sigma_1^*L \otimes \sigma_2^*L = L^{\otimes 23}$$ for all $L \in \langle L_1, L_2 \rangle$.
$K3$ surface with three involutions
-----------------------------------
If we define a $K3$ surface in ${\mathbb{P}}_1 \times {\mathbb{P}}^1 \times {\mathbb{P}}^1$, then Néron-Severi group is of rank $3$. Thus we expect that the dimension of the family of $K3$ surface is reduced by $1$; the number of parameters for defining equations $$\sum_{0\geq i \geq j \geq 2}\sum_{0\geq k \geq l\geq 2}\sum_{0\geq m \geq n \geq 2} A_{ijklmmn}I x_{i}xji y_k y_l z_m z_n$$ is $27-1$ and the dimension of ${\operatorname{PGL}}_2$, the isometry group of each ${\mathbb{P}}^1$, is 3. Hence the dimension of the family of $K3$ such surfaces is $26 - 3 - 3 - 3 = 17$.
\[K3III\] Let $S = {\mathbb{P}}^1 \times {\mathbb{P}}^1 \times {\mathbb{P}}^1$ be $K3$ surface, a hypersurface of bidegree $(2,2,2)$ with three involutions $\iota_1, \iota_2, \iota_3$. Let $\pi_i$ be the projection map onto $i$-th component and $L_i = \pi_i^*{\mathcal{O}}_{{\mathbb{P}}^1}(1)$. Then, we have $$\iota_i^*L_j = L_j~\text{for}~i\neq j, \quad \iota_i^*L_i = L_i^{\otimes -1} \otimes L_j^{\otimes2} \otimes L_k^{\otimes2}.$$ Hence, let $L =L_1 \otimes L_2 \otimes L_3$ and get $$\iota_1^*L \otimes \iota_2^*L \otimes \iota_3^*L = L^{\otimes 5}.$$ So, $(S, \{\iota_1, \iota_2, \iota_3\})$ is polarizable.
\[K3IIIB\] Let $S$ be the $K3$ surface defined on Example \[K3III\]. Consider $\tau_1 = \iota_3\circ \iota_2 \circ \iota_1$, $\tau_2 = \tau_1^{-1}$. Then, $$\begin{array}{c@{~=~}l@{~=~}l@{~=~}l}
\tau_1^*L_1 & \iota_3^* \iota_2 (L_1^{\otimes -1} \otimes L_2^{\otimes 2} \otimes L_3^{\otimes 2}) &
\iota_3^*(L_1^{\otimes 3} \otimes L_2^{\otimes -2} \otimes L_3^{\otimes 6}) &
L_1^{\otimes 15} \otimes L_2^{\otimes 10} \otimes L_3^{\otimes -6} \\
\tau_1^*L_2 & \iota_3^*\iota_2^* (L_2) & \iota_3^*(L_1^{\otimes 2} \otimes L_2^{\otimes -1} \otimes L_3^{\otimes 2}) &
L_1^{\otimes 6} \otimes L_2^{\otimes 3} \otimes L_3^{\otimes -2} \\
\tau_1^*L_3 & \iota_3^*\iota_2^* (L_3) & \iota_3^* L_3 & L_1^{\otimes 2} \otimes L_2^{\otimes 2} \otimes L_3^{\otimes -1}
\end{array}$$ $$\begin{array}{c@{~=~}l@{~=~}l@{~=~}l}
\tau_2^*L_1 & \iota_1^*\iota_2^* (L_1) & \iota_1^* L_1 & L_1^{\otimes -1} \otimes L_2^{\otimes 2} \otimes L_3^{\otimes 2}\\
\tau_2^*L_2 & \iota_1^*\iota_2^* (L_2) & \iota_1^*(L_1^{\otimes 2} \otimes L_2^{\otimes -1} \otimes L_3^{\otimes 2}) &
L_1^{\otimes -2} \otimes L_2^{\otimes 3} \otimes L_3^{\otimes 6} \\
\tau_2^*L_3 & \iota_1^* \iota_2 (L_1^{\otimes 2} \otimes L_2^{\otimes 2} \otimes L_3^{\otimes -1}) &
\iota_1^*(L_1^{\otimes 3} \otimes L_2^{\otimes -2} \otimes L_3^{\otimes 6}) & L_1^{\otimes -6} \otimes L_2^{\otimes 10} \otimes L_3^{\otimes 15} \\
\end{array}$$
$$\begin{aligned}
[\tau_1^*\otimes \tau_2^*](L_1^{\otimes a} \otimes L_2^{\otimes b} \otimes L_3^{\otimes c})
&=& (L_1^{\otimes 15a+6b+2c} \otimes L_2^{\otimes 10a+3b+2c} \otimes L_3^{\otimes -6a-2b-c}) \\
& & \otimes (L_1^{\otimes -a-2b-6c} \otimes L_2^{\otimes 2a+3b+ 10c } \otimes L_3^{\otimes 2a+6b+15c}) \\
&=& L_1^{\otimes 14a+4b-4c} \otimes L_2^{\otimes 12a+6b+12c} \otimes L_3^{\otimes -4a+4b+14c}
\end{aligned}$$
Therefore, let $L = L_1 \otimes L_2^{\otimes 2} \otimes L_3$. Then, $$\tau_1^*L \otimes \tau_2^*L \sim L^{\otimes 18}$$ and hence $(S, \tau_1, \tau_1^{-1})$ is polarizable. More precisely, let $L_{\alpha, \beta} = (L_1 \otimes L_2)^{\otimes \alpha} \otimes (L_1^{-1} \otimes L_3)^{\otimes \beta}$, then $$\tau_1^*L_{\alpha, \beta} \otimes \tau_2^*L_{\alpha, \beta} \sim L_{\alpha, \beta}^{\otimes 18}.$$
Similarly, automorphisms $\tau' = \iota_1\circ \iota_3 \circ \iota_2$, $\tau''= \iota_2 \circ \iota_1 \circ \iota_3$ with their inverses will generate polarizable dynamical systems respectively.
\[ce\] Consider the following case; let $S$ be a $K3$ surface defined on Example \[K3III\]. Define $\sigma_1 = \iota_2 \circ \iota_1$ and $\sigma_2 = \iota_1 \circ \iota_2 = \sigma_1^{-1}$. Then, $(S, \sigma_1, \sigma_2)$ is a polarizable dynamical system: $$\begin{array}{c@{~=~}l@{~=~}l}
\sigma_1^*L_1 & \iota_2^* (L_1^{\otimes -1} \otimes L_2^{\otimes 2} \otimes L_3^{\otimes 2}) &
L_1^{\otimes 3} \otimes L_2^{\otimes -2} \otimes L_3^{\otimes 6} \\
\sigma_1^*L_2 & \iota_2^* (L_2) & L_1^{\otimes 2} \otimes L_2^{\otimes -1} \otimes L_3^{\otimes 2} \\
\sigma_1^*L_3 & \iota_2^* (L_3) & L_3
\end{array}$$ Therefore, $$\begin{aligned}
[\sigma_1^*\otimes \sigma_2^*](L_1^{\otimes a} \otimes L_2^{\otimes b} \otimes L_3^{\otimes c})
&=& (L_1^{\otimes 3a+2b} \otimes L_2^{\otimes -2a-b} \otimes L_3^{\otimes 6a+2b+c}) \\
& & \otimes (L_1^{\otimes -a-2b} \otimes L_2^{\otimes 2a+3b } \otimes L_3^{\otimes 2a+6b+c}) \\
&=& L_1^{\otimes 2a} \otimes L_2^{\otimes 2b} \otimes L_3^{\otimes 8a+8b+2c}
\end{aligned}$$ Therefore, $L_3$ is the only combination of $L_1, L_2$ and $L_3$ which makes linear equivalence; $$\sigma_1^*L_3 \otimes \sigma_2^* L_3 \sim L_3^{\otimes 2}$$ and hence $(S, \{\sigma_1, \sigma_2\})$ is not a polarizable dynamical system in Kawaguchi’s sense. Similarly, $(S, \{\iota_1\circ \iota_3, \iota_3\circ \iota_1\})$, $(S, \{\iota_2\circ \iota_3, \iota_3\circ \iota_2\})$ are not polarizable.
$K3$ surface with three involutions, of the Picard number $4$.
--------------------------------------------------------------
Let $S = {\mathbb{P}}^1 \times {\mathbb{P}}^1 \times {\mathbb{P}}^1$ be $K3$ surface, a hypersurface of bidegree $(2,2,2)$ of the Picard number $4$. Then, we have another involution $\iota_4$ of order $2$, which is a group inverse at each elliptic curve fibers of $S$. Then, ${\operatorname{Pic}}(S) = \langle L_1, L_2, L_3, L_4 \rangle$ where $L_4$ corresponds to $-2$-curve class containing $(x,0,0)$. (See [@BM] for details.) Define an automorphisms $\tau = \iota_1 \circ \iota_2 \circ \iota_4$ and $\tau^{-1} = \iota_4 \circ \iota_2 \circ \iota_1$. Then, $$\iota_4^*L_1 = L_1, \iota_4^*L_4 = L_4, \quad \iota_4^*L_j = L_1^{\otimes 8} \otimes L_i^{-1} \otimes L_4.$$ Therefore, $$\tau^*L \otimes {\tau^{-1}}^* L = L^{\otimes 30}$$ if $$L = (L_1^{\otimes 4} \otimes L_2^{\otimes 4} \otimes L_3^{\otimes 3})^{\otimes \alpha} \otimes
(L_1^{\otimes 8} \otimes L_2^{\otimes 2} \otimes L_4^{\otimes 3})^{\otimes \beta}$$ for some $\alpha, \beta$. Similarly, $\tau' = \iota_1 \circ \iota_3 \circ \iota_4$ with its inverse will generate a polarizable dynamical system. But, $\zeta = \iota_2 \circ \iota_3 \circ \iota_4$ or $\zeta' = \iota_3 \circ \iota_2 \circ \iota_4$ only polarized by $q=2$ or $q=-2$. Also, $\eta = \iota_1 \circ \iota_4\circ \iota_2$, $\eta' = \iota_1 \circ \iota_4 \circ \iota_2$ is not polarized by any ample line bundles.
Dynamical Equidistribution on polarizable dynamical systems of several morphisms
================================================================================
The equidistribution of small points for polarizable morphisms is almost proved. Kawaguchi proved that the dynamical system of several morphisms generates the dynamical adelic metric which is semipositive. Thus, the equidistribution of small points of a polarizable dynamical system is an easy consequence of Yuan’s results. In this section, we will briefly check Kawaguchi’s and Yuan’s results to confirm the equidistribution theory for dynamical system of several morphisms.
Let $W$ be a projective variety defined over a number field $K$, let $\overline{L} = (L,||\cdot||)$ be an adelic ample line bundle on $W$, an ample line bundle $L$ with a semipositive adelic metric$||\cdot||$. Then we define *a height of subvarieties corresponding to $\overline{L}$* to be $${h}_{\overline{L}} (Y) := \dfrac{c_1({\overline{L}})^{d+1}}{(d+1){\operatorname{ord}}_{\overline{L}} Y},$$ where $Y$ is a subvariety of $W$ of dimension $d$ and $c_1$ is the curvature form.
Let $W$ be a projective variety, let $\overline{L}$ be an adelic ample line bundle and let ${h}_{\overline{L}}$ be the height function for closed subvarieties of $W$ corresponding line bundle $\overline{L}$. Suppose $\{x_m\}$ is a sequence of points. Then we say *$\{x_m\}$ is generic* if any infinite subsequence is not contained in a closed subvariety. We say*$\{x_m\}$ is small* if ${h}_{\overline{L}}(x_m)$ converges to ${h}_{\overline{L}}(W)$.
\[yuan\] Suppose that $W$ is a projective variety of dimension $n$ over a number field $K$, and $\overline{L}$ is a metrized line bundle over $W$with semipositive adelic metric. Let $\{x_m\}$ be an infinite sequence of closed point in $W$ which is generic and small with respect to $h_{\overline{L}}$. Then for any place $v$ of $K$, the Galois orbit of sequence $\{x_m\}$ are equidistributed in the analytic space $W^{an}_{K_{v}}$ with respect to the canonical measure $d\mu_{v} = c_1(\overline{L})^n_v / \deg_{\overline{L}} W$: $$\dfrac{1}{\deg x_m} \sum_{y \in \Gamma x_m} \delta_y ~\text{weakly converges to}~ d\mu_{v}.$$
In Theorem \[yuan\], we should assume that $L$ is $\mathbb{Q}$-divisor. Actually we will use the integral model $(\mathcal{W}, \mathcal{L})$ of $(W,L^e)$ where $\mathcal{L}$ on the generic fiber $\mathcal{W}_{\mathbb{Q}}$ is $L^e$. Thus, $L \in {\operatorname{Pic}}(W)\otimes \mathbb{Q}$. Using $\mathcal{L}$ instead of $L$ implies this fact.
\[kawa\] Let $W$ be a projective variety defined over a number field $K$, Let $L$ be an ample line bundle on $W$, and let $M = \{\phi_1, \cdots \phi_t\}$ be a set of endomorphisms on $W$. Suppose that $(W,M)$ is polarizable with respect to $L$: $$\bigotimes_{i=1}^t \phi_i^*L = L^{\otimes q}$$ where $q>t$. Then,
1. There is a unique continuous metric $||\cdot ||_M$, called the admissible metric on $L$ with $||\cdot ||^q_M = \tau^*(\phi_1^*|| \cdot ||^q_M \cdots \phi_t^* ||\cdot ||^q_M)$ where $\tau : L^{\otimes q} \rightarrow \bigotimes \phi_i^*L$ is an isomorphism.
2. Let $\overline{L}= (L, ||\cdot||_M)$ be the line bundle with admissible metric. Then, there exists a unique real-valued function $$\widehat{h}_{\overline{L}} : W(\overline{K}) \rightarrow \mathbb{R}$$ with the following properties:
1. $\widehat{h}_{\overline{L}}$ is a Weil height corresponding to $L$.
2. $\displaystyle \sum_{i=1}^k \widehat{h}_{\overline{L}} \bigl( \phi_i(x)\bigr) = q \cdot \widehat{h}_{\overline{L}} (x)$ for all $x \in W(\overline{K})$.
3. $\widehat{h}_{\overline{L}} \geq 0$ for all $x\in W(\overline{K})$.
The condition $q>t$ is necessary because the number of $N$-combinations of $\phi_i$’s in $M$ is $t^N$ while the growth rate of height is $q^N$. More precisely, the canonical height, if exists, defined by several morphism is of the form $$\lim_{N\rightarrow \infty} \dfrac{1}{q^N} \sum_{F \in \mathcal{M}_N} h_{\overline{L}} \bigl( F(P) \bigr).$$ Therefore, if $q\leq k$, then it may not shrink at prepriodic points. For example, if $q=k$ and $P$ is a common fixed point of $M$ with nontrivial height value, then $$\lim_{N\rightarrow \infty} \dfrac{1}{q^N} \sum_{F \in \mathcal{M}_N} h_{\overline{L}}\bigl( F(P) \bigr) = h_{\overline{L}}(P)$$ so that $P$ may not be a root of the canonical height. Thus, even we can build a semipositive metric, it is not compatible with the original dynamical system.
Now, combining previous two Lemmas, we get equidistribution of small points for the polarizable dynamical system of several morphisms:
Let $W$ be a projective variety of dimension $n$ over a number field $K$, let $L$ be an ample line bundle and let $M= \{ \phi_1, \cdots, \phi_t\}$ be a finite set of endomorphisms on $W$. Suppose that $(W,M)$ is a polarizable with some integer $q>t$ and $\{ x_m\}$ be a generic and small sequence. Then, the a sequence of probability measure on the Galois orbit of $x_m$ weakly converges to the dynamical measure at every place $v$: $$\dfrac{1}{\deg x_m} \sum_{y\in \Gamma_{x_m}} \delta_y \rightarrow \mu_{M,v}$$ where $\Gamma_{x_m}$ is the Galois orbit of $x_m$ and $\mu_{M,v} = \dfrac{c_1(L)^n_v}{\deg_L W}$ is the probability $M$-invariant measure on the analytic space $W_{K_v}^{an}$.
Theorem \[kawa\] says that we have the dynamical adelic metric on $L$. Such metric is semipositive because it is defeind by the limit of positive metrics. Since we assume that $q$ is rational number, we may assume that $\mathcal{L}$ is $\mathbb{Q}$ divisor and hence we can apply Theorem \[yuan\], to get the desired result.
Periodic points of automorphisms
================================
In previous section, we have the equidistribution of small points for the polarizable dynamical systems of several morphisms. To show the periodic points are equidistributed, we can make the generic sequence and hence we should show that the set of periodic points is Zariski dense. We can prove it in two ways: Fakhruddin [@Fa] introduce algebro-geometric proof by Poonen. I will introduce the arithmetic proof.
\[induction\] Let $\sigma :W \rightarrow W$ be an automorphism on a surface $W$ such that $(W, \{ \sigma, \sigma^{-1}\})$ is polarizable. Then, $(W, \{ \sigma^m, \sigma^{-m}\})$ is also polarizable for all $m \in \mathbb{Z}$.
Let $$\sigma^*L \otimes {\sigma^{-1}}^*L = L^{\otimes q}\quad \text{where}~ q>2.$$ Then, $${\sigma^2}^*L \otimes {\sigma^{-2}}^*L = L^{\otimes q^2-2}$$ By induction, suppose $$L_m = {\sigma^l}^*m \otimes {\sigma^{-l}}^*L = L^{\otimes q_m}\quad \text{where}~ q_m>2$$ and $q_{m} - q_{m-1}>2$ holds for $m=m_0-1, m_0$. Then, Then, $${\sigma}^*L_{m_0} \otimes {\sigma^{-1}}^*L_{m_0}
= L^{\otimes q \cdot q_{m_0}}.$$ On the other hand, $${\sigma}^*L_{m_0} \otimes {\sigma^{-1}}^*L_{m_0}
= {\sigma^{{m_0}+1}}^*L \otimes {\sigma^{{m_0}-1}}^*L \otimes {\sigma^{-{m_0}+1}}^*L \otimes {\sigma^{-{m_0}-1}}^*L$$ and hence $${\sigma^{{m_0}+1}}^*L \otimes {\sigma^{-{m_0}-1}}^*L = L^{\otimes q\cdot q_{m_0} - q_{{m_0}-1}}.$$ Therefore, $${\sigma^{{m_0}+1}}^*L \otimes {\sigma^{-{m_0}-1}}^*L = L^{\otimes q_{m_0}}$$ where $q_{m_0} = q\cdot q_{{m_0}+1} - q_{{m_0}} \geq 2q_{{m_0}} - q_{{m_0}-1}>2$. Moreover, $q_{{m_0}+1} - q_{{m_0}} = (q-1)q_{{m_0}} - q_{{m_0}-1} > q_{m_0} - q_{{m_0}-1} >2$.
\[dense\] Let $W$ be a projective variety defined over a number field $K$, let $M = \{\phi, \phi^{-1}\}$ be an automorphism and its inverse on $W$. Suppose that $(W,M)$ is polarizable with some integer $q>2$. Then, ${\operatorname{Per}}(\phi)$ is Zariski dense.
Suppose that the Zariski closure of $\mathcal{S}$ is $\mathbf{C} = \{ C_1, \cdots , C_r\}$ where $C_i$ are irreducible curves. Then, $\sigma(C_i)$ is still in $\mathcal{C}$ and hence $\phi$ works as maps $\mathbf{C}$ into itself. Thus, there is a curve fixed by $\sigma^{m}$ for some $m>0$. Without loss of generality, $\sigma^m(C_1) = C_1$. Furthermore, since $\sigma$ is an automorphism, $\sigma^m_*C_1 = C_1$. Similarly, $\sigma^{-m}_*C_1 = C_1$. Then, $$L \cdot C_1 = L^d \cdot \sigma^m_* C_1 = {\sigma^m}^*L \cdot C_1 \quad L \cdot C_1 = L \cdot \sigma^m_* C_1 = {\sigma^m}^*L \cdot C_1$$ and hence $$L^{\otimes 2} \cdot C_1 = {\sigma^m}^*L \cdot C_1 + {\sigma^m}^*L \cdot C_1 = \left({\sigma^m}^*L \otimes {\sigma^m}^*L \right)\cdot C_1.$$
On the other hand, by Lemma \[induction\], $(W, \{\sigma^m, \sigma^{m-}\})$ is also polarizable. Thus, $$L^{\otimes 2} \cdot C_1 = L^{\otimes q_m} \cdot C_1 \quad \text{where}~q_m >2$$ and hence $L \cdot C_1 = 0$, which contradicts to the assumption that $L$ is ample.
Let $\sigma :W \rightarrow W$ be an automorphism on a projective surface $W$ such that $(W, \{ \sigma, \sigma^{-1}\})$ is polarizable with respect to an ample line bundle $L$. Then, a generic sequence $\{ x_m\} \subset {\operatorname{Per}}(\sigma)$ is small.
We know that $\widehat{h}_{\overline{L}}(x_m)=0$ if $x_m \in {\operatorname{Per}}(\sigma)$. Thus, we only have to show that $\widehat{h}_{\overline{L}}(W)=0$.
[@Z1 Theorem 1.10] says that $$e_1(L) \geq
\widehat{h}_{\overline{L}}(W) \geq \dfrac{1}{n} \sum_{i=1}^n e_i (\overline{L})$$ where $$e_i(\overline{L})
= \sup_{\substack{Y \subsetneq W \\ \operatorname{Codim} Y = i}} \inf_{x \in W \setminus Y} \widehat{h}_{\overline{L}} (x).$$
By Theorem \[dense\], ${\operatorname{Per}}(\sigma)$ is Zariski Dense in $W$ and hence $e_i(\overline{L}) =0$ for all $i = 1 \cdots n$ and hence $\widehat{h}_{\overline{L}}(W) =0$.
Now, Theorem \[dense\] say that we can build a generic and small sequence of periodic points. Therefore, we can prove the equidistribution of periodic point:
Let $W$ be a projective variety defined over a number field $K$, let $M = \{\sigma, \sigma^{-1}\}$ be an automorphism and its inverse on $W$. Suppose that $(W,M)$ is polarizable with some integer $q>2$. Then, ${\operatorname{Per}}(\sigma)$ is equidistributed.
It is an easy consequence of Theorem 1.2. and Theorem 1.3: by Theorem 1.3, we can build a generic and small sequence $\{x_m\}$ in ${\operatorname{Preper}}(\sigma)$. And, the sequence of probability measure on Galois orbit of $x_m$ weakly converges to the dynamical measure by Theorem 1.2.
[ABCDEFG]{}
Bilu, Y., [*Limit distribution of small points on algebraic tori*]{}, /duke Math., 89 (1997), 465–476
Baragar, A.; McKinnon, David., [*K3 surfaces, rational curves, and rational points*]{}, J. Number Theory 130 (2010), no. 7, 1470–1479
Baker, M. H.; Rumely, R.,[*Equidistribution of small points, rational dynamics, and potential theory*]{}, Ann. Inst. Fourier (Grenoble) 56 (2006), no. 3, 625–688
Chamert-Loir, A., [*Mesures et equidistribution sur les espaces de Berkovich*]{}, J. Reine Angew. Math. 595 (2006), 215–235
Fulton, W., [*Intersection theory* ]{}, Second edition, Springer-Verlag, Berlin, 1998.
Fakhruddin, N.,[*Questions on self maps of algebraic varieties*]{}, J. Ramanujan Math. Soc. 18 (2003), no. 2, 109–122
Favre, C.; Rivera-Letelier, J., [*Equidistribution quantitative des points de petite hauteur sur la droite projective*]{}, Math. Ann. 335 (2006), no. 2, 311–361
Kawaguchi, S., [*Canonical heights, invariant currents, and dynamical eigensystems of morphisms for line bundles*]{}, J. Reine Angew. Math. 597 (2006), 135–173,
Silverman, J. H. [*The arithmetic of dynamical system*]{}, Springer, 2007.
Silverman, J. H., [*Rational points on K3 surfaces: a new canonical height.*]{}, Invent. Math., 105(2) (1991), 347–373
Silverman, J. H.; Hindry, M. [*Diophantine geometry, An introduction*]{}, Springer, 2000.
Szpiro, L.; Ullmo, E.; Zhang, S., [*Equirepartition des petits points*]{}, Invent. Math. 127 (1997), no. 2, 337–347
Yuan, X., [*Big line bundles over arithmetic varieties*]{}, Invent. Math. 173, no. 3 (2008), 603–649
Zhang, S., [*Small points and adelic metrics*]{} J. Algebraic Geom. 4 (1995), no. 2, 281?300
| {
"pile_set_name": "ArXiv"
} |
---
subtitle: Acknowledgments
---
[**Super-weak force and neutrino masses** ]{}
[**Zoltán Trócsányi**]{},\
[Institute for Theoretical Physics, Eötvös Loránd University,\
Pázmány Péter sétány 1/A, H-1117 Budapest, Hungary and\
MTA-DE Particle Physics Research Group ]{}\
E-mail: [[email protected]]{}
[**Abstract**]{}
> 10000 We consider an anomaly free extension of the standard model gauge group $G_{\rm SM}$ by an abelian group to $G_{\rm SM}\otimes U(1)_Z$. The condition of anomaly cancellation is known to fix the $Z$-charges of the particles, but two. We fix one remaining charge by allowing for all possible Yukawa interactions of the known left handed neutrinos and new right-handed ones that obtain their masses through interaction with a new scalar field with spontaneously broken vacuum. We discuss some of the possible consequences of the model. Assuming that the new interaction is responsible for the observed differences between the standard model prediction for the anomalous magnetic moment of the muon and its measured value, we constrain the size of the new gauge coupling, the mass of the new gauge boson and the vacuum expectation value of the new scalar field.
2018
Introduction {#sec:intro}
============
The remarkable experimental success of the standard model of elementary particle interaction [@Weinberg:1967tq] leaves very little room for the explanation of the observed deviations from it. This success story has culminated in the discovery of the Higgs particle [@Aad:2012tfa; @Chatrchyan:2012xdj], which could not have happened without the immense theoretical input to the design of the accelerator and the experiments. With this discovery also a new era of particle physics has arrived as there is no established model that can guide us to new discoveries. Therefore, theories that might incorporate the existing deviations from the standard model are desirable.
The most outstanding experimental observations that cannot be explained by the standard model are the (i) abundance of dark matter in the universe; (ii) non-vanishing neutrino masses; (iii) leptogenesis;[^1] (iv) accelerating expansion of the universe, signalling the existence of dark energy [@Tanabashi:2018oca].[^2] In addition to (i)–(iv), (v) inflation in the early universe is also considered a fairly established fact, although there is no direct proof for it. All these facts have to be explained by such an extension of the standard model that respects (a) the high precision confirmation of the standard model at collider experiments (b) and the lack of finding new particles beyond the Higgs boson by the LHC experiments [@ATLAS; @CMS]. There is one more feature of the standard model, the metastability of vacuum [@Bezrukov:2009db; @Degrassi:2012ry] that does not necessarily require new physics, but if new physics exists, it should not worsen the stability, but possibly push the vacuum to the stability region.
In addition to the experimental success of the standard model, it is also highly efficient being based on the concepts of local gauge invariance and spontaneous symmetry breaking [@Englert:1964et; @Higgs:1964pj]. The only exception of economical description is the relatively large number of Yukawa couplings of the fermions needed to explain their masses. The generation of the fermion masses however, is also highly efficient in the sense that it uses the same spontaneous symmetry breaking of the scalar field to which all other particles owe their masses. In this spirit, it is reasonable to expect that the non-vanishing masses of the neutrinos should be explained by Yukawa couplings, too. Also, the choice of the gauge groups and number of family replications look arbitrary and presently these are determined by phenomenology only.
Clearly, the neutrino masses must play a fundamental role in the possible extensions of the standard model. As the gauge and mass eigenstates of the neutrinos differ, they must feel a second force to the gauge interaction. The second force can be a Yukawa coupling to a scalar. Such explanation of neutrino masses in general requires the assumption of the existence of right-handed neutrinos and perhaps a new scalar field.
In the spirit of economy and level of arbitrariness explained above, in this article we propose an extension of the zoo of particles in the standard model with three right-handed neutrinosand the gauge symmetry of the standard model Lagrangian $G_{\ensuremath{\mathrm{SM}}}= SU(3)_{\ensuremath{\mathrm{c}}}\otimes SU(2)_{{\mathrm{L}}}\otimes U(1)_Y$ to $G_{\ensuremath{\mathrm{SM}}}\otimes U(1)_Z$. Such extensions have already been considered in the literature extensively (for an incomplete set of popular examples and their studies see [@Schabinger:2005ei; @Pospelov:2007mp; @Basso:2008iv]). In particular, it was shown that the charge assignment of the matter fields is constrained by the requirement of anomaly cancellations up to two free charges [@Appelquist:2002mw]. To define the model completely, one has to take a specific choice for these remaining free charges. In this article we propose that the mechanism for the generation of neutrino masses fixes the values of the $U(1)_Z$ charges up to an overall scale that can be embedded in the $U(1)_Z$ coupling.
The difference between our proposal and existing studies is two-fold. The model proposed here introduces a new force along the same principles as the known forces are included in the standard model: all renormalizable terms that are allowed by the underlying gauge symmetry are present, but no other symmetry than the extra $U(1)_Z$ is assumed. Our primary goal is not the prediction of new observable phenomena at collider experiments, but first focus only on the unexplained phenomena (i–iv), with respecting the observations (a) and (b). As the deviations from the standard model are related to the intensity and cosmic frontiers of particle physics, we assume that the new $U(1)_Z$ interaction is secluded from the standard model by a small coupling. Thus we propose the model in a region of the parameter space that has received little attention before.
Definition of the model {#sec:model}
=======================
Fermion sector
--------------
We consider the usual three fermion families of the standard model extended with one right-handed Dirac neutrino in each family.[^3] We introduce the notation \_[q,1]{}\^f &= [$\left$]{}(
[c]{} U\^f\
D\^f
[$\right$]{})\_\_[q,2]{}\^f = U\_\^f ,\_[q,3]{}\^f = D\_\^f\
\_[l,1]{}\^f &= [$\left$]{}(
[c]{} \^f\
\^f
[$\right$]{})\_\_[l,2]{}\^f = \_\^f ,\_[l,3]{}\^f = \_\^f \[eq:psi\_j\] for the chiral quark fields $\psi_q$ and chiral lepton fields $\psi_l$. In [Eq.(\[eq:psi\_j\])]{} L and R denote the left and right-handed projections of the same field, [^4] \_[/]{}\_= [$\left$]{}(1\_[5]{}[$\right$]{})P\_[/]{} . \[eq:LRprojections\] , which differ from the charged fermions in the sense that [*the left and right-handed fields are projections of different fields*]{} \_= P\_\_[,1]{} \_= P\_\_[,2]{} . These two fields are Dirac spinors composed of single Weyl spinors, \_[,1]{} = [$\left$]{}(
[c]{} \_\
[$\mathrm{i}$]{}\_2\^\*\_
[$\right$]{}) ,\_[,2]{} = [$\left$]{}(
[c]{} [$\mathrm{i}$]{}\_2\^\*\_\
\_
[$\right$]{}) . Then the field content in family $f$ ($f=1$, 2 or 3) consists of two quarks, $U_f$, $D_f$, a neutrino $\nu_f$ and a charged lepton $\ell_f$ where $U_f$ is the generic notation for the u-type quarks $U_1 = $ u, $U_2 = $ c, $U_3 = $ t, while $D_f$ is that for d-type quarks, $D_1=$ d, $D_2 = $ s and $D_3 = $ b. The charged leptons $\ell_f$ can be $\ell_1
= e$, $\ell_2 = \mu$ or $\ell_3 = \tau$ and $\nu_f$ are the corresponding neutrinos, $\nu_1=\nu_e$, $\nu_2=\nu_\mu$, $\nu_3=\nu_\tau$.
For a matrix $U\in G_{{\ensuremath{\mathrm{SM}}}}\otimes U(1)_Z$ the three generic fields in [Eq.(\[eq:psi\_j\])]{} transform as U\_[1]{}[$\left$]{}(x[$\right$]{}) &= [$\mathrm{e}$]{}\^[[$\mathrm{i}$]{}[$\left$]{}(x[$\right$]{})]{} [$\mathrm{e}$]{}\^[[$\mathrm{i}$]{}y\_[1]{}[$\left$]{}(x[$\right$]{})]{} [$\mathrm{e}$]{}\^[[$\mathrm{i}$]{}z\_[1]{}[$\left$]{}(x[$\right$]{})]{} \_[1]{}(x) =12 (\_1,\_2,\_3)\
U\_[j]{}[$\left$]{}(x[$\right$]{}) &= [$\mathrm{e}$]{}\^[[$\mathrm{i}$]{}y\_[j]{}[$\left$]{}(x[$\right$]{})]{} [$\mathrm{e}$]{}\^[[$\mathrm{i}$]{}z\_[j]{}[$\left$]{}(x[$\right$]{})]{}\_[j]{}(x) j=2,3 and ${{\mbox{\boldmath $\alpha$}}} = (\alpha_1,\alpha_2,\alpha_3)$, with $\alpha_i$, $\beta$, $\zeta\in\mathbb{R}$. The matrices $\tau_{i}$ are the Pauli matrices, $y_j$ is the hypercharge, while $z_j$ denotes the $Z$-charge of the field $\psi_j$. There is a lot of freedom how to choose the $Z$-charges. In this article we make two assumptions that fix these completely. The first is that the charges do not depend on the families, which is also the case in the standard model.[^5] With this assumption the assignment for the $Z$-charges of the fermions can be expressed using two free numbers $Z_1$ and $Z_2$ of the $U$ quark fields if we want a model free of gauge and gravity anomalies. The rest of the charges must take values as given in [Table \[tab:charges\]]{} [@Appelquist:2002mw].
.
field $SU(3)_{\ensuremath{\mathrm{c}}}$ $SU(2)_{{\mathrm{L}}}$ $y_j$ $z_j$ $z_j$ $r_j=z_j/z_\phi-y_j$
----------------------------------------------------- ----------------------------------- ------------------------ ------------ ------------- ------------ ----------------------
$U_{{\mathrm{L}}}$, $D_{{\mathrm{L}}}$ 3 2 $\frac16$ $Z_1$ $\frac16$ 0
\[2mm\] $U_{{\mathrm{R}}}$ 3 1 $\frac23$ $Z_2$ $\frac76$ $\frac12$
\[2mm\] $D_{{\mathrm{R}}}$ 3 1 $-\frac13$ $2 Z_1-Z_2$ $-\frac56$ $-\frac12$
\[2mm\] $\nu_{{\mathrm{L}}}$, $\ell_{{\mathrm{L}}}$ 1 2 $-\frac12$ $-3 Z_1$ $-\frac12$ 0
\[2mm\] $\nu_{{\mathrm{R}}}$ 1 1 0 $Z_2-4 Z_1$ $\frac12$ $\frac12$
\[2mm\] $\ell_{{\mathrm{R}}}$ 1 1 $-1$ $-2Z_1-Z_2$ $-\frac32$ $-\frac12$
\[2mm\] $\phi$ 1 2 $\frac12$ $z_\phi$ 1 $\frac12$
\[2mm\] $\chi$ 1 1 0 $z_\chi$ $-1$ $-1$
: Assignments for the representations (for $SU(N)$) and charges (for $U(1)$) of fermion and scalar fields of the complete model. The charges $y_j$ denote the eigenvalue of $Y/2$, with $Y$ being the hypercharge operator and $z_j$ denote the supercharges of the fields $\psi_j$ of [Eq.(\[eq:psi\_j\])]{} ($j=1$, 2, 3). The right-handed Dirac neutrinos $\nu_{{{\mathrm{R}}}}$ are sterile under the $G_{\rm SM}$ group. The sixth column gives a particular realization of the $U(1)_Z$ charges, motivated below, and the last one is added for later convenience.[]{data-label="tab:charges"}
The Dirac Lagrangian summed over the family replications, [$\mathcal{L}$]{}\_&= [$\mathrm{i}$]{}\_[f=1]{}\^[3]{}\_[j=1]{}\^[3]{}( \^f\_[q,j]{}(x)\_[j]{}\_[q,j]{}\^f(x) + \^f\_[l,j]{}(x)\_[j]{}\_[l,j]{}\^f(x) ) ,\
D\^\_[j]{} &= \^ + [$\mathrm{i}$]{}g\_\^ + [$\mathrm{i}$]{}g\_Yy\_[j]{}B\^ + [$\mathrm{i}$]{}g\_Zz\_[j]{}Z\^ \[eq:LD\] is invariant under local $G=G_{\rm SM}\otimes U(1)_Z$ gauge transformations, provided the five gauge fields introduced in the covariant derivative transform as \^(x)& \^(x)= U(x)\^(x)U\^(x) +[$\left$]{}\[\^U(x)[$\right$]{}\]U\^(x)\
B\^& B\^(x)=B\^(x)-\^(x)\
Z\^& Z\^(x)=Z\^(x)-\^(x) \[eq:WBXtransformations\] where $U(x)=\exp{\ensuremath{\left}}[{\ensuremath{\mathrm{i}}}{{{\mbox{\boldmath $T$}}}}\cdot{{\mbox{\boldmath $\alpha$}}}{\ensuremath{\left}}(x{\ensuremath{\right}}){\ensuremath{\right}}]$. The gauge invariant kinetic term for these vector fields is $${\ensuremath{\mathcal{L}}}_{B,Z,W}=
-\frac{1}{4}B_{\mu\nu}B^{\mu\nu}
-\frac{1}{4}Z_{\mu\nu}Z^{\mu\nu}
-\frac{1}{4}{{\mbox{\boldmath $W$}}}_{\mu\nu}\cdot{{\mbox{\boldmath $W$}}}^{\mu\nu},
\label{eq:LBZW}$$ with $B_{\mu\nu} = \partial_{\mu}B_{\nu}-\partial_{\nu}B_{\mu}
\equiv\partial_{[\mu}B_{\nu]}$, $Z_{\mu\nu} = \partial_{[\mu}Z_{\nu]}$ and ${{\mbox{\boldmath $W$}}}_{\mu\nu} = \partial_{[\mu}{{\mbox{\boldmath $W$}}}_{\nu]}
-g\,{{\mbox{\boldmath $W$}}}_{\mu}\times{{\mbox{\boldmath $W$}}}_{\nu}$. The field strength ${{{\mbox{\boldmath $T$}}}}\cdot{{\mbox{\boldmath $W$}}}_{\mu\nu}$ transforms covariantly under $G$ transformations, ${{{\mbox{\boldmath $T$}}}}\cdot{{\mbox{\boldmath $W$}}}_{\mu\nu}\stackrel{G}{\longrightarrow}
U(x)\,{{{\mbox{\boldmath $T$}}}}\cdot{{\mbox{\boldmath $W$}}}_{\mu\nu}\,U^{\dag}(x)$, but $B_{\mu\nu}$ and $Z_{\mu\nu}$ are invariant, hence a kinetic mixing term of the $U(1)$ fields is also allowed by gauge invariance: - B\_ Z\^. \[eq:kinetic-mixing\] We can get rid of this mixing term by redefining the $U(1)$ fields using the transformation [$\left$]{}(
[c]{} B\_’\
Z\_’
[$\right$]{}) = [$\left$]{}(
[cc]{} 1 & [\_[Z]{}]{}\
0 & [\_[Z]{}]{}
[$\right$]{}) [$\left$]{}(
[c]{} B\_\
Z\_
[$\right$]{}) ,[\_[Z]{}]{}= . In terms of the redefined fields, the covariant derivative becomes D\^\_j= \^ +[$\mathrm{i}$]{}g\_\^+[$\mathrm{i}$]{}g\_Yy\_[j]{}B\^ +[$\mathrm{i}$]{}(g\_Z’z\_[j]{} - g\_Y’y\_[j]{})Z\^ \[eq:cov-dev\] where $g_Y' = g_Y \tan{\theta_{Z}}= \epsilon g_Y + {{\mathrm{O}}}(\epsilon^3)$ and $g_Z' = g_Z/\cos{\theta_{Z}}= g_Z + {{\mathrm{O}}}(\epsilon^2)$. Thus the effect of the kinetic mixing is to change the couplings of the matter fields to the vector field $Z^\mu$. Note that we cannot immediately combine the coupling factor $(g_Z'\,z_{j} - g_Y'\,y_{j})$ into a single product of a coupling and a charge. We shall discuss this issue further below.
Gauge symmetry forbids mass terms for gauge bosons. Fermion masses must also be absent because $$m\,\bar{\psi}\psi =
m\,\bar{\psi}_{{\mathrm{L}}}\psi_{{\mathrm{R}}}+m\,\bar{\psi}_{{\mathrm{R}}}\psi_{{\mathrm{L}}},$$ but the $\psi_{{\mathrm{L}}}$, $\psi_{{\mathrm{R}}}$ fields transform differently under $G$. Thus the $G$-invariant Lagrangian describes massless fields in contradiction to observation.
Scalar sector
-------------
To solve the puzzle of missing masses we proceed similarly as in the standard model, but in addition to the usual Brout-Englert-Higgs (BEH) field $\phi$ that is an $SU(2)_{{\mathrm{L}}}$-doublet $$\phi={\ensuremath{\left}}(\!\!\begin{array}{c}
\phi^{+} \\
\phi^{0}
\end{array}\!\!{\ensuremath{\right}}) =
\frac{1}{\sqrt{2}}
{\ensuremath{\left}}(\!\!\begin{array}{c}
\phi_{1}+{\ensuremath{\mathrm{i}}}\phi_{2} \\
\phi_{3}+{\ensuremath{\mathrm{i}}}\phi_{4}
\end{array}\!\!
{\ensuremath{\right}})
\,,$$ we also introduce another complex scalar $\chi$ that transforms as a singlet under $G_{\rm SM}$ transformations. The gauge invariant Lagrangian of the scalar fields is [$\mathcal{L}$]{}\_[,]{} = \[D\_ \]\^\* D\_\^ + \[D\_ \]\^\* D\_\^ - V(,) \[eq:Lphichi\] where the covariant derivative for the scalar $s$ ($s=\phi$, $\chi$) is D\_[s]{}\^ = \^ +[$\mathrm{i}$]{}g\_\^+[$\mathrm{i}$]{}g\_Yy\_[s]{}B\^ +[$\mathrm{i}$]{}(g\_Z’z\_[s]{} - g\_Y’y\_[s]{})Z\^ \[eq:cov-dev-scalar\] and the potential energy V(,) = V\_0 - \_\^2 ||\^2 - \_\^2 ||\^2 + [$\left$]{}(||\^2, ||\^2[$\right$]{}) [$\left$]{}(
[cc]{} \_& 2\
2 & \_
[$\right$]{}) [$\left$]{}(
[c]{} ||\^2\
||\^2
[$\right$]{}), \[eq:V\] in addition to the usual quartic terms, introduces a coupling term $-\lambda |\phi|^2 |\chi|^2$ of the scalar fields in the Lagrangian. For the doublet $|\phi|$ denotes the length $\sqrt{|\phi^+|^2 + |\phi^0|^2}$. The value of the additive constant $V_0$ is irrelevant for particle dynamics, but may be relevant for inflationary scenarios, hence we allow for its non-vanishing value. In order that this potential energy be bounded from below, we have to require the positivity of the self-couplings, $\lambda_\phi$, $\lambda_\chi>0$. The eigenvalues of the coupling matrix are \_= 12 (\_+\_) , while the corresponding un-normalized eigenvectors are u\^[(+)]{} = [$\left$]{}(
[c]{} (\_+ - \_)\
1
[$\right$]{}) u\^[(-)]{} = [$\left$]{}(
[c]{} (\_- - \_)\
1
[$\right$]{}) . As $\lambda_+>0$ and $\lambda_-<\lambda_+$, in the physical region the potential can be unbounded from below only if $\lambda_-<0$ and $u^{(-)}$ points into the first quadrant, which may occur only when $\lambda<0$. In this case, to ensure that the potential is bounded from below, one also has to require that the coupling matrix be positive definite, which translates into the condition 4 \_\_- \^2 > 0 . \[eq:positivity\] With these conditions satisfied, we can find the minimum of the potential energy at field values $\phi= v/\sqrt{2}$ and $\chi = w/\sqrt{2}$ where the vacuum expectation values (VEVs) are v = ,w = . \[eq:VEVs\] Using the VEVs, we can express the quadratic couplings as \_\^2 = \_v\^2 + w\^2 ,\_\^2 = \_w\^2 + v\^2 , \[eq:scalarmasses\] so those are both positive if $\lambda > 0$. If $\lambda < 0$, the constraint (\[eq:positivity\]) ensures that the denominators of the VEVs in [Eq.(\[eq:VEVs\])]{} are positive, so the VEVs have non-vanishing real values only if 2\_\_\^2 - \_\^2 > 0 2\_\_\^2 - \_\^2 > 0 \[eq:muXconditions\] simultaneously, which can be satisfied if at most one of the quadratic couplings is smaller than zero. We summarize the possible cases for the signs of the couplings in [Table \[tab:scalarcouplings\]]{}.
.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$\Theta(\lambda)$ $\Theta(\lambda_\phi)\!$ $\!\Theta(\lambda_\chi)$ $\Theta(4 \lambda_\phi \lambda_\chi - \lambda^2)$ $\Theta(\mu_\phi^2)\,\Theta(\mu_\chi^2)$ $\Theta(2\lambda_\chi \mu_\phi^2 - \lambda \mu_\chi^2)
\Theta(2\lambda_\phi \mu_\chi^2 - \lambda \mu_\phi^2)$
------------------- -------------------------- -------------------------- --------------------------------------------------- ------------------------------------------ ---------------------------------------------------------
1 1 1 unconstrained 1 unconstrained
\[2mm\] 0 1 1 1 1 unconstrained
0 1 1 1 0 1
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: Possible signs of the couplings in the scalar potential $V(\phi,\chi)$ in order to have two non-vanishing real VEVs. $\Theta$ is the step function, $\Theta(x)=1$ if $x>0$ and 0 if $x<0$.[]{data-label="tab:scalarcouplings"}
After spontaneous symmetry breaking of $G \to SU(3)_{\ensuremath{\mathrm{c}}}\otimes U(1)_Q$, we use the following convenient parametrization for the scalar fields: $$\phi=\frac{1}{\sqrt{2}}\,{\ensuremath{\mathrm{e}}}^{{\ensuremath{\mathrm{i}}}{{{\mbox{\boldmath $T$}}}}\cdot{{\mbox{\boldmath $\xi$}}}(x)/v}
{\ensuremath{\left}}(\!\!\begin{array}{c} 0 \\ v+h'(x) \end{array}\!\!{\ensuremath{\right}})
\quad\mbox{and}\quad
\chi(x) =
\frac{1}{\sqrt{2}}\,{\ensuremath{\mathrm{e}}}^{{\ensuremath{\mathrm{i}}}\eta(x)/w}\big(w + s'(x)\big)
\,.
\label{eq:BEHparametrization}$$ We can use the gauge invariance of the model to choose the unitary gauge when ’(x)= [$\left$]{}(
[c]{} 0\
v+h’(x)
[$\right$]{}) ’(x) = (w + s’(x)) and the vector fields are transformed according to [Eq.(\[eq:WBXtransformations\])]{}. With this gauge choice, the scalar kinetic term contains quadratic terms of the gauge fields from which one can identify mass parameters of the massive standard model gauge bosons proportional to the vacuum expectation value $v$ of the BEH field and also that of a massive vector boson $Z^{\prime\mu}$ proportional to $w$.
We can diagonalize the mass matrix (quadratic terms) of the two real scalars ($h'$ and $s'$) by the rotation [$\left$]{}(
[c]{} h\
s
[$\right$]{}) = [$\left$]{}(
[cr]{} [\_S]{}& -[\_S]{}\
[\_S]{}& [\_S]{}
[$\right$]{}) [$\left$]{}(
[c]{} h’\
s’
[$\right$]{}) where for the scalar mixing angle ${\theta_S}\in (-\frac\pi4,\frac\pi4)$ we find (2[\_S]{}) = - . The masses of the mass eigenstates $h$ and $s$ are M\_[h/H]{} = [$\left$]{}(\_v\^2 + \_w\^2 [$\right$]{})\^[1/2]{} \[eq:Mhs\] where $M_h \leq M_H$ by convention. At this point either $h$ or $H$ can be the standard model Higgs boson. A more detailed analysis of this scalar sector but within a different $U(1)_Z$ model can be found in [Ref.[@Duch:2015jta]]{} and for the present model in [Ref.[@Peli:2019xwv]]{}.
As $M_h$ must be positive, the condition v\^2 w\^2 (4 \_\_- \^2) > 0 \[eq:Mhpositivity\] has to be fulfilled. If both VEVs are greater than zero, then this condition reduces to the positivity constraint (\[eq:positivity\]), but with different meaning. [Eq.(\[eq:positivity\])]{} is required to ensure that the potential be bounded from below if $\lambda<0$, which has to be fulfilled at any scale. For $\lambda>0$, the potential is bounded from below even without requiring the constraint (\[eq:positivity\]). The inequality in (\[eq:Mhpositivity\]) ensures $M_h > 0$, which does not have to be fulfilled up to the Planck scale. It may happen that $w$ vanishes at some scale, above which there is only one massive scalar particle.
The VEV of the BEH field and the mass of the Higgs boson are known experimentally, $v = 246.22$GeV and $m_H = 125.09$GeV \cite{}. Introducing the abbreviation ${\ensuremath{\lambda_{{\ensuremath{\mathrm{SM}}}}}}= \frac12 m_H^2/v^2 \simeq 0.129$, we distinguish two cases at the weak scale: (i) $\lambda_\phi > {\ensuremath{\lambda_{{\ensuremath{\mathrm{SM}}}}}}$ and (ii) ${\ensuremath{\lambda_{{\ensuremath{\mathrm{SM}}}}}}> \lambda_\phi$. If $\lambda_\phi > {\ensuremath{\lambda_{{\ensuremath{\mathrm{SM}}}}}}$, then $M_H > m_H$, so only $h$ can be the Higgs particle ($\sin(2{\theta_S})$, $\cos(2{\theta_S}) > 0$) and M\_h = m\_H ,M\_H = m\_H . \[eq:MhMs\] The positivity of $M_H^2$, in addition to the constraint in (\[eq:Mhpositivity\]), also requires that 4 (\_-[$\lambda_{{\ensuremath{\mathrm{SM}}}}$]{}) \_-\^2 > 0 \_> [$\lambda_{{\ensuremath{\mathrm{SM}}}}$]{}+ , which is stronger than (\[eq:Mhpositivity\]).
In case (ii), $m_H^2 > 2 \lambda_\phi v^2 > M_h^2$, so only $H$ can be the Higgs particle ($\sin(2{\theta_S})$, $\cos(2{\theta_S}) < 0$) and we can express the masses of the scalars as in [Eq.(\[eq:MhMs\])]{}, with $h$ and $H$ interchanged, or explicitly M\_h = m\_H M\_H = m\_H , \[eq:MsMh\] which does not require any further constraint to (\[eq:Mhpositivity\]).
We see that the signs in the rotation matrix are tied to the sign of $\lambda_\phi - {\ensuremath{\lambda_{{\ensuremath{\mathrm{SM}}}}}}$: (2[\_S]{}) = (\_- [$\lambda_{{\ensuremath{\mathrm{SM}}}}$]{}) 2 ,(2[\_S]{}) = (\_- [$\lambda_{{\ensuremath{\mathrm{SM}}}}$]{}) 2 . Finally, we can relate the new VEV $w$ to the BEH VEV $v$ and the four couplings ${\ensuremath{\lambda_{{\ensuremath{\mathrm{SM}}}}}}$, $\lambda_\phi$, $\lambda_\chi$, $\lambda$ using [Eq.(\[eq:Mhs\])]{} as w\^2 (4 (\_- [$\lambda_{{\ensuremath{\mathrm{SM}}}}$]{}) \_- \^2) = 4 v\^2 [$\lambda_{{\ensuremath{\mathrm{SM}}}}$]{}(\_-[$\lambda_{{\ensuremath{\mathrm{SM}}}}$]{}) , \[eq:w2\] which shows that $w^2 = 0$ if and only if $\lambda_\phi={\ensuremath{\lambda_{{\ensuremath{\mathrm{SM}}}}}}$.
NOTE: As ${\ensuremath{\lambda_{{\ensuremath{\mathrm{SM}}}}}}$ runs (decreases) faster than $\lambda_\phi$, $w^2=0$ can happen only if ${\ensuremath{\lambda_{{\ensuremath{\mathrm{SM}}}}}}>\lambda_\phi$ at small scales. When ${\ensuremath{\lambda_{{\ensuremath{\mathrm{SM}}}}}}$ reaches $\lambda_\phi$, $w^2$ and $M_h$ vanish, only the BEH vacuum remains non-zero.
Fermion masses {#ssec:Yukawas}
--------------
We already discussed that explicit mass terms of fermions would break $SU(2)_{{\mathrm{L}}}\otimes U(1)_Y$ invariance. However, we can introduce gauge-invariant fermion-scalar Yukawa interactions[^6] [$\mathcal{L}$]{}\_ = -[$\left$]{}\[c\_[D]{} |[Q]{}\_D\_+c\_[U]{} |[Q]{}\_U\_+c\_|[L]{}\_\_[$\right$]{}\] + \[eq:LY\] where h.c. means hermitian conjugate terms and the parameters $c_{D},\,c_{U},\,c_{\ell}$ are called Yukawa couplings that are matrices in family indices and summation over the families is understood implicitly. The dot product abbreviates scalar products of $SU(2)$ doublets: |[Q]{}\_[$\left$]{}(|[U]{},|[D]{}[$\right$]{})\_[$\left$]{}(
[c]{} \^[[$\left$]{}(+[$\right$]{})]{}\
\^[[$\left$]{}(0[$\right$]{})]{}
[$\right$]{}) ,|[Q]{}\_ [$\left$]{}(|[U]{},|[D]{}[$\right$]{})\_[$\left$]{}(
[r]{} \^[[$\left$]{}(0[$\right$]{})]{}\
-\^[[$\left$]{}(+[$\right$]{})]{}
[$\right$]{}) and $\bar{L} \equiv {\ensuremath{\left}}(\bar{\nu}_{\ell},\,\bar{\ell}{\ensuremath{\right}})$. The $Z$-charge of the BEH field is constrained by $U(1)_Z$ invariance of the Yukawa terms to $z_\phi = Z_2 - Z_1$, which works simultaneously for all three terms.
After spontaneous symmetry breaking and fixing the unitary gauge, this Yukawa Lagrangian becomes $${\ensuremath{\mathcal{L}}}_{{{\mathrm{Y}}}} =
-\frac{1}{\sqrt{2}}{\ensuremath{\left}}(v+h(x){\ensuremath{\right}})
{\ensuremath{\left}}[c_{D}\,\bar{D}_{{\mathrm{L}}}D_{{\mathrm{R}}}+ c_{U}\,\bar{U}_{{\mathrm{L}}}U_{{\mathrm{R}}}+ c_{\ell}\,\bar{\ell}_{{\mathrm{L}}}\ell_{{\mathrm{R}}}{\ensuremath{\right}}]
+{\rm h.c.}
\label{eq:LYcharged}$$ We see that there are mass terms with mass matrices $M_{i}=\frac{c_{i}v}{\sqrt{2}}$ where $i=D$, $U$, $\ell$: $${\ensuremath{\mathcal{L}}}_{{{\mathrm{Y}}}} =
- {\ensuremath{\left}}(1+\frac{h(x)}{v}{\ensuremath{\right}})
{\ensuremath{\left}}[\bar{D}_{{\mathrm{L}}}\,M_{D}\,D_{{\mathrm{R}}}+ \bar{U}_{{\mathrm{L}}}\,M_{U}\,U_{{\mathrm{R}}}+ \bar{\ell}_{{\mathrm{L}}}\,M_{\ell}\,\ell_{{\mathrm{R}}}{\ensuremath{\right}}]
+{\rm h.c.}
\label{eq:chargedmasses}$$ The general complex matrices $M_{i}$ can be diagonalized employing bi-unitary transformations. The diagonal elements on the basis of mass eigenstates provide the mass parameters of the fermions. Due to the bi-unitary transformation the left and right-handed components of the fermion field are different linear combinations of the mass eigenstates.
The neutrino oscillation experiments suggest non-vanishing neutrino masses and the weak and mass eigenstates of the left-handed neutrinos do not coincide. In principle, the charge assignment of our model allows for the following gauge invariant Yukawa terms of dimension four operators for the neutrinos [$\mathcal{L}$]{}\^\_ = -\_[i,j]{}( (c\_)\_[ij]{} |[L]{}\_[i,]{}\_[j,]{} + 12 (c\_)\_[ij]{} \_[j,]{}) + [h.c.]{} \[eq:nuYukawa\] for arbitrary values of $Z_1$ and $Z_2$ if the superscript $c$ denotes the charge conjugate of the field, $\nu^c = -{\ensuremath{\mathrm{i}}}\gamma_2 \nu^*$ and the $Z$-charge of the right-handed neutrinos and the new scalar satisfy the relation $z_\chi = -2 z_{\nu_{{\mathrm{R}}}}$. There are two natural choices to fix the $Z$-charges: (i) the left- and right-handed neutrinos have the same charge, or (ii) those have opposite charges. In the first case we have Z\_2-4 Z\_1 = -3 Z\_1 , which is solved by $Z_1 = Z_2$ and it leads to the charge assignment of the $U(1)_{B-L}$ extension of the standard model, studied in detail (see for instance, [@Basso:2011hn] and references therein). In the second case Z\_2-4 Z\_1 = 3 Z\_1 , \[eq:Z2-Z1\] which is solved by $Z_1 = Z_2/7$. As the overall scale of the $Z$-charges depends only on the value of the gauge coupling $g_Z'$, we set $Z_2$ freely. For instance, choosing $Z_2=7/6$ implies $Z_1 = 1/6$ and the $Z$-charge of the BEH scalar is z\_= 1 , \[eq:zphi\] while that of the new scalar is z\_= -1 = -z\_.
While we cannot exclude the infinitely many cases when the magnitudes of $Z$-charges of the left- and right-handed neutrinos differ, we find natural to assume that [Eq.(\[eq:Z2-Z1\])]{} is valid. The corresponding $Z$-charges are given explicitly in the sixth column of [Table \[tab:charges\]]{}.
After the spontaneous symmetry breaking of the vacuum of the scalar fields [Eq.(\[eq:nuYukawa\])]{} leads to the following mass terms for the neutrinos: [$\mathcal{L}$]{}\^\_ = -12 \_[i,j]{}\[eq:numasses1\] where M(h,s)\_[ij]{} = [$\left$]{}(
[cc]{} 0 & m\_(1+)\
m\_(1+) & M\_(1+)
[$\right$]{})\_[ij]{} , \[eq:massmatrix1\] with complex $m_{{\mathrm{D}}}$ and real $M_{{\mathrm{M}}}$ being symmetric $3\times 3$ matrices, so $M(0,0)$ is a complex symmetric $6\times 6$ matrix. The diagonal elements of the mass matrix $M(0,0)$ provide Majorana mass terms for the left-handed and right-handed neutrinos. Thus we conclude that the model predicts [*vanishing masses of the left-handed neutrinos*]{} at the fundamental level.
The off-diagonal elements represent interaction terms that look formally like Dirac mass terms, $- \sum_{i,j} \overline{\nu_{i,{{\mathrm{L}}}}} (m_{{\mathrm{D}}})_{ij} \nu_{j,{{\mathrm{R}}}} + $ h.c. After spontaneous symmetry breaking the quantum numbers of the particles $\nu^c_{i,{{\mathrm{L}}}}$ and $\nu_{i,{{\mathrm{R}}}}$ are identical, hence they can mix. Thus the propagating states will be a mixture of the left- and right-handed neutrinos. Those states can be obtained by the diagonalization of the full matrix $M(0,0)$, for which a possible parametrization is given for instance in [Ref.[@Blennow:2011vn]]{}.
In order to understand the structure of the matrix $M(0,0)$ better, we first diagonalize the matrices $m_{{\mathrm{D}}}$ and $M_{{\mathrm{M}}}$ separately by a unitary transformation and an orthogonal one. Defining ’\_[,i]{} = \_j (U\_)\_[ij]{} \_[,j]{} ’\_[,i]{} = \_j (O\_)\_[ij]{} \_[,j]{} , we can rewrite the neutrino Yukawa Lagrangian as [$\mathcal{L}$]{}\^\_ = -12 \_[i,j]{}\[eq:numasses2\] where M’(h,s) = [$\left$]{}(
[cc]{} 0 & m V (1+)\
V\^m (1+) & M (1+)
[$\right$]{}) . \[eq:massmatrix2\] In [Eq.(\[eq:massmatrix2\])]{} $m$ and $M$ are real diagonal matrices, while $V = U_{{\mathrm{L}}}^T O_{{\mathrm{R}}}$ is a unitary matrix, $V V^\dag = 1$, so $M'(0,0)$ is Hermitian with real eigenvalues that are the masses of the mass eigenstates of neutrinos. In general, $M'(0,0)$ may have 15 independent parameters: $m_i$ and $M_i$ ($i=1$, 2 ,3), while there are three Euler angles and six phases $V$. Three phases can be absorbed into the definition of $\nu'_{{\mathrm{L}}}$.
Assuming the hierarchy $m_i \ll M_j$, we can integrate out the right-handed (heavy) neutrinos and obtain a gauge invariant higher dimensional operator of the form [$\mathcal{L}$]{}\_[dim-5]{}\^= - \_[i,j]{} (|[L]{}\_[i,]{}) (\^L\_[j,]{})\^. After spontaneous symmetry breaking of the BEH vacuum and choosing the unitary gauge, this an effective higher dimensional operator with Majorana mass terms for the left-handed neutrinos [$\mathcal{L}$]{}\_[dim-5]{}\^= - 12 \_[i]{} m\_[,i]{} (1+)\^2 ( ’\_[i,]{} + [h.c.]{}) . The Majorana masses m\_[,i]{} = are suppressed by the ratios $m_i/M_i$ as compared to $m_i$. The latter have a similar role in the Lagrangian as the mass parameters of the charged leptons, so one may assume $m_i \sim$ O(100keV), while the masses of the right-handed neutrinos can be naturally around O(100GeV), so that $m_i/M_i\sim{{\mathrm{O}}}(10^{-6\pm1})$ and $m_{{{\mathrm{M}}},i} \lesssim 0.1$eV. Thus if $m_i \ll M_i$, then the [*mixing between the light and heavy neutrinos will be very small, the $\nu'_{i,{{\mathrm{L}}}}$ can be considered as the mass eigenstates that are mixtures of the left-handed weak eigenstates, and whose masses can be small naturally*]{} as suggested by phenomenological observations.
As we can only observe neutrinos together with their flavours through their charged current interactions, it is more natural to use the flavour eigenstates than the mass eigenstates. In the flavour basis, the couplings of the leptons to the W boson are diagonal: [$\mathcal{L}$]{}\^[()]{}\_[CC]{} = - \_f \^f \^\^f\_+ [h.c.]{} , \[eq:LleptonCC\] with summation over the three lepton flavours $f = e$, $\mu$ and $\tau$. The same charged current interactions in mass basis $\nu_{{{\mathrm{L}}},i} = (U_{\rm PMNS})_{if} \nu_{{\mathrm{L}}}^f$, contains the Pontecorvo-Maki-Nakagawa-Sakata matrix $U_{\rm PMNS}$, [$\mathcal{L}$]{}\^[()]{}\_[CC]{} = - \_[i,f=1]{}\^3 (U\_[PMNS]{})\_[if]{}\^\_\^f + [h.c.]{} , just like the charged current quark interactions contain the Cabibbo-Kobayashi-Maskawa matrix. If the heavy neutrinos are integrated out, then the matrix $U_{{\mathrm{L}}}$ coincides with the PMNS matrix. For propagating degrees of freedom, such as in the case of travelling neutrinos over macroscopic distances, one should use mass eigenstates $\nu_{{{\mathrm{L}}},i}$ and the PMNS matrix becomes the source of neutrino oscillations in flavour space. However, in the case of elementary particle scattering processes involving the left-handed neutrinos, one can work using the flavour basis, i.e. with [Eq.(\[eq:LleptonCC\])]{} because the effect of their masses can be neglected.
Re-parametrization into right-handed and mixed couplings
--------------------------------------------------------
Having set the $Z$-charges of the matter fields, we can re-parametrize the couplings to $Z'$ using the new coupling g’\_[ZY]{} = g\_Z’ - g\_Y’ = . \[eq:gZYdef\] Then the covariant derivative in [Eq.(\[eq:cov-dev\])]{} becomes D\^\_j= \^ +[$\mathrm{i}$]{}g\_\^+[$\mathrm{i}$]{}y\_[j]{} g\_Y B\^ +[$\mathrm{i}$]{}[$\left$]{}(r\_j g\_Z’ + y\_[j]{} g’\_[ZY]{} [$\right$]{})Z\^ \[eq:cov-dev2\] where $r_j = z_j - y_j$ and its values are given explicitly in the last column of [Table \[tab:charges\]]{}. Thus, if a $U(1)_Z$ extension of $G_{\rm SM}$ is free of gauge and gravity anomalies and the $Z$-charges of left and right-handed fields are the opposite, then it is equivalent to a $U(1)_{{\mathrm{R}}}$ extension with tree-level mixed coupling $g_{ZY}'$ [@delAguila:1995rb], related to the kinetic mixing parameter ${\theta_{Z}}$ by [Eq.(\[eq:gZYdef\])]{}.
Particle phenomenology of the standard model suggests that the interaction of the fermions through the $Z'$ vector boson must be suppressed significantly. The origin of such a suppression can be either a small coupling to $Z'$ or the large mass of $Z'$. Usual studies in the literature focus on the latter case. Here we explore the former possibility.
The complete Lagrangian is the sum of the pieces given in Eqs. (\[eq:LD\]), (\[eq:LBZW\]), (\[eq:Lphichi\]), (\[eq:LY\]), (\[eq:nuYukawa\]), [$\mathcal{L}$]{}= [$\mathcal{L}$]{}\_+ [$\mathcal{L}$]{}\_[B,Z,W]{} + [$\mathcal{L}$]{}\_[,]{} + [$\mathcal{L}$]{}\_+ [$\mathcal{L}$]{}\_\^\[eqn:Lfull\] with covariant derivative given in [Eq.(\[eq:cov-dev2\])]{}, i.e. the kinetic mixing of [Eq.(\[eq:kinetic-mixing\])]{} is also taken into account.
Mixing in the neutral gauge sector
----------------------------------
The neutral gauge fields of the standard model and the $Z'$ mix, which leads to mass eigenstates $A_\mu$, $Z_\mu$ and $T_\mu$ (not to be confused with the isospin components $T_i$, $i=1$, 2, 3). The mixing is described by a $3\times 3$ mixing matrix as [$\left$]{}(
[c]{} W\^3\_\
B’\_\
Z’\_
[$\right$]{}) = [$\left$]{}(
[rrr]{} [\_]{}[\_[T]{}]{}& [\_]{}[\_[T]{}]{}& [\_]{}\
-[\_]{}[\_[T]{}]{}&-[\_]{}[\_[T]{}]{}& [\_]{}\
-[\_[T]{}]{}& [\_[T]{}]{}& 0
[$\right$]{}) [$\left$]{}(
[c]{} Z\_\
T\_\
A\_
[$\right$]{}) . For the Weinberg mixing angle ${\theta_{\mathrm{W}}}$ we have the usual value $\sin{\theta_{\mathrm{W}}}= g_Y/\sqrt{g_{{\mathrm{L}}}^2 + g_Y^2}$. We introduce the notion of reduced coupling defined by $\gamma_i = g_i/g_{{\mathrm{L}}}$, i.e. $\gamma_{{\mathrm{L}}}=1$. Then we have [\_]{}= ,[\_]{}= \[eq:sW-cW\] and for the mixing angle ${\theta_{T}}$ of the $Z'$ boson we find [\_[T]{}]{}&= \^[1/2]{} ,\
[\_[T]{}]{}&= \^[1/2]{} , \[eq:sT-cT\] so $\tan(2 {\theta_{T}}) = 2 \kappa/(1 - \kappa^2 - \tau^2)$, with = ,= 2 \[eq:kappa-tau\] and = is the ratio of the scalar vacuum expectation values (not a scalar mixing angle). For small values of the new couplings $\gamma'_{ZY}$ and $\gamma_Z'$, implying small $\kappa$, we have [\_[T]{}]{}= +(\^2,\^3). \[eq:thetaT=kappa\]
The charged current interactions remain the same as in the standard model. The neutral current Lagrangian can be written in the form [$\mathcal{L}$]{}\_[NC]{} = [$\mathcal{L}$]{}\_[QED]{} + [$\mathcal{L}$]{}\_[Z]{} + [$\mathcal{L}$]{}\_[T]{} where the first term is the usual Lagrangian of QED, [$\mathcal{L}$]{}\_[QED]{} = -e A\_J\^\_[em]{} ,J\^\_[em]{} = \_[f=1]{}\^[3]{}\_[j=1]{}\^[3]{} e\_j ( \^f\_[q,j]{}(x)\^\_[q,j]{}\^f(x) + \^f\_[l,j]{}(x)\^\_[l,j]{}\^f(x)) , \[eq:LQED\] the second one is a neutral current coupled to the ${{\mathrm{Z}}^{0}}$ boson, [$\mathcal{L}$]{}\_[Z]{} = -e Z\_( [\_[T]{}]{}J\^\_[Z]{} - [\_[T]{}]{}J\^\_[T]{}) = -e Z\_J\^\_[Z]{} + ([\_[T]{}]{}) \[eq:LZ0NC\] and the third one is the neutral current coupled to the ${{\mathrm{T}}^{0}}$ boson, [$\mathcal{L}$]{}\_[T]{} = -e T\_( [\_[T]{}]{}J\^\_[Z]{} + [\_[T]{}]{}J\^\_[T]{}) = -e T\_J\^\_[T]{} + ([\_[T]{}]{}) . \[eq:LTNC\] In [Eq.(\[eq:LQED\])]{} $e$ is the electric charge unit and $e_j$ is the electric charge of field $\psi_j$ in units of $e$. In [Eqs.(\[eq:LZ0NC\]) and (\[eq:LTNC\])]{} $J^\mu_{Z}$ is the usual neutral current, J\^\_[Z]{} = \_[f=1]{}\^[3]{}\_[j=1]{}\^[3]{} ( \^f\_[q,j]{}(x)\^\_[q,j]{}\^f(x) + \^f\_[l,j]{}(x)\^\_[l,j]{}\^f(x)) , \[eq:JZ0\] while the new neutral current has the same dependence on fermion dynamics with different coupling strength: J\^\_[T]{} = \_[f=1]{}\^[3]{}\_[j=1]{}\^[3]{} ( \^f\_[q,j]{}(x)\^\_[q,j]{}\^f(x) + \^f\_[l,j]{}(x)\^\_[l,j]{}\^f(x)) . \[eq:JT\] We can rewrite these currents as vector–axialvector currents using the non-chiral fields $\psi_f$ J\^\_[X]{} = \_[f]{} \_f(x)\^(v\_f\^[(X)]{} - a\_f\^[(X)]{}\_5)\_f(x) ,X = ZT , \[eq:v-acurrents\] with vector couplings $v_f^{(X)}$ and axialvector couplings $a_f^{(X)}$ given in [Appendix \[sec:FeynmanRules\]]{} and the summation runs over all quark and lepton flavours. Clearly, the QED current $J^\mu_{\rm em}$ can also be written using non-chiral fields in the form of [Eq.(\[eq:v-acurrents\])]{} with $v_f^{(\rm em)} = e_f$ and $a_f^{(\rm em)} = 0$.
As the dependence on the couplings and charges of the neutral currents in [Eqs.(\[eq:JZ0\]) and (\[eq:JT\])]{} are very different for different fermion fields, the only way that the standard model phenomenology is not violated by the extended model if ${\theta_{T}}$ is small, which supports the expansions used in [Eqs.(\[eq:LZ0NC\]) and (\[eq:LTNC\])]{}.
To define the perturbation theory of this model explicitly, we present the Feynman rules in [Appendix \[sec:FeynmanRules\]]{}.
Masses of the gauge bosons
--------------------------
The photon is massless, while the masses of the massive neutral bosons are M\_[Z]{} = M\_[W]{} \^[1/2]{} \[eq:MZ0\] and M\_[T]{} = M\_[W]{} \^[1/2]{} \[eq:MT\] where $M_W=\frac12 v g_L$ and we assumed $M_{T} < M_{Z}$. Indeed, in order to have $M_{Z}$ within the experimental uncertainty of the known measured value, we need ${\theta_{T}}\simeq 0$, which justifies the expansions at $\kappa=0$, M\_[Z]{} = [$\left$]{}(1 + (\^2)[$\right$]{}) and M\_[T]{} = [$\left$]{}(1 + (\^2)[$\right$]{}) M\_[Z’]{} where we used [Eq.(\[eq:thetaT=kappa\])]{} and $M_{Z'} = w g_Z'$. Thus $\tau$ can also be written as the ratio of the masses of the two massive neutral gauge bosons, = [\_]{} , \[eq:tau\] justifying our assumption on the hierarchy of masses. In fact, unless $w \gg v$, we find $M_{T} \ll M_{Z}$.
Free parameters
---------------
There are five parameters in the scalar sector, $\lambda_\phi$, $\lambda_\chi$, $\lambda$, $v$ and $w$ that has to be determined experimentally, while the values of $\mu_\phi$ and $\mu_\chi$ (at tree level) are given in [Eq.(\[eq:scalarmasses\])]{}. However, it is more convenient to use parameters that can be measured more directly, for instance, M\_h, M\_H, [\_S]{}, v = ( G\_[F]{})\^[-1/2]{} , of which we know two from measurements: one of the scalar masses and Fermi’s constant.
In addition to the neutrino Yukawa couplings (or neutrino masses and PMNS mixing parameters), there are five free parameters in the model that we choose as the mass of the new scalar particle $M_h$ or $M_H$ (the other being fixed by the mass of the Higgs boson), the scalar and vector mixing angles, the ratio of the vacuum expectation values and (essentially) the new gauge coupling: [\_S]{}, [\_[T]{}]{}, , . \[eq:freeparameters\] It can be shown [@Peli:2019xwv] that requiring stable vacuum up to the Planck scale, the Higgs particle coincides with the scalar $h$ and according to a one-loop analysis of the running scalar couplings $M_h$ falls into the range \[144,558\]GeV.
The other parameters can be expressed in terms of the free ones as follows: $w = v \tan\beta$, \_&= 1[2 v\^2]{} [$\left$]{}(M\_[h/H]{}\^2 \^2[\_S]{}+ M\_[H/h]{}\^2 \^2 [\_S]{}[$\right$]{}) ,\
\_&= 1[2 w\^2]{} [$\left$]{}( M\_[H/h]{}\^2 \^2[\_S]{}+ M\_[h/H]{}\^2 \^2[\_S]{}[$\right$]{}) ,\
&= (2[\_S]{}) (first indices are to be used if $\lambda_\phi v^2 < \lambda_\chi w^2$, the second ones otherwise) and [\_[Z]{}]{}&= ,\
&= (2[\_[T]{}]{}) (-1) = (1-\^2) [\_[T]{}]{}+ ([\_[T]{}]{}\^3) ,\
\_Z’ &= ,\_Y’ = ,’\_[ZY]{} = . \[eq:parameters\]
Possible consequences
=====================
Our hope in devising this model is to explain the established experimental observations listed in the introduction. We envisage the following scenario:
- The massive ${{\mathrm{T}}^{0}}$ vector boson is a natural candidate for WIMP dark matter if it is sufficiently stable, i.e. its mass is below the threshold of electron-positron pair production, which requires that the new force is super-weak, $\tau \sim 10^{-5}$. Such a light vector boson is not yet excluded by beam-dump experiments (see for instance [@Alexander:2016aln] and references therein). We study the most recent exclusion limits in [Sect.[\[sec:BaBar-N664\]]{}]{}. A new technology to search for electron recoils from the interaction of sub-GeV dark matter particles with electrons in silicon start to become sensitive to dark matter searches of mass as low as about 500keV [@Crisler:2018gci].
- Majorana neutrino mass terms for the right-handed neutrinos and Yukawa interactions between the left- and right-handed neutrinos and the BEH vacuum are generated by the spontaneous symmetry breaking of the scalar fields as outlined in [Sect.[\[ssec:Yukawas\]]{}]{}. This scenario provides a possible origin of neutrino oscillations and effective Majorana mass terms for the left-handed neutrinos.
- The neutrino Yukawa terms provide a source for the PMNS matrix as shown in in [Sect.[\[ssec:Yukawas\]]{}]{}, which in turn can produce leptogenesis (and hence baryogenesis).
- The vacuum of the $\chi$ scalar has a charge $z_j=-1$ (or $r_j = -1$) that may be a source of the current accelerated expansion of the universe.
- The second scalar together with the established BEH field can cause hybrid inflation.
In order that the model makes these explanations credible, we have to find answer to the following question: [*Is there any region of the parameter space of the model that is not excluded by experimental results, both established in standard model phenomenology and elsewhere?*]{} Of course, answering this question requires studies well beyond the scope of a single article. Here we shall focus on the constraints over the parameter space that can be obtained from the standard model phenomenology and in particular from the anomalous magnetic moment of the muon and searches for light neutral vector boson.
Anomalous magnetic moment of the muon {#sec:muong-2}
=====================================
There is a long standing deviation between the experimental result and predicted standard model value of the anomalous magnetic moment of the muon [@Bennett:2006fi], a\_\^[(exp)]{} - a\_\^[(SM)]{} = 268(76)10\^[-11]{} . \[eq:amuexp\] Here we assume that this difference–which will be tested by the improved precision of future experiments–is due to the effect of the new gauge boson to the anomalous magnetic moment and we estimate the allowed values for the ratio $\tan\beta$ of the vacuum expectation values and that of the mixed coupling $\gamma_{ZY}'$ and the right coupling $\gamma_Z'$, ’\_Z = = 1- = 1- . \[eq:rhoZ’\] Note that if $\rho'_Z$ were vanishing, then the new gauge boson couples only to right-handed fermions, while $\rho'_Z = 1$ implies vanishing kinetic mixing when $\tau = \kappa/\tan\beta$, so the number of free parameters reduces by one.
As the new $U(1)_Z$ sector may influence the standard model phenomenology only within the current experimental uncertainties, the new gauge coupling must be small. Therefore, the use of first order perturbation theory is justified. At one-loop accuracy, the only new contributions to the anomaly constant $a_\mu = (g_\mu-2)/2$ emerge from to the modified $Z\bar{\mu}\mu$ interaction and the new interaction $T\bar{\mu}\mu$, both presented in the Appendix. The only new Feynman graph is a triangle with the exchange of a ${{\mathrm{T}}^{0}}$ boson between the muon legs, which is formally identical to the triangle with the exchange of a ${{\mathrm{Z}}^{0}}$ boson between the muon legs as shown in [Fig.[\[fig:muon-graph\]]{}]{}. There is also a graph with the exchange of the new scalar, but that is suppressed by the super-weak new gauge coupling as compared to the (also negligible) contribution from the exchange of the Higgs boson. Consequently, the computation follows the same steps as in the case of the electroweak corrections [@Jackiw:1972jz; @Bars:1972pe; @Bardeen:1972vi; @Fujikawa:1972fe], so we present only the result for the exchange of a massive U(1) gauge boson $X$ $(X = {{\mathrm{Z}}^{0}})$ or ${{\mathrm{T}}^{0}}$):
![\[fig:muon-graph\] Feynman diagram containing the effect of the new vector boson on the anomalous magnetic moment of the muon at one loop accuracy. ](muon-graph.pdf){width="20.00000%"}
a\^[(X)]{}\_(h\_f,[\_[T]{}]{}) = , \[eq:amuZ\] with coefficients $C_X^+$ and $C_X^-$ given for both gauge bosons in the Appendix in terms of flavour dependent constants $g_f^\pm$ and $h_f^\pm$ defined in [Eq.(\[eq:gfs\])]{}. For the muon g\_\^+ = ,g\_\^- = ,h\_\^- = - ,h\_\^+ = -(1+2\_Z’) . The contribution of the ${{\mathrm{Z}}^{0}}$ boson in the standard model is recovered by setting $h_f^\pm = 0$ and ${\theta_{T}}= 0$. Thus, the complete new contribution to the $a_\mu$ in this model is given by a\_= a\^[(T+SM)]{}\_- a\^[(SM)]{}\_= a\^[(0)]{}\_(h\_f,[\_[T]{}]{}) - a\^[(0)]{}\_(0,0) + a\^[(0)]{}\_(h\_f,[\_[T]{}]{}) .
As mentioned before, the standard model phenomenology requires ${\theta_{T}}\simeq 0$, which justifies the expansion in ${\theta_{T}}$: a\_= ( + ([\_[T]{}]{},\_Z’))\^2 \[eq:Damu\] where we used [Eqs.(\[eq:MZ0\]) and (\[eq:MT\])]{} together with [Eq.(\[eq:thetaT=kappa\])]{} and the definitions in [Eqs.(\[eq:sT-cT\]) and (\[eq:kappa-tau\])]{}. According to [Ref.[@Tanabashi:2018oca]]{}, numerically 155.510\^[-11]{} . \[eq:factor\]
The deviation in [Eq.(\[eq:amuexp\])]{} is explained by the contribution in [Eq.(\[eq:Damu\])]{} if $\rho_Z'$ and $\tan\beta$ are confined to the region determined by ( 0.76848(1 + \_Z’) - 0.5 ) , and shown in [Fig.[\[fig:rho-tanbeta\]]{}]{}. We see that smaller difference between the data and the standard model prediction essentially implies larger $\tan\beta$, i.e. larger $w$.
![\[fig:rho-tanbeta\] Allowed region of parameter space assuming that the ${{\mathrm{T}}^{0}}$ boson is responsible for the deviation between the measured value of the anomalous magnetic moment of the muon and the prediction for it in the standard model. The band represents the uncertainty derived from the experimental uncertainty of the measurements. ](rho-tanbeta.pdf){width="50.00000%"}
Exclusion limits from searches for invisibly decaying light neutral vector boson {#sec:BaBar-N664}
================================================================================
There is a long list of experiments devised to search directly for light neutral vector boson. The results of the most recent ones put severe constraint in the plane of kinetic mixing parameter $\epsilon$ and mass of the boson $M_T$ [@Lees:2017lec; @NA64:2019imj] in the framework when the new gauge boson couples only by the kinetic mixing term to the fields of the standard model, which means that its coupling to the fermions is purely vector like. It is interesting to test whether the parameter space allowed in [Fig.[\[fig:rho-tanbeta\]]{}]{}, i.e. the region favoured by the muon magnetic moment anomaly at present, has any overlap with the still allowed region in these direct search experiments. As in the model presented here the ${{\mathrm{T}}^{0}}$ boson has both $v-a$ couplings to the fermions (see [Eq.(\[eq:v-acurrents\])]{}), one has to derive an effective kinetic mixing parameter ${\epsilon_{\mathrm{eff}}}$ that can be used in vector like interactions.
In the case of the BaBar experiment the search channel is an associated production of a light neutral vector boson with a photon in electron-positron annihilation [@Lees:2017lec]. The massive boson is assumed to decay invisibly to the detector, hence the signal is a single photon plus missing energy and momentum. At lowest order in perturbation theory the production channel is given by the diagrams shown in [Fig.[\[fig:BaBar\]]{}]{}.
![\[fig:BaBar\] Feynman diagrams producing a photon and a ${{\mathrm{T}}^{0}}$ boson in electron-positron annihilation. ](BaBar-graph.pdf){width="40.00000%"}
The corresponding cross section, ([$\mathrm{e}$]{}\^+[$\mathrm{e}$]{}\^- A’) = 1[2s]{} [$\mathrm{d}$]{}\_2 |[$\mathcal{M}$]{}\_[[$\mathrm{e}$]{}\^+[$\mathrm{e}$]{}\^- A’]{}|\^2 , is proportional to the square of the kinetic mixing parameter $\epsilon$. As the ${{\mathrm{T}}^{0}}$ boson couples to the fermions with $v-a$ type couplings, the cross section for the ${\ensuremath{\mathrm{e}}}^+{\ensuremath{\mathrm{e}}}^- \to \gamma {{\mathrm{T}}^{0}}$ process, ([$\mathrm{e}$]{}\^+[$\mathrm{e}$]{}\^- 0) = 1[2s]{} [$\mathrm{d}$]{}\_2 |[$\mathcal{M}$]{}\_[[$\mathrm{e}$]{}\^+[$\mathrm{e}$]{}\^- 0]{}|\^2 , is proportional to $\Big(v_{\ensuremath{\mathrm{e}}}^{(T)}\Big)^2 + \Big(a_{\ensuremath{\mathrm{e}}}^{(T)}\Big)^2$. Hence, we can define the effective kinetic mixing parameter for the ${\ensuremath{\mathrm{e}}}^+{\ensuremath{\mathrm{e}}}^- \to \gamma {{\mathrm{T}}^{0}}$ production channel as [\_]{}= = = .
In the case of the NA64 experiment the production channel corresponds to the crossing of the BaBar production channel [@NA64:2019imj], $({\ensuremath{\mathrm{e}}}^+{\ensuremath{\mathrm{e}}}^- \to \gamma A') \longrightarrow ({\ensuremath{\mathrm{e}}}^- \gamma^* \to {\ensuremath{\mathrm{e}}}^- A')$ where the virtual photon emerges from the nucleus on which the electron scatters. As a result, the effective kinetic mixing parameter is the same as in the BaBar experiment.
Using Eqs. (\[eq:sW-cW\]), (\[eq:kappa-tau\]) and (\[eq:tau\]), we can translate the preferred region in [Fig.[\[fig:rho-tanbeta\]]{}]{} into the preferred region in the ${\epsilon_{\mathrm{eff}}}- M_T$ plane, shown in [Fig.[\[fig:eps-mT\]]{}]{}. We see that the direct searches for the light neutral gauge boson allows for the interpretation of the deviation $a_{\mu}^{(\rm exp)} - a_{\mu}^{(\rm SM)} = 268(76)\cdot 10^{-11}$ with the existence of the ${{\mathrm{T}}^{0}}$ boson only if it has mass below the electron-pair threshold, which supports our previous assumption about the smallness of the mass $M_T$, and consequently of the coupling $g_Z'$, i.e. the new force is super weak.
![\[fig:eps-mT\] Region in the ${\epsilon_{\mathrm{eff}}}- M_T$ plane favoured by the muon magnetic moment anomaly at several values of the $\gamma_Z'$ coupling vs. the exclusion limits set by the BaBar [@NA64:2019imj] and NA64 [@Lees:2017lec] experiments.](eps-mT.pdf){width="70.00000%"}
Conclusions
===========
In this paper we collected the well established experimental observations that cannot be explained by the standard model of particle interactions. We have then proposed an anomaly free extension by a $U(1)_Z$ gauge group, which is the simplest possible model. We also assumed the existence of a new complex scalar field with $Z$-charge only (i.e. neutral with respect to the standard model interactions) and three right-handed neutrinos. In order to fix the $Z$-charges of the particle spectrum we assumed that the left- and right-handed neutrinos have opposite $Z$-charges. Thus such a model predicts the existence of (i) a massive neutral vector boson, (ii) a massive scalar particle and (iii) three massive right-handed neutrinos. The left-handed neutrinos remain massless as in the standard model, but their Yukawa interactions with the BEH field and the right-handed neutrinos provide a field theoretical basis for explaining neutrino oscillations and predict effective Majorana masses for the propagating mass eigenstates.
We have discussed how the new neutral gauge field $Z^\mu$ mixes with those of the standard model ($B^\mu$ and $W_3^\mu$) and argued that the mixing results in a new vector boson ${{\mathrm{T}}^{0}}$ of a small mass related to the small new gauge coupling and small mixing with the standard model vector fields. We also presented the Feynman rules of the model in unitary gauge and collected the new free parameters.
In order that the predictions of the model be credible, we have to answer whether there is any region of the parameter space that is not excluded by experimental results established in standard model phenomenology or elsewhere. To answer such a question with satisfaction, studies well beyond the scope of a single article are needed, which forecasts an exciting research project. As a first check, we computed the contribution of the new vector boson to the anomalous magnetic moment of the muon and used the difference $\Delta a_\mu$ between the prediction of the standard model and the measured values to constrain the parameter space of the model. Comparing the allowed parameter space to the exclusion limits set by direct searches for invisible decays of dark photons by the NA64 experiment, we found that the present value of $\Delta a_\mu$ allows for the existence of the ${{\mathrm{T}}^{0}}$ boson if its mass is below the threshold of electron-positron pair production. An analysis of the ultraviolet behaviour of the scalar couplings is presented in [Ref.[@Peli:2019xwv]]{}.
I am grateful G. Cynolter, D. Horváth, S. Iwamoto, A. Kardos and S. Katz for their constructive criticism on the manuscript. This work was supported by grant K 125105 of the National Research, Development and Innovation Fund in Hungary.
[10]{}
S. Weinberg, [*[A Model of Leptons]{}*]{}, [*Phys. Rev. Lett.*]{} [**19**]{} (1967) 1264–1266. Collaboration, G. Aad [*et. al.*]{}, [*[Observation of a new particle in the search for the Standard Model Higgs boson with the ATLAS detector at the LHC]{}*]{}, [*Phys. Lett.*]{} [**B716**]{} (2012) 1–29 \[[[1207.7214]{}](http://arXiv.org/abs/1207.7214)\]. Collaboration, S. Chatrchyan [*et. al.*]{}, [*[Observation of a new boson at a mass of 125 GeV with the CMS experiment at the LHC]{}*]{}, [ *Phys. Lett.*]{} [**B716**]{} (2012) 30–61 \[[[1207.7235]{}](http://arXiv.org/abs/1207.7235)\]. Collaboration, M. Tanabashi [*et. al.*]{}, [ *[Review of Particle Physics]{}*]{}, [*Phys. Rev.*]{} [**D98**]{} (2018), no. 3 030001. https://twiki.cern.ch/twiki/bin/view/AtlasPublic/StandardModelPublicResults
https://twiki.cern.ch/twiki/bin/view/CMSPublic/PhysicsResultsCombined
F. Bezrukov and M. Shaposhnikov, [*[Standard Model Higgs boson mass from inflation: Two loop analysis]{}*]{}, [*JHEP*]{} [**07**]{} (2009) 089 \[[[0904.1537]{}](http://arXiv.org/abs/0904.1537)\]. G. Degrassi, S. Di Vita, J. Elias-Miro, J. R. Espinosa, G. F. Giudice, G. Isidori and A. Strumia, [*[Higgs mass and vacuum stability in the Standard Model at NNLO]{}*]{}, [*JHEP*]{} [**08**]{} (2012) 098 \[[[1205.6497]{}](http://arXiv.org/abs/1205.6497)\]. F. Englert and R. Brout, [*[Broken Symmetry and the Mass of Gauge Vector Mesons]{}*]{}, [*Phys. Rev. Lett.*]{} [**13**]{} (1964) 321–323. \[,157(1964)\]. P. W. Higgs, [*[Broken Symmetries and the Masses of Gauge Bosons]{}*]{}, [ *Phys. Rev. Lett.*]{} [**13**]{} (1964) 508–509. \[,160(1964)\]. R. M. Schabinger and J. D. Wells, [*[A Minimal spontaneously broken hidden sector and its impact on Higgs boson physics at the large hadron collider]{}*]{}, [*Phys. Rev.*]{} [**D72**]{} (2005) 093007 \[[[hep-ph/0509209]{}](http://arXiv.org/abs/hep-ph/0509209)\]. M. Pospelov, A. Ritz and M. B. Voloshin, [*[Secluded WIMP Dark Matter]{}*]{}, [*Phys. Lett.*]{} [**B662**]{} (2008) 53–61 \[[[0711.4866]{}](http://arXiv.org/abs/0711.4866)\]. L. Basso, A. Belyaev, S. Moretti and C. H. Shepherd-Themistocleous, [ *[Phenomenology of the minimal B-L extension of the Standard model: Z’ and neutrinos]{}*]{}, [*Phys. Rev.*]{} [**D80**]{} (2009) 055030 \[[[0812.4313]{}](http://arXiv.org/abs/0812.4313)\]. T. Appelquist, B. A. Dobrescu and A. R. Hopper, [*[Nonexotic neutral gauge bosons]{}*]{}, [*Phys. Rev.*]{} [**D68**]{} (2003) 035012 \[[[hep-ph/0212073]{}](http://arXiv.org/abs/hep-ph/0212073)\]. M. Duch, B. Grzadkowski and M. McGarrie, [*[A stable Higgs portal with vector dark matter]{}*]{}, [*JHEP*]{} [**09**]{} (2015) 162 \[[[1506.08805]{}](http://arXiv.org/abs/1506.08805)\]. Z. Péli and Z. Trócsányi, [*[Stability of the vacuum as constraint on $U$(1) extensions of the standard model]{}*]{}, [[1902.02791]{}](http://arXiv.org/abs/1902.02791). L. Basso, [*[Phenomenology of the minimal B-L extension of the Standard Model at the LHC]{}*]{}. PhD thesis, Southampton U., 2011. [[1106.4462]{}](http://arXiv.org/abs/1106.4462). M. Blennow and E. Fernandez-Martinez, [*[Parametrization of Seesaw Models and Light Sterile Neutrinos]{}*]{}, [*Phys. Lett.*]{} [**B704**]{} (2011) 223–229 \[[[1107.3992]{}](http://arXiv.org/abs/1107.3992)\]. F. del Aguila, M. Masip and M. Perez-Victoria, [*[Physical parameters and renormalization of U(1)-a x U(1)-b models]{}*]{}, [*Nucl. Phys.*]{} [**B456**]{} (1995) 531–549 \[[[ hep-ph/9507455]{}](http://arXiv.org/abs/hep-ph/9507455)\]. J. Alexander [*et. al.*]{}, [*[Dark Sectors 2016 Workshop: Community Report]{}*]{}, 2016. [[1608.08632]{}](http://arXiv.org/abs/1608.08632). Collaboration, M. Crisler, R. Essig, J. Estrada, G. Fernandez, J. Tiffenberg, M. Sofo haro, T. Volansky and T.-T. Yu, [*[SENSEI: First Direct-Detection Constraints on sub-GeV Dark Matter from a Surface Run]{}*]{}, [*Phys. Rev. Lett.*]{} [**121**]{} (2018), no. 6 061803 \[[[1804.00088]{}](http://arXiv.org/abs/1804.00088)\]. Collaboration, G. W. Bennett [*et. al.*]{}, [*[Final Report of the Muon E821 Anomalous Magnetic Moment Measurement at BNL]{}*]{}, [*Phys. Rev.*]{} [**D73**]{} (2006) 072003 \[[[hep-ex/0602035]{}](http://arXiv.org/abs/hep-ex/0602035)\]. R. Jackiw and S. Weinberg, [*[Weak interaction corrections to the muon magnetic moment and to muonic atom energy levels]{}*]{}, [*Phys. Rev.*]{} [ **D5**]{} (1972) 2396–2398. I. Bars and M. Yoshimura, [*[Muon magnetic moment in a finite theory of weak and electromagnetic interaction]{}*]{}, [*Phys. Rev.*]{} [**D6**]{} (1972) 374–376. W. A. Bardeen, R. Gastmans and B. E. Lautrup, [*[Static quantities in Weinberg’s model of weak and electromagnetic interactions]{}*]{}, [*Nucl. Phys.*]{} [**B46**]{} (1972) 319–331. K. Fujikawa, B. W. Lee and A. I. Sanda, [*[Generalized Renormalizable Gauge Formulation of Spontaneously Broken Gauge Theories]{}*]{}, [*Phys. Rev.*]{} [ **D6**]{} (1972) 2923–2943. J. P. Lees [*et al.*]{} \[BaBar Collaboration\], [*Search for Invisible Decays of a Dark Photon Produced in ${e}^{+}{e}^{-}$ Collisions at BaBar*]{}, [*Phys. Rev. Lett. *]{} [**119**]{} (2017) no.13, 131804 \[[[1702.03327]{}](http://arXiv.org/abs/1702.03327)\]. D. Banerjee [*et al.*]{}, [*Dark matter search in missing energy events with NA64*]{}, \[[[1906.00176]{}](http://arXiv.org/abs/1906.00176)\].
Feynman rules {#sec:FeynmanRules}
=============
The Feynman rules of the model are obtained from the complete Lagrangian in [Eq.(\[eqn:Lfull\])]{}. For studying the UV behaviour of the model, it is convenient to use the Feynman rules before SSB, while for low energy phenomenology the rules after SSB are needed. In this paper we concentrated only on a simple application of the latter that did not require renormalization, so rules in the unitary gauge were sufficient. The propagators of the new fields are related trivially to those of the standard fields. Thus, we present only the vertices, neglecting the rules related to QCD, which are unchanged.
Feynman rules after SSB {#feynman-rules-after-ssb .unnumbered}
-----------------------
We present the rules in unitary gauge.
- Gauge field interactions:
- The cubic gauge field interactions of fields $V_{1,\alpha}$, $V_{2,\beta}$ and $V_{3,\gamma}$ with all-incoming kinematics, $p^\mu+q^\mu+r^\mu=0$ are $\Gamma_{\alpha,\,\beta,\,\gamma}{\ensuremath{\left}}(p,q,r{\ensuremath{\right}})=
{\ensuremath{\mathrm{i}}}e C V_{\alpha,\,\beta,\,\gamma}{\ensuremath{\left}}(p,q,r{\ensuremath{\right}})$ where $$V_{\alpha,\,\beta,\,\gamma}{\ensuremath{\left}}(p,q,r{\ensuremath{\right}})=
{\ensuremath{\left}}(p-q{\ensuremath{\right}})_{\gamma}g_{\alpha\beta}
+{\ensuremath{\left}}(q-r{\ensuremath{\right}})_{\alpha}g_{\beta\gamma}
+{\ensuremath{\left}}(r-p{\ensuremath{\right}})_{\beta}g_{\alpha\gamma}\,,$$ while $C$ depends on the type of the gauge bosons participating in the interaction as follows $$\begin{array}{l|l}
\hline \hline
V_1 V_2 V_3 & C \\
\hline
\gamma W^+ W^- & 1 \\
Z W^+ W^- & \displaystyle{\frac{{\cos{\theta_{\mathrm{W}}}}}{{\sin{\theta_{\mathrm{W}}}}}}{\cos{\theta_{T}}}\\[4mm]
T W^+ W^- & \displaystyle{\frac{{\cos{\theta_{\mathrm{W}}}}}{{\sin{\theta_{\mathrm{W}}}}}}{\sin{\theta_{T}}}\\
\hline \hline
\end{array}$$
- The quartic gauge field interactions of fields $V_{1,\alpha}$, $V_{2,\beta}$, $V_{3,\gamma}$ and $V_{4,\delta}$ are\
$\Gamma_{\alpha,\,\beta,\,\gamma,\,\delta} = {\ensuremath{\mathrm{i}}}e^2 C\,
{\ensuremath{\left}}[2 g_{\alpha\beta}g_{\gamma\delta}
- g_{\alpha\gamma}g_{\beta\delta}
- g_{\alpha\delta}g_{\beta\gamma}{\ensuremath{\right}}]$ where $C$ again depends on the type of the gauge bosons participating in the interaction as follows $$\begin{array}{l|l}
\hline \hline
V_1 V_2 V_3 V_4 & C \\
\hline
W^+ W^- \gamma \gamma & -1 \\
W^+ W^- \gamma Z & -\displaystyle{\frac{{\cos{\theta_{\mathrm{W}}}}}{{\sin{\theta_{\mathrm{W}}}}}}{\cos{\theta_{T}}}\\[4mm]
W^+ W^- \gamma T & -\displaystyle{\frac{{\cos{\theta_{\mathrm{W}}}}}{{\sin{\theta_{\mathrm{W}}}}}}{\sin{\theta_{T}}}\\[4mm]
W^+ W^- Z Z & -{\ensuremath{\left}}(\displaystyle{\frac{{\cos{\theta_{\mathrm{W}}}}}{{\sin{\theta_{\mathrm{W}}}}}}{\cos{\theta_{T}}}{\ensuremath{\right}})^2 \\[4mm]
W^+ W^- T Z & -{\ensuremath{\left}}(\displaystyle{\frac{{\cos{\theta_{\mathrm{W}}}}}{{\sin{\theta_{\mathrm{W}}}}}}{\ensuremath{\right}})^2{\cos{\theta_{T}}}{\sin{\theta_{T}}}\\[4mm]
W^+ W^- T T & -{\ensuremath{\left}}(\displaystyle{\frac{{\cos{\theta_{\mathrm{W}}}}}{{\sin{\theta_{\mathrm{W}}}}}}{\sin{\theta_{T}}}{\ensuremath{\right}})^2 \\[4mm]
W^+ W^+ W^- W^- & \displaystyle{\frac{1}{({\sin{\theta_{\mathrm{W}}}})^2}} \\
\hline \hline
\end{array}$$
- Scalar interactions: We denote the standard model Higgs boson by ${\ensuremath{\mathcal{H}}}$, while the new one by ${\ensuremath{\mathcal{S}}}$.
- Cubic scalar interactions can be either of the form ${\ensuremath{\mathrm{i}}}e \frac{C}{3!} S^3$ where $C$ depends on the type of the scalar boson participating in the interaction: $$\begin{array}{l|l}
\hline \hline
S S S & \qquad C \\
\hline
{\ensuremath{\mathcal{H}}}{\ensuremath{\mathcal{H}}}{\ensuremath{\mathcal{H}}}& \displaystyle{-\frac32\frac{M_h^2 \cos^2{\theta_S}+ M_H^2\sin^2 {\theta_S}}{{\sin{\theta_{\mathrm{W}}}}M_W}}\\[4mm]
{\ensuremath{\mathcal{S}}}{\ensuremath{\mathcal{S}}}{\ensuremath{\mathcal{S}}}& \displaystyle{-\frac32\frac{M_h^2 \sin^2{\theta_S}+ M_H^2\cos^2 {\theta_S}}{{\sin{\theta_{\mathrm{W}}}}M_W \tan\beta}}\\
\hline \hline
\end{array}$$ or of the form ${\ensuremath{\mathrm{i}}}e \frac{C}{2!} S S S'$ where $C$ depends on the type of the $S$ boson participating in the interaction: $$\begin{array}{l|l}
\hline \hline
S S S' & \qquad C \\
\hline
{\ensuremath{\mathcal{H}}}{\ensuremath{\mathcal{H}}}{\ensuremath{\mathcal{S}}}& \displaystyle{-\sin{\theta_S}\cos{\theta_S}\frac{M_H^2 - M_h^2}{2 {\sin{\theta_{\mathrm{W}}}}M_W}}\\[4mm]
{\ensuremath{\mathcal{S}}}{\ensuremath{\mathcal{S}}}{\ensuremath{\mathcal{H}}}& \displaystyle{-\sin{\theta_S}\cos{\theta_S}\frac{M_H^2 - M_h^2}{2 {\sin{\theta_{\mathrm{W}}}}M_W \tan\beta}}\\
\hline \hline
\end{array}$$ Recall that $M_{H/h}$ is the mass of the heavier/lighter scalar.
- The quartic scalar interactions are either of the form ${\ensuremath{\mathrm{i}}}e^2 \frac{C}{4!}S^4$ where $C$ depends on the type of the scalar bosons participating in the interaction as follows $$\begin{array}{l|l}
\hline \hline
S S S S & \qquad C \\
\hline
{\ensuremath{\mathcal{H}}}{\ensuremath{\mathcal{H}}}{\ensuremath{\mathcal{H}}}{\ensuremath{\mathcal{H}}}& \displaystyle{-\frac34\frac{M_h^2 \cos^2{\theta_S}+ M_H^2\sin^2 {\theta_S}}{({\sin{\theta_{\mathrm{W}}}}M_W)^2}}\\[4mm]
{\ensuremath{\mathcal{S}}}{\ensuremath{\mathcal{S}}}{\ensuremath{\mathcal{S}}}{\ensuremath{\mathcal{S}}}& \displaystyle{-\frac34\frac{M_h^2 \sin^2{\theta_S}+ M_h^2\cos^2 {\theta_S}}{({\sin{\theta_{\mathrm{W}}}}M_W \tan\beta)^2}}\\
\hline \hline
\end{array}$$ or of the form ${\ensuremath{\mathrm{i}}}e^2 \frac{C}{2!\,2!}{\ensuremath{\mathcal{H}}}^2 {\ensuremath{\mathcal{S}}}^2$ where $$C = -\frac34\frac{M_h^2 - M_h^2}{({\sin{\theta_{\mathrm{W}}}}M_W)^2 \tan\beta}\,.$$
- Mixed gauge field-scalar interactions:
- The cubic gauge field-scalar interactions of fields $V_{1,\alpha}$, $V_{2,\beta}$ and $S$ are ${\ensuremath{\mathrm{i}}}e g_{\alpha\beta} C$ where $C$ depends on the types of the fields participating in the interaction as follows\
$$\begin{array}{ll}
\hline \hline
V_1 V_2 S & C \\
\hline
W^+ W^- {\ensuremath{\mathcal{H}}}& \displaystyle{\frac{M_W}{{\sin{\theta_{\mathrm{W}}}}}} \\[4mm]
Z Z {\ensuremath{\mathcal{H}}}& \displaystyle{\frac{M_W}{{\sin{\theta_{\mathrm{W}}}}}\frac{({\cos{\theta_{T}}}+\kappa{\sin{\theta_{T}}})^2}{({\cos{\theta_{\mathrm{W}}}})^2}} \\[4mm]
T T {\ensuremath{\mathcal{H}}}& \displaystyle{\frac{M_W}{{\sin{\theta_{\mathrm{W}}}}}\frac{({\sin{\theta_{T}}}-\kappa{\cos{\theta_{T}}})^2}{({\cos{\theta_{\mathrm{W}}}})^2}} \\[4mm]
T Z {\ensuremath{\mathcal{H}}}& \displaystyle{\frac{M_W}{{\sin{\theta_{\mathrm{W}}}}}\frac{({\sin{\theta_{T}}}-\kappa{\cos{\theta_{T}}})({\cos{\theta_{T}}}+\kappa{\sin{\theta_{T}}})}{({\cos{\theta_{\mathrm{W}}}})^2}} \\[4mm]
Z Z {\ensuremath{\mathcal{S}}}& \displaystyle{\frac{M_W}{{\sin{\theta_{\mathrm{W}}}}\tan\beta}\frac{(\tau {\sin{\theta_{T}}})^2}{({\cos{\theta_{\mathrm{W}}}})^2}} \\[4mm]
T T {\ensuremath{\mathcal{S}}}& \displaystyle{\frac{M_W}{{\sin{\theta_{\mathrm{W}}}}\tan\beta}\frac{(\tau {\cos{\theta_{T}}})^2}{({\cos{\theta_{\mathrm{W}}}})^2}} \\[4mm]
T Z {\ensuremath{\mathcal{S}}}& \displaystyle{-\frac{M_W}{{\sin{\theta_{\mathrm{W}}}}}}\frac{\tau^2 {\sin{\theta_{T}}}{\cos{\theta_{T}}}}{({\cos{\theta_{\mathrm{W}}}})^2} \\[4mm]
\hline \hline
\end{array}$$
- Quartic gauge field-scalar interactions $V_\alpha V_\beta SS:$ ${\ensuremath{\mathrm{i}}}e^2 g_{\alpha\beta} C$ where $C$ depends on the type of the gauge boson participating in the interaction as follows $$\begin{array}{l|l}
\hline \hline
V_1V_2SS & C \\
\hline
W^+ W^- {\ensuremath{\mathcal{H}}}{\ensuremath{\mathcal{H}}}& \displaystyle{\frac{1}{2({\sin{\theta_{\mathrm{W}}}})^2}} \\[4mm]
Z Z {\ensuremath{\mathcal{H}}}{\ensuremath{\mathcal{H}}}& \displaystyle{\frac{({\cos{\theta_{T}}}+\kappa {\sin{\theta_{T}}})^2}{2({\cos{\theta_{\mathrm{W}}}}{\sin{\theta_{\mathrm{W}}}})^2}} \\[4mm]
T T {\ensuremath{\mathcal{H}}}{\ensuremath{\mathcal{H}}}& \displaystyle{\frac{({\sin{\theta_{T}}}-\kappa {\cos{\theta_{T}}})^2}{2({\cos{\theta_{\mathrm{W}}}}{\sin{\theta_{\mathrm{W}}}})^2}} \\[4mm]
T Z {\ensuremath{\mathcal{H}}}{\ensuremath{\mathcal{H}}}& \displaystyle{\frac{({\sin{\theta_{T}}}-\kappa {\cos{\theta_{T}}})({\cos{\theta_{T}}}+\kappa {\sin{\theta_{T}}})}{2({\cos{\theta_{\mathrm{W}}}}{\sin{\theta_{\mathrm{W}}}})^2}} \\[4mm]
Z Z {\ensuremath{\mathcal{S}}}{\ensuremath{\mathcal{S}}}& \displaystyle{\frac{(\tau {\sin{\theta_{T}}})^2}{2({\cos{\theta_{\mathrm{W}}}}{\sin{\theta_{\mathrm{W}}}}\tan\beta)^2}} \\[4mm]
T T {\ensuremath{\mathcal{S}}}{\ensuremath{\mathcal{S}}}& \displaystyle{\frac{(\tau {\cos{\theta_{T}}})^2}{2({\cos{\theta_{\mathrm{W}}}}{\sin{\theta_{\mathrm{W}}}}\tan\beta)^2}} \\[4mm]
T Z {\ensuremath{\mathcal{S}}}{\ensuremath{\mathcal{S}}}& \displaystyle{-\frac{\tau^2 {\sin{\theta_{T}}}{\cos{\theta_{T}}}}{2({\cos{\theta_{\mathrm{W}}}}{\sin{\theta_{\mathrm{W}}}}\tan\beta)^2}} \\
\hline \hline
\end{array}$$
- Gauge field-fermion interactions $V_\alpha \bar{f}_i f_j$: $-{\ensuremath{\mathrm{i}}}e \gamma_\alpha (C^-P_- + C^+P_+)$ where $C^\pm$ depend on the type of the gauge boson participating in the interaction, the flavour $f$ of fermions and family number $i$ and $j$ as follows $$\begin{array}{l|l|l}
\hline \hline
V\bar{f}_if_j & C^+ & C^- \\
\hline
\gamma \bar{f}_if_j & e_f \delta_{ij} & e_f \delta_{ij} \\
Z \bar{f}_if_j & (g_f^+ {\cos{\theta_{T}}}- h_f^+ {\sin{\theta_{T}}}) \delta_{ij}& (g_f^- {\cos{\theta_{T}}}- h_f^- {\sin{\theta_{T}}}) \delta_{ij}\\
T \bar{f}_if_j & (g_f^+ {\sin{\theta_{T}}}+ h_f^+ {\cos{\theta_{T}}}) \delta_{ij}& (g_f^- {\sin{\theta_{T}}}+ h_f^- {\cos{\theta_{T}}}) \delta_{ij}\\
W^+ \bar{u}_id_j & 0 & \displaystyle{\frac{1}{\sqrt{2} {\sin{\theta_{\mathrm{W}}}}}} V_{ij}\\[4mm]
W^- \bar{d}_ju_i & 0 & \displaystyle{\frac{1}{\sqrt{2} {\sin{\theta_{\mathrm{W}}}}}} V^\dag_{ij}\\[4mm]
W^+ \bar{\nu}_i\ell_j & 0 & \displaystyle{\frac{1}{\sqrt{2} {\sin{\theta_{\mathrm{W}}}}}} \delta_{ij}\\[4mm]
W^- \bar{\ell}_j\nu_i & 0 & \displaystyle{\frac{1}{\sqrt{2} {\sin{\theta_{\mathrm{W}}}}}} \delta_{ij}\\
\hline \hline
\end{array}$$ where g\_f\^+ = - e\_f ,g\_f\^- = ,h\_f\^= \[eq:gfs\] where $R_f^+ = 1/2$ for $U^f$ or $\nu^f$, $R_f^+ = -1/2$ for $D^f$ or $\ell^f$ and $R_f^- = 0$. The vector and axial vector couplings of the ${{\mathrm{Z}}^{0}}$ boson read as $$\bsp
&
v^{(Z)}_f =
\frac12\Big(g_f^-+g_f^+\Big) {\cos{\theta_{T}}}- \frac12\Big(h_f^-+h_f^+\Big) {\sin{\theta_{T}}}\\[2mm] &\quad
= \frac{\Big(T_f^3-2({\sin{\theta_{\mathrm{W}}}})^2 e_f\Big) {\cos{\theta_{T}}}- \gamma_Z' \Big(2 \rho_Z' e_f + (1 - \rho_Z') r_f\Big){\cos{\theta_{\mathrm{W}}}}{\sin{\theta_{T}}}}{2{\sin{\theta_{\mathrm{W}}}}{\cos{\theta_{\mathrm{W}}}}}\,,
\\[2mm] &\quad
= \frac{T_f^3-2({\sin{\theta_{\mathrm{W}}}})^2 e_f}{2{\sin{\theta_{\mathrm{W}}}}{\cos{\theta_{\mathrm{W}}}}} + {{\mathrm{O}}}({\theta_{T}})
\,, \\[2mm] &
a^{(Z)}_f =
\frac12\Big(g_f^--g_f^+\Big) {\cos{\theta_{T}}}- \frac12\Big(h_f^--h_f^+\Big) {\sin{\theta_{T}}}\\[2mm] &\quad
= \frac{T_f^3 {\cos{\theta_{T}}}+ \gamma_Z' (1 - \rho_Z') r_f {\cos{\theta_{\mathrm{W}}}}{\sin{\theta_{T}}}}{2{\sin{\theta_{\mathrm{W}}}}{\cos{\theta_{\mathrm{W}}}}}
= \frac{T_f^3}{2{\sin{\theta_{\mathrm{W}}}}{\cos{\theta_{\mathrm{W}}}}} + {{\mathrm{O}}}({\theta_{T}})
\,,
\esp$$ while those of the ${{\mathrm{T}}^{0}}$ boson are & v\^[(T)]{}\_f =\
&=\
&= + ([\_[T]{}]{}) ,\
& a\^[(T)]{}\_f =\
&= - + ([\_[T]{}]{}) , \[eq:v-acouplings\] with $\rho_Z' = \gamma_{ZY}'/\gamma_Z'$ defined in [Eq.(\[eq:rhoZ’\])]{}.
- $H\bar{f}_i f_j$ vertex: ${\ensuremath{\mathrm{i}}}e C$ where $$C = - \delta_{ij} \frac{1}{2{\sin{\theta_{\mathrm{W}}}}} \frac{m_{f,i}}{M_W}\,.$$
- $S\overline{\nu^c}_{{{\mathrm{R}}},i} \nu_{{{\mathrm{R}}},j}$ vertex: ${\ensuremath{\mathrm{i}}}e C$ where $$C = - \delta_{ij} \frac{1}{2{\sin{\theta_{\mathrm{W}}}}\tan\beta} \frac{m_{\nu_{{\mathrm{R}}},i}}{M_W}\,.$$
[^1]: Baryogenesis can be explained in the standard model provided leptogenesis occurs, which is called lepto-baryogenesis.
[^2]: There are numerous other deviations of experimental results from precision predictions, but to date none has reached the significance of discovery.
[^3]: We find natural to assume one extra neutrino in each family although known observations do not exclude other possibilities.
[^4]: The Weyl spinors of $\nu_{{\mathrm{L}}}$ and $\nu_{{\mathrm{R}}}$ can be embedded into [*different*]{} Dirac spinors, leading to Majorana neutrinos, without essential changes in the model. However, the negative results of the experiments searching for neutrinoless double $\beta$-decay make the Majorana nature of neutrinos increasingly unlikely.
[^5]: Several recent observations hint at violation of lepton flavour universality, which may be taken into account in our model by choosing family dependent $Z$-charges. However, those results are controversial at present, so we neglect them.
[^6]: We distinguish the hypercharge $Y$ from the index referring to Yukawa terms using different type of letters.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Computational methods are proposed for solving a convex quadratic program (QP). Active-set methods are defined for a particular primal and dual formulation of a QP with general equality constraints and simple lower bounds on the variables. In the first part of the paper, two methods are proposed, one primal and one dual. These methods generate a sequence of iterates that are feasible with respect to the equality constraints associated with the optimality conditions of the primal-dual form. The primal method maintains feasibility of the primal inequalities while driving the infeasibilities of the dual inequalities to zero. The dual method maintains feasibility of the dual inequalities while moving to satisfy the primal inequalities. In each of these methods, the search directions satisfy a KKT system of equations formed from Hessian and constraint components associated with an appropriate column basis. The composition of the basis is specified by an active-set strategy that guarantees the nonsingularity of each set of KKT equations. Each of the proposed methods is a conventional active-set method in the sense that an initial primal- or dual-feasible point is required. In the second part of the paper, it is shown how the quadratic program may be solved as a coupled pair of primal and dual quadratic programs created from the original by simultaneously shifting the simple-bound constraints and adding a penalty term to the objective function. Any conventional column basis may be made optimal for such a primal-dual pair of shifted-penalized problems. The shifts are then updated using the solution of either the primal or the dual shifted problem. An obvious application of this approach is to solve a shifted dual QP to define an initial feasible point for the primal (or *vice versa*). The computational performance of each of the proposed methods is evaluated on a set of convex problems from the [[CUTE]{}st]{} test collection.'
author:
- 'Anders Forsgren [^1]'
- 'Philip E. Gill [^2]'
- Elizabeth Wong
bibliography:
- 'pdqpreferences.bib'
date: |
The final publication is available at Springer via\
http://dx.doi.org/10.1007/s10107-015-0966-2
title: 'PRIMAL AND DUAL ACTIVE-SET METHODS FOR CONVEX QUADRATIC PROGRAMMING'
---
Introduction
============
We consider the formulation and analysis of active-set methods for a convex quadratic program ([QP]{}) of the form $$\label{eqn-QP-defined}
\begin{array}{l@{\hspace{10pt}}l}
{{\displaystyle{\mathop{\operator@font{minimize}}}_{x\in{\mathbb{R}}^n,\,y\in{\mathbb{R}}^m}}} & {{\textstyle\frac12}}x{^T\!}Hx + {{\textstyle\frac12}}y{^T\!}M y +
c{^T\!}x \\[5pt]
{\mathop{\operator@font{subject\ to}}}& A x + M y = b, \quad x\ge 0,
\end{array}$$ where $A$, $b$, $c$, $H$ and $M$ are constant, with $H$ and $M$ symmetric positive semidefinite. In order to simplify the theoretical discussion, the inequalities of (\[eqn-QP-defined\]) involve nonnegativity constraints only. However, the methods to be described are easily extended to treat all forms of linear constraints. (Numerical results are given for problems with constraints in the form $x\subL \le x \le x{{_{\scriptscriptstyle U}}}$ and $b\subL \le Ax \le b{{_{\scriptscriptstyle U}}}$, for fixed vectors $x\subL$, $x{{_{\scriptscriptstyle U}}}$, $b\subL$ and $b{{_{\scriptscriptstyle U}}}$.) If $M =0$, the [QP]{} (\[eqn-QP-defined\]) is a conventional convex quadratic program with constraints defined in standard form. A regularized quadratic program may be obtained by defining $M=\mu I$ for some small positive parameter $\mu$. (For applications that require the solution of a regularized [QP]{} see, e.g., [@AltG99; @SWri98; @GilR13].)
Active-set methods for quadratic programming problems of the form (\[eqn-QP-defined\]) solve a sequence of linear equations that involve the $y$-variables and a subset of the $x$-variables. Each set of equations constitutes the optimality conditions associated with an equality-constrained quadratic subproblem. The goal is to predict the optimal active set, i.e., the set of constraints that are satisfied with equality, at the solution of the problem. A conventional active-set method has two phases. In the first phase, a feasible point is found while ignoring the objective function; in the second phase, the objective is minimized while feasibility is maintained. A useful feature of active-set methods is that they are well-suited for “warm starts”, where a good estimate of the optimal active set is used to start the algorithm. This is particularly useful in applications where a sequence of quadratic programs is solved, e.g., in a sequential quadratic programming method or in an ODE- or PDE-constrained problem with mesh refinement. Other applications of active-set methods for quadratic programming include mixed-integer nonlinear programming, portfolio analysis, structural analysis, and optimal control.
In Section \[sec-background\], the primal and dual forms of a convex quadratic program with constraints in standard form are generalized to include general lower bounds on both the primal and dual variables. These problems constitute a primal-dual pair that includes problem (\[eqn-QP-defined\]) and its associated dual as a special case. In Sections \[sec-primal\] and \[sec-dual\], an active-set method is proposed for each of the primal and dual forms associated with the generalized problem of Section \[sec-background\]. Both of these methods provide a sequence of iterates that are feasible with respect to the equality constraints associated with the optimality conditions of the primal-dual problem pair. The primal method maintains feasibility of the primal inequalities while driving the infeasibilities of the dual inequalities to zero. By contrast, the dual method maintains feasibility of the dual inequalities while moving to satisfy the primal inequalities. In each of these methods, the search directions satisfy a [KKT]{} system of equations formed from Hessian and constraint components associated with an appropriate column basis. The composition of the basis is specified by an active-set strategy that guarantees the nonsingularity of each set of [KKT]{} equations.
The methods formulated in Sections \[sec-primal\]–\[sec-dual\] define conventional active-set methods in the sense that an initial feasible point is required. In Section \[sec-primal-dual\], a method is proposed that solves a pair of coupled quadratic programs created from the original by simultaneously shifting the simple-bound constraints and adding a penalty term to the objective function. Any conventional column basis can be made optimal for such a primal-dual pair of shifted-penalized problems. The shifts are then updated using the solution of either the primal or the dual shifted problem. An obvious application of this idea is to solve a shifted dual [QP]{} to define an initial feasible point for the primal, or *vice-versa*. In addition to the obvious benefit of using the objective function while getting feasible, this approach provides an effective method for finding a dual-feasible point when $H$ is positive semidefinite and $M = 0$. Finding a dual-feasible point is relatively straightforward for the strictly convex case, i.e., when $H$ is positive definite. However, in the general case, the dual constraints for the phase-one linear program involve entries from $H$ as well as $A$, which complicates the formulation of the phase-one method considerably.
Finally, in Section \[sec-numerical-results\] some numerical experiments are presented for a simple [[Matlab]{}]{} implementation of a coupled primal-dual method applied to a set of convex problems from the [[CUTE]{}st]{} test collection [@GouOT15; @GouOT03].
There are a number of alternative active-set methods available for solving a [QP]{} with constraints written in the format of problem (\[eqn-QP-defined\]). Broadly speaking, these methods fall into three classes defined here in the order of increasing generality: (i) methods for strictly convex quadratic programming ($H$ symmetric positive definite) [@GolI83; @GilGMSW84; @Pow85; @Sto86; @BarB06]; (ii) methods for convex quadratic programming ($H$ symmetric positive semidefinite) [@GilMS06a; @Bol97; @Huy08; @Mae10; @Won11]; and (iii) methods for general quadratic programming (no assumptions on $H$ other than symmetry) [@Bea67; @Fle71; @GilM78a; @BeaB78; @BunK80; @Hoy86; @GilMSW90; @GilMSW91; @Gou91; @Fle00; @GouT02; @GouT02b; @Won11; @GilW15]. Of the methods specifically designed for convex quadratic programming, only the methods of Boland [@Bol97] and Wong [@Won11 Chapter 4] are dual active-set methods. Some existing active-set quadratic programming solvers include [[`QPOPT`]{}]{} [@GilMS95], [`QPSchur`]{} [@BarB06], [[`SQOPT`]{}]{} [@GilMS06a], [[`SQIC`]{}]{} [@GilW15] and [`QPA`]{} (part of the [`GALAHAD`]{} software library) [@GouOT03b].
The primal active-set method proposed in Section \[sec-primal\] is motivated by the methods of Fletcher [@Fle71], Gould [@Gou91], and Gill and Wong [@GilW15], which may be viewed as methods that extend the properties of the simplex method to general quadratic programming. At each iteration, a direction is computed that satisfies a *nonsingular* system of linear equations based on an estimate of the active set at a solution. The equations may be written in symmetric form and involve both the primal and dual variables. In this context, the purpose of the active-set strategy is not only to obtain a good estimate of the optimal active set, but also to ensure that the systems of linear equations that must be solved at each iteration are nonsingular. This strategy allows the application of any convenient linear solver for the computation of the iterates. In this paper, these ideas are applied to convex quadratic programming. The resulting sequence of iterates is the same as that generated by an algorithm for general [QP]{}, but the structure of the iteration is different, as is the structure of the linear equations that must be solved. Similar ideas are used to formulate the new dual active-set method proposed in Section \[sec-dual\].
The proposed primal, dual, and combined primal-dual methods use a “conventional” active-set approach in the sense that the constraints remain unchanged during the solution of a given [QP]{}. Alternative approaches that use a parametric active-set method have been proposed by Best [@Bes82; @Bes96], Ritter [@Rit67; @Rit81], Ferreau, Bock and Diehl [@FerBD08], Potschka [et al.]{} [@PotKBS10], and implemented in the [[`qpOASES`]{}]{} package by Ferreau [et al.]{} [@FerKPBD14]. Primal methods based on the augmented Lagrangian method have been proposed by Delbos and Gilbert [@DelG05], Chiche and Gilbert [@ChiG15], and Gilbert and Joannopoulos [@GilL15]. The use of shifts for the bounds have been suggested by Cartis and Gould [@CarG06] in the context of interior methods for linear programming. Another class of active-set methods that are convergent for strictly convex quadratic programs have been considered by Curtis, Han, and Robinson [@CurHR14].
#### Notation and terminology.
Given vectors $a$ and $b$ with the same dimension, $\min(a,b)$ is a vector with components $\min(a_i,b_i)$. The vectors $e$ and $e_j$ denote, respectively, the column vector of ones and the $j$th column of the identity matrix $I$. The dimensions of $e$, $e_i$ and $I$ are defined by the context. Given vectors $x$ and $y$, the column vector consisting of the components of $x$ augmented by the components of $y$ is denoted by $(x,y)$.
Background {#sec-background}
==========
Although the purpose of this paper is the solution of quadratic programs of the form (\[eqn-QP-defined\]), for reasons that will become evident in Section \[sec-primal-dual\], the analysis will focus on the properties of a pair of problems that may be interpreted as a primal-dual pair of [QP]{}s associated with problem (\[eqn-QP-defined\]). It is assumed throughout that the matrix ${\big(\, A \ \ M \,\big)}$ associated with the equality constraints of problem (\[eqn-QP-defined\]) has full row rank. This assumption can be made without loss of generality, as shown in Proposition \[prop-AMfullrowrank\] of the Appendix. The paper involves a number of other basic theoretical results that are subsidiary to the main presentation. The proofs of these results are relegated to the Appendix.
Formulation of the primal and dual problems
-------------------------------------------
For given constant vectors $q$ and $r$, consider the pair of convex quadratic programs $$({\mbox{\small PQP}}_{q,r})\quad
\begin{array}{ll}
{{\displaystyle{\mathop{\operator@font{minimize}}}_{x,y}}} & {{\phantom-}}{{\textstyle\frac12}}x{^T\!}H x + {{\textstyle\frac12}}y{^T\!}M y + c{^T\!}x + r{^T\!}x\\
{\mathop{\operator@font{subject\ to}}}& {\settowidth{\argwidth}{$\null- Hx + A{^T\!}y + z = c,$}
\makebox[\argwidth][l]{$\null{{\phantom-}}Ax + M y = b,$}} {\;\;}{\;\;\;}x \ge -q,
\end{array}
\]
and
\[
({\mbox{\small DQP}}_{q,r}) \quad
\begin{array}{ll}
{{\displaystyle{\mathop{\operator@font{maximize}}}_{x,y,z}}}& -{{\textstyle\frac12}}x{^T\!}H x - {{\textstyle\frac12}}y{^T\!}M y + b{^T\!}y - q{^T\!}z\\
{\mathop{\operator@font{subject\ to}}}& {\settowidth{\argwidth}{$\null- Hx + A{^T\!}y + z = c,$}
\makebox[\argwidth][l]{$\null- Hx + A{^T\!}y + z = c,$}} {\;\;}{\;\;\;}z \ge -r.
\end{array}$$ The following result gives joint optimality conditions for the triple $(x, y,z)$ such that $(x$, $y)$ is optimal for $({\mbox{\small PQP}}_{q,r})$, and $(x, y, z)$ is optimal for $({\mbox{\small DQP}}_{q,r})$. If $q$ and $r$ are zero, then $({\mbox{\small PQP}}_{0,0})$ and $({\mbox{\small DQP}}_{0,0})$ are the primal and dual problems associated with (\[eqn-QP-defined\]). For arbitrary $q$ and $r$, $({\mbox{\small PQP}}_{q,r})$ and $({\mbox{\small DQP}}_{q,r})$ are essentially the dual of each other, the difference is only an additive constant in the value of the objective function.
\[prop-regQPopt\] Let $q$ and $r$ denote constant vectors in ${\mathbb{R}}^n$. If $(x$, $y$, $z)$ is a given triple in ${\mathbb{R}}^n \times {\mathbb{R}}^m \times {\mathbb{R}}^n$, then $(x$, $y)$ is optimal for $({\mbox{\small PQP}}_{q,r})$ and $(x$, $y$, $z)$ is optimal for $({\mbox{\small DQP}}_{q,r})$ if and only if
\[eqn-optcond\] $$\begin{aligned}
H x + c - A{^T\!}y - z & = 0, \label{eqn-optgradLzero} \\
A x + M y - b & = 0, \label{eqn-optfeasprimallin} \\
x + q &\ge 0, \label{eqn-optfeasprimalbound} \\
z + r &\ge 0, \label{eqn-optnonnegmult} \\
(x+q)^T(z+r) & = 0. \label{eqn-optcomp}\end{aligned}$$
In addition, the optimal objective values satisfy ${\mathop{\operator@font optval}}({\mbox{\small PQP}}_{q,r})-{\mathop{\operator@font optval}}({\mbox{\small DQP}}_{q,r})=-q{^T\!}r$. Finally, [(\[eqn-optcond\])]{} has a solution if and only if the sets $$\big\{ (x,y,z) {:}- H x + A{^T\!}y + z = c, {\;}z \ge -r \big\} {{\;\;}\text{and}{\;\;}}
\big\{ (x,y) {:}A x + M y = b, {\;}x \ge -q \big\}$$ are both nonempty.
Let the vector of Lagrange multipliers for the constraints $Ax + M y - b=0$ be denoted by ${\skew3\widetilde y}$. Without loss of generality, the Lagrange multipliers for the bounds $x + q \ge 0$ of $({\mbox{\small PQP}}_{q,r})$ may be written in the form $z + r$, where $r$ is the given fixed vector $r$. With these definitions, a Lagrangian function $L(x,y,{\skew3\widetilde y},z)$ associated with $({\mbox{\small PQP}}_{q,r})$ is given by $$\begin{gathered}
L(x,y,{\skew3\widetilde y},z) = {{\textstyle\frac12}}x{^T\!}H x + (c + r){^T\!}x + {{\textstyle\frac12}}y{^T\!}M y -
{\skew3\widetilde y}{^T\!}(Ax + M y - b) \\ - (z + r){^T\!}(x+q).\end{gathered}$$ Stationarity of the Lagrangian with respect to $x$ and $y$ implies that
\[eqn-gradLzero\] $$\begin{aligned}
Hx + c + r - A{^T\!}{\skew3\widetilde y}- z - r &= Hx + c - A{^T\!}{\skew3\widetilde y}- z = 0, \label{eqn-gradLzeroI}\\
M y - M{\skew3\widetilde y}&= 0. \label{eqn-gradLzeroII}\end{aligned}$$
The optimality conditions for $({\mbox{\small PQP}}_{q,r})$ are then given by: (i) the feasibility conditions (\[eqn-optfeasprimallin\]) and (\[eqn-optfeasprimalbound\]); (ii) the nonnegativity conditions (\[eqn-optnonnegmult\]) for the multipliers associated with the bounds $x+q\ge 0$; (iii) the stationarity conditions (\[eqn-gradLzero\]); and (iv) the complementarity conditions (\[eqn-optcomp\]). The vector $y$ appears only in the term $M y$ of (\[eqn-optfeasprimallin\]) and (\[eqn-gradLzeroII\]). In addition, (\[eqn-gradLzeroII\]) implies that $M y = M {\skew3\widetilde y}$, in which case we may choose $y={\skew3\widetilde y}$. This common value of $y$ and ${\skew3\widetilde y}$ must satisfy (\[eqn-gradLzeroI\]), which is then equivalent to (\[eqn-optgradLzero\]). The optimality conditions (\[eqn-optcond\]) for $({\mbox{\small PQP}}_{q,r})$ follow directly.
With the substitution ${\skew3\widetilde y}= y$, the expression for the Lagrangian may be rearranged so that $$\label{eqn-Lyy}
L(x,y,y,z)
= -{{\textstyle\frac12}}x{^T\!}H x - {{\textstyle\frac12}}y{^T\!}M y + b{^T\!}y - q{^T\!}z + (Hx + c - A{^T\!}y -z){^T\!}x - q{^T\!}r.$$ Taking into account (\[eqn-gradLzero\]) for $y={\skew3\widetilde y}$, the dual objective is given by (\[eqn-Lyy\]) as $-{{\textstyle\frac12}}x{^T\!}H x - {{\textstyle\frac12}}y{^T\!}M y +
b{^T\!}y - q{^T\!}z-q{^T\!}r$, and the dual constraints are $Hx + c - A{^T\!}y - z = 0$ and $z+r\ge 0$. It follows that $({\mbox{\small DQP}}_{q,r})$ is equivalent to the dual of $({\mbox{\small PQP}}_{q,r})$, the only difference is the constant term $-q{^T\!}r$ in the objective, which is a consequence of the shift $z+r$ in the dual variables. Consequently, strong duality for convex quadratic programming implies ${\mathop{\operator@font optval}}({\mbox{\small PQP}}_{q,r})-{\mathop{\operator@font optval}}({\mbox{\small DQP}}_{q,r})=-q{^T\!}r$. In addition, the variables $x$, $y$ and $z$ satisfying (\[eqn-optcond\]) are feasible for $({\mbox{\small PQP}}_{q,r})$ and $({\mbox{\small DQP}}_{q,r})$ with the difference in the objective function value being $-q{^T\!}r$. It follows that $(x,y,z)$ is optimal for $({\mbox{\small DQP}}_{q,r})$ as well as $({\mbox{\small PQP}}_{q,r})$. Finally, feasibility of both $({\mbox{\small PQP}}_{q,r})$ and $({\mbox{\small DQP}}_{q,r})$ is both necessary and sufficient for the existence of optimal solutions.
Optimality conditions and the KKT equations
-------------------------------------------
The proposed methods are based on maintaining index sets ${\mathcal{B}}$ and ${\mathcal{N}}$ that define a partition of the index set ${\mathcal{I}}= \{1$, $2$, …, $n\}$, i.e., ${\mathcal{I}}= {\mathcal{B}}\cup {\mathcal{N}}$ with ${\mathcal{B}}\cap {\mathcal{N}}= \emptyset$. Following standard terminology, we refer to the subvectors $x{_{\scriptscriptstyle B}}$ and $x{_{\scriptscriptstyle N}}$ associated with an arbitrary $x$ as the basic and nonbasic variables, respectively. The crucial feature of ${\mathcal{B}}$ is that it defines a unique solution $(x,y,z)$ to the equations $$\label{eqn-B-N-lin}
\begin{alignedat}{2}
H x + c - A^T y - z &= 0, \quad & x{_{\scriptscriptstyle N}}+ q{_{\scriptscriptstyle N}}&= 0, \\
A x + M y - b &= 0, \quad & z{_{\scriptscriptstyle B}}+ r{_{\scriptscriptstyle B}}&= 0.
\end{alignedat}$$ For the symmetric Hessian $H$, the matrices $H{_{\scriptscriptstyle\mathit{BB}}}$ and $H{_{\scriptscriptstyle\mathit{NN}}}$ denote the subset of rows and columns of $H$ associated with the sets ${\mathcal{B}}$ and ${\mathcal{N}}$, respectively. The unsymmetric matrix of components $h_{ij}$ with $i\in{\mathcal{B}}$ and $j\in {\mathcal{N}}$ will be denoted by $H{_{\scriptscriptstyle\mathit{BN}}}$. Similarly, $A{_{\scriptscriptstyle B}}$ and $A{_{\scriptscriptstyle N}}$ denote the matrices of columns of $A$ associated with ${\mathcal{B}}$ and ${\mathcal{N}}$ respectively. With this notation, the equations (\[eqn-B-N-lin\]) may be written in partitioned form as $$\begin{alignedat}{2}
H{_{\scriptscriptstyle\mathit{BB}}}x{_{\scriptscriptstyle B}}+ H{_{\scriptscriptstyle\mathit{BN}}}x{_{\scriptscriptstyle N}}+ c{_{\scriptscriptstyle B}}- A{_{\scriptscriptstyle B}}^T y - z{_{\scriptscriptstyle\mathit{B}}{^{\null}}}&= 0, & \quad x{_{\scriptscriptstyle N}}+ q{_{\scriptscriptstyle N}}&= 0,\\
H{_{\scriptscriptstyle\mathit{BN}}}^T x{_{\scriptscriptstyle\mathit{B}}{^{\null}}}+ H{_{\scriptscriptstyle\mathit{NN}}}x{_{\scriptscriptstyle N}}+ c{_{\scriptscriptstyle N}}- A{_{\scriptscriptstyle N}}^T y - z{_{\scriptscriptstyle\mathit{N}}{^{\null}}}&= 0, & z{_{\scriptscriptstyle B}}+ r{_{\scriptscriptstyle B}}&= 0,\\
A{_{\scriptscriptstyle B}}x{_{\scriptscriptstyle B}}+ A{_{\scriptscriptstyle N}}x{_{\scriptscriptstyle N}}+ M y - b &= 0.
\end{alignedat}$$ Eliminating $x{_{\scriptscriptstyle N}}$ and $z{_{\scriptscriptstyle B}}$ from these equations using the equalities $x{_{\scriptscriptstyle N}}+ q{_{\scriptscriptstyle N}}= 0$ and $z{_{\scriptscriptstyle B}}+ r{_{\scriptscriptstyle B}}= 0$ yields the symmetric equations $$\label{eqn-xBy}
{\begin{pmatrix} H{_{\scriptscriptstyle\mathit{BB}}}& A{_{\scriptscriptstyle B}}^T \\
A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}x{_{\scriptscriptstyle B}}\\
- y \end{pmatrix}}
= {\begin{pmatrix} H{_{\scriptscriptstyle\mathit{BN}}}q{_{\scriptscriptstyle\mathit{N}}{^{\null}}}- c{_{\scriptscriptstyle B}}- r{_{\scriptscriptstyle B}}\\
A{_{\scriptscriptstyle N}}q{_{\scriptscriptstyle N}}+ b \end{pmatrix}}$$ for $x{_{\scriptscriptstyle B}}$ and $y$. It follows that (\[eqn-B-N-lin\]) has a unique solution if and only if (\[eqn-xBy\]) has a unique solution. Therefore, if ${\mathcal{B}}$ is chosen to ensure that (\[eqn-B-N-lin\]) has a unique solution, it must follow from (\[eqn-xBy\]) that the matrix $K{_{\scriptscriptstyle B}}$ such that $$\label{eqn-submin-KKT}
K{_{\scriptscriptstyle B}}= {\begin{pmatrix} H{_{\scriptscriptstyle\mathit{BB}}}& A{_{\scriptscriptstyle B}}^T \\
A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}$$ is nonsingular. Once $x{_{\scriptscriptstyle B}}$ and $y$ have been computed, the $z{_{\scriptscriptstyle N}}$-variables are given by $$\label{eqn-zN}
z{_{\scriptscriptstyle N}}= H{_{\scriptscriptstyle\mathit{BN}}}^T x{_{\scriptscriptstyle\mathit{B}}{^{\null}}}- H{_{\scriptscriptstyle\mathit{NN}}{^{\null}}}q{_{\scriptscriptstyle\mathit{N}}{^{\null}}}+ c{_{\scriptscriptstyle\mathit{N}}{^{\null}}}- A{_{\scriptscriptstyle N}}^T y.$$ As in Gill and Wong [@GilW15], any set ${\mathcal{B}}$ such that $K{_{\scriptscriptstyle B}}$ is nonsingular is referred to as a *second-order consistent basis*. Methods that impose restrictions on the eigenvalues of $K{_{\scriptscriptstyle B}}$ are known as inertia-controlling methods. (For a description of inertia-controlling methods for general quadratic programming, see, e.g., Gill [et al.]{} [@GilMSW91], and Gill and Wong [@GilW15].)
The two methods proposed in this paper, one primal, one dual, generate a sequence of iterates that satisfy the equations (\[eqn-B-N-lin\]) for some partition ${\mathcal{B}}$ and ${\mathcal{N}}$. If the conditions (\[eqn-B-N-lin\]) are satisfied, the additional requirement for fulfilling the optimality conditions of Proposition \[prop-regQPopt\] are $x{_{\scriptscriptstyle B}}+ q{_{\scriptscriptstyle B}}\ge 0$ and $z{_{\scriptscriptstyle N}}+ r{_{\scriptscriptstyle N}}\ge 0$. The primal method of Section \[sec-primal\] imposes the restriction that $x{_{\scriptscriptstyle B}}+ q{_{\scriptscriptstyle B}}\ge 0$, which implies that the sequence of iterates is primal feasible. In this case the method terminates when $z{_{\scriptscriptstyle B}}+r{_{\scriptscriptstyle B}}\ge 0$ is satisfied. Conversely, the dual method of Section \[sec-dual\] imposes dual feasibility by means of the bounds $z{_{\scriptscriptstyle N}}+ r{_{\scriptscriptstyle N}}\ge 0$ and terminates when $x{_{\scriptscriptstyle B}}+q{_{\scriptscriptstyle B}}\ge 0$.
In both methods, an iteration starts and ends with a second-order consistent basis, and comprises one or more *subiterations*. In each subiteration an index $l$ and index sets ${\mathcal{B}}$ and ${\mathcal{N}}$ are known such that ${\mathcal{B}}\cup \{ l \} \cup {\mathcal{N}}= \{1$, $2$, …, $n\}$. This partition defines a search direction $({{{{\mathit\Delta}}}x},{{{{\mathit\Delta}}}y},{{{{\mathit\Delta}}}z})$ that satisfies the identities $$\label{eqn-dxdydz}
\begin{alignedat}{2}
H {{{{\mathit\Delta}}}x}- A{^T\!}{{{{\mathit\Delta}}}y}- {{{{\mathit\Delta}}}z}&= 0, \qquad& {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle N}}&= 0, \\
A {{{{\mathit\Delta}}}x}+ M{{{{\mathit\Delta}}}y}&= 0, \qquad& {{{{\mathit\Delta}}}z}{_{\scriptscriptstyle B}}&= 0.
\end{alignedat}$$ As $l\not\in{\mathcal{B}}$ and $l\not\in{\mathcal{N}}$, these conditions imply that neither ${{{{\mathit\Delta}}}x}_l$ nor ${{{{\mathit\Delta}}}z}_l$ are restricted to be zero. The conditions ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle N}}= 0$ and ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle B}}= 0$ imply that (\[eqn-dxdydz\]) may be expressed in the partitioned-matrix form $${\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T & {\settowidth{\argwidth}{$a_l^T$}
\makebox[\argwidth][c]{$1$}} & \\[1pt]
h{_{{\scriptscriptstyle\mathit{Bl}}}}& H{_{\scriptscriptstyle\mathit{BB}}}& A{_{\scriptscriptstyle B}}^T & & \\[1pt]
h{_{{\scriptscriptstyle\mathit{Nl}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BN}}}^T & A{_{\scriptscriptstyle N}}^T & &{\settowidth{\argwidth}{$A{_{\scriptscriptstyle N}}^T$}
\makebox[\argwidth][c]{$I$}}\\[1pt]
a_l & A{_{\scriptscriptstyle B}}& -M & & \end{pmatrix}}
{\begin{pmatrix}{{\phantom-}}{{{{\mathit\Delta}}}x}_l \\
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}y}\\
- {{{{\mathit\Delta}}}z}_l \\
- {{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}\end{pmatrix}}
= {\begin{pmatrix} 0 \\[1pt] 0 \\[1pt] 0 \\[1pt] 0 \end{pmatrix}},$$ where $h{_{ll}}$ denotes the $l$th diagonal of $H$, and the column vectors $h{_{{\scriptscriptstyle\mathit{Bl}}}}$ and $h{_{{\scriptscriptstyle\mathit{Nl}}}}$ denote the column vectors of elements $h_{il}$ and $h_{jl}$ with $i\in{\mathcal{B}}$, and $j\in{\mathcal{N}}$, respectively. It follows that ${{{{\mathit\Delta}}}x}_l$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$, ${{{{\mathit\Delta}}}y}$ and ${{{{\mathit\Delta}}}z}_l$ satisfy the homogeneous equations
\[eqn-hom\] $$\begin{gathered}
{\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T & {\settowidth{\argwidth}{$a_l^T$}
\makebox[\argwidth][c]{$1$}}\\
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T & \\
a_l & A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}_l \\
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}y}\\
- {{{{\mathit\Delta}}}z}_l \end{pmatrix}}
= {\begin{pmatrix} 0 \\ 0 \\ 0 \end{pmatrix}}, \label{eqn-homa} \\
\intertext{and ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}$ is given by}
{{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}= h{_{{\scriptscriptstyle\mathit{Nl}}}{^{\null}}}{{{{\mathit\Delta}}}x}{_l{^{\null}}}+ H{_{\scriptscriptstyle\mathit{BN}}}^T {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle\mathit{B}}{^{\null}}}- A{_{\scriptscriptstyle N}}^T {{{{\mathit\Delta}}}y}. \label{eqn-homb}\end{gathered}$$
The properties of these equations are established in the next subsection.
The linear algebra framework
----------------------------
This section establishes the linear algebra framework that serves to emphasize the underlying symmetry between the primal and dual methods. It is shown that the search direction for the primal and the dual method is a nonzero solution of the homogeneous equations (\[eqn-homa\]), i.e., every direction is a nontrivial null vector of the matrix of (\[eqn-homa\]). In particular, it is shown that the null-space of (\[eqn-homa\]) has dimension one, which implies that the solution of (\[eqn-homa\]) is unique up to a scalar multiple. The length of the direction is then completely determined by fixing either ${{{{\mathit\Delta}}}x}_l =1$ or ${{{{\mathit\Delta}}}z}_l=1$. The choice of which component to fix depends on whether or not the corresponding component in a null vector of (\[eqn-homa\]) is nonzero. The conditions are stated precisely in Propositions \[prop-KBnonsing\] and \[prop-Klnonsing\] below.
The first result shows that the components ${{{{\mathit\Delta}}}x}_l$ and ${{{{\mathit\Delta}}}z}_l$ of any direction $({{{{\mathit\Delta}}}x}$, ${{{{\mathit\Delta}}}y}$, ${{{{\mathit\Delta}}}z})$ satisfying the identities (\[eqn-dxdydz\]) must be such that ${{{{\mathit\Delta}}}x}_l{{{{\mathit\Delta}}}z}_l\ge 0$.
\[prop-dotproduct\] If the vector $({{{{\mathit\Delta}}}x},{{{{\mathit\Delta}}}y},{{{{\mathit\Delta}}}z})$ satisfies the identities $$\begin{aligned}
H {{{{\mathit\Delta}}}x}- A{^T\!}{{{{\mathit\Delta}}}y}- {{{{\mathit\Delta}}}z}&= 0, \\
A {{{{\mathit\Delta}}}x}+ M{{{{\mathit\Delta}}}y}&= 0,\end{aligned}$$ then ${{{{\mathit\Delta}}}x}^T{{{{\mathit\Delta}}}z}= {{{{\mathit\Delta}}}x}{^T\!}H {{{{\mathit\Delta}}}x}+ {{{{\mathit\Delta}}}y}{^T\!}M {{{{\mathit\Delta}}}y}\ge 0$. Moreover, given an index $l$ and index sets ${\mathcal{B}}$ and ${\mathcal{N}}$ such that ${\mathcal{B}}\cup\{l\}\cup{\mathcal{N}}= \{1$, $2$, …, $n\}$ with ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle N}}=0$ and ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle B}}= 0$, then ${{{{\mathit\Delta}}}x}_l{{{{\mathit\Delta}}}z}_l ={{{{\mathit\Delta}}}x}{^T\!}H {{{{\mathit\Delta}}}x}+ {{{{\mathit\Delta}}}y}{^T\!}M {{{{\mathit\Delta}}}y}\ge 0$.
Premultiplying the first identity by ${{{{\mathit\Delta}}}x}^T$ and the second by ${{{{\mathit\Delta}}}y}^T$ gives $${{{{\mathit\Delta}}}x}{^T\!}H {{{{\mathit\Delta}}}x}- {{{{\mathit\Delta}}}x}{^T\!}A{^T\!}{{{{\mathit\Delta}}}y}- {{{{\mathit\Delta}}}x}{^T\!}{{{{\mathit\Delta}}}z}= 0, {\quad\text{and}\quad}
{{{{\mathit\Delta}}}y}{^T\!}A {{{{\mathit\Delta}}}x}+ {{{{\mathit\Delta}}}y}{^T\!}M{{{{\mathit\Delta}}}y}= 0.$$ Eliminating the term ${{{{\mathit\Delta}}}x}{^T\!}A{^T\!}{{{{\mathit\Delta}}}y}$ gives ${{{{\mathit\Delta}}}x}{^T\!}H {{{{\mathit\Delta}}}x}+ {{{{\mathit\Delta}}}y}{^T\!}M {{{{\mathit\Delta}}}y}=
{{{{\mathit\Delta}}}x}{^T\!}{{{{\mathit\Delta}}}z}$. By definition, $H$ and $M$ are symmetric positive semidefinite, which gives ${{{{\mathit\Delta}}}x}{^T\!}{{{{\mathit\Delta}}}z}\ge 0$. In particular, if ${\mathcal{B}}\cup\{l\}\cup{\mathcal{N}}=
\{ 1$, $2$, …, $n \}$, with ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle N}}= 0$ and ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle B}}=0$, it must hold that ${{{{\mathit\Delta}}}x}{^T\!}{{{{\mathit\Delta}}}z}={{{{\mathit\Delta}}}x}_l{{{{\mathit\Delta}}}z}_l\ge 0$.
The set of vectors $({{{{\mathit\Delta}}}x}_l$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$, ${{{{\mathit\Delta}}}y}$, ${{{{\mathit\Delta}}}z}_l$, ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}})$ satisfying the equations (\[eqn-hom\]) is completely characterized by the properties of the matrices $K{_{\scriptscriptstyle B}}$ and $K_l$ such that $$\label{eqn-KB-Kl-defined}
K{_{\scriptscriptstyle B}}= {\begin{pmatrix} H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \\
A{_{\scriptscriptstyle B}}& -M \end{pmatrix}} {\quad\text{and}\quad}
K_l = {\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T \\
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \\
a_l & A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}.$$ The properties are summarized by the results of the following two propositions.
\[prop-KBnonsing\] Assume that $K{_{\scriptscriptstyle B}}$ is nonsingular. Let ${{{{\mathit\Delta}}}x}_l$ be a given nonnegative scalar.
1. \[prop-KBnonsing:dx-zero\] If ${{{{\mathit\Delta}}}x}_l = 0$, then the only solution of [(\[eqn-hom\])]{} is zero, i.e., ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}=0$, ${{{{\mathit\Delta}}}y}=0$, ${{{{\mathit\Delta}}}z}_l=0$ and ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}=0$.
2. \[prop-KBnonsing:dx-pos\] If ${{{{\mathit\Delta}}}x}_l > 0$, then the quantities ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$, ${{{{\mathit\Delta}}}y}$, ${{{{\mathit\Delta}}}z}_l$ and ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}$ of [(\[eqn-hom\])]{} are unique and satisfy the equations $$\label{eqn-KB}
\begin{aligned}
{\begin{pmatrix} H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \\
A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}y}\end{pmatrix}}
&= -{\begin{pmatrix} h{_{{\scriptscriptstyle\mathit{Bl}}}}\\
a_l \end{pmatrix}} {{{{\mathit\Delta}}}x}_l, \\ {{{{\mathit\Delta}}}z}_l &= h{_{ll}{^{\null}}}{{{{\mathit\Delta}}}x}{_l{^{\null}}}+ h{_{{\scriptscriptstyle\mathit{Bl}}}}^T {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle\mathit{B}}{^{\null}}}- a_l{^T\!}{{{{\mathit\Delta}}}y},\\ {{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}&= h{_{{\scriptscriptstyle\mathit{Nl}}}{^{\null}}}{{{{\mathit\Delta}}}x}{_l{^{\null}}}+ H{_{\scriptscriptstyle\mathit{BN}}}^T {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle\mathit{B}}{^{\null}}}- A{_{\scriptscriptstyle N}}^T {{{{\mathit\Delta}}}y}. \end{aligned}$$ Moreover, either
1. \[prop-KBnonsing:dz-pos\] $K_l$ is nonsingular and ${{{{\mathit\Delta}}}z}_l>0$, or
2. \[prop-KBnonsing:dz-zero\] $K_l$ is singular and ${{{{\mathit\Delta}}}z}_l=0$, in which case it holds that ${{{{\mathit\Delta}}}y}=0$, ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}=0$, and the multiplicity of the zero eigenvalue of $K_l$ is one, with corresponding eigenvector $({{{{\mathit\Delta}}}x}_l,{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}},0)$.
Proposition \[prop-dotproduct\] implies that ${{{{\mathit\Delta}}}z}_l\ge 0$ if ${{{{\mathit\Delta}}}x}_l>0$, which implies that the statement of the proposition includes all possible values of ${{{{\mathit\Delta}}}z}_l$. The second and third blocks of the equations (\[eqn-homa\]) imply that $$\label{eqn-dxzero}
{\begin{pmatrix} h{_{{\scriptscriptstyle\mathit{Bl}}}}\\
a_l \end{pmatrix}} {{{{\mathit\Delta}}}x}_l
+ {\begin{pmatrix} H{_{\scriptscriptstyle\mathit{BB}}}& A{_{\scriptscriptstyle B}}^T \\
A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}y}\end{pmatrix}}
= {\begin{pmatrix} 0 \\
0 \end{pmatrix}}.$$ As $K{_{\scriptscriptstyle B}}$ is nonsingular by assumption, the vectors ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$ and ${{{{\mathit\Delta}}}y}$ must constitute the unique solution of (\[eqn-dxzero\]) for a given value of ${{{{\mathit\Delta}}}x}_l$. Furthermore, given ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$ and ${{{{\mathit\Delta}}}y}$, the quantities ${{{{\mathit\Delta}}}z}_l$ and ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}$ of (\[eqn-KB\]) are also uniquely defined. The specific value ${{{{\mathit\Delta}}}x}_l=0$, gives ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}=0$ and ${{{{\mathit\Delta}}}y}=0$, so that ${{{{\mathit\Delta}}}z}_l=0$ and ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}=0$. It follows that ${{{{\mathit\Delta}}}x}_l$ must be nonzero for at least one of the vectors ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$, ${{{{\mathit\Delta}}}y}$, ${{{{\mathit\Delta}}}z}_l$ or ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}$ to be nonzero.
Next it is shown that if ${{{{\mathit\Delta}}}x}_l>0$, then either (\[prop-KBnonsing:dz-pos\]) or (\[prop-KBnonsing:dz-zero\]) must hold. For (\[prop-KBnonsing:dz-pos\]), it is necessary to show that if ${{{{\mathit\Delta}}}x}_l>0$ and $K_l$ is nonsingular, then ${{{{\mathit\Delta}}}z}_l >0$. If $K_l$ is nonsingular, the homogeneous equations (\[eqn-homa\]) may be written in the form $$\begin{aligned}
\label{eqn-Klnonsing}
{\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T \\
h{_{{\scriptscriptstyle\mathit{Bl}}}}& H{_{\scriptscriptstyle\mathit{BB}}}& A{_{\scriptscriptstyle B}}^T \\
a_l & A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}_l \\
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}y}\end{pmatrix}}
&= {\begin{pmatrix} 1 \\ 0 \\ 0 \end{pmatrix}} {{{{\mathit\Delta}}}z}_l,\end{aligned}$$ which implies that ${{{{\mathit\Delta}}}x}_l$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$ and ${{{{\mathit\Delta}}}y}$ are unique for a given value of ${{{{\mathit\Delta}}}z}_l$. In particular, if ${{{{\mathit\Delta}}}z}_l =0$ then ${{{{\mathit\Delta}}}x}_l = 0$, which would contradict the assumption that ${{{{\mathit\Delta}}}x}_l >0$. If follows that ${{{{\mathit\Delta}}}z}_l$ must be nonzero. Finally, Proposition \[prop-dotproduct\] implies that if ${{{{\mathit\Delta}}}z}_l$ is nonzero and ${{{{\mathit\Delta}}}x}_l>0$, then ${{{{\mathit\Delta}}}z}_l>0$ as required.
For the first part of (\[prop-KBnonsing:dz-zero\]), it must be shown that if $K_l$ is singular, then ${{{{\mathit\Delta}}}z}_l=0$. If $K_l$ is singular, it must have a nontrivial null vector $(p_l$, $p{_{\scriptscriptstyle B}}$, $-u)$. Moreover, every null vector must have a nonzero $p_l$, because otherwise $(p{_{\scriptscriptstyle B}}$, $-u)$ would be a nontrivial null vector of $K{_{\scriptscriptstyle B}}$, which contradicts the assumption that $K{_{\scriptscriptstyle B}}$ is nonsingular. A fixed value of $p_l$ uniquely defines $p{_{\scriptscriptstyle B}}$ and $u$, which indicates that the multiplicity of the zero eigenvalue must be one. A simple substitution shows that $(p_l$, $p{_{\scriptscriptstyle B}}$, $-u$, $v_l)$ is a nontrivial solution of the homogeneous equation (\[eqn-homa\]) such that $v_l= 0$. As the subspace of vectors satisfying (\[eqn-homa\]) is of dimension one, it follows that every solution is unique up to a scalar multiple. Given the properties of the known solution $(p_l$, $p{_{\scriptscriptstyle B}}$, $-u$, $0)$, it follows that every solution $({{{{\mathit\Delta}}}x}_l$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$, $-{{{{\mathit\Delta}}}y}$, $-{{{{\mathit\Delta}}}z}_l)$ of (\[eqn-homa\]) is an eigenvector associated with the zero eigenvalue of $K_l$, with ${{{{\mathit\Delta}}}z}_l = 0$.
For the second part of (\[prop-KBnonsing:dz-zero\]), if ${{{{\mathit\Delta}}}z}_l = 0$, the homogeneous equations (\[eqn-homa\]) become $$\label{eqn-Klsing}
{\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T \\
h{_{{\scriptscriptstyle\mathit{Bl}}}}& H{_{\scriptscriptstyle\mathit{BB}}}& A{_{\scriptscriptstyle B}}^T \\
a_l & A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}_l \\
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}y}\end{pmatrix}}
= {\begin{pmatrix} 0 \\ 0 \\ 0 \end{pmatrix}}.$$ As $K_l$ is singular in (\[eqn-Klsing\]), Proposition \[propA-nonsing\] of the Appendix implies that $$\label{eqn-KlsingII}
{\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T \\
h{_{{\scriptscriptstyle\mathit{Bl}}}}& H{_{\scriptscriptstyle\mathit{BB}}}\\
a_l & A{_{\scriptscriptstyle B}}\end{pmatrix}}
{\begin{pmatrix} {{{{\mathit\Delta}}}x}_l \\
{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\end{pmatrix}}
= {\begin{pmatrix} 0 \\ 0 \\ 0 \end{pmatrix}}, {\quad\text{and}\quad}
{\begin{pmatrix} a_l^T \\
A{_{\scriptscriptstyle B}}^T \\
-M \end{pmatrix}} {{{{\mathit\Delta}}}y}= {\begin{pmatrix} 0 \\ 0 \\ 0 \end{pmatrix}}.$$ The nonsingularity of $K{_{\scriptscriptstyle B}}$ implies that ${\big(\, A{_{\scriptscriptstyle B}}\ \ -M \,\big)}$ has full row rank, in which case the second equation of (\[eqn-KlsingII\]) gives ${{{{\mathit\Delta}}}y}=0$. It follows that every eigenvector of $K_l$ associated with the zero eigenvalue has the form $({{{{\mathit\Delta}}}x}_l$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$, $0)$. It remains to show that ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}=0$. If Proposition \[propA-diag\] of the Appendix is applied to the first equation of (\[eqn-KlsingII\]), then it must hold that $${\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T \\
h{_{{\scriptscriptstyle\mathit{Bl}}}}& H{_{\scriptscriptstyle\mathit{BB}}}\\
h{_{{\scriptscriptstyle\mathit{Nl}}}}& H{_{\scriptscriptstyle\mathit{BN}}}^T \end{pmatrix}}
{\begin{pmatrix} {{{{\mathit\Delta}}}x}_l \\
{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\end{pmatrix}}
= {\begin{pmatrix} 0 \\ 0 \\ 0 \end{pmatrix}}.$$ It follows from the definition of ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}$ in (\[eqn-KB\]) that ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle\mathit{N}}{^{\null}}}= h{_{{\scriptscriptstyle\mathit{Nl}}}{^{\null}}}{{{{\mathit\Delta}}}x}{_l{^{\null}}}+ H{_{\scriptscriptstyle\mathit{BN}}}^T {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle\mathit{B}}{^{\null}}}- A{_{\scriptscriptstyle N}}^T {{{{\mathit\Delta}}}y}= 0$, which completes the proof.
\[prop-Klnonsing\] Assume that $K_l$ is nonsingular. Let ${{{{\mathit\Delta}}}z}_l$ be a given nonnegative scalar.
1. \[prop-Klnonsing:dz-zero\] If ${{{{\mathit\Delta}}}z}_l = 0$, then the only solution of [(\[eqn-hom\])]{} is zero, i.e., ${{{{\mathit\Delta}}}x}_l=0$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}=0$, ${{{{\mathit\Delta}}}y}=0$ and ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}=0$.
2. \[prop-Klnonsing:dz-pos\] If ${{{{\mathit\Delta}}}z}_l > 0$, then the quantities ${{{{\mathit\Delta}}}x}_l$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$, ${{{{\mathit\Delta}}}y}$ and ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}$ of [(\[eqn-hom\])]{} are unique and satisfy the equations
\[eqn-Klsolve\] $$\begin{aligned}
{\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T \\
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \\
a_l & A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}
{\renewcommand{\arraystretch}{0.95} \left({\mskip -9.5 mu}\begin{array}{l} {{\phantom-}}{{{{\mathit\Delta}}}x}_l \\
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}y}\end{array}{\mskip -9.5 mu}\right)}
&= {\begin{pmatrix} 1 \\
0 \\
0 \end{pmatrix}} {{{{\mathit\Delta}}}z}_l, \label{eqn-KBmain} \\
{{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}&= H{_{{\scriptscriptstyle\mathit{Nl}}}{^{\null}}}{{{{\mathit\Delta}}}x}{_l{^{\null}}}+ H{_{\scriptscriptstyle\mathit{BN}}}^T {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle\mathit{B}}{^{\null}}}- A{_{\scriptscriptstyle N}}^T {{{{\mathit\Delta}}}y}. \label{eqn-dzN}\end{aligned}$$
Moreover, either
1. \[prop-Klnonsing:dx-pos\] $K{_{\scriptscriptstyle B}}$ is nonsingular and ${{{{\mathit\Delta}}}x}_l>0$, or
2. \[prop-Klnonsing:dx-zero\] $K{_{\scriptscriptstyle B}}$ is singular and ${{{{\mathit\Delta}}}x}_l=0$, in which case, it holds that ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}=0$ and the multiplicity of the zero eigenvalue of $K{_{\scriptscriptstyle B}}$ is one, with corresponding eigenvector $(0,{{{{\mathit\Delta}}}y})$.
In Proposition \[prop-dotproduct\] it is established that ${{{{\mathit\Delta}}}x}_l\ge 0$ if ${{{{\mathit\Delta}}}z}_l>0$, which implies that the statement of the proposition includes all possible values of ${{{{\mathit\Delta}}}x}_l$.
It follows from (\[eqn-homa\]) that ${{{{\mathit\Delta}}}x}_l$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$, and ${{{{\mathit\Delta}}}y}$ must satisfy the equations $$\label{eqn-dzzero}
{\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T \\
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \\
a_l & A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}_l \\
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}y}\end{pmatrix}}
= {\begin{pmatrix} {{{{\mathit\Delta}}}z}_l \\
0 \\
0 \end{pmatrix}}.$$ Under the given assumption that $K_l$ is nonsingular, the vectors ${{{{\mathit\Delta}}}x}_l$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$ and ${{{{\mathit\Delta}}}y}$ are uniquely determined by (\[eqn-dzzero\]) for a fixed value of ${{{{\mathit\Delta}}}z}_l$. In addition, once ${{{{\mathit\Delta}}}x}_l$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$ and ${{{{\mathit\Delta}}}y}$ are defined, ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}$ is uniquely determined by (\[eqn-dzN\]). It follows that if ${{{{\mathit\Delta}}}z}_l=0$, then ${{{{\mathit\Delta}}}x}_l=0$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}=0$, ${{{{\mathit\Delta}}}y}=0$ and ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}=0$.
It remains to show that if ${{{{\mathit\Delta}}}z}_l>0$, then either (\[prop-Klnonsing:dx-pos\]) or (\[prop-Klnonsing:dx-zero\]) must hold. If $K{_{\scriptscriptstyle B}}$ is singular, then Proposition \[propA-nonsing\] of the Appendix implies that there must exist $u$ and $v$ such that $${\begin{pmatrix} H{_{\scriptscriptstyle\mathit{BB}}}\\
A{_{\scriptscriptstyle B}}\end{pmatrix}} u
= {\begin{pmatrix} 0 \\
0 \end{pmatrix}} {\quad\text{and}\quad}
{\begin{pmatrix} A{_{\scriptscriptstyle B}}^T \\
-M \end{pmatrix}} v
= {\begin{pmatrix}0 \\ 0 \end{pmatrix}}.$$ Proposition \[propA-diag\] of the Appendix implies that the vector $u$ must also satisfy $h{_{{\scriptscriptstyle\mathit{Bl}}}}^T u=0$. If $u$ is nonzero, then $(0,u,0)$ is a nontrivial null vector for $K_l$, which contradicts the assumption that $K_l$ is nonsingular. It follows that ${\big(\, H{_{\scriptscriptstyle\mathit{BB}}}\ \ A{_{\scriptscriptstyle B}}^T \,\big)}$ has full row rank and the singularity of $K{_{\scriptscriptstyle B}}$ must be caused by dependent rows in ${\big(\, A{_{\scriptscriptstyle B}}\ \ -M \,\big)}$. The nonsingularity of $K_l$ implies that ${\big(\, a_l \ \ A{_{\scriptscriptstyle B}}\ \ -M\,\big)}$ has full row rank and there must exist a vector $v$ such that $v{^T\!}a_l\ne 0$, $v{^T\!}A{_{\scriptscriptstyle B}}=0$ and $v{^T\!}M=0$. If $v$ is scaled so that $v{^T\!}a_l=-{{{{\mathit\Delta}}}z}_l$, then $(0,0,-v)$ must be a solution of (\[eqn-dzzero\]). It follows that ${{{{\mathit\Delta}}}x}_l=0$, $v = {{{{\mathit\Delta}}}y}$, and $(0,{{{{\mathit\Delta}}}y})$ is an eigenvector of $K{_{\scriptscriptstyle B}}$ associated with a zero eigenvalue. The nonsingularity of $K_l$ implies that $v$ is unique given the value of the scalar ${{{{\mathit\Delta}}}z}_l$, and hence the zero eigenvalue has multiplicity one.
Conversely, ${{{{\mathit\Delta}}}x}_l=0$ implies that $({{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}},{{{{\mathit\Delta}}}y})$ is a null vector for $K{_{\scriptscriptstyle B}}$. However, if $K{_{\scriptscriptstyle B}}$ is nonsingular, then the vector is zero, contradicting (\[eqn-KBmain\]). It follows that $K{_{\scriptscriptstyle B}}$ must be singular.
A Primal Active-Set Method for Convex QP {#sec-primal}
========================================
In this section a primal-feasible method for convex [QP]{} is formulated. Each iteration begins and ends with a point $(x,y,z)$ that satisfies the conditions $$\label{eqn-primalxyz}
\begin{alignedat}{3}
H x + c - A{^T\!}y -z &= 0, & \qquad x{_{\scriptscriptstyle N}}+ q{_{\scriptscriptstyle N}}&= 0, & \qquad x{_{\scriptscriptstyle B}}+ q{_{\scriptscriptstyle B}}&\ge 0, \\
Ax + M y - b &= 0, & \qquad z{_{\scriptscriptstyle B}}+ r{_{\scriptscriptstyle B}}&= 0,
\end{alignedat}$$ for appropriate second-order consistent bases. The purpose of the iterations is to drive $(x,y,z)$ to optimality by driving the dual variables to feasibility (i.e., by driving the negative components of $z{_{\scriptscriptstyle N}}+
r{_{\scriptscriptstyle N}}$ to zero). Methods for finding ${\mathcal{B}}$ and ${\mathcal{N}}$ at the initial point are discussed in Section \[sec-primal-dual\].
An iteration consists of a group of one or more consecutive *subiterations* during which a specific dual variable is made feasible. The first subiteration is called the *base* subiteration. In some cases only the base subiteration is performed, but, in general, additional *intermediate* subiterations are required.
At the start of the base subiteration, an index $l$ in the nonbasic set ${\mathcal{N}}$ is identified such that $z_l+r_l < 0$. The idea is to remove the index $l$ from ${\mathcal{N}}$ (i.e., ${\mathcal{N}}\gets {\mathcal{N}}\setminus \{l\}$) and attempt to increase the value of $z_l+r_l$ by taking a step along a primal-feasible direction $({{{{\mathit\Delta}}}x}_l$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$, ${{{{\mathit\Delta}}}y}$, ${{{{\mathit\Delta}}}z}_l)$. The removal of $l$ from ${\mathcal{N}}$ implies that ${\mathcal{B}}\cup\{l\}\cup{\mathcal{N}}= \{ 1$, $2$, …, $n\}$ with ${\mathcal{B}}$ second-order consistent. This implies that $K{_{\scriptscriptstyle B}}$ is nonsingular and the (unique) search direction may be computed as in (\[eqn-KB\]) with ${{{{\mathit\Delta}}}x}_l=1$.
If ${{{{\mathit\Delta}}}z}_l > 0$, the step ${\alpha{_{\ast}}{^{\null}}}= -(z_l+r_l)/{{{{\mathit\Delta}}}z}_l$ gives $z_l+{\alpha{_{\ast}}{^{\null}}}{{{{\mathit\Delta}}}z}_l + r_l = 0$. Otherwise, ${{{{\mathit\Delta}}}z}_l=0$, and there is no finite value of $\alpha$ that will drive $z_l + \alpha {{{{\mathit\Delta}}}z}_l + r_l$ to its bound, and ${\alpha{_{\ast}}{^{\null}}}$ is defined to be $+\infty$. Proposition \[propA-dirs\] of the Appendix implies that the case ${{{{\mathit\Delta}}}z}_l
=0$ corresponds to the primal objective function being linear and decreasing along the search direction.
Even if ${{{{\mathit\Delta}}}z}_l$ is positive, it is not always possible to take the step ${\alpha{_{\ast}}{^{\null}}}$ and remain primal feasible. A positive step in the direction $({{{{\mathit\Delta}}}x}_l$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$, ${{{{\mathit\Delta}}}y}$, ${{{{\mathit\Delta}}}z}_l)$ must increase $x_l$ from its bound, but may decrease some of the basic variables. This makes it necessary to limit the step to ensure that the primal variables remain feasible. The largest step length that maintains primal feasibility is given by $${\alpha{_{\operator@font{max}}}}= \min_{ i {:}{{{{\mathit\Delta}}}x}_i < 0} {\;}\frac{x_i + q_i}{-{{{{\mathit\Delta}}}x}_i}.$$ If ${\alpha{_{\operator@font{max}}}}$ is finite, this value gives $x_k + {\alpha{_{\operator@font{max}}}}{{{{\mathit\Delta}}}x}_k + q_k =
0$, where $k$ is the index $k = {\mathop{\operator@font{argmin}}}_{i {:}{{{{\mathit\Delta}}}x}_i < 0}{\;}(x_i +
q_i)/(-{{{{\mathit\Delta}}}x}_i)$. The overall step length is then $\alpha =
\min\big({\alpha{_{\ast}}{^{\null}}},{\alpha{_{\operator@font{max}}}}\big)$. An infinite value of $\alpha$ indicates that the primal problem $({\mbox{\small PQP}}_{q,r})$ is unbounded, or, equivalently, that the dual problem $({\mbox{\small DQP}}_{q,r})$ is infeasible. In this case, the algorithm is terminated. If the step $\alpha = {\alpha{_{\ast}}{^{\null}}}$ is taken, then $z_l+ \alpha {{{{\mathit\Delta}}}z}_l+r_l=0$, the subiterations are terminated with no intermediate subiterations and ${\mathcal{B}}\gets {\mathcal{B}}\cup \{ l \}$. Otherwise, $\alpha = {\alpha{_{\operator@font{max}}}}$, and the basic and nonbasic sets are updated as ${\mathcal{B}}\gets {\mathcal{B}}\setminus \{k\}$ and ${\mathcal{N}}\gets {\mathcal{N}}\cup
\{k\}$ giving a new partition ${\mathcal{B}}\cup\{l\}\cup{\mathcal{N}}= \{ 1$, $2$, …, $n\}$. In order to show that the equations associated with the new partition are well-defined, it is necessary to show that allowing $z_k$ to move does not give a singular $K_l$. Proposition \[propA-primalnonsing\] of the Appendix shows that the submatrix $K_l$ associated with the updated ${\mathcal{B}}$ and ${\mathcal{N}}$ is nonsingular for the cases ${{{{\mathit\Delta}}}z}_l>0$ and ${{{{\mathit\Delta}}}z}_l=0$.
Because the removal of $k$ from ${\mathcal{B}}$ does not alter the nonsingularity of $K_l$, it is possible to add $l$ to ${\mathcal{B}}$ and thereby define a unique solution of the system (\[eqn-B-N-lin\]). However, if $z_l+r_l<0$, additional intermediate subiterations are required to drive $z_l + r_l$ to zero. In each of these subiterations, the search direction is computed by choosing ${{{{\mathit\Delta}}}z}_l=1$ in Proposition \[prop-Klnonsing\]. The step length ${\alpha{_{\ast}}{^{\null}}}$ is given by ${\alpha{_{\ast}}{^{\null}}}= -(z_l+r_l)/{{{{\mathit\Delta}}}z}_l$ as in the base subiteration above, but now ${\alpha{_{\ast}}{^{\null}}}$ is always finite because ${{{{\mathit\Delta}}}z}_l=1$. Similar to the base subiteration, if no constraint is added, then $z_l+{\alpha{_{\ast}}{^{\null}}}{{{{\mathit\Delta}}}z}_l+r_l=0$. Otherwise, the index of another blocking variable $k$ is moved from ${\mathcal{B}}$ to ${\mathcal{N}}$. Proposition \[propA-primalnonsing\] implies that the updated matrix $K_l$ is nonsingular at the end of an intermediate subiteration. As a consequence, the intermediate subiterations may be repeated until $z_l +
r_l$ is driven to zero.
At the end of the base subiteration or after the intermediate subiterations are completed, it must hold that $z_l + r_l = 0$ and the final $K_l$ is nonsingular. This implies that a new iteration may be initiated with the new basic set ${\mathcal{B}}\cup \{l\}$ defining a nonsingular $K{_{\scriptscriptstyle B}}$.
The primal active-set method is summarized in Algorithm \[alg-primalqp\] below. The convergence properties of Algorithm \[alg-primalqp\] are established in Section \[sec-primal-dual\], which concerns a general primal algorithm that includes Algorithm \[alg-primalqp\] as a special case.
Find $(x,y,z)$ satisfying conditions (\[eqn-primalxyz\]) for some second-order consistent basis ${\mathcal{B}}$;
${\mathcal{N}}\gets {\mathcal{N}}\setminus \{l\}$; ; ; ${\mathcal{B}}\gets {\mathcal{B}}\cup \{l\}$;
${{{{\mathit\Delta}}}x}_l \gets 1$; Solve ${\begin{pmatrix} H{_{\scriptscriptstyle\mathit{BB}}}& A{_{\scriptscriptstyle B}}^T \\
A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}y}\end{pmatrix}}
= -{\begin{pmatrix} h{_{{\scriptscriptstyle\mathit{Bl}}}}\\
a_l \end{pmatrix}}$; ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}\gets h{_{{\scriptscriptstyle\mathit{Nl}}}{^{\null}}}{{{{\mathit\Delta}}}x}{_l{^{\null}}}+ H{_{\scriptscriptstyle\mathit{BN}}}^T {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle\mathit{B}}{^{\null}}}- A{_{\scriptscriptstyle N}}^T {{{{\mathit\Delta}}}y}$; ${{{{\mathit\Delta}}}z}_l \gets h{_{ll}{^{\null}}}{{{{\mathit\Delta}}}x}{_l{^{\null}}}+ h{_{{\scriptscriptstyle\mathit{Bl}}}}^T {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle\mathit{B}}{^{\null}}}- a_l^T {{{{\mathit\Delta}}}y}$; ${\alpha{_{\ast}}{^{\null}}}\gets -(z_l+r_l)/{{{{\mathit\Delta}}}z}_l$; ${\displaystyle}{\alpha{_{\operator@font{max}}}}\gets \min_{i {:}{{{{\mathit\Delta}}}x}_i < 0}{\;\;}(x_i + q_i)/(-{{{{\mathit\Delta}}}x}_i)$; ${\displaystyle}k \gets {\mathop{\operator@font{argmin}}}_{i {:}{{{{\mathit\Delta}}}x}_i < 0}{\;\;}(x_i + q_i)/(-{{{{\mathit\Delta}}}x}_i)$; $\alpha \gets \min\big({\alpha{_{\ast}}{^{\null}}},\alpha{_{\operator@font{max}}}\big)$; ${\mathbf{stop}}$; $x_l \gets x_l + \alpha {{{{\mathit\Delta}}}x}_l$; $x{_{\scriptscriptstyle B}}\gets x{_{\scriptscriptstyle B}}+ \alpha {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$; $y \gets y + \alpha {{{{\mathit\Delta}}}y}$; $z_l \gets z_l + \alpha {{{{\mathit\Delta}}}z}_l$; $z{_{\scriptscriptstyle N}}\gets z{_{\scriptscriptstyle N}}+ \alpha {{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}$; ${\mathcal{B}}\gets {\mathcal{B}}\setminus \{k\}$; ${\mathcal{N}}\gets {\mathcal{N}}\cup \{k\}$; ;
${{{{\mathit\Delta}}}z}_l \gets 1$; Solve ${\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T \\
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \\
a_l & A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}
{\renewcommand{\arraystretch}{0.95} \left({\mskip -9.5 mu}\begin{array}{l} {{\phantom-}}{{{{\mathit\Delta}}}x}_l \\
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}y}\end{array}{\mskip -9.5 mu}\right)}
= {\begin{pmatrix} 1 \\
0 \\
0 \end{pmatrix}}$; ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}\gets H{_{{\scriptscriptstyle\mathit{Nl}}}{^{\null}}}{{{{\mathit\Delta}}}x}{_l{^{\null}}}+ H{_{\scriptscriptstyle\mathit{BN}}}^T {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}- A{_{\scriptscriptstyle N}}^T {{{{\mathit\Delta}}}y}$; ${\displaystyle}{\alpha{_{\ast}}{^{\null}}}\gets -(z_l+r_l)$; ${\displaystyle}\alpha{_{\operator@font{max}}}\gets \min_{i {:}{{{{\mathit\Delta}}}x}_i < 0} {\;\;}(x_i + q_i)/(-{{{{\mathit\Delta}}}x}_i)$; ${\displaystyle}k \gets {\mathop{\operator@font{argmin}}}_{i {:}{{{{\mathit\Delta}}}x}_i < 0}{\;\;}(x_i + q_i)/(-{{{{\mathit\Delta}}}x}_i)$; $\alpha \gets \min\big({\alpha{_{\ast}}{^{\null}}},\alpha{_{\operator@font{max}}}\big)$; $x_l \gets x_l + \alpha {{{{\mathit\Delta}}}x}_l$; $x{_{\scriptscriptstyle B}}\gets x{_{\scriptscriptstyle B}}+ \alpha {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$; $y \gets y + \alpha {{{{\mathit\Delta}}}y}$; $z_l \gets z_l + \alpha {{{{\mathit\Delta}}}z}_l$; $z{_{\scriptscriptstyle N}}\gets z{_{\scriptscriptstyle N}}+ \alpha {{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}$; ${\mathcal{B}}\gets {\mathcal{B}}\setminus \{k\}$; ${\mathcal{N}}\gets {\mathcal{N}}\cup \{k\}$; ;
A Dual Active-Set Method for Convex QP {#sec-dual}
======================================
Each iteration of the dual active-set method begins and ends with a point $(x,y,z)$ that satisfies the conditions $$\label{eqn-dualxyz}
\begin{alignedat}{2}
H x + c - A{^T\!}y -z &= 0, & \qquad x{_{\scriptscriptstyle N}}+ q{_{\scriptscriptstyle N}}&= 0, \\
Ax + M y - b &= 0, & \qquad z{_{\scriptscriptstyle B}}+ r{_{\scriptscriptstyle B}}&= 0, & \qquad
z{_{\scriptscriptstyle N}}+ r{_{\scriptscriptstyle N}}&\ge 0,
\end{alignedat}$$ for appropriate second-order consistent bases. For the dual method, the purpose is to drive the primal variables to feasibility (i.e., by driving the negative components of $x + q$ to zero).
An iteration begins with a base subiteration in which an index $l$ in the basic set ${\mathcal{B}}$ is identified such that $x_l+q_l < 0$. The corresponding dual variable $z_l$ may be increased from its current value $z_l = -r_l$ by removing the index $l$ from ${\mathcal{B}}$, and defining ${\mathcal{B}}\gets {\mathcal{B}}\setminus \{l\}$. Once $l$ is removed from ${\mathcal{B}}$, it holds that ${\mathcal{B}}\cup\{l\}\cup{\mathcal{N}}= \{1$, $2$,…, $n\}$. The resulting matrix $K_l$ of (\[eqn-KB-Kl-defined\]) is nonsingular, and the unique direction $({{{{\mathit\Delta}}}x}_l,{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}},{{{{\mathit\Delta}}}y})$ may be computed with ${{{{\mathit\Delta}}}z}_l=1$ in Proposition \[prop-Klnonsing\].
If ${{{{\mathit\Delta}}}x}_l > 0$, the step ${\alpha{_{\ast}}{^{\null}}}= -(x_l+q_l)/{{{{\mathit\Delta}}}x}_l$ gives $x_l +
{\alpha{_{\ast}}{^{\null}}}{{{{\mathit\Delta}}}x}_l + q_l = 0$. Otherwise, ${{{{\mathit\Delta}}}x}_l=0$ and Proposition \[propA-dirs\] of the Appendix implies that the dual objective function is linear and increasing along $({{{{\mathit\Delta}}}x},{{{{\mathit\Delta}}}y},{{{{\mathit\Delta}}}z})$. In this case ${\alpha{_{\ast}}{^{\null}}}= +\infty$. As $x_l + q_l$ is increased towards zero, some nonbasic dual variables may decrease and the step must be limited by ${\alpha{_{\operator@font{max}}}}= \min_{ i {:}{{{{\mathit\Delta}}}z}_i < 0} {\;}(z_i + r_i)(-{{{{\mathit\Delta}}}z}_i)$ to maintain feasibility of the nonbasic dual variables. This gives the step $\alpha = \min\big({\alpha{_{\ast}}{^{\null}}},{\alpha{_{\operator@font{max}}}}\big)$. If $\alpha = +\infty$, the dual problem is unbounded and the iteration is terminated. This is equivalent to the primal problem $({\mbox{\small PQP}}_{q,r})$ being infeasible. If $\alpha={\alpha{_{\ast}}{^{\null}}}$, then $x_l+ \alpha {{{{\mathit\Delta}}}x}_l+q_l=0$. Otherwise, it must hold that $\alpha = {\alpha{_{\operator@font{max}}}}$ and ${\mathcal{N}}$ and ${\mathcal{B}}$ are redefined as ${\mathcal{N}}={\mathcal{N}}\setminus \{k\}$ and ${\mathcal{B}}={\mathcal{B}}\cup \{k\}$, where $k$ is the index $k = {\mathop{\operator@font{argmin}}}_{ i {:}{{{{\mathit\Delta}}}z}_i < 0}{\;\;}(z_i + r_i)/(-{{{{\mathit\Delta}}}z}_i)$. The partition at the new point satisfies ${\mathcal{B}}\cup\{l\}\cup{\mathcal{N}}= \{1$, $2$, …, $n\}$. Proposition \[propA-dualnonsing\] of the Appendix shows that the new $K{_{\scriptscriptstyle B}}$ is nonsingular for both of the cases ${{{{\mathit\Delta}}}x}_l>0$ and ${{{{\mathit\Delta}}}x}_l=0$.
If $x_l+q_l<0$ at the new point, then at least one intermediate subiteration is necessary to drive $x_l + q_l$ to zero. The nonsingularity of $K{_{\scriptscriptstyle B}}$ implies that the search direction may be computed with ${{{{\mathit\Delta}}}x}_l=1$ in Proposition \[prop-KBnonsing\]. As in the base subiteration, the step length is ${\alpha{_{\ast}}{^{\null}}}= -(x_l+q_l)/{{{{\mathit\Delta}}}x}_l$, but in this case ${\alpha{_{\ast}}{^{\null}}}$ can never be infinite because ${{{{\mathit\Delta}}}x}_l=1$. If no constraint index is added to ${\mathcal{B}}$, then $x_l+\alpha {{{{\mathit\Delta}}}x}_l+q_l=0$. Otherwise, the index $k$ of a blocking variable is moved from ${\mathcal{N}}$ to ${\mathcal{B}}$. Proposition \[propA-dualnonsing\] of the Appendix implies that the updated $K{_{\scriptscriptstyle B}}$ is nonsingular at the end of an intermediate subiteration. Once $x_l + q_l$ is driven to zero, the index $l$ is moved to ${\mathcal{N}}$ and a new iteration is started.
The dual active-set method is summarized in Algorithm \[alg-dualqp\] below. Its convergence properties are discussed in Section \[sec-relaxed-dual\].
Find $(x,y,z)$ satisfying conditions (\[eqn-dualxyz\]) for some second-order consistent basis ${\mathcal{B}}$;
${\mathcal{B}}\gets {\mathcal{B}}\setminus \{l\}$; ; ; ${\mathcal{N}}\gets {\mathcal{N}}\cup \{l\}$;
${{{{\mathit\Delta}}}z}_l \gets 1$; Solve $ {\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T \\
h{_{{\scriptscriptstyle\mathit{Bl}}}}& H{_{\scriptscriptstyle\mathit{BB}}}& A{_{\scriptscriptstyle B}}^T \\
a_l & A{_{\scriptscriptstyle B}}& -M\end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}_l \\
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}y}\end{pmatrix}}
= {\begin{pmatrix} 1 \\
0 \\
0 \end{pmatrix}}$; ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}\gets h{_{{\scriptscriptstyle\mathit{Nl}}}{^{\null}}}{{{{\mathit\Delta}}}x}{_l{^{\null}}}+ H{_{\scriptscriptstyle\mathit{BN}}}^T {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle\mathit{B}}{^{\null}}}- A{_{\scriptscriptstyle N}}^T {{{{\mathit\Delta}}}y}$; ${\alpha{_{\ast}}{^{\null}}}\gets -(x_l+q_l)/{{{{\mathit\Delta}}}x}_l$; ${\displaystyle}{\alpha{_{\operator@font{max}}}}\gets \min_{i {:}{{{{\mathit\Delta}}}z}_i < 0}{\;\;}(z_i + r_i)/(-{{{{\mathit\Delta}}}z}_i)$; ${\displaystyle}k \gets {\mathop{\operator@font{argmin}}}_{ i {:}{{{{\mathit\Delta}}}z}_i < 0}{\;\;}(z_i + r_i)/(-{{{{\mathit\Delta}}}z}_i)$; $\alpha \gets \min\big({\alpha{_{\ast}}{^{\null}}},{\alpha{_{\operator@font{max}}}}\big)$; ${\mathbf{stop}}$; $x_l \gets x_l + \alpha {{{{\mathit\Delta}}}x}_l$; $x{_{\scriptscriptstyle B}}\gets x{_{\scriptscriptstyle B}}+ \alpha {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$; $y \gets y + \alpha {{{{\mathit\Delta}}}y}$; $z_l \gets z_l + \alpha {{{{\mathit\Delta}}}z}_l$; $z{_{\scriptscriptstyle N}}\gets z{_{\scriptscriptstyle N}}+ \alpha {{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}$; ${\mathcal{B}}\gets {\mathcal{B}}\cup \{k\}$; ${\mathcal{N}}\gets {\mathcal{N}}\setminus \{k\}$; ;
${{{{\mathit\Delta}}}x}_l \gets 1$; Solve ${\begin{pmatrix} H{_{\scriptscriptstyle\mathit{BB}}}& A{_{\scriptscriptstyle B}}^T \\
A{_{\scriptscriptstyle B}}& -M\end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}y}\end{pmatrix}}
= -{\begin{pmatrix} h{_{{\scriptscriptstyle\mathit{Bl}}}}\\
a_l \end{pmatrix}}$; ${{{{\mathit\Delta}}}z}_l \gets h{_{ll}{^{\null}}}{{{{\mathit\Delta}}}x}{_l{^{\null}}}+ h{_{{\scriptscriptstyle\mathit{Bl}}}}^T {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle\mathit{B}}{^{\null}}}- a_l^T {{{{\mathit\Delta}}}y}$; ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}\gets h{_{{\scriptscriptstyle\mathit{Nl}}}{^{\null}}}{{{{\mathit\Delta}}}x}{_l{^{\null}}}+ H{_{\scriptscriptstyle\mathit{BN}}}^T {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle\mathit{B}}{^{\null}}}- A{_{\scriptscriptstyle N}}^T {{{{\mathit\Delta}}}y}$; ${\alpha{_{\ast}}{^{\null}}}\gets -(x_l + q_l)$; ${\displaystyle}{\alpha{_{\operator@font{max}}}}\gets \min_{ i {:}{{{{\mathit\Delta}}}z}_i < 0}{\;\;}(z_i + r_i)/(-{{{{\mathit\Delta}}}z}_i)$; ${\displaystyle}k \gets {\mathop{\operator@font{argmin}}}_{ i {:}{{{{\mathit\Delta}}}z}_i < 0}{\;\;}(z_i + r_i)/(-{{{{\mathit\Delta}}}z}_i);$ $\alpha \gets \min\big({\alpha{_{\ast}}{^{\null}}},{\alpha{_{\operator@font{max}}}}\big)$; $x_l \gets x_l + \alpha {{{{\mathit\Delta}}}x}_l$; $x{_{\scriptscriptstyle B}}\gets x{_{\scriptscriptstyle B}}+ \alpha {{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$; $y \gets y + \alpha {{{{\mathit\Delta}}}y}$; $z_l \gets z_l + \alpha {{{{\mathit\Delta}}}z}_l$; $z{_{\scriptscriptstyle N}}\gets z{_{\scriptscriptstyle N}}+ \alpha {{{{\mathit\Delta}}}z}{_{\scriptscriptstyle N}}$; ${\mathcal{B}}\gets {\mathcal{B}}\cup \{k\}$; ${\mathcal{N}}\gets {\mathcal{N}}\setminus \{k\}$; ;
Combining Primal and Dual Active-Set Methods {#sec-primal-dual}
============================================
The primal active-set method proposed in Section \[sec-primal\] may be used to solve $({\mbox{\small PQP}}_{q,r})$ for a given initial second-order consistent basis satisfying the conditions (\[eqn-primalxyz\]). An appropriate initial point may be found by solving a conventional phase-1 linear program. Alternatively, the dual active-set method of Section \[sec-dual\] may be used in conjunction with an appropriate phase-1 procedure to solve the quadratic program $({\mbox{\small PQP}}_{q,r})$ for a given initial second-order consistent basis satisfying the conditions (\[eqn-dualxyz\]). In this section a method is proposed that provides an alternative to the conventional phase-1/phase-2 approach. It is shown that a pair of coupled quadratic programs may be created from the original by simultaneously shifting the bound constraints. Any second-order consistent basis can be made optimal for such a primal-dual pair of shifted problems. The shifts are then updated using the solution of either the primal or the dual shifted problem. An obvious application of this approach is to solve a shifted dual [QP]{} to define an initial feasible point for the primal, or *vice-versa*. This strategy provides an alternative to the conventional phase-1/phase-2 approach that utilizes the [QP]{} objective function while finding a feasible point.
Finding an initial second-order-consistent basis {#sec-SOC-basis}
------------------------------------------------
For the methods described in Section \[sec-initshift\] below, it is possible to define a simple procedure for finding the initial second-order consistent basis ${\mathcal{B}}$ such that the matrix $K{_{\scriptscriptstyle B}}$ of (\[eqn-submin-KKT\]) is nonsingular. The required basis may be obtained by finding a symmetric permutation ${{{\mathit\Pi}}}$ of the “full” KKT matrix $K$ such that $$\label{eqn-KKTpart}
{{{\mathit\Pi}}}^T K {{{\mathit\Pi}}}= {{{\mathit\Pi}}}^T{\begin{pmatrix} H & \; A^T \\
A & -M \end{pmatrix}}{{{\mathit\Pi}}}= {\begin{pmatrix} H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& \;A{_{\scriptscriptstyle B}}^T & H{_{\scriptscriptstyle\mathit{BN}}{^{\null}}}\\
A{_{\scriptscriptstyle B}}& -M & A{_{\scriptscriptstyle N}}\\
H{_{\scriptscriptstyle\mathit{BN}}}^T & \;A{_{\scriptscriptstyle N}}^T & H{_{\scriptscriptstyle\mathit{NN}}{^{\null}}}\end{pmatrix}},$$ where the leading principal block $2\times 2$ submatrix is of the form (\[eqn-submin-KKT\]). The full row-rank assumption on ${\big(\, A \ \
-M \,\big)}$ ensures that the permutation (\[eqn-KKTpart\]) is well defined, see [@For02 Section 6]. In practice, the permutation may be determined using any method for finding a symmetric indefinite factorization of $K$, see, e.g., [@BunP71; @Fle76; @BunK77]. Such methods use symmetric interchanges that implicitly form the nonsingular matrix $K{_{\scriptscriptstyle B}}$ by deferring singular pivots. In this case, $K{_{\scriptscriptstyle B}}$ may be defined as any submatrix of the largest nonsingular principal submatrix obtained by the factorization. (There may be further permutations within ${{{\mathit\Pi}}}$ that are not relevant to this discussion; for further details, see, e.g., [@FOrM93; @For02; @DufR83; @Duf04].) The permutation ${{{\mathit\Pi}}}$ defines the initial ${\mathcal{B}}$-${\mathcal{N}}$ partition of the columns of $A$, i.e., it defines an initial second-order consistent basis.
Initializing the shifts {#sec-initshift}
-----------------------
Given a second-order consistent basis, it is straightforward to create shifts $(q{^{(0)}},r{^{(0)}})$ and corresponding $(x,y,z)$ so that $q{^{(0)}}\ge0$, $r{^{(0)}}\ge 0$ and $(x,y,z)$ are optimal for $({\mbox{\small PQP}}_{q{^{(0)}},r{^{(0)}}})$ and $({\mbox{\small DQP}}_{q{^{(0)}},r{^{(0)}}})$. First, choose nonnegative vectors $q{^{(0)}}{_{\scriptscriptstyle N}}$ and $r{^{(0)}}{_{\scriptscriptstyle B}}$. (Obvious choices are $q{^{(0)}}{_{\scriptscriptstyle N}}= 0$ and $r{^{(0)}}{_{\scriptscriptstyle B}}= 0$.) Define $z{_{\scriptscriptstyle B}}= -r{^{(0)}}{_{\scriptscriptstyle B}}$, $x{_{\scriptscriptstyle N}}= -q{^{(0)}}{_{\scriptscriptstyle N}}$, and solve the nonsingular [KKT]{}-system (\[eqn-xBy\]) to obtain $x{_{\scriptscriptstyle B}}$ and $y$, and compute $z{_{\scriptscriptstyle N}}$ from (\[eqn-zN\]). Finally, let $q{^{(0)}}{_{\scriptscriptstyle B}}\ge \max\{-x{_{\scriptscriptstyle B}},0\}$ and $r{^{(0)}}{_{\scriptscriptstyle N}}\ge
\max\{-z{_{\scriptscriptstyle N}},0\}$. Then, it follows from Proposition \[prop-regQPopt\] that $x$, $y$ and $z$ are optimal for the problems $({\mbox{\small PQP}}_{q{^{(0)}},r{^{(0)}}})$ and $({\mbox{\small DQP}}_{q{^{(0)}},r{^{(0)}}})$, with $q{^{(0)}}\ge 0$ and $r{^{(0)}}\ge 0$. If $q{^{(0)}}$ and $r{^{(0)}}$ are zero, then $x$, $y$ and $z$ are optimal for the original problem.
Solving the original problem by removing the shifts
---------------------------------------------------
The original problem may now be solved as a pair of shifted quadratic programs. Two alternative strategies are proposed. The first is a “primal first” strategy in which a shifted primal quadratic program is solved, followed by a dual. The second is an analogous “dual first” strategy.
The “primal-first” strategy is summarized as follows.
1. Find ${\mathcal{B}}$, ${\mathcal{N}}$, $q{^{(0)}}$, $r{^{(0)}}$, $x$, $y$, $z$, as described in Sections \[sec-SOC-basis\] and \[sec-initshift\].
2. Set $q{^{(1)}} = q{^{(0)}}$, $r{^{(1)}}=0$. Solve $({\mbox{\small PQP}}_{q,0})$ using the primal active-set method.
3. Set $q{^{(2)}} = 0$, $r{^{(2)}}=0$. Solve $({\mbox{\small DQP}}_{0,0})$ using the dual active-set method.
In steps (1) and (2), the initial ${\mathcal{B}}$–${\mathcal{N}}$ partition and initial values of $x$, $y$, and $z$ are defined as the final ${\mathcal{B}}$–${\mathcal{N}}$ partition and final values of $x$, $y$, and $z$ from the preceding step.
The “dual-first” strategy is defined in an analogous way.
1. Find ${\mathcal{B}}$, ${\mathcal{N}}$, $q{^{(0)}}$, $r{^{(0)}}$, $x$, $y$, $z$, as described in Section \[sec-SOC-basis\] and \[sec-initshift\].
2. Set $q{^{(1)}}=0$, $r{^{(1)}} = r{^{(0)}}$. Solve $({\mbox{\small DQP}}_{0,r})$ using the dual active-set method.
3. Set $q{^{(2)}}=0$, $r{^{(2)}}=0$. Solve $({\mbox{\small PQP}}_{0,0})$ using the primal active-set method.
As in the “primal-first” strategy, the initial ${\mathcal{B}}$–${\mathcal{N}}$ partition and initial values of $x$, $y$, and $z$ for steps (1) and (2), are defined as the final ${\mathcal{B}}$–${\mathcal{N}}$ partition and final values of $x$, $y$, and $z$ from the preceding step.
(The strategies of solving two consecutive quadratic programs may be generalized to a sequence of more than two quadratic programs, where we alternate between primal and dual active-set methods, and eliminate the shifts in more than two steps.)
In order for these approaches to be well-defined, a simple generalization of the primal and dual active-set methods of Algorithms \[alg-primalqp\] and \[alg-dualqp\] is required.
Relaxed initial conditions for the primal QP method. {#sec-relaxed-primal}
----------------------------------------------------
For Algorithm \[alg-primalqp\], the initial values of ${\mathcal{B}}$, ${\mathcal{N}}$, $q$, $r$, $x$, $y$, and $z$ must satisfy conditions (\[eqn-primalxyz\]). However, the choice of $r = r{^{(2)}} = 0$ in Step (2) of the dual-first strategy may give some negative components in the vector $z{_{\scriptscriptstyle B}}+ r{_{\scriptscriptstyle B}}$. This possibility may be handled by defining a simple generalization of Algorithm \[alg-primalqp\] that allows initial points satisfying the conditions $$\label{eqn-primalxyzForShifts}
\begin{alignedat}{3}
H x + c - A{^T\!}y - z &= 0, & \qquad x{_{\scriptscriptstyle N}}+ q{_{\scriptscriptstyle N}}&= 0, & \qquad x{_{\scriptscriptstyle B}}+ q{_{\scriptscriptstyle B}}&\ge 0, \\
Ax + M y - b &= 0, & \qquad z{_{\scriptscriptstyle B}}+ r{_{\scriptscriptstyle B}}&\le 0,
\end{alignedat}$$ instead of the conditions (\[eqn-primalxyz\]). In Algorithm \[alg-primalqp\], the index $l$ identified at the start of the primal base subiteration is selected from the set of nonbasic indices such that $z_j + r_j < 0$. In the generalized algorithm, the set of eligible indices for $l$ is extended to include indices associated with negative values of $z{_{\scriptscriptstyle B}}+ r{_{\scriptscriptstyle B}}$. If the index $l$ is deleted from ${\mathcal{B}}$, the associated matrix $K_l$ is nonsingular, and intermediate subiterations are executed until the updated value satisfies $z_l+r_l=0$. At this point, the index $l$ is returned ${\mathcal{B}}$. The method is summarized in Algorithm \[alg-primalqpForShifts\].
Find $(x,y,z)$ satisfying conditions (\[eqn-primalxyzForShifts\]) for some second-order consistent basis ${\mathcal{B}}$;
${\mathcal{N}}\gets {\mathcal{N}}\setminus \{l\}$; ; ${\mathcal{B}}\gets {\mathcal{B}}\setminus \{l\}$; ; ${\mathcal{B}}\gets {\mathcal{B}}\cup \{l\}$;
This section concludes with a convergence result for the primal method of Algorithm \[alg-primalqpForShifts\]. In particular, it is shown that the algorithm is well-defined, and terminates in a finite number of iterations if $({\mbox{\small PQP}}_{q,r})$ is *nondegenerate*. We define nondegeneracy to mean that a nonzero step in the $x$-variables is taken at each iteration of Algorithm \[alg-primalqpForShifts\] that involves a base subiteration. A sufficient condition on $({\mbox{\small PQP}}_{q,r})$ for this to hold is that the gradients of the equality constraints and active bound constraints are linearly independent at each iterate. See, e.g., Fletcher [@Fle93] for further discussion of these issues. As the active-set strategy uses the same criteria for adding and deleting variables as those used in the simplex method, standard pivot selection rules used to avoid cycling in linear programming, such as lexicographical ordering, least-index selection or perturbation may be applied directly to the method proposed here (see, e.g., Bland [@Bla77], Charnes [@Cha52], Dantzig, Orden and Wolfe [@DanOW55], and Harris [@Har73]).
\[thm-primalConvergenceWithShifts\] Given a primal-feasible point $(x_0,y_0,z_0)$ satisfying conditions [(\[eqn-primalxyzForShifts\])]{} for a second-order consistent basis ${\mathcal{B}}_0$, then [Algorithm \[alg-primalqpForShifts\]]{} generates a sequence of second-order consistent bases ${\{{\mskip 1 mu}{\mathcal{B}}_j{\mskip 1 mu}\}}_{j>0}$. Moreover, if problem [$({\mbox{\small PQP}}_{q,r})$]{} is nondegenerate, then [ Algorithm \[alg-primalqpForShifts\]]{} finds a solution of [ $({\mbox{\small PQP}}_{q,r})$]{} or determines that [$({\mbox{\small DQP}}_{q,r})$]{} is infeasible in a finite number of iterations.
Assume that $(x,y,z)$ satisfies the conditions (\[eqn-primalxyzForShifts\]) for the second-order consistent basis ${\mathcal{B}}$. Propositions \[prop-KBnonsing\] and \[prop-Klnonsing\] imply that the [KKT]{} matrices associated with subsequent base and intermediate iterations are nonsingular, in which case each basis is second-order consistent. Let ${{\mathcal{B}}^{\scriptscriptstyle<}}$ denote the index set ${{\mathcal{B}}^{\scriptscriptstyle<}}=\{ i
\in{\mathcal{B}}: z_i + r_i < 0\}$, and let ${{\skew3\widetilde r}}$ be the vector ${{\skew3\widetilde r}}_i=r_i$, $i\not \in {{\mathcal{B}}^{\scriptscriptstyle<}}$, and ${{\skew3\widetilde r}}_i=-z_i$, $i\in {{\mathcal{B}}^{\scriptscriptstyle<}}$. These definitions imply that ${{\skew3\widetilde r}}_i=-z_i>
-z_i + z_i + r_i = r_i$, for every $i\in {{\mathcal{B}}^{\scriptscriptstyle<}}$. It follows that ${{\skew3\widetilde r}}\ge r$, and the feasible region of $({\mbox{\small DQP}}_{q,r})$ is a subset of the feasible region of $({\mbox{\small DQP}}_{q,{{\skew3\widetilde r}}})$. In addition, if $r$ is replaced by ${{\skew3\widetilde r}}$ in (\[eqn-primalxyz\]), the only difference is that $z{_{\scriptscriptstyle B}}+{{\skew3\widetilde r}}{_{\scriptscriptstyle B}}=0$, i.e., the initial point for (\[eqn-primalxyzForShifts\]) is a stationary point with respect to $({\mbox{\small PQP}}_{q,{{\skew3\widetilde r}}})$.
The first step of the proof is to show that after a finite number of iterations of Algorithm \[alg-primalqpForShifts\], one of three possible events must occur: (i) the cardinality of the set ${{\mathcal{B}}^{\scriptscriptstyle<}}$ is decreased by at least one; (ii) a solution of problem $({\mbox{\small PQP}}_{q,r})$ is found; or (iii) $({\mbox{\small DQP}}_{q,r})$ is declared infeasible. The proof will also establish that if (i) does not occur, then either (ii) or (iii) must hold after a finite number of iterations.
Assume that (i) never occurs. This implies that the index $l$ selected in the base subiteration can never be an index in ${{\mathcal{B}}^{\scriptscriptstyle<}}$ because at the end of such an iteration, it would belong to ${\mathcal{B}}$ with $z_l+r_l=0$, contradicting the assumption that the cardinality of ${{\mathcal{B}}^{\scriptscriptstyle<}}$ never decreases. For the same reason, it must hold that $k\not\in {{\mathcal{B}}^{\scriptscriptstyle<}}$ for every index $k$ selected to be moved from ${\mathcal{B}}$ to ${\mathcal{N}}$ in any subiteration, because an index can only be moved from ${\mathcal{N}}$ to ${\mathcal{B}}$ by being selected in the base subiteration. These arguments imply that $z_i=-{{\skew3\widetilde r}}_i$, with $i\in{{\mathcal{B}}^{\scriptscriptstyle<}}$, throughout the iterations. It follows that the iterates may be interpreted as being members of a sequence constructed for solving $({\mbox{\small PQP}}_{q,{{\skew3\widetilde r}}})$ with a fixed ${{\skew3\widetilde r}}$, where the initial stationary point is given, and each iteration gives a new stationary point. The nondegeneracy assumption implies that $\alpha {{{\mathit\Delta}}}x\ne 0$ for at least one subiteration. For the base subiteration, ${{{\mathit\Delta}}}x_l >
0$, and it follows from Proposition \[prop-Klnonsing\] that ${{{\mathit\Delta}}}x\ne 0$ if and only if ${{{\mathit\Delta}}}x_l>0$ for an intermediate subiteration. Therefore, Proposition \[propA-dirs\] shows that the objective value of $({\mbox{\small PQP}}_{q,{{\skew3\widetilde r}}})$ is strictly decreasing for a subiteration where $\alpha {{{\mathit\Delta}}}x\ne 0$. In addition, the objective value of $({\mbox{\small PQP}}_{q,{{\skew3\widetilde r}}})$ is nonincreasing at each subiteration, so a strict overall improvement of the objective value of $({\mbox{\small PQP}}_{q,{{\skew3\widetilde r}}})$ is obtained at each iteration. As there are only a finite number of stationary points, Algorithm \[alg-primalqpForShifts\] either solves $({\mbox{\small PQP}}_{q,{{\skew3\widetilde r}}})$ or concludes that $({\mbox{\small DQP}}_{q,{{\skew3\widetilde r}}})$ is infeasible after a finite number of iterations. If $({\mbox{\small PQP}}_{q,{{\skew3\widetilde r}}})$ is solved, then $z{_{\scriptscriptstyle N}}+r{_{\scriptscriptstyle N}}\ge
0$, because ${{\skew3\widetilde r}}_j = r_j$ for $j\in{\mathcal{N}}$. Hence, Algorithm \[alg-primalqpForShifts\] can not proceed further by selecting an $l\in{\mathcal{N}}$, and the only way to reduce the objective is to select an $l$ in ${\mathcal{B}}$ such that $z_j + r_j < 0$. Under the assumption that (i) does not occur, it must hold that no eligible indices exist and ${{\mathcal{B}}^{\scriptscriptstyle<}}=\emptyset$. However, in this case $({\mbox{\small PQP}}_{q,r})$ has been solved with ${{\skew3\widetilde r}}=r$, and (ii) must hold. If Algorithm \[alg-primalqpForShifts\] declares $({\mbox{\small DQP}}_{q,{{\skew3\widetilde r}}})$ to be infeasible, then $({\mbox{\small DQP}}_{q,r})$ must also be infeasible because the feasible region of $({\mbox{\small DQP}}_{q,r})$ is contained in the feasible region of $({\mbox{\small DQP}}_{q,{{\skew3\widetilde r}}})$. In this case $({\mbox{\small DQP}}_{q,r})$ is infeasible and (iii) occurs.
Finally, if (i) occurs, there is an iteration at which the cardinality of ${{\mathcal{B}}^{\scriptscriptstyle<}}$ decreases and an index is removed from ${{\mathcal{B}}^{\scriptscriptstyle<}}$. There may be more than one such index, but there is at least one $l$ moved from ${{\mathcal{B}}^{\scriptscriptstyle<}}$ to ${\mathcal{B}}\backslash{{\mathcal{B}}^{\scriptscriptstyle<}}$, or one $k$ moved from ${{\mathcal{B}}^{\scriptscriptstyle<}}$ to ${\mathcal{N}}$. In either case, the cardinality of ${{\mathcal{B}}^{\scriptscriptstyle<}}$ is decreased by at least one. After such an iteration, the argument given above may be repeated for the new set ${{\mathcal{B}}^{\scriptscriptstyle<}}$ and new shift ${{\skew3\widetilde r}}$. Applying this argument repeatedly gives the result that the situation (i) can occur only a finite number of times.
It follows that (ii) or (iii) must occur after a finite number of iterations, which is the required result.
Relaxed initial conditions for the dual QP method. {#sec-relaxed-dual}
--------------------------------------------------
Analogous to the primal case, the choice of $q = q{^{(2)}} = 0$ in Step (2) of the primal-first strategy may give some negative components in the vector $x{_{\scriptscriptstyle N}}+ q{_{\scriptscriptstyle N}}$. In this case, the conditions (\[eqn-dualxyz\]) on the initial values of ${\mathcal{B}}$, ${\mathcal{N}}$, $q$, $r$, $x$, $y$, and $z$ are relaxed so that $$\label{eqn-dualxyzForShifts}
\begin{alignedat}{3}
H x + c - A{^T\!}y - z &= 0, &\qquad x{_{\scriptscriptstyle N}}+ q{_{\scriptscriptstyle N}}&\le 0, \\
Ax + M y - b &= 0, &\qquad z{_{\scriptscriptstyle B}}+ r{_{\scriptscriptstyle B}}& = 0, &\qquad z{_{\scriptscriptstyle N}}+ r{_{\scriptscriptstyle N}}&\ge 0.
\end{alignedat}$$ Similarly, the set of eligible indices may be extended to include indices associated with negative values of $x{_{\scriptscriptstyle N}}+ q{_{\scriptscriptstyle N}}$. If the index $l$ is from ${\mathcal{N}}$, the associated matrix $K{_{\scriptscriptstyle B}}$ is nonsingular, and intermediate subiterations are executed until the updated value satisfies $x_l+q_l=0$. At this point, the index $l$ is returned ${\mathcal{N}}$. The method is summarized in Algorithm \[alg-dualqpForShifts\].
Find $(x,y,z)$ satisfying conditions (\[eqn-dualxyzForShifts\]) for some second-order consistent ${\mathcal{B}}$; ${\mathcal{B}}\gets {\mathcal{B}}\setminus \{l\}$; ; ${\mathcal{N}}\gets {\mathcal{N}}\setminus \{l\}$; ; ${\mathcal{N}}\gets {\mathcal{N}}\cup \{l\}$;
A convergence result analogous to Theorem \[thm-primalConvergenceWithShifts\] holds for the dual algorithm. In this case, the nondegeneracy assumption concerns the linear independence of the gradients of the equality constraints and active bounds for $({\mbox{\small DQP}}_{q,r})$.
\[thm-dualConvergenceWithShifts\] Given a dual-feasible point $(x_0,y_0,z_0)$ satisfying conditions [(\[eqn-dualxyzForShifts\])]{} for a second-order consistent basis ${\mathcal{B}}_0$, then [Algorithm \[alg-dualqpForShifts\]]{} generates a sequence of second-order consistent bases ${\{{\mskip 1 mu}{\mathcal{B}}_j
{\mskip 1 mu}\}}_{j>0}$. Moreover, if problem [$({\mbox{\small DQP}}_{q,r})$]{} is nondegenerate, then [Algorithm \[alg-dualqpForShifts\]]{} either solves [$({\mbox{\small DQP}}_{q,r})$]{} or concludes that [$({\mbox{\small PQP}}_{q,r})$]{} is infeasible in a finite number of iterations.
The proof mirrors that of Theorem \[thm-primalConvergenceWithShifts\] for the primal method.
Practical Issues {#sec-general-format}
================
As stated, the primal quadratic program has lower bound zero on the $x$-variables. This is for notational convenience. This form may be generalized in a straightforward manner to a form where the $x$-variables has both lower and upper bounds on the primal variables, i.e., $b\subL \le x \le b{{_{\scriptscriptstyle U}}}$, where components of $b\subL$ can be $-\infty$ and components of $b{{_{\scriptscriptstyle U}}}$ can be $+\infty$. Given primal shifts $q\subL$ and $q{{_{\scriptscriptstyle U}}}$, and dual shifts $r\subL$ and $r{{_{\scriptscriptstyle U}}}$, we have the primal-dual pair $$({\mbox{\small PQP}}_{q,r}){\;}\begin{array}{ll}
{{\displaystyle{\mathop{\operator@font{minimize}}}_{x,y}}} &{{\phantom-}}{{\textstyle\frac12}}x{^T\!}H x + {{\textstyle\frac12}}y{^T\!}M y + c{^T\!}x + (r\subL - r{{_{\scriptscriptstyle U}}}){^T\!}x\\
{\mathop{\operator@font{subject\ to}}}&{\settowidth{\argwidth}{$- Hx + A{^T\!}y + z\subL - z{{_{\scriptscriptstyle U}}}= c$}
\makebox[\argwidth][l]{${{\phantom-}}Ax + M y = b,$}}
{\;\;}{\;\;\;}b\subL - q\subL \le x \le b{{_{\scriptscriptstyle U}}}+ q{{_{\scriptscriptstyle U}}},
\end{array}
\]
and
\[
({\mbox{\small DQP}}_{q,r}) {\;}\begin{array}{ll}
{{\displaystyle{\mathop{\operator@font{maximize}}}_{x,y,z\subL,z{{_{\scriptscriptstyle U}}}}}}&-{{\textstyle\frac12}}x{^T\!}H x - {{\textstyle\frac12}}y{^T\!}M y + b{^T\!}y + (b\subL - q\subL){^T\!}z\subL - (b{{_{\scriptscriptstyle U}}}+ q{{_{\scriptscriptstyle U}}}){^T\!}z{{_{\scriptscriptstyle U}}}\\
{\mathop{\operator@font{subject\ to}}}& {\settowidth{\argwidth}{$- Hx + A{^T\!}y + z\subL - z{{_{\scriptscriptstyle U}}}= c$}
\makebox[\argwidth][l]{$- Hx + A{^T\!}y + z\subL - z{{_{\scriptscriptstyle U}}}= c,$}}
{\;\;}{\;\;\;}z\subL \ge -r\subL, {\;\;\;}z{{_{\scriptscriptstyle U}}}\ge -r{{_{\scriptscriptstyle U}}}.
\end{array}$$ An infinite bound has neither a shift nor a corresponding dual variable. For example, if the $j$th components of $b\subL$ and $b{{_{\scriptscriptstyle U}}}$ are infinite, then the corresponding variable $x_j$ is free. In the procedure given in Section \[sec-SOC-basis\] for finding the first second-order consistent basis ${\mathcal{B}}$, it is assumed that variables with indices not selected for ${\mathcal{B}}$ are initialized at one of their bounds. As a free variable has no finite bounds, any index $j$ associated with a free variable should be selected for ${\mathcal{B}}$. However, this cannot be guaranteed in practice, and in the next section it is shown that the primal and dual [QP]{} methods may be extended to allow a free variable to be fixed temporarily at some value.
If the [QP]{} is defined in the general problem format of Section \[sec-general-format\], then any free variable not selected for ${\mathcal{B}}$ has no upper or lower bound and must be temporarily fixed at some value $x_j = {\skew{2.8}\bar x}_j$ (say). The treatment of such “temporary bounds” involves some additional modifications to the primal and dual methods of Sections \[sec-relaxed-primal\] and \[sec-relaxed-dual\].
Each temporary bound $x_j = {\skew{2.8}\bar x}_j$ defines an associated dual variable $z_j$ with initial value ${\skew{2.8}\bar z}_j$. As the bound is temporary, it is treated as an equality constraint, and the desired value of $z_j$ is zero. Initially, an index $j$ corresponding to a temporary bound is assigned a primal shift $q_j=0$ and a dual shift $r_j=-{\skew{2.8}\bar z}_j$, making ${\skew{2.8}\bar x}_j$ and ${\skew{2.8}\bar z}_j$ feasible for the shifted problem. In both the primal-first and dual-first approaches, the idea is to drive the $z_j$-variables associated with temporary bounds to zero in the primal and leave them unchanged in the dual.
In a primal problem, regardless of whether it is solved before or after the dual problem, an index $j$ corresponding to a temporary bound for which $z_j\ne 0$ is considered eligible for selection as $l$ in the base subiteration, i.e., the index can be selected regardless of the sign of $z_j$. Once selected, $z_j$ is driven to zero and $j$ belongs to ${\mathcal{B}}$ after such an iteration. In addition, as $x_j$ has no finite bounds, $j$ will remain in ${\mathcal{B}}$ throughout the iterations. Hence, at termination of a primal problem, any index $j$ corresponding to a temporarily bounded variable must have $z_j=0$. If the maximum step length at a base subiteration is infinite, the dual problem is infeasible, as in the case of a regular bound.
In a dual problem, the dual method is modified so that the dual variables associated with temporary bounds remain fixed throughout the iterations. At any subiteration, if it holds that ${{{{\mathit\Delta}}}z}_j\ne 0$ for some temporary bound, then no step is taken and one such index $j$ is moved from ${\mathcal{N}}$ to ${\mathcal{B}}$. Consequently, a move is made only if ${{{{\mathit\Delta}}}z}_j= 0$ for every temporary bound $j$. It follows that the dual variables for the temporary bounds will remain unaltered throughout the dual iterations. Note that an index $j$ corresponding to a temporary bound is moved from ${\mathcal{N}}$ to ${\mathcal{B}}$ at most once, and is never moved back because the corresponding $x_j$-variable has no finite bounds. If the maximum step length at a base subiteration is infinite, it must hold that ${{{{\mathit\Delta}}}z}_j=0$ for all temporary bounds $j$, and the primal problem is infeasible.
The discussion above implies that a pair of primal and dual problems solved consecutively will terminate with $z_j=0$ for all indices $j$ associated with temporary bounds. This is because $z_j$ is unchanged in the dual problem and driven to zero in the primal problem.
Numerical Examples {#sec-numerical-results}
==================
This section concerns a particular formulation of the combined primal-dual method of Section \[sec-primal-dual\] in which either a “primal-first” or “dual-first” strategy is selected based on the initial point. In particular, if the point is dual feasible, then the “dual-first” strategy is used, otherwise, the “primal-first” strategy is selected. Some numerical experiments are presented for a simple [[Matlab]{}]{} implementation applied to a set of convex problems from the [[CUTE]{}st]{} test collection (see Bongartz, [et al.]{}[@BonCGT95], and Gould, Orban and Toint [@GouOT03; @GouOT15]).
The test problems {#sec-test-problems}
-----------------
Each [QP]{} problem in the [[CUTE]{}st]{} test set may be written in the form $${{\displaystyle{\mathop{\operator@font{minimize}}}_{x}}} {\;\;}{{\textstyle\frac12}}x{^T\!}{\skew3\widehat H}x + c{^T\!}x {\;\;\;}{\mathop{\operator@font{subject\ to}}}{\;\;}\ell \le {\begin{pmatrix} x \\ {\skew6\widehat A}x \end{pmatrix}} \le u,$$ where $\ell$ and $u$ are constant vectors of lower and upper bounds, and ${\skew6\widehat A}$ has dimension $m\times n$. In this format, a fixed variable or equality constraint has the same value for its upper and lower bound. Each problem was converted to the equivalent form $$\label{eqn-QP-standard}
{{\displaystyle{\mathop{\operator@font{minimize}}}_{x,s}}} {\;\;}{{\textstyle\frac12}}x{^T\!}{\skew3\widehat H}x + c{^T\!}x {\;\;\;}{\mathop{\operator@font{subject\ to}}}{\;\;}{\skew6\widehat A}x - s = 0,
{\;\;\;}\ell \le {\begin{pmatrix} x \\ s \end{pmatrix}} \le u,$$ where $s$ is a vector of slack variables. With this formulation, the [QP]{} problem involves simple upper and lower bounds instead of nonnegativity constraints. It follows that the matrix $M$ is zero, but the full row-rank assumption on the constraint matrix is satisfied because the constraint matrix $A$ takes the form ${\big(\, {\skew6\widehat A}\ \ -I \,\big)}$ and has rank $m$.
Numerical results were obtained for a set of 121 convex [QP]{}s in standard interface format (SIF). The problems were selected based on the dimension of the constraint matrix $A$ in (\[eqn-QP-standard\]). In particular, the test set includes all [QP]{} problems for which the smaller of $m$ and $n$ is of the order of 500 or less. This gave 121 [QP]{}s ranging in size from [`BQP1VAR`]{} (one variable and one constraint) to [`LINCONT`]{} (1257 variables and 419 constraints).
The implementation
------------------
The combined primal-dual active-set method was implemented in [[Matlab]{}]{} as Algorithm [`PDQP`]{}. For illustrative purposes, results were obtained for [`PDQP`]{} and the [QP]{} solver [[`SQOPT`]{}]{} [@GilMS06a], which is a Fortran implementation of a conventional two-phase (primal) active-set method for large-scale [QP]{}. Both [`PDQP`]{} and [[`SQOPT`]{}]{} use the method of variable reduction, which implicitly transforms a [KKT]{} system of the form (\[eqn-xBy\]) into a block-triangular system. The general [QP]{} constraints ${\skew6\widehat A}x - s = 0$ are partitioned into the form $B x{^{\scriptscriptstyle\mathit{B}}} + S
x{^{\scriptscriptstyle\mathit{S}}} + A{_{\scriptscriptstyle N}}x{_{\scriptscriptstyle N}}= 0$, where $B$ is square and nonsingular, with $A{_{\scriptscriptstyle B}}={\big(\, B \ \ S \,\big)}$ and $x{_{\scriptscriptstyle B}}= (x{^{\scriptscriptstyle\mathit{B}}},x{^{\scriptscriptstyle\mathit{S}}})$. The vectors $x{^{\scriptscriptstyle\mathit{B}}}$, $x{^{\scriptscriptstyle\mathit{S}}}$, $x{_{\scriptscriptstyle N}}$ are the associated basic, superbasic, and nonbasic components of $(x,s)$ (see Gill, Murray and Saunders [@GilMS05]). If $H$ denotes the Hessian ${\skew3\widehat H}$ of (\[eqn-QP-standard\]) augmented by the zero rows and columns corresponding to the slack variables, then the reduced Hessian $Z{^T\!}H Z$ is defined in terms of the matrix $Z$ such that $$Z = P {\begin{pmatrix}-B{^{-1}}S \\ I \\ 0\end{pmatrix}},$$ where $P$ permutes the columns of ${\big(\, {\skew6\widehat A}\ \ -I \,\big)}$ into the order ${\big(\, B \ \ S \ \ A{_{\scriptscriptstyle N}}\,\big)}$. The matrix $Z$ is used only as an operator, i.e., it is not stored explicitly. Products of the form $Zv$ or $Z{^T\!}u$ are obtained by solving with $B$ or $B^T$. With these definitions, the resulting block lower-triangular system has diagonal blocks $Z{^T\!}H Z$, $B$ and $B^T$.
The initial nonsingular $B$ is identified using an LU factorization of $A^T$. The resulting $Z$ is used to form $Z{^T\!}H Z$, and a partial Cholesky factorization with interchanges is be used to find an upper-triangular matrix $R$ that is the factor of the largest nonsingular leading submatrix of $Z{^T\!}H Z$. If $Z{_{\scriptscriptstyle R}}$ denotes the columns of $Z$ corresponding to $R$, and $Z$ is partitioned as $Z = {\big(\, Z{_{\scriptscriptstyle R}}\ \ Z{_{\scriptscriptstyle A}}\,\big)}$, then the index set ${\mathcal{B}}$ consisting of the union of the column indices of $B$ and the indices corresponding to $Z{_{\scriptscriptstyle R}}$ defines an appropriate initial second-order consistent basis.
All [[`SQOPT`]{}]{} runs were made using the default parameter options. Both [`PDQP`]{} and [[`SQOPT`]{}]{} are terminated at a point $(x,y,z)$ that satisfies the optimality conditions of Proposition \[prop-regQPopt\] modified to conform to the constraint format of (\[eqn-QP-standard\]). The feasibility and optimality tolerances are given by ${\epsilon_{\mathtt{fea}}}= 10^{-6}$ and $\epsilon{_{\mathtt{opt}}}= 10^{-6}$, respectively. For a given $\epsilon{_{\mathtt{opt}}}$, [`PDQP`]{} and [[`SQOPT`]{}]{} terminate when $$\max_{i\in{\mathcal{B}}} {|z_i|} \le \epsilon{_{\mathtt{opt}}}{{\|y\|}_{\infty}{^{\null}}}, {\quad\text{and}\quad}
\left\{
\begin{array}{c@{\hspace{10pt}}l}
z_i \ge -\epsilon{_{\mathtt{opt}}}{{\|y\|}_{\infty}{^{\null}}} & \mbox{if $x_i \ge - \ell_i$, $i\in{\mathcal{N}}$;} \\[2ex]
z_i \le {{\phantom-}}\epsilon{_{\mathtt{opt}}}{{\|y\|}_{\infty}{^{\null}}} & \mbox{if $x_i \le{{\phantom-}}u_i$, $i\in{\mathcal{N}}$.}
\end{array}
\right.$$ Both [`PDQP`]{} and [[`SQOPT`]{}]{} use the [[EXPAND]{}]{} anti-cycling procedure of Gill [et al.]{} [@GilMSW89a] to allow the variables $(x,s)$ to move outside their bounds by as much as ${\epsilon_{\mathtt{fea}}}$. The [[EXPAND]{}]{} procedure does not guarantee that cycling will never occur (see Hall and McKinnon [@HalM96] for an example). Nevertheless, in many years of use, the authors have never known [[EXPAND]{}]{} to cycle on a practical problem.
Numerical results {#sec-result-summary}
-----------------
[`PDQP`]{} and [[`SQOPT`]{}]{} were applied to the 121 problems considered in Section \[sec-test-problems\]. A summary of the results is given in Table \[tab:SQvsPDQP-auto\]. The first four columns give the name of the problem, the number of linear constraints `m`, the number of variables `n`, and the optimal objective value `Objective`. The next two columns summarize the [[`SQOPT`]{}]{} result for the given problem, with `Phs1` and `Itn` giving the phase-one iterations and iteration total, respectively. The last four columns summarize the results for [`PDQP`]{}. The first column gives the total number of primal and dual iterations `Itn`. The second column gives the order in which the primal and dual algorithms were applied, with `PD` indicating the “primal-first” strategy, and `DP` the “dual-first” strategy. The final two columns, headed by `p-Itn`, and `d-Itn`, give the iterations required for the primal method and the dual method, respectively.
Of the 121 problems tested, two ([`LINCONT`]{} and [`NASH`]{}) are known to be infeasible. This infeasibility was identified correctly by both [[`SQOPT`]{}]{} and [`PDQP`]{}. In total, [[`SQOPT`]{}]{} solved 117 of the remaining 119 problems, but declared (incorrectly) that problems [`RDW2D51U`]{} and [`RDW2D52U`]{} are unbounded. [`PDQP`]{} solved the same number of problems, but failed to achieve the required accuracy for the problems [`RDW2D51B`]{} and [`RDW2D52F`]{}. In these two cases, the final objective values computed by [`PDQP`]{} were `1.0947648E-02` and `1.0491239E-02` respectively, instead of the optimal values `1.0947332e-02` and `1.0490828e-02`. (The five [`RDW2D5*`]{} problems in the test set are known to be difficult to solve, see Gill and Wong [@GilW15].)
Figure \[fig:pp\] gives a performance profile (in $\log_2$ scale) for the iterations required by [`PDQP`]{} and [[`SQOPT`]{}]{}. (For more details on the use of performance profiles, see Dolan and Moré [@DolM02].) The figure profiles the total iterations for [`PDQP`]{}, the number of phase-2 iterations for [[`SQOPT`]{}]{}, and the sum of phase-1 and phase-2 iterations for [[`SQOPT`]{}]{}. Some care must be taken when interpreting the results in the profile. First, the [[CUTE]{}st]{} test set contains several groups made up of similar variants of the same problem. In this situation, the profiles can be skewed by the fact that a method will tend to exhibit similar performance on all the problems in the group. For example, [`PDQP`]{} performs significantly better than [[`SQOPT`]{}]{} on all four [`JNLBRNG*`]{} problems, but significantly worse on all 12 [`LISWET*`]{} problems.
Second, the phase-1 search direction for [[`SQOPT`]{}]{} requires the computation of the vector $-Z Z{^T\!}{\skew3\widehat g}(x)$, where ${\skew3\widehat g}(x)$ is the gradient of the sum of infeasibilities of the bound constraints at $x$. This implies that a phase-1 iteration for [[`SQOPT`]{}]{} requires solves with $B$ and $B^T$, compared to solves with $B$, $B^T$ and $Z{^T\!}H Z$ for a phase-2 iteration. As every iteration for [`PDQP`]{} requires the solution of a [KKT]{} system, if the number of superbasic variables is not small, a phase-1 iteration of [[`SQOPT`]{}]{} requires considerably less work than an iteration of [`PDQP`]{}. It follows that the total iterations for [`PDQP`]{} and [[`SQOPT`]{}]{} are not entirely comparable. In particular a profile that would provide an accurate comparison with [`PDQP`]{} lies somewhere in-between the two [[`SQOPT`]{}]{} profiles shown.
Notwithstanding these remarks, the profile indicates that [`PDQP`]{} has comparable overall performance to a primal method that ignores the objective function while finding an initial feasible point. This provides some preliminary evidence that a combined primal-dual active set method can be an efficient and reliable alternative to conventional two-phase active-set methods. The relative performance of the proposed method is likely to increase when solving a sequence of related [QP]{}s for which the initial point for one [QP]{} is close to being the solution for the next. In this case, regardless of whether a primal or dual method is being used to solve the [QP]{}, the initial point may start off being primal or dual feasible, or the number of primal or dual infeasibilities may be small. This is typically the case for [QP]{} subproblems arising in sequential quadratic programming methods or mixed-integer [QP]{}.
Figure \[fig:bar2\] provides a bar graph of the so-called “outperforming factors” for iterations, as proposed by Morales [@Mor02]. On the $x$-axis, each bar corresponds to a particular test problem, with the problems listed in the order of Table \[tab:SQvsPDQP-auto\]. The $y$-axis indicates the factor ($\log_2$ scaled) by which one solver outperformed the other. A bar in the positive region indicates that [`PDQP`]{} outperformed [[`SQOPT`]{}]{}. A negative bar means [[`SQOPT`]{}]{} performed better. A positive (negative) dark grey bar denotes a failure in [[`SQOPT`]{}]{} ([`PDQP`]{}). Light grey bars denote a zero iteration count for a solver.
![\[fig:bar2\] Outperforming factors for total iterations for each of the 121 `CUTEst` QP problems solved using [`PDQP`]{} and [[`SQOPT`]{}]{}.](pdqp-pp2-bw.pdf){width=".97\textwidth"}
![\[fig:bar2\] Outperforming factors for total iterations for each of the 121 `CUTEst` QP problems solved using [`PDQP`]{} and [[`SQOPT`]{}]{}.](pdqp-bar2-bw.pdf){width=".97\textwidth"}
Summary and Conclusions
=======================
A pair of two-phase active-set methods, one primal and one dual, are proposed for convex quadratic programming. The methods are derived in terms of a general framework for solving a convex quadratic program with general equality constraints and simple lower bounds on the variables. In each of the methods, the search directions satisfy a [KKT]{} system of equations formed from Hessian and constraint components associated with an appropriate column basis. The composition of the basis is specified by an active-set strategy that guarantees the nonsingularity of each set of [KKT]{} equations. In addition, a combined primal-dual active set method is proposed in which a shifted dual [QP]{} is solved for a feasible point for the primal (or *vice versa*), thereby avoiding the need for an initial feasibility phase that ignores the properties of the objective function. This approach provides an effective method for finding a dual-feasible point when the QP is convex but not strictly convex. Preliminary numerical experiments indicate that this combined primal-dual active set method can be an efficient and reliable alternative to conventional two-phase active-set methods. Future work will focus on the application of the proposed methods to situations in which a series of related [QP]{}s must be solved, for example, in sequential quadratic programming methods and methods for mixed-integer nonlinear programming.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors would like to thank two referees for constructive comments that significantly improved the presentation.
Appendix
========
The appendix concerns some basic results used in previous sections. The first result shows that the nonsingularity of a [KKT]{} matrix may be established by checking that the two row blocks ${\big(\, H \ \ A^T \,\big)}$ and ${\big(\, A \ \ -M \,\big)}$ have full row rank.
\[propA-nonsing\] Assume that $H$ and $M$ are symmetric, positive semidefinite matrices. The vectors $u$ and $v$ satisfy $$\label{eqn-KKT1}
{\begin{pmatrix} H & A^T \\
A & -M \end{pmatrix}}
{\begin{pmatrix}{{\phantom-}}u \\
- v \end{pmatrix}}
= {\begin{pmatrix} 0 \\
0 \end{pmatrix}}$$ if and only if $$\label{eqn-KKT2}
{\begin{pmatrix} H \\
A \end{pmatrix}} u
= {\begin{pmatrix} 0 \\
0 \end{pmatrix}} {{\;\;}\text{and}{\;\;}}
{\begin{pmatrix} A^T \\
-M \end{pmatrix}} v
= {\begin{pmatrix} 0 \\
0 \end{pmatrix}}.$$
If (\[eqn-KKT2\]) holds, then (\[eqn-KKT1\]) holds, which establishes the “if” direction. Now assume that $u$ and $v$ are vectors such that (\[eqn-KKT1\]) holds. Then, $$u{^T\!}H u - u{^T\!}A{^T\!}v = 0, {\quad\text{and}\quad}
v{^T\!}A u + v{^T\!}M v = 0.$$ Adding these equations gives the identity $u{^T\!}H u + v{^T\!}M v = 0$. But then, the symmetry and semidefiniteness of $H$ and $M$ imply $u{^T\!}H u=0$ and $v{^T\!}M v = 0$. This can hold only if $H u = 0$ and $M v=0$. If $Hu=0$ and $M v=0$, (\[eqn-KKT1\]) gives $A{^T\!}v=0$ and $Au=0$, which implies that (\[eqn-KKT2\]) holds, which completes the proof.
The next result shows that when checking a subset of the columns of a symmetric positive semidefinite matrix for linear dependence, it is only the diagonal block that is of importance. The off-diagonal block may be ignored.
\[propA-diag\] Let $H$ be a symmetric, positive semidefinite matrix partitioned as $$H = {\begin{pmatrix} H_{11} & H_{12} \\[1pt]
H_{12}^T & H_{22}{^{\null}}\end{pmatrix}}.$$ Then, $${\begin{pmatrix} H_{11} \\[1pt]
H_{12}^T \end{pmatrix}} u
= {\begin{pmatrix} 0 \\
0 \end{pmatrix}} {{\;\;}\text{if and only if}{\;\;}}
H_{11} u = 0.$$
If $H$ is positive semidefinite, then $H_{11}$ is positive semidefinite, and it holds that $${\begin{pmatrix} 0 \\
0 \end{pmatrix}}
= {\begin{pmatrix} H_{11} \\[1pt]
H_{12}^T \end{pmatrix}} u
= {\begin{pmatrix} H_{11} & H_{12} \\[1pt]
H_{12}^T & H_{22}{^{\null}}\end{pmatrix}}
{\begin{pmatrix} u \\
0 \end{pmatrix}}$$ if and only if $$0 = {\begin{pmatrix} u^T & 0 \end{pmatrix}}
{\begin{pmatrix} H_{11} & H_{12} \\[1pt]
H_{12}^T & H_{22}{^{\null}}\end{pmatrix}}
{\begin{pmatrix} u \\[1pt]
0 \end{pmatrix}}
= u{^T\!}H_{11} u$$ if and only if $H_{11} u = 0$, as required.
In the following propositions, the distinct integers $k$ and $l$, together with integers from the index sets ${\mathcal{B}}$ and ${\mathcal{N}}$ define a partition of ${\mathcal{I}}= \{1$, $2$, …, $n\}$, i.e., ${\mathcal{I}}= {\mathcal{B}}\cup \{k\} \cup
\{l\} \cup {\mathcal{N}}$. If $w$ is any $n$-vector, the $n{_{\scriptscriptstyle B}}$-vector $w{_{\scriptscriptstyle B}}$ and $w{_{\scriptscriptstyle N}}$-vector $w{_{\scriptscriptstyle N}}$ denote the vectors of components of $w$ associated with ${\mathcal{B}}$ and ${\mathcal{N}}$. For the symmetric Hessian $H$, the matrices $H{_{\scriptscriptstyle\mathit{BB}}}$ and $H{_{\scriptscriptstyle\mathit{NN}}}$ denote the subset of rows and columns of $H$ associated with the sets ${\mathcal{B}}$ and ${\mathcal{N}}$ respectively. The unsymmetric matrix of components $h_{ij}$ with $i\in{\mathcal{B}}$ and $j\in {\mathcal{N}}$ will be denoted by $H{_{\scriptscriptstyle\mathit{BN}}}$. Similarly, $A{_{\scriptscriptstyle B}}$ and $A{_{\scriptscriptstyle N}}$ denote the matrices of columns associated with ${\mathcal{B}}$ and ${\mathcal{N}}$.
The next result concerns the row rank of the ${\big(\, A \ \ -M \,\big)}$ block of the [KKT]{} matrix.
\[propA-AMfullrowrank\] If the matrix ${\big(\, a_l \ \ a_k \ \ A{_{\scriptscriptstyle B}}\ \ -M \,\big)}$ has full row rank, and there exist ${{{{\mathit\Delta}}}x}_l$, ${{{{\mathit\Delta}}}x}_k$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$, and ${{{{\mathit\Delta}}}y}$ such that $a_l {{{{\mathit\Delta}}}x}_l + a_k {{{{\mathit\Delta}}}x}_k
+ A{_{\scriptscriptstyle B}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}+ M {{{{\mathit\Delta}}}y}=0$ with ${{{{\mathit\Delta}}}x}_k\ne 0$, then ${\big(\, a_l \ \ A{_{\scriptscriptstyle B}}\ \ -M \,\big)}$ has full row rank.
It must be established that $u{^T\!}{\begin{pmatrix} a_l & A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}=0$ implies that $u=0$. For a given $u$, let $\gamma = -u{^T\!}a_k$, so that $${\begin{pmatrix} u^T & \gamma \end{pmatrix}}
{\begin{pmatrix} a_l & a_k & A{_{\scriptscriptstyle B}}& -M \\
& 1 \end{pmatrix}}
= {\begin{pmatrix} 0 & 0 & 0 & 0 \end{pmatrix}}.$$ Then, $$0
= {\begin{pmatrix} u^T & \gamma \end{pmatrix}}
{\begin{pmatrix} a_l & a_k & A{_{\scriptscriptstyle B}}& -M \\
& 1 \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}_l \\
{{\phantom-}}{{{{\mathit\Delta}}}x}_k \\
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}y}\end{pmatrix}}
= \gamma \,{{{{\mathit\Delta}}}x}_k.$$ As ${{{{\mathit\Delta}}}x}_k\ne0$, it must hold that $\gamma = 0$, in which case $$u^T {{\bigl(\; a_l {\;\;}a_k {\;\;}A{_{\scriptscriptstyle B}}{\;\;}-M \;\bigr)}} = 0.$$ As ${\big(\, a_l \ \ a_k \ \ A{_{\scriptscriptstyle B}}\ \ -M \,\big)}$ has full row rank by assumption, it follows that $u=0$ and ${\big(\, a_l \ \ A{_{\scriptscriptstyle B}}\ \ -M \,\big)}$ must have full row rank.
An analogous result holds concerning the ${\big(\, H \ \ A^T \,\big)}$ block of the [KKT]{} matrix.
\[propA-HATfullrowrank\] If ${\big(\, H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}\ \ A{_{\scriptscriptstyle B}}^T \,\big)}$ has full row rank, and there exist quantities ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle N}}$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$, ${{{{\mathit\Delta}}}y}$, and ${{{{\mathit\Delta}}}z}_k$ such that $$\label{eqn-HATfullrowrank}
{\begin{pmatrix} h{_{{\scriptscriptstyle\mathit{Nk}}}}^T & h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T & 1 \\[1pt]
h{_{\scriptscriptstyle\mathit{BN}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle N}}\\
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
-{{{{\mathit\Delta}}}y}\\
-{{{{\mathit\Delta}}}z}_k \end{pmatrix}}
= {\begin{pmatrix} 0 \\
0 \end{pmatrix}},$$ with ${{{{\mathit\Delta}}}z}_k \ne 0$, then the matrix $${\begin{pmatrix} h{_{kk}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T \\[1pt]
h{_{{\scriptscriptstyle\mathit{Bk}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \end{pmatrix}}$$ has full row rank.
Let ${\big(\, \mu \ \ v^T \,\big)}$ be any vector such that $${{\bigl(\, \mu {\;\;}v^T \,\bigr)}}
{\begin{pmatrix} h{_{{\scriptscriptstyle\mathit{Nk}}}}^T & h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T \\[1pt]
h{_{\scriptscriptstyle\mathit{BN}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \end{pmatrix}}
= {\begin{pmatrix}0 & 0 & 0 \end{pmatrix}}.$$ The assumed identity (\[eqn-HATfullrowrank\]) gives $$0 = {{\bigl(\, \mu {\;\;}v^T \,\bigr)}}
{\begin{pmatrix} h{_{{\scriptscriptstyle\mathit{Nk}}}}^T & h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T \\[1pt]
h{_{\scriptscriptstyle\mathit{BN}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle N}}\\
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
-{{{{\mathit\Delta}}}y}\end{pmatrix}}
= \mu \,{{{{\mathit\Delta}}}z}_k.$$ As ${{{{\mathit\Delta}}}z}_k\ne 0$ by assumption, it must hold that $\mu=0$. The full row rank of ${\big(\, H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}\ \ A{_{\scriptscriptstyle B}}^T \,\big)}$ then gives $v = 0$ and $${\begin{pmatrix} h{_{{\scriptscriptstyle\mathit{Nk}}}}^T & h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T \\[1pt]
h{_{\scriptscriptstyle\mathit{BN}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \end{pmatrix}}$$ must have full row rank. Proposition \[propA-nonsing\] implies that this is equivalent to $${\begin{pmatrix} h{_{kk}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T \\[1pt]
h{_{{\scriptscriptstyle\mathit{Bk}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \end{pmatrix}}$$ having full row rank.
The next proposition concerns the primal subiterations when the constraint index $k$ is moved from ${\mathcal{B}}$ to ${\mathcal{N}}$. In particular, it is shown that the $K_l$ matrix is nonsingular after a subiteration.
\[propA-primalnonsing\] Assume that $({{{{\mathit\Delta}}}x}_l$, ${{{{\mathit\Delta}}}x}_k$, ${{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}$, $-{{{{\mathit\Delta}}}y}$, $-{{{{\mathit\Delta}}}z}_l)$ is the unique solution of the equations $$\label{eqn-primaldir}
{\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{kl}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T & {{\phantom-}}1 \\[1pt]
h{_{kl}{^{\null}}}& h{_{kk}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T & \\[1pt]
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T & \\[1pt]
a_l & a_k & A{_{\scriptscriptstyle B}}& -M & \\[1pt]
1 & & & & -1 \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}_l \\[1pt]
{{\phantom-}}{{{{\mathit\Delta}}}x}_k \\[1pt]
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\[1pt]
-{{{{\mathit\Delta}}}y}\\[1pt]
-{{{{\mathit\Delta}}}z}_l \end{pmatrix}}
= {\begin{pmatrix} 0 \\[1pt]
0 \\[1pt]
0 \\[1pt]
0 \\[1pt]
1 \end{pmatrix}},$$ and that ${{{{\mathit\Delta}}}x}_k\ne 0$. Then, the matrices $K_l$ and $K_k$ are nonsingular, where $$K_l = {\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T \\[1pt]
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \\[1pt]
a_l & A{_{\scriptscriptstyle B}}& -M \end{pmatrix}} {\quad\text{and}\quad}
K_k = {\begin{pmatrix} h{_{kk}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T \\[1pt]
h{_{{\scriptscriptstyle\mathit{Bk}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \\[1pt]
a_k & A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}.$$
By assumption, the equations (\[eqn-primaldir\]) have a unique solution with ${{{{\mathit\Delta}}}x}_k\ne 0$. This implies that there is no solution of the overdetermined equations $$\label{eqn-baseprimal}
{\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{kl}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T &{{\phantom-}}1 \\[1pt]
h{_{kl}{^{\null}}}& h{_{kk}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T & \\[1pt]
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T & \\[1pt]
a_l & a_k & A{_{\scriptscriptstyle B}}& -M & \\[1pt]
1 & & & & -1 \\[1pt]
& 1 & & & \end{pmatrix}}
{\begin{pmatrix}{{\phantom-}}{{{{\mathit\Delta}}}x}_l \\[1pt]
{{\phantom-}}{{{{\mathit\Delta}}}x}_k \\[1pt]
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\[1pt]
- {{{{\mathit\Delta}}}y}\\[1pt]
- {{{{\mathit\Delta}}}z}_l \end{pmatrix}}
= {\begin{pmatrix} 0 \\[1pt]
0 \\[1pt]
0 \\[1pt]
0 \\[1pt]
1 \\[1pt]
0 \end{pmatrix}}.$$ Given an arbitrary matrix $D$ and nonzero vector $f$, the fundamental theorem of linear algebra implies that if $D w = f$ has no solution, then there exists a vector $v$ such that $v{^T\!}f \ne 0$. The application of this result to (\[eqn-baseprimal\]) implies the existence of a nontrivial vector $({{{{\mathit\Delta}}}{\skew3\widetilde x}}_l$, ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_k$, ${{{{\mathit\Delta}}}{\skew3\widetilde x}}{_{\scriptscriptstyle B}}$, $-{{{{\mathit\Delta}}}{\skew3\widetilde y}}$, $-{{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l$, $-{{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_k)$ such that $$\label{eqn-baseprimal-alt}
{\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{kl}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T & {{\phantom-}}1 & \\[1pt]
h{_{kl}{^{\null}}}& h{_{kk}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T & &{\settowidth{\argwidth}{$a_l^T$}
\makebox[\argwidth][c]{$1$}}\\[1pt]
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T & & \\[1pt]
a_l & a_k & A{_{\scriptscriptstyle B}}& -M & & \\[1pt]
1 & & & & -1 & \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}{\skew3\widetilde x}}_l \\
{{\phantom-}}{{{{\mathit\Delta}}}{\skew3\widetilde x}}_k \\
{{\phantom-}}{{{{\mathit\Delta}}}{\skew3\widetilde x}}{_{\scriptscriptstyle B}}\\
- {{{{\mathit\Delta}}}{\skew3\widetilde y}}\\
- {{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l \\
- {{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_k \end{pmatrix}}
= {\begin{pmatrix} 0 \\[1pt]
0 \\[1pt]
0 \\[1pt]
0 \\[1pt]
0 \end{pmatrix}},$$ with ${{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l\ne 0$. The last equation of (\[eqn-baseprimal-alt\]) gives ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_l + {{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l=0$, in which case ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_l {{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l =
-{{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l^2 <0$ because ${{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l\ne0$. Any solution of (\[eqn-baseprimal-alt\]) may be viewed as a solution of the equations $H {{{{\mathit\Delta}}}{\skew3\widetilde x}}- A{^T\!}{{{{\mathit\Delta}}}{\skew3\widetilde y}}- {{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}= 0$, $A {{{{\mathit\Delta}}}{\skew3\widetilde x}}+ M {{{{\mathit\Delta}}}{\skew3\widetilde y}}= 0$, ${{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}{_{\scriptscriptstyle B}}=0$, and ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_i = 0$ for $i\in \{ 1$, $2$, …, $n\}
\setminus \{ l \} \setminus \{ k \}$. An argument similar to that used to establish Proposition \[prop-dotproduct\] gives $${{{{\mathit\Delta}}}{\skew3\widetilde x}}_l {{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l + {{{{\mathit\Delta}}}{\skew3\widetilde x}}_k {{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_k \ge 0,$$ which implies that ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_k {{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_k > 0$, with ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_k\ne0$ and ${{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_k\ne 0$.
As the search direction is unique, it follows from (\[eqn-primaldir\]) that ${\big(\, h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}\ \ H{_{{\scriptscriptstyle\mathit{Bk}}}{^{\null}}}\ \ H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}\ \ A{_{\scriptscriptstyle B}}^T \,\big)}$ has full row rank, and Proposition \[propA-diag\] implies that ${\big(\, H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}\ \ A{_{\scriptscriptstyle B}}^T \,\big)}$ has full row rank. Hence, as ${{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l\ne0$, it follows from (\[eqn-baseprimal-alt\]) and Proposition \[propA-HATfullrowrank\] that the matrix $${\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{kl}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T \\[1pt]
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \end{pmatrix}}$$ has full row rank, which is equivalent to the matrix $${\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T \\[1pt]
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \end{pmatrix}}$$ having full row rank by Proposition \[propA-diag\],
Again, the search direction is unique and (\[eqn-primaldir\]) implies that ${\big(\, a_l \ \ a_k \ \ A{_{\scriptscriptstyle B}}\ \ -M \,\big)}$ has full row rank. As ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_k\ne 0$, Proposition \[propA-AMfullrowrank\] implies that ${\big(\, a_l \ \ A{_{\scriptscriptstyle B}}\ \ -M \,\big)}$ must have full row rank. Consequently, Proposition \[propA-nonsing\] implies that $K_l$ is nonsingular.
As ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_k$, ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_l$, ${{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_k$ and ${{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l$ are all nonzero, the roles of $k$ and $l$ may be reversed to give the result that $K_k$ is nonsingular.
The next proposition concerns the situation when a constraint index $k$ is moved from ${\mathcal{N}}$ to ${\mathcal{B}}$ in a dual subiteration. In particular, it is shown that the resulting matrix $K{_{\scriptscriptstyle B}}$ defined after a subiteration is nonsingular.
\[propA-dualnonsing\] Assume that there is a unique solution of the equations $$\label{eqn-dualdir}
{\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{kl}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T & {{\phantom-}}1 \\[1pt]
h{_{kl}{^{\null}}}& h{_{kk}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T & & {{\phantom-}}1 \\[1pt]
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T & \\[1pt]
a_l & a_k & A{_{\scriptscriptstyle B}}& -M & \\[1pt]
1 & & & & -1 \\[1pt]
& 1 \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}_l \\
{{\phantom-}}{{{{\mathit\Delta}}}x}_k \\
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
-{{{{\mathit\Delta}}}y}\\
-{{{{\mathit\Delta}}}z}_l \\
-{{{{\mathit\Delta}}}z}_k \end{pmatrix}}
= {\begin{pmatrix} 0 \\[1pt]
0 \\[1pt]
0 \\[1pt]
0 \\[1pt]
1 \\[1pt]
0 \end{pmatrix}},$$ with ${{{{\mathit\Delta}}}z}_k\ne 0$. Then, the matrices $K_l$ and $K_k$ are nonsingular, where $$K_l = {\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T \\[1pt]
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \\[1pt]
a_l & A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}, {\quad\text{and}\quad}
K_k = {\begin{pmatrix} h{_{kk}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T \\[1pt]
h{_{{\scriptscriptstyle\mathit{Bk}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T \\[1pt]
a_k & A{_{\scriptscriptstyle B}}& -M \end{pmatrix}}.$$
As (\[eqn-dualdir\]) has a unique solution with ${{{{\mathit\Delta}}}z}_k\ne 0$, there is no solution of $$\label{eqn-basedual}
{\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{kl}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T & {{\phantom-}}1 \\[1pt]
h{_{kl}{^{\null}}}& h{_{kk}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T & \\[1pt]
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T & \\[1pt]
a_l & a_k & A{_{\scriptscriptstyle B}}& -M & \\[1pt]
1 & & & & -1 \\[1pt]
& 1 \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}x}_l \\
{{\phantom-}}{{{{\mathit\Delta}}}x}_k \\
{{\phantom-}}{{{{\mathit\Delta}}}x}{_{\scriptscriptstyle B}}\\
-{{{{\mathit\Delta}}}y}\\
-{{{{\mathit\Delta}}}z}_l \end{pmatrix}}
= {\begin{pmatrix} 0 \\[1pt]
0 \\[1pt]
0 \\[1pt]
0 \\[1pt]
1 \\[1pt]
0 \end{pmatrix}}.$$ The fundamental theorem of linear algebra applied to (\[eqn-basedual\]) implies the existence of a solution of $$\label{eqn-basedual-alt}
{\begin{pmatrix} h{_{ll}{^{\null}}}& h{_{kl}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bl}}}}^T & a_l^T & {{\phantom-}}1 \\[1pt]
h{_{kl}{^{\null}}}& h{_{kk}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}}^T & a_k^T & &{\settowidth{\argwidth}{$a_l^T$}
\makebox[\argwidth][c]{$1$}}\\[1pt]
h{_{{\scriptscriptstyle\mathit{Bl}}}{^{\null}}}& h{_{{\scriptscriptstyle\mathit{Bk}}}{^{\null}}}& H{_{\scriptscriptstyle\mathit{BB}}{^{\null}}}& A{_{\scriptscriptstyle B}}^T & \\[1pt]
a_l & a_k & A{_{\scriptscriptstyle B}}& -M & \\[1pt]
1 & & & & -1 \end{pmatrix}}
{\begin{pmatrix} {{\phantom-}}{{{{\mathit\Delta}}}{\skew3\widetilde x}}_l \\
{{\phantom-}}{{{{\mathit\Delta}}}{\skew3\widetilde x}}_k \\
{{\phantom-}}{{{{\mathit\Delta}}}{\skew3\widetilde x}}{_{\scriptscriptstyle B}}\\
-{{{{\mathit\Delta}}}{\skew3\widetilde y}}\\
-{{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l \\
-{{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_k \end{pmatrix}}
= {\begin{pmatrix} 0 \\[1pt]
0 \\[1pt]
0 \\[1pt]
0 \\[1pt]
0 \end{pmatrix}},$$ with ${{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l\ne0$. It follows from (\[eqn-basedual-alt\]) that ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_l+{{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l=0$. As ${{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l\ne0$, this implies ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_l {{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l<0$. The solution of (\[eqn-basedual-alt\]) may be regarded as a solution of the homogeneous equations $H {{{{\mathit\Delta}}}x}- A{^T\!}{{{{\mathit\Delta}}}y}- {{{{\mathit\Delta}}}z}= 0$, $A {{{{\mathit\Delta}}}x}+ M {{{{\mathit\Delta}}}y}= 0$, with ${{{{\mathit\Delta}}}z}_i = 0$, for $i\in{\mathcal{B}}$, and ${{{{\mathit\Delta}}}x}_i =
0$, for $i\in\{1,\dots,n\}\setminus\{k\}\setminus\{l\}$. Hence, Proposition \[prop-dotproduct\] gives $${{{{\mathit\Delta}}}{\skew3\widetilde x}}_l {{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l + {{{{\mathit\Delta}}}{\skew3\widetilde x}}_k {{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_k \ge 0,$$ so that ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_k {{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_k > 0$. Hence, it must hold that ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_k\ne0$ and ${{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_k\ne 0$.
As ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_k\ne0$, ${{{{\mathit\Delta}}}{\skew3\widetilde x}}_l\ne0$, ${{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_k\ne0$ and ${{{{\mathit\Delta}}}{\skew{2.8}\widetilde z}}_l\ne0$, the remainder of the proof is analogous to that of Proposition \[propA-primalnonsing\].
The next result gives expressions for the primal and dual objective functions in terms of the computed search directions.
\[propA-dirs\] Assume that $(x,y,z)$ satisfies the primal and dual equality constraints $$H x + c - A{^T\!}y - z = 0, {\quad\text{and}\quad} Ax + M y - b = 0.$$ Consider the partition $\{1$, $2$, …, $n\}= {\mathcal{B}}\cup\{l\}\cup{\mathcal{N}}$ such that $x{_{\scriptscriptstyle N}}+q{_{\scriptscriptstyle N}}=0$ and $z{_{\scriptscriptstyle B}}+r{_{\scriptscriptstyle B}}=0$. If the components of the direction $({{{{\mathit\Delta}}}x}$, ${{{{\mathit\Delta}}}y}$, ${{{{\mathit\Delta}}}z})$ satisfy [(\[eqn-dxdydz\])]{}, then the primal and dual objective functions for [$({\mbox{\small PQP}}_{q,r})$]{} and [$({\mbox{\small DQP}}_{q,r})$]{}, i.e., $$\begin{aligned}
f{_{\scriptscriptstyle P}}(x,y) &= {{\phantom-}}{{\textstyle\frac12}}x{^T\!}H x + {{\textstyle\frac12}}y{^T\!}M y + c{^T\!}x + r{^T\!}x \\
f{_{\scriptscriptstyle D}}(x,y,z) &= - {{\textstyle\frac12}}x{^T\!}H x - {{\textstyle\frac12}}y{^T\!}M y + b{^T\!}y - q{^T\!}z,\end{aligned}$$ satisfy the identities $$\begin{aligned}
f{_{\scriptscriptstyle P}}(x+\alpha{{{{\mathit\Delta}}}x},y+\alpha{{{{\mathit\Delta}}}y})
&= f{_{\scriptscriptstyle P}}(x,y)+ {{{{\mathit\Delta}}}x}_l(z_l+r_l) \alpha + {{\textstyle\frac12}}{{{{\mathit\Delta}}}x}_l {{{{\mathit\Delta}}}z}_l \alpha^2, \\
f{_{\scriptscriptstyle D}}(x+\alpha{{{{\mathit\Delta}}}x},y+\alpha{{{{\mathit\Delta}}}y},z+\alpha{{{{\mathit\Delta}}}z})
&= f{_{\scriptscriptstyle D}}(x,y,z) - {{{{\mathit\Delta}}}z}_l(x_l+q_l) \alpha - {{\textstyle\frac12}}{{{{\mathit\Delta}}}x}_l {{{{\mathit\Delta}}}z}_l \alpha^2.\end{aligned}$$
The directional derivative of the primal objective function is given by
\[eqn-Pdirder\] $$\begin{aligned}
{\begin{pmatrix} {{{{\mathit\Delta}}}x}^T & {{{{\mathit\Delta}}}y}^T \end{pmatrix}} {\nabla\!}f{_{\scriptscriptstyle P}}(x,y)
&= {\begin{pmatrix} {{{{\mathit\Delta}}}x}^T & {{{{\mathit\Delta}}}y}^T \end{pmatrix}}
{\begin{pmatrix} Hx + c + r \\
M y \end{pmatrix}} \notag \\
&= {\begin{pmatrix} {{{{\mathit\Delta}}}x}^T & {{{{\mathit\Delta}}}y}^T \end{pmatrix}}
{\begin{pmatrix} A{^T\!}y + z + r \\
M y \end{pmatrix}} \label{eqn-Pdirderb}\\
&= (A{{{{\mathit\Delta}}}x}+ M{{{{\mathit\Delta}}}y}){^T\!}y + {{{{\mathit\Delta}}}x}{^T\!}( z + r)
= {{{{\mathit\Delta}}}x}_l ( z_l + r_l ), \label{eqn-Pdirderd}\end{aligned}$$
where the identity $Hx+c = A{^T\!}y +z$ has been used in (\[eqn-Pdirderb\]) and the identities $A{{{{\mathit\Delta}}}x}+ M{{{{\mathit\Delta}}}y}=0$, ${{{{\mathit\Delta}}}x}_N=0$ and $z{_{\scriptscriptstyle B}}+r{_{\scriptscriptstyle B}}=0$ have been used in (\[eqn-Pdirderd\]).
The curvature in the direction $({{{{\mathit\Delta}}}x},{{{{\mathit\Delta}}}y})$ is given by $$\label{eqn-Pdircurv}
{\begin{pmatrix} {{{{\mathit\Delta}}}x}^T & {{{{\mathit\Delta}}}y}^T \end{pmatrix}} {\nabla^2\!}f{_{\scriptscriptstyle P}}(x,y) {\begin{pmatrix} {{{{\mathit\Delta}}}x}\\ {{{{\mathit\Delta}}}y}\end{pmatrix}}
= {\begin{pmatrix} {{{{\mathit\Delta}}}x}^T & {{{{\mathit\Delta}}}y}^T \end{pmatrix}}
{\begin{pmatrix} H & \\
& M \end{pmatrix}}
{\begin{pmatrix} {{{{\mathit\Delta}}}x}\\ {{{{\mathit\Delta}}}y}\end{pmatrix}}
= {{{{\mathit\Delta}}}x}_l {{{{\mathit\Delta}}}z}_l,$$ where the last equality follows from Proposition \[prop-dotproduct\].
The directional derivative of the dual objective function is given by
\[eqn-Ddirder\] $$\begin{aligned}
{\begin{pmatrix} {{{{\mathit\Delta}}}x}^T & {{{{\mathit\Delta}}}y}^T & {{{{\mathit\Delta}}}z}^T \end{pmatrix}} {\nabla\!}f{_{\scriptscriptstyle D}}(x,y,z)
&= {\begin{pmatrix} {{{{\mathit\Delta}}}x}^T & {{{{\mathit\Delta}}}y}^T & {{{{\mathit\Delta}}}z}^T \end{pmatrix}}
{\begin{pmatrix} -Hx \\
-My + b \\
-q \end{pmatrix}} \\
&= - {{{{\mathit\Delta}}}x}{^T\!}H x + {{{{\mathit\Delta}}}y}{^T\!}(-My + b) - {{{{\mathit\Delta}}}z}^T q \\
&= - (A{^T\!}{{{{\mathit\Delta}}}y}+ {{{{\mathit\Delta}}}z}){^T\!}x + {{{{\mathit\Delta}}}y}{^T\!}(-My + b)
- {{{{\mathit\Delta}}}z}^T q \label{eqn-Ddirderc} \\
&= -{{{{\mathit\Delta}}}y}{^T\!}(Ax + My - b)- {{{{\mathit\Delta}}}z}{^T\!}(x+q) \\
&= - {{{{\mathit\Delta}}}z}_l (x_l+q_l), \label{eqn-Ddirdere}\end{aligned}$$
where the identity $H {{{{\mathit\Delta}}}x}-A{^T\!}{{{{\mathit\Delta}}}y}-{{{{\mathit\Delta}}}z}=0$ has been used in (\[eqn-Ddirderc\]) and the identities $Ax + My - b = 0$, $x{_{\scriptscriptstyle N}}+q{_{\scriptscriptstyle N}}=0$ and ${{{{\mathit\Delta}}}z}{_{\scriptscriptstyle B}}=0$ have been used in (\[eqn-Ddirdere\]).
As $z$ only appears linearly in the dual objective function, it follows from the structure of the Hessian matrices of $f{_{\scriptscriptstyle P}}(x,y)$ and $f{_{\scriptscriptstyle D}}(x,y,z)$ in combination with (\[eqn-Pdircurv\]) that $$\begin{aligned}
{\begin{pmatrix} {{{{\mathit\Delta}}}x}^T & {{{{\mathit\Delta}}}y}^T & {{{{\mathit\Delta}}}z}^T\end{pmatrix}} {\nabla^2\!}f{_{\scriptscriptstyle D}}(x,y,z)
{\begin{pmatrix} {{{{\mathit\Delta}}}x}\\ {{{{\mathit\Delta}}}y}\\ {{{{\mathit\Delta}}}z}\end{pmatrix}}
&= -{\begin{pmatrix} {{{{\mathit\Delta}}}x}^T & {{{{\mathit\Delta}}}y}^T\end{pmatrix}}{\nabla^2\!}f{_{\scriptscriptstyle P}}(x,y) {\begin{pmatrix} {{{{\mathit\Delta}}}x}\\ {{{{\mathit\Delta}}}y}\end{pmatrix}} \\
&= -{{{{\mathit\Delta}}}x}_l {{{{\mathit\Delta}}}z}_l.
$$
The final result shows that there is no loss of generality in assuming that ${\big(\, A \ \ M \,\big)}$ has full row rank in $({\mbox{\small PQP}}_{q,r})$.
\[prop-AMfullrowrank\] There is no loss of generality in assuming that ${\big(\, A \ \ M \,\big)}$ has full row rank in [$({\mbox{\small PQP}}_{q,r})$]{}.
Let $(x$, $y$, $z)$ be any vector satisfying (\[eqn-optgradLzero\]) and (\[eqn-optfeasprimallin\]). Assume that ${\big(\, A \ \ M \,\big)}$ has linearly dependent rows, and that ${\big(\, A \ \ M \,\big)}$ and $b$ may be partitioned conformally such that $${{\bigl(\, A {\;\;}M \,\bigr)}}
= {\begin{pmatrix}A_1 & M_{11} & M_{12} \\[1pt]
A_2{^{\null}}& M_{12}^T & M_{22}{^{\null}}\end{pmatrix}}, {\quad\text{and}\quad}
b
= {\begin{pmatrix} b_1 \\ b_2 \end{pmatrix}},$$ with ${\big(\, A_{1} \ \ M_{11} \ \ M_{12} \,\big)}$ having full row rank, and $$\label{eqn-AMpart}
{\begin{pmatrix} A_2{^{\null}}& M_{12}^T & M_{22}{^{\null}}\end{pmatrix}}
= N {\begin{pmatrix} A_1 & M_{11} & M_{12}\end{pmatrix}},$$ with $A_1\in{\mathbb{R}}^{m_1\times n}$ and $A_2\in{\mathbb{R}}^{m_2\times n}$ for some matrix $N\in{\mathbb{R}}^{m_2\times m_1}$. From the linear dependence of the rows of ${\big(\, A \ \ M \,\big)}$, it follows that $x$, $y$ and $z$ satisfy (\[eqn-optgradLzero\]) and (\[eqn-optfeasprimallin\]) if and only if $$\begin{aligned}
Hx+c - A_1^T y_1{^{\null}}- A_2^T y_2{^{\null}}- z &= 0, \\
A_1 x + M_{11} y_1 + M_{12} y_2 - b_1&= 0 {\quad\text{and}\quad} b_2 = N b_1.\end{aligned}$$ It follows from (\[eqn-AMpart\]) that $M_{12} = M_{11} N^T$ and $A_2^T=A_1^T N^T$, so that $x$, $y$ and $z$ satisfy (\[eqn-optgradLzero\]) and (\[eqn-optfeasprimallin\]) if and only if $$\begin{aligned}
Hx + c - A_1^T( y_1{^{\null}}+ N{^T\!}y_2{^{\null}}) - z &= 0, \\
A_1 x + M_{11}( y_1 + N{^T\!}y_2) - b_1 &= 0 {{\;\;}\text{and}{\;\;}} b_2 = N b_1.\end{aligned}$$ We may now define ${\skew3\widetilde y}_1=y_1+N{^T\!}y_2$ and replace (\[eqn-optfeasprimallin\]) and (\[eqn-optgradLzero\]) by the system $$\begin{aligned}
H x + c - A_1^T{\skew3\widetilde y}_1{^{\null}}- z &= 0, \\
A_1 x + M_{11}{\skew3\widetilde y}_1 - b_1 &= 0.\end{aligned}$$ By assumption, ${\big(\, A_{1} \ \ M_{11} \ \ M_{12} \,\big)}$ has full row rank. Proposition \[propA-diag\] implies that ${\big(\, A_{1} \ \ M_{11} \,\big)}$ has full row rank. This gives an equivalent problem for which ${\big(\, A_{1} \ \ M_{11} \,\big)}$ has full row rank.
[^1]: Optimization and Systems Theory, Department of Mathematics, KTH Royal Institute of Technology, SE-100 44 Stockholm, Sweden (`[email protected]`). Research supported by the Swedish Research Council (VR).
[^2]: Department of Mathematics, University of California, San Diego, La Jolla, CA 92093-0112 (`[email protected]`, `[email protected]`). Research supported in part by Northrop Grumman Aerospace Systems, and National Science Foundation grants DMS-1318480 and DMS-1361421.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Automatic text summarisation has drawn considerable interest in the field of software engineering. It can improve the efficiency of software developers, enhance the quality of products, and ensure timely delivery. In this paper, we present our initial work towards automatically generating human-like multi-document summaries from heterogeneous software artefacts. Our analysis of the text properties of 545 human-written summaries from 15 software engineering projects will ultimately guide heuristics searches in the automatic generation of human-like summaries.'
author:
- 'Mahfouth Alghamdi, Christoph Treude, Markus Wagner'
title: 'Toward Human-Like Summaries Generated from Heterogeneous Software Artefacts'
---
<ccs2012> <concept> <concept\_id>10011007.10011074.10011784</concept\_id> <concept\_desc>Software and its engineering Search-based software engineering</concept\_desc> <concept\_significance>500</concept\_significance> </concept> </ccs2012>
[12]{}
\#1 \#1[\#1]{}\#1 \#1 \#1 \#1 \#1[\#1]{} \#1[\#1]{}
[Allahyari]{} . . ().
[Chand2015emo]{} . . , (), .
[Dabbish]{} . . In . .
[Hahn]{} . . , (), .
[harman:sbse]{} . . , (), .
[Luhn]{} . . , (), .
[Maaten]{} . . (), .
[Moreno]{} . . In . .
[Rastkar]{} . . , (), .
[Rigby]{} . . In .
[Sridhara]{} . . In .
[Treude2015]{} . . In .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Mapping the QCD phase boundary and locating critical end point still remains as an open problem in strong interaction physics. Predictions about the co-ordinates of the critical point in the $(T, \mu_B)$ plane, from different QCD motivated models show a wide variation. Lattice QCD calculations are also available, that give an estimation of the critical point for chiral phase transition, where the transition changes its nature from rapid cross over to first order transition. Recently co-ordinates of the critical point for deconfinement phase transition are claimed to be found as an endpoint of the first order phase transition line, in Bag model scenario. In the present paper we have shown that Bag model gives a complete first order phase transition line in the $(T, \mu_B)$ plane, and one can not have any point where the transition changes its nature.'
author:
- Partha Pratim Bhaduri
- Victor Roy
- 'A. K. Chaudhuri'
title: 'The phase diagram of QCD in a BAG+HRG based equation of state: appearance of a pseudo-critical point'
---
In recent years there is much interest in QCD phase diagram and in particular about the location of QCD critical end point. That a critical end point exists in the QCD phase diagram can be argued as follows: finite temperature lattice simulation of baryon free ($\mu_B=0$) QCD indicate a cross-over transition at $T_c\approx 170\pm 20$ MeV [@Cheng:2007jq], [@Soltz:2009bc], [@Aoki:2009sc], [@Aoki:2006br]. On the other hand there are indications that at zero temperature but at finite baryon density, QCD has a 1st order transition [@Asakawa:1989bq], [@Berges:1998rc], [@Barducci:1989eu]. Only with the existence of a critical end point where the 1st order transition ends, the cross over transition at $(T,\mu_B=0)$ can be reconciled with a 1st order transition at $(T=0,\mu_B)$. Unfortunately, lattice simulations at finite baryon density is plugged with the well known fermion sign problem and at the present state of art, location of QCD phase boundary or the critical end point is beyond the lattice simulations. Several methods have been devised to surpass the fermion problem, e.g Taylor expansion [@Gavai:2004sd], [@Gavai:2003mf], re-weighting [@Barbour:1997bh], [@Allton:2002zi], [@Fodor:2001pe], imaginary chemical potential [@de; @Forcrand:2002ci], [@de; @Forcrand:2003hx], [@deForcrand:2006pv]. However, it is uncertain whether the methods mentioned above are valid at large baryon density. QCD phase diagram has been studied in QCD inspired models, e.g. instanton models, linear sigma model, Nambu-Jona-Lasinio (NJL) or Polyakov loop extended Nambu-Jona-Lasinio models [@Berges:1998rc], [@Alford:1997zt], [@Rapp:1997zu], [@Asakawa:1989bq], [@Meisinger:1995ih], [@Ghosh:2006qh]. In QCD inspired model, in general quark degrees of freedom are integrated out. The model results depend on the cut off parameter. Depending on the model parameters, QCD phase diagram and location of critical end point vary widely.
In a recent work, Singh et. al. [@Singh:2009jd], have obtained the QCD phase boundary for de-confinement transition by using Gibb’s equilibrium criterion for first order phase transition between QGP and hadronic phase. The hadronic phase is modeled by a gas of interacting hadrons, where the geometrical size of the baryons is explicitly incorporated as the excluded volume correction, in a thermodynamically consistent manner. All baryons, mesons and their resonances having masses up to 2 GeV are included in calculation and strangeness conservation is taken into account by equating the net strangeness to zero. Mesons are considered as point-like particles, where as an equal volume $v_{ex}=\frac{4\pi}{3}r_c^3$ is assumed for each type of baryon with a hard-core radius $r_c = 0.6 - 0.8 fm$. In addition, full quantum statistics is included in the partition function of the grand canonical ensemble which helps to navigate the region of the phase diagram with low T and high $\mu_B$. Thermodynamic functions of weakly interacting quark matter is obtained by a simple Bag model equation of state (EOS) with perturbative corrections of the order of $\alpha_{s}^{3/2}$ in strong interaction coupling constant $\alpha_{s}$. Having modeled the two phases, the QCD phase boundary is obtained advocating Gibb’s equilibrium criterion for first order phase transition. Gibb’s criterion demands that at the transition point the pressure of the hadronic state $P_H(T,\mu_B)$ and that of the qgp state $P_Q(T,\mu_B)$ becomes equal.
Starting from a low but non-zero value of T and large value of $\mu_B$, gradually they have moved towards large T and small $\mu_B$. At a given T, a corresponding value of $\mu_B$ is found out at which the pressure equality holds and beyond which QGP pressure dominates. It has been observed that the line of co-existence between two phases ends at a point which is being interpreted as the end point of the 1st order transition line, or the critical end point. Thus the precise co-ordinates of the QCD critical point has been claimed to be estimated for deconfinement phase transition in the $\mu_B-T$ plane. The critical values of temperature and chemical potential are found to be $(T_c = 160 MeV, \mu_c = 156 MeV)$ for hard-core radius $r_c = 0.6 fm$ and Bag constant $B^{1/4} = 216 ~MeV$. For $r_c = 0.8 fm$ and $B^{1/4} = 200 ~MeV$, the point is shifted down in temperature to $(T_c = 146 MeV, \mu_c = 156 MeV)$. Variation in QCD scale parameter$\Lambda$ is reported to give insignificant change in the location of the phase boundary. Considering that earlier investigation [@Satz], [@Dixit], [@Cleymans1], [@Magas], [@Castorina] of phase diagram in terms of Bag model EOS and hadronic resonance gas fails to detect the QCD critical point, we have studied the model in detail. We find that the simple interpretation of minimum of the chemical potential below which co-existence line ceases to exist as the critical end point is misleading and appears due to lack of sufficient numerical precision.
In the present work we have employed the same Bag model EOS as used in [@Singh:2009jd] to compute the thermodynamic parameters of the partonic phase. In this model, QGP is assumed to consist of massless quarks (u,d), their antiquarks and gluons only and the pressure in the QGP phase takes the form
$$\begin{split}
P_{QGP}
= \frac{37}{90}\pi^2 T^4 + \frac1{9}\mu_B^2 T^2 + \frac{\mu_B^4}{162 \pi^2}
\\
- \alpha_S\left[\frac{11}{9}\pi T^4 + \frac{2}{9\pi^2}\mu_B^2 T^2 + \frac1{81\pi^3}\mu_B^4\right ]
\\
+\frac{8\alpha_S^{3/2} T}{3\pi^2\sqrt{2\pi}}{\left[\frac{8\pi^2 T^2}{3} +\frac2{9}\mu_B^2\right ]^{3/2}} - B
\end{split}$$
where $\mu_B$, T dependence of $\alpha_S$ can be given as \[15\]:
$$\alpha_S
= \frac{12\pi}{29}\left[ln (\frac{0.089 \mu_B^2 + 15.622 T^2}{\Lambda^2})\right]^{-1}$$
\
Here we have used $B^{1/4} = 200 ~MeV$ and $\Lambda = 100 ~MeV$ in our calculation.
For simulating the hadronic phase we considered a hadron resonance gas model (HRG) which includes all hadrons and their resonances having masses up to 2.5 GeV. Excluded volume correction has also been incorporated following the prescription by Cleymans and Suhonen [@Cleymans2]. In this approach, for a given eigen volume $v_{ex}$, the excluded volume corrected pressure, is obtained as,
$$P^{excl}_H(T,\mu_B)=\frac{P^{id}_B(T,\mu_B)}{1+v_{ex} n_B^{id}(T,\mu_B)} + P^{id}_M(T,\mu_B) \\$$
Here $P^{id}_{B(M)} (T,\mu_B)$ is the pressure of the corresponding the ideal gas having point like baryons (mesons) and can be computed as
$$\begin{aligned}
P^{id}_{B(M)}(T,\mu_B)&=&
\sum_i\frac{g_i}{(2\pi)^3}\int d^3p \frac{p^2}{3\sqrt{m_i^2+p^2}} \nonumber\\
&\times&\frac{1}{e^{(\sqrt{m_i^2+p^2}-\mu_i n_i)/T}\pm 1}\end{aligned}$$
where the sum is over all the baryonic (mesonic) species and their resonances included in the calculation.
Though not thermodynamically consistent, but the resulting reduced pressure obtained by this approach is in close agreement with the approach employed by Singh et al [@Madhukar], which preserves thermodynamical consistency. Having computed the pressure of the two phases, we then obtain the phase diagram, in $\mu_B-T$ plane, by searching for zeros of the equation,
$$\label{eq5}
\Delta P(T,\mu_B) = P_{QGP} (T,\mu_B)-P^{excl}_H (T,\mu_B)=0$$
The locus of the zeros is identified as the phase boundary in $\mu_B-T$ plane along which QGP and the hadronic phase can co-exist. The numerical zero-search can be employed in two ways. We can fix a temperature $T_0$ and vary $\mu_B$ over the specified range in steps of $\delta\mu_B$ and calculate $\Delta P(T_0,\mu_B)$ and $\Delta P(T_0,\mu_B+\delta\mu_B)$. If they are of opposite sign then there must a value of $\mu_B=\mu_B^0$, such that $\Delta P(T_0,\mu_B^0)=0$, and which can be calculated using straight line interpolation. Then ($T_0,\mu_B^0$) denotes a first order phase transition point in the $\mu_B-T$ plane, provided $\Delta P(T_0,\mu_B+\delta\mu_B) > \Delta P(T_0,\mu_B)$. We can then go to another temperature $T_0+\delta T$ and get the corresponding value of $\mu_B^0$. In this way we can trace the first order phase transition line in $\mu_B-T$ plane. Let us name this method as $\mu_B$-scan. This is identical to the method employed by Singh et. al. to construct the QCD phase digram. On the contrary one can in principle fix a baryon chemical potential $\mu_B^0$ and scan the temperature axis identically to obtain the corresponding transition temperatures. We can call this approach as T-scan.
We have employed both methods namely T-scan and $\mu_B$-scan to draw the phase diagram in the $\mu_B-T$ plane. For both cases we have fixed the step length of scan as $\delta T = \delta\mu_B = 5$ MeV. The resulting phase boundaries are shown in Fig. \[fig1\].
{height="6.5cm" width="6.5cm"} \[fig:F2a\] {height="6.5cm" width="6.5cm"} \[fig:F2b\] {height="6.5cm" width="6.5cm"} \[fig:F2c\] {height="6.5cm" width="6.5cm"} \[fig:F2d\]
The first order phase transition line obtained from $\mu_B$ scan ends at some non-zero value of $\mu_B$, the so called critical point of the de-confinement phase transition. However T-scan with same step length yields a phase boundary that reaches up to $\mu_B=0$. Hence the end point of the first order phase transition line as obtained through $\mu_B$ scan can no longer be interpreted as the true critical point of de-confinement phase transition. Rather it is a pseudo-critical point, which appears due to insufficient numerical resolution of the method employed for the construction of the phase diagram. The absence of 1st order phase transition points (zeros of $\Delta P (T, \mu_B)$) beyond some minimum of $\mu_B$ can be understood by looking at the structure of the phase diagram obtained through T-scan. As the line approaches towards the T axis it becomes more and more flat in T. Hence the zeros of $\Delta P$ becomes more and more closer in T and lie with in the search length for a step length of $\delta T=5$ Mev. If one gradually decreases the step length, more and more zeros start to appear in the low $\mu_B$ region and corresponding curve eventually moves towards the $\mu_B = 0$ axis. This becomes evident from Fig. \[fig2\], where we have plotted the phase transition line for gradually decreasing values of $\delta T$. Thus the appearance of the end point of the first order phase transition line in $\mu_B-T$ plane, is not physical. This is rather a pseudo-critical point, coming into the picture due to lack of numerical detectability.
Indeed, one can argue that for bag model type equation of state, it is possible to tune the bag pressure such that Eq.\[eq5\] is satisfed for zero baryons density. In Fig. \[fig3\], for $B^{1/4}$=(50-300) MeV, we have plotted the pressure difference $\Delta P(T) = P_{QGP}(T,\mu_B=0) - P_{H}(T,\mu_B=0)$ as a fuction of temeprature. One observes that for $50 MeV \leq B^{1/4} \leq 270$ MeV, a co-existence phase exists. At low temperatures, QGP pressure is smaller than the hadron pressure and $\Delta P(T)$ is negative. At high temperature QGP pressure dominates over the hadronic pressure resulting a positive $\Delta P(T)$. However, the transition temperaure shifts to higher $T$ with increasing $B$ and there by indicating a Bag pressure dependent first order de-confinement phase transition at zero baryon chemical potential. At sufficiently high (low) values of B $(B^{1/4}= 300 (50) MeV)$, the QGP pressure is always lower (higher) than the hadronic pressure (at $\mu_B = 0$), and one can not get any first order phase transition. But over a reasonable range of Bag constants, one can always get a solution of $\Delta P(T,0) = 0$ and hence a first order hadron to QGP phase transition at $\mu_B = 0$. At very high values of temperature $\Delta P(T)$ is seen to decrease with T. This can be attributed to the larger degrees of freedom associated with the hadronic phase compared to the QGP because of the exponential growth of the hadrons and resonances at very high T, resulting in a higher pressure in HG than QGP.This indicates a reversal phase transition from QGP to hadron gas at a still higher temperature [@Madhukar]. But once the system goes over to the QGP phase, due to rise in temperature, it should continue to stay in that phase owing to the asymptotic freedom of QCD. Moreover, it is expected that the hadronic interactions are become significant when hadrons are closely packed in a hot and dense hadron gas. This anomalous behavior arises owing to the treatment of the hadronic phase as an ideal gas of non-interacting point-like hadrons. As a result of this assumption, the thermal production of an arbitrarily large number of hadrons in a given volume at very high T (or $mu_B$) is possible and eventually leads to infinitely large energy densities and pressure. In fact a simple remedy to this problem is the inclusion of finite, proper volume for each hadron, which leads to a hard-core repulsion among themselves at very high temperature and/or density and thereby limiting the number of hadrons in the system so that its volume is completely filled with particles. This is the so called excluded excluded volume correction, where the repulsive force is being incorporated by assigning a geometrical hard-core volume to each hadron. This finally leads to the reduction of the effective pressure and energy density of the hadron gas particularly at high temperature and/or density. We do have incorporated the excluded volume effect in our calculation following the model by Cleymans and Suhonen [@Cleymans2]. In this model the the volume correction term is proportional to the net baryon density and thus becomes in-effective at $mu_B = 0$. Hence the hadron gas behaves like an ideal gas and shows the indications of the reverse phase transition at very high temperature.
That our observation of the absence of any true critical point in HRG+Bag model based EOS can be justified further through following arguments. The critical point is the end point of the first order phase transition line and at the critical point transition is believed to be second order. A first order phase transition is always associated with non-zero finite latent heat, which in fact measures the discontinuity in entropy at the transition point. For a second order transition, entropy changes smoothly along the transition point and hence latent heat is zero. In Fig. \[fig4\], we have shown the temperature variation of Helmholtz free energy density ($f=\varepsilon-Ts$) at $\mu_B=0$. Both $f_{had}$ and $f_{QGP}$ decreases as temperature increases. Since the system always follows the path of minimum free energy, hence the system will undergo a transition from hadron phase to partonic phase at $T=129.6$ MeV. Since $s=-\frac{df}{dT}$, hence a kink in f(T) at $T=129.6 MeV$ would indicate a discontinuity in entropy density thereby ensuring the zero chemical potential, temperature driven transition, to be of first order.
To conclude, we would like to clarify that we are not doubting the existence of the critical point in the QCD phase diagram. The ab-initio Lattice QCD calculations have indeed proved that such a point does exist. Though its precise location in the $T-\mu_B$ is a matter of ongoing debate as predictions from different groups vary wildly. Our only motivation is to prove that the explicit construction of the first order phase separation boundary through comparison of a HRG EOS with Bag model EOS, can not give any estimation of the critical end point for de-confinement phase transition.
[99]{}
M. Cheng [*et al.*]{}, Phys. Rev. D [**77**]{}, 014511 (2008) \[arXiv:0710.0354 \[hep-lat\]\]. R. A. Soltz \[HotQCD Collaboration\], Nucl. Phys. A [**830**]{}, 725C (2009) \[arXiv:0908.1951 \[hep-lat\]\].
Y. Aoki, S. Borsanyi, S. Durr, Z. Fodor, S. D. Katz, S. Krieg and K. K. Szabo, JHEP [**0906**]{}, 088 (2009) \[arXiv:0903.4155 \[hep-lat\]\]. Y. Aoki, Z. Fodor, S. D. Katz and K. K. Szabo, Phys. Lett. B [**643**]{}, 46 (2006) \[arXiv:hep-lat/0609068\]. M. Asakawa and K. Yazaki, Nucl. Phys. A [**504**]{}, 668 (1989). J. Berges and K. Rajagopal, Nucl. Phys. B [**538**]{}, 215 (1999) \[arXiv:hep-ph/9804233\]. A. Barducci, R. Casalbuoni, S. De Curtis, R. Gatto and G. Pettini, Phys. Rev. D [**41**]{}, 1610 (1990). R. V. Gavai and S. Gupta, Phys. Rev. D [**71**]{}, 114014 (2005) \[arXiv:hep-lat/0412035\]. R. V. Gavai and S. Gupta, Phys. Rev. D [**68**]{}, 034506 (2003) \[arXiv:hep-lat/0303013\].
I. M. Barbour, S. E. Morrison, E. G. Klepfish, J. B. Kogut and M. P. Lombardo, Phys. Rev. D [**56**]{}, 7063 (1997) \[arXiv:hep-lat/9705038\]. C. R. Allton [*et al.*]{}, Phys. Rev. D [**66**]{}, 074507 (2002) \[arXiv:hep-lat/0204010\]. Z. Fodor and S. D. Katz, JHEP [**0203**]{}, 014 (2002) \[arXiv:hep-lat/0106002\]. P. de Forcrand and O. Philipsen, Nucl. Phys. B [**642**]{}, 290 (2002) \[arXiv:hep-lat/0205016\]. P. de Forcrand and O. Philipsen, Nucl. Phys. B [**673**]{}, 170 (2003) \[arXiv:hep-lat/0307020\]. P. de Forcrand and O. Philipsen, JHEP [**0701**]{}, 077 (2007) \[arXiv:hep-lat/0607017\]. M. G. Alford, K. Rajagopal and F. Wilczek, Phys. Lett. B [**422**]{}, 247 (1998) \[arXiv:hep-ph/9711395\]. R. Rapp, T. Schafer, E. V. Shuryak and M. Velkovsky, Phys. Rev. Lett. [**81**]{}, 53 (1998) \[arXiv:hep-ph/9711396\].
P. N. Meisinger and M. C. Ogilvie, Phys. Lett. B [**379**]{}, 163 (1996) \[arXiv:hep-lat/9512011\]. S. K. Ghosh, T. K. Mukherjee, M. G. Mustafa and R. Ray, Phys. Rev. D [**73**]{}, 114007 (2006) \[arXiv:hep-ph/0603050\]. J. I. Kapusta, Phys. Rev. C [**81**]{}, 055201 (2010) \[arXiv:1001.3650 \[nucl-th\]\]. J. Kapusta, J. Phys. Conf. Ser. [**230**]{}, 012016 (2010) \[arXiv:1005.0860 \[nucl-th\]\]. E. S. Bowman and J. I. Kapusta, Phys. Rev. C [**79**]{}, 015202 (2009) \[arXiv:0810.0042 \[nucl-th\]\]. P. K. Srivastava, S. K. Tiwari and C. P. Singh, Phys. Rev. D [**82**]{}, 014023 (2010) \[arXiv:1002.4780 \[hep-ph\]\]. C. P. Singh, P. K. Srivastava and S. K. Tiwari, Phys. Rev. D [**80**]{}, 114508 (2009) \[arXiv:0908.0194 \[hep-ph\]\].
G. Boyd, J. Engels, F. Karsch, E. Laermann, C. Legeland, M. Lutgemeier and B. Petersson, Nucl. Phys. B [**469**]{}, 419 (1996) \[arXiv:hep-lat/9602007\]. R. D. Pisarski, Prog. Theor. Phys. Suppl. [**168**]{}, 276 (2007) \[arXiv:hep-ph/0612191\]. A. Bazavov [*et al.*]{}, Phys. Rev. D [**80**]{}, 014504 (2009) \[arXiv:0903.4379 \[hep-lat\]\]. F. Becattini, J. Manninen and M. Gazdzicki, Phys. Rev. C [**73**]{}, 044905 (2006) \[arXiv:hep-ph/0511092\].
F. Karsch, and H. Satz, Phys. Rev. D [**21**]{}, 1168 (1980)
V. V. Dixit, F. Karsch and H. Satz, Phys. Lett. B [**101**]{}, 412 (1981)
J. Cleymans and K. Redlich, H. Satz and E. Suhonen, Z. Phys. C [**33**]{}, 151 (1986)
V. Magas, and H. Satz, Eur. Phys. J.C [**32**]{}, 115 (2003)
P. Castorina,K. Redlich and H. Satz, Eur. Phys. J.C [**59**]{}, 67 (2009)
J. Cleymans and E. Suhonen, Z. Phys. C [**37**]{}, 51 (1987)
M. Mishra, and C. P. Singh, Phys. Rev. C [**76**]{}, 024908 (2007) \[arXiv:hep-ph/0708.0488\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Using photon conversions to $e^+e^-$ pairs, the energy spectrum of inclusive photons from $\psi(2S)$ radiative decays is measured by BESII at the Beijing Electron-Positron Collider. The $\chi_{cJ}(1P)$ states (J=0,1,2) are clearly observed with energy resolution between 2.3 to 3.8 $\hbox{MeV}$, and their masses and the spin-averaged $\chi_{cJ}$ mass are determined to be $M_{\chi_{c0}}=3414.21\pm 0.39\pm 0.27$, $M_{\chi_{c1}}=3510.30\pm 0.14\pm 0.16$, $M_{\chi_{c2}}=3555.70\pm
0.59\pm 0.39$ and $M(^3P_{cog})=3524.85\pm 0.32\pm 0.30$ $\hbox{MeV}/c^2$.
author:
- |
M. Ablikim$^{1}$, J. Z. Bai$^{1}$, Y. Ban$^{11}$, J. G. Bian$^{1}$, X. Cai$^{1}$, J. F. Chang$^{1}$, H. F. Chen$^{17}$, H. S. Chen$^{1}$, H. X. Chen$^{1}$, J. C. Chen$^{1}$, Jin Chen$^{1}$, Jun Chen$^{7}$, M. L. Chen$^{1}$, Y. B. Chen$^{1}$, S. P. Chi$^{2}$, Y. P. Chu$^{1}$, X. Z. Cui$^{1}$, H. L. Dai$^{1}$, Y. S. Dai$^{19}$, Z. Y. Deng$^{1}$, L. Y. Dong$^{1}$$^a$, Q. F. Dong$^{15}$, S. X. Du$^{1}$, Z. Z. Du$^{1}$, J. Fang$^{1}$, S. S. Fang$^{2}$, C. D. Fu$^{1}$, H. Y. Fu$^{1}$, C. S. Gao$^{1}$, Y. N. Gao$^{15}$, M. Y. Gong$^{1}$, W. X. Gong$^{1}$, S. D. Gu$^{1}$, Y. N. Guo$^{1}$, Y. Q. Guo$^{1}$, Z. J. Guo$^{16}$, F. A. Harris$^{16}$, K. L. He$^{1}$, M. He$^{12}$, X. He$^{1}$, Y. K. Heng$^{1}$, H. M. Hu$^{1}$, T. Hu$^{1}$, G. S. Huang$^{1}$$^b$, X. P. Huang$^{1}$, X. T. Huang$^{12}$, X. B. Ji$^{1}$, C. H. Jiang$^{1}$, X. S. Jiang$^{1}$, D. P. Jin$^{1}$, S. Jin$^{1}$, Y. Jin$^{1}$, Yi Jin$^{1}$, Y. F. Lai$^{1}$, F. Li$^{1}$, G. Li$^{2}$, H. H. Li$^{1}$, J. Li$^{1}$, J. C. Li$^{1}$, Q. J. Li$^{1}$, R. Y. Li$^{1}$, S. M. Li$^{1}$, W. D. Li$^{1}$, W. G. Li$^{1}$, X. L. Li$^{8}$, X. Q. Li$^{10}$, Y. L. Li$^{4}$, Y. F. Liang$^{14}$, H. B. Liao$^{6}$, C. X. Liu$^{1}$, F. Liu$^{6}$, Fang Liu$^{17}$, H. H. Liu$^{1}$, H. M. Liu$^{1}$, J. Liu$^{11}$, J. B. Liu$^{1}$, J. P. Liu$^{18}$, R. G. Liu$^{1}$, Z. A. Liu$^{1}$, Z. X. Liu$^{1}$, F. Lu$^{1}$, G. R. Lu$^{5}$, H. J. Lu$^{17}$, J. G. Lu$^{1}$, C. L. Luo$^{9}$, L. X. Luo$^{4}$, X. L. Luo$^{1}$, F. C. Ma$^{8}$, H. L. Ma$^{1}$, J. M. Ma$^{1}$, L. L. Ma$^{1}$, Q. M. Ma$^{1}$, X. B. Ma$^{5}$, X. Y. Ma$^{1}$, Z. P. Mao$^{1}$, X. H. Mo$^{1}$, J. Nie$^{1}$, Z. D. Nie$^{1}$, S. L. Olsen$^{16}$, H. P. Peng$^{17}$, N. D. Qi$^{1}$, C. D. Qian$^{13}$, H. Qin$^{9}$, J. F. Qiu$^{1}$, Z. Y. Ren$^{1}$, G. Rong$^{1}$, L. Y. Shan$^{1}$, L. Shang$^{1}$, D. L. Shen$^{1}$, X. Y. Shen$^{1}$, H. Y. Sheng$^{1}$, F. Shi$^{1}$, X. Shi$^{11}$$^c$, H. S. Sun$^{1}$, J. F. Sun$^{1}$, S. S. Sun$^{1}$, Y. Z. Sun$^{1}$, Z. J. Sun$^{1}$, X. Tang$^{1}$, N. Tao$^{17}$, Y. R. Tian$^{15}$, G. L. Tong$^{1}$, G. S. Varner$^{16}$, D. Y. Wang$^{1}$, J. Z. Wang$^{1}$, K. Wang$^{17}$, L. Wang$^{1}$, L. S. Wang$^{1}$, M. Wang$^{1}$, P. Wang$^{1}$, P. L. Wang$^{1}$, S. Z. Wang$^{1}$, W. F. Wang$^{1}$$^d$, Y. F. Wang$^{1}$, Z. Wang$^{1}$, Z. Y. Wang$^{1}$, Zhe Wang$^{1}$, Zheng Wang$^{2}$, C. L. Wei$^{1}$, D. H. Wei$^{1}$, N. Wu$^{1}$, Y. M. Wu$^{1}$, X. M. Xia$^{1}$, X. X. Xie$^{1}$, B. Xin$^{8}$$^b$, G. F. Xu$^{1}$, H. Xu$^{1}$, S. T. Xue$^{1}$, M. L. Yan$^{17}$, F. Yang$^{10}$, H. X. Yang$^{1}$, J. Yang$^{17}$, Y. X. Yang$^{3}$, M. Ye$^{1}$, M. H. Ye$^{2}$, Y. X. Ye$^{17}$, L. H. Yi$^{7}$, Z. Y. Yi$^{1}$, C. S. Yu$^{1}$, G. W. Yu$^{1}$, C. Z. Yuan$^{1}$, J. M. Yuan$^{1}$, Y. Yuan$^{1}$, S. L. Zang$^{1}$, Y. Zeng$^{7}$, Yu Zeng$^{1}$, B. X. Zhang$^{1}$, B. Y. Zhang$^{1}$, C. C. Zhang$^{1}$, D. H. Zhang$^{1}$, H. Y. Zhang$^{1}$, J. Zhang$^{1}$, J. W. Zhang$^{1}$, J. Y. Zhang$^{1}$, Q. J. Zhang$^{1}$, S. Q. Zhang$^{1}$, X. M. Zhang$^{1}$, X. Y. Zhang$^{12}$, Y. Y. Zhang$^{1}$, Yiyun Zhang$^{14}$, Z. P. Zhang$^{17}$, Z. Q. Zhang$^{5}$, D. X. Zhao$^{1}$, J. B. Zhao$^{1}$, J. W. Zhao$^{1}$, M. G. Zhao$^{10}$, P. P. Zhao$^{1}$, W. R. Zhao$^{1}$, X. J. Zhao$^{1}$, Y. B. Zhao$^{1}$, Z. G. Zhao$^{1}$$^e$, H. Q. Zheng$^{11}$, J. P. Zheng$^{1}$, L. S. Zheng$^{1}$, Z. P. Zheng$^{1}$, X. C. Zhong$^{1}$, B. Q. Zhou$^{1}$, G. M. Zhou$^{1}$, L. Zhou$^{1}$, N. F. Zhou$^{1}$, K. J. Zhu$^{1}$, Q. M. Zhu$^{1}$, Y. C. Zhu$^{1}$, Y. S. Zhu$^{1}$, Yingchun Zhu$^{1}$$^f$, Z. A. Zhu$^{1}$, B. A. Zhuang$^{1}$, X. A. Zhuang$^{1}$, B. S. Zou$^{1}$.\
(BES Collaboration)\
title: ' Precise measurement of spin-averaged $\chi_{cJ}(1P)$ mass using photon conversions in $\psi(2S) \rightarrow \gamma \chi_{cJ}$ '
---
Introduction
============
Precise measurements of the spectrum and the decay properties of charmonia are essential to test Potential QCD models and QCD based approaches [@eichten]. There is renewed interest since the discovery of the X(3872) [@X3872] and the observations of the expected $\eta_C(2S)$ and $h_C$ $(^1P_1)$ states [@etac; @and; @hc], and there has been recent progress, both theoretically and experimentally [@QCD; @review]. There are more accurate determinations of the charmonium mass spectrum and radiative transition rates using both a relativistic quark model with relativistic corrections of order $v^2/c^2$ [@ebert] and a potential model with a semirelativistic approach [@Radford]. The $\psi(2S)$ mass and width have been redetermined with an updated radiative correction [@artamonov], and newly measured with better precision [@Aulchenko]. In addition to previous measurements of $\chi_{cJ}$ states [@pdg04], two $\chi_{c0}$ measurements by E835 [@bagnosco] and new $\chi_{cJ}$ (J=0,1,2) measurements by CLEO [@CLEO; @chicJ] have been recently published. Improved precision on $\chi_{cJ}$ masses is important for the determination of the singlet-triplet splitting, $M(^1P_1)-M(^3P_{cog})$, which is predicted by lattice QCD and nonrelativistic QCD [@Godfrey; @and; @Rosner]. Here $M(^3P_{cog})$ is the spin-averaged $^3P_J$ mass for the $\chi_{cJ}$ states ($J=0,1,2$).
In this paper, results on the $\chi_{cJ}$ masses (J=0,1,2) and widths (J=0,1) from a measurement of the energy spectrum of inclusive photons in $\psi(2S)$ radiative decays, using photon conversions to improve the energy resolution, are presented. The measurement uses $14\times
10^6$ $\psi(2S)$ events collected with the upgraded Beijing Spectrometer (BESII) at the BEPC Collider.
BES Detector and Monte Carlo simulation
=======================================
The BESII detector is described elsewhere [@bes2]. A 12-layer vertex chamber (VC) surrounding the beam pipe provides hit information in trigger criteria for charged tracks. Charged particle momenta are determined with a resolution of $\sigma _p/p = 1.78\%
\sqrt{1+p^2}$ ($p$ in $\hbox{\rm GeV}/c$) in a 40-layer cylindrical drift chamber (MDC). Particle identification is accomplished by measurements of ionization ($dE/dx$) in the MDC and time-of-flight (TOF) in a barrel-like array of 48 scintillation counters. The $dE/dx$ resolution is $\sigma_{dE/dx}=
8\%$; the TOF resolution is $\sigma_{TOF}$=200 ps for hadrons. A 12-radiation-length barrel shower counter (BSC) measures energies of photons with a resolution of $\sigma_E/E=21\%/\sqrt{E}$ ($E$ in GeV). A solenoidal coil supplies a 0.4 Tesla magnetic field over the tracking volume.
A Geant3 based Monte Carlo (MC) SIMBES [@liuhm], which simulates the detector response including interactions of secondary particles in the detector material, is used to determine the energy resolution and detection efficiency of photons reconstructed from their converted $e^+e^-$ pairs, as well as to optimize selection criteria and estimate backgrounds. Under the assumption of a pure E1 transition for the $\psi(2S)\rightarrow \gamma\chi_{cJ}$ decays, the polar angle ($\theta$) distributions of the photons are given by $1+k\cos^2\theta$ with $k = 1,-\frac{1}{3},\frac{1}{13}$ for $J=0,1,2$, respectively [@karl].
Good energy resolution for low energy photons is essential for precise measurements of $\chi_{cJ}$ masses and widths from fitting the photon spectrum of $\psi(2S)$ radiative decays. Momentum resolution of about 1.6 to 4.1 $\hbox{MeV}/c$ can be obtained for low momentum electrons from 60 to 250 $\hbox{MeV}/c$. Photons from $\psi(2S)\rightarrow \gamma \chi_{cJ}$ decays have energies of about 261, 171, and 128 $\hbox{MeV}$ for the $\chi_{cJ}$ final states (J=0,1,2), and the electrons produced in photon conversions occur in this low momentum region.
Photon reconstruction and selection
===================================
We choose two oppositely charged tracks with each track having a good helix fit and a polar angle with $|\cos\theta|< 0.8$. The intersection of the electron and positron trajectories in the xy-plane (the beam line is the $z$ axis) is determined, and this point is taken as the photon conversion point (CP). The photon conversion length $R_{xy}$ is defined as the distance from the beam line to the CP in the xy-plane. Fig. \[rxy\] shows the $R_{xy}$ distribution for photon conversions to $e^+e^-$ pairs in the BESII detector for hadronic events in the 58 $\times 10^6$ $J/\psi$ event sample. The two broad peaks in Fig. \[rxy\] correspond to the beampipe region, where the beampipe, the VC, and inner wall of the MDC are located. Combinatorial background from charged hadron tracks is also seen in the $R_{xy}<2$ cm region. Equivalent materials in the beampipe wall, VC, VC outer wall, and the MDC inner wall are $0.536$, $0.657$,$0.375$, and $1.107$ in units of $0.01X_0$ [@hongt], respectively, where $X_0$ is a radiation length. The electron and positron directions are calculated at the photon conversion point, and their momenta are corrected to that point.
Good photons are selected. The photon conversion length must lie within the beampipe region, $2 < R_{xy}< 22$ cm, and the invariant mass of an $e^+e^-$ pair is required to satisfy $M_{e^+e^-} < 20$ $\hbox{MeV}/c^2$. Combinatorial background from charged hadron tracks is further removed by requiring $\cos\theta_{defl} > 0.9$ , where $\theta_{defl}$ is the deflection angle between the photon momentum and photon track (a vector from the beam to the CP). To suppress background from beam-gas and beam-pipe interactions, the total energy in the event must satisfy $E_{tot} > E_{beam}/2$ and momentum asymmetry must satisfy $dp_{asym} < 0.9$. Here $dp_{asym}$ is defined as a ratio of the vector sum to the scalar sum of the momenta of all charged and neutral tracks in the event. The observed photon energy spectrum from the $\psi(2S)$ data after the selection of good photons is shown in Fig. \[plot fit data\]. The spectrum shows the $\chi_{cJ}$ states plus a large background. The sharp drop at low energy is mainly caused by low photon detection efficiency.
$dE/dx$ correction and photon energy scale
==========================================
Energy loss dE/dx by ionization for electrons traversing a small thickness of material with energy above a few tens of $\hbox{MeV}$ can be described by the Bethe-Bloch equation [@pdg04]. The dE/dx correction for charged particles, produced near the beamline and traversing the whole beampipe region, should take into account the full thickness of material in the region. However, $e^+e^-$ pairs from photon conversions are mostly produced in the region where the VC outer wall and the MDC inner wall are located. Thus the effective thickness of material between the location, where a pair is produced, and first layer of the MDC wires must be estimated for each electron pair. The procedure to make $dE/dx$ corrections for electrons has two steps : (1) A preliminary $dE/dx$ correction using half the full thickness of all the materials in the beampipe region is made. Good photons are reconstructed, and their conversion lengths $R_{xy}$ are calculated. (2) The final $dE/dx$ corrections are estimated based on these $R_{xy}$.
The energy scale of photons reconstructed from $e^+e^-$ pairs is studied using simulated MC events and data. $63\times 10^6$ $\pi^0$ signal events are generated using MC technique, with same momentum and polar angular distributions as that found from $\pi^0$ data sample. A sample of $\pi^0$ mesons decaying to two photons with both photons converting to $e^+e^-$ pairs is selected from $58\times 10^6$ $J/\psi$ events. To suppress hadron contamination, electron identification is required and good photons are selected. Background is further suppressed with additional requirements on the photon energy, $E_{\gamma} \leq 1GeV$, and the opening angle between the two photons, $0.75<|\cos\theta_{\gamma\gamma}|<0.97$. The invariant mass distribution of two photons for both MC and data, after the specific $dE/dx$ correction for electrons described above, is fitted with the improved Crystal Ball (ICB) function (defined in section \[Physics problem\]) plus a first order polynomial background. The results are shown in Fig. \[plot fit pi0 mass\]. The resulting $\pi^0$ masses $(134.47\pm 0.42)$ $\hbox{MeV}/c^2$ in data and $(134.86\pm 0.20)$ $\hbox{MeV}/c^2$ in MC are consistent with the PDG value of $134.98$ $\hbox{MeV}/c^2$ [@pdg04]. The corresponding mass resolutions $(5.70\pm 0.58)$ $\hbox{MeV}/c^2$ in data and $(5.55\pm 0.21)$ $\hbox{MeV}/c^2$ in MC agree within errors. The $\chi^2$/D.F. (degree of freedom) for the fits are 126/103 in data and 117/140 in MC.
Physics problem and detector resolution {#Physics problem}
=======================================
The energy of a photon from $\psi(2S)\rightarrow
\gamma\chi_{cJ}$ decay is given by $$\begin{aligned}
E_\gamma &=& (M^2_{\psi(2S)}-M^2_{\chi_{cJ}}) / 2M_{\psi(2S)},
\label{energyg}\end{aligned}$$ where $M_{\psi(2S)}$ and $M_{\chi_{cJ}}$ are the masses of the $\psi(2S)$ and $\chi_{cJ}$, respectively. The $\psi(2S)$ and $\chi_{cJ}$ widths must be taken into account; $M_{\psi(2S)}$ and $M_{\chi_{cJ}}$ are described by Breit-Wigner functions (2D problem). By taking $x={M_{\psi(2S)}}$, the probability density function (pdf) for the photon energy $E_\gamma$ can be written as [@mfisz] $$\begin{aligned}
f_{pdf}(E_\gamma) = \int BW(x) BW(M_{\chi_{cJ}})
\frac{x}{M_{\chi_{cJ}}} dx,
\label{egpdf2}\end{aligned}$$ where $M_{\chi_{cJ}}$ depends on $E_\gamma$ by Eq. (\[energyg\]).
As a result of traversing material in the beam pipe region, the electron energy is smeared due to energy loss by ionization, and a long tail on the low side of the energy distribution is induced by bremsstrahlung radiation. Multiple scattering of electrons, especially at large angles, gives tails on both sides of the photon energy distribution of photon conversions. The photon energy resolution from photon conversions can be nicely modeled by our Geant3 MC simulation, and well fitted by the ICB function. The original Crystal Ball (CB) function has a Gaussian in its central and upper energy region but long tail at lower energy region [@mnfit]. The improved CB function is defined as same as the CB function but has an additional tail at its upper side. The photon energy distributions from large MC samples of $\psi(2S) \rightarrow
\gamma\chi_{cJ}$ decays $(J=0,1,2)$, with zero widths for both the $\psi(2S)$ and $\chi_{cJ}$ states, are fitted to ICB functions and shown in Fig. \[photon resolution\]. The $\chi^2$/D.F. from the fits are 103.8/93, 37.7/53 and 47.6/43 corresponding to $\chi_{c0}$, $\chi_{c1}$, $\chi_{c2}$ states. Five parameters in the ICB function, the photon energy resolution and four empirical parameters to describe the tails on the lower and upper sides are determined from the fits and used as input parameters in the detector resolution function for each decay mode. Photon energy resolutions for the $\psi(2S) \rightarrow
\gamma\chi_{cJ}$ decays (J=0,1,2) are found to be $3.78\pm 0.04$, $2.58\pm 0.05$, and $2.26\pm 0.11$ $\hbox{MeV}$, respectively.
The energy dependencies of the photon detection efficiency and resolution are determined using MC simulation in the energy range between 100 and 400 (300, 220) $\hbox{MeV}$ for $\psi(2S) \rightarrow
\gamma\chi_{c0}~(\gamma\chi_{c1},\gamma\chi_{c2})$ decays. The efficiency includes the effects of detector geometry, MDC tracking, photon reconstruction, and the spin-dependent $\cos\theta$ distribution.
Fitting three photon spectrum
=============================
Background contamination from $\chi_{c0} \rightarrow \gamma J/\psi$ decay is negligible due to its small branching fraction. To exclude photons from the $\chi_{c1,c2}
\rightarrow \gamma J/\psi$ decays, the photon energy range in the fits is chosen to be within $90~\hbox{MeV} \le E_\gamma \le 350~\hbox{MeV}$. An enhancement from the decay $\psi(2S) \rightarrow \eta J/\psi$ with $\eta\rightarrow \gamma\gamma$ in a region of $180$ to $400~\hbox{MeV}$ is estimated by MC method, and subtracted from the data, according to the measured total number of $\psi(2S)$ events and known branching fractions. The smooth background under the signal photon lines can be described by a threshold function from the $Mn\_Fit$ package [@mnfit]. $$\begin{aligned}
bg_{threshold}(x) &=& B\cdot (x-x_0)^{\alpha}\cdot e^{c_1(x-x_0)+c_2(x-x_0)^2},
\label{threshold fun}\end{aligned}$$ where B, $x_0$, and $\alpha$ are normalization factor, threshold, and power; $c_1$ and $c_2$ are coefficients linear and quadratic in x. The threshold function has been used to fit backgound with a threshold at the lower or upper side of an observed distribution by some experiments [@Babar].
Considering the physical photon energy $x\equiv E_\gamma$ and its error $y \equiv \triangle E_\gamma$ due to detector resolution, the measured photon energy is $u=x + y$ and its pdf function can be written as : $$\begin{aligned}
h_{pdf}(u) = \int c_{E_1}(u-y)\cdot c_{eff}(u-y)\cdot f_{pdf}(u-y)\cdot g_{res}(y)dy
= \int c_{E_1}(x)\cdot c_{eff}(x)\cdot f_{pdf}(x)\cdot g_{res}(u-x)dx,
\label{pdf function of photon energy}\end{aligned}$$ where $c_{E_1}(x) = x^3/E^3_{\gamma,\chi_{cJ}}$, $c_{eff}(x)=\epsilon_{\chi_{cJ}}(x)/\epsilon(E_{\gamma,\chi_{cJ}})$, $f_{pdf}(x)$ is defined with Eq. (\[egpdf2\]), and $g_{res}(y)$ is the ICB resolution function. With the assumption that the $E1$ electric dipole transition for $\psi(2S) \rightarrow \gamma \chi_{cJ}$ decays (J=0,1,2) dominates, an $E^3_\gamma$ energy dependence is included in the folded signal shape. The detection efficiency $\epsilon_{\chi_{cJ}}(x)$ and energy resolution as a function of photon energy are included in the fitting. Normalization factors $E_{\gamma,\chi_{cJ}}$ and $\epsilon(E_{\gamma,\chi_{cJ}})$ are photon energy corresponding to fitted $\chi_{cJ}$ mass and efficiency at the photon energy, respectively. Notice that parameters, masses $M_{\chi_{cJ}}$ and widths $\Gamma_{\chi_{cJ}}$ (J=0,1,2), are implicitly contained in the $f_{pdf}(x)$ function, and detector resolution and tail parameters are in the $g_{res}(y)$ function. The likelihood function, $Lk(u;M_{\chi_{cJ}},\Gamma_{\chi_{cJ}})$, is constructed with three $\chi_{cJ}$ signals plus threshold background: $$\begin{aligned}
Lk(u;M_{\chi_{cJ}},\Gamma_{\chi_{cJ}}) = bg_{threshold}(u) +
\sum_{J=0}^{2} A_J\cdot h_{pdf,\chi_{cJ}}(u;M_{\chi_{cJ}},\Gamma_{\chi_{cJ}}).
\label{pdf function of photon energy}\end{aligned}$$ Here $A_J$ is the observed area in each $\chi_{cJ}$ signal.
An input-output test is performed to verify the accuracy of the fitting algorithm for the 2D problem using MC events. The energy dependencies of the photon detection efficiency and resolution are included in the fitting procedure. A sample of MC events for the $\psi(2S) \rightarrow \gamma\chi_{cJ}$ decays with non-zero width for both the $\psi(2S)$ and $\chi_{cJ}$ are produced. The photon energy distribution is fitted with the 2D pdf function convoluted with the ICB resolution. The resulting masses and widths of the $\chi_{cJ}$ states in this test are consistent with the MC input parameters.
A combined fit of the three photon spectra corresponding to the $\psi(2S)\rightarrow \gamma\chi_{c0}, \gamma\chi_{c1},
\gamma\chi_{c2}$ decays is performed to three 2D pdf functions (see Eq. (\[egpdf2\])), each convoluted with its ICB resolution function, plus threshold background. The $\chi_{c2}$ width is fixed in the fit due to the limited statistics. The $\chi^2$ and D.F. (degrees of freedom) from the fit are 521.0 and (520-13), where total number of free paramaters is 13. The effect of the beam energy spread [@beam; @energy; @spread] in the measurement is also included, but is found to be negligible due to the narrow width of the $\psi(2S)$ state. A study shows that the bin size ($0.5$ or $0.2~\hbox{MeV}$) in the binned fits slightly affects the fitted masses and widths of the $\chi_{cJ}$ states. The difference in the results due to different bin sizes are added to the systematic error. The results of the binned fit (0.5 $\hbox{MeV}/$bin) are shown in Fig. \[plot fit data\].
Systematic errors
=================
Samples of QED radiative two photon events with one photon converting to an $e^+e^-$ pair are selected for both data and MC simulation. The two photons are required to be emitted back-to-back. The fitted photon energy in data is different from the expected MC value by $-1.2\sigma$ ($\sigma$ = 0.86 $\hbox{MeV}$), which has a relative error at the same level as a correction factor $s = 0.9975\pm 0.0007$ for the magnetic field [@field; @correction]. Thus a relative precision of 0.0007 is added as the systematic error in the photon energy determination.
The selection of $\pi^0\rightarrow\gamma\gamma$ events with both photons converting to $e^+e^-$ pairs from $58\times 10^6$ $J/\psi$ events yields a data sample of $503$ $\pi^0$ mesons. A MC sample of $\pi^0$ mesons is generated with the same momentum and polar angular distributions as found from the $\pi^0$ data sample. The $\pi^0$ mass resolutions determined from the data and MC are in good agreement; their difference is $0.15\pm 0.62~\hbox{MeV}/c^2$. We assume that the photon energy resolution and uncertainty in the direction of the photon momentum each contribute half in the $\pi^0$ mass resolution. Hence, the difference $\triangle\sigma_{M_{\pi^0}}$ of the $\pi^0$ mass resolutions between MC and data from the uncertainty of the photon energy resolution lies within $(-0.29,
+0.59)~\hbox{MeV}/c^2$ with a probability of $68.3\%$. We assume $\triangle\sigma_{E_\gamma}(\chi_{cJ}) /
\sigma_{E_\gamma}(\chi_{cJ},MC) = \triangle\sigma_{M_{\pi^0}} /
\sigma_{M_\pi^0}(MC)$, where $\sigma_{E_\gamma}(\chi_{cJ},MC)$ and $\triangle\sigma_{E_\gamma}(\chi_{cJ})$ are the MC photon energy resolution and the difference between MC and data for $\chi_{cJ}$ final states, and $\sigma_{M_\pi^0}(MC)$ and $\triangle\sigma_{M_{\pi^0}}$ are $\pi^0$ mass resolution in MC and the difference between MC and data. Thus $1\sigma$ confidence intervals of $\triangle\sigma_{E_\gamma}(\chi_{cJ})$ for the $\psi(2S)\rightarrow\gamma\chi_{c0},\gamma\chi_{c1}$ decays are estimated to be $(-0.20,+0.40)$ and $(-0.14,+0.27)$ $\hbox{MeV}$, which are used to estimate systematic errors in the determination of the $\chi_{c0}$ and $\chi_{c1}$ widths.
The effect of the background shape uncertainty is studied using $\psi(2S)$ data and $\psi(2S)\rightarrow anything$ MC [@chenjc]. The relative differences in background shape parameters between floated and fixed widths of the $\chi_{c0,c1}$ states are determined in fits for MC data, and fed back to correct background parameters in the fits for data. The difference between results for $\psi(2S)$ data with the background shape floated and fixed is taken as a systematic error. In addition, our MC study with non-zero width of both $\psi(2S)$ and $\chi_{cJ}$ shows that differences in the fitted masses from input values for the $\chi_{c0}$ and $\chi_{c1}$ are $0.12\pm0.06$ and $0.06\pm0.03~\hbox{MeV}/c^2$, while that for the $\chi_{c2}$ is as large as $0.31\pm0.06~\hbox{MeV}/c^2$. The differences are attributed to uncertainties in the energy loss correction for low momentum electrons. The systematic errors, including the contributions from the uncertainties of the photon detection efficiency, are summarized in the Table \[table sys error\].
source $M_{\chi_{c0}}$ $\Gamma_{\chi_{c0}}$ $M_{\chi_{c1}}$ $\Gamma_{\chi_{c1}}$ $M_{\chi_{c2}}$
-------------------------------------- ----------------- ---------------------- ----------------- ---------------------- -----------------
background shape $0.03$ $0.8$ $0.02$ $0.07$ $0.04$
correction in magnetic field $0.19$ $0.13$ $0.09$
MC simulation in $\sigma_{E_\gamma}$ $^{+0.29}_{-0.76}$ $^{+0.25}_{-0.77}$
different bin size $0.02$ $0.02$ $0.01$ $0.01$ $0.02$
photon energy correction $0.18$ $0.09$ $0.37$
efficiency uncertainty $0.04$ $0.03$ $0.01$ $0.00$ $0.04$
error of $M_{\psi(2S)}$ $0.034$ $0.034$ $0.034$
total $0.27$ $^{+0.85}_{-1.10}$ $0.16$ $^{+0.26}_{-0.77}$ $0.39$
: Summary of systematic errors in the determination of the $\chi_{cJ}$ masses and widths (in $\hbox{MeV}/c^2$).
\[table sys error\]
Results and discussion
======================
With good energy resolution for low energy photons obtained using photon conversions, the precise measurement of the masses and widths of $\chi_{cJ}~(J=0,1,2)$ states from inclusive $\psi(2S)$ radiative decays can be obtained. The masses and widths are determined to be $M_{\chi_{c0}}=3414.21\pm 0.39\pm 0.27$, $M_{\chi_{c1}}=3510.30\pm
0.14\pm 0.16$, $M_{\chi_{c2}}=3555.70\pm 0.59\pm 0.39$ $\hbox{MeV}/c^2$, $\Gamma_{\chi_{c0}}=12.6{^{+1.5}_{-1.6}}{^{+0.9}_{-1.1}}$ and $\Gamma_{\chi_{c1}}=1.39{^{+0.40}_{-0.38}}{^{+0.26}_{-0.77}}$ $\hbox{MeV}/c^2$. The mass splittings in the $\chi_{cJ}(1P)$ triplet and their ratio are found to be $\triangle M_{21} = M_{\chi_{c2}} -
M_{\chi_{c1}} = 45.40\pm0.61\pm0.42~\hbox{MeV}/c^2$, $\triangle M_{10}
= M_{\chi_{c1}} - M_{\chi_{c0}} = 96.09\pm0.41\pm0.31~\hbox{MeV}/c^2$ and $\rho (\chi_c) = \triangle M_{21} / \triangle M_{10} = 0.472\pm 0.006\pm
0.004$. For the first time, the spin-averaged $^3P_J$ mass (weighted with the factors $2J+1$) for the $\chi_{cJ}$ states is precisely measured in one experiment and determined to be $M(^3P_{cog})=3524.85\pm 0.32\pm 0.30$ $\hbox{MeV}/c^2$. The first errors in the results are statistical and the second are systematic. Correlations are taken into account in estimations of both statistical and systematic errors for the $\triangle M_{21}$, $\triangle M_{10}$, $\rho (\chi_c)$ and $M(^3P_{cog})$. Correlation coefficients between mass parameters for statistical error are obtained from the error matrix of the combined fit, and that for systematic error are all assumed to be 1.
The $\chi_{cJ}$ masses (J=0,1,2) determined here are consistent with the recent measurements by CLEO [@CLEO; @chicJ], but have smaller systematic errors. The precisions for the $\chi_{c0}$ and $\chi_{c1}$ masses are compatible with those of previous measurements by E835 [@bagnosco] and E760 [@pdg04], while that for the $\chi_{c2}$ mass is not as good as theirs due to low statistics. Note that our $\chi_{c0}$ mass is lower than that measured by the E835 via $\chi_{c0}\rightarrow \gamma J/\psi$ decay by $1.2~\hbox{MeV}$ (corresponding to $1.8\sigma$), but agrees with their later measurement via $\chi_{c0}\rightarrow \pi^0\pi^0$ decay. The width of the $\chi_{cJ}$ states (J=0,1) determined here are also consistent with their values; larger errors in our widths are caused by limited statistics for both signal photons and inclusive $\pi^0$ mesons.
Acknowledgment
==============
The BES collaboration thanks the staff of BEPC for their hard efforts and the members of IHEP computing center for their helpful assistance, and also T.P. Li for helpful discussion on 2D pdf function. This work is supported in part by the National Natural Science Foundation of China under contracts Nos. 19991480,10225524,10225525, the Chinese Academy of Sciences under contract No. KJ 95T-03, the 100 Talents Program of CAS under Contract Nos. U-11, U-24, U-25, and the Knowledge Innovation Project of CAS under Contract Nos. U-602, U-34(IHEP); by the National Natural Science Foundation of China under Contract No.10175060(USTC),and No.10225522(Tsinghua University); and by the U.S. Department of Energy under Contract No.DE-FG02-04ER41291 (U Hawaii).
[dd]{}
E. Eichten [*et al.*]{}, Phys. Rev. D [**21**]{}, 203 (1980) and D [**17**]{}, 3090 (1978); V. A. Novikov, [*et al.*]{}, Phys. Rep. [**41**]{}, 1 (1978); C. Quigg and J. L. Rosner, Phys. Rep. [**56**]{}, 167 (1979); G. Bali, K. Schilling and A. Wachter, Phys. Rev. D [**56**]{}, 2566 (1997).
Belle, S. K. Choi [*et al.*]{}, Phys. Rev. Lett. [**91**]{}, 262001 (2004). Babar, B. Aubert [*et al.*]{}, hep-ex/0406022.
Belle, S. .K Choi [*et al.*]{}, Phys. Rev. Lett. [**89**]{}, 102001 (2002); CLEO, A. Tomaradze, hep-ex/0410090; E835, C. Patrignani, hep-ex/0410085.
N. Brambilla [*et al.*]{}, Heavy Quarkonium Physics, hep-ph/0412158.
D. Ebert, R. N. Faustov and V. O. Galkin, Phys. Rev. D67, 014027 (2003) and D [**62**]{}, 034014 (2000).
S. F. Radford and W .W. Repko, hep-ph/0409290, Sep 24, 2004.
OLYA and MD-1, A. S. Artamonov [*et al.*]{}, Phys. Lett. B [**474**]{}, 427 (2000).
KEDR, V. M. Aulchenko [*et al.*]{}, Phys. Lett. B [**573**]{}, 63 (2003); BES, J.Z. Bai et al., Phys. Lett. B [**550**]{}, 24 (2002);
S. Eidelman [*et al.*]{} (Particle Data Group), Phys. Lett. B [**592**]{}, 1 (2004).
E835, S. Bagnosco [*et al.*]{}, Phys. Lett. B [**533**]{}, 237 (2002); M. Andreotti [*et al.*]{}, Phys. Rev. Lett. [**91**]{}, 091801 (2003).
CLEO, S.B. Athar [*et al.*]{}, Phys. Rev. D [**60**]{} 112002(2004). From their measured photon energies, one may estimate the $\chi_{cJ}$ masses using Eq. (\[energyg\]) in the text.
S. Godfrey and J. L. Rosner, Phys. Rev. D [**66**]{}, 014012(2002); S. Godfrey, hep-ph/0501083.
BES, J. Z. Bai. [*et al.*]{}, Nucl. Instr. Meth. [**A458**]{}, 627 (2001); Nucl. Instr. Meth. [**A344**]{}, 319 (1994).
H.M. Liu [*et al.*]{}, “The BESII Detector Simulation”, to be submitted to NIM.
G. Karl, S. Meshkov, and J.L. Rosner, Phys. Rev. D [**13**]{}, 1203 (1976).
T. Hong [*et al.*]{}, High Energy Phys. and Nucl. Phys., V [**25**]{}, 617 (2001).
M. Fisz, Probability and Mathematic Statistics, Berlin, 1958.
I.C. Brock, A Fitting and Plotting Package Using MINUIT, version 4.07, Dec. 22th, 2000.
For instance, see : Babar collab., Phys. Rev. D65, 091104(2002).
The spread in the center-of mass energy at $\psi(2S)$ energy is $1.3~\hbox{MeV}$. See also BES Collab., Phys. Lett. B [**550**]{}, 24 (2002).
A correction factor $s = 0.9975\pm 0.0007$ for magnetic field at the BESII is determined with mass measurement of the $J/\psi$ reconstructed from $\psi(2S) \rightarrow \pi^+\pi^- J/\psi$ and $J/\psi \rightarrow \mu^+\mu^-$ decays.
J.C. Chen [*et al.*]{}, Phys. Rev. D [**62**]{}, 034003 (2000).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We perform a statistical study focused on void environments. We examine galaxy density profiles around voids in the SDSS, finding a correlation between void–centric distance to the shell of maximum density and void radius when a maximum in overdensity exists. We analyze voids with and without a surrounding over-dense shell in the SDSS. We find that small voids are more frequently surrounded by over-dense shells whereas the radial galaxy density profile of large voids tends to rise smoothly towards the mean galaxy density. We analyse the fraction of voids surrounded by overdense shells finding a continuous trend with void radius. The differences between voids with and without an overdense shell around them can be understood in terms of whether the voids are, on average, in the process of collapsing or continuing their expansion, respectively, in agreement with previous theoretical expectations. We use numerical simulations coupled to semi-analytic models of galaxy formation in order to test and interpret our results. The very good agreement between the mock catalog results and the observations provides additional support to the viability of a $\Lambda$CDM model to reproduce the large scale structure of the universe as defined by the void network, in a way which has not been analysed previously.'
author:
- |
\
$^1$ Instituto de Astronomía Teórica y Experimental, UNC-CONICET, Córdoba Argentina.\
$^2$ Observatorio Astronómico de Córdoba, UNC, Argentina.\
$^3$ Departamento de Astronomía y Astrofísica, Pontificia Universidad Católica de Chile, Santiago, Chile.\
bibliography:
- 'draft1.bib'
---
large-scale structure of the Universe – methods: data analysis, observational, statistics
Introduction {#S_intro}
============
{width="\textwidth"}
Although the large scale galaxy and matter distributions are dominated by different structures such as groups, clusters, filaments, walls etc., most of the volume of the Universe is occupied by voids: large underdense regions are the prominent structures at large scales.
These large scale underdensities have been identified and analyzed in numerical simulations and in galaxy catalogues [@hoffman_shaham; @hausman; @fillmore_goldreich; @icke_1984; @bertschinger; @pellegrini; @kauffmann_voids_1991; @blumenthal_largest_1992; @slezak_1993; @el-ad_1997; @el-ad_piran_1997; @el-ad_piran_2000; @aikio_maehoenen_1998; @muller_2000; @plionis_basilakos_2002; @hoyle_vogeley_2002; @sheth_hierarchy_2004; @hoyle_voids_2004; @ceccarelli_voids_2006; @tikhonov_2006; @patiri_2006; @furlanetto_piran_2006; @Neyrinck_2008; @aragon-calvo_unfolding_2010; @pan2012; @sutter_public_2012]. Despite the fact that the agreement between void finders is weaker for small and medium scale voids to a first approximation their more prominent properties are similar. In general, similar properties of voids are found regardless of the diversity of identification methods and galaxy sample properties. For a comparison of the different techniques adopted see @colberg_aspen-amsterdam_2008.
It is also important to stress the fact that galaxies and haloes trace the void distribution in a similar fashion. @padilla_spatial_2005 have studied voids defined by the spatial distribution of haloes and galaxies finding that they have comparable general statistical and dynamical properties, such as abundances, correlation functions and velocity fields.
Moreover, the statistics of void and matter distributions are strongly related [@white79] and therefore voids can provide simple and useful information on the clustering pattern, giving clues on the formation and evolution of overdense structures. Observations of the large scale structure traced by voids can also be used to constrain cosmological models [e.g. @peebles_void_2001; @park_challenge_2012; @kolokotronis_supercluster_2002; @colberg_voids_2005; @lavaux_precision_2010; @bos_darkness_2012; @biswas_voids_2010; @benson_galaxy_2003; @park_challenge_2012; @bos_less_2012; @hernandez-monteagudo_signature_2012; @clampitt_voids_2012; @sutter_first_2012] and to shed light on the mechanisms of galaxy evolution and its dependence on the large scale environment [@lietzen_environments_2012; @hahn_properties_2007; @hahn_evolution_2007; @ceccarelli_low_2012; @ceccarelli_large-scale_2008].
Voids in the galaxy distribution, in a first approximation, can be described as simple underdense regions which have nearly spherical shapes and isotropic expansion motions [@icke_1984; @weygaert_bertschinger_1996; @padilla_spatial_2005; @ceccarelli_voids_2006]. Nevertheless, in more detailed analyses it became clear that as voids are not isolated structures but rather part of an intricate network, their structure and dynamics are more complicated [@bertschinger; @melott_shandarin_1990; @mathis_white_2002; @colberg_etal_2005; @Shandarin_etal_2006; @platen_etal_2008; @aragon-calvo_hierarchical_2012]. The formation and evolution of voids is strongly affected by their surrounding large scale environment [@sheth_hierarchy_2004]. The void distribution also evolves accordingly, as matter collapses to form structures and galaxies dissipate from voids, making a supercluster-void network [@frisch_evolution_1995; @einasto_supercluster_1997; @einasto_multimodality_2012]. In order to deepen our understanding of the nature of voids and the evolution of their properties, it is crucial to take into account the large scale structure where they are embedded [@more_a_do_about_nothing].
Essentially, the hierarchy of voids arises by the assembly of matter in the growing nearby structures. @sheth_hierarchy_2004 suggest that while some voids remain as underdense regions, other voids fall in on themselves due to the collapse of dense structures surrounding them. According to this scenario, void evolution exhibits two opposite processes, expansion and collapse, being the dominant process determined by the global density around the voids. The distinction between these two types of void behavior depends on the surrounding environment. It is expected that the large underdense regions with surrounding overdense shells will undergo a ”void-in-cloud” evolution mode. These voids are likely to be squeezed as the surrounding structures tend to collapse onto them.
On the other hand, voids in an environment more similar to the global background density will expand and remain as underdense regions following a so called ”void-in-void” mode. In this scenario, it could be expected that largest voids at present remain stable, and thus are unlikely to be surrounded by overdense regions that are massive enough to produce a contraction. However, many of the smallest voids at present may show surrounding overdense shells. Inspired in this schematic scenario we analyze the void-size dependence of the relative population of voids embedded in low density and overdense regions.
This paper is organized as follows: in Section \[S\_data\] we describe the galaxy and mock catalogs and the corresponding void catalogues. In Section \[S\_modes\] we provide the analysis of the environments of voids in the SDSS. A comparison of observational results to the numerical simulation and the mock catalogue is given in Section \[S\_teo\]. Finally, we discuss our results in Section \[S\_concl\].
Data {#S_data}
=====
Galaxy catalogue and void sample
--------------------------------
The observational data used in this work were extracted from the Main Galaxy Sample [@strauss_spectroscopic_2002] of the Sloan Digital Sky Survey data release 7 [<span style="font-variant:small-caps;">SDSS-DR7,</span> @abazajian_seventh_2009]. The SDSS contains CCD imaging data in five photometric bands [UGRIZ, @fukugita_sloan_1996; @smith_ugriz_2002]. The <span style="font-variant:small-caps;">SDSS-DR7</span> spectroscopic catalogue comprises in this release 929,555 galaxies with a limiting magnitude of .
The voids in the galaxy distributions are identified using volume complete samples. We adopt a limiting redshift $z=0.08$ and maximum absolute magnitude in the $r-band$ $M_r=-19.2$. The limiting redshift of the sample is chosen on the basis of a compromise between the quality of the void sample, determined by the dilution of the sample of galaxies, and the number of voids, required to be large in order to achieve statistically significant results.
We have explored the effects of shot noise on the void identification in both real data and simulations in @ceccarelli_voids_2006. We found that for low-density galaxy samples, the identification of small voids ($R_{void} <$ 10 h$^{-1}$ Mpc) is limited by shot noise, however we can identify voids as small as these in denser galaxy samples that can be obtained by restricting the analysis to low redshifts ($z \lessapprox 0.1$ in SDSS). Our sample is redshift limited to have a large enough galaxy density to avoid the effects of shot noise on small voids (down to 5 h$^{-1}$ Mpc), but also a large enough volume to provide a statistically significant number of objects. We apply the void finding algorithm described in @padilla_spatial_2005 and @ceccarelli_voids_2006 to this volume limited sample. In order to prevent effects of survey geometry and redshift limits on the process of void identification we avoid including voids near the survey edges. The algorithm identifies the largest spherical regions where the overall density contrast is at most $\Delta=-0.9$, and are not contained in any other sphere satisfying the same condition. According to this procedure, a void is located at the centre of an underdense sphere and has a scale size equal to the sphere radius. We obtain 131 voids in the SDSS sample with radii ranging from 5 h$^{-1}$Mpc to 22 h$^{-1}$Mpc.
Large scale numerical simulation {#ss:sim_mock}
--------------------------------
We use the snapshot corresponding to $z=0.0$ of the Millennium simulation [@springel_simulations_2005; @lemson_halo_2006] and the associated semi-analytic model of galaxy formation by @bower_flip_2008. The Millennium simulation utilizes a $\Lambda$CDM cosmological model with , , , , and based on WMAP observations [@spergel_first-year_2003] and 2dF Galaxy Redshift Survey [@colless_2df_2001]. The simulation follows the evolution of $2160^3$ particles, each with 8.6 $\times\, 10^8\, \mathrm h^{-1}M_{\odot}$ through a comoving box of side . The dark matter haloes of the Millennium simulation were used to follow the simulated growth of galaxies by implementing a semi-analytic model of galaxy formation [<span style="font-variant:small-caps;">GALFORM</span>, @bower_flip_2008], which generates a population of galaxies within the simulation box. We used the full simulation box to identify voids and study the effects of dilution and redshift space distortion in the sample of voids. The semi-analytic galaxy catalogue from the simulation box comprises 2783 voids.
We use a mock catalogue, constructed by selecting galaxies within the simulation box. The mock catalogue is constructed by first positioning an observer in a random position within the numerical simulation box, and then reproducing the selection function and angular mask of the SDSS from this position. This results in a mock catalogue of galaxies of similar properties and observational biases to those of the SDSS catalogue. Positions in real space and peculiar velocities are available to test possible projection biases and to quantify the effects of redshift space distortions. The mock catalogue also provides information on SDSS photometric magnitudes, star formation rates and total stellar masses, based on computations from the semi–analytic model of galaxy formation [@bower_flip_2008]. This mock catalogue will be used in this work to calibrate our statistical methods, to interpret the data, and to detect any systematic biases in our procedures. In order to do this, we will treat the mock catalogue in exactly the same way as the real data. We found 113 mock voids, with radii ranging from to .
![Integrated density contrast profiles around voids of radii in the range 6-8 h$^{-1}$Mpc (upper panels) and 10-12 h$^{-1}$Mpc in SDSS (lower panels). The profiles in the left panels correspond to voids surrounded by overdense shells, i.e., S-Type voids. Right panels are for R-Type voids, showing a non–decreasing or ”rising” profile. Solid thick lines correspond to the mean density contrast profiles and black dots show the values of density contrast for each individual void as a function of distance. Dotted lines indicate the mean galaxy density, i.e., zero density contrast.[]{data-label="F_SDSS_prof"}](fig2.eps){width="84mm"}
![ ***Upper panels:*** Integrated number of voids as a function of radius in SDSS (solid lines, filled circles) and in the mock (dashed lines, open circles) catalogues, for voids with (S-type) and without (R-type) a surrounding overdense shell (left and right panels respectively). ***Lower panels:*** Fraction of total volume occupied by voids smaller than $ {\mathrm {R_{void}^{max}}}$ for voids with and without a surrounding overdense shell in SDSS (solid lines, filled circles) and in the mock (dashed lines, open circles) catalogues. []{data-label="F_SDSS_vol"}](fig3.ps){width="84mm"}
Large scale regions surrounding voids in galaxy catalogue {#S_modes}
=========================================================
@sheth_hierarchy_2004 proposed that void evolution is strongly determined by the larger regions surrounding them. In this framework, voids embedded in overdense environments tend to experience gravitational collapse more likely than expansion. Consequently, it is expected that most of the small voids with an overdense shell surrounding them have sank inward by the present epoch. Larger voids, on the other hand, are probably expanding in concordance with the formation of large structures shaping them. We analyze the relative abundance of voids embedded in overdense environments for different void sizes in the galaxy distribution. Such abundances are used in this section to test model predictions about voids in the SDSS spectroscopic galaxy catalogue.
Void Density
------------
@padilla_spatial_2005 provide mean relations between the maximum density, its radial distance to the void centre, and the void radius. In order to further explore these relations we study the integrated density contrast profiles of galaxies as a function of void radius and distance to the void centre. We compute for each void the integrated number density contrast of galaxies, $\Delta_g$, as:
$$\Delta_g (\mathrm d) =
\frac{\mathrm{n_g(d)}-\langle \mathrm{n} \rangle}{\langle \mathrm{n}
\rangle},$$
where $\langle \mathrm{n} \rangle$ is the mean number density of galaxies in the catalogue and $\mathrm{n_g(d)}$ is the number density of galaxies within a void–centered sphere of radius $d$. In order to compensate for the effects due to the survey limits and geometry on our density calculations we use appropriate random catalogues to estimate densities.
In the left panel of Figure \[F\_SDSS\_dens\] we show $\Delta_g$ as a function of void-centric distance (d) for all voids with size (${\mathrm{R_{void}}}$) in the range (grey solid lines). The thick black solid line represents the mean integrated density contrast for this void subsample, hereafter $\langle\Delta_g\rangle$. By inspection of this panel, it can be noticed that there is a significant variation of individual curves around the mean. We also highlight with thick dashed lines two profiles that illustrate two distinct behaviours: profiles with a noticeable maximum density contrast and a decline at larger distances; and those showing a non-decreasing profile that tends to the mean density at large distances from the void centres.
![ Fraction of S-Type voids as a function of void radius. Filled dots represent the fractions for the SDSS sample, grey boxes the corresponding fractions obtained in the mock sample and grey triangles the values computed in the simulation box, in real space. The solid line indicates the linear regression fit to SDSS data. Error bars represent the 68% confidence interval of the binomial distribution.[]{data-label="F_SDSS_frac"}](fig4.eps){width="84mm"}
In the right panel of Figure \[F\_SDSS\_dens\], we show $\langle\Delta_g\rangle$ as a function of $\mathrm{(R_{void},d)}$ for voids with radii in the range $5<{\mathrm{R_{void}/h^{-1}\,Mpc}}<16$. An element in this matrix represents $\langle\Delta_g\rangle$, averaged over voids in a given void size bin, integrated up to a given maximum void-centric distance d. Therefore, columns in this matrix are the mean integrated density contrast profile as a function of void size and projected distance. The superimposed dashed lines represent different density contrast levels. The solid line corresponds to a fit to the zero density contrast iso-contour in the range $5<{\mathrm{R_{void}/h^{-1}\,Mpc}}<16$, shown for future reference (see Section \[S\_teo\]). Underdense and overdense regions are shown with different colours, ranging from blue to orange for negative to positive density contrasts, respectively. The white color corresponds to a zero density contrast, i.e., regions where the integrated density is equal to the mean global density. This panel shows the general trend of the average behaviour of void profiles depending on void sizes. In the mean, the integrated density contrast profile of small voids exhibits a prominent shell, whereas in the case of large voids there is a smooth behaviour towards the global mean galaxy density.
The results shown in the left panel of Figure \[F\_SDSS\_dens\] suggest that it is possible to classify voids according to their large-scale radial density profiles allowing for a subdivision of the void sample into two types of voids. We notice that mean integrated density contrast profiles can be defined for ${\mathrm{R_{void}}}$ intervals. These average curves have a well defined maximum at a distance $\mathrm {d_{max}}$ from the void centre, except for the largest voids that exhibit an asymptotically increasing profiles. When there is not a clear local maximum, we adopt $\mathrm {d_{max}}=3
{\mathrm{R_{void}}}$. This choice is justified by the fact that small and intermediate size voids have $\mathrm {d_{max}} \sim 3
{\mathrm{R_{void}}}$ (right panel of Figure \[F\_SDSS\_dens\]). Thus, we classify voids into two subsamples according to positive or negative values of the integrated density contrast at $\mathrm{d_{max}}$. Voids surrounded by an overdense shell (hereafter S-Type voids) correspond to $\Delta_g(\mathrm {d_{max}}) > 0$. On the other hand R-Type voids ($\Delta_g(\mathrm{ d_{max}}) <
0$) correspond to voids with continuously rising density profiles.
In the left panels of Fig. \[F\_SDSS\_prof\] we show the galaxy density profile for S-Type voids with radii in two ranges $6$ h$^{-1}$Mpc $< {\mathrm{R_{void}}}< 8$ h$^{-1}$Mpc and $10$ h$^{-1}$Mpc $<
{\mathrm{R_{void}}}< 12$ h$^{-1}$Mpc. The solid lines correspond to the mean density whereas dots correspond to the individual void profiles. The profiles of R-Type voids are shown in the right panel of this figure. It can be appreciated the similarity of the inner radial density profiles of S and R-Type voids, regardless of their different large scale environment. For the total sample of voids, we obtain 59 S-Type and 82 R-Type voids. The solid lines in the upper panels of Figure \[F\_SDSS\_vol\] show the cumulative number of voids corresponding to the S-Type (left panel) and R-Type sample (right panel) in the SDSS catalogue. It can be noticed the opposite trends distinguishing both samples: while the quantity of S-Type voids decreases as the radius increases, that of R-Type voids monotonically increases. This is consistent with the theoretical results indicating that large voids are unlikely to be surrounded by overdense shells [@more_a_do_about_nothing].
The relative volume occupied by voids smaller than a given maximum void radius $\mathrm{R_{void}^{max}}$ is shown in the lower panels of Figure \[F\_SDSS\_vol\], for S and R-types separately. As it can easily be seen in the left lower panel, approximately 10$\%$ of the catalogue is occupied by S-type voids (solid lines, filled dots). On the other hand, the volume occupied by R-Type voids increases with radius (solid lines and filled dots in the lower right panel of Figure \[F\_SDSS\_vol\]) and they approximately occupy half of the total volume, in agreement with the well known statement that most of the volume of the Universe is filled by voids.
{width="45.00000%" height="45.00000%"} {width="45.00000%" height="45.00000%"}
{width="45.00000%" height="45.00000%"} {width="45.00000%" height="45.00000%"}
Fraction of S and R-Type voids
------------------------------
In this subsection we analyse the fraction of S and R-Type voids as a function of void radius. The results are displayed in Fig. \[F\_SDSS\_frac\] for SDSS data and the results of the mock catalogue (discussed in section \[S\_teo\]). Error-bars correspond to the 68% confidence interval of the binomial ditribution. It can be seen in this figure that most of the smallest voids are surrounded by an overdense region while larger voids tend to have smoothly rising density profiles. A similar behaviour has been reported by @sheth_hierarchy_2004, based on a void sample identified in numerical simulations. They provide a theoretical framework based on the excursion set formalism where the void population is twofold, according to their evolutionary processes. Our classification of voids density profiles resembles the idea of a void-in-void and void-in-cloud dichotomy in the evolutionary processes presented by @sheth_hierarchy_2004. We have found that the relative number of voids in S-Type decreases as the size increases. The comparison with simulation results is detailed in the next section.
$\Lambda$CDM comparison {#S_teo}
=======================
In this section we use the mock catalogue constructed using the numerical simulation and identify and classify voids using the same methods as applied to the observations in the previous sections. The Millennium simulation, described in Section \[S\_data\], provides three dimensional positions and velocities of semi–analytic galaxies. This allows to analyze the properties of voids defined by galaxy positions for different types of large-scale surrounding regions. The simulation and the mock catalogue can also be used to test possible systematics in the definition and properties of voids, which can change from real to redshift space [@schmidt_size_2001; @ryden_voids_1996].
With the aim to compare the results of galaxy catalogue to numerical results, we study the behaviour of the void integrated density profiles $\Delta_{g}(d)$ as a function of both void radius (${\mathrm{R_{void}}}$) and void–centric distance ($d$), in the mock catalogue and in the simulation box. The semi–analytic sample of galaxies comprises all galaxies from the simulation box. As is described in subsection \[ss:sim\_mock\], the mock galaxy catalogue is constructed from the semi–analytic sample of galaxies, and with the same angular mask and redshift space effects than the SDSS spectroscopic galaxy catalogue. This allows a better understanding of the spread in the behaviours of the different void profiles. In panel (a) and (c) of Fig. \[F\_MOCK\_dens\] we show the integrated galaxy overdensity profiles of voids with sizes ranging in both, the mock catalogue and the simulation box. In order to avoid shot–noise in the estimation of the density profile, we require galaxy counts to be above 100 in each radial distance bin. Due to this limitation, the curves are trimmed at small radial distances. As can be seen in panel (a), the individual curves span a wide range of behaviours for the galaxy number overdensity $\Delta_g$ as a function of centre void distance $d$. Similar features are observed, although with better coverage, in the sample of voids in the simulation box (panel c of Fig. \[F\_MOCK\_dens\]). Given the high number of voids found in the simulation volume we show just a random subset of void profiles (gray lines). The curves can again be subdivided in groups showing two behaviours, that of a prominent maximum indicating the presence of an averaged overdensity in a surrounding shell, and the characteristic smooth rise to $\Delta=0$ at large distances. In order to illustrate these distinct behaviours, the black dashed lines in panel (a) show the profiles of two voids that are clearly classified into S and R types. The black solid line in both panels, a and c, represents the mean integrated density profile $\langle\Delta_g\rangle$ of the mock and simulation samples, respectively. In the case of the simulation box, due to the larger volume covered, the number of voids in the sample is larger, allowing a more detailed analysis of the behaviour of void profiles. For instance, in panel (c) the color map represents a 2-D histogram of the number of voids in bins of the integrated galaxy density profile ($\Delta_g$) and the void–centric distance ($d$), where redder colours indicate higher number counts, and blue colour accounts for one measurement in that bin. As can be seen, the lack of red coloured bins at separations $<30$h$^{-1}$Mpc shows that the distribution of overdensity values becomes broader. It should be noticed that each curve is sampled at a discrete set of values for the void-centric distance (the bins in $d$), giving rise to the noisy aspect of the colour map for low density values.
In the above paragraph we have illustrated for a fixed range of void sizes () the features observed in void profiles. We also provide in figure \[F\_MOCK\_dens\] an analysis of the mean void density profile $\langle\Delta_g\rangle$ as a function of void size. This is shown in panels (b) and (b), where mean galaxy overdensity is presented as a function of the void radius (${\mathrm{R_{void}}}$) and the distance to the void centre ($d$) in the mock catalogue and in the simulation box, respectively. Orange colours represent overdense regions ($\langle\Delta_g\rangle > 0$) whereas cyan colours correspond to underdense regions (negative overdensities). The dashed lines represent isodensity contours as labeled in the figure. The Solid lines in both panels represent an approximation to the $\Delta_g=0$ isodensity contour in the mock sample, also shown for comparison in the simulation sample (panel d). As it can be noticed, small voids tend to exhibit a surrounding overdense shell in their mean profile, whereas larger voids show continuously rising mean density profiles. It is worth to mention that all the observational effects taken into account in the construction of the mock catalogue do not seem to affect in an appreciable way the mock results. This arises from comparing the mock results (upper panels in figure \[F\_MOCK\_dens\]) with the simulation results (lower panels). In order to examine cosmic variance effects on our results, we have also constructed several mock catalogues by placing observers in unconnected locations in the simulation box. We examine the density profiles around voids. In general they show similar behaviour with small differences between different mocks which are comparable to the difference between mock and observational data (see Figures \[F\_SDSS\_dens\] and \[F\_MOCK\_dens\]).We stress the fact that our theoretical results are in good agreement withobservations as can be seen when comparing to Figure \[F\_SDSS\_dens\].
We select S and R-Type void samples in the mock catalogue following the same criteria described in section \[S\_modes\] obtaining 48 and 65 S and R-Type voids respectively. The density profiles corresponding to S and R-Type voids are shown in Figure \[fig:prof\_mock\] (left and right panels respectively). In Figure \[F\_SDSS\_frac\] we present the fraction of S and R-Type voids with respect to the total number as a function of void size in the mock catalogue, where the squares indicate the fraction and error-bars represent the confidence interval of the binomial distribution. It is relevant that the mock behaviour is indistinguishable from the observations. We also show the results obtained selecting galaxies brighter than $M_r =$ -19.2 in the simulation box (triangles in figure \[F\_SDSS\_frac\]). Thus, this sample of simulated galaxies have the same absolute limiting magnitude than the volume-complete samples of SDSS and mock galaxies.
Based on analytical formulations for the evolution of inhomogeneities on the mass distribution in the Universe [e.g. @peebles_principles_1993] and the theoretical analysis of void evolution [@sheth_hierarchy_2004] it is natural to expect a dependence of the peculiar velocity field around voids with the presence of a surrounding over-dense shell. In order to examine this effect, we have studied the mean peculiar velocity around voids traced by the semi-analytic galaxies in the full simulation box, for S and R-Type voids separately. In Figure \[F\_MOCK\_vel\] we show the mean velocity profiles of these two subsamples of voids in the mock catalogue, where the dashed line indicates S-Type and the solid line R-Type voids. We adopt positive velocities to indicate expansion and negative velocities for infall motions. As can be seen in the figure, S-Type voids show substantial infall velocities (raising to ) at the same distances where the overdense shells are located (d/${\mathrm{R_{void}}}\sim$ 3) whereas at smaller distances (d/${\mathrm{R_{void}}}\lesssim$ 2) the velocity field is characterized by expansion. On the other hand, R-Type voids, lacking an overdense shell, show only expansion velocities. Both samples show significant expansion velocities at void shells (0.8 $\leq$ d/${\mathrm{R_{void}}}\leq$ 1.2), with a maximum (velocities $\sim$ 300 km s$^{-1}$) reached at d/${\mathrm{R_{void}}}\simeq$1.
![ Radial integrated density profile around voids in the Mock catalogue with void radii in the ranges 6-8 h$^{-1}$Mpc (upper panels) and 10-12 h$^{-1}$Mpc (lower panels). The left panels correspond to voids surrounded by large scale overdense shells and the right panels correspond to voids in large scale underdense regions. Solid lines correspond to the mean density and dots show the individual voids. Dotted lines indicate the mean galaxy density. []{data-label="fig:prof_mock"}](fig6.eps){width="84mm"}
![ Mean radial velocity profile for voids surrounded (dashed line) and not surrounded (solid line) by overdense shells in the mock catalogue. Void radii are in the range 6-22 h$^{-1}$ Mpc. The dotted line indicates a zero mean velocity (v = 0 km s$^{-1}$). []{data-label="F_MOCK_vel"}](fig7.ps){width="84mm"}
Summary and discussion {#S_concl}
======================
We have performed a statistical study of the void phenomenon focussing on void environments. We have examined the distribution of galaxies around voids in the SDSS, by computing the integrated density contrast profile. There is a correlation between void–centric distance to the shell of maximum density and void radius. We defined separation criterion to characterize voids according to their surrounding environment, giving rise to S-Type (Shell) and R-Type (Rising) voids. We found that small voids are more frequently surrounded by over-dense shells. On the other hand, larger voids are more likely classified as R-Type, i.e., with a non–decreasing integrated density contrast profile, which smoothly rises towards the mean galaxy density. The fraction of voids surrounded by overdense shells continuously decreases as the void size increases.
In order to test and interpret the observed properties of voids we identified and analyzed voids in a numerical simulation with a semi-analytic galaxy catalog. We used the same procedures than those applied to the observational galaxy catalog to identify and analyze voids, particularly when using the mock catalog. We have computed the velocity curves for the two types of voids, in the full box and in the mock sample. The results suggest that there is a relation between our separation criterion and the evolution of voids, as was suggested previously by @sheth_hierarchy_2004. The good agreement obtained between the SDSS voids sample and the mock catalog provides additional support to the viability of a $\Lambda$CDM model to reproduce the large scale structure of the universe as defined by the void network. In this work we give the first observational evidence of multiple modes in the void hierarchy, which have been predicted in numerical simulations, see for example [@more_a_do_about_nothing] and references therein.
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank the referee, Paul Sutter, for the thorough, constructive and helpful comments and suggestions on the manuscript, which greatly improved this work. This work has been partially supported by Consejo de Investigaciones Científicas y Técnicas de la República Argentina (CONICET) and the Secretaría de Ciencia y Técnica de la Universidad Nacional de Córdoba (SeCyT). LC, DP and ML acknowledge research fellowships from CONICET. NP acknowledges support from Fondecyt 1110328 and BASAL-PFB06 “Centro de Astronomía y Tecnologías afines”.
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. The Participating Institutions are the American Museum of Natural History, Astrophysical Institute Potsdam, University of Basel, University of Cambridge, Case Western Reserve University, University of Chicago, Drexel University, Fermilab, the Institute for Advanced Study, the Japan Participation Group, Johns Hopkins University, the Joint Institute for Nuclear Astrophysics, the Kavli Institute for Particle Astrophysics and Cosmology, the Korean Scientist Group, the Chinese Academy of Sciences (LAMOST), Los Alamos National Laboratory, the Max-Planck-Institute for Astronomy (MPIA), the Max-Planck-Institute for Astrophysics (MPA), New Mexico State University, Ohio State University, University of Pittsburgh, University of Portsmouth, Princeton University, the United States Naval Observatory, and the University of Washington.
The Millennium Simulation databases used in this paper and the web application providing online access to them were constructed as part of the activities of the German Astrophysical Virtual Observatory.
Some of the plots presented in this work were made by using R Software.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
G.M. Shore\
Department of Physics\
University of Wales, Swansea\
Swansea SA2 8PP, U.K.\
E-mail:
title: 'The $U(1)_A$ Anomaly and QCD Phenomenology[^1]'
---
Introduction
============
The $U(1)_A$ anomaly has played an important historical role in establishing QCD as the theory of the strong interactions. The description of radiative decays of the pseudoscalar mesons in the framework of a gauge theory requires the existence of the electromagnetic axial anomaly and determines the number of colours to be $N_c = 3$. The compatibility of the symmetries of QCD with the absence of a ninth light pseudoscalar meson – the so-called ‘$U(1)_A$ problem’ – in turn depends on the contribution of the colour gauge fields to the anomaly. More recently, it has become clear how the anomaly-mediated link between quark dynamics and gluon topology (the non-perturbative dynamics of topologically non-trivial gluon configurations) is the key to understanding a range of phenomena in polarised QCD phenomenology, most notably the ‘proton spin’ sum rule for the first moment of the structure function $g_1^p$.
In this paper, based on original research performed in a long-standing collaboration with Gabriele Veneziano, we review the role of the $U(1)_A$ anomaly in describing a wide variety of phenomena in QCD, ranging from the low-energy dynamics of the pseudoscalar mesons to sum rules in polarised deep-inelastic scattering. The aim is to show how these experiments reveal subtle aspects of quantum field theory, in particular topological gluon dynamics, which go beyond simple current algebra or parton model interpretations.
We begin in section 2 with a brief review of the essential theoretical toolkit: anomalous chiral Ward identities, Zumino transforms, the renormalisation group, and the range of expansion schemes associated with large $N_c$, notably the OZI approximation. Then, in section 3, we build on Veneziano’s seminal 1979 paper [@Veneziano:1979ec] to describe how the pseudoscalar mesons saturate the Ward identities in a way compatible with both the renormalisation group and large-$N_c$ constraints and derive a generalisation of the famous Witten-Veneziano mass formula for the $\eta'$ which incorporates, but goes beyond, the original large-$N_c$ derivation [@Shore:1999tw; @Shore:2006mm].
In section 4, we turn to QCD phenomenology and describe how this intuition on the resolution of the $U(1)_A$ problem allows a quantitative description of low-energy pseudoscalar meson physics, especially radiative decays, the determination of the pseudoscalar decay constants, and meson-nucleon couplings. We review the $U(1)_A$ extension of the Goldberger-Treiman formula first proposed by Veneziano [@Veneziano:1989ei] as the key to understanding the ‘proton spin’ problem and test an important hypothesis on the origin of OZI violations and their relation to the renormalisation group. Low-energy $\eta$ and $\eta'$ physics is currently an active experimental field and we explain the importance of an accurate determination of the couplings $g_{\eta NN}$ and $g_{\eta' NN}$ in elucidating the role of gluon topology in QCD.
All of these low-energy phenomena have counterparts in high-energy, polarised deep-inelastic scattering. This enables us to formulate a new sum rule for the first moment of the polarised photon structure function $g_1^\c$ (section 6). The dependence of this sum rule on the invariant momentum of the off-shell target photon measures the form factors of the 3-current AVV Green function and encodes a wealth of information about the realisation of chiral symmetry in QCD, while its asymptotic limit reflects both the electromagnetic and colour $U(1)_A$ anomalies. We show how this sum rule, which we first proposed in 1992 [@Narison:1992fd; @Shore:1992pm], may soon be tested if the forthcoming generation of high-luminosity $e^+ e^-$ colliders, currently conceived as $B$ factories, are run with polarised beams [@Shore:2004cb].
The most striking application of these ideas is, however, to the famous ‘proton spin’ problem, which originated with the observation of the violation of the Ellis-Jaffe sum rule for the first moment of the polarised proton structure function $g_1^p$ by the EMC collaboration at CERN in 1988. This experiment, and its successors at SLAC, DESY (HERMES) and CERN (SMC, COMPASS) determined the axial charge $a^0$ of the proton. In the simple valence-quark parton model, this can be identified with the quark spin and its observed suppression led to an intense experimental and theoretical search over two decades for the origin of the proton spin. In fact, as Veneziano was the first to understand [@Veneziano:1989ei], $a^0$ does not measure spin in QCD itself and its suppression is related to OZI violations induced by the $U(1)_A$ anomaly.
In a series of papers, summarised in section 5, we have shown how $a^0$ decouples from the real angular momentum sum rule for the proton (the form factors for this sum rule are given by generalised parton distributions (GPDs) which can be extracted from less inclusive measurements such as deeply-virtual Compton scattering) and is instead related to the gluon topological susceptibility [@Shore:1990zu; @Shore:1991dv]. The experimentally observed suppression is a manifestation of topological charge screening in the QCD vacuum. In a 1994 paper with Narison [@Narison:1994hv], using QCD spectral sum rule methods, we were able to compute the slope of the topological susceptibility and give a quantitative prediction for $a^0$. Our prediction, $a^0 = 0.33$, has within the last few months been spectacularly confirmed by the latest data on the deuteron structure function from the COMPASS and HERMES collaborations.
Hopefully, this impressive new evidence for topological charge screening will provide fresh impetus to experimental ‘spin’ physics - first, to verify the real angular momentum sum rule by measuring the relevant GPDs, and second, to pursue the programme of target-fragmentation studies in semi-inclusive DIS at polarised $ep$ colliders which we have proposed as a further test of our understanding of the $g_1^p$ sum rule [@Shore:1997tq].
0.5cm This review has been prepared in celebration of the 65th birthday of Gabriele Veneziano. I first met Gabriele when I came to Geneva as a CERN fellow in 1981. In fact, our first interaction was across a tennis court, in a regular Friday doubles match with Daniele Amati and Toine Van Proeyen. I like to think that in those days I could show Gabriele a thing or two about tennis – physics, of course, was a different matter. It has been my privilege through these ensuing 25 years to collaborate with one of the most brilliant and innovative physicists of our generation. But it has also been fun. As all his collaborators will testify, his good humour, generosity to younger colleagues, and enthusiasm in thinking out solutions to the deepest and most fundamental problems in particle physics and cosmology make working with Gabriele not only intellectually rewarding but hugely enjoyable.
In his contribution to the ‘Okubofest’ in 1990 [@Veneziano:Okubo], Gabriele concluded an account of the relevance of the OZI rule to $g_1^p$ by hoping that he had ‘made Professor Okubo happy’. In turn, I hope that this review will make Gabriele happy: happy to recall how his original ideas on the $U(1)_A$ problem have grown into a quantitative description of anomalous QCD phenomenology, and happy at the prospect of new discoveries from a rich programme of experimental physics at future polarised colliders. It is my pleasure to join all the contributors to this volume in wishing him a happy birthday.
The $U(1)_A$ anomaly and the topological susceptibility
=======================================================
We begin by reviewing some essential features of the $U(1)_A$ anomaly, chiral Ward identities and the renormalisation group, placing particular emphasis on the role of the gluon topological susceptibility. As we shall see, the anomaly provides the vital link between quark dynamics and gluon topology which is essential in understanding a range of phenomena in polarised QCD phenomenology.
Anomalous chiral Ward identities {#sec 2.1}
--------------------------------
An anomaly arises when a symmetry which is present in the classical limit cannot be consistently imposed in a quantum field theory. The original example of an anomaly, and one which continues to have far-reaching implications for the phenomenology of QCD, is the famous Adler-Bell-Jackiw axial anomaly [@Adler:1969; @BellJackiw:1969; @AdlerBardeen:1969], which was first understood in its present form in 1969. In fact, calculations exhibiting what we now recognise as the anomaly had already been performed much earlier by Steinberger in his analysis of meson decays [@Steinberger:1949] and by Schwinger [@Schwinger:1951].
Anomalies manifest themselves in a number of ways. The original derivations of the axial anomaly involved the impossibility of simultaneously imposing conservation of both vector and axial currents due to regularisation issues in the AVV triangle diagram in QED. More generally, they arise as anomalous contributions to the commutation relations in current algebra. A modern viewpoint, due to Fujikawa [@Fujikawa:1979], sees anomalies as due to the non-invariance of the fermionic measure in the path integral under transformations corresponding to a symmetry of the classical Lagrangian. In this approach, the result of a chiral transformation $q \rta e^{i\a^a T^a \c_5}q$ on the quark fields in the QCD generating functional $W[V_{\m 5}^a,V_\m^a,\o,S_5^a,S^a]$ defined as[^2]
$$e^{iW} = \int \DD A \DD\bar q \DD q ~\exp \biggl[i\int dx\bigl(
{\cal L}_{\rm QCD} + V^{\m a}_5 J_{\m5}^a +
V^{\m a} J_\m^a + \o Q + S_5^a \phi_5^a
+ S^a \phi^a\bigr)\biggr]
\label{eq:ba}$$
is $$\int \DD A \DD\bar q \DD q~\Bigl[\pl^\m J_{\m5}^a - \sqrt{2n_f} \d^{a0} Q
-d_{abc}m^b\phi_5^c - \d \Bigl(\int d^4x {\cal L}_{\rm QCD}\Bigr)\Bigr]~
\exp \biggl[ \ldots \biggr] = 0
\label{eq:bb}$$ The terms in the square bracket are simply those arising from Noether’s theorem, including soft breaking by the quark masses, with the addition of the anomaly involving the gluon topological charge density $Q$. Re-expressing the chiral variation of the elementary fields in terms of a variation with respect to the sources $V_{\m5}^a, V_\m^a, \theta,
S_5^a, S^a$ then gives the functional form of the anomalous chiral Ward identities: $$\begin{aligned}
&\pl_\m W_{V_{\m5}^a} &- \sqrt{2n_f} \d_{a0} W_{\o}-
d_{abc} m^b W_{S_5^c}
\nonumber\\
&{}&+ f_{abc} V_\m^b W_{V_{\m 5}^c}
+ f_{abc} V_{\m 5}^b W_{V_\m^c}
+ d_{abc} S^b W_{S_5^c}
- d_{abc} S_5^b W_{S^c}
= 0
\label{eq:bc}\end{aligned}$$ where we have abbreviated functional derivatives as suffices. This is the key to all the results derived in this section. It makes precise the familiar statement of the anomaly as $$\pl^\m J_{\m5}^a - \sqrt{2n_f} Q \d_{a0} - d_{abc}m^b \phi_5^c ~\sim~0
\label{eq:bd}$$
The chiral Ward identities for two and higher-point Green functions are found by taking functional derivatives of eq.(\[eq:bc\]) with respect to the sources. The complete set of identities for two-point functions is given in our review [@Shore:1998dm]. As an example, we find[^3] $$\pl_\m W_{V_{\m5}^a S_5^b} - \sqrt{2n_f}\d_{a0}W_{\o S_5^b} -
M_{ac}W_{S_5^c S_5^b} - \Phi_{ab} ~=~ 0
\label{eq:be}$$ which in more familiar notation reads $$\pl^\m \langle 0| T~ J_{\m 5}^a~ \phi_5^b |0\rangle - \sqrt{2n_f}\d_{a0}
\langle 0| T~ Q~\phi_5^b|0\rangle - d_{adc}m^d \langle 0|
T~\phi_5^c ~ \phi_5^b |0\rangle - d_{abc}\langle \phi^c\rangle = 0
\label{eq:bf}$$
The anomaly breaks the original $U(n_f)_L \times U(n_f)_R$ chiral symmetry to $SU(n_f)_L \times SU(n_f)_R \times U(1)_V / Z_{n_f}^V$ and the quark condensate spontaneously breaks this further to the coset $SU(n_f)_L \times SU(n_f) / SU(n_f)_V$. Goldstone’s theorem follows immediately. In the chiral limit, there are $(n_f^2 - 1)$ massless Nambu-Goldstone bosons, which acquire masses of order $\sqrt{m}$ for non-zero quark mass. There is no flavour singlet Nambu-Goldstone boson since the corresponding current is anomalous.
The zero-momentum Ward identities are especially important here, since they control the low-energy dynamics. With the assumption that there are no exactly massless particles coupling to the currents, we find $$\begin{aligned}
&\sqrt{2n_f} \d_{a0} W_{\o\o} + M_{ac} W_{S_5^c \o} = 0
\nonumber\\
&\sqrt{2n_f} \d_{a0} W_{\o S_5^b} + M_{ac} W_{S_5^c S_5^b} + \Phi_{ab} = 0
\label{eq:bg}\end{aligned}$$
Another key element of our analysis will be the chiral Ward identities for the effective action $\C[V_{\m 5}^a,V_\m^a, Q, \phi_5^a, \phi^a]$, defined as the generating functional for vertices which are 1PI with respect to the set of fields $Q, \phi_5^a$ and $\phi^a$ but [*not*]{} the currents $J_{\m 5}^a$, $J_\m^a$. This is achieved using the partial Legendre transform (or Zumino transform): $$\C[V_{\m5}^a, V_\m^a, Q, \phi_5^a, \phi^a]~=~ W[V_{\m5}^a, V_\m^a, \o,
S_5^a, S^a] - \int dx~\Bigl(\o Q + S_5^a \phi_5^a + S^a \phi^a \Bigr)
\label{eq:bh}$$ The chiral Ward identities for $\C$ are $$\begin{aligned}
&\pl_\m \C_{V_{\m5}^a} &- \sqrt{2n_f} \d_{a0} Q - d_{abc}m^b \phi_5^c
\nonumber\\
&{}&+ f_{abc} V_{\m}^b \C_{V_{\m5}^c}
+ f_{abc} V_{\m5}^b \C_{V_{\m}^c}
- d_{abc} \phi_5^c \C_{\phi^b}
+ d_{abc} \phi^c \C_{\phi_5^b} = 0
\label{eq:bi}\end{aligned}$$ Again, the zero-momentum identities for the two-point vertices play an important role: $$\begin{aligned}
&\Phi_{ac}\C_{\phi_5^c Q} - \sqrt{2n_f} \d_{a0} = 0
\nonumber\\
&\Phi_{ac} \C_{\phi_5^c \phi_5^b} - M_{ab} = 0
\label{eq:bj}\end{aligned}$$ These will be used in section 3 to construct an effective action which captures the low-energy dynamics of QCD in the pseudoscalar sector.
Topological susceptibility
--------------------------
The connection with topology arises through the identification of the gluon operator $Q$ in the anomaly with a topological charge density. $Q$ is a total divergence: $$Q ~=~ {\a_s\over8\pi}~{\rm tr}~G_{\m\n} \tilde G^{\m\n} ~=~ \pl^\m K_\m
\label{eq:bk}$$ where $K_\m$ is the Chern-Simons current, $$K_\m = {\a_s\over4\pi}\e_{\m\n\r\s}{\rm tr}\bigl(A^\n G^{\r\s} -
{1\over3}gA^\n [A^\r,A^\s]\bigr)
\label{eq:bl}$$ Nevertheless, the integral over (Euclidean) spacetime of $Q$ need not vanish. In fact, for gauge field configurations such as instantons which become pure gauge at infinity, $$\int d^4 x~Q ~=~ n \in {\bf Z}
\label{eq:bm}$$ where the integer $n$ is the topological winding number, an element of the homotopy group $\pi_3(SU(N_c))$.
The form of the anomaly is then understood as follows. Under a chiral transformation, the fermion measure in the path integral transforms as (for one flavour) $$\DD \bar q \DD q ~\rta~ e^{-2i\a\int dx \varphi_i^\dagger \c_5 \varphi_i}~
\DD \bar q \DD q ~=~ \exp^{-2i\a (n_+ - n_-)}~\DD\bar q \DD q
\label{eq:bn}$$ where $\varphi_i$ is a basis of eigenfunctions of the Dirac operator $\Dslash$ in the background gauge field. The non-zero eigenvalues are chirality paired, so the Jacobian only depends on the difference $(n_+ - n_-)$ of the positive and negative chirality zero modes of $\Dslash$. Finally, the index theorem relates the anomaly to the topological charge density: $${\rm ind}\Dslash ~=~ n_+ - n_- ~=~ \int d^4 x ~Q
\label{eq:bo}$$
The topological susceptibility $\chi(p^2)$ is defined as the two-point Green function of $Q$, viz. $$\chi(p^2) = i\int dx~e^{ipx}\langle 0|T~Q(x)~Q(0)|0\rangle
\label{eq:bp}$$ We are primarily concerned with the zero-momentum limit $\chi(0) =
W_{\o\o}(0)$. Combining eqs.(\[eq:bg\]) gives the crucial Ward identity satisfied by $\chi(0)$: $$2n_f \chi(0) = M_{0a} W_{S_5^a S_5^b} M_{b0} + (M\Phi)_{00}
\label{eq:bq}$$ that is, $$n_f^2 \int dx~\langle 0|T~Q(x)~Q(0)|0\rangle ~=~
\int dx~m^a m^b\langle 0|T~\phi_5^a(x)~\phi_5^b(0)|0\rangle
~+~ m^a \langle \phi^a\rangle
\label{eq:br}$$ Determining exactly how this is satisfied in QCD is at the heart of the Witten-Veneziano approach to the $U(1)_A$ problem [@Witten:1979vv; @Veneziano:1979ec].
The zero-momentum Ward identities allow us to write a precise form for the topological susceptibility in QCD in terms of just one unknown dynamical constant [@DiVecchia:1980ve]. To derive this, recall that the matrix of two-point vertices is simply the inverse of the two-point Green function matrix, so in the pseudoscalar sector we have the following inversion formula: $$\begin{aligned}
&\C_{QQ} = - \Bigl(W_{\o\o} - W_{\o S_5^a} (W_{S_5 S_5})_{ab}^{-1}
W_{S_5^b \o} \Bigr)^{-1}
\label{eq:bs}\end{aligned}$$ Using the identities (\[eq:bg\]) and (\[eq:bq\]), this implies that at zero momentum $$\C_{QQ}^{-1} = - \chi \Bigl(1 - 2n_f \chi (M\Phi)_{00}^{-1} \Bigr)^{-1}
\label{eq:bt}$$ and inverting this relation gives $$\chi = - \C_{QQ}^{-1} \Bigl(1 - 2n_f \C_{QQ}^{-1}
(M\Phi)_{00}^{-1} \Bigr)^{-1}
\label{eq:bu}$$ Finally, substituting for $(M\Phi)_{00}^{-1}$ using the definitions above, we find the following important identity which determines the quark mass dependence of the topological susceptibility in QCD: $$\chi(0) = -A \biggl(1 - A \sum_q {1\over m_q \la \bar q q\ra}\biggr)^{-1}
\label{eq:bv}$$ where we identify the non-perturbative coefficient $A$ as $\C_{QQ}^{-1}$.
Notice immediately how this expression exposes the well-known result that $\chi(0)$ vanishes if any quark mass is set to zero. In section 3, we will see how it also clarifies the role of the $1/N_c$ expansion in the $U(1)_A$ problem.
Renormalisation group
---------------------
The conserved current corresponding to a non-anomalous symmetry is not renormalised and has vanishing anomalous dimension. However, an anomalous current such as the flavour singlet axial current $J_{\m5}^0$ is renormalised. The composite operator renormalisation and mixing in the $J_{\m5}^0, Q$ sector is as follows [@Espriu:1982bw]: $$\begin{aligned}
&J_{\m5R}^0 = Z J_{\m5B}^0 ~~~~~~~~~~
\nonumber\\
&Q_R = Q_B - {1\over\sqrt{2n_f}}(1-Z) \pl^\m J_{\m5B}^0
\label{eq:bw}\end{aligned}$$ Notice the form of the mixing of the operator $Q$ with $\pl^\m J_{\m5}^0$ under renormalisation. This ensures that the combination $\bigl(\pl^\m J_{\m5}^0 - \sqrt{2n_f} Q\bigr)$ occurring in the $U(1)_A$ anomaly equation is RG invariant. The chiral Ward identities therefore take precisely the same form expressed in terms of the bare or renormalised operators, making precise the notion of ‘non-renormalisation of the anomaly’. We may therefore interpret the above Ward identities, which were derived in terms of the bare operators, as identities for the renormalised composite operators (and omit the suffix ${}_R$ for notational simplicity).
The renormalisation group equation (RGE) for the generating functional $W[V_{\m5}^a, V_\m^a, \o, S_5^a, S^a]$ follows immediately from the definitions (\[eq:bw\]) of the renormalised composite operators. Including also a standard multiplicative renormalisation $Z_\phi =
Z_m^{-1}$ for the pseudoscalar and scalar operators $\phi_5^a$ and $\phi^a$ and denoting the anomalous dimensions corresponding to $Z$ and $Z_\phi$ by $\c$ and $\c_\phi$ respectively, we find[^4] $$\DD W = \c\Bigl(V_{\m 5}^0 -
{1\over\sqrt{2n_f}}\pl_\m \theta\Bigr)W_{V_{\m5}^0}
+ \c_\phi\Bigl(S_5^a W_{S_5^a} + S^a W_{S^a}\Bigr) + \ldots
\label{eq:bx}$$ where $\DD = \Bigl(\m{\pl\over\pl\m} + \b{\pl\over\pl g} - \c_m\sum_q
m_q{\pl\over\pl m_q}\Bigr)\Big|_{V,\theta,S_5,S}$.
The RGEs for Green functions are found by functional differentiation of eq.(\[eq:bx\]) and can be simplified using the Ward identities. For example, for $W_{\o\o}$ we find $$\DD W_{\o\o} ~=~ 2\c W_{\o\o} + 2\c {1\over\sqrt{2n_f}}M_{0b}W_{\o S_5^b}
+ \ldots
\label{eq:by}$$ At zero momentum, we can then use the first identity in eq.(\[eq:bg\]) to prove that the topological susceptibility $\chi(0)$ is RG invariant, $$\DD \chi(0) ~=~ 0
\label{eq:bz}$$ which is consistent with its explicit expression (\[eq:bv\]).
A similar RGE holds for the effective action $\C[V_{\m 5}^a,V_\m^a,Q,\phi_5^a,\phi_5]$, which allows the scaling behaviour of the proper vertices involving $Q$ and $\phi_5^a$ to be determined [@Shore:1990wp; @Shore:1991dv; @Shore:1991pn]. This reads $$\DD\C = \c\Bigl(V_{\m5}^0 - {1\over\sqrt{2n_f}}\C_Q\pl_\m\Bigr)
\C_{V_{\m5}^0} - \c_\phi\Bigl(\phi_5^a \C_{\phi_5^a} +
\phi^a \C_{\phi^a}\Bigr) +\ldots
\label{eq:bzz}$$ An immediate consequence is that $\DD \C_{QQ} = 0$ at zero momentum, which ensures the compatibility of (\[eq:bv\]) with the RG invariance of $\chi(0)$.
$1/N_c$, the topological expansion and OZI
------------------------------------------
The final theoretical input into our analysis of the $U(1)_A$ problem and phenomenological implications of the anomaly concerns the range of dynamical approximation schemes associated with the large-$N_c$ limit. At various points we will refer either to the original large-$N_c$ expansion of ’t Hooft [@Hooft:1974], the topological expansion introduced by Veneziano [@Veneziano:TE] and the OZI limit [@Okubo; @Zweig; @Iizuka]. A very clear summary of the distinction between them is given in Veneziano’s ‘Okubofest’ review [@Veneziano:Okubo], which we follow here.
In terms of Feynman diagrams, the leading order in the large $N_c$, fixed $n_f$ (’t Hooft) limit is the most restrictive of these approximations, including only planar diagrams with sources on a single quark line and no further quark loops (Fig. \[fig:largeN\]).
![A typical Feynman diagram allowed in the large-$N_c$ limit. The dots on the quark loop represent external sources.[]{data-label="fig:largeN"}](largeN.eps){height="3.3cm"}
A better approximation to QCD is the quenched approximation familiar in lattice gauge theory. This is a small $n_f$ expansion at fixed $N_c$, i.e. excluding quark loops but allowing non-planar diagrams (Fig. \[fig:quenchtop\]).
An alternative is the topological expansion, which allows any number of internal quark loops, but restricts to planar diagrams at leading order. Provided the sources remain attached to the same quark line, this corresponds to taking large $N_c$ at fixed $n_f/N_c$. This means that quarks and gluons are treated democratically and the order of approximation is determined solely by the topology of the diagrams (Fig. \[fig:quenchtop\]).
![Feynman diagrams allowed in the quenched approximation (left) or leading order in the topological expansion (right).[]{data-label="fig:quenchtop"}](quenched.eps "fig:"){height="3.3cm"} ![Feynman diagrams allowed in the quenched approximation (left) or leading order in the topological expansion (right).[]{data-label="fig:quenchtop"}](topological.eps "fig:"){height="3.3cm"}
Finally, the OZI approximation is a still closer match to full QCD with dynamical quarks than either the leading order quenched or topological expansions. Non-planar diagrams and quark loops are retained, but diagrams in which the external sources are connected to different quark loops are still excluded (Fig. \[fig:OZI\]). This means that amplitudes which involve purely gluonic intermediate states are suppressed. This is the field-theoretic basis for the original empirical OZI rule.
![Feynman diagrams allowed (left) and forbidden (right) by the OZI rule.[]{data-label="fig:OZI"}](OZI.eps "fig:"){height="3.3cm"} ![Feynman diagrams allowed (left) and forbidden (right) by the OZI rule.[]{data-label="fig:OZI"}](OZIbreak.eps "fig:"){height="3cm"}
In each of these large-$N_c$ expansions, except the topological expansion where $n_f/N_c$ is fixed, the $U(1)_A$ anomaly does not contribute at leading order. More precisely, the anomalous contribution $\langle 0|T~Q~\phi_5^b|0\rangle$ in the chiral Ward identity (\[eq:bf\]) is suppressed by $O(1/N_c)$ relative to the current term $\langle 0|T~J_{\m 5}^0 ~\phi_5^b|0\rangle$. This means that the flavour singlet current is conserved, Goldstone’s theorem applies, and conventional PCAC methods can be used to understand the dynamics of the Green functions with a full set of $(n_f^2 - 1)$ massless bosons in the chiral limit. Taking this as a starting point, we can then learn about the spectral decomposition of the actual QCD Green functions as we relax from the leading-order limits. In particular, this leads us to the famous Witten-Veneziano mass formula for the $\eta'$ meson [@Witten:1979vv; @Veneziano:1979ec].
The behaviour of the topological susceptibility at large $N_c$ is central to this analysis. It is clear from looking at planar diagrams that at leading order in $1/N_c$, $\chi(0)$ in QCD coincides with the topological susceptibility $\chi(0)|_{\rm YM}$ in the corresponding pure Yang-Mills theory. Referring now to the explicit expression (\[eq:bv\]) for $\chi(0)$, large-$N_c$ counting rules give $A = O(1)$ while $\langle\bar q q\rangle = O(N_c)$. It follows that [*for non-zero quark masses*]{}, $$\chi(0) ~=~ - A ~+~ O(n_f/N_c)
\label{eq:baa}$$ where $A = \C_{QQ}^{-1}$ is identified as $-\chi(0)_{\rm YM} + O(1/N_c)$. On the other hand, if we consider the limit of $\chi(0)$ for $m_q\rta 0$ at finite $N_c$, then we have $$\chi(0)|_{m_q\rta 0} ~=~ 0
\label{eq:bbb}$$ The ’t Hooft large-$N_c$ limit is therefore not smooth in QCD; the $N_c \rta \infty$ and $m_q \rta 0$ limits do not commute [@Witten:1979vv; @Veneziano:1979ec; @DiVecchia:1980ve]. This is remedied in the topological expansion, where quark loops are retained and the $O(n_f/N_c)$ contribution in eq.(\[eq:baa\]) allows the smooth chiral limit $\chi(0) \rta 0$ even for large $N_c$.
‘$U(1)_A$ without instantons’
=============================
The $U(1)_A$ problem has a long history, pre-dating QCD itself, and has been an important stimulus to new theoretical ideas involving anomalies and gluon topology.
At its simplest, the original ‘$U(1)_A$ problem’ in current algebra is relatively straightforwardly resolved by the existence of the anomalous contributions to the chiral Ward identities (anomalous commutators in current algebra) and the consequent absence of a ninth light Nambu-Goldstone boson in $n_f=3$ QCD.[^5] However, a full resolution requires a much more detailed understanding of the dynamics of the pseudoscalar sector and the role of topological fluctuations in the anomalous Green functions.
In this section, we review the analysis of the $U(1)_A$ problem presented by Veneziano in his seminal 1974 paper, ‘$U(1)_A$ without instantons’ [@Veneziano:1979ec].[^6] As well as deriving the eponymous mass formula relating the $\eta'$ mass to the topological susceptibility, the essential problem resolved in ref.[@Veneziano:1979ec] is how to describe the dynamics of the Green functions of the pseudoscalar operators in QCD in terms of a spectral decomposition compatible with the $n_f$, $N_c$, $\theta$ and quark mass dependence imposed by the anomalous Ward identities.
First, recall that in the absence of the anomaly, there will be light pseudoscalar mesons $\eta^\a$ coupling derivatively to the currents with decay constants $f^{a\a}$, i.e. $\langle 0|J_{\m5}^a|\eta^\a\rangle = i p_\m f^{a\a}$. (We use the notation $\eta^\a$ to denote the physical mesons $\pi^0,\eta$ and $\eta'$, while the $SU(3)$ index $a = 3,8,0$.) The mass matrix $\m^2_{\a\b}$ satisfies the Dashen, Gell-Mann–Oakes–Renner (DGMOR) relation [@Gell-Mann:1968rz; @Dashen:1969eg] $$f^{a\a} \m^2_{\a\b} f^{T\b b} ~=~ - (M\Phi)_{ab} ~~~~~~~(\rm no~anomaly)
\label{eq:ccca}$$ The consequences of the anomaly are determined by the interaction of the pseudoscalar fields $\phi_5^a$ with the topological charge density $Q$ and the subsequent mixing. This gives rise to an additional contribution to the masses. Moreover, we can no longer identify the flavour singlet decay constant by the coupling to $J_{\m5}^0$ since this is not RG invariant. Instead, the physical decay constants $f^{a\a}$ are defined in terms of the couplings of the $\eta^\a$ to the pseudoscalar fields through the relation $f^{a\a}\la 0| \phi_5^b |\eta^\a\ra = d_{abc}\la\phi^c\ra$. This coincides with the usual definition except in the flavour singlet case.
The most transparent way to describe how all this works is to use an effective action $\C[Q,\phi_5^a]$ constructed to satisfy the anomalous chiral Ward identities. It is important to emphasise from the outset that this is an effective action in the sense of section \[sec 2.1\], i.e. the generating functional for vertices which are 1PI with respect to the set of fields $Q,\phi_5^a$ only. The choice of fields is designed to capture the degrees of freedom essential for the dynamics.[^7] A different choice (or linear combination) redefines the physical meaning of the vertices, so it is important that the final choice of fields in $\C$ results in vertices which are most directly related to physical couplings.
The simplest effective action consistent with the anomalous Ward identities and the renormalisation group is $$\begin{aligned}
&\C[Q,\phi_5^a] ~&=~ \int dx~\biggl({1\over2A}Q^2 ~+~
Q\bigl(\sqrt{2n_f}\d_{0a} - B_a \pl^2 \bigr)\Phi_{ab}^{-1}\phi_5^b
~~~~~~~~~~~~~~~~~~~~~~~
\nonumber\\
&{}&~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+~{1\over2} \phi_5^a \Phi^{-1}_{ac}\bigl((M\Phi)_{cd} - C_{cd}\pl^2 \bigr)
\Phi^{-1}_{db}\phi_5^b \biggr)
\label{eq:cca}\end{aligned}$$ The constants $C_{ab}$ and $B_a$ are related to $\C_{V_{\m5}^a V_{\n5}^b}$ and $\C_{V_{\m5}^a Q}$ respectively. The inclusion of the term with $B_a$ is unusual, but is required for consistency with the RGEs derived from (\[eq:bzz\]) beyond zero momentum.
0.1cm This form of $\C[Q,\phi_5^a]$ encodes three key dynamical assumptions: 0.1cm
$\bullet$ Pole dominance. We assume that the Green functions are dominated by the contribution of single-particle poles associated with the pseudoscalar mesons [*including*]{} the flavour singlet. This extends the usual PCAC assumption to the singlet sector.
$\bullet$ Smoothness. We assume that pole-free dynamical quantities such as the decay constants and couplings (1PI vertices) are only weakly momentum-dependent in the range from $p=0$ to their on-shell values. This allows us to impose relations derived from the zero-momentum Ward identities, provided this is compatible with the renormalisation group.
$\bullet$ Topology. There must exist topologically non-trivial fluctuations which can give a non-vanishing value to $\chi(0)|_{\rm YM}$ in pure gluodynamics. This is required to give the non-vanishing coefficient in the all-important ${1\over 2A}Q^2$ term in $\C[Q,\phi_5^a]$. Notice that we do not require a kinetic term for $Q$, which would be associated with a (presumed heavy) pseudoscalar glueball.
0.1cm The second derivatives of $\C[Q,\phi_5^a]$ are $$\begin{aligned}
&\left(\matrix{\C_{QQ} &\C_{Q\phi_5^b}\cr
&{}\cr
\C_{\phi_5^a Q} &\C_{\phi_5^a
\phi_5^b} \cr}\right) =
\left(\matrix{A^{-1}
&(\sqrt{2n_f}\d_{0d} + B_d p^2) \Phi_{db}^{-1} \cr
&{}\cr
\Phi_{ac}^{-1}(\sqrt{2n_f}\d_{c0} + B_c p^2)
& \Phi^{-1}_{ac}\bigl((M\Phi)_{cd} + C_{cd}p^2\bigr)\Phi^{-1}_{db}\cr}
\right)
\nonumber\\
&{}
\nonumber\\
\label{eq:ccb}\end{aligned}$$ The corresponding Green functions (composite operator propagators) are given by inversion: $$\left(\matrix{W_{\o\o} & W_{\o S_5^b}\cr
W_{S_5^a \o} & W_{S_5^a S_5^b} \cr}\right) ~~=~~ -
\left(\matrix{\C_{QQ} &\C_{Q\phi_5^b}\cr
\C_{\phi_5^a Q} &\C_{\phi_5^a \phi_5^b} \cr}\right)^{-1}
\label{eq:ccc}$$ and we find, to leading order in $p^2$, $$\begin{aligned}
&W_{\o\o} ~&=~ -A~ \tilde\D^{-1}
\nonumber\\
&W_{\o S_5^b} ~&=~ W_{S_5^b \o} ~\simeq~ \sqrt{2n_f} A \D_{0d}^{-1}\Phi_{db}
\nonumber\\
&W_{S_5^a S_5^b} ~&=~ - \Phi_{ac}~ \D_{cd}^{-1}~ \Phi_{db}
\label{eq:ccd}\end{aligned}$$ where $$\tilde \D ~=~ 1 - \Bigl(2n_f A \d_{a0}\d_{0b}
+ \sqrt{2n_f} A(\d_{a0}B_b + B_a \d_{0b})p^2\Bigr)
\bigr(M\Phi + C p^2\bigr)^{-1}_{ab}
\label{eq:cce}$$ and $$\D_{ab} ~=~ \Bigl(C_{ab}
- \sqrt{2n_f} A (\d_{a0}B_b + B_a \d_{0b})\Bigr)p^2
+ (M\Phi)_{ab} - 2n_f A~\d_{a0}\d_{0b}~~~
\label{eq:ccf}$$
In this form, however, the propagator matrix is not diagonal and the operators are not normalised so as to couple with unit decay constants to the physical states. It is therefore convenient to make a change of variables in $\C$ so that it is written in terms of operators which are more closely identified with the physical states. We do this is in two stages, since the intermediate stage allows us to make direct contact with the discussion in ref.[@Veneziano:1979ec] and will play an important role in some of the phenomenological applications considered later.
0.1cm First, we define rescaled fields $\hat\eta^\a$ whose kinetic terms, before mixing with $Q$, are canonically normalised. That is, we set $$\hat\eta^\a ~=~ \hat f^{T\a a} \Phi_{ab}^{-1} \phi_5^b
\label{eq:ccg}$$ with the ‘decay constants’ $\hat f^{a\a}$ defined such that ${d\over dp^2}\C_{\hat\eta^\a \hat\eta^\b}|_{p=0} = \d_{\a\b}$. This implies $$(\hat f \hat f^T)_{ab} ~=~ C_{ab} ~=~
{d\over dp^2} W_{S_D^a S_D^b}\Big|_{p=0}
\label{eq:cch}$$ where $D^a = \sqrt{2n_f}\d_{a0} Q + M_{ab}\phi_5^b$ is the divergence of the current $J_{\m5}^a$. In the chiral limit, this reduces in the flavour singlet sector to $$(\hat f \hat f^T)_{00} ~=~ {d\over dp^2} \chi(p^2)\Big|_{p=0} ~=~ \chi'(0)
\label{eq:cci}$$ a result which plays a vital role in understanding the ‘proton spin’ problem. Notice however that the $\hat f^{a\a}$ are [*not*]{} RG invariant: in fact, $\DD \hat f^{a\a} = \c \d_{a0} \hat f^{a\a}$. The effective action $\C[Q,\hat\eta^\a]$ is: $$\begin{aligned}
&\C[Q,\hat\eta^a] ~&=~ \int dx~\biggl({1\over2A}Q^2 ~+~
Q \bigl(\sqrt{2n_f} \d_{0a} - B_a \pl^2 \bigr)
(\hat f^{-1})^{a\a} \hat\eta^\a ~~~~~~~~~~~
\nonumber\\
&{}&~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+~{1\over2} \hat\eta^\a\bigl(-\pl^2
+ \hat f^{-1T}M\Phi \hat f^{-1} \bigr)_{\a\b}\hat\eta^\b \biggr)
\label{eq:ccj}\end{aligned}$$ In this form, the $\hat\eta^\a$ are the canonically normalised fields corresponding to the ‘would-be Nambu-Goldstone bosons’ in the absence of the anomaly, before they acquire an additional anomaly-induced mass. In the framework of the large-$N_c$ or OZI approximations, they would correspond to true Nambu-Goldstone bosons. The singlet $\hat\eta^0$ is what we have therefore referred to in our previous papers as the ‘OZI boson’ $\eta'_{OZI}$. As we see later, the naive current algebra relations hold when expressed in terms of the $\hat \eta^\a$ and $\hat f^{a\a}$, though these do [*not*]{} correspond to physical states or decay constants.
0.1cm The physical particle masses are identified with the poles in the two-point Green functions (\[eq:ccd\]). We immediately see that due to mixing with the topological charge density $Q$, the physical pseudoscalar meson mass $m^2_{\a\b}$ is shifted from its original value. From the pole in eq.(\[eq:ccf\]), we immediately find $$f^{a\a} m_{\a\b}^2 f^{T\b b} = - (M\Phi)_{ab}
+ 2n_f A \d_{a0} \d_{b0}
\label{eq:cck}$$ where we identify the physical, RG-invariant decay constants as $$(f f^T)_{ab} ~=~ (\hat f \hat f^T)_{ab}
- \sqrt{2n_f}A(\d_{a0} B_b + B_a \d_{0b})
\label{eq:cciii}$$ Eq.(\[eq:cck\]) is the key result. It generalises the original DGMOR relations (\[eq:ccca\]) to the flavour-singlet sector with the anomaly properly incorporated and the renormalisation group constraints satisfied. It represents a generalisation of the Witten-Veneziano mass formula which makes no direct reference to large-$N_c$ arguments but depends only on the three dynamical assumptions stated above [@Shore:1999tw].
With this clarification of the distinction between the physical decay constants $f^{a\a}$ and the RG non-invariant $\hat f^{a\a}$, we can rewrite eq.(\[eq:cce\]) for the topological susceptibility $\chi(p^2) = W_{\o\o}(p^2)$ as $$\chi(p^2) ~=~ -A \Bigl[1 -
{\rm tr}\bigl((\hat f \hat f^T - f f^T)p^2 + 2n_f A {\bf 1}_{00}\bigr)
\bigl(\hat f \hat f^T p^2 + M\Phi\bigr)^{-1} \Bigr]^{-1}
\label{eq:ccl}$$ It is clear that in the zero-momentum limit, this expression successfully reproduces eq.(\[eq:bv\]) for $\chi(0)$. For one flavour, the formula simplifies to $$\chi(p^2) ~=~ -A \bigl(\hat f \hat f^T p^2 + M\Phi\bigr)
\Bigl[f f^T p^2 + M\Phi + 2n_f A \Bigr]^{-1}~~~~~(n_f = 1)
\label{eq:ccm}$$ showing clearly the pole at the shifted mass $m^2$ of eq.(\[eq:cck\]). The occurrence of both $\hat f^{a\a}$ and $f^{a\a}$ in these expressions allows them to satisfy the RGE (\[eq:by\]) for the topological susceptibility, which requires $\DD \chi(p^2) = O(p^2)$.
0.1cm The second stage is to make a change of variable which diagonalises the propagator matrix, so as to give the most direct possible relation between the operators and the physical states. Choosing $$\begin{aligned}
&G ~&=~Q - W_{\o S_5^a} W_{S_5^a S_5^b}^{-1} \phi_5^b~~
~\simeq~ Q + \sqrt{2n_f} A \Phi_{0b}^{-1} \phi_5^b
\nonumber\\
&\eta^\a ~&=~ f^{T\a a} \Phi_{ab}^{-1} \phi_5^b
\label{eq:ccn}\end{aligned}$$ defines the effective action $\C[G,\eta^\a]$ as $$\C[G,\eta^\a] ~=~ \int dx~ \biggl({1\over 2A} G^2 ~+~
{1\over2} \eta^\a (-\pl^2 -m^2)_{\a\b} \eta^\b ~\biggr)
\label{eq:cco}$$ with $m^2_{\a\b}$ given by eq.(\[eq:cck\]). The corresponding propagators are $$\begin{aligned}
&\langle 0|T~G~G|0\rangle &= - A
\nonumber\\
&\langle 0|T~ \eta^\a ~\eta^\b|0\rangle &=
{-1\over p^2 - m_{\eta^\a}^2}\d^{\a\b}
\label{eq:ccp}\end{aligned}$$ where with no loss of generality we have taken $m^2_{\a\b}$ diagonal.
Notice also that the states mix in the complementary way to the operators. In particular, the mixing for the states corresponding to eq.(\[eq:ccn\]) for the fields $G$ and $\eta^\a$ is $$\begin{aligned}
&|G\rangle ~&=~ |Q\rangle
\nonumber\\
&|\eta^\a\rangle ~&=~ (f^{-1})^{\a a}\bigl( \Phi_{ab} |\phi_5^b\rangle -
\sqrt{2n_f} A \d_{a0}|Q\rangle \bigr)
\label{eq:ccq}\end{aligned}$$ In this sense, we see that we can regard the physical $\eta'$ (and, with $SU(3)$ breaking, the $\eta$) as an admixture of quark and gluon components, while the unphysical state $|G\rangle$ is purely gluonic.
An immediate corollary is the following relation, which we will use repeatedly in deriving alternative forms of the current algebra identities for the pseudoscalar mesons: $$\Phi_{ab}{\d\over\d\phi_5^b} ~=~ \hat f^{a\a}{\d\over\d\hat\eta^\a} ~=~
f^{a\a}{\d\over\d\eta^a} + \sqrt{2n_f} A \d_{a0} {\d\over\d G}
\label{eq:ccr}$$
0.1cm The formulation in terms of $\C[G,\eta^\a]$ is exactly what we need to construct a simple ‘$U(1)_A$ PCAC’ with which to interpret the low-energy phenomenology of the pseudoscalar mesons. We turn to this in the next section.
Here, we focus on the intermediate formulation $\C[Q,\hat\eta^\a]$ in order to describe Veneziano’s analysis of the $U(1)_A$ problem in the framework of the large-$N_c$ and topological expansions. The starting point is the anomalous Ward identity (\[eq:br\]) for the topological susceptibility: $$n_f^2 \int dx~\langle 0|T~Q(x)~Q(0)|0\rangle ~=~
\int dx~m^a m^b\langle 0|T~\phi_5^a(x)~\phi_5^b(0)|0\rangle
~+~ m^a \langle \phi^a\rangle
\label{eq:ccs}$$ The essential problem is how to understand this relation in terms of a spectral decomposition in the context of the $1/N_c$ expansion.
Assuming that $\chi(0)_{\rm YM} = -A + O(1/N_c)$ is non-vanishing at $O(1)$, the l.h.s. is $O(n_f^2)$ in leading order in $1/N_c$. On the other hand, the r.h.s. includes the condensate term of $O(n_f N_c m)$. To resolve this apparent paradox, we have to go beyond leading order in $1/N_c$ and consider the quark loop contributions which are included in the topological expansion. Although these are formally suppressed by powers of $(n_f/N_c)$, they contain light intermediate states which can enhance the order of the Green function. As we have seen above, these light states are just the ‘OZI bosons’ $|\hat\eta^\a\rangle$ with masses $\m^2_{\a\b}$ of $O(n_f m)$. Inserting these intermediate states, we therefore find that: $$\chi(p^2) ~=~
\chi(p^2)|_{\rm YM} -\langle 0|Q|\hat\eta^\a\rangle
{1\over (p^2 - \m^2)_{\a\b}}\langle \hat\eta^\b|Q|0\rangle ~+~ \ldots
\label{eq:cct}$$ where the coupling $\langle 0|Q|\hat\eta^\a\rangle$ is $O(\sqrt{n_f/N_c})$.
Approximating $\chi(p^2)_{\rm YM} \sim -A$ (a low-momentum smoothness assumption) and $\langle 0|Q|\hat\eta^\a\rangle \sim \sqrt{2n_f}A(f^{-1})^{0\a}$, then summing the series of intermediate state contributions, we find $$\chi(p^2) ~\simeq~
- {A\over 1 - 2n_f A \Bigl(f(p^2 - \m^2) f^{T}\Bigr)^{-1}_{00}}
\label{eq:cp}$$ This expression reproduces eq.(13) of ref.[@Veneziano:1979ec]. Clearly, it is dominated by the physical pseudoscalar pole with anomaly-induced mass given by eq.(\[eq:cck\]). It does not completely recover our more precise expression (\[eq:ccl\]) because of the approximation for the coupling of $Q$ to the $|\hat \eta^\a\rangle$, which misses the subtleties related to the introduction of $B_a$ in the effective action $\C[Q,\hat\eta^\a]$ and the distinction of $\hat f^{a\a}$ and $f^{a\a}$. These are effects of higher order in $1/N_c$ but, as we have seen, they are necessary to establish full RG consistency and will prove to be important for phenomenology.
To see how a term with the $O(n_f N_c m)$ dependence of the condensate can arise in $n_f^2 \chi(0)$, notice from eq.(\[eq:cck\]) that the physical pseudoscalar mass squared $m^2_{\eta^\a}$ has two contributions, the first of $O(m)$ from the conventional quark mass term and the new, anomaly-induced contribution of $O(n_f/N_c)$. If we are in a regime where the anomaly contribution dominates ($m < \L_{\rm QCD}/N_c$), then it follows that the above expression for $\chi(0)$ indeed becomes of $O(n_f^{-1}N_c m)$.
The original Witten-Veneziano mass formula for the $\eta'$ is the large-$N_c$ limit of eq.(\[eq:cck\]). In the chiral limit there is no flavour mixing and the singlet mass is given by $$m_{\eta'}^2 ~=~ {1\over (f^{0\eta'})^2} 2n_f A ~=~
-{2n_f\over f_{\pi}^2} \chi(0)_{\rm YM} + O((n_f/N_c)^2)
\label{eq:cq}$$ This formula provided the first link between the $\eta'$ mass and gluon topology. For an alternative recent derivation in the context of a $n_f/N_c$ expansion, see also ref.[@Giusti:2001xh].
What we learn from all this is that the Green functions in the anomalous chiral Ward identities admit a consistent spectral decomposition in terms of a full set of $(n_f^2 - 1)$ pseudoscalar mesons, provided they satisfy the generalised DGMOR mass formula (\[eq:cck\]) [*including*]{} the all-important anomaly term. The presence of these light poles can enhance the apparent order of the Green functions, as is familiar with Nambu-Goldstone bosons, and the anomaly-induced $O(n_f/N_c)$ contribution to $m_{\eta^\a}^2$ is critical in ensuring complete consistency with the Ward identities.
Similar considerations apply to the resolution of apparent paradoxes in the $\theta$-dependence of some Green functions. For example [@Veneziano:1979ec], we can show from the anomalous Ward identities that the condensate satisfies $$\sum_q m_q\langle\bar q q\rangle|_{\theta} ~\equiv~
m^a \langle \phi^a \rangle ~=~
\cos(\theta/n_f)~ m^a \langle \phi^a\rangle|_{\theta =0}
\label{eq:cr}$$ This implies $$\begin{aligned}
&{\pl^2\over\pl\theta^2}{} m^a \langle \phi^a\rangle|_{\theta =0} ~&=~
- m^a \int dx \int dy~ \langle 0|T~Q(x)~Q(y)~\phi^a(0)|0\rangle
\nonumber\\
&{}~&=~
-{1\over n_f^2}~ m^a \langle \phi^a \rangle|_{\theta =0}
\label{eq:cs}\end{aligned}$$ Here, the Green function is superficially of $O(n_f/N_c)$ while the r.h.s. is $O(N_c/n_f)$. The resolution is simply that it contains pseudoscalar intermediate states contributing two light poles with $m^2 \sim O(n_f/N_c)$. So once again we see how the spectral decomposition in terms of the full set of pseudoscalar mesons, including the flavour singlet, ensures consistency with the anomalous Ward identities.
Pseudoscalar mesons {#mesons}
===================
This theoretical analysis provides the basis for an extension of the conventional PCAC or chiral Lagrangian description of the phenomenology of the pseudoscalar mesons to the flavour singlet sector. In this section[^8] we describe the role of the $U(1)_A$ anomaly in the radiative decays of $\pi^0, \eta$ and $\eta'$ and derive the $U(1)_A$ Goldberger-Treiman relation, first proposed by Veneziano as a resolution of the ‘proton spin’ problem.
$U(1)_A$ Dashen, Gell-Mann–Oakes–Renner relations
-------------------------------------------------
The extension of the DGMOR relations to the $U(1)_A$ sector follows from the application of the three key dynamical assumptions used above (viz. pole dominance by the nonet of pseudoscalar mesons, smoothness of decay constants and couplings over the range from zero to on-shell momentum, and the existence of topologically non-trivial gluon dynamics) to the anomalous chiral Ward identities.
The fundamental $U(1)_A$ DGMOR relation $$f^{a\a} m_{\a\b}^2 f^{T\b b} = -M_{ac} \Phi_{cb}
+ 2n_f A \d_{a0} \d_{b0}
\label{eq:daaa}$$ has been derived above in the course of the general discussion of the $U(1)_A$ problem. In order to make this section self-contained, we give a brief and direct derivation here.
Recall that the physical meson fields are given as $\eta^\a = f^{T\a a} \Phi_{ab}^{-1} \phi_5^b$, with the decay constants defined so that the propagator $W_{S_\eta^\a S_\eta^\b} = -1/(p^2 - m_{\eta}^2)_{\a\b}$. It follows immediately that at zero momentum, $$f^{a\a} m_{\a\b}^2 f^{T\b b} = \Phi_{ac} (W_{S_5 S_5})_{cd}^{-1} \Phi_{db}
\label{eq:dbbb}$$ Using the chiral Ward identities of section 2 together with the identification (\[eq:bu\]) of the topological susceptibility, we can then show $$\begin{aligned}
&\Phi_{ac} (W_{S_5 S_5})_{cd}^{-1} \Phi_{db}
&= (\Phi M)_{ac} \bigl(M W_{S_5 S_5} M\bigr)_{cd}^{-1} (M\Phi)_{db}
\nonumber\\
&{}&= (M\Phi)_{ac} \Bigl(-(M\Phi) + 2n_f \chi(0) {\bf 1}_{00}
\Bigr)_{cd}^{-1} (M\Phi)_{db}
\nonumber\\
&{}&= -(M\Phi)_{ab} + 2n_f \C_{QQ}^{-1} ~\d_{a0} \d_{b0}
\label{eq:dccc}\end{aligned}$$ proving the result (\[eq:daaa\]).
Expanding this out, and assuming the mixed decay constants $f^{0\pi}, f^{8\pi}, f^{3\eta}, f^{3\eta'}$ are all negligible, we have $$\begin{aligned}
&(f^{0\eta'})^2 m_{\eta'}^2 + (f^{0\eta})^2 m_\eta^2 ~&=~
- {2\over3}\bigl(m_u \la\bar u u\ra + m_d \la\bar d d\ra
+ m_s \la\bar s s\ra \bigr) + 6 A ~~~~~~
\label{eq:da}\\
\nonumber\\
&f^{0\eta'} f^{8\eta'} m_\eta'^2 + f^{0\eta} f^{8\eta} m_{\eta}^2 ~&=~
- {\sqrt2\over3} \bigl(m_u \la\bar u u\ra + m_d \la\bar d d\ra
- 2 m_s \la\bar s s\ra \bigr)
\label{eq:db}\\
\nonumber\\
&(f^{8\eta'})^2 m_\eta'^2 + (f^{8\eta})^2 m_{\eta}^2 ~&=~
-{1\over3}\bigl(m_u \la\bar u u\ra + m_d \la\bar d d\ra +
4 m_s \la\bar s s\ra \bigr)
\label{eq:dc}\\
\nonumber\\
&~~~~~~~~~~~~~~~~~~~~~~~
f_\pi^2 m_\pi^2 ~&=~ - (m_u \la \bar u u \ra + m_d \la \bar d d \ra)
\label{eq:dd}\end{aligned}$$ and we can add the standard DGMOR relation for the $K^+$, $$f_K^2 m_K^2 ~=~ - (m_u \la \bar u u \ra + m_s \la \bar s s \ra)~~~~
\label{eq:de}$$
We emphasise that these formulae, as well as the radiative decay and $U(1)_A$ Goldberger-Treiman relations derived below, do not depend at all on the $1/N_c$ expansion. In particular, the constant $A$ appearing in the flavour singlet formula is defined as the non-perturbative parameter determining the topological susceptibility $\chi(0)$ in QCD according to the exact identity (\[eq:bv\]). As explained above, large-$N_c$ ideas do indeed provide a rationale for extending the familiar PCAC assumptions of pole dominance and smoothness to the flavour singlet channel, but these assumptions can be tested independently against experimental data.
The most useful form of these relations for phenomenology is to assume exact $SU(2)$ flavour symmetry and eliminate the quark masses and condensates in favour of $f_\pi, f_K, m_{\pi}^2$ and $m_K^2$ in the DGMOR relations for the $\eta$ and $\eta'$. This gives $$\begin{aligned}
&(f^{0\eta'})^2 m_{\eta'}^2 + (f^{0\eta})^2 m_\eta^2 ~&=~
{1\over3} \bigl(f_\pi^2 m_\pi^2 + 2 f_K^2 m_K^2\bigr) + 6 A
\label{eq:df}\\
\nonumber\\
&f^{0\eta'} f^{8\eta'} m_\eta'^2 + f^{0\eta} f^{8\eta} m_{\eta}^2 ~&=~
{2\sqrt2\over3}\bigl(f_\pi^2 m_\pi^2 - f_K^2 m_K^2\bigr)
\label{eq:dg}\\
\nonumber\\
&(f^{8\eta'})^2 m_\eta'^2 + (f^{8\eta})^2 m_{\eta}^2 ~&=~
-{1\over3}\bigl(f_\pi^2 m_\pi^2 - 4 f_K^2 m_K^2\bigr)
\label{eq:dh}\\
\nonumber\end{aligned}$$
We can also now clarify the precise relation of these results to the Witten-Veneziano formula for the mass of the $\eta'$ in its non-vanishing quark mass form, viz. $$m_{\eta'}^2 + m_{\eta}^2 - 2 m_K^2 ~=~ - {6\over f_\pi^2} \chi(0)|_{YM}
\label{eq:di}$$ Of course, only the $m_{\eta'}^2$ term on the l.h.s. is present in the chiral limit. Substituting in the explicit values for the masses in this formula gives a prediction [@Veneziano:1979ec] for the topological susceptibility, $\chi(0) \simeq -(180 ~{\rm MeV})^4$, which as we see below is remarkably close to the subsequently calculated lattice result.
If we now add the DGMOR formulae (\[eq:df\]) and (\[eq:dh\]), we find $$(f^{0\eta'})^2 m_{\eta'}^2 + (f^{0\eta})^2 m_\eta^2 +
(f^{8\eta})^2 m_\eta^2 + (f^{8\eta'})^2 m_{\eta'}^2 - 2 f_K^2 m_K^2
~=~ 6A
\label{eq:dj}$$ which we repeat is valid to all orders in $1/N_c$. To reduce this to its Witten-Veneziano approximation, we impose the large-$N_c$ limit to approximate the QCD topological charge parameter $A$ with $-\chi(0)|_{YM}$ as explained in section 2.4. We then set the ‘mixed’ decay constants $f^{0\eta}$ and $f^{8\eta'}$ to zero and all the remaining decay constants $f^{0\eta'}, f^{8\eta}$ and $f_K$ equal to $f_\pi$. With these approximations, we recover eq.(\[eq:di\]). Eventually, after we have found explicit experimental values for all these quantities, we will be able to demonstrate quantitatively how good an approximation the large-$N_c$ Witten-Veneziano formula is to the generalised $U(1)_A$ DGMOR relation in full QCD.
Radiative decay formulae for $\pi^0,\eta,\eta' \rta \c\c$ {#radiative}
---------------------------------------------------------
Radiative decays of the pseudoscalar mesons are of particular interest as they are controlled by the electromagnetic $U(1)_A$ anomaly, $$\pl^\m J_{\m 5}^a - M_{ab} \phi_5^b - \sqrt{2n_f} Q \d_{a0} - a_{\rm em}^a
{\a\over8\pi} F^{\m\n} \tilde F_{\m\n} ~=~ 0
\label{eq:dk}$$ where $F_{\m\n} = \pl_\m A_\n - \pl_\n A_\m$ is the usual electromagnetic field strength and the anomaly coefficients $a_{\rm em}^a$ are determined by the quark charges. The generating functional $\C[V_{\m 5}^a, V_\m^a, Q, \phi_5^a,
\phi^a, A_\m]$ of 1PI vertices including the photon satisfies the Ward identity $$\begin{aligned}
&\pl_\m \C_{V_{\m5}^a} &- \sqrt{2n_f} \d_{a0} Q
- a_{\rm em}^a {\a\over8\pi} F^{\m\n} \tilde F_{\m\n}
- d_{abc}m^b \phi_5^c
\nonumber\\
&{}&+ f_{abc} V_{\m}^b \C_{V_{\m5}^c}
+ f_{abc} V_{\m5}^b \C_{V_{\m}^c}
- d_{abc} \phi_5^c \C_{\phi^b}
+ d_{abc} \phi^c \C_{\phi_5^b} = 0
\label{eq:dl}\end{aligned}$$
To derive the radiative decay formulae, we first differentiate this identity with respect to the photon field $A_\m$. This gives $$ip_\m \C_{V_{\m 5}^a A^\l A^\r} ~+~\Phi_{ab} \C_{\phi_5^b A^\l A^\r}
~=~ - a_{\rm em}^a {\a\over\pi}~ \e_{\m\n\l\r} k_1^\m k_2^\n
\label{eq:dm}$$ where $k_1, k_2$ are the momenta of the two photons. Notice that the mass term does not contribute directly to this formula. From its definition as 1PI w.r.t. the pseudoscalar fields, the vertex $\C_{V_{\m 5}^a A^\l A^\r}$ does not have a pole at $p^2 = 0$, even in the massless limit, so we find simply $$\Phi_{ab} \C_{\phi_5^b A^\l A^\r}\bigl|_{p=0}
~=~ - a_{\rm em}^a {\a\over\pi}~ \e_{\m\n\l\r} k_1^\m k_2^\n
\label{eq:dn}$$
The radiative couplings $g_{\eta^\a \c\c}$ for the physical mesons $\eta^\a = \pi^0,\eta,\eta'$ are defined as usual from the decay amplitude $\langle \c\c|\eta^\a\rangle$. With the PCAC assumptions already discussed, they can be identified with the 1PI vertices as follows: $$\langle \c\c|\eta^\a\rangle ~=~ -i g_{\eta^\a \c\c}~\e_{\m\n\l\r}
k_1^\m k_2^\n \e^\l (k_1) \e^\r (k_2) ~=~
i \C_{\eta^\a A_\l A_\r} \e^\l (k_1) \e^\r (k_2)
\label{eq:do}$$ Re-expressing eq.(\[eq:dn\]) in terms of the canonically normalised ‘OZI bosons’ $\hat\eta^\a$, we therefore have the first form of the decay formula, $$\hat f^{a\a} g_{\hat\eta^\a \c\c} ~=~ a_{\rm em}^a {\a\over\pi}
\label{eq:dp}$$
Then, rewriting this in terms of the physical pseudoscalar couplings $g_{\eta^\a \c\c}$ and decay constants according to the relation (\[eq:ccr\]) gives the final form for the generalised $U(1)_A$ PCAC formula describing radiative pseudoscalar decays, incorporating both the electromagnetic and colour anomalies: $$f^{a\a} g_{\eta^\a \c\c} + \sqrt{2n_f} A g_{G\c\c} \d_{a0} ~=~
a_{\rm em}^a {\a\over\pi}
\label{eq:dr}$$ Expanding this formula, we have $$\begin{aligned}
f^{0\eta'} g_{\eta'\c\c} + f^{0\eta} g_{\eta\c\c} + {\sqrt6} A g_{G\c\c}
~=~ a_{\rm em}^0 {\a\over\pi}
\label{eq:ds}\\
\nonumber\\
f^{8\eta'}g_{\eta'\c\c} + f^{8\eta} g_{\eta\c\c}
~=~ a_{\rm em}^8 {\a\over\pi}~~~~~~~~~~~~~~~~
\label{eq:dt}\\
\nonumber\\
f_\pi g_{\pi\c\c} ~=~ a_{\rm em}^3 {\a\over\pi}
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\label{eq:du}\end{aligned}$$ where $a_{\rm em}^0 = {2\sqrt2\over3\sqrt3}N_c$, $a_{\rm em}^8 = {1\over3\sqrt3}N_c$ and $a_{\rm em}^3 = {1\over3}N_c$.
The new element in the flavour singlet decay formula is the gluonic coupling parameter $g_{G\c\c}$. It takes account of the fact that because of the anomaly-induced mixing with the gluon topological density $Q$, the physical $\eta'$ is not a true Nambu-Goldstone boson so the naive PCAC formulae must be modified. $g_{G\c\c}$ is [*not*]{} a physical coupling and must be regarded as an extra parameter to be fitted to data, although in view of the identifications in eq.(\[eq:ccq\]) it may reasonably be thought of as the coupling of the photons to the gluonic component of the $\eta'$.
The renormalisation group properties of these relations are readily derived from the RGE (\[eq:bzz\]) for $\C$. In the ‘OZI boson’ form, the unphysical coupling $g_{\hat\eta^\a \c\c}$ satisfies the complementary RGE to the decay constant $\hat f^{a\a}$ so the combination is RG invariant: $$\DD \hat f^{a\a} = \c \d_{a0} \hat f^{a\a} ~~~~~~~~~~~~~
\DD \bigl(\hat f^{a\a} g_{\hat\eta^\a \c\c}\bigr) = 0
\label{eq:duuu}$$ In contrast, [*all*]{} the decay constants and couplings in the relation (\[eq:dr\]) can be shown to be separately RG invariant, including the gluonic coupling $g_{G\c\c}$ [@Shore:1991pn; @Shore:2001cs].
The renormalisation group, OZI and $1/N_c$ : a conjecture {#sec:conj}
---------------------------------------------------------
Although these $U(1)_A$ PCAC relations have been derived purely on the basis of the pole dominance and smoothness assumptions, we will nevertheless find it useful in practical applications to exploit their OZI or large-$N_c$ behaviour, in conjunction with the renormalisation group.
The basic idea is that violations of the OZI rule, or equivalently anomalous large-$N_c$ behaviour, are generally related to the existence of the $U(1)_A$ anomaly. Moreover, we can identify the quantities which will be particularly sensitive to the anomaly as those which have RGEs involving the anomalous dimension $\c$. We therefore conjecture that the dependence of Green functions and 1PI vertices on $\c$ will be an important guide in identifying propagators and couplings which are likely to show violations of the OZI rule and those for which the OZI (or large-$N_c$) limit should be a good approximation [@Shore:1991dv; @Shore:1991pn].
As an example, the large-$N_c$ order of the quantities in the flavour singlet decay relation (\[eq:ds\]) is as follows: $f^{a\a} = O(\sqrt{N_c})$ for all the decay constants, $g_{\eta^\a\c\c} = O(\sqrt{N_c})$, $g_{G\c\c} = O(1)$, $a_{\rm em}^a = O(N_c)$ and the topological susceptibility parameter $A = O(1)$. The renormalisation group behaviour is especially simple, with both the meson and gluonic couplings $g_{\eta^\a \c\c}$ and $g_{G\c\c}$ as well as the decay constants being RG invariant. Putting this together, we find that all the terms in the decay formula are of $O(N_c)$ except the anomalous contribution $A g_{G\c\c}$ which is $O(1)$. Since it is RG invariant and independent of the anomalous dimension $\c$, we conjecture that it is a quantity for which the OZI (or large-$N_c$) approximation should be reliable so we expect it to be numerically small compared with the other contributions. In the next section, we test this against experiment.
As we shall see later, this conjecture has far-reaching implications for a range of predictions related to the anomaly, particularly in the interpretation of the $U(1)_A$ Goldberger-Treiman relation and associated ideas on the first moment sum rules for $g_1^p$ and $g_1^\c$ in deep-inelastic scattering.
Phenomenology
-------------
After all this theoretical development, we finally turn to experiment and use the data on the radiative decays $\eta,\eta' \rta \c\c$ to deduce values for the pseudoscalar meson decay constants $f^{0\eta'}$, $f^{0\eta}$, $f^{8\eta'}$ and $f^{8\eta}$ from the set of decay formulae (\[eq:ds\]), (\[eq:dt\]) and $U(1)_A$ DGMOR relations (\[eq:df\])-(\[eq:dh\]). We will also find the value of the unphysical coupling parameter $g_{G\c\c}$ and test the realisation of the $1/N_c$ expansion in real QCD.
The two-photon decay widths are given by $$\C\bigl(\eta'(\eta)\rta\c\c\bigr) ~=~ {m_{\eta'(\eta)}^3 \over64\pi~}
|g_{\eta'(\eta)\c\c}|^2
\label{eq:dv}$$ The current experimental data, quoted in the Particle Data Group tables [@PDG], are $$\C(\eta'\rta\c\c) ~=~ 4.28 \pm 0.19 ~{\rm KeV}
\label{eq:dw}$$ which is dominated by the 1998 L3 data [@L3] on the two-photon formation of the $\eta'$ in $e^+ e^- \rta e^+ e^- \pi^+ \pi^- \c$, and $$\C(\eta\rta\c\c) ~=~ 0.510 \pm 0.026 ~{\rm KeV}
\label{eq:dx}$$ which arises principally from the 1988 Crystal Ball [@Crystal] and 1990 ASP [@ASP] results on $e^+ e^- \rta e^+ e^- \eta$. From this data, we deduce the following results for the couplings $g_{\eta'\c\c}$ and $g_{\eta\c\c}$: $$g_{\eta'\c\c} ~=~ 0.031 \pm 0.001 ~{\rm GeV}^{-1}
\label{eq:dy}$$ and $$g_{\eta\c\c} ~=~ 0.025 \pm 0.001 ~{\rm GeV}^{-1}
\label{eq:dz}$$ which may be compared with $g_{\pi\c\c} = 0.024 \pm 0.001 ~{\rm GeV}$.
We also require the pseudoscalar meson masses: $$\begin{aligned}
m_{\eta'} ~=~ 957.78 \pm 0.14 ~{\rm MeV} ~~~~~~~~
m_{\eta} ~=~ 547.30 \pm 0.12 ~{\rm MeV} \nonumber\\
m_K ~=~ 493.68 \pm 0.02 ~{\rm MeV} ~~~~~~~~
m_\pi ~=~ 139.57 \pm 0.00 ~{\rm MeV}
\label{eq:daa}\end{aligned}$$ and the decay constants $f_\pi$ and $f_K$. These are defined in the standard way, so we take the following values (in our normalisations) from the PDG [@PDG]: $$f_K ~=~ 113.00 \pm 1.03 ~{\rm MeV} ~~~~~~~~
f_\pi ~=~ 92.42 \pm 0.26 ~{\rm MeV}
\label{eq:dbb}$$ giving $f_K/f_\pi = 1.223 \pm 0.012$.
The octet decay constants $f^{8\eta}$ and $f^{8\eta'}$ are obtained from eqs.(\[eq:dh\]) and (\[eq:dt\]). This leaves three remaining equations which determine the singlet decay constants $f^{0\eta'}, f^{0\eta}$ and the gluonic coupling $g_{G\c\c}$ in terms of the QCD topological susceptibility parameter $A$. This dependence is plotted in Figs. \[fig:feta\] and \[fig:twogamma\].
![The decay constants $f^{0\eta'}$ and $f^{0\eta}$ as functions of the non-perturbative parameter $A = (x~{\rm MeV})^4$ which determines the topological susceptibility in QCD.[]{data-label="fig:feta"}](Fetap.eps "fig:"){height="3.3cm"} 0.7cm ![The decay constants $f^{0\eta'}$ and $f^{0\eta}$ as functions of the non-perturbative parameter $A = (x~{\rm MeV})^4$ which determines the topological susceptibility in QCD.[]{data-label="fig:feta"}](Feta.eps "fig:"){height="3.3cm"}
![This shows the relative sizes of the contributions to the flavour singlet radiative decay formula (\[eq:ds\]) expressed as functions of the topological susceptibility parameter $A = (x~{\rm MeV})^4$. The dotted (black) line denotes ${2\sqrt2\over\sqrt3} {\a_{\rm em}\over\pi}$. The dominant contribution comes from the term $f^{0\eta'} g_{\eta'\c\c}$, denoted by the long-dashed (green) line, while the short-dashed (blue) line denotes $f^{0\eta} g_{\eta\c\c}$. The contribution from the gluonic coupling, ${\sqrt6} A g_{G\c\c}$, is shown by the solid (red) line.[]{data-label="fig:twogamma"}](Twogamma.eps){height="4.5cm"}
To make a definite prediction, we need a theoretical input value for the topological susceptibility. In time, lattice calculations in full QCD with dynamical fermions should be able to determine the parameter $A$. For the moment, however, only the topological susceptibility in pure Yang-Mills theory is known accurately. The most recent value [@DelDebbio:2004ns] is $$\chi(0)|_{YM} ~=~ -(191 \pm 5 ~{\rm MeV})^4 ~=~
- (1.33 \pm 0.14)\times 10^{-3} ~{\rm GeV}^4
\label{eq:dcc}$$ This supersedes the original value $\chi(0)|_{YM} \simeq -(180~{\rm MeV})^4$ obtained some time ago [@DiGiacomo:1990ij]. Similar estimates are also obtained using QCD spectral sum rule methods [@Narison:1990cz]. At this point, therefore, we have to make an approximation and so we assume that the $O(1/N_c)$ corrections in the identification $$A ~=~ \chi(0)_{\rm YM} + O(1/N_c)
\label{eq:ddd}$$ are numerically small. With this provisional input for $A$, we can then determine the full set of decay constants: $$\begin{aligned}
{}\nonumber\\
f^{0\eta'} ~=~ 104.2 \pm 4.0 ~{\rm MeV} ~~~~~~~~~~
f^{0\eta} ~=~ 22.8 \pm 5.7 ~{\rm MeV} \nonumber\\
f^{8\eta'} ~=~-36.1 \pm 1.2 ~{\rm MeV} ~~~~~~~~
f^{8\eta} ~=~ 98.4 \pm 1.4 ~{\rm MeV}
\label{eq:dee}\end{aligned}$$ and $$g_{G\c\c} ~=~ - 0.001 \pm 0.072 ~{\rm GeV}^{-4}
\label{eq:dff}$$ It is striking how close both the diagonal decay constants $f^{0\eta'}$ and $f^{8\eta}$ are to $f_{\pi}$. Predictably, the off-diagonal ones $f^{0\eta'}$ and $f^{8\eta'}$ are strongly suppressed.
It is also useful to quote these results in the two-angle parametrisation normally used in phenomenology. Defining, $$\left(\matrix{f^{0\eta'} &f^{0\eta}\cr f^{8\eta'} &f^{8\eta}}\right)
~~=~~\left(\matrix{f_0 \cos\theta_0 &-f_0 \sin\theta_0\cr
f_8 \sin\theta_8 &f_8 \cos\theta_8}\right)
\label{eq:dgg}$$ we find $$\begin{aligned}
f_0 ~=~ 106.6 \pm 4.2 ~{\rm MeV} ~~~~~~~~
f_8 ~=~ 104.8 \pm 1.3 ~{\rm MeV} \nonumber\\
\theta_0 ~=~ -12.3 \pm 3.0 ~{\rm deg} ~~~~~~~~
\theta_8 ~=~ -20.1 \pm 0.7 ~{\rm deg}
\label{eq:dhh}\end{aligned}$$ that is $${f_0\over f_\pi} ~=~ 1.15 \pm 0.05 ~~~~~~~~
{f_8\over f_\pi} ~=~ 1.13 \pm 0.02
\label{eq:dii}$$
Given these results, we can now investigate how closely our expectations based on OZI or $1/N_c$ reasoning are actually realised by the experimental data. With the input value (\[eq:dcc\]) for $A$, the numerical magnitudes and $1/N_c$ orders of the terms in the flavour singlet decay relation are as follows (see Fig. \[fig:twogamma\]): $$\begin{aligned}
f^{0\eta'} g_{\eta'\c\c} ~[N_c;~3.23]
+ f^{0\eta} g_{\eta\c\c} ~[N_c;~0.57]
+ {\sqrt6} A g_{G\c\c} ~[1;~-0.005 \pm 0.23]
\nonumber\\
=~~a_{\rm em}^0 {\a_{\rm em}\over\pi} ~[N_c;~3.79]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\label{eq:djj}\end{aligned}$$ The important point is that the gluonic contribution $g_{G\c\c}$, which is suppressed by a power of $1/N_c$ compared to the others, is also experimentally small. The near-vanishing for the chosen value of $A$ is presumably a coincidence, but we see from Fig. \[fig:twogamma\] that across a reasonable range of values of the topological susceptibility it is still contributing no more than around $10\%$, in line with our expectations for a RG-invariant, OZI-suppressed quantity.
It is also interesting to see how the $1/N_c$ approximation is realised in the $U(1)_A$ DGMOR generalisation (\[eq:dj\]) of the Witten-Veneziano formula (\[eq:di\]). Here we find $$\begin{aligned}
(f^{0\eta'})^2 m_{\eta'}^2 ~[N_c;~9.96]
~+~ (f^{0\eta})^2 m_\eta^2 ~[N_c;~0.15]
~+~(f^{8\eta'})^2 m_\eta'^2 ~[N_c;~1.19]
\nonumber\\
~+~ (f^{8\eta})^2 m_{\eta}^2 ~[N_c;~2.90]
- 2 f_K^2 m_K^2 [N_c;-6.22]
~=~ 6A ~[1;~7.98]
\nonumber\\
\label{eq:dkk}\end{aligned}$$ This confirms the picture that the anomaly-induced contribution of $O(1/N_c)$ to $m_{\eta'}^2$, which gives a sub-leading $O(1)$ effect in $(f^{0\eta'})^2 m_{\eta'}^2$, is in fact numerically dominant and matched by the $O(1)$ topological susceptibility term $6A$. Away from the chiral limit, the conventional non-anomalous terms are all of $O(N_c)$ and balance as expected. The surprising numerical accuracy of the Witten-Veneziano formula (\[eq:br\]) is seen to be in part due to a cancellation between the underestimates of $f^{8\eta'}$ (taken to be 0) and $f_K$ (set equal to $f_\pi$). This emphasises, however, that great care must be taken in using the formal order in the $1/N_c$ expansion as a guide to the numerical importance of a physical quantity, especially in the $U(1)_A$ channel.
Nevertheless, the fact that the RG-invariant, OZI-suppressed coupling $g_{G\c\c}$ is experimentally small is a very encouraging result. It increases our confidence that we are able to identify quantities where the OZI, or leading $1/N_c$, approximation is likely to be numerically good. It also shows that $g_{G\c\c}$ gives a contribution to the decay formula which is entirely consistent with its picturesque interpretation as the coupling of the photons to the anomaly-induced gluonic component of the $\eta'$. [*A posteriori*]{}, the fact that its contribution is at most $10\%$ explains the general success of previous theoretically inconsistent phenomenological parametrisations of $\eta'$ decays in which the naive current algebra formulae omitting the gluonic term are used.
However, while the flavour singlet decay formula is well-defined and theoretically consistent, it is necessarily non-predictive. To be genuinely useful, we would need to find another process in which the same coupling enters. The problem here is that, unlike the decay constants which are universal, the coupling $g_{G\c\c}$ is process-specific just like $g_{\eta'\c\c}$ or $g_{\eta\c\c}$. There are of course many other processes to which our methods may be applied such as $\eta'(\eta)\rta V\c$, where $V$ is a flavour singlet vector meson $\rho,\omega,\phi$, or $\eta'
(\eta)\rta \pi^+ \pi^- \c$. The required flavour singlet formulae may readily be written down, generalising the naive PCAC formulae. However, each will introduce its own gluonic coupling, such as $g_{GV\c}$. Although strict predictivity is lost, our experience with the two-photon decays suggests that these extra couplings will give relatively small, at most $O(10-20\%$), contributions if like $g_{G\c\c}$ they can be identified as RG invariant and $1/N_c$ suppressed. This observation restores at least a reasonable degree of predictivity to the use of PCAC methods in the $U(1)_A$ sector.
$U(1)_A$ Goldberger-Treiman relation
------------------------------------
A further classic application of PCAC is to the pseudoscalar couplings of the nucleon. For the pion, the relation between the axial-vector form factor of the nucleon and the pion-nucleon coupling $g_{\pi NN}$ is the famous Goldberger-Treiman relation. Here, we present its generalisation to the flavour singlet sector, which involves the anomaly and gluon topology. This $U(1)_A$ Goldberger-Treiman relation was first proposed by Veneziano [@Veneziano:1989ei] in an investigation of the ‘proton-spin’ problem and further developed in refs.[@Shore:1990zu; @Shore:1991dv; @Narison:1998aq; @Shore:2006mm].
The axial-vector form factors are defined from $$\langle N|J_{\mu 5}^a|N\rangle ~=~
2m_N \Bigl( G_A^a(p^2) s_\mu + G_P^a(p^2) p.s p_\mu \Bigr)
\label{eq:dll}$$ where $s_\mu = \bar u \c_\mu \c_5 u /2m_N$ is the covariant spin vector. In the absence of a massless pseudoscalar, only the form factors $G_A^a(0)$ contribute at zero momentum.
Expressing the matrix element in terms of the 1PI vertices derived from the generating functional $\C[V_{\m5}^a, V_\m^a, Q, \phi_5^a,\phi^a]$, including spectator fields $N, \bar N$ for the nucleon, we have $$\langle N|J_{\mu 5}^a|N\rangle ~=~ \bar u\Bigl(
\C_{V_5^{\m a} \bar N N} + W_{V_5^{\m a} \o} \C_{Q \bar N N}
+ W_{V_5^{\m a} S_5^b} \C_{\phi_5^b \bar N N}\Bigr) u
\label{eq:dmmm}$$ Note that this expansion relies on the specific definition(\[eq:bh\]) of $\C$ as a partial Legendre transform.
We also need the following relation, valid for all momenta, which is derived directly from the fundamental anomalous chiral Ward identity (\[eq:bi\]) for $\C$: $$\pl_\m \C_{V_{\m5}^a \bar N N} ~=~ - \Phi_{ab} \C_{\phi_5^b \bar N N}
\label{eq:dnnn}$$ Now, taking the divergence of eq.(\[eq:dmmm\]), using this Ward identity and then[^9] taking the zero-momentum limit, noting that the propagators vanish at zero momentum since there is no massless pseudoscalar, gives $$2m_N G_A^a(0)~ \bar u \c_5 u ~=~
i \bar u ~ \Phi_{ab} \C_{\phi_5^b \bar N N}\big|_{p=0} u
\label{eq:dooo}$$
The meson-nucleon couplings are related to the 1PI vertices by $$\langle N|\eta^\a N\rangle ~=~ g_{\eta^\a NN}~ \bar u \c_5 u
~=~ i \bar u \C_{\eta^\a \bar N N} u
\label{eq:dppp}$$ Re-expressing eq.(\[eq:dooo\]) in terms of the canonically normalised ‘OZI boson’ field $\hat \eta^\a$, we therefore derive $$2m_N G_A^a(0) ~=~ \hat f^{a\a} g_{\hat\eta^\a NN}
\label{eq:dqqq}$$ This relation will be useful to us when we consider the ‘proton spin’ problem.
All that now remains to cast this into its final form is to make the familiar change of variables from $Q,\hat\eta^\a$ to $G,\eta^\a$, where $\eta^\a$ are interpreted as the physical mesons. We therefore find the generalised $U(1)_A$ Goldberger-Treiman relation: $$2m_N G_A^a(0) ~=~ f^{a\a} g_{\eta^\a NN} ~+~ \sqrt{2n_f} A g_{GNN}\d_{a0}
\label{eq:dmm}$$ For the individual components, this is $$\begin{aligned}
&2m_N G_A^3 ~&=~ f_\pi g_{\pi NN}
\label{eq:dnn} \\
&2m_N G_A^8 ~&=~ f^{8\eta'} g_{\eta' NN} ~+~ f^{8\eta} g_{\eta NN}
\label{eq:doo} \\
&2m_N G_A^0 ~&=~ f^{0\eta'} g_{\eta' NN} ~+~ f^{0\eta} g_{\eta NN}
~+~ \sqrt{6} A g_{GNN}~~~~~~
\label{eq:dpp} \end{aligned}$$
The renormalisation group properties of these relations are described in great detail in ref.[@Shore:1991dv]. It is clear that the flavour singlet axial coupling $G_A^0$ satisfies a homogeneous RGE and scales with the anomalous dimension $\c$ corresponding to the multiplicative renormalisation of $J_{\m5}^0$. In the form (\[eq:dqqq\]), RG consistency is simply achieved by $$\DD \hat f^{a\a} ~=~ \c \d_{a0} \hat f^{a\a}~~~~~~~~~~~~~~
\DD g_{\hat\eta^\a NN} ~=~ 0
\label{eq:drrr}$$ All the scale dependence is in the decay constant $\hat f^{0\a}$ while the the coupling $g_{\hat\eta^\a NN}$ of the ‘OZI boson’ to the nucleon is RG invariant (in contrast to $g_{\hat\eta^\a \c\c}$). In the final form (\[eq:dmm\]) involving the physical decay constants, a careful analysis shows that apart from $G_A^0(0)$ the only other non RG-invariant quantity is the gluonic coupling $g_{GNN}$, which is required to satisfy the following non-homogeneous RGE to ensure the self-consistency of eq.(\[eq:dpp\]): $$\DD g_{GNN} ~=~ \c \Bigl(g_{GNN} + {1\over\sqrt{2n_f}}{1\over A}
f^{0\a} g_{\eta^\a NN}\Bigr)
\label{eq:dsss}$$
The large-$N_c$ behaviour in the flavour singlet relation is as follows: $G_A^0 = O(N_c)$, $f^{0\eta}, f^{0\eta'} = O(\sqrt{N_c})$, $A = O(1)$, $g_{\eta NN}, g_{\eta' NN} = O(\sqrt{N_c})$, $g_{GNN} = O(1)$. So the final term $A g_{GNN}$ is $O(1)$, suppressed by a power of $1/N_c$ compared to all the others, which are $O(N_c)$.
We see that, like $g_{G\c\c}$, the gluonic coupling $g_{GNN}$ is suppressed at large $N_c$ relative to the corresponding meson couplings. However, unlike $g_{G\c\c}$ which is RG invariant, $g_{GNN}$ has a complicated RG non-invariance and depends on the anomaly-induced anomalous dimension $\c$. The conjecture in section \[sec:conj\] then suggests that while the OZI or large-$N_c$ approximation should be a good guide to the value of $g_{G\c\c}$, we may expect significant OZI violations for $g_{GNN}$. We would therefore not be surprised to find that $g_{GNN}$ makes a sizeable numerical contribution to the $U(1)_A$ Goldberger-Treiman relation.
We now try to test these expectations against the experimental data. We first introduce a notation that has become standard in the literature on deep-inelastic scattering. There, the axial couplings are written as $$G_A^3 ~=~ {1\over2}~ a^3 ~~~~~~~~
G_A^8 ~=~ {1\over 2\sqrt{3}}~ a^8 ~~~~~~~~
G_A^0 ~=~ {1\over\sqrt{6}}~ a^0
\label{eq:dqq}$$ where the $a^a$ have a simple interpretation in terms of parton distribution functions.
Experimentally, $$a^3 ~=~ 1.267 \pm 0.004 ~~~~~~~~
a^8 ~=~ 0.585 \pm 0.025
\label{eq:dss}$$ from low-energy data on nucleon and hyperon beta decay. The latest result[^10] for $a^0$ quoted by the COMPASS collaboration [@COMPASS] from deep-inelastic scattering data is $$a^0|_{Q^2 \rta\infty} ~=~ 0.33 \pm 0.06
\label{eq:dtt}$$ with a similar result from HERMES [@HERMES].
The OZI expectation is that $a^0 = a^8$. In the context of DIS, this is a prediction of the simple quark model, where it is known as the Ellis-Jaffe sum rule [@Ellis:1973kp]). We return to this in the context of the ‘proton spin’ problem in section 5 but for now we concentrate on the low-energy phenomenology of the pseudoscalar meson-nucleon couplings.
The original Goldberger-Treiman relation (\[eq:dnn\]) gives the following value for the pion-nucleon coupling, $$g_{\pi NN} ~=~ 12.86 \pm 0.06
\label{eq:duu}$$ consistent to within about $5\%$ with the experimental value $13.65 (13.80)
\pm 0.12$ (depending on the dataset used [@Bugg:2004cm]). In an ideal world where $g_{\eta NN}$ and $g_{\eta' NN}$ were both known, we would now verify the octet formula (\[eq:doo\]) then determine the gluonic coupling $g_{GNN}$ from the singlet Goldberger-Treiman relation (\[eq:dpp\]). However, the experimental situation with the $\eta$ and $\eta'$-nucleon couplings is far less clear. One would hope to determine these couplings from the near threshold production of the $\eta$ and $\eta'$ in nucleon-nucleon collisions, i.e. $pp\rightarrow pp\eta$ and $pp\rightarrow pp\eta'$, measured for example at COSY-II [@Moskal:2004cm; @Bass:2001ix; @Moskal:2004nw]. However, the $\eta$ production is dominated by the $N(1535) S_{11}$ nucleon resonance which decays to $N\eta$, and as a result very little is known about $g_{\eta NN}$ itself. The detailed production mechanism of the $\eta'$ is not well understood. However, since there is no known baryonic resonance decaying into $N\eta'$, we may simply assume that the reaction $pp\rightarrow pp\eta'$ is driven by the direct coupling supplemented by heavy-meson exchange. This allows an upper bound to be placed on $g_{\eta' NN}$ and on this basis ref.[@Moskal:1998pc] quotes $g_{\eta' NN}< 2.5$. This is supported by an analysis [@Nakayama:2005ts] of very recent data from CLAS [@Dugger:2005du] on the photoproduction reaction $\c p \rightarrow p \eta'$. Describing the cross-section data with a model comprising the direct coupling together with $t$-channel meson exchange and $s$ and $u$-channel resonances, it is found that equally good fits can be obtained for several values of $g_{\eta' NN}$ covering the whole region $0 < g_{\eta' NN} < 2.5$.
In view of this experimental uncertainty, we shall use the octet and singlet Goldberger-Treiman relations to plot the predictions for $g_{\eta NN}$ and $g_{GNN}$ as a function of the ill-determined $\eta'$-nucleon coupling in the experimentally allowed range $0 < g_{\eta' NN} < 2.5$. The results (again taking the value (\[eq:dcc\]) for $A$) are given in Fig. \[fig:NNcouplings\]. In Fig. \[fig:GTformula\] we have shown the relative magnitudes of the various contributions to the flavour-singlet formula.
![These figures show the dimensionless $\eta$-nucleon coupling $g_{\eta NN}$ and the gluonic coupling $g_{GNN}$ in units of ${\rm GeV}^{-3}$ expressed as functions of the experimentally uncertain $\eta'$-nucleon coupling $g_{\eta' NN}$, as determined from the flavour octet and singlet Goldberger-Treiman relations (\[eq:doo\]) and (\[eq:dpp\]).[]{data-label="fig:NNcouplings"}](getaNNrev.eps "fig:"){height="3.3cm"} 0.7cm ![These figures show the dimensionless $\eta$-nucleon coupling $g_{\eta NN}$ and the gluonic coupling $g_{GNN}$ in units of ${\rm GeV}^{-3}$ expressed as functions of the experimentally uncertain $\eta'$-nucleon coupling $g_{\eta' NN}$, as determined from the flavour octet and singlet Goldberger-Treiman relations (\[eq:doo\]) and (\[eq:dpp\]).[]{data-label="fig:NNcouplings"}](gGNNrev.eps "fig:"){height="3.3cm"}
![This shows the relative sizes of the contributions to the $U(1)_A$ Goldberger-Treiman relation from the individual terms in eq.(\[eq:dpp\]), expressed as functions of the coupling $g_{\eta' NN}$. The dotted (black) line denotes $2m_N G_A^0$. The long-dashed (green) line is $f^{0\eta'}g_{\eta' NN}$ and the short-dashed (blue) line is $f^{0\eta}g_{\eta NN}$. The solid (red) line shows the contribution of the novel gluonic coupling, $\sqrt{6}Ag_{GNN}$, where $A$ determines the QCD topological susceptibility.[]{data-label="fig:GTformula"}](UA1GTrev.eps){height="4.5cm"}
What we learn from this is that for values of $g_{\eta' NN}$ approaching the upper end of the experimentally allowed range, the contribution of the OZI-suppressed gluonic coupling $g_{GNN}$ is quite large. The variation of $f^{0\eta'} g_{\eta' NN}$ over the allowed range is compensated almost entirely by the variation of $\sqrt{6} g_{GNN}$, with the $f^{0\eta} g_{\eta NN}$ contribution remaining relatively constant.
For example, if experimentally we found $g_{\eta' NN} \simeq 2.5$, which corresponds to the cross-sections for $pp\rta pp\eta'$ and $\c p \rta p\eta'$ being almost entirely determined by the direct coupling, then we would have $g_{\eta NN} \simeq 4.14$ and $g_{GNN} \simeq -31.2~
{\rm GeV}^{-3}$. In terms of the contributions to the $U(1)_A$ Goldberger-Treiman relation, this would give (in GeV) $$\begin{aligned}
2m_N G_A^0 [N_c;~0.25] ~=~
f^{0\eta'} g_{\eta' NN} [N_c;~0.26] ~+~
f^{0\eta} g_{\eta NN} [N_c;~0.09] ~~~~~~~
\nonumber\\
~+~ \sqrt{6} A g_{GNN} [O(1);~-0.10]~~~~
\label{eq:dvv}\end{aligned}$$ The anomalously small value of $G_A^0$ compared to its OZI value (the OZI approximation is $2m_N G_A^0\big|_{\rm OZI}
= \sqrt{2}~2m_N G_A^8 = 0.45$) is then due to the partial cancellation of the sum of the meson-nucleon coupling terms by the gluonic coupling $g_{GNN}$. Although formally $O(1/N_c)$ suppressed, numerically it gives a major contribution to the large OZI violation in $G_A^0$. This would give some support to our conjecture and provide further evidence that we are able to predict the location of large OZI violations using the renormalisation group as a guide.
Of course, it may be that experimentally we eventually find a value for $g_{\eta' NN} \simeq 1.5$, in the region where $g_{GNN}$ contributes only around $10\%$ or less. Although surprising, this would open the possibility that all gluonic couplings of type $g_{GXX}$ are close to zero, which could be interpreted as implying that the gluonic component of the $\eta'$ wave function is simply small. Clearly, a reliable determination of $g_{\eta' NN}$, or equivalently $g_{\eta NN}$, would shed considerable light on the $U(1)_A$ dynamics of QCD.
Topological charge screening and the ‘proton spin’
==================================================
So far, we have focused on the implications of the $U(1)_A$ anomaly for low-energy QCD phenomenology. However, the anomaly also plays a vital role in the interpretation of high-energy processes, in particular polarised deep-inelastic scattering.
In this section, we discuss one of the most intensively studied topics in QCD of the last two decades - the famous, but misleadingly named, ‘proton spin’ problem. We review the interpretation initially proposed by Veneziano [@Veneziano:1989ei] and developed by us in a series of papers exploring the relation with the $U(1)_A$ GT relation and gluon topology [@Shore:1990zu; @Shore:1991dv; @Shore:1994zh]. In subsequent work with Narison, we were able to quantify our prediction by using QCD spectral sum rules to compute the slope $\chi'(0)$ of the topological susceptibility [@Narison:1994hv; @Narison:1998aq]. Remarkably, the most recent experimental data from the COMPASS [@COMPASS] and HERMES [@HERMES] collaborations, released in September 2006, now confirms our original 1994 numerical prediction [@Narison:1994hv].
The $g_1^p$ and angular momentum sum rules
------------------------------------------
The ‘proton spin’ problem concerns the sum rule for the first moment of the polarised proton structure function $g_1^p$. This is measured in polarised DIS experiments through the inclusive processes $\m p \rta \m X$ (EMC, SMC, COMPASS at CERN) or $e p \rta e X$ (SLAC, HERMES at DESY) together with similar experiments on a deuteron target. The polarisation asymmetry of the cross-section is expressed as $$x{d\D\s\over dx dy}~=~{Y_P\over2}{16\pi^2\a^2\over s} g_1^p(x,Q^2)~~+~~
O\Bigl({M^2 x^2\over Q^2}\Bigr)
\label{eq:ea}$$ with conventional notation: $Q^2 = -q^2$ and $x = {Q^2/2p_2.q}$ are the Bjorken variables, where $p_2$, $q$ are the momenta of the target proton and incident virtual photon respectively, $y = {Q^2/xs}$ and $Y_p = {(2-y)/y}$.
According to standard theory, $g_1^p$ is determined by the proton matrix element of two electromagnetic currents carrying a large spacelike momentum. The sum rule for the first moment of $g_1^p$ is derived from the twist 2, spin 1 terms in the operator product expansion for the currents: $$J^\l(q) J^\r(-q) \SIMQ
2 \e^{\l\r\n\m} {q_\n\over Q^2} ~\Bigl[\D C_1^{NS}(\a_s)
\Bigl(J_{\m5}^3 + {1\over\sqrt3} J_{\m5 }^8\Bigr)
+ {2\sqrt{2}\over\sqrt{3}} \D C_1^{S}(\a_s) J_{\m5}^0 \Bigr]
\label{eq:eb}$$ where $\D C_1^{NS}$ and $\D C_1^S$ are Wilson coefficients and $J_{\m5}^a$ ($a=3,8,0$) are the renormalised axial currents, with the normalisations defined in section 2. It is the occurrence of the axial currents in this OPE that provides the link between the $U(1)_A$ anomaly and polarised DIS. The sum rule is therefore: $$\C^p_1(Q^2) \equiv
\int_0^1 dx~ g_1^p(x,Q^2)
= {1\over12} \D C_1^{NS} \Bigl( a^3
+ {1\over3} a^8 \Bigr) + {1\over9} \D C_1^{\rm S} a^0(Q^2)
\label{eq:ec}$$ where the axial charges $a^3$, $a^8$ and $a^0(Q^2)$ are defined in terms of the forward proton matrix elements as in eq.(\[eq:dqq\]). Here, we have explicitly shown the $Q^2$ scale dependence associated with the RG non-invariance of $a^0(Q^2)$.
Since the flavour non-singlet axial charges are known from low-energy data, a measurement of the first moment of $g_1^p$ amounts to a determination of the flavour singlet $a^0(Q^2)$. At the time of the original EMC experiment in 1988 [@EMC] the theoretical expectation based on the quark model was that $a^0 = a^8$. The resulting sum rule for $g_1^p$ is known as the Ellis-Jaffe sum rule [@Ellis:1973kp]. The great surprise of the EMC measurement was the discovery that in fact $a^0$ is significantly suppressed relative to $a^8$, and indeed the earliest results suggested it could even be zero. However, the reason the result sent shockwaves through both the theoretical and experimental communities (to date, the EMC paper has over 1300 citations) was the interpretation that this implies that the quarks contribute only a fraction of the total spin of the proton.
In fact, this interpretation relies on the simple valence quark model of the proton and is [*not*]{} true in QCD, where the axial charge decouples from the real angular momentum sum rule for the proton. Rather, as we shall show, the suppression of $a^0(Q^2)$ reflects the dynamics of gluon topology and appears to be largely independent of the structure of the proton itself. Precisely, it is a manifestation of [*topological charge screening*]{} in the QCD vacuum.
0.2cm The angular momentum sum rule is derived by taking the forward matrix element of the conserved angular momentum current $M^{\m\n\l}$, defined in terms of the energy-momentum tensor as $$M^{\m\n\l} ~=~ x^{[\n}T^{\l]\m} + \pl_\r X^{\r\m\n\l}
\label{eq:eea}$$ The inclusion of the arbitrary tensor $X^{\r\m\n\l}$ just reflects the usual freedom in QFT of defining conserved currents. This gives us some flexibility in attempting to write $M^{\m\n\l}$ as a sum of local operators, suggesting interpretations of the total angular momentum as a sum of ‘components’ of the proton spin. In fact, however, it is not possible to write $M^{\m\n\l}$ as a sum of operators corresponding to quark and gluon spin and angular momentum in a gauge-invariant way. The best decomposition is [@Jaffe:1989jz; @Shore:1999be; @Shore:2000ca] $$M^{\m\n\l} ~=~ O_1^{\m\n\l} + O_2^{\m[\l}x^{\n]} + O_3^{\m[\l}x^{\n]}
+ \ldots
\label{eq:eeb}$$ where the dots denote terms whose forward matrix elements vanish. Here, $$\begin{aligned}
&O_1^{\m\n\l} &= {1\over2}\e^{\m\n\l\s}\bar q \c_\s \c_5 q =
{1\over2}\e^{\m\n\l\s}\sqrt{2n_f}J_{\s 5}^0
\nonumber\\
&O_2^{\m\l} &= i \bar q \c^\m \hD{}^\l q
\nonumber\\
&O_3^{\m\l} &= F^{\m\r}F_\r{}^\l
\label{eq:eec}\end{aligned}$$ At first sight, $O_1^{\m\n\l}$ looks as if it could be associated with ‘quark spin’, since for [*free*]{} Dirac fermions the spin operator coincides with the axial vector current. $O_2^{\m[\l}x^{\m]}$ would correspond to ‘quark orbital angular momentum’, leaving $O_3^{\m[\l}x^{\n]}$ as ‘gluon total angular momentum’. Any further decomposition of the gluon angular momentum is necessarily not gauge invariant.
The forward matrix elements of these operators may be expressed in terms of form factors and, as we showed in ref.[@Shore:1999be], this exhibits an illuminating cancellation. After some analysis, we find: $$\begin{aligned}
&\langle p,s|O_1^{\m\n\l}|p,s\rangle ~&=~ a^0 m_N \e^{\m\n\l\s}s_\s
\nonumber\\
&\langle p,s|O_2^{\m[\l}x^{\n]}|p,s\rangle ~&=~
J_q {1\over2m_N}p_\r p^{\{\m} \e^{[\l\}\n]\r\s} s_\s -
a^0 m_N \e^{\m\n\l\s}s_\s
\nonumber\\
&\langle p,s|O_3^{\m[\l}x^{\n]}|p,s\rangle ~&=~
J_g {1\over2m_N}p_\r p^{\{\m} \e^{[\l\}\n]\r\s} s_\s
\label{eq:eed}\end{aligned}$$ The angular momentum sum rule for the proton is then just $${1\over2}~=~ J_q + J_g
\label{eq:eee}$$ where the Lorentz and gauge-invariant form factors $J_q$ and $J_g$ may reasonably be thought of as representing quark and gluon total angular momentum. However, even this interpretation is not at all rigorous, not least because $J_q$ and $J_g$ mix under renormalisation and scale as $${d\over d\ln Q^2} \left(\matrix{J_q\cr J_g\cr}\right) ~=~
{\a_s\over4\pi} \left(\matrix{-{8\over3}C_F & {2\over3}n_f\cr
{8\over3}C_F & -{2\over3}n_f\cr}\right)~
\left(\matrix{J_q\cr J_g\cr}\right)
\label{eq:eef}$$ Only the total angular momentum is Lorentz, gauge and scale invariant.[^11]
The crucial observation, however, is that the axial charge $a^0$ explicitly [*cancels*]{} from the angular momentum sum rule. $a^0$ is an important form factor, which relates the first moment of $g_1^p$ to gluon topology via the $U(1)_A$ anomaly, but it is [*not*]{} part of the angular momentum sum rule for the proton.
Just as $a^0$ can be measured in polarised inclusive DIS, the form factors $J_q$ and $J_g$ can be extracted from measurements of unpolarised generalised parton distributions (GPDs) in processes such as deeply-virtual Compton scattering $\c^* p \rta \c p$. These can also in principle be calculated in lattice QCD. The required identifications with GPDs are given in ref.[@Shore:1999be].
QCD parton model
----------------
Before describing our resolution of the ‘proton spin’ problem, we briefly review the parton model interpretation of the first moment sum rule for $g_1^p$.
In the simplest form of the parton model, the proton structure at large $Q^2$ is described by parton distributions corresponding to free valence quarks only. The polarised structure function is given by $$g_1^p(x) = {1\over2} \sum_{i=1}^{n_f} ~e_i^2 ~\D q_i(x)
\label{eq:ed}$$ where $\D q_i(x)$ is defined as the difference of the distributions of quarks (and antiquarks) with helicities parallel and antiparallel to the nucleon spin. It is convenient to work with the conventionally-defined flavour non-singlet and singlet combinations $\D q^{NS}$ and $\D q^S$ (often also written as $\D \Sigma$).
In this model, the first moment of the singlet quark distribution $\D q^S = \int_0^1 dx~\D q^S(x)$ can be identified as the sum of the helicities of the quarks. Interpreting the structure function data [*in this model*]{} then leads to the conclusion that the quarks carry only a small fraction of the spin of the proton. There is indeed a real contradiction between the experimental data and the [*free valence quark*]{} parton model.
However, this simple model leaves out many important features of QCD, the most important being gluons, RG scale dependence and the chiral $U_A(1)$ anomaly. When these effects are included, in the QCD parton model, the naive identification of $\D q^S$ with spin no longer holds and the experimental results for $g_1^p$ can be accommodated, though not predicted.
In the QCD parton model, the polarised structure function is written in terms of both quark and gluon distributions as follows: $$\begin{aligned}
g_1^p(x,Q^2) = ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
~~~~~~~~~~~~~~~~~~~~~~~~~~~\nonumber\\
\int_{x}^{1} {du\over u}~ {1\over9}
\Bigl[\D C^{NS}\Bigl({x\over u}\Bigr)\D q^{NS}(u,t)
+ \D C^{S}\Bigl({x\over u}\Bigr)\D q^S(u,t) +
\D C^{g}\Bigl({x\over u}\Bigr)\D g(u,t) \Bigr]
\nonumber\\
{}
\label{eq:ehhh}\end{aligned}$$ where $\D C^S, \D C^g$ and $\D C^{NS}$ are perturbatively calculable functions related to the Wilson coefficients and the quark and gluon distributions have [*a priori*]{} a $t=\ln Q^2/\L^2$ dependence determined by the RG evolution, or DGLAP, equations. The first moment sum rule is therefore $$\C_1^p(Q^2) = {1\over9}\Bigl[\D C_1^{NS}\D q^{NS}
+ \D C_1^S \D q^S + \D C_1^g \D g \Bigr]
\label{eq:eg}$$ Comparing with eq.(\[eq:ec\]), we see that the axial charge $a^0(Q^2)$ is identified with a linear combination of the first moments of the singlet quark and gluon distributions. It is often, though not always, the case that the moments of parton distributions can be identified in one-to-one correspondence with the matrix elements of local operators. The polarised first moments are special in that two parton distributions correspond to the same local operator.
The RG evolution equations for the first moments of the parton distributions are derived from the matrix of anomalous dimensions for the lowest spin, twist 2 operators. This introduces an inevitable renormalisation scheme ambiguity in the definitions of $\D q$ and $\D g$, and their physical interpretation is correspondingly nuanced. The choice closest to our own analysis is the ‘AB’ scheme [@Ball:1995td] where the parton distributions have the folowing RG evolution: $$\begin{aligned}
&{d\over d\ln Q^2} \D q^{NS} = 0 ~~~~~~~~~~
{d\over d\ln Q^2} \D q^S = 0
\nonumber\\
&{} \nonumber\\
&{d\over d\ln Q^2} {\a_s\over2\pi}\D g(Q^2) =
\c \Bigl({\a_s\over2\pi}\D g(Q^2)
-{1\over 3} \D q^S\Bigr)
\label{eq:eh}\end{aligned}$$ which requires $\D C_1^g = {3\a_s\over2\p}\D C_1^S$. It is then possible to make the following identifications with the axial charges: $$\begin{aligned}
&a^3 &= \D u - \D d
\nonumber\\
%{}&{}\nonumber\\
&a^8 &= \D u + \D d - 2 \D s
\nonumber\\
%{}&{}\nonumber\\
&a^0(Q^2) &= \D u + \D d + \D s - {3\a_s\over2\pi} \D g(Q^2)
\label{eq:ei}\end{aligned}$$ with $\D q^S = \D u + \D d + \D s$. Notice that in the AB scheme, all the scale dependence of the axial charge $a^0(Q^2)$ is assigned to the gluon distribution $\D g(Q^2)$.
This was the identification originally introduced for the first moments by Altarelli and Ross [@Altarelli:1988nr], and resolves the ‘proton spin’ problem in the context of the QCD parton model. In this scheme, the Ellis-Jaffe sum rule follows from the assumption that in the proton both $\D s$ and $\D g(Q^2)$ are zero, which is the natural assumption in the free valence quark model. This is equivalent to the OZI approximation $a^0(Q^2) = a^8$. However, in the full QCD parton model, there is no reason why $\D g(Q^2)$, or even $\D s$, should be zero in the proton. Indeed, given the different scale dependence of $a^0(Q^2)$ and $a^8$, it would be unnatural to expect this to hold in QCD itself.
An interesting conjecture [@Altarelli:1988nr] is that the observed suppression in $a^0(Q^2)$ is due overwhelmingly to the gluon distribution $\D g(Q^2)$ alone. Although by no means a necessary consequence of QCD, this is a reasonable expectation given that it is the anomaly (which is due to the gluons and is responsible for OZI violations) which is responsible for the scale dependence in $a^0(Q^2)$ and $\D g(Q^2)$ whereas the $\D q$ are scale invariant. This would be in the same spirit as our conjecture on OZI violations in low-energy phenomenology in section \[sec:conj\]. To test this, however, we need to find a way to measure $\D g(Q^2)$ itself, rather than the combination $a^0(Q^2)$. The most direct option is to extract $\D g(x,Q^2)$ from processes such as open charm production, $\c^* g \rta c \bar c$, which is currently being intensively studied by the COMPASS [@Procureur:2006sg], STAR [@STAR] and PHENIX [@PHENIX] collaborations.
Topological charge screening {#topscreen}
----------------------------
We now describe a less conventional approach to deep-inelastic scattering based entirely on field-theoretic concepts. In particular, the role of parton distributions is taken over by the 1PI vertices of composite operators introduced above. (For a review, see ref.[@Shore:1998dn]).
Once again, the starting point is the use of the OPE to express the moments of a generic structure function $F(x,Q^2)$ as $$\int_0^1 dx~x^{n-1} F(x,Q^2) ~=~ \sum_A C_A^n(Q^2)~
\langle p|\OO_A^n(0)|p\rangle
\label{eq:ej}$$ where $\OO_A^n$ denotes the set of lowest twist, spin $n$ operators and $C_A^n(Q^2)$ are the corresponding Wilson coefficients. The next step is to introduce a new set of composite operators $\tilde \OO_B$, chosen to encompass the physically relevant degrees of freedom, and write the matrix element as a product of two-point Green functions and 1PI vertices as follows: $$\int_0^1 dx~x^{n-1} F(x,Q^2) ~=~ \sum_A \sum_B C_A^n(Q^2)~
\langle 0|T~\OO_A^n~\tilde \OO_B|0\rangle~\C_{\tilde \OO_B pp}
\label{eq:ek}$$ This decomposition splits the structure function into three parts – first, the Wilson coefficients $C_A^n(Q^2)$ which can be calculated in perturbative QCD; second, non-perturbative but [*target independent*]{} Green functions which encode the dynamics of the QCD vacuum; third, non-perturbative vertex functions which characterise the target by its couplings to the chosen operators $\tilde \OO_B$.[^12]
Now specialise to the first moment sum rule for $g_1^p$. For simplicity, we first present the analysis for the chiral limit, where there is no flavour mixing. Using the anomaly (\[eq:bd\]), we can express the flavour singlet contribution to the sum rule as $$\C_1^p(Q^2)_{singlet} ~\equiv~
\int_0^1 dx~g_1^p(x,Q^2)_{singlet} ~=~
{2\over3} {1\over 2m_N}~ \D C_1^S(\a_s)~ \langle p|Q|p\rangle
\label{eq:el}$$ The obvious choice for the operators $\tilde\OO_B$ in this case are the flavour singlet pseudoscalars and it is natural to choose the ‘OZI boson’ field $\hat\eta^0 = \hat f^{00}{1\over \langle\bar q q\rangle}\phi_5^0$, which is normalised so that $d/dp^2~ \C_{\hat\eta^0 \hat\eta^0}\big|_{p=0} = 1$. As we have seen in eq.(\[eq:drrr\]), the corresponding 1PI vertex is then RG invariant. Writing the 1PI vertices in terms of nucleon couplings as in eq.(\[eq:dppp\]), we find $$\C_1^p(Q^2)_{singlet} ~=~
{2\over3} {1\over 2m_N} \D C_1^S(\a_s)~
\Bigl(\langle 0|T~Q~Q|0\rangle~g_{QNN} ~+~
\langle 0|T~Q~\hat\eta^0|0\rangle~g_{\hat\eta^0 NN}\Bigr)
\label{eq:elll}$$
![Illustration of the decomposition of the matrix element $\langle p |Q|p\rangle$ into two-point Green functions and 1PI vertices. The Green function in the first diagram is $\chi(0)$; in the second it is $\sqrt{\chi'(0)}$.[]{data-label="fig:semifignew"}](semifigQ.eps){height="2.8cm"}
Recalling that the matrix of two-point Green functions is given by the inversion formula $$\left(\matrix{W_{\o\o} & W_{\o S_{\hat\eta^0}}\cr
W_{S_{\hat\eta^0} \o} & W_{S_{\hat\eta^0} S_{\hat\eta^0}} \cr}\right)
~=~ -
\left(\matrix{\C_{QQ} &\C_{Q \hat\eta^0}\cr
\C_{\hat\eta^0 Q} &\C_{\hat\eta^0 \hat\eta^0} \cr}\right)^{-1}
\label{eq:em}$$ and using the normalisation condition for $\hat\eta^0$, we can easily show that at zero momentum, $$W_{\o S_{\hat\eta^0}}^2 ~=~ {d\over dp^2} W_{\o\o}\big|_{p=0}
\label{eq:en}$$
Finally, therefore, we can represent the first moment of $g_1^p$ in the following, physically intuitive form: $$\C_1^p(Q^2)_{singlet} ~=~
{2\over3} {1\over 2m_N}~ \D C_1^S(\a_s)~
\Bigl(\chi(0)~g_{QNN} ~+~
\sqrt{\chi'(0)}~g_{\hat\eta^0 NN}\Bigr)
\label{eq:eo}$$ This shows that the first moment is determined by the gluon topological susceptibility in the QCD vacuum as well as the couplings of the proton to the pseudoscalar operators $Q$ and $\hat\eta^0$. In the chiral limit, $\chi(0) = 0$ so the first term vanishes. The entire flavour singlet contribution is therefore simply $$\C_1^p(Q^2)_{singlet} ~=~
{2\over3} {1\over 2m_N}~ \D C_1^S(\a_s)~
\sqrt{\chi'(0)}~g_{\hat\eta^0 NN}
\label{eq:eooo}$$ The 1PI vertex $g_{\hat\eta^0 NN}$ is RG invariant, and we see from eq.(\[eq:by\]) that [*in the chiral limit*]{} the slope of the topological susceptibility scales with the anomalous dimension $\c$, viz. $${d\over d\ln Q^2} \sqrt{\chi'(0)} ~=~ \c ~\sqrt{\chi'(0)}
\label{eq:ep}$$ ensuring consistency with the RGE for the flavour singlet axial charge.
The formulae (\[eq:eo\]) and (\[eq:eooo\]) are our key result. They show how the first moment of $g_1^p$ can be factorised into couplings $g_{QNN}$ and $g_{\hat\eta^0 NN}$ which carry information on the proton structure, and Green functions which characterise the QCD vacuum. In the case of $g_1^p$, the Green functions reduce simply to the topological susceptibility $\chi(0)$ and its slope $\chi'(0)$. We now argue that the experimentally observed suppression in the first moment of $g_1^p$ is due [*not*]{} to a suppression in the couplings, but to the vanishing of the topological susceptibility $\chi(0)$ and an anomalously small value for its slope $\chi'(0)$. This is what we refer to as [*topological charge screening*]{} in the QCD vacuum.
The justification follows our now familiar conjecture on the relation between OZI violations and RG scale dependence. We expect the source of OZI violations to be in those quantities which are sensitive to the anomaly, as identified by their scaling dependence on the anomalous dimension $\c$, in this case $\chi'(0)$. In contrast, it should be a good approximation to use the OZI value for the RG-invariant vertex $g_{\hat\eta^0 NN}$, that is $g_{\hat\eta^0 NN} \simeq \sqrt{2}
g_{\hat\eta^8 NN}$. The corresponding OZI value for $\sqrt{\chi'(0)}$ would be $f_\pi/\sqrt{6}$. This gives our key formula for the flavour singlet axial charge: $${a^0(Q^2)\over a^8} ~\simeq~ {\sqrt{6}\over f_\pi}~\sqrt{\chi'(0)}
\label{eq:eq}$$
The corresponding prediction for the first moment of $g_1^p$ is $$\C_1^p(Q^2)_{singlet} ~=~ {1\over 9} \D C_1^S(\a_s)~a^8~
{\sqrt{6}\over f_\pi}~\sqrt{\chi'(0)}
\label{eq:er}$$
The final step is to compute the slope of the topological susceptibility. In time, lattice gauge theory should provide an accurate measurement of $\chi'(0)$. However, this is a particlarly difficult correlator for lattice methods since it requires a simulation of QCD with light dynamical fermions and algorithms that implement topologically non-trivial configurations in a sufficiently fast and stable way. Instead, we have estimated the value of $\chi'(0)$ using the QCD spectral sum rule method. Full details and discussion of this computation can be found in refs.[@Narison:1994hv; @Narison:1998aq]. The result is: $$\sqrt{\chi'(0)} ~=~ 26.4 \pm 4.1 ~{\rm MeV}
\label{eq:es}$$
This gives our final prediction for the flavour singlet axial charge and the complete first moment of $g_1^p$: $$\begin{aligned}
&a^0\big|_{Q^2=10 {\rm GeV}^2} ~&=~ 0.33 \pm 0.05
\label{eq:et}\\
&\C_1^p\big|_{Q^2=10 {\rm GeV}^2} ~&=~ 0.144 \pm 0.009
\label{eq:eu}\end{aligned}$$ Topological charge screening therefore gives a suppression factor of approximately 0.56 in $a^0$ compared to its OZI value $a^8 = 0.585$.
0.2cm In the decade since we made this prediction, the experimental measurement has been somewhat lower than this value, in the range $a^0 \simeq 0.20-0.25$. This would have suggested there is also a significant OZI violation in the nucleon coupling $g_{\hat\eta^0 NN}$ itself, implicating the proton structure in the anomalous suppression of $\C_1^p$. Very recently, however, the COMPASS and HERMES collaborations have published new results on the deuteron structure function which spectacularly confirm our picture that topological charge screening in the QCD vacuum is the dominant suppression mechanism.
![COMPASS and SMC data for the deuteron structure function $g_1^d(x)$. Statistical error bars are shown with the data points. The shaded band shows the systematic error.[]{data-label="fig:compassg1d"}](compassg1d.eps){height="5.0cm"}
This new data is shown in Fig. \[fig:compassg1d\]. This is based on data collected by COMPASS at CERN in the years 2002-2004 and has only recently been published. The accuracy compared to earlier SMC data at small $x$ is significantly improved and the dip in $x g_1^d$ around $x\sim 10^{-2}$ suggested by the SMC data is no longer present (Fig. \[fig:compassg1d\]). This explains the significantly higher value for $a^0$ found by COMPASS compared to SMC. From this data, COMPASS quote the first moment for the proton-neutron average $g_1^N =
(g_1^p + g_1^n)/2$ as [@COMPASS] $$\C_1^N\big|_{Q^2=3{\rm GeV}^2} ~=~ 0.050 \pm 0.003(stat) \pm 0.002(evol)
\pm 0.005(syst)
\label{eq:ev}$$ Extracting the flavour singlet axial charge from the analogue of eq.(\[eq:ec\]) for $\C_1^N$ then gives $$a^0\big|_{Q^2=3{\rm GeV}^2} ~=~ 0.35 \pm 0.03(stat) \pm 0.05(syst)
\label{eq:ew}$$ or evolving to the $Q^2\rta\infty$ limit, $$a^0\big|_{Q^2\rta\infty} ~=~ 0.33 \pm 0.03(stat) \pm 0.05(syst)
\label{eq:ex}$$ Similar results are found by HERMES, who quote [@HERMES] $$a^0\big|_{Q^2=5{\rm GeV}^2} ~=~ 0.330 \pm 0.011(th) \pm 0.025(exp)
\pm 0.028(evol)
\label{eq:ey}$$ The agreement with our prediction (\[eq:et\]) is striking.
0.2cm To close this section, we briefly comment on the extension of our analysis beyond the chiral limit. In this case, the operator $\sqrt{2n_f}Q$ in eq.(\[eq:el\]) is replaced by the full divergence of the flavour singlet axial current, viz. $D^0 = \sqrt{2n_f}Q + d_{0bc}m^b \phi_5^c$. Separating the matrix element $\langle p|D^0|p\rangle$ into Green functions and 1PI vertices, we find from the zero-momentum Ward identities that $\langle 0|T~D^0~Q|0\rangle = 0$ so the contribution from $g_{QNN}$ still vanishes. The other Green function is $\langle 0|T~D^0~\hat\eta^\a|0\rangle
= -\hat f^{0\a}$, so the first moment sum rule becomes $$\C_1^p(Q^2)_{singlet} ~=~
{1\over9} {1\over 2m_N}~ \D C_1^S(\a_s)~\sqrt{6}~
\hat f^{0\a}g_{\hat\eta^\a NN}
\label{eq:eaa}$$ It is clear that this is simply an alternative derivation of the $U(1)$ GT relation (\[eq:dqqq\]) for $a^0$. We could equally use the alternative form (\[eq:dmm\]) to write $$\C_1^p(Q^2)_{singlet} ~=~
{1\over9} {1\over 2m_N}~ \D C_1^S(\a_s)~\sqrt{6}~
\Bigl(f^{0\a}g_{\eta^\a NN} + \sqrt{6}A g_{GNN}\Bigr)
\label{eq:ebb}$$ Recalling the RGE (\[eq:dsss\]) for $g_{GNN}$, we see that this bears a remarkable similarity to the expression for $a^0$ in terms of parton distributions in the AB scheme, eq.(\[eq:ei\]). This was first pointed out in ref.[@Shore:1990zu; @Shore:1991dv].
Manipulating the zero-momentum Ward identities in a similar way to that explained above in the chiral limit now shows that we can express the decay constants $\hat f^{a\a}$ in terms of vacuum Green functions as follows (see eq.(\[eq:cch\]): $$(\hat f \hat f^T)_{ab} ~=~
{d\over dp^2} \langle 0|T~D^a ~D^b|0\rangle\big|_{p=0}
\label{eq:ecc}$$ However, for non-zero quark masses there is flavour mixing amongst the ‘OZI bosons’ $\hat\eta^\a$ and we cannot extract the decay constants simply by taking a square root, as was the case in writing $\hat f^{00} = \sqrt{\chi'(0)}$ in the chiral limit. Nevertheless, in ref.[@Narison:1998aq] we estimated the decay constants and form factors in the approximation where we use eq.(\[eq:ecc\]) with the full divergence $D^a$ but neglect flavour mixing. Assuming OZI for the couplings, this gives the estimate $${a^0(Q^2)\over a^8} ~\simeq~ \sqrt{6}~ {\hat f^{00} \over \hat f^{88}}
\label{eq:edd}$$ where we take $$\hat f^{00} \simeq
\sqrt{{{}^{}\over{}^{}}}{d\over dp^2} \langle 0|T~D^0 ~D^0|0\rangle\big|_{p=0}
~~~~~~~~~~~~~~
\hat f^{88} \simeq
\sqrt{{{}^{}\over{}^{}}}{d\over dp^2} \langle 0|T~D^8 ~D^8|0\rangle\big|_{p=0}
\label{eq:eeeee}$$ Evaluating the Green functions using QCD spectral sum rules gives $$\begin{aligned}
&a^0\big|_{Q^2=10 {\rm GeV}^2} ~&=~ 0.31 \pm 0.02
\label{eq:eff}\\
&\C_1^p\big|_{Q^2=10 {\rm GeV}^2} ~&=~ 0.141 \pm 0.005
\label{eq:egg}\end{aligned}$$ As we have seen in the last section, flavour mixing can be non-negligible in the phenomenology of the pseudoscalar mesons, so we should be a little cautious in over-estimating the accuracy of these estimates. (The quoted errors do not include this systematic effect.) Nevertheless, the fact that they are consistent with those obtained in the chiral limit reinforces our confidence that the flavour singlet axial charge is relatively insensitive to the quark masses and that eqs.(\[eq:et\]) and (\[eq:eu\]) indeed provide an accurate estimate of the first moment of $g_1^p$.
The observation that the ‘proton spin’ sum rule could be explained in terms of an extension of the Goldberger-Treiman relation to the flavour singlet sector was made in Veneziano’s original paper [@Veneziano:1989ei]. This pointed out for the first time that the suppression in $a^0$ was an OZI-breaking effect. Since the Goldberger-Treiman relation connects the pseudovector form factors with the pseudoscalar channel, where it is known that there are large OZI violations for the flavour singlet, it becomes natural to expect similar large OZI violations also in $a^0$. This is the fundamental intuition which we have developed into a quantitative resolution of the ‘proton spin’ problem.
Semi-inclusive polarised DIS
----------------------------
While the agreement between our prediction for the first moment of $g_1^p$ and experiment is now impressive, it would still be interesting to find other experimental tests of topological charge screening. A key consequence of this mechanism is that the OZI violation observed in $a^0$ is not a property specifically of the proton, but is [*target independent*]{}. This leads us to look for ways to make measurements of the polarised structure functions of other hadronic targets besides the proton and neutron. We now show how this can effectively be done by studying semi-inclusive DIS $eN \rta ehX$ in the target fragmentation region (Fig. \[fig:semiincl\]).
![Semi-inclusive DIS $eN\rta ehX$ in the target fragmentation region. In the equivalent current fragmentation process, the detected hadron $h$ is emitted from the hard collision with $\c$. The right-hand figure shows a simple Reggeon exchange model valid for $z\sim 1$, where $h$ carries a large target energy fraction. []{data-label="fig:semiincl"}](semiincl.eps){height="4.1cm"}
The differential cross-section in the target fragmentation region can be written analogously to eq.(\[eq:ea\]) in terms of fracture functions: $$x{d \D \s^{target}\over dx dy dz dt} = {Y_P\over2}{4\pi\a^2\over s}
\D M_1^{hN}(x,z, t, Q^2)
\label{eq:ehh}$$ where $x = Q^2/2p_2.q$, $x_{\BB} = Q^2/2k.q$, $z=p_2^{\prime}.q/p_2.q$ so that $1-z = x/x_{\BB}$, and the invariant momentum transfer $t = K^2 = -k^2$, where $k$ is the momentum of the struck parton. For $K^2 \ll Q^2$, $z \simeq
E_h/E_N$ (in the photon-nucleon CM frame) is the energy fraction of the target nucleon carried by the detected hadron $h$.
$\D M_1^{hN}$ is the fracture function [@Trentadue:1993ka] equivalent of the inclusive structure function $g_1^N$, so in the same way as in eq.(\[eq:ed\]) we have $$\D M_1^{hN}(x,z, t, Q^2) = {1\over2}\sum_i e_i^2
\D M_i^{hN}(x,z, t, Q^2)
\label{eq:eii}$$ Here, $\D M_i^{hN}(x,z, t, Q^2)$ is an extended fracture function, introduced by Grazzini, Trentadue and Veneziano [@Grazzini:1997ih], which carries an explicit dependence on $t$. One of the advantages of these fracture functions is that they satisfy a simple, homogeneous RG evolution equation analogous to the usual inclusive parton distributions.
Our proposal [@Shore:1997tq; @deFlorian:1997th] (see also [@Shore:1997ck]) is to study semi-inclusive DIS in the kinematical region where the detected hadron $h$ ($\pi$, $K$ or $D$) carries a large target energy fraction, i.e. $z$ approaching 1, with a small invariant momentum transfer $t$. In this region, it is useful to think of the target fragmentation process as being simply modelled by a single Reggeon exchange (see Fig. \[fig:semiincl\]), i.e. $$\D M_1^{hN}(x,z, t, Q^2)\big|_{z\sim 1} ~\simeq~ F(t) (1-z)^{-2\a_{\BB}(t)}
g_1^{\BB}(x_{\BB},t,Q^2)
\label{eq:ejj}$$ If we consider ratios of cross-sections, the dynamical Reggeon emission factor $F(t)(1-z)^{-2\a_{\BB}(t)}$ will cancel and we will be able to isolate the ratios of $g_1^{\BB}(x_{\BB},t,Q^2)$ for different effective targets $\BB$. Although single Reggeon exchange is of course only an approximation to the more fundamental QCD description in terms of fracture functions (see ref.[@Grazzini:1999vz] for a more technical discussion), it shows particularly clearly how observing semi-inclusive processes at large $z$ with particular choices of $h$ and $N$ amounts in effect to performing inclusive DIS on virtual hadronic targets $\BB$. Since our predictions will depend only on the $SU(3)$ properties of $\BB$, together with target independence, they will hold equally well when $\BB$ is interpreted as a Reggeon rather than a pure hadron state.
The idea is therefore to make predictions for the ratios $\RR$ of the first moments of the polarised fracture functions $\int_0^{1-z} dx \D M_1^{hN}(x,z, t, Q^2)$ or equivalently $\int_0^1 dx_{\BB} g_1^{\BB}(x_{\BB},t, Q^2)$ for various reactions. The first moments $\C_1^{\cal B}$ are calculated as in eq.(\[eq:ec\]) in terms of the axial charges $a^3$, $a^8$ and $a^0(Q^2)$ for a state with the $SU(3)$ quantum numbers of ${\cal B}$. We then use topological charge screening to say that $a^0(Q^2) \simeq s(Q^2) a^0\big|_{\rm OZI}$, i.e. the flavour singlet axial charge is suppressed relative to its OZI value by a universal, target-independent, suppression factor $s(Q^2)$. From our calculation of $\sqrt{\chi'(0)}$ and the experimental results for $g_1^p$, we have $s|_{Q^2=10{\rm GeV}^2} \simeq 0.33/0.585 = 0.56$.
Some of the more interesting predictions obtained in ref.[@Shore:1997tq] are as follows. The ratio $${\cal R} \biggl({en\rta e\pi^+ X\over ep\rta e\pi^- X}\biggr)_{z\sim 1}
~\simeq~~ {2s-1\over2s+2}
\label{eq:ekk}$$ is calculated by comparing $\C_1$ for the $\D^-$ and $\D^{++}$. It is particularly striking because the physical value of $s(Q^2)$ is close to one half, so the ratio becomes very small. For strange mesons, on the other hand, the ratio depends on whether the exchanged object is in the $\bf 8$ (where the reduced matrix elements involve the appropriate $F/D$ ratio) or $\bf 10$ representation, so the prediction is less conclusive, viz. $${\cal R}\biggl({en\rta eK^+ X\over ep\rta eK^0 X}\biggr)_{z\sim 1} ~\simeq~~
{2s-1-3(2s-1)F/D \over 2s-1-3(2s+1)F/D}~~~({\bf 8}) ~~~~~~{\rm or}~~~~~~
{2s-1\over 2s+1}~~~({\bf 10})
\label{eq:ell}$$ which we find by comparing $\C_1$ for either the $\Sigma^-$ and $\Sigma^+$ in the ${\bf 8}$ representation or $\Sigma^{*-}$ and $\Sigma^{*+}$ in the ${\bf 10}$. For charmed mesons, we again find $${\cal R} \biggl({en\rta eD^0 X\over ep\rta eD^- X}\biggr)_{z\sim 1} ~\simeq~~
{2s-1\over2s+2}
\label{eq:emm}$$ corresponding to the ratio for $\Sigma_c^0$ to $\Sigma_c^{++}$.
At the other extreme, for $z$ approaching 0, the detected hadron carries only a small fraction of the target nucleon energy. In this limit, the ratio $\RR$ of the fracture function moments becomes simply the ratio of the structure function moments for $n$ and $p$, i.e. using the current experimental values, $\RR_{z\sim 0} ~\simeq~ \C_1^n / \C_1^p = -0.30$. This is to be compared with the corresponding OZI or Ellis-Jaffe value of $-0.12$.
The differences between the OZI, or valence quark model, expectations and our predictions based on topological charge screening can therefore be quite dramatic and should give a very clear experimental signal. In ref.[@deFlorian:1997th], together with De Florian, we analysed the potential for realising these experiments in some detail. Since we require particle identification in the target fragmentation region, fixed-target experiments such as COMPASS or HERMES are not appropriate. The preferred option is a polarised $ep$ collider.
The first requirement is to measure particles at extremely small angles ($\theta \le 1$ mrad), corresponding to $t$ less than around 1 ${\rm GeV}^2$. This has already been achieved at HERA in measurements of diffractive and leading proton/neutron scattering using a forward detection system known as the Leading Proton Spectrometer (LPS). The technique for measuring charged particles involves placing detectors commonly known as ‘Roman Pots’ inside the beam pipe itself.
The next point is to notice that the considerations above apply equally to $\r$ as to $\pi$ production, since the ratios ${\cal R}$ are determined by flavour quantum numbers alone. The particle identification requirements will therefore be less stringent, especially as the production of leading strange mesons from protons or neutrons is strongly suppressed. However, we require the forward detectors to have good acceptance for both positive and negatively charged mesons $M = \pi,\r$ in order to measure the ratio (\[eq:ekk\]).
The reactions with a neutron target can be measured if the polarised proton beam is replaced by polarised ${}^3 He$. In this case, if we assume that ${}^3He = Ap + Bn$, the cross section for the production of positive hadrons $h^+$ measured in the LPS is given by $$\s\bigl({}^3He \rta h^+\bigr) \simeq A \s\bigl(p\rta h^+\bigr)
+ B \s\bigl(n \rta p\bigr) + B\s\bigl(n\rta M^+\bigr)
\label{eq:enn}$$ The first contribution can be obtained from measurements with the proton beam. However, to subtract the second one, the detectors must have sufficient particle identification at least to distinguish protons from positively charged mesons.
Finally, estimates of the total rates [@deFlorian:1997th] suggest that around $1\%$ of the total DIS events will contain a leading meson in the target fragmentation region where a LPS would have non-vanishing acceptance ($z>0.6$) and in the dominant domain $x<0.1$. The relevant cross-sections are therefore sufficient to allow the ratios ${\cal R}$ to be measured.
The conclusion is that while our proposals undoubtedly pose a challenge to experimentalists, they are nevertheless possible. Given the theoretical importance of the ‘proton spin’ problem and the topological charge screening mechanism, there is therefore strong motivation to perform target fragmentation experiments at a future polarised $ep$ collider [@DeRoeck:2001ir].
Polarised two-photon physics and a sum rule for $g_1^\c$
========================================================
The $U(1)_A$ anomaly plays a vital role in another sum rule arising in polarised deep-inelastic scattering, this time for the polarised photon structure function $g_1^\c(x_\c,Q^2;K^2)$. For real photons, the first moment of $g_1^\c$ vanishes as a consequence of electromagnetic current conservation [@Bass:1992]. For off-shell photons, we proposed a sum rule in 1992 [@Narison:1992fd; @Shore:1992pm] whose dependence on the virtual momentum of the target photon encodes a wealth of information about the anomaly, chiral symmetry breaking and gluon dynamics in QCD. This is of special current interest since, given the ultra-high luminosity of proposed $e^+ e^-$ colliders designed as $B$ factories, a detailed measurement of our sum rule is about to become possible for the first time.
The first moment sum rule for $g_1^\c$
--------------------------------------
The polarised structure function $g_1^\c$ is measured in the process $e^+ e^- \rta e^+ e^- X$, which at sufficiently high energy is dominated by the two-photon interaction shown in Fig. \[fig:g1gammakin\]. The deep-inelastic limit is characterised by $Q^2\rta \infty$ with $x = Q^2/2p_2.q$ and $x_\c = Q^2/2k.q$ fixed, where $Q^2 = -q^2$, $K^2 = -k^2$ and $s = (p_1+p_2)^2$. The target photon is assumed to be relatively soft, $K^2 \ll Q^2$.
![Kinematics for the two-photon DIS process $e^+ e^- \rta e^+ e^- X$.[]{data-label="fig:g1gammakin"}](g1gammakin.eps){height="4.5cm"}
We are interested in the dependence of the photon structure function $g_1^\c(x_\c,Q^2;K^2)$ on the invariant momentum $K^2$ of the target photon. Experimentally, this is given by $K^2 \simeq E E'_2\theta_2^2$ where $E'_2$ and $\theta_2$ are the energy and scattering angle of the target electron. For the values $K^2\sim m_\r^2$ of interest in the sum rule, the target electron is nearly-forward and $\theta_2$ is very small. If it can be tagged, then the virtuality $K^2$ is simply determined from $\theta_2$; otherwise $K^2$ can be inferred indirectly from a measurement of the total hadronic energy.
The total cross-section $\s$ and the spin asymmetry $\D\s$ can be expressed formally in terms of ‘electron structure functions’ as follows [@Narison:1992fd] $$\s~=~2\pi\a^2~{1\over s}~\int_0^\infty {dQ^2\over Q^2}\int_0^1
{dx\over x^2}~\Bigl[F_2^e ~{1\over y}\Bigl(1 -y +{y^2\over2}\Bigr)
- F_L^e {y\over2} \Bigr]
\label{eq:fa}$$ $$\D\s~=~2\pi\a^2~{1\over s}~\int_0^\infty {dQ^2\over Q^2} \int_0^1
{dx\over x}~g_1^e \Bigl(1- {y\over2}\Bigr)~~~~~~~~~~~~~~~~~~~~~~~~~
\label{eq:fb}$$ where $\s = {1\over 2}(\s_{++} + \s_{+-})$ and $\D\s = {1\over 2}(\s_{++} -
\s_{+-})$ with $+,-$ referring to the electron helicities. The parameter $y = Q^2/x s \ll 1$ and only the leading order terms are retained below.
These electron structure functions can be expressed as convolutions of the photon structure functions with appropriate splitting functions. In particular, we have $$g_1^e(x,Q^2)~=~{\a\over2\pi}\int_0^\infty{dK^2\over K^2} \int_{x}^1
{dx_\c\over x_\c} \D P_{\c e}\Bigl({x\over x_\c}\Bigr)~
g_1^\c(x_\c,Q^2;K^2)
\label{eq:fc}$$ where $\D P_{\c e}(x) = (2 - x)$. This allows us to relate the $x_\c$-moments of the photon structure functions to the $x$-moments of the cross-sections. For the first moment of $g_1^\c$, we find: $$\int_0^1 dx~x {d^3\D\s\over dQ^2 dx dK^2} ~~=~~
{3\over2} \a^3 {1\over s Q^2 K^2}
\int_0^1 dx_\c~g_1^\c(x_\c,Q^2;K^2)
\label{eq:fd}$$
The first moment sum rule follows, as for the proton, by using the OPE (\[eq:eb\]) to express the product of electromagnetic currents for the incident photon in terms of the axial currents $J_{\m5}^a$. The matrix elements $\langle \c^*(k)|J_{\m5}^a|\c^*(k)\rangle$ with the target photon are then expressed in terms of the 3-current AVV Green function involving one axial and two electromagnetic currents. We define form factors for this fundamental correlator as follows: $$\begin{aligned}
&-i\langle0|J_{\m5}^a(p) J_\l(k_1) J_\r(k_2)|0\rangle ~~&=~~
A_1^a ~\e_{\m\l\r\a}k_1^\a ~+~ A_2^a ~\e_{\m\l\r\a}k_2^\a
\nonumber\\
&{}&~+A_3^a ~\e_{\m\l\a\b}k_1^\a k_2^\b k_{2\r} ~+~
A_4^a ~\e_{\m\r\a\b}k_1^\a k_2^\b k_{1\l}
\nonumber\\
&{}&~+A_5^a ~\e_{\m\l\a\b}k_1^\a k_2^\b k_{1\r} ~+~
A_6^a ~\e_{\m\r\a\b}k_1^\a k_2^\b k_{2\l}
\nonumber\\
&{}&{}\label{eq:fe}\end{aligned}$$ where the six form factors are functions of the invariant momenta, i.e. $A_i^a = A_i^a(p^2,k_1^2,k_2^2)$. We also abbreviate $A_i^a(0,k^2,k^2)
= A_i^a(K^2)$.
0.1cm The first moment sum rule for $g_1^\c$ is then [@Narison:1992fd]: $$\int_0^1 dx_\c~g_1^\c(x_\c,Q^2;K^2) ~=~ 4\pi\a \sum_{a=3,8,0} \D C_1^a(Q^2)
\Bigl(A_1^a(K^2) - A_2^a(K^2)\Bigr)
\label{eq:ff}$$ where the Wilson coefficients are related to those in eq.(\[eq:ec\]) by $\D C_1^3 = \D C_1^{NS}$, $\D C_1^8 = {1\over\sqrt3}\D C_1^{NS}$ and $\D C_1^0 = {2\sqrt2\over\sqrt3}\D C_1^{S}$.[^13]
0.1cm Now, just as the sum rule for the proton structure function $g_1^p$ could be related to low-energy meson-nucleon couplings via the $U(1)_A$ Goldberger-Treiman relations, we can relate this sum rule for $g_1^\c$ to the pseudoscalar meson radiative decays using the analysis in section \[radiative\]. Introducing the [*off-shell*]{} radiative pseudoscalar couplings for photon virtuality $K^2$, we define form factors $$F^a(K^2) ~=~ 1 - \Bigl(a^a_{\rm em}{\a\over\pi}\Bigr)^{-1}~
\hat f^{a\a} g_{\hat\eta^\a \c\c}(K^2)
\label{eq:fg}$$ or alternatively, $$\begin{aligned}
&F^3(K^2) ~&=~ 1 - \Bigl(a^3_{\rm em}{\a\over\pi}\Bigr)^{-1}~
f_\pi g_{\pi\c\c}(K^2)
\nonumber\\
&F^8(K^2) ~&=~ 1 - \Bigl(a^8_{\rm em} {\a\over\pi}\Bigr)^{-1}~
\Bigl(f^{8\eta} g_{\eta\c\c}(K^2) + f^{8\eta'}g_{\eta'\c\c}(K^2) \Bigr)
\nonumber\\
&F^0(K^2) ~&=~ 1 - \Bigl(a^0_{\rm em} {\a\over\pi}\Bigr)^{-1}~
\Bigl(f^{0\eta}g_{\eta\c\c}(K^2) + f^{0\eta'} g_{\eta'\c\c}(K^2)
+ \sqrt{6}A g_{G\c\c}(K^2) \Bigr)
\nonumber\\
&{}&{}\label{eq:fh}\end{aligned}$$ where the $a^a_{\rm em}$ are the electromagnetic $U(1)_A$ anomaly coefficients defined earlier. We may then rewrite the sum rule as $$\int_0^1 dx_\c~g_1^\c(x_\c,Q^2;K^2) ~=~ {1\over2}{\a\over\pi}
\sum_{a=3,8,0} \D C_1^a(Q^2)~
a^a_{\rm em}~F^a(K^2)
\label{eq:fi}$$
0.1cm The dependence of the $g_1^\c$ on the invariant momentum $K^2$ of the target photon reflects many key aspects of both perturbative and non-perturbative QCD dynamics. For on-shell photons, $K^2=0$, we have simply [@Bass:1992; @Narison:1992fd] $$\int_0^1 dx_\c~g_1^\c(x_\c,Q^2;K^2=0) ~=~ 0
\label{eq:fj}$$ This is a consequence of electromagnetic current conservation. This follows simply by taking the divergence of eq.(\[eq:fe\]) and observing that in the limit $p\rta 0$, both $A_1$ and $A_2$ are of $O(K^2)$.[^14]
In the asymptotic limit where $K^2 \ll m_\r^2$, a relatively straightforward renormalisation group analysis combined with the anomaly equation shows that, for the flavour non-singlets, the $A_i^a$ tend to the value ${1\over2}{\a\over\pi}a_{\rm em}^a$. while in the flavour singlet sector, $A_i^0$ has an additional factor depending on the anomalous dimension $\c$. Using the explicit expressions for the Wilson coefficients, we find $$\begin{aligned}
\int_0^1{} dx_\c~g_1^\c(x_\c,Q^2;K^2\ll m_\r^2)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\nonumber\\
=~{1\over6} {\a\over\pi} \Bigl(1 - {\a_s(Q^2)\over\pi}\Bigr)
\biggl(a_{\rm em}^3 + {1\over\sqrt3}a_{\rm em}^8 +
{2\sqrt2\over\sqrt3}a_{\rm em}^0
\exp\biggl[ \int_{t(K)}^{t(Q)}{} dt'~\c(\a_s(t'))\biggr] \biggr)
\nonumber\\
{}\nonumber\\
=~{1\over3} {\a\over\pi}
\biggl[1 ~-~ {4\over9}{1\over \ln Q^2/\L^2} ~+~
{16\over81} \biggl({1\over \ln Q^2/\L^2}- {1\over \ln K^2/\L^2}
\biggr) \biggr]~~~~~~~~~~~~~~~~~~~~~
\label{eq:fk}\end{aligned}$$ The asymptotic limit is therefore determined by the electromagnetic $U(1)_A$ anomaly, with logarithmic corrections reflecting the anomalous dimension of the flavour singlet current due to the colour $U(1)_A$ anomaly. (See also ref.[@Sasaki:2006bt] for a NNLO analysis.)
In between these limits, the first moment of $g_1^\c$ provides a measure of the form factors defining the 3-current $AVV$ Green function, which encodes a great deal of information about the dynamics of QCD, especially the non-perturbative realisation of chiral symmetry [@Shore:1992pm]. Equivalently, in the form (\[eq:fi\]), it measures the momentum dependence of the off-shell radiative couplings of the pseudoscalar mesons as the form factors $F^a(K^2)$ vary from 0 to 1.
Just as for $g_1^p$, we can again isolate a dependence on the topological susceptibility through the identification of the flavour singlet decay constant $\hat f^{00}$ in eq.(\[eq:fg\]) with $\sqrt{\chi'(0)}$ in the chiral limit. This time, however, it is unlikely to be a good approximation to set the corresponding coupling $g_{\hat\eta^0 \c\c}$ equal to its OZI value since it is not RG invariant. A more promising approximation is to recall from section 4 that the RG invariant gluonic coupling $g_{G\c\c}(0)$ is OZI suppressed and likely to be small. This was confirmed by the phenomenological analysis. If we assume this is also true of the off-shell coupling, then we may approximate the sum rule for $g_1^\c$ entirely in terms of the off-shell couplings of the physical mesons $\pi^0$, $\eta$ and $\eta'$.
In general, the momentum dependence of the form factors $(A_1^a - A_2^a)$ or $F^a$ will depend on the fermions contributing to the AVV Green function [@Shore:1992pm]. In the case of leptons, or heavy quarks, the crossover scale as the form factors $F^a(K^2)$ rise from 0 to 1 with increasing $K^2$ will be given by the fermion mass. For the light quarks, however, we expect the crossover scale to be a typical hadronic scale $\sim m_\r$ rather than $m_{u,d,s}$. This can be justified by a rough OPE argument and is consistent with old ideas of vector meson dominance [@Shore:1992pm; @Ueda:2006cp]. This behaviour would be an interesting manifestation of the spontaneous breaking of chiral symmetry.
Once again, therefore, we see a close relation between the realisation of sum rules in high-energy deep-inelastic scattering and low-energy meson physics. All these issues are discussed at some length in our earlier papers, but here we now turn our attention to the vital question of whether the $g_1^\c$ sum rule can be measured in current or future collider experiments [@Shore:2004cb].
Cross-sections and spin asymmetries at polarised B factories
------------------------------------------------------------
The spin-dependent cross-sections for the two-photon DIS process $e^+ e^- \rta e^+ e^- X$ were analysed in refs.[@Narison:1992fd; @Shore:2004cb] taking account of the experimental cuts on the various kinematical parameters. Keeping the lower cut on $Q^2$ as a free parameter, we found the following results for the total cross-section and spin asymmetry: $$\s ~\simeq~
0.5 \times 10^{-8}~{1\over Q^2_{\rm min}}~\log{Q^2_{\rm min}\over \L^2}
~\biggl(\log{s\over Q^2_{\rm min}}\biggr)^2
\label{eq:ffa}$$ and $${\D\s\over\s} ~=~
{1\over2}~{Q^2_{\rm min}\over s}~ \log{s\over 4 Q^2_{\rm min}}~
\biggl[ 1 + \log{s\over 4\L^2}\biggl(\log{Q^2_{\rm min}\over\L^2}\biggr)^{-1}
\biggr]
\label{eq:ffb}$$ In order to measure the $g_1^\c$ sum rule, we need to find collider parameters such that the spin asymmetry is significant in a kinematic region where the total cross-section is still large. A useful statistical measure of the significance of the asymmetry is that $\sqrt{L\s}\D\s/\s \gg 1$, where $L$ is the luminosity.
When we first proposed the first moment sum rule for $g_1^\c$, the luminosity available from the then current accelerators was inadequate to allow it to be studied. For example, for a polarised version of LEP operating at $s=10^4~{\rm GeV}^2$ with an annual integrated luminosity of $L=100~{\rm pb}^{-1}$, and optimising the cut at $Q_{\rm min}^2 = 10~{\rm GeV}^2$, we only have $\s \simeq 35~{\rm pb}$ and $\D\s/\s \simeq 0.01$. The corresponding annual event rate would be $3.5\times 10^3$ and the statistical significance $\sqrt{L\s} \D\s/\s \simeq 0.5$, so even a reliable measurement of the spin asymmetry could not be made.
Clearly, a hugely increased luminosity is required and this has now become available with proposals for machines with projected annual integrated luminosities measured in inverse attobarns. However, as noted in ref.[@Narison:1992fd], if this increased luminosity is associated with increased CM energy, then the $1/s$ factor in the spin asymmetry (\[eq:ffb\]) sharply reduces the possibility of extracting $g_1^\c$. There is also a competition as $Q^2_{\rm min}$ is varied between increasing spin asymmetry and decreasing total cross-section. This is particularly evident when we analyse the potential of the ILC [@ILCone; @ILCtwo] for measuring the sum rule [@Shore:2004cb]. We find that even optimising the $Q^2_{\rm min}$ cut, the spin asymmetry is still only of order $\D\s/\s \simeq 0.002$ when $\s$ itself has fallen to around $15~{\rm pb}$. While, given the high luminosity, this would allow a measurement of the first moment of $g_1^\c$ integrated over $K^2$, a detailed study of the $K^2$-dependence of the sum rule requires a much greater spin asymmetry.
This leads us to consider instead the new generation of ultra-high luminosity $e^+ e^-$ colliders. Although these are envisaged as $B$ factories, these colliders operating with polarised beams would, as we now show, be extremely valuable for studying polarisation phenomena in QCD. As an example of this class, we take the proposed SuperKEKB collider. (The analysis for PEPII is very similar, the main difference being the additional ten-fold increase in luminosity in the current SuperKEKB proposals.)
SuperKEKB is an asymmetric $e^+ e^-$ collider with $s= 132~{\rm GeV}^2$, corresponding to electron and positron beams of 8 and 3.5 GeV respectively. The design luminosity is $5\times 10^{35}$ ${\rm cm}^{-2} s^{-1}$, which gives an annual integrated luminosity of $5~{\rm ab}^{-1}$ [@SuperKEKB]. To see the effects of the experimental cut on $Q_{\rm min}^2$ in this case, we have plotted the total cross-section and the spin asymmetry in Fig. \[fig:SuperKEKBfig\], in the range of $Q_{\rm min}^2$ from 1 to $10~{\rm GeV}^2$. In this range $\s$ is falling like $1/Q_{\rm min}^2$ while $\D\s/\s$ rises to what is actually a maximum at $Q_{\rm min}^2=10~{\rm GeV}^2$.
![The left-hand graph shows the total cross-section $\s$ (in pb) at SuperKEKB as the experimental cut $Q^2_{\rm min}$ is varied from 1 to 10 ${\rm GeV}^2$. The right-hand graph shows the spin asymmetry $\D\s/\s$ over the same range of $Q^2_{\rm min}$.[]{data-label="fig:SuperKEKBfig"}](Pbsigmarev.eps "fig:"){height="3.2cm"} 0.5cm ![The left-hand graph shows the total cross-section $\s$ (in pb) at SuperKEKB as the experimental cut $Q^2_{\rm min}$ is varied from 1 to 10 ${\rm GeV}^2$. The right-hand graph shows the spin asymmetry $\D\s/\s$ over the same range of $Q^2_{\rm min}$.[]{data-label="fig:SuperKEKBfig"}](Deltasigmarev.eps "fig:"){height="3.2cm"}
Taking $Q_{\rm min}^2 =5~{\rm GeV}^2$, we find $\s \simeq 12.5~{\rm pb}$ with spin asymmetry $\D\s/\s \simeq 0.1$. The annual event rate is therefore $6.25 \times 10^7$, with $\sqrt{L\s} \D\s/\s \simeq 750$. This combination of a very high event rate and the large $10\%$ spin asymmetry means that SuperKEKB has the potential not only to measure $\D\s$ but to access the full first moment sum rule for $g_1^\c$ itself. Recall from eq.(\[eq:fd\]) that to measure $\int_0^1 dx~g_1^\c(x,Q^2;K^2)$ we need not just $\D\s$ but the fully differential cross-section w.r.t. $K^2$ as well as $x$ and $Q^2$ if the interesting non-perturbative QCD physics is to be accessed. To measure this, we need to divide the data into sufficiently fine $K^2$ bins in order to plot the explicit $K^2$ dependence of $g_1^\c$, while still maintaining the statistical significance of the asymmetry. The ultra-high luminosity of SuperKEKB ensures that the event rate is sufficient, while its moderate CM energy means that the crucial spin asymmetry is not overly suppressed by its $1/s$ dependence.
Our conclusion is that the new generation of ultra-high luminosity, moderate energy $e^+ e^-$ colliders, currently conceived as $B$ factories, could also be uniquely sensitive to important QCD physics if run with polarised beams. In particular, they appear to be the only accelerators capable of accessing the full physics content of the sum rule for the first moment of the polarised structure function $g_1^\c(x,Q^2;K^2)$. The richness of this physics, in particular the realisation of chiral symmetry breaking, the manifestations of the axial $U(1)_A$ anomaly and the role of gluon topology, provides a strong motivation for giving serious consideration to an attempt to measure the $g_1^\c$ sum rule at these new colliders.
[**Acknowledgements**]{} 0.3cm
In addition to Gabriele, I would like to thank Daniel de Florian, Massimiliano Grazzini, Stephan Narison and Ben White for their collaboration on the original research presented here. This paper has been prepared with the partial support of PPARC grant PP/D507407/1.
[99.]{}
G. Veneziano, Nucl. Phys. **B159**, 213 (1979). G. M. Shore, Nucl. Phys. **B569**, 107 (2000), \[arXiv:hep-ph/9908217\]. G. M. Shore, Nucl. Phys. **B744**, 34 (2006), \[arXiv:hep-ph/0601051\]. G. Veneziano, Mod. Phys. Lett. **A4**, 1605 (1989). S. Narison, G. M. Shore and G. Veneziano, Nucl. Phys. **B391**, 69 (1993). G. M. Shore and G. Veneziano, Mod. Phys. Lett. **A8**, 373 (1993). G. M. Shore, Nucl. Phys. **B712**, 411 (2005), \[arXiv:hep-ph/0412192\]. G. M. Shore and G. Veneziano, Phys. Lett. **B244**, 75 (1990). G. M. Shore and G. Veneziano, Nucl. Phys. **B381**, 23 (1992). S. Narison, G. M. Shore and G. Veneziano, Nucl. Phys. **B433**, 209 (1995), \[arXiv:hep-ph/9404277\]. G. M. Shore and G. Veneziano, Nucl. Phys. **B516**, 333 (1998), \[arXiv:hep-ph/9709213\]. G. Veneziano: “The ‘spin’ of the proton and the OZI limit of QCD.” In: *From Symmetries to Strings: Forty Years of Rochester Conferences*, ed. A. Das (World Scientific, 1990), 13-26.
S. L. Adler, Phys. Rev. **177**, 2426 (1969).
J. S. Bell and R. Jackiw, Nuovo Cimento **60A**, 47 (1969).
S. L. Adler and W.A. Bardeen, Phys. Rev. **182**, 1517 (1969).
J. Steinberger, Phys. Rev. **76**, 1180 (1949).
J. Schwinger, Phys. Rev. **82**, 664 (1951).
K. Fujikawa, Phys. Rev. Lett. **42**, 1195 (1979); Phys. Rev. **D21**, 2848 (1980), erratum-ibid. **D22**, 1499 (1980).
G. M. Shore, “$U(1)_A$ problems and gluon topology: anomalous symmetry in QCD”, In: *Hidden Symmetries and Higgs Phenomena*, Zuoz Summer School, Switzerland, 1998, pp 201-223; arXiv:hep-ph/9812354. E. Witten, Nucl. Phys. **B156**, 269 (1979). P. Di Vecchia and G. Veneziano, Nucl. Phys. **B171**, 253 (1980). D. Espriu and R. Tarrach, Z. Phys. **C16**, 77 (1982). G. M. Shore, Nucl. Phys. **B362**, 85 (1991). G. M. Shore and G. Veneziano, Nucl. Phys. **B381**, 3 (1992). G. ’t Hooft, Nucl. Phys. **B72**, 461 (1972).
G. Veneziano, Phys. Lett. **52B**, 220 (1974); Nucl. Phys. **B117**, 519 (1976).
S. Okubo, Phys. Lett. **5**, 165 (1963).
G. Zweig, CERN report 8419/TH412 (1964).
J. Iizuka, Prog. Theor. Phys. Suppl. **37-38**, 21 (1966).
S. Weinberg, Phys. Rev. **D11**, 3583 (1975).
G. ’t Hooft, Phys. Rev. **D14**, 3432 (1976); \[Erratum-ibid. **D18**, 2199 (1978)\]. R. J. Crewther, Riv. Nuovo Cim. **2N8**, 63 (1979). G. A. Christos, Phys. Rept. **116**, 251 (1984). G. ’t Hooft, Phys. Rept. **142**, 357 (1986). M. Gell-Mann, R. J. Oakes and B. Renner, Phys. Rev. **175**, 2195 (1968).
R. F. Dashen, Phys. Rev. **183**, 1245 (1969).
C. Rosenzweig, J. Schechter and C. G. Trahern, Phys. Rev. **D21**, 3388 (1980). P. Di Vecchia, F. Nicodemi, R. Pettorino and G. Veneziano, Nucl. Phys. **B181**, 318 (1981). K. Kawarabayashi and N. Ohta, Nucl. Phys. **B175**, 477 (1980). P. Herrera-Siklody, J. I. Latorre, P. Pascual and J. Taron, Nucl. Phys. **B497**, 345 (1997); Phys. Lett. **B419**, 326 (1998).
H. Leutwyler, Nucl. Phys. Proc. Suppl. **64**, 223 (1998), \[arXiv:hep-ph/9709408\]. R. Kaiser and H. Leutwyler, Eur. Phys. J. **C17**, 623 (2000), \[arXiv:hep-ph/0007101\]. L. Giusti, G. C. Rossi, M. Testa and G. Veneziano, Nucl. Phys. **B628** (2002) 234 \[arXiv:hep-lat/0108009\]. G. M. Shore, Phys. Scripta **T99**, 84 (2002), \[arXiv:hep-ph/0111165\]. Particle Data Group, Review of Particle Properties, Phys. Lett. **B592**, 1 (2004).
M. Acciarri [*et al.*]{}, L3 Collaboration, Phys. Lett. **B418**, 399 (1998).
D. A. Williams [*et al.*]{}, Crystal Ball Collaboration, Phys. Rev. **D38**, 1365 (1988).
N. A. Roe [*et al.*]{}, ASP Collaboration, Phys. Rev. **D41**, 17 (1990).
L. Del Debbio, L. Giusti and C. Pica, Phys. Rev. Lett. **94**, 032003 (2005). \[arXiv:hep-th/0407052\]. A. Di Giacomo, Nucl. Phys. Proc. Suppl. **23B**, 191 (1991). S. Narison, Phys. Lett. **B255**,101 (1991); Z. Phys. **C26**, 209 (1984). S. Narison, G. M. Shore and G. Veneziano, Nucl. Phys. **B546**, 235 (1999), \[arXiv:hep-ph/9812333\]. V. Y. Alexakhin [*et al.*]{} \[COMPASS Collaboration\], “The deuteron spin-dependent structure function g1(d) and its first moment,” arXiv:hep-ex/0609038. A. Airapetian [*et al.*]{} \[HERMES Collaboration\], “Precise determination of the spin structure function g(1) of the proton, deuteron and neutron,” arXiv:hep-ex/0609039. G. Mallot, S. Platchkov and A. Magnon, CERN-SPSC-2005-017; SPSC-M-733.
E. S. Ageev [*et al.*]{} \[COMPASS Collaboration\], Phys. Lett. **B612**, 154 (2005), \[arXiv:hep-ex/0501073\]. J. R. Ellis and R. L. Jaffe, Phys. Rev. **D9**, 1444 (1974), \[Erratum-ibid. **D10**, 1669 (1974)\]. D. V. Bugg, Eur. Phys. J. **C33**, 505 (2004). P. Moskal, “Hadronic interaction of eta and eta$'$ mesons with protons,” arXiv:hep-ph/0408162. S. D. Bass, Phys. Scripta **T99**, 96 (2002), \[arXiv:hep-ph/0111180\]. P. Moskal [*et al.*]{}, Int. J. Mod. Phys. **A20**, 1880 (2005), \[arXiv:hep-ex/0411052\]. P. Moskal [*et al.*]{}, Phys. Rev. Lett. **80**, 3202 (1998), \[arXiv:nucl-ex/9803002\]. K. Nakayama and H. Haberzettl, “Analyzing eta$'$ photoproduction data on the proton at energies of 1.5GeV – 2.3GeV,” arXiv:nucl-th/0507044. M. Dugger \[CLAS Collaboration\], “S=0 pseudoscalar meson photoproduction from the proton,” arXiv:nucl-ex/0512005. G. M. Shore, Nucl. Phys. Proc. Suppl. **39BC**, 101 (1995), \[arXiv:hep-ph/9410383\]. J. Ashman [et al.]{}, Phys. Lett. **B206**, 364 (1988); Nucl. Phys. **B328**, 1 (1990).
R. L. Jaffe and A. Manohar, Nucl. Phys. **B337**, 509 (1990). G. M. Shore and B. E. White, Nucl. Phys. **B581**, 409 (2000), \[arXiv:hep-ph/9912341\]. G. M. Shore, Nucl. Phys. Proc. Suppl. **96**,171 (2001), \[arXiv:hep-ph/0007239\]. B. L. G. Bakker, E. Leader and T. L. Trueman, Phys. Rev. **D70**, 114001 (2004), \[arXiv:hep-ph/0406139\]. R. D. Ball, S. Forte and G. Ridolfi, Phys. Lett. **B378**, 255 (1996), \[arXiv:hep-ph/9510449\]. G. Altarelli and G. G. Ross, Phys. Lett. **B212**, 391 (1988). S. Procureur \[COMPASS Collaboration\], “New measurement of Delta(G)/G at COMPASS,” arXiv:hep-ex/0605043. R. Fatemi \[STAR Collaboration\], “Using jet asymmetries to access Delta(G), the gluon helicity distribution of the proton at STAR,” arXiv:nucl-ex/0606007. Y. Fukao \[PHENIX Collaboration\], “The overview of the spin physics at RHIC-PHENIX experiment,” AIP Conf. Proc. **842**, 321 (2006), \[arXiv:hep-ex/0607033\]. G. M. Shore, “The proton spin crisis: Another ABJ anomaly?”, In: *From the Planck length to the Hubble radius*, Erice 1998, ed. A. Zichichi, World Scientific, Singapore, pp 79-105; arXiv:hep-ph/9812355. L. Trentadue and G. Veneziano, Phys. Lett. **B323**, 201 (1994). M. Grazzini, L. Trentadue and G. Veneziano, Nucl. Phys. **B519**, 394 (1998), \[arXiv:hep-ph/9709452\]. D. de Florian, G. M. Shore and G. Veneziano, “Target fragmentation at polarized HERA: A test of universal topological charge screening in QCD,” In: *Proceedings of the 1997 Workshop with Polarized Protons at Hera*, ed. A. de Roeck and T. Gehrmann, Hamburg/Zeuthen 1997, pp 696-703; arXiv:hep-ph/9711353. G. M. Shore, Nucl. Phys. Proc. Suppl. **64**, 167 (1998), \[arXiv:hep-ph/9710367\]. M. Grazzini, G. M. Shore and B. E. White, Nucl. Phys. **B555**, 259 (1999), \[arXiv:hep-ph/9903530\]. A. De Roeck, Nucl. Phys. Proc. Suppl. **105**, 40 (2002), \[arXiv:hep-ph/0110335\]. S.D. Bass, Int. J. Mod. Phys. **A7**, 6039 (1992).
K. Sasaki, T. Ueda and T. Uematsu, Phys. Rev. **D73**, 094024 (2006), \[arXiv:hep-ph/0604130\]. T. Ueda, T. Uematsu and K. Sasaki, Phys. Lett. **B640**, 188 (2006), \[arXiv:hep-ph/0606267\]. F. Richard [*et al.*]{}, “TESLA: The Superconducting electron positron linear collider with an integrated X-ray laser laboratory. Technical Design Report, Part I”; hep-ph/0106314.
M. Woods [*et al.*]{}, “Luminosity, Energy and Polarization Studies for the Linear Collider”, In: *Proc. 5th International Workshop on Electron-Electron Interactions at TeV Energies*, Santa Cruz, 2003; physics/0403037.
A. G. Akeroyd [*et al.*]{} (SuperKEKB Physics Working Group), “Physics at Super B Factory”; hep-ex/0406071.
[^1]: To appear in the volume [*String Theory and Fundamental Interactions*]{}, published in honour of Gabriele Veneziano on his 65th birthday, eds. M. Gasperini and J. Maharana, Lecture Notes in Physics, Springer, Berlin/Heidelberg 2007.
[^2]: Our notation follows that of ref.[@Shore:2006mm]. The currents and pseudoscalar fields $J_{\m5}^a$, $Q$, $\phi_5^a$ together with the scalar $\phi^a$ are defined by $$\begin{aligned}
&J_{\m5}^a = \bar q \c_\m \c_5 T^a q ~~~~~~~~
&J_\m^a = \bar q \c_\m T^a q ~~~~~~~~
Q = {\a_s\over8\pi} {\rm tr} G_{\m\n} \tilde G^{\m\n}
\nonumber\\
&\phi_5^a = \bar q \c_5 T^a q ~~~~~~~~~
&\phi^a = \bar q T^a q ~~~~~~~~~~~~\end{aligned}$$ where $G_{\m\n}$ is the field strength for the gluon field. Here, $T^i = {1\over2}\l^i$ are flavour $SU(n_f)$ generators, and we include the singlet $U(1)_A$ generator $T^0 = {\bf 1}/\sqrt{2n_f}$ and let the index $a = 0, i$. With this normalisation, ${\rm tr} T^a T^b
= {1\over2}\d^{ab}$ for all the generators $T^a$. This accounts for the rather unconventional factor $\sqrt{2n_f}$ in the anomaly equation but has the advantage of giving a consistent normalisation to the full set of decay constants including the flavour singlets $f^{0\eta'}$ and $f^{0\eta}$.
We will only need to consider fields where $i$ corresponds to a generator in the Cartan sub-algebra, so that $a = 3, 8, 0$ for $n_f = 3$ quark flavours. We define $d$-symbols by $\{T^a,T^b\} = d_{abc} T^c$. For $n_f = 3$, the explicit values are $d_{000} =
d_{033} = d_{088} = d_{330} = d_{880} = \sqrt{2/3}, ~d_{338} = d_{383} =
-d_{888} = \sqrt{1/3}$.
[^3]: We use the following $SU(3)$ notation for the quark masses and condensates: $$\left(\matrix{m_u &0 &0 \cr
0 &m_d &0 \cr
0 &0 &m_s \cr}\right)
= \sum_{a=0,3,8} m^a T^a$$ and $$\left(\matrix{ \langle \bar u u\rangle &0 &0 \cr 0 &\langle \bar d d\rangle
&0 \cr
0 &0 &\langle \bar s s\rangle \cr}\right) =
2 \sum_{0,3,8} \langle \phi^a \rangle T^a$$ where $\langle \phi^a\rangle$ is the VEV of $\phi^a = \bar{q}~ T^a q$. It is also convenient to use the compact notation $$M_{ab} = d_{acb} m^c
~~~~~~~~~~~~~~~~~~
\Phi_{ab} = d_{abc} \langle \phi^c\rangle$$
[^4]: The notation $+\ldots$ refers to additional terms which are required to produce the contact term contributions to the RGEs for $n$-point Green functions and vertices of composite operators. These are discussed fully in refs.[@Shore:1990wp; @Shore:1991dv; @Shore:1991pn], but will be omitted here for simplicity. They vanish at zero-momentum.
[^5]: The existence of a light flavour-singlet Nambu-Goldstone boson would produce a rapid off-shell variation in the $\eta \rta 3\pi$ decay amplitude, in contradiction with the experimental data [@Weinberg:1975].
[^6]: For reviews of the instanton approach to the resolution of the $U(1)_A$ problem, see e.g. refs.[@'tHooft:1976fv; @Crewther:1978kq; @Christos:1984tu; @'tHooft:1986nc].
[^7]: Note especially the frequently misunderstood point that the choice of fields in $\C$ is not required to be in any sense a complete set, nor does the restriction to a given set of fields constitute an approximation. Before imposing dynamical simplifications, the identities derived from $\C$ are [*exact*]{} - increasing the set of basis fields simply changes the definitions of the 1PI vertices. The effective action considered here is therefore different from the non-linear chiral Lagrangians incorporating the large-$N_c$ approach to the pseudoscalar mesons constructed by a number of groups. See, for example, refs. [@Rosenzweig:1979ay; @DiVecchia:1980ve; @DiVecchia:1980sq; @Kawarabayashi:1980dp; @Herrera-Siklody:1996pm; @Leutwyler:1997yr; @Kaiser:2000gs].
[^8]: This section is based on the presentation in ref.[@Shore:2006mm], where we extend and update our original work [@Shore:1999tw; @Shore:2001cs] to include a detailed comparison with experimental data.
[^9]: The $p\rta 0$ limit is delicate, as is the case for the derivation of the conventional Goldberger-Treiman relation, and should be taken in this order. Literally at $p=0$, both sides vanish since $\bar u \c_5 u = 0$.
[^10]: This supersedes the result $a^0|_{Q^2=4{\rm GeV}^2} ~=~ 0.237{}^{+0.024}_{-0.029}$ quoted by COMPASS in 2005 [@Mallot; @Ageev:2005gh], which we used as input into our analysis of the phenomenology of the $U(1)_A$ GT relation in ref.[@Shore:2006mm]. The fits presented here are updated from those of ref.[@Shore:2006mm] to take account of this. For a further discussion of the experimental situation, see section 5.
[^11]: For a careful discussion of the parton interpretation of longitudinal and transverse angular momentum sum rules, see ref.[@Bakker:2004ib]. This confirms our assertion that the axial charge $a^0$ is not to be identified with quark helicities in the parton model.
[^12]: We emphasise again that this decomposition of the matrix elements into products of Green functions and 1PI vertices is [*exact*]{}, independent of the choice of the set of operators $\tilde{\OO}_B$. In particular, it is not necessary for $\tilde{\OO}_B$ to be in any sense a complete set. If a different choice is made, the vertices $\C_{\tilde{\OO}_B pp}$ themselves change, becoming 1PI with respect to a different set of composite fields. In practice, the set of operators $\tilde{\OO}_B$ should be as small as possible while still capturing the essential degrees of freedom. A good choice can also result in vertices $\C_{\tilde{\OO}_B pp}$ which are both RG invariant and closely related to low-energy physical couplings.
[^13]: Explicitly, $$\D C_1^{NS} = {1\over3}\Bigl(1- {\a_s(Q^2)\over\pi}\Bigr), ~~~~~~~
\D C_1^{S} = {1\over3}\Bigl(1- {\a_s(Q^2)\over\pi}\Bigr)~
\exp \int_0^{t(Q)} dt' ~\c(\a_s(t'))$$ at leading order, where $t(Q) = {1\over2}\ln{Q^2\over\m^2}$ and $\c = -{3\over4}{\a_s^2\over(4\pi)^2}$ is the anomalous dimension corresponding to the $U(1)_A$ current renormalisation.
[^14]: Electromagnetic current conservation in eq.(\[eq:fe\]) implies $$A_1^a = A_3^a k_2^2 + A_5^a {1\over2}(p^2 - k_1^2 - k_2^2),~~~~~~~~~
A_2^a = A_4^a k_1^2 + A_6^a {1\over2}(p^2 - k_1^2 - k_2^2)$$ The chiral limit is special since the form factors can have massless poles and is considered in detail in ref.[@Shore:1992pm]. The sum rule (\[eq:fj\]) still holds.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the results of numerical calculations of the channelling and undulator radiation generated by an ultra-relativistic positron channelling along a crystal plane, which is periodically bent. The bending might be due either to the propagation of a transverse acoustic wave through the crystal, or due to the static strain as it occurs in superlattices. The periodically bent crystal serves as an undulator. We investigate the dependence of the intensities of both the ordinary channelling and the undulator radiations on the parameters of the periodically bent channel with simultaneous account for the dechannelling effect of the positrons. We demonstrate that there is a range of parameters in which the undulator radiation dominates over the channelling one and the characteristic frequencies of both types of radiation are well separated. This result is important, because the undulator radiation can be used to create a tunable source of X-ray and $\gamma$-radiation.'
address:
- 'Institut für Theoretische Physik der Johann Wolfgang Goethe-Universität, 60054 Frankfurt am Main, Germany'
- 'Department of Physics, St.Petersburg State Maritime Technical University, Leninskii prospect 101, St. Petersburg 198262, Russia'
- '§A.F.Ioffe Physical-Technical Institute of the Academy of Sciences of Russia, Polytechnicheskaya 26, St. Petersburg 194021, Russia'
author:
- 'Wolfram Krause, Andrei V. Korol, Andrey V. Solov’yov§, and Walter Greiner'
---
In this Letter we present for the first time the results of calculations of the total spectrum of emitted photons accompanying the channelling process of ultra-relativistic charged particles through a crystal which is periodically bent. The present consideration is a further step in investigating new phenomenon which was described recently in [@JPG; @laser] and was called Acoustically Induced Radiation (AIR). It was noted that the periodic pattern of crystal bendings (which can be achieved either through propagation of a transverse acoustic wave, AW, or by using static periodically strained crystalline structures [@laser; @Uggerhoj2000]) gives rise to a new mechanism of electromagnetic emission of the undulator type, in addition to a well-known ordinary channelling radiation [@Kumakhov].
Without any loss of generality further in this Letter we consider the case of a dynamic periodic bending of a crystal by means of transverse AW, as it is illustrated in . Under the action of the transverse AW propagating along the $z$-direction, which defines the center line of initially straight channel (not plotted in the figure) the channel becomes periodically bent. Provided certain conditions are fulfilled [@laser] the beam of particles, which enters the crystal at a small incident angle with respect to the curved crystallographic plane, will penetrate through the crystal following the bendings of its channel. This results in transverse oscillations of the beam particles ([*additional*]{} to the oscillations inside the channel due to the action of the interplanar force). These oscillations become an effective source of spontaneous radiation of undulator type due to the constructive interference of the photons emitted from similar parts of the trajectory. It was demonstrated [@laser] that the system “ultra-relativistic charged particle + periodically bent crystal” serves as a new type of undulator, and, consequently, as a new source of undulator radiation of high intensity, monochromaticity and of a particular pattern of the angular-frequency distribution.
As it was pointed out in [@laser] this scheme leads, in addition to the spontaneous radiation, to a possibility to generate stimulated emission, similar to the one known for a free electron laser [@Madey71] in which the periodicity of a trajectory of an ultra-relativistic projectile is achieved by applying spatially periodic magnetic field. In connection with the stimulated AIR it was noted in [@laser] that to achieve noticeable degree of amplification one has to operate with a positron bunch of a high volume density, which, nevertheless, turned out to be achievable in modern accelerators as discussed in [@Zhang1999].
The main subject of our previous studies [@JPG; @laser] was the characteristics of the undulator AIR itself. Due to this reason we primarily investigate the case of a high-amplitude AW, $a\gg d$ with $d$ standing for the interplanar spacing, propagating through the crystal. Less attention, except for some qualitative estimates, has been paid to the detailed investigation of the mutual influence of two types of radiation, the ordinary channelling radiation and the AIR, in forming the total spectrum of the radiation emitted by a channeled positron. This topic is addressed in the present Letter where we report our first quantitative results on numerical calculation of the spectrum of the emitted photons with both mechanisms taken into account simultaneously. We demonstrate that there are ranges of (i) the parameters of AW, which are the amplitude, $a$, the wavelength $\lambda$, and the sound velocity $V$, (ii) the energies of projectile positron, $\varepsilon$, (iii) the crystal parameters, which include the length of a crystal, the constituent atoms and the types of crystallographic planes, inside which
- the characteristic frequencies of the AIR and the ordinary channelling radiation are well separated,
- the intensity of AIR is essentially higher than that of the ordinary channelling radiation,
- the radiative spectrum is stable towards the total losses of the particle (in the case of a positron it is primarily the radiative ones)
These items, except the last one which was considered in [@losses], are discussed below in the Letter.
To conclude the introductory part we mention that in the past the problem of evaluating the total spectrum of radiation formed in a bent crystal was considered in several publications [@Taratin; @Arutyunov] in the case of a projectile channelling in a crystal bent with a constant curvature radius. In our Letter for the first we investigate the problem for the periodically bent channel. In full the results of our research will be published elsewhere [@new_full]. Below we present the essential points.
An adequate approach to the problem of the radiation emission by an ultra-relativistic particle moving in an external field was developed by Baier and Katkov in the late 1960s [@Baier67] and was called “operator quasi-classical method” by the authors. The details of that formalism can be found in [@Baier; @Land4]. The advantage of this method is that it allows to use the classical trajectory for the particle in an external field and, simultaneously, it takes into account the effect of the radiative recoil.
For particles with spin $s=1/2$ the energy radiated into a given direction ${\bf n}$ summed over the polarizations of the photon and the projectile is given by (the CGS system is adopted throughout the paper) $$\d E \equiv
{\d E \over \hbar\d \omega\, \d \Omega_{\bf n} }=
\frac{\alpha\, \omega^2}{4 \pi^2}
\int_0^\tau \d t_1 \int_0^\tau \d t_2 \
\e^{\i \omega^{\prime} \varphi(t_1,t_2)}
\ f(t_1,t_2),
\label{wkb_1}$$ where $\alpha\approx 1/137$ is the fine structure constant, $\varphi(t_1,t_2)=t_1-t_2- {\bf n}\cdot({\bf r}(t_1)-{\bf r}(t_2))/c$, and $f(t_1,t_2)=\left\{\left[ 1+(1+u)^2\right] \left({\bf v}(t_1){\bf
v}(t_2)/c^2-1\right)+u^2\gamma^{-2}\right\}/2$, where $c$ is the velocity of light, $\gamma=\varepsilon/m$ is the relativistic factor and $u=\hbar \omega/(\varepsilon - \hbar \omega)$.
The main advantage of the the quasi-classical approach is that to calculate the angular-spectral distribution of the radiation one only needs to know the time dependencies of classical radius-vector ${\bf
r}(t)$ and the velocity ${\bf v}(t)$ of projectile. In connection with a positron channelling through a periodically bent crystal the above relations impose, if applied directly, some restrictions on the projectile energy, on the parameters of the crystal (channel) and the AW.
The channelling process in a bent crystal takes place if the centrifugal force in the channel is less than the maximal force due to the interplanar field [@Tsyganov]. For a periodically bent crystal the maximal centrifugal force is equal to $m \gamma v^2/R_{\rm
min}$, with $v\approx c$ and $R_{\rm min}$ being a minimum curvature radius of the bent channel. Hence, the following condition must be fulfilled [@JPG; @laser] $$m \gamma c^2/R_{\rm min} < U_{\rm max}^{\prime}.
\label{1}$$ where $U_{\rm max}^{\prime}$ stands for the maximum gradient of the interplanar field. For an acoustically bent channel $R_{\rm min}
=(\lambda/2\pi)^2/a$, therefore, the inequality (\[1\]) bounds together the characteristics of the crystal, $U_{\rm max}^{\prime}$, the AW amplitude and the wavelength, and the relativistic factor $\gamma$. A comprehensive study of the allowed ranges of all parameters involved in (\[1\]) was carried out in [@JPG; @laser].
Second requirement which has to be fulfilled to make eq. (\[wkb\_1\]) directly applicable to calculating the spectrum accompanying the channelling process concerns the upper limit of integration $\tau$, which is related to the crystal length $L$ through $L=c\tau$, and which, for given value of the AW wavelength, defines the number of the undulator periods as $N=L/\lambda$. The AIR acquires specific features of the undulator-type radiation (such as the monochromaticity and particular pattern of the angular-frequency distribution, see e.g. [@Baier]) provided $N\gg 1$. The length of a crystal is subject to a physical condition that the positron bunch stays inside the channel when penetrating through the crystal on the scale of $L$. In reality, the parasitic effect, the dechannelling leads to a decrease in the volume density of the channeled particles $n(z)$ with penetration distance $z$, and roughly satisfies the exponential decay law for both straight and bent channels (see [@Gemmell; @Biryukov]), $n(z) = n_0\, \exp\left(- z/ L_d\right)$, where $n_0$ is the volume density t the entrance, and $L_d$ is the dechannelling length, which for given crystal and channel depends on a positron energy and on the curvature radius. For a periodically bent crystal $L_d(\gamma, R)$ can be estimated as [@laser; @Biryukov; @Komarov] $$L_d(\gamma, R)
=
\left(1 - R_c/R_{\rm min}\right)^2 \,
{256 \over 9\pi^2}\,{ a_{\rm TF} \over r_{\rm cl} }\,
{ d \over L_c }\, \gamma
\label{stim9}$$ where $r_{\rm cl}$ is the classical radius of an electron, $a_{\rm
TF}=0.8853 Z^{-1/3} a_0$ ($a_0$ is the Bohr radius) and $I = 16
Z^{0.9}$ eV are, respectively, the Thomas-Fermi radius and ionization potential of the crystal atoms, $Z$ is the atomic number, $d$ is the interplanar distance in the lattice. The quantity $R_c= \varepsilon/
U_{\rm max}^{\prime}$ is the critical (minimal) radius consistent with the channelling condition in a bent crystal (\[1\]). The quantity $L_c =\ln\left(\sqrt{2\gamma}\, mc^2/I\right) - 23/24$ is the Coulomb logarithm characterizing the ionization losses of an ultra-relativistic positron in amorphous media with account for the density effect (see e.g. [@Komarov; @Sternheimer]).
The AIR will have a pronounced pattern of undulator-type radiation provided the number of the periods is large on the scale of $L_d$. Combined with (\[stim9\]) this condition leads to another restriction on the parameters involved $$N = \left(1 - R_c/R_{\rm min}\right)^2 \,
{256 \over 9\pi^2}\,{ a_{\rm TF} \over r_{\rm cl} }\,
{ d \over L_c }\, {\gamma \over \lambda} \gg 1
\label{2}$$ where $\lambda$ can be related to the AW frequency $\nu$ through $\lambda=V/\nu$ with $V$ standing for the sound velocity.
illustrates the restrictions which are imposed on the values of $a$ and $\nu$ by inequalities (\[1\]) and (\[2\]) in the case of $\varepsilon=0.5$ GeV planar channelling in $Si$ along (110) crystallographic planes. The diagonal straight lines correspond to various values (as indicated) of the parameter $C=\varepsilon/(R_{\rm min}U_{\rm max}^{\prime}) \leq 1$ consistent with the channelling condition (\[1\]). The curved lines correspond to various values (as indicated) of the number of undulator periods $N$ related to the dechannelling length $L_d$ through eq. (\[2\]). The horizontal lines mark the values of the AW amplitude equal to $d$ (with $d=1.92\, \AA$ being the (110) interplanar distance in $Si$) and to $10\,d$. The vertical line marks the value $\nu = 200$ MHz of the AW frequency for which the spectra (\[wkb\_1\]) were calculated. We used the value $V=4.67\times 10^5$ cm/s for the velocity of sound in $Si$ (this value was obtained by using the data from [@Mason]). Thus, the AW wave-length used in our calculations equals to $\lambda=2.33\times 10^{-3}$ cm.
The calculated spectra of the radiation emitted in the forward direction (with respect to the $z$-axis, see ) for photon energies from 45 keV to 1.5 MeV are presented in figures \[Fig.3\]. The details of the analytical evaluation of the right-hand side of (\[wkb\_1\]) and its numerical implementation as well as more extended results of the calculation will be published soon [@new_full]. Here we briefly sketch the numeric procedure and present a discussion of the exhibited results. The AW frequency, the number of undulator periods and crystal length were fixed at $\nu =
200$ MHz, $N=15$ and $L=N\, \lambda = 3.5\times 10^{-2}$ cm. The ratio $a/d$ of the AW amplitude to the interplanar spacing was varied within the interval $a/d = 0\dots 10$. The case $a/d=0$ corresponds to the straight channel.
To evaluate the spectral distribution (\[wkb\_1\]) the following procedure was adopted.
Firstly, the spectrum was calculated for individual trajectories. These were obtained by solving the relativistic equations of motion with both the interplanar and the centrifugal potentials taken into account. We used the continuum approximation [@Lindhard] to describe the projectile positron – lattice atoms interaction. Within these scheme we considered two frequently used [@Gemmell] analytic forms for the interplanar potential, the harmonic and the Molière potentials calculated at the temperature $T=150$ K to account for the thermal vibrations of the lattice atoms. For each $a/d$ value by changing the initial values of the entrance coordinate $y^{(0)}$ and the initial velocities along the $y$-axis (see ) we left only those trajectories which corresponded to the case of stable channelling through the whole crystal length $L$. We call a trajectory as a “stable” one if moving along it the particle does not approach crystalline planes at a distance less than the Thomas-Fermi radius $a_{\rm TF}$ ($a_{\rm TF}=0.194$ Å for a $Si$ atom). This allowed us to totally disregard, at least on the scale $L\sim L_d$, the random scattering of a projectile by lattice electrons (see e.g. [@Gemmell; @Biryukov]). Thus, for each $a/d$ value, we defined the ranges of the initial coordinates $y^{(0)}\in
[-d/2+a_{\rm TF}, d/2-a_{\rm TF}]$ and the velocities $v_y^{(0)}$, and, correspondingly, the initial phase volume $\Phi^{(0)}(a/d)=\oint
p_y^{(0)} \d y^{(0)}$ (where $p_y^{(0)}$ stands for initial transverse momentum) for which the corresponding classical trajectories are stable when channelling through the whole crystal length $L$. Then, discretizing the found initial phase volume $\Phi^{(0)}(a/d)$ by choosing $N_{y^{(0)}}\times N_{v_y^{(0)}}$ points $(y^{(0)},
p_y^{(0)})\in \Phi^{(0)}(a/d)$, the individual spectra $\d E(y^{(0)},
p_y^{(0)})$ were calculated for each pair of the initial coordinate and velocity. Finally, for each $a/d$ value we evaluated the averaged spectrum defined as follows: $$\langle \d E \rangle
=
{1 \over \Phi^{(0)}(a/d=0)}\,
\oint_{\Phi^{(0)}(a/d)} \left[\d E(y^{(0)}, p_y^{(0)})\right]\,
p_y^{(0)} \d y^{(0)}
\label{3}$$ Here, the integration is carried out over the phase volume $\Phi^{(0)}(a/d)$, and the integral is scaled by the phase volume $\Phi^{(0)}(a/d=0)$ of stable trajectories in straight channel. The ratio $\Phi^{(0)}(a/d)/\Phi^{(0)}(a/d=0)$ describes the number of particles channelled through the bent crystal relative to the number of particles channelled through the straight one. Hence, it is convenient to use the quantity $\langle \d E \rangle$ to compare the spectra produced by effectively different number of projectiles as it happens for different $a/d$ values.
Figures \[Fig.3\] correspond to the spectra (\[3\]) obtained as outlined above. The results presented were calculated by using the Molière approximation for interplanar potential.
The first graph in corresponds to the case of zero amplitude AW (the ratio $a/d=0$) and, hence, presents the spectral dependence of the ordinary channelling radiation only. The asymmetric shape of the calculated ordinary channelling radiation peak, which is due to the strong anharmonic character of the Molière potential, bears close resemblance with the experimentally measured spectra [@Uggerhoj1993]. The spectrum starts at $\hbar\omega\approx 960$ keV, reaches its maximum value at $1190$ keV, and steeply cuts off at $1200$ keV. This peak corresponds to the radiation into the first harmonic of the ordinary channelling radiation (see e.g. [@Kumakhov1989]), and there is almost no radiation into higher harmonics. The latter fact is consistent with general theory of dipole radiation by ultra-relativistic particles undergoing quasi-periodic motion (see e.g. [@Baier]). Dipole approximation is valid provided the corresponding undulator parameter $p_c = 2\pi
\gamma(a_c/\lambda_c)$ is much less than 1. In this relation $a_c$ and $\lambda_c$ stand for the characteristic scales of, correspondingly, the amplitude and the wave-length of the quasi-periodic trajectory. For the channelling motion one can estimate $a_c$ as $d/2$, and $\lambda_c=c \tau_c$, where $\tau_c\sim
2\pi\sqrt{m\gamma/U^{\prime\prime}}$ standing for the characteristic period of the channelling oscillations (using the harmonic approximation for the interplanar potential one gets $U^{\prime\prime}
\sim 8\, U_0/d^2$ where $U_0$ is the depth of the interplanar potential well). In the case of 0.5 GeV positron channeled along (110) planes in $Si$ one has $\gamma\approx 10^3$, $U_0 = 23$ eV, $d=1.92$ Å [@Baier]. Hence, $p_c\approx 0.2 \ll 1$ and all the channelling radiation is concentrated within some interval in the vicinity of the energy of the first harmonic. The latter one can estimate as (see e.g. [@Baier]) $\omega_c^{(1)}\sim4\pi\gamma^2/\tau_c\sim 4\gamma^2
c/d\sqrt{U_0/\varepsilon}$ arriving at the value $\hbar\omega_c^{(1)}\approx 1190$ keV which is exactly the calculated maximum value $1190$ keV. More accurate estimates (the details are presented in [@new_full]), based on the account for the dependence of the channeling oscillation period $\tau_c$ on the amplitude $a_c$ of the oscillations, also reproduce the calculated position and the width of the peak of the channeling radiation.
Increasing the $a/d$ ratio leads to the modifications in the spectrum of radiation. The changes which occur manifest themselves via three main features, (i) the lowering of the ordinary channelling radiation peak, (ii) the gradual increase of the intensity of undulator radiation due to the crystal bending, (iii) the appearing of additional structure (the sub-peaks) in the vicinity of the first harmonic of the ordinary channelling radiation. Let us discuss all these features.
The decrease in the intensity of the ordinary channelling radiation with the increase of the $a/d$ ratio is related to the simple fact that the growth of the AW amplitude leads to lowering of the allowed maximum value of the channelling oscillations amplitude $a_c$ (this are measured with respect to the centerline of the bent channel) [@losses]. Indeed, inside the bent channel the motion of the particle is subject to the action of the effective potential $$U_{eff}(\rho) = U(\rho) - {\varepsilon \over R(z)}\, \rho
\label{2.25}$$ where $\rho\in [-d/2,d/2]$ is the (local) distance from the centerline and $R(z)$ is the (local) curvature radius of the channel. For a channel bent by a transverse harmonic AW $R =
\left[(\lambda/2\pi)^2/a\right]\,\sin(2\pi z/\lambda)$. The particle could be trapped into the channelling mode provided its total energy associated with the transverse motion is less the minimal value, $U_{eff}(\rho_0)$, of the two maxima points of the asymmetric potential well described by (\[2.25\]) [@losses; @Biryukov]. The potential $U_{eff}(\rho)$ reaches the magnitude of $U_{eff}(\rho_0)$ at some point $\rho_0$ which satisfies the condition $|\rho_0|< d/2$ and the absolute value of $\rho_0$ decreases with the growth of $a$. Hence, the larger the channel is bent the lower the allowed values of the channelling oscillations amplitude are, and, consequently, the less intensive is the channelling radiation, which is, essentially, proportional to the squared amplitude [@Baier]. Let us note, that since we have restricted the range of the $a_c$ values by imposing the condition $a_c < d/2-a_{TF}$ (see the discussion above ), then the decrease of the intensity of channelling radiation occurs starting with some non-zero value of the AW amplitude for which $|\rho_0|$ also becomes less than $d/2-a_{TF}$.
These arguments explain the fact that the intensities of the ordinary channelling radiation for $a/d=0$ and $a/d=1$ are much alike, while for $a/d>1$ it starts loosing the magnitude.
The undulator radiation (the AIR) related to the motion of the particle along the centerline of periodically bent channel is absent in the case of straight channel (the graph $a/d=0$), and is almost invisible for comparatively small amplitudes of the AW (see the graph for $a/d=1$). With $a$ increasing the peaks corresponding to AIR are becoming more prominent, and for large $a$ values ($a/d \sim 10$) two additional features appear: the intensity of the first harmonic of the AIR becomes larger than the intensity of the ordinary channelling radiation, and there appears radiation into the third harmonic of the AIR.
The positions and the widths of the AIR peaks can be quite accurately estimated as follows (see [@JPG; @laser]): $\omega^{(n)} = 8\pi
\gamma^2 c \lambda^{-1}\, n /(2+p^2)$, where the integer $n=1,2\dots$ enumerates the harmonics, and $p=2\pi\gamma(a/\lambda)$ stands for the parameter of the undulator related to the periodicity of the channel bending. The width of each peak $\Delta \omega=
(1/N)(\omega^{(n)}/n)$ is independent on $n$. Substituting the values of $\gamma$, $\lambda$ and $d$ indicated above one expresses the undulator parameter via the ratio $a/d$: $p\approx 0.05 (a/d)$. Even for the largest considered value $a/d=10$ the parameter $p$ is less than 1, thus making the radiation into higher harmonics of the AIR almost negligible compared with the intensity radiated into the fundamental harmonic $n=1$. The latter is located at $\hbar\omega^{(1)} \approx 90$ keV having the width $\hbar\Delta
\omega \approx 6$ keV which is almost 40 times less than the width of the peak of the channeling radiation. These values depend neither on the ratio $a/d=10$ nor on the type of the interplanar potential.
As mentioned, all graphs in refer to the forward emission. Therefore, in accordance with general theory of the undulator radiation (see e.g. [@Baier]), the second peak of the AIR, which is mostly pronounced in the case $a/d=10$, corresponds to the third harmonic of AIR, and is located at $\hbar\omega^{(2)}
\approx 270$ keV. The intensities radiated into the fundamental and the third harmonics are equal to $1.10\times 10^{5}$ sr$^{-1}$ and $7.12\times 10^{3}\ \mathrm{sr}^{-1}$, respectively. Their ratio is approximately equal to $p^{-4}$ which is also in accordance with general theory.
The intensity of the AIR gradually increases with the AW amplitude. More precisely, $\d E_{\rm AIR} \propto p^2\propto (a/d)^2$ (in the case $p<1$), and this tendency one can observe comparing the AIR peaks in the graphs corresponding to $a/d=4,6,8,10$.
It is important to note that the positions of sharp AIR peaks, their narrow widths, and the radiated intensity are, practically, insensitive to the choice of the approximation used to describe the interplanar potential. In addition, provided the condition (\[1\]) is fulfilled, the AIR peaks are well separated (in the photon energy scale) from the peaks of the channeling radiation. Therefore, if being interested in the spectral distribution of the AIR only, one may disregard the channeling oscillations and to assume that the projectile moves along the centerline of the bent channel [@JPG; @laser]. The above statements are illustrated by where we compare the results of calculations of the total spectrum (\[3\]) in vicinity of the first harmonic of AIR in the case $a/d=10$. All parameters are the same as in . The filled and open circles represent the results of evaluation of the right-hand sides of (\[1\]) and (\[3\]) accompanied by numerical solution of the equations of motion for the projectile within the Molière (filled circles) and the harmonic (open circles) approximations for the interplanar potential. The solid line corresponds to the AIR radiation only (in this case the numerical procedures are simplified considerably, leading to the reduction, by orders of magnitude, of the CPU time). It is clearly seen that the more sophisticated treatment almost does not change the profile of the peak obtained by means of simple formulae describing purely AIR radiation [@JPG; @laser]. Moreover, the minor changes in the position and the height of the peak can be easily accounted for within the framework of the cited formalism by introducing the effective undulator parameter [@losses; @new_full] and (in the case of the harmonic approximation) the effective undulator amplitude [@new_full].
Thus, in vicinity of the AIR peaks there is no coupling of the two mechanisms of the radiation. On the contrary, the AIR strongly affects the spectrum in the photon energy range corresponding to the channeling radiation. The discussion of this effect lies beyond the scope of this Letter and is presented in [@new_full]. Here we only mention that, as it is clearly seen from the graphs $a/d=6,8,10$ in , for large values of $a/d$ the peaks of the ordinary channeling radiation acquire additional structure: there appear sub-peaks separated (in the case of the forward emission) by the interval $\delta \omega = 2\omega^{(1)}$.
The presented results of the numerical calculations of the total spectrum of radiation formed in an acoustically bent crystal clearly demonstrate the validity of the statement made in [@JPG; @laser; @losses] that the AIR and the ordinary channeling radiation occur in essentially different ranges of the emitted photons energies, allowing, thus, to investigate the AIR properties separately from the ordinary channeling radiation.
The positions of the peaks of the AIR as well as their intensity can be easily varied by changing the energy of projectile, the crystal type, the type of crystallographic plane, and the parameters of periodic pattern of crystal bendings. This statement by no means is restricted to the case of the acoustically bent crystal, but has more general nature. The treatment and the main results presented above can be applied, in particular, to consider the undulator radiation in statically bent crystals [@laser; @Uggerhoj2000]. Whatever way is chosen to prepare periodically bent crystalline lattice it is worth investigating, both theoretically and experimentally, not only the spontaneous emission of the AIR-type, but the stimulated emission as well [@laser].
The authors express their gratitude to Professor E. Uggerh[ø]{}j for sending a copy of the paper [@Uggerhoj2000] prior to its publication. We want to thank the administrative staff of the Physics CIP Computer Cluster at University of Munich for the possibility of using some of their computers for numerical calculations. The research was supported by DFG, GSI, and BMBF. AVK and AVS acknowledge the support from the Alexander von Humboldt Foundation.
References {#references .unnumbered}
==========
[99]{} Korol A V, Solov’yov A V, and Greiner W 1998 [*J.Phys.G.: Nucl. Part. Phys.*]{} [**24**]{} L45 Korol A V, Solov’yov A V, and Greiner W 1999 [*Int. J. Mod. Phys.*]{} E [**8**]{} 49 Mikkelsen U and Uggerh[ø]{}j E 2000 [*Nucl. Inst. and Meth.*]{} B [**160**]{} 435 Kumakhov M A 1976 [*Phys. Lett.*]{} A[**57**]{} 17 Zhang Q-R 1999 [*Int. J. Mod. Phys.*]{} E [**8**]{} 493 Madey J M J 1971 [*J.Appl.Phys.*]{} [**42**]{} 1906 Korol A V, Solov’yov A V, and Greiner W 2000 [*Int. J. Mod. Phys.*]{} E [**9**]{} 77 Taratin A M and Vorobiev S A 1988 [*Nucl. Inst. and Meth.*]{} B [**31**]{} 551; 1989 [*Nucl. Inst. and Meth.*]{} B [**42**]{} 41 Arutyunov V A, Kudryashov N A, Samsonov V M, and Strikhanov M N 1991 [*Zh. Tehn. Fiz.*]{} [**61**]{} 32; 1991 [*Zh. Tehn. Fiz.*]{} [**61**]{} 1; 1991 [*Nucl. Phys.*]{} B [**363**]{} 283 Krause W, Korol A V, Solov’yov A V, and Greiner W 2000 to be submitted to [*Phys. Rev.*]{} E Baier V N and Katkov V M 1967 [*Zh. Eksp. Teor. Fiz.*]{} [**53**]{} 1478 (Engl. transl. 1968 [*Sov. Phys. – JETP*]{} [**26**]{} 854); 1968 [*Zh. Eksp. Teor. Fiz.*]{} [**55**]{} 1542 (Engl. transl. 1969 [*Sov. Phys. – JETP*]{} [**28**]{} 807) Baier V N, Katkov V M, and Strakhovenko V M 1998 [*High Energy Electromagnetic Processes in Oriented Single Crystals*]{} (Singapore: World Scientific) Berestetskii V B, Lifshitz E M, and Pitaevskii L P 1982 [*Quantum Electrodynamics*]{} (Oxford: Pergamon) Tsyganov E N 1976 [*Fermilab preprint*]{} TM-682, 684 (Batavia) Gemmell D S 1974 [*Rev. Mod. Phys.*]{} [**46**]{} 129 Biruykov V M, Chesnokov Y A, and Kotov V I 1996 [*Crystal Channelling and its Application at High-Energy Accelerators*]{} (Berlin: Springer) Komarov F F 1979 [*Phys. Stat. Sol.*]{} (b) [**96**]{} 555 Sternheimer R M 1966 [*Phys. Rev.*]{} [**145**]{} 247 Mason W P 1972 Acoustic properties of Solids [*American Institute of Physics Handbook*]{} 3rd edn (New York: McGraw-Hill) Lindhard J 1965 [*Kong. Danske Vid. Selsk. mat.-fys. Medd.*]{} [**34**]{} 14 Uggerh[ø]{}j E 1993 [*Radiation Effects and Defects in Solids*]{} [**25**]{} 3 Kumakhov M A and Komarov F F 1989 [*Radiation From Charged Particles in Solids*]{} (New York: American Institute of Physics)
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Daniel Bhatti
- Joachim von Zanthier
- 'Girish S. Agarwal'
title: Superbunching and Nonclassicality as new Hallmarks of Superradiance
---
Introduction {#introduction .unnumbered}
============
Dicke [@Dicke(1954)] predicted that if an ensemble of two-level atoms is prepared in a collective state where half of the atoms are in the excited state and half of the atoms are in the ground state the spontaneous emission is proportional to the square of the number of atoms as if the particles would radiate coherently in phase like synchronized antennas [@Eberly(1975)]. To analyze the phenomenon Dicke introduced the concept of collective spins where $N$ two level atoms are described by the collective spin eigenstates $\ket{N/2, M}$, with $M$ running from $M = -N/2, \ldots, +N/2$ in steps of unity. Among these states the state $\ket{N/2, 0}$ radiates with an intensity $N^2$ times as strong as that of a single atom. The origin of superradiance is difficult to see since all states $\ket{N/2, M}$ exhibit no macroscopic dipole moment whereas such a dipole moment is commonly assumed to be required for a radiation rate proportional to $N^2$. The reason is that the Dicke states display strong quantum entanglement. The entangled character of the states is particularly apparent for the case of two two-level atoms where the individual atomic states are labeled by $\ket{e_l}$ and $\ket{g_l}$, $l = 1, 2$, for the excited and ground state, respectively. In this case the Dicke state $\ket{1,0} = 1/\sqrt{2} \left( \ket{e_1, g_2} + \ket{g_1, e_2} \right)$, also known as the Bell state or the EPR state, is clearly maximally entangled. For three atoms one of the Dicke states is denoted by $\ket{3/2,-1/2}$, which in current language would be the $W$ state $1/\sqrt{3} \left( \ket{e_1, g_2, g_3} + \ket{g_1, e_2, g_3} + \ket{g_1, g_2, e_3} \right)$ [@Dur(2000)]. The single excited generalized $W$ state, where only one atom is excited and $N-1$ atoms are in the ground state, is also known to be fully entangled and plays a particularly important role for single-photon superradiance [@Scully(2006); @Scully(2007); @Scully(2008); @Scully(2009); @Scully(2009)Super]. In fact, it has been recognized that most of the important aspects of superradiance [@Dicke(1954); @Eberly(1970); @Rehler(1971); @AGARWAL(1974); @Haroche(1982)] can be studied by examining samples in single excited generalized $W$ states [@Nienhuis(1987); @Scully(2006); @Scully(2007); @Scully(2009)Super; @Scully(2009); @Svidzinsky(2010); @Wiegner(2011)PRA; @Kaiser(2012); @Kaiser(2013); @Rohlsberger(2013)] as the emission from these states possesses all the features of superradiance that originally were calculated for samples with an arbitrary number of excitations [@Scully(2007); @Wiegner(2011)PRA]. The spatial features of one photon superradiance have been extensively studied for example from the perspective of *timed Dicke states* [@Scully(2006); @Scully(2007); @Scully(2009)Super] and also the spectral and temporal aspects have been investigated in a large variety of systems [@Scully(2009); @Svidzinsky(2010); @Cleland(2010); @Simmonds(2011); @Nori(2011); @Nori(2012); @Rohlsberger(2013)]. Note that a number of recent works [@Thiel(2007); @Vitanov(2008); @Vitanov(2008)a; @Zeilinger(2009); @Weinfurter(2009); @Wineland(2009); @Lemr(2009)] have also discussed how single excited generalized $W$ states for a small number of atoms can be prepared in the laboratory.
The single excited generalized $W$ state does however not allow one to study the quantum statistical properties of superradiance. In order to explore these aspects the system must emit at least two photons. Only then one has access to the photon-photon correlations which display amongst others the particular quantum characteristics of the spontaneously scattered radiation [@Glauber(1963)Quantum; @Mandel(1977); @WALLS(1994)]. To this end it is required to investigate what we will term two-photon superradiance from generalized $W$ states with $n_{e}\geq2$ excitations. In the present paper we show that in two-photon superradiance the emitted radiation can exhibit both bunched as well as nonclassical and antibunched light depending on the angle of observation, i.e., the position of the detectors collecting the scattered photons, and on the particular $W$ state, i.e., the number of atoms $N$ and the number of excitations $n_{e}$, considered. In particular, in certain cases it is also possible to observe the phenomenon of superbunching, i.e., photon-photon correlations larger than those maximally measurable for classical light sources. In all the cases the mean intensity displays the familiar features of superradiance produced by the corresponding $W$ state. While we derive our results for two-photon superradiance for arbitrary generalized $W$ states we focus in this paper on systems in doubly excited $W$ states; the outcomes for arbitrary $W$ states with more than two excitations are presented in the Supplementary Information section.
Note that bunching in the radiation of generalized $W$ states can be explained semi-classically. However, the phenomenon of superbunching as well as the emission of nonclassical light, first demonstrated in 1977 [@Mandel(1977)], can only be understood in a quantum mechanical description [@Mandel(1977); @Paul(1982)]. The latter is a feature arising from the light’s particle nature where photon fluctuations become smaller than for coherent light. Demonstrating nonclassicality in the light of arbitrary $W$ states thus directly leads to a manifestation of the particular quantum mechanical characteristics of these superradiant states. We finally discuss also the spatial cross correlations of photons in two-photon superradiance. Here, likewise, superbunching and nonclassicality can be observed.
We note that recent experiments by Jahnke et al. with quantum dots in a cavity have already reported the observation of superbunching [@Auffeves(2011); @Jahnke(tbp)]. In this paper we bring out for a simple model system in free space the reasons for the appearance of this phenomenon, a curio which does not commonly occur, the squeezed vacuum being one of the rare examples [@Boitier(2011)].
Results {#sec:system .unnumbered}
=======
To focus on the key aspects of two-photon superradiance we consider a linear system of $N$ equidistantly aligned identical emitters, e.g., atoms or ions with upper state $\ket{e_{l}}$ and ground state $\ket{g_{l}}$, $l = 1, \ldots , N$, trapped in a linear arrangement [@Meschede(2006); @Wineland(2008); @Blatt(2011); @Rauschenbeutel(2014)] at positions ${{\mathbf{R}}}_{l}$ with spacing $d \gg \lambda$ such that the dipole-dipole coupling between the particles can be neglected (see Fig. \[fig:detection\_scheme\]). The atoms are assumed to be prepared initially in a generalized $W$ state with $n_{e}$ excitations, i.e., in the state $\ket{W_{n_{e},N}}=\binom{N}{n_{e}}^{-\frac{1}{2}}\sum_{\{\alpha_{l}\}=\mathcal{P} {\{l\}}} \prod_{i=1}^{n_{e}}\ket{e_{\alpha_{i}}}\prod_{i=n_{e}+1}^{N}\ket{g_{\alpha_{i}}} $, where $\mathcal{P} {\{l\}}$ denotes all permutations of the set of atoms $\{l\}=\{1,2,\hdots,N\}$. In what follows we study the second order correlation functions at equal times emitted by the atoms in the described $W$ state. To this end two detectors are placed at positions ${\mathbf{r}}_{1}$ and ${\mathbf{r}}_{2}$ in the far field each measuring a single photon coincidentally, i.e., within a small time window much smaller than the lifetime of the upper state. To simplify the calculations we suppose that the emitters and the detectors are in one plane and that the atomic dipole moments of the transition $\ket{e_{l}} \rightarrow \ket{g_{l}}$ are oriented perpendicular to this plane (see Fig. \[fig:detection\_scheme\]).
![Scheme of considered setup: $N$ two-level atoms are aligned on the x-axis, whereby neighboring atoms are separated by a distance $d$. The intensity in the far field $I({\mathbf{r}}_{1})$ is measured by a single detector at position ${\mathbf{r}}_{1}$, whereas the second order correlation function $G^{(2)}({\mathbf{r}}_{1},{\mathbf{r}}_{2})$ is measured by two detectors at ${\mathbf{r}}_{1}$ and ${\mathbf{r}}_{2}$.[]{data-label="fig:detection_scheme"}](superbunching_antibunching_8-figure0){width="55.00000%"}
Due to the far field condition and therefore the inability to identify the individual photon sources, the electric field operator at ${{\mathbf{r}}}_j$ takes the form [@Thiel(07)] $\left[ \hat{E}^{(-)}({\mathbf{r}}_j) \right]^\dagger \! = \hat{E}^{(+)}({\mathbf{r}}_j) \sim \sum_{l=1}^{N} e^{-i\,\varphi_{lj}} \;\hat{s}^{-}_l = \sum_{l=1}^{N} e^{-i\,l\,\delta_{j}} \;\hat{s}^{-}_l$, where $\hat{s}_{l}^{-}=\ket{g_{l}}\!\bra{e_{l}}$ is the atomic lowering operator for atom $l$, and $\varphi_{lj} = - k \, \frac{{\mathbf{r}}_j\cdot{\mathbf{R}}_l}{r_j} = l\, k\, d \, \sin\theta_{j} \cos\phi_j =l\,\delta_{j}$ the relative optical phase accumulated by a photon emitted by source $l$ and recorded by detector $j$ with respect to a photon emitted at the origin. Note that the field operators have been chosen dimensionless as all dimension defining prefactors cancel out in the normalized correlation functions. The first and second order spatial correlation functions at equal times are defined as [@Glauber(1963)Quantum] $G^{(1)}({\mathbf{r}}_{1})=\left<\hat{E}^{(-)}({\mathbf{r}}_{1})\hat{E}^{(+)}({\mathbf{r}}_{1})\right>$ and $G^{(2)}({\mathbf{r}}_{1},{\mathbf{r}}_{2})=\left<\hat{E}^{(-)}({\mathbf{r}}_{1})\hat{E}^{(-)}({\mathbf{r}}_{2})\hat{E}^{(+)}({\mathbf{r}}_{2})\hat{E}^{(+)}({\mathbf{r}}_{1})\right> $, respectively, where $G^{(1)}({\mathbf{r}}_{1})$ is proportional to the mean intensity of the emitted radiation, i.e., $G^{(1)}({\mathbf{r}}_{1})\sim I({\mathbf{r}}_{1}) $. To compare the photon statistics of various systems radiating with different intensities we further introduce the normalized second order correlation function [@Glauber(1963)Quantum] $g^{(2)}({\mathbf{r}}_{1},{\mathbf{r}}_{2})=G^{(2)}({\mathbf{r}}_{1},{\mathbf{r}}_{2})/(G^{(1)}({\mathbf{r}}_{1})\,G^{(1)}({\mathbf{r}}_{2}))$. For the state $\ket{W_{n_{e},N}}$ the first order correlation function in the configuration of Fig. \[fig:detection\_scheme\] has been calculated [@Wiegner(2011)PRA] to $$\begin{aligned}
G^{(1)}_{n_{e},N}(\delta_{1})&= \frac{n_{e}(n_{e}-1)}{N-1} + \frac{N\, n_{e}(N-n_{e})}{N-1} \chi^{2}(\delta_{1}) \ ,
\label{eq:intdefinition}
\end{aligned}$$ where $\chi(x)=\frac{\sin(\frac{Nx}{2})}{N\sin(\frac{x}{2})}$ corresponds to the normalized far-field intensity distribution of a coherently illuminated $N$-slit grating [@BORN(1999)]. As it is well-known from classical optics, the distribution $\chi^{2}(x)$ is strongly peaked in particular directions. The fact that $\chi^{2}(x)$ appears in the context of spontaneous emission of atoms in generalized $W$ states as in Eq. (\[eq:intdefinition\]) has been coined by Dicke *spontaneous emission of coherent radiation* or simply *superradiance* [@Dicke(1954)]. Note that even though $\chi^{2}(x)$ displays pronounced maxima in certain directions the distribution assumes also very small values and even vanishes in other directions what has been interpreted as *subradiance* of the states $\ket{W_{n_{e},N}}$ [@Wiegner(2011)PRA].
From Eq. (\[eq:intdefinition\]) we find that the intensity distribution of the state $\ket{W_{1,N}}$ with only one excitation ($n_{e}=1$) simplifies to $G^{(1)}_{1,N}(\delta_{1})=N\chi^{2}(\delta_{1})$, with $G^{(1)}_{1,N}(\delta_{1}) \rightarrow N \ \ \text{for} \ \delta_{1}=0$, displaying a maximal visibility $\mathcal{V}=1$ and a peak value $N$ times the intensity of a single atom. This kind of single photon superradiance has been extensively studied in the past [@Scully(2006); @Scully(2007); @Scully(2009)Super; @Scully(2009); @Svidzinsky(2010); @Wiegner(2011)PRA; @Kaiser(2012); @Kaiser(2013); @Rohlsberger(2013)]. Its particular superradiant characteristics have been shown to result from quantum path interferences occurring due to the particular interatomic correlations of the collective state $\ket{W_{1,N}}$ [@Wiegner(2011)PRA].
As discussed below the strong correlations of the states $\ket{W_{n_{e},N}}$ may lead to photon-photon correlations with $g_{n_{e},N}^{(2)}(\delta_{1},\delta_{1})> 2$ as well as $g_{n_{e},N}^{(2)}(\delta_{1},\delta_{1})<1$, corresponding to superbunched as well as nonclassical light, respectively. Note that from the form of normalized second order correlation function $g_{n_{e},N}^{(2)}(\delta_{1},\delta_{2})$ it is obvious that bunching necessitates small intensities, i.e., $G^{(1)}(\delta_{1}) \ll G^{(2)}(\delta_{1},\delta_{1})$, whereas nonclassical light requires small values of the two-photon correlation function, i.e., $G^{(2)}(\delta_{1},\delta_{1})\ll G^{(1)}(\delta_{1})$.
In what follows we study two-photon superradiance for the simplest form of generalized $W$ states, i.e., $W$ states with only two excitations, as the main features of two-photon superradiance can already be observed for this configuration; two-photon superradiance for arbitrary generalized $W$ states $\ket{W_{n_{e},N}}$ with $n_{e} > 2$ is discussed in the Supplementary Information section.
Note that an atomic ensemble in the state $\ket{W_{n_{e}=2,N}}$ can be prepared for example by using photon pairs generated in a down conversion process. When sent on a $50:50$ beam splitter the photon pair would then produce two photons at either of the two output ports of the beam splitter, i.e., if one port delivers zero photons then the other one has two photons [@HongOuMandel(1987)]. Assuming perfect detection efficiency, a photon pair not registered at one output port of the beam splitter and not registered at the other one after having passed the atomic ensemble would then herald the absorption of two photons by the atomic system.
For $n_{e}=2$ the normalized second order correlation function for the considered configuration of Fig. \[fig:detection\_scheme\] takes the form $$\begin{aligned}
& g_{2,N}^{(2)}(\delta_{1},\delta_{2})= \frac{N(N-1)}{2} \frac{\left( N \chi(\delta_{1})\chi(\delta_{2}) - \chi(\delta_{1}+\delta_{2}) \right)^{2}}{\left(1 + N(N-2)\chi^2(\delta_{1})\right) \left(1 + N(N-2)\chi^2(\delta_{2})\right)} \ .
\label{eq:Gcompletene2}
\end{aligned}$$ This expression will be investigated in detail in the following subsections.
Superbunching in Two-photon Superradiance {#sec:ne2d1d1 .unnumbered}
-----------------------------------------
In this section we investigate whether bunching in two-photon superradiance, in particular the phenomenon of superbunching with $g_{n_{e},N}^{(2)}(\delta_{1},\delta_{2})> 2$, can be observed in the radiation produced by states of the form $\ket{W_{n_{e}=2,N}}$. We start to explore the photon-photon correlations with the two detectors placed at equal positions, i.e., with the two spontaneously emitted photons recorded in the same mode; photon-photon cross correlations are studied thereafter.
![Second order correlation function $g^{(2)}_{2,N}(\delta_{1},\delta_{1})$ (left) and first order correlation function, i.e., the intensity distribution, $G^{(1)}(\delta_{1})$ (right) of the radiation emitted by $N$ two-level atoms in the doubly excited $W$ state $\ket{W_{2,N}}$ for $N=3$ (dotted), $N=4$ (dashed), $N=6$ (solid). Superbunching is observed at $\delta_{1}=\pi$ for $N \geq 4$, while antibunching occurs at detector positions $\delta_{1}$ fulfilling $N \chi^{2}(\delta_{1})=\chi(2\delta_{1})$. For increasing $N$ superbunching becomes stronger, similar to single photon superradiance of the state $\ket{W_{1,N}}$. Comparing the two plots one can see that for high values of $G^{(1)}(\delta_{1})$ small values of $g^{(2)}_{2,N}(\delta_{1},\delta_{1})$ are obtained and vice versa.[]{data-label="fig:ne2d1d1"}](ne2d1d1_small){width="80.00000%"}
According to Eq. (\[eq:Gcompletene2\]) the second order correlation function for the state $\ket{W_{2,N}}$ with two detectors placed at the same position takes the form $$\begin{aligned}
g^{(2)}_{2,N}(\delta_{1},\delta_{1})= \frac{ N(N-1)\, \left( N\,\chi^{2}(\delta_{1})-\chi(2\delta_{1}) \right)^{2}}{2\, \left( 1 + N(N-2)\, \chi^{2}(\delta_{1}) \right)^{2}} \ .
\label{eq:g2ne2d1d1}
\end{aligned}$$
In order to access whether the system displays bunching for this configuration we have to search for values $g^{(2)}_{2,N}(\delta_{1},\delta_{1}) > 1$. Hitherto, we choose detector positions for which the values of the first order correlation function $G^{(1)}(\delta_{1})$ remain smaller than the unnormalized second order correlation function $G^{(2)}_{2,N}(\delta_{1},\delta_{1})$, i.e., locations where the intensity is low. This occurs for example at $\delta_{1}=\pi$, where $g^{(2)}_{2,N}(\delta_{1},\delta_{1})$ attains its maximal value (see Fig. \[fig:ne2d1d1\]). To investigate this outcome quantitatively we have to study the case of even and odd $N$ separately since $\chi(\pi)$ and $\chi(2\pi)$ yield different results in these two cases [@note(1)]. For an even number $N$ of atoms one obtains the following two identities $\chi(\pi)=0$, $\chi(2\pi)=-1$, from which we deduce $g^{(2)}_{2,N_{even}}(\pi,\pi)=N(N-1)/2$, leading to bunched light in case that $N_{even}\geq 4$ (see Fig. \[fig:ne2d1d1\]). In fact, for $N_{even}\geq 4$, we even obtain superbunching, i.e., $g^{(2)}_{2,N_{even}}(\pi,\pi) > 2$, what surpasses the maximum value achievable with classical light. Note that $g^{(2)}_{2,N_{even}}(\pi,\pi)$ as a function of $N$ has in principal no upper limit as it increases $\sim N^{2}$ for $N\gg1$. This means that we can produce principally unlimited values of superbunching if we add more and more atoms in the ground state to the system [@Auffeves(2011); @Jahnke(tbp)].
In case of odd $N$ the above identities read $\chi(\pi)=\pm 1/N$, $\chi(2\pi)= +1$. what leads to the a maximal value of the second oder correlation function of $g^{(2)}_{2,N_{odd}}(\pi,\pi)=N(N-1)/8$. Here we obtain bunched and superbunched light in case that $N_{odd} \geq 5$. Again, as in the case of even $N$, there is no upper limit for the maximum value of the two-photon correlation function, as once more we have $g^{(2)}_{2,N_{odd}}(\pi,\pi) \sim N^{2}$.
Nonclassicality in Two-photon Superradiance {#sec:ne2d1d1anti .unnumbered}
-------------------------------------------
In this section we investigate whether for the initial state $\ket{W_{2,N}}$ and two detectors at equal positions we can obtain nonclassical light, a sub-Poissonian photon statistics and antibunching in two-photon superradiance. As known from the radiation of a single atom [@Mandel(1977); @Paul(1982)] a sub-Poissonian photon statistics derives from the discrete nature of the scattered radiation and can be explained only in a quantum mechanical treatment of resonance fluorescence. A complete vanishing of the second order correlation function at equal times for the state $\ket{W_{2,N}}$ indicates that $g^{(2)}_{2,N}(\delta_{1},\delta_{1}) = 0$ at $\tau = 0$ what proves true antibunching, whereas values $g^{(2)}_{2,N}(\delta_{1},\delta_{1}) < 1$ result from the nonclassical nature of the radiation scattered by the state $\ket{W_{2,N}}$.
Noting that the second order correlation function $g^{(2)}_{2,N}(\delta_{1},\delta_{1})$ displays a visibility of $\mathcal{V}=1$ (see Fig. \[fig:ne2d1d1\]) there must be indeed detector positions where $g^{(2)}_{2,N}(\delta_{1},\delta_{1}) = 0$. More specifically, according to Eq. ($\ref{eq:g2ne2d1d1}$), the photon-photon correlation function vanishes independently of the atom number $N$ in case that the detectors are located at positions $\tilde{\delta}_{1}=a\frac{2\pi}{N}$, with $ a = 1,2, \ldots < N/2$, as in this case $\chi(\tilde{\delta}_{1})=\chi(2 \tilde{\delta}_{1})=0$. Towards these positions the atomic system thus radiates photons which display complete antibunching [@note(2)]. If the two detectors are located at $\delta_{1}=0$ we have $g^{(2)}_{2,N}(\delta_{1},\delta_{1}) < 1$ as long as $N\geq 3$, as in this case the photon-photon correlation function takes the form $g^{(2)}_{2,N}(0,0)=N/(2(N-1))$ (cf. Eq. (\[eq:g2ne2d1d1\])). Hence, in those directions the atomic system emits nonclassical light with photon number fluctuations smaller than those for coherent light.
![Left: Second order correlation function $g_{n_{e},N}^{(2)}(\delta_{1},\delta_{2})$ for $n_{e}=2$, $\delta_{2}=-\delta_{1}$ and $N=3$ (dotted), $N=4$ (dashed), $N=6$ (solid). The function displays both superbunching and antibunching with a maximal visibilities of $\mathcal{V}=1$. It can be seen that the superbunching effect increases with increasing $N$, while for $\delta_{1}=0$ the second order correlation function converges to $1/2$. Right: First order correlation function $G^{(1)}_{2,N}(\delta_{1})$ for $N=3$ (dotted), $N=4$ (dashed), $N=6$ (solid). From the figure it can be seen that for small values of $G^{(1)}_{2,N}(\delta_{1})$ high values of $g_{n_{e},N}^{(2)}(\delta_{1},\delta_{2})$ are obtained. In contrast to the case $\delta_{1}=\delta_{2}$ several superbunching peaks occur.[]{data-label="fig:ne2d1minusd1"}](ne2d1minusd1_small){width="80.00000%"}
Two-photon Cross Correlations {#sec:crosscorrelations .unnumbered}
-----------------------------
Finally we study the spatial cross correlations in two-photon superradiance for atoms in the state $\ket{W_{2,N}}$, i.e, the behavior of the second order correlation function $g^{(2)}_{2,N}(\delta_{1}, \delta_{2})$ in case that the scattered photons are recorded at different positions $\delta_{1} \neq \delta_{2}$. We start to explore the particular configuration of counter-propagating detectors, i.e., detectors at positions $\delta_{1}= - \delta_{2}$, followed by the case where $\delta_{2}$ is fixed and only $\delta_{1}$ is varied.
For $\delta_{1}= -\delta_{2}$ (i.e., where we choose $\theta_1 = \theta_2$, $\phi_2 = 0$ and $\phi_1 = \pi$), the second order correlation function reads (cf. Eq. (\[eq:Gcompletene2\])) $$g^{(2)}_{2,N}(\delta_{1},-\delta_{1})= \frac{N (N-1)\, \left(N\, \chi^{2}(\delta_{1}) - 1 \right)^{2} }{ 2\, \left( 1 + N(N-2)\, \chi^{2}(\delta_{1}) \right)^{2} } \ .
\label{eq:g2ne2d1minusd1}$$ To determine the possibilities for superbunching in cross correlations of the scattered photons we look for the maximum of Eq. (\[eq:g2ne2d1minusd1\]). This is attained at $\bar{\delta}_{1}=a\frac{2\pi}{N}$, with $a = 1, \ldots, N/2$, in which case $\chi(\delta_{1})$ vanishes and the second order correlation function reads $g^{(2)}_{2,N}(\bar{\delta},-\bar{\delta})=N(N-1)/2$. This result is principally identical to $g^{(2)}_{2,N_{even}}(\pi,\pi)$. However, in the case of counter-propagating detectors it is valid for arbitrary $N$, i.e., for $N$ even or odd. Here, the threshold for superbunching is exceeded if $N\geq 3$ (see Fig. \[fig:ne2d1minusd1\]). In the same configuration also complete antibunching can be obtained. This occurs for $\hat{\delta}_{1}$ such that $\chi^{2}(\hat{\delta}_{1}) = 1/N$ (cf. Eq. (\[eq:g2ne2d1minusd1\])), in which case the photon correlation function vanishes identically, i.e., $g^{(2)}_{2,N}(\hat{\delta}_{1},-\hat{\delta}_{1})= 0$.
Another interesting configuration is the case when $\delta_{2}$ is fixed and only $\delta_{1}$ is varied. If we fix $\delta_{2}=0$ the photon-photon cross correlation function takes the form (cf. Eq. (\[eq:Gcompletene2\])) $$g^{(2)}_{2,N}(\delta_{1},0) = \frac{N(N-1)\chi^{2}(\delta_{1})}{2 + 2N(N-2)\chi^{2}(\delta_{1})} \ ,
\label{eq:g2ne2delta10}$$ which can maximally take a value of $1$ (for $N=2$ and $\delta_{1} = 0$), whereas for $N > 2$ the cross correlation function remains always $< 1$. In Fig. \[fig:ne2d10\] the corresponding photon-photon cross correlations are shown for $N = 2, 4, 6$ for the entire range $\delta_{1} \in [0, \pi]$. It can be seen that by increasing $N$ the second order correlation function $g^{(2)}_{2,N}(\delta_{1},0)$ decreases in absolute values. The reason is the following: since $G^{(2)}_{2,N}(\delta_{1},0)$ and $G^{(1)}_{2,N}(\delta_{1})$ depend on $\chi(\delta_{1})$ in a similar way the overall behavior of $g^{(2)}_{2,N}(\delta_{1},0)$ is determined by the prefactors of $\chi(\delta_{1})$ in the numerator and denominator of $g^{(2)}_{2,N}(\delta_{1},0)$, the ratio of which decreases with increasing $N$ and converges to $1/2$ for $N \gg 1$.
![Left: Second order correlation function $g_{n_{e},N}^{(2)}(\delta_{1},\delta_{2})$ for $n_{e}=2$ where one detector is fixed at $\delta_{2}=0$. For all $N$ (in the plot: $N=2$ (solid), $N=4$ (dashed), $N=6$ (dotted)) the two-photon correlation function remains smaller than one for any $\delta_{1}\neq 0$. For $\delta_{1}=0$ the maximal value of one is attained only in the case $N=2$. Right: First order correlation function $G^{(1)}_{2,N}(\delta_{1})$ for $N=2$ (solid), $N=4$ (dashed), $N=6$ (dotted).[]{data-label="fig:ne2d10"}](ne2d10_small){width="80.00000%"}
Discussion {#discussion .unnumbered}
==========
In conclusion we investigated for a prototype ensemble of $N$ identical non-interacting two-level atoms prepared in collective superradiant generalized $W$-states with $n_{e}$ excitations the particular quantum statistical properties of the emitted radiation. Such investigations require the collective system to have at least two excitations as we explore the photon-photon correlations of the scattered light. We derived conditions for which the atomic system emits bunched and even superbunched light, as well as nonclassial and antibunched radiation. Here, superbunching refers to values of the normalized second order correlation function $g^{(2)}_{2,N}(\delta_{1}, \delta_{1}) > 2$ and antibunching to values $g^{(2)}_{2,N}(\delta_{1}, \delta_{1}) = 0$. In some cases the results were obtained under the condition that the number of atoms in the ensemble exceed a certain threshold. For example, the smallest number of atoms producing superbunching in the state $\ket{W_{2,N}}$ is $N=3$; similar results were derived for $N$ atoms with arbitrary number of excitations, as shown in the Supplementary Information section. Note that in coherently driven atomic systems superbunching can be observed already for $N = 2$. For example, in [@Skornia(2001)] it was demonstrated that arbitrarily high values of $g^{(2)}_{N = 2}(\delta_{1}, \delta_{1})$ can be produced for $\delta_{1} = \pi$ in case that the two atoms are very weakly excited, as in this case $g^{(2)}_{N = 2}(\pi, \pi) \sim 1/\Omega^4$, where $\Omega$ is the Rabi frequency (in units of the spontaneous decay rate $\gamma$). The effect is even stronger if the two atoms are subject to a strong dipole-dipole interaction as in dipole blockade systems; here $g^{(2)}_{N = 2}(\pi, \pi) \sim \delta_0^2/\Omega^4$ where $\delta_0$ is the level shift of the doubly excited state (again in units of $\gamma$) [@Bastin(2010)]. In the last part of the paper we finally investigated the spatial cross correlations in two-photon superradiance, i.e., the second order correlation functions $g^{(2)}_{2,N}(\delta_{1}, \delta_{2})$ for detector positions $\delta_{1} \neq \delta_{2}$. Here, again, positions were found where $g^{(2)}_{2,N}(\delta_{1}, \delta_{2}) > 2$ and $g^{(2)}_{2,N}(\delta_{1}, \delta_{2}) = 0$, corresponding in this case to superbunching and antibunching of the cross correlations of the scattered photons.
[10]{} url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
. ** ****, ().
& ** (, ).
, & . ** ****, ().
, , & . ** ****, ().
. ** ****, ().
, & . ** ****, ().
. ** ****, ().
& . ** ****, ().
& . ** ****, ().
& . ** ****, ().
. In **, vol. of **, (, ).
& . ** ****, ().
& . ** ****, ().
, & . ** ****, ().
, & . ** ****, ().
, & . ** ****, ().
, , & . ** ****, ().
. ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
, & . ** ****, ().
, , , & . ** ****, ().
, , , & . ** ****, ().
& . ** ****, ().
& . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
, , & . ** ****, ().
& . ** ****, ().
. ** ****, ().
, & . ** ****, ().
& ** (, ).
. ** ****, ().
, , , & . ** ****, ().
*et al.* ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
& . ** ****, ().
*et al.* . ** ****, ().
, , , & . ** ****, ().
*et al.* . ** ****, ().
& ** (, ).
, & . ** ****, ().
, , , & . ** ****, ().
, & . ** ****, ().
Acknowledgements {#acknowledgements .unnumbered}
================
GSA and JvZ thank Frank Jahnke for fruitful discussions and the disclosure of documents prior to publication. DB and JvZ warmly thank the hospitality of GSA at various stays at the Oklahoma State University. DB thanks the Friedrich-Alexander Universität Erlangen-Nürnberg for a travel grant “Hochschule International” and the German Academic Exchange Service for a PROMOS travel grant. The authors gratefully acknowledge funding by the Erlangen Graduate School in Advanced Optical Technologies (SAOT) by the German Research Foundation (DFG) in the framework of the German excellence initiative.
Author contributions statement {#author-contributions-statement .unnumbered}
==============================
DB did all the calculations while the ideas were formulated by GSA. JvZ helped with thoughts in possible experimental implementation. JvZ primarily wrote the manuscript for publication which was reviewed by both DB and GSA.
Additional information {#additional-information .unnumbered}
======================
The authors declare no competing financial interests.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We discuss the design and operation of a confocal rheometer, formed by integrating an Anton Paar MCR301 stress-controlled rheometer with a Leica SP5 laser scanning confocal microscope. Combining two commercial instruments results in a system which is straightforward to assemble that preserves the performance of each component with virtually no impact on the precision of either device. The instruments are configured so that the microscope can acquire time-resolved, three-dimensional volumes of a sample whose bulk viscoelastic properties are being measured simultaneously. We describe several aspects of the design and, to demonstrate the system’s capabilities, present the results of a few common measurements in the study of soft materials.'
author:
- 'S. K. Dutta'
- 'A. Mbi'
- 'Richard C. Arevalo'
- 'Daniel L. Blair'
title: Development of a confocal rheometer for soft and biological materials
---
Introduction
============
Complex fluids exhibit unique mechanical properties that are determined by the physicochemical details of their constituent components. One common example is a colloidal dispersion, where nano- to micrometer sized particles are suspended within a fluid – [ *e.g.*]{} milk, paint and blood are all colloidal dispersions. The [ *structuring*]{} of the fluid has two predominant effects: an enhancement of the fluid viscosity and, when the concentration of the dispersed material is high enough, the appearance of a frequency dependent elastic modulus. These characteristics are determined using rheology, the measurement science of quantifying the response of fluid based materials to an applied stress or strain.
Rheology techniques are classified into two main types known as bulk and micro, each having specific benefits and limitations. Bulk measurements often require relatively large sample volumes and, depending on the measuring system, can suffer from a somewhat limited range due to the inherent inertia of the tooling. The main advantage of bulk rheology is that the highly nonlinear mechanical behavior of soft materials can be directly determined. Passive microrheology utilizes the energy spectrum of thermal fluctuations and provides very localized structural information over a tremendous dynamical range that is limited only by the acquisition rate of the measurement.[@Dasgupta:2002uw] The primary drawbacks of microrheology are understanding the implications of incorporating tracer particles and the limitations set by the magnitude of the thermal fluctuations; effectively, microrheology is limited to nearly homogeneous systems with very small elastic moduli. Overall, rheology is extremely powerful as a characterization tool for a broad class of biologically derived or chemically synthesized materials. However, in many instances where bulk- and micro- rheology are applied, information about the role of structure, either inherent or influenced by boundary conditions, are essentially unknowable. Limited access to structure results in a great deal of uncertainty about the microscopic origins of the mechanical response.
The first instruments specifically developed for optically quantifying the structural response of complex fluids to an externally applied shear stress were based on X-ray and neutron scattering. [@Panine:2003bq; @Sasa:2010hb; @Porcar:2011gc] Scattering methods are particularly powerful for investigating average structural changes, such as conformational changes in protein networks[@Weigandt:2009je] and the bulk phase behavior of worm-like micelles.[@Liberatore:2006dq] If the material is inherently disordered, which is the case for most soft and biological materials, scattering can only provide spatially averaged information that generally precludes details about localized structural rearrangements driven by thermal excitations and external stresses. A natural extension of the scattering approach is to directly measure the real space structural response of complex fluids under shear through the use of optical microscopy. This need has driven the development of new optical-rheology platforms with ever increasing sophistication and versatility.[@Bender:1995dr; @Cohen:2004hp; @Besseling09a; @Basu11a; @Paredes11a; @Cheng11a; @Boitte:2013dn] Access to time-resolved, three-dimensional information is crucial for an accurate quantification of the microscopic structure that ultimately determines material properties; connecting macroscopic observables, such as shear and bulk moduli, to the relevant physical interactions and structure is a cornerstone of modern materials science.
The instrumentation challenge remains clear and open: provide a measurement device that combines high resolution, high magnification, real space, time resolved spatial information in three dimensions that is coupled with simultaneous high resolution mechanical deformations that can be easily reproduced by research groups. Precision measurements of the structure and mechanical properties of soft materials requires sub-micrometer spatial resolution and nanonewton-meter torque resolution. Luckily, devices that independently attain these levels of precision are commercially available.
The technology of fast laser scanning and spinning disk confocal microscopy techniques (LSCM or SDCM) has matured dramatically over the past twenty years. Innovations made to confocal microscopy, through commercial and academic partnerships, are providing unprecedented gains in imaging resolution at ever increasing acquisition speeds and at steadily decreasing costs.[@Hell:2003if] Confocal microscopy has emerged as a powerful tool in soft materials physics,[@Lu:2008fo] as it provides three-dimensional reconstructions of structures at sub-micrometer resolution. The principles of confocal microscopy are straightforward; by discriminating out of focus light, sharp two-dimensional images are “stacked” in the third dimension, providing time resolved volumetric data. These image volumes can then be rapidly analyzed at the sub-voxel level using advanced processing techniques.[@Crocker96a; @MPT] There are many commercial implementations from all of the major microscopy companies, and a number of component built systems that are generally based on spinning disk platforms.
Concomitant to the advances of confocal microscopy, stress-controlled rheometer technology has also advanced, providing new standards for sensitivity and stability. Moreover, stress-controlled rheometers are ideally suited as development platforms due to the combination of low friction, feedback-controlled, inductive motors that provide precise torque and position encoders that measure displacements. In nearly all implementations, the the motor/encoder systems are integrated into the [*upper*]{} tool. This compact design is in contrast to strain-controlled platforms where one tool rotates to provide displacement, while the other tool responds to the stress that propagates through the sample. The single tool configuration provides the flexibility to incorporate versatile modifications of the static portion of the lower tooling.
What follows below is a detailed description of what is needed to reproduce our confocal rheometer system. A key component of our plan was to produce a functional device within a very short period of time with a limited amount of machine work. To attain this does require the use of a particular rheometer (Anton Paar MCR series) and therefore emulating this system with other devices may prove difficult. We do feel that this system can be reproduced using devices from other manufacturers if the specific guidelines we provide are transferred. We will discuss a number of design criteria that guided our development and a series of data that will help motivate the intended demand. We also quantify a few unavoidable limitations of these systems and confirm some of the capabilities through recent publications.[@Schmoller10a; @Arevalo11a; @Holland12a]
Design Principles
=================
The plates of shearing devices should remain parallel at a fixed gap throughout their entire range of motion, whether that corresponds to a full rotation for a rheometer or a maximum displacement for a linear shear stage. The limitations of these devices are set by the tolerances attainable through computer numerical control (CNC) machining methods; in most instances, CNC machining can attain $2\ {\ensuremath{\mu \mathrm{m}}}$ precision. Most modern rheometers are produced within these specifications and therefore can reproducibly attain gaps to within $20\ {\ensuremath{\mu \mathrm{m}}}$ while remaining functional and allowing easy interchange between tools. Tool runout and parallelity are dramatically compromised if the relative orientation of each tool is not maintained for all applications, leading to a reduction in the reliability of reported rheological data. Therefore, when developing a new rheological system, either with a top-down design using existing technology or a bottom-up design by assembling custom components, each new implementation must at least match these specifications.
We have chosen a top-down approach for constructing a confocal rheometer. The primary considerations for this decision were time and functionality. Having a working device within a year, from conception to implementation, was highly desirable and attainable. By using devices with guaranteed factory specifications, we could forgo years of engineering and benchmarking. Furthermore, commercial instruments potentially provide a more user friendly system, making training, operation, and collaboration simpler.
![\[FPhoto\] Photograph of the confocal rheometer. The base plate of an Anton Paar MCR301 rheometer is replaced with a metal cup. Optical access for a Leica SP5 confocal microscope is provided by a glass coverslip mounted in the cup, which serves as the rheometer bottom plate. The field of view of the microscope can be changed by moving the rheometer on a manual three-axis translation stage. The inset shows the device that clamps the cup to the microscope stand to reduce vibrations.](Figure1-eps-converted-to.pdf){width="3.0in"}
Our system, shown in Fig. \[FPhoto\], consists of an MCR301 stress-controlled rheometer from Anton Paar GmbH and an SP5 LSCM from Leica Microsystems; it was designed in collaboration with both companies. Optical access to the sample for the microscope is provided from below by a glass coverslip which also serves as the bottom plate for the rheometer. This coverslip is rigidly mounted to the rheometer via a metal cup; as a result, the two devices maintain autonomous functionality even when joined together.
The rheometer modifications made for this application do not dictate a choice of microscope manufacturer or confocal head style. However, if an attempt is made to duplicate this design using another rheometer manufacturer, care must be taken to match the mounting and tool specifications. Maintaining gap tolerance and tool runout are critical features for simultaneous rheology and visualization. We feel that if a microscope stage or any other independent platform is substituted for the bottom plate of the rheometer, the task of maintaining a gap on the order of $50\ {\ensuremath{\mu \mathrm{m}}}$ across a tool diameter of $25\ \mathrm{mm}$ becomes dramatically more difficult. In our design, we are able to machine fixed components at tolerances that are within the manufacturers specifications, providing us reliable gaps of $h \ge
20 \mu$m.
System Components
=================
Rheometer
---------
The rheometer for the system was modified at the factory by Anton Paar to be compatible with our design. This primarily involved the relocation of the front control panel and removal of the lower front section of the rheometer; this provides unimpeded access below the lower tool platform. The result is a horizontal platform that accepts all standard and custom manufactured bottom plate accessories. These modifications, and the large distance from the tool rotational axis to the front of the rheometer body, provide sufficient clearance for the microscope when the two instruments are mounted next to each other.
![\[FDrawings\] Cut-away drawings of the (a) confocal rheometer assembly (including the microscope objective, metal cup that mounts to the rheometer, glass coverslip, and rheometer tool) and (b) the baseplate at the bottom of the cup. (c) Magnified cross section of the sample region.](Figure2-eps-converted-to.pdf){width="3.0in"}
A custom stainless steel cup bolts to the rheometer platform and positions the bottom plate coverslip so that the platform does not interfere with the microscope; see Fig. \[FDrawings\](a). The cup consists of a factory-supplied mounting flange, a cylindrical section that bolts to the flange, and an interchangeable baseplate that bolts to the cylindrical section. The baseplate, an example of which is drawn in Fig. \[FDrawings\](b), registers the coverslip and has a slotted opening to provide optical access for the microscope. The underside of the baseplate is machined to a sharp edge around the perimeter of the slot; this allows a high magnification immersion objective to reach the coverslip, thus preserving its full working distance. A flat acrylic ring, fitted with a rubber O-ring, clamps the circular coverslip (40 mm in diameter) in place. A cross-sectional view of the sample region is shown in Fig. \[FDrawings\](c).
We use standard measuring system tools from Anton Paar. They must, however, be 150 mm in length in order to reach the coverslip. The cup assembly can accommodate tools, in either plate or cone geometries, with diameters of up to 25 mm.
The cup can be covered to limit evaporation from aqueous samples and is water tight so that biological samples can be immersed in media. A bath circulator flows heated or chilled fluid through copper coils that wrap around the cup to regulate the temperature of the entire cup and sample.
Microscope {#SScope}
----------
The Leica confocal uses a standard DMI6000B inverted microscope with the upper illumination arm removed. The vertical ($z$) position of the objective is controlled by a piezo-based focusing attachment from piezosystem jena. It is important to take into account the further geometrical constraints that the body of the piezo puts on the design of the rheometer cup. Alternatively, motor control of the nosepiece will suffice, but will lead to slower acquisition times for confocal stacks.
Images are acquired with a raster point scanner, which can be operated in an 8 kHz resonant mode. The acquisition speed can be effectively doubled with a bi-directional $x$ scan; in this case, a typical image stack with a resolution of $256 \times 256 \times 100$ voxels can be acquired in roughly 2 s. In addition, a series of three-dimensional stacks can be imaged in a bi-directional $z$ mode, where the order of $z$ slices is reversed on alternate stacks, preventing the piezo from resetting abruptly.
Translation Stage and Instrument Coupling
-----------------------------------------
To adjust the position of the rheometer relative to the microscope body, and thus the imaging field of view, we designed a mounting stage that provides leveling and translation capabilities. The stage is comprised of two aluminum plates separated by three fine-threaded, ball-bearing-tipped screws for leveling the rheometer. These plates are held together with three springs. The springs are removable and provide access to the bottom plate which bolts directly to two orthogonal Edmund Optics 38-180 translation stages. The translation stages are then mounted to a plate that bolts to the breadboard of the air table (Technical Manufacturing Corporation, model 63-543) on which the entire confocal rheometer sits. In order to place the rheometer’s tool over the objective, the stage must be rotated about it’s vertical center-line axis at an angle of $12.5^\circ$ relative to the the microscope – [*N.B.*]{} this is the case for Leica 6000 series inverted microscopes, however each different manufacturer should be tested for orientation requirements.
Through the use of a vibrationally isolated table, mechanical noise from external sources is diminished for the microscope and rheometer separately. However any residual relative motion between the instruments can degrade the imaging quality. The biggest source of this motion comes from the active components of the rheometer itself, with dominant frequencies near 100 Hz. To minimize relative motion, a “soft-coupling” clamp, shown in the inset of Fig. \[FPhoto\], connects the microscope base and the rheometer cup. This clamp is adjustable in the horizontal plane, so the rheometer can be positioned as needed, and is attached to the microscope stand through the stage mounting positions. While the vibrational noise with the clamp (with an amplitude under 100 nm, as directly measured from rapidly acquired microscope images) is still larger than when using a traditional microscope stage, the current conditions do not hinder the sort of measurements described in the next section.
![\[FRCalibrate\] Radial position calibration. The measured linear shear velocity $v_t$ (symbols) of the rheometer tool displays the expected linear dependence (solid line) on the position $X$ of the rheometer stage.](Figure3-eps-converted-to.pdf){width="3.0in"}
Placing the rheometer on its translation stage only roughly constrains its position, which simplifies assembly of the instrument. To accurately determine the radial position of the microscope objective with respect to the central axis of the rheometer tool, we rely on imaging the surface of the tool directly. For instance, this position can be determined by measuring the local linear velocity $v_t$ of the tool due to a rotation of known angular velocity. We verified this approach with the measurement shown in Fig. \[FRCalibrate\]. Images of the tool were acquired in reflectance while the rheometer was set to a steady shear rate of 0.330 1/s with a gap of $50\ {\ensuremath{\mu \mathrm{m}}}$. We measured $v_t$ at several locations across the face of the tool as the stage was moved in one dimension. As this path was chosen to run through the rheometer axis, $v_t$ should have a linear dependence on the stage position $X$. The fit shown with a solid line in Fig. \[FRCalibrate\] yields a shear rate of 0.331 1/s, where the discrepancy from the nominal value reflects the size of the mismatch in the spatial and temporal calibrations of the two instruments.
Experimental Results
====================
To demonstrate the capabilities of the instrument as well as provide some context for the discussion of other design issues, we next present a few representative measurements. Throughout the section, the local velocity, vorticity, and gradient axes at the imaging location will be referred to as $x$, $y$, and $x$, respectively.
Oscillatory Measurements
------------------------
Many biological polymer networks have complex rheological properties that play important roles in structural integrity and cell motility. Furthermore, individual fiber bundles can often be imaged, opening up the possibility of linking bulk behavior and various geometrical properties of these sparse disordered networks. We now show a simple example of how the viscoelasticity of a gel, as quantified by the dynamic shear moduli, is reflected in its structure.
![\[FCollagen\] Polymerization of collagen. (a)-(c) The images show snapshots in the $xz$-plane of a sample during polymerization; the scale bar indicates $30\ {\ensuremath{\mu \mathrm{m}}}$. (d) The storage $G^\prime$ (squares) and loss $G^{\prime\prime}$ (circles) moduli both plateau as the gel is formed. The arrows indicate the times at which the images were acquired.](Figure4-eps-converted-to.pdf){width="3.0in"}
Figure \[FCollagen\] shows data taken during the polymerization of a fluorescently-labeled collagen network (1 mg/mL concentration, 0.12 ionic strength). Once the polymerization was initiated and the sample was loaded, the rheometer gap was set to $90\ {\ensuremath{\mu \mathrm{m}}}$ with a 25 mm parallel plate tool. The storage $G^\prime$ and loss $G^{\prime\prime}$ moduli were continuously monitored using a 0.5% strain amplitude, 1 Hz oscillation. The results, shown in Fig. \[FCollagen\](d), suggest that the polymerization took roughly 4000 s.
At the same time, three-dimensional image stacks were periodically acquired at a radial position of 8 mm. Figure \[FCollagen\](a)-\[FCollagen\](c) show images in the $xz$-plane taken at the times indicated by the arrows in the rheology plot. Each panel is the maximum projection along the $y$-axis of a section of the stack $9\ {\ensuremath{\mu \mathrm{m}}}$ thick and shows the full extent of the sample in $z$, from coverslip to tool. From these images, where brighter colors indicate a stronger fluorescence signal, the formation of fibers can be followed. It is interesting to note that the background noise signal diminishes as more fluorescently-labeled collagen is incorporated into the network.
For this measurement, the polymerization was carried out on a clean glass coverslip. However, if adhesion is a concern for larger strains, we have found it possible to chemically treat the coverslips to make them hydrophilic (or hydrophobic, when needed) without affecting the imaging quality in any appreciable way.
Steady Flow Measurements
------------------------
Many colloidal systems and other structured fluids display interesting behavior under a continuous shear. This may include ordering at the level of individual constituent particles or the formation of large-scale shear bands. The confocal rheometer can provide some insight into how these structural properties affect bulk rheology under a wide range of shears.
![\[FFlow\] Flow properties of two fluids. (a) Rheological flow curves for honey (circles) and a compressed emulsion (squares) show Newtonian and Herschel-Bulkley behavior, respectively. The $z$-dependent flow profiles for (b) honey and (c) the emulsion have a very different dependence on the shear rate; here, the average shear velocity $v_x$ (normalized by the local tool speed $v_t$) is plotted for local shear rates of $10^{-3}$ (squares), 1 (triangles), and $10^3$ (circles) $1 / \mathrm{s}$. The images inset in (b) show an example of the analysis used to extract $v_x$ for the fastest rate.](Figure5-eps-converted-to.pdf){width="3.0in"}
A comparison of the flow behavior for honey, a nearly Newtonian fluid, and an oil-in-water emulsion (compressed to a volume fraction of 0.60), which has a yield stress and shear thins, is shown in Fig. \[FFlow\]. A clean coverslip worked well for the honey, but this surface lead to complete boundary slip for the emulsion drops. We found that a robust solution was to lithographically define a square grid of posts on a coverslip using SU-8, a negative photoresist that adheres well to glass. Image degradation can be avoided by matching the index of refraction of the posts to that of the sample. As there was also considerable slip on the bare metal rheometer tool, we attached a similarly roughened coverslip to its surface.
The difference in the bulk behavior of the materials is clearly seen in the flow curves of Fig. \[FFlow\](a). The shear stress $\sigma$ is nearly linear in the strain rate $\dot{\gamma}$ for the honey (circles), corresponding to a constant viscosity of 6.4 Pa s. On the other hand, the emulsion (squares) closely follows a Herschel-Bulkley form, as shown by the solid fit line, with $\sigma =
47 + 17 \dot{\gamma}^{0.47}$.
A simple way to characterize the spatial properties of a steady state flow is with the average velocity $v_x$ in the shear direction as a function of the position $z$ above the coverslip. Such flow profiles are shown in Fig. \[FFlow\](b) and (c) for the two fluids for three different shear rates. The velocities are normalized by the local tool speed $v_t$; the position is normalized by the rheometer gap $h$, which was set to $100\ {\ensuremath{\mu \mathrm{m}}}$. The flow was measured by following 1 [$\mu \mathrm{m}$]{} fluorescent tracer beads mixed in with the samples.
As seen in Fig. \[FFlow\], the honey undergoes a roughly affine deformation (indicated by the solid line) for all $\dot{\gamma}$, as expected in a flow where the shear stress is independent of $z$. The emulsion flows in a similar fashion for the highest shear rate, but displays strong shear localization near the coverslip at the lowest shear rate.
The advanced engineering of both the confocal and rheometer are needed to acquire data over the wide dynamic range shown in the figure. For instance, the torque needed by the rheometer to produce the shears in Fig. \[FFlow\](a) for the honey varies by six orders of magnitude.
In terms of imaging, for the slow rate, the samples move slowly enough for full three-dimensional stacks to be acquired over time. With these in hand, traditional particle finding and tracking algorithms can be used to extract three-dimensional velocity fields from the individual tracers.[@Crocker96a; @MPT] At higher shear, where it is no longer possible to acquire full stacks, two-dimensional images can be rapidly acquired at fixed $z$. From these images, particle image velocimetry techniques yield values of $v_x$ and $v_y$ averaged over the $xy$-plane.
At still higher shear, even a single image is distorted by the raster scanning process, as the fluid moves a significant amount in the time it takes to acquire one line. In this case, a single velocity component of a flow, assumed to be uniform in the direction of scanning, can be inferred from the relative displacement between pairs of image lines needed to recover circular particles. This type of analysis was used for the highest shear rates for both fluids; an example of the image processing is shown in the inset to Fig. \[FFlow\](b). The raw image is at the top, while the one below it shows the recovered image due to an average flow of $6700\ {\ensuremath{\mu \mathrm{m}}}/ \mathrm{s}$. The frequency of the resonant scanner limits the velocity to which this technique can be applied.
When acquiring a three-dimensional stack with a confocal microscope, as needed for a flow profile measurement, it can be critical to know the exact value of $z$ at which each slice is taken. As there are several changes in the index of refraction along the imaging path, knowledge of the objective position is insufficient. In particular, the index mismatch between the immersion fluid (with index $n_i$) and the sample (with $n_s$) introduces a number of effects, including reduced signal intensity, a degradation in resolution, and a shift in the position of the focal plane.[@Carlsson91a]
![\[FFocalShift\] Focal shift measurements. The ratio $\Delta f / \Delta z$ between changes in the position of the focal plane and the objective increases roughly linearly with the index of refraction $n_s$ of the sample being imaged. Measurements are shown for two objectives which use oil (solid circles) and water (open circles) as the immersion fluid.](Figure6-eps-converted-to.pdf){width="3.0in"}
We measured the focal shift for two objectives and a selection of samples that filled the rheometer’s gap, as shown in Fig. \[FFocalShift\]. For each sample, a linear dependence was found between the rheometer’s gap and the position of the objective needed to bring the surface of the tool into focus. The focal shift can be quantified by the slope $\Delta f / \Delta z$ of this relationship, where a change in the objective height $\Delta z$ leads to the focal plane moving by $\Delta f$. As expected, the values are nearly reproduced by $n_s / n_i$ for two $63\times$ objectives using oil (solid circle, solid line) and water (open, dashed) immersion fluid. This correction was applied to the flow profiles in Fig. \[FFlow\]; in fact, the water-immersion objective had to be moved 89 and $95\ {\ensuremath{\mu \mathrm{m}}}$ for the honey and emulsion, respectively, to cover the actual $100\ {\ensuremath{\mu \mathrm{m}}}$ gap. While the focal shift can be measured with far less instrumentation, the confocal rheometer is well suited to the task, particularly given the ease of setting a variable gap, and can provide a value for any specific combination of objective and sample.
Performing the flow measurements presented a few other challenges, particularly for the emulsion. A consequence of providing optical access to the sample is that the coverslip that serves as the bottom rheometer plate can deflect over the area where it is unsupported by the metal baseplate. This can occur in two ways. For one, loading a stiff sample can result in a significant force on the glass originating from the tool. Additionally, for objectives that require an immersion fluid, the resultant coupling can lead to a deflection of the coverslip when the objective moves in the $z$-direction. This problem is exacerbated by the presence of the rheometer tool, which imposes a fixed boundary plane. If the coverslip moves for either reason, it imposes a stress on the sample which can result, for example, in the rearrangement of emulsion droplets.
There are several ways to mitigate these issues. To minimize deflection, a rheometer cup baseplate with a single small hole \[rather than one with a wide slot, as show in Fig. \[FDrawings\](b)\] can be used to reduce the unsupported coverslip area. Using water or a low viscosity oil as the immersion fluid greatly reduces the objective coupling. Ill effects of the coupling can be further reduced with the bi-directional $z$ scanning described in Sec. \[SScope\], which is essential for imaging colloidal systems.
The severity of the deflection depends strongly on the properties of the sample, so it is important to monitor its effects. In particular, we have found that the normal force sensor of the rheometer is sensitive to very small deflections of the coverslip, including those caused by the objective during sequential three-dimensional stack acquisition. When loading a stiff sample, the rheometer control software can lower the tool at a slow enough rate to allow the sample to fill the gap uniformly. In addition, for a structured fluid that displays yielding behavior, a slow rotation or oscillation of the tool can improve this relaxation. In practice, we modify the loading profile to minimize the coverslip deflection, as measured by the normal force sensor and direct imaging with the microscope.
![\[FLoading\] Sample loading. The (a) normal force and (b) coverslip deflection were monitored while a compressed emulsion was loaded under the different conditions described in the text. In the main plots, the gap $h$ decreases with time to a final value of 100 [$\mu \mathrm{m}$]{}. The insets show the relaxation of the coverslip when the tool was rotated after loading, as a function of time $t$.](Figure7-eps-converted-to.pdf){width="3.0in"}
An example of this monitoring is shown in Fig. \[FLoading\], where the normal force $F_N$ and coverslip location $z_0$ (where $z_0 = 0$ before the sample is loaded) were measured during the loading of a compressed emulsion to a gap $h$ of $100\ {\ensuremath{\mu \mathrm{m}}}$. Both of these values were quite high when the tool was lowered at a rate of $50\ {\ensuremath{\mu \mathrm{m}}}/\mathrm{s}$ (solid lines). As shown in the insets, the coverslip relaxes quickly when the tool is rotated at a shear rate of 5 1/s after the gap has been set to its final value. The maximum normal force is smaller when loading at $5\ {\ensuremath{\mu \mathrm{m}}}/\mathrm{s}$ with a 1 1/s rotation (dashed lines). Alternatively, the deflection can be minimized by using a baseplate with a single hole (dotted lines). This baseplate also limits deflection during measurements after the sample has been loaded. Thus the baseplate and loading protocol can be chosen to meet the needs of each experiment.
Summary
=======
We have described a confocal rheometer comprised of two commercial instruments. While the Anton Paar rheometer did have to be modified to provide adequate access for the Leica microscope, taking advantage of the significant engineering of the instruments simplified assembly of the combined system and went a long way to ensuring successful operation. In addition to providing standard bulk viscoelastic measurement capability, the rheometer has many advantages over a standard shear cell, including a normal force sensor that is particularly useful during loading, easy measurement profile definition, a wide range of applied torque, and a gap that can be precisely controlled over the course of an experiment.
We feel that due to the large number of groups that have incorporated microscopy and rheology as equipment for their research, that this system can be implemented with a modest amount of additional machine work and engineering. We hope that our straightforward design principles will be easily transferable.
Acknowledgments
===============
We are indebted to L. Der for his insight on a variety of design issues and machining expertise. This work was funded by the generous support of Georgetown University and the National Science Foundation Grant \# DMR-0847490.
[21]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, , , , , ****, ().
, , , ****, ().
, , , , , , , , ****, ().
, , , , , ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, , , , ****, ().
, , , , , , ****, ().
, , , ****, ().
, , , , ****, ().
, , , , , , ****, ().
, ****, ().
, , , , , , ****, ().
, ****, ().
.
, , , , , ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
A stratified bundle is a fibered space in which strata are classical bundles and in which attachment of strata is controlled by a structure category ${{\bf\mathfrak{{F}}}}$ of fibers. Well known results on fibre bundles are shown to be true for stratified bundles; namely the pull back theorem, the bundle theorem and the principal bundle theorem.
*AMS SC*: 55R55 (Fiberings with singularities); 55R65 (Generalizations of fiber spaces and bundles); 55R70 (Fiberwise topology); 55R10 (Fibre bundles); 18F15 (Abstract manifolds and fibre bundles); 54H15 (Transformation groups and semigroups); 57S05 (Topological properties of groups of homeomorphisms or diffeomorphisms).
*Keywords*: Stratified fibre bundle, structure category, function space of admissible maps, principal bundle
author:
- 'Hans-Joachim Baues [^1] and Davide L. Ferrario [^2]'
title: Stratified fibre bundles
---
Introduction {#section:introduction}
============
A stratified bundle is a filtered fibered space $X=\{X_i, i\geq 0\}$ for which the complements $X_i{\smallsetminus\/}X_{i-1}$, termed strata, are fibre bundles. Moreover the attachment of strata is controlled by a structure category ${{\bf\mathfrak{{F}}}}$ of fibers. For example the tangent bundle of a stratified manifold is a stratified vector bundle, see [@sm]. Moreover for a compact smooth $G$-manifold $M$ the projection $M\to M/G$ to the orbit space is a stratified bundle by results of Davis [@davis]. In this paper we prove three basic properties: The *pull back theorem* shows that certain pull backs of stratified bundles are again stratified bundles, see \[theo:pullback\]. The *bundle theorem* \[theo:bundle\] states that stratified bundles are the same as classical fibre bundles in case the structure category is a groupoid. Moreover the *principal bundle theorem* \[theo:principal\] shows that the function space of admissible maps, $X^V$, is a stratified bundle and the ${{\bf\mathfrak{{F}}}}^{\mathrm{op}}$-diagram $V\mapsto X^V$ plays the role of the associated principal bundle, see \[theo:principalbundle\]. In the proofs of the results we have to consider the intricate compatibility of quotient topology, product topology, and compact-open topology in function spaces. As application we study in [@sm] algebraic constructions on stratified vector bundles (like the direct sum, tensor product, exterior product) and we show that stratified vector bundles lead to a stratified $K$-theory generalizing the Atiyah-Hirzebruch $K$-theory. This was the main motivation for the proof of the basic properties in this paper.
The structure of the paper is the following. In section \[section:fibrespaces\] the category ${\mbox{${{\bf\mathfrak{{F}}}}$-${{{\mathbf{Top}}}}$}}$ of fibre families is introduced. In section \[section:cwcomplexes\] the proper notion of CW-complex in ${\mbox{${{\bf\mathfrak{{F}}}}$-${{{\mathbf{Top}}}}$}}$ is exploited, and this leads to the definition of stratified bundles in section \[section:stratifiedbundles\]. In more standard language of stratification theory, the stratified bundles here introduced should be called CW-stratified bundles. In the same section the pull back theorem \[theo:pullback\] is introduced. The proof of this theorem, like the proofs of \[theo:bundle\], \[theo:principal\] and \[theo:principalbundle\], is forward-referencing: its proof will be complete only in section \[section:pullback\]. In section \[sec:bundle\] we introduce the bundle theorem \[theo:bundle\]. Its proof will be done in sections \[sec:bundle\] and \[section:pushout\]. The principal bundle theorem \[theo:principalbundle\] (and also theorem \[theo:principal\]) are stated and introduced in section \[section:principalbundle\]. Its proof will be the content of sections \[section:9\] (where we study NNEP pairs) and \[section:10\] (where NKC categories are introduced). At the end, in the short section \[section:NKC\], some examples of NKC categories are given. Thus, after this introduction in section \[section:introduction\], the paper consists mainly of three parts: a first part — sections \[section:fibrespaces\] and \[section:cwcomplexes\] — with some preliminaries; then a second part — sections \[section:stratifiedbundles\], \[sec:bundle\] and \[section:principalbundle\] — with our definition of (CW)-stratified bundle and the main theorems; finally, a third part – sections \[section:pullback\], \[section:pushout\], \[section:9\], \[section:10\] and \[section:NKC\] – with the actual proofs of the main theorems, together with some remarks and further propositions.
Fibre families {#section:fibrespaces}
==============
Let ${{\bf\mathfrak{{F}}}}$ be a small category together with a functor ${F}{\colon}{{\bf\mathfrak{{F}}}}\to {\mathbf{Top}}$ to the category ${\mathbf{Top}}$ of topological spaces. We assume that the functor ${F}$ satisfies the assumption:
> $(*)$ For every object $V$ in ${{\bf\mathfrak{{F}}}}$ the space ${F}(V)$ is locally compact, second-countable and Hausdorff.
Then ${{\bf\mathfrak{{F}}}}$ is termed *structure category* and ${F}$ is a *fibre functor* on ${{\bf\mathfrak{{F}}}}$. To be precise a structure category is a pair $({{\bf\mathfrak{{F}}}},{F})$, but in general we omit to write explicitely ${F}$. If ${F}$ is a faithful functor then ${{\bf\mathfrak{{F}}}}$ is a topological enriched category in which morphism sets have the compact open topology. In many examples the functor ${F}$ is actually the inclusion of a subcategory ${{\bf\mathfrak{{F}}}}$ of ${\mathbf{Top}}$ so that in this case we need not to mention the fibre functor ${F}$.
A *fibre family* with fibres in ${{\bf\mathfrak{{F}}}}$ (or a $({{\bf\mathfrak{{F}}}},{F})$-family) is a topological space $X$, termed *total space*, together with a map $p_X{\colon}X \to \bar X$, termed *projection* to the *base space* $\bar X$, and for every $b\in \bar X$ a selected homeomorphism $\Phi_b{\colon}p_X^{-1}b \approx {F}X_b$ where $X_b$ is an object in ${{\bf\mathfrak{{F}}}}$, called *fibre*, depending on $b\in \bar X$. The homeomorphism $\Phi_b$ is termed *chart* at $b$. The family $(p_X{\colon}X \to \bar X, X_b, \Phi_b, b\in \bar X)$ is denoted simply by $X$. Each object $V$ in ${{\bf\mathfrak{{F}}}}$ yields the *point family*, or *${{\bf\mathfrak{{F}}}}$-point*, also denoted by $V$ given by the map $p_V{\colon}{F}V\to *$ where $*$ is the singleton space.
Given two ${{\bf\mathfrak{{F}}}}$-families $X$ and $Y$ a *${{\bf\mathfrak{{F}}}}$-map* from $X$ to $Y$ is a pair of maps $(f,\bar f)$ such that the following diagram [$$\begin{aligned}\xymatrix{X \ar[r]^f
\ar[d]_{p_X} & Y \ar[d]^{p_Y} \\
\bar X \ar[r]^{\bar f} & \bar Y,
}\end{aligned}$$]{}commutes, and such that for every $b\in \bar X$ the composition given by the dotted arrow of the diagram [$$\begin{aligned}\xymatrix{p_X^{-1}(b)\ar[r]^{f|p^{-1}(b)}
& p_Y^{-1}(\bar f b) \ar[d]^{\Phi_{\bar f b}} \\
{F}X_b
\ar[u]^{\Phi_b^{-1}}
\ar@{.>}[r]^{} & {F}Y_{\bar f b}
}\end{aligned}$$]{}is a morphism in the image of the functor ${F}$. That is, there exists a morphism $\phi{\colon}X_b \to Y_{\bar f b}$ in ${{\bf\mathfrak{{F}}}}$ such that the dotted arrow is equal to ${F}(\phi)$. We will often denote ${F}X_b$ by $X_b$ and it will be clear from the context whether $X_b$ denotes an object in ${{\bf\mathfrak{{F}}}}$ or a space in ${\mathbf{Top}}$ given by the functor ${F}$. If a ${{\bf\mathfrak{{F}}}}$-map $f=(f,\bar f)$ is a ${{\bf\mathfrak{{F}}}}$-isomorphism then $f$ and $\bar f$ are homeomorphisms but the converse need not be true.
If $X$ is a ${{\bf\mathfrak{{F}}}}$-family and $Z$ is a topological space, then $X\times Z$ is a ${{\bf\mathfrak{{F}}}}$-family with projection $p_{X\times Z} = p\times 1_Z\colon X\times Z \to \bar X \times Z$. The fibre over a point $(b,z)\in \bar X\times Z$ is equal to $p_{X}^{-1}(b)\times \{z\}$; using the chart $\Phi_b:p_{X}^{-1}(b) \to X_b$ the chart $\Phi_{(b,z)}$ is defined by $(x,z) \mapsto \Phi_b(x) \in X_b$ where of course we set $(X\times Z)_{(b,z)} = X_b$. In particular, by taking $Z=I$ the unit interval we obtain the cylinder object $X\times I$ and therefore the notion of *homotopy*: two ${{\bf\mathfrak{{F}}}}$-maps $f_0,f_1\colon X \to Y$ are ${{\bf\mathfrak{{F}}}}$-homotopic (in symbols $f_0{\sim}f_1$) if there is a ${{\bf\mathfrak{{F}}}}$-map $F\colon X\times I \to Y$ such that $f_0=Fi_0$ and $f_1=Fi_1$. Here $i_0$ and $i_1$ are the inclusions $X \to X\times I$ at the levels $0$ and $1$ respectively.
Let ${\mbox{${{\bf\mathfrak{{F}}}}$-${{{\mathbf{Top}}}}$}}$ be the category consisting of ${{\bf\mathfrak{{F}}}}$-families $p_X: X \to \bar X$ in ${\mathbf{Top}}$ and ${{\bf\mathfrak{{F}}}}$-maps. Homotopy of ${{\bf\mathfrak{{F}}}}$-maps yields a natural equivalence relation $\sim$ on ${\mbox{${{\bf\mathfrak{{F}}}}$-${{{\mathbf{Top}}}}$}}$ so that the homotopy category $({\mbox{${{\bf\mathfrak{{F}}}}$-${{{\mathbf{Top}}}}$}})/{_{\sim}}$ is defined.
Consider a structure category ${{\bf\mathfrak{{F}}}}$ with fibre functor ${F}$. In general it is not assumed that ${F}$ is faithful, but it is easy to see that the category ${\mbox{${{\bf\mathfrak{{F}}}}$-${{{\mathbf{Top}}}}$}}$ is equivalent to ${{\bf\mathfrak{{F}}}}'$-${\mathbf{Top}}$, where ${{\bf\mathfrak{{F}}}}'={{\bf\mathfrak{{F}}}}/{_\equiv}$ is the category with the same objects as ${{\bf\mathfrak{{F}}}}$ and morphisms given by equivalence classes of morphisms, with $f\equiv f' \iff {F}(f) = {F}(f')$. We call ${{\bf\mathfrak{{F}}}}'$ the *faithful image* category of ${F}$.
If $V$ is an object in ${{\bf\mathfrak{{F}}}}$ and $\bar X$ is a space in ${\mathbf{Top}}$, then the projection onto the first factor $p_1{\colon}X=\bar X\times V \to \bar X$ yields the *product family* with fibre $V$; the charts $\Phi_b{\colon}\{b\}\times V \to V=X_b$ are given by projection and $X_b=V$ for all $b\in \bar X$. If $X$ is a ${{\bf\mathfrak{{F}}}}$-family ${{\bf\mathfrak{{F}}}}$-isomorphic to a product family then $X$ is said to be a *trivial ${{\bf\mathfrak{{F}}}}$-bundle*. In general a *${{\bf\mathfrak{{F}}}}$-bundle* is a locally trivial family of fibres, i.e. a family $X$ over $\bar X$ such that every $b\in \bar X$ admits a neighborhood $U$ for which $X|U$ is trivial. Here $X|U$ is the *restriction* of the family $X$ defined by $U\subset \bar X$.
Given a family $Y$ with projection $p_Y\colon Y \to \bar Y$ and a map $\bar f\colon \bar X \to \bar Y$, the pull-back $X=\bar f^*Y$ is the total space of a family of fibres given by the vertical dotted arrow of the following pull-back diagram. [$$\begin{aligned}\xymatrix{X =
\bar f^* Y
\ar@{.>}[r]
\ar@{.>}[d]
& Y
\ar[d]^{p}
\\
\bar X
\ar[r]^{\bar f}
& \bar Y.\\
}\end{aligned}$$]{}The charts are defined as follows: For every $b\in \bar X$ let $X_b= Y_{\bar fb}$, and $\Phi_b: p_X^{-1}(b) \to X_b$ the composition $p_X^{-1}(b)\to p_Y^{-1}(\bar fb) \approx Y_{\bar fb}=X_b$ where the map $p_X^{-1}(b)\to p_Y^{-1}(\bar fb)$ is a homeomorphism since $X$ is a pull-back.
A ${{\bf\mathfrak{{F}}}}$-map $i{\colon}A \to Y$ is termed a *closed inclusion* if $\bar i{\colon}\bar A \to \bar Y$ is an inclusion, $\bar i \bar A$ is closed in $\bar Y$ and the following diagram is a pull-back: [$$\begin{aligned}\xymatrix{i^*Y = A \ar[r]^{i} \ar[d]_{p_A} & Y \ar[d]^{p_Y}\\
\bar A \ar[r]^{\bar i} & \bar Y \\
}\end{aligned}$$]{}Hence a closed inclusion $i{\colon}A \to Y$ induces homeomorphisms on fibres.
The push-out construction can be extended to the category ${\mbox{${{\bf\mathfrak{{F}}}}$-${{{\mathbf{Top}}}}$}}$, provided the push-out is defined via a closed inclusion.
\[lemma:pushout\] Given ${{\bf\mathfrak{{F}}}}$-families $A$, $X$, $Y$ and ${{\bf\mathfrak{{F}}}}$-maps $f{\colon}A \to X$, $i{\colon}A\to Y$ with $i$ a closed inclusion the push-out diagram in ${\mbox{${{\bf\mathfrak{{F}}}}$-${{{\mathbf{Top}}}}$}}$ [$$\begin{aligned}\xymatrix{A \ar[r]^f \ar@{->}[d]_{i} & X \ar@{->}[d]\\
Y \ar[r] & Z \\
}\end{aligned}$$]{}exists and $X\to Z$ is a closed inclusion.
The push-out is obtained by the following commutative diagram, [$$\begin{aligned}\xymatrix{
A \ar[rrr]^f
\ar[rd]^{p_A}
\ar@{ >->}[ddd]_i
& & & X \ar[ddd]
\ar[ld]^{p_X}
\\
& \bar A \ar[r]^{\bar f} \ar@{ >->}[d]^{\bar i} & \bar X \ar[d] \\
& \bar Y \ar[r] & \bar Z \\
Y \ar[rrr] \ar[ur]^{p_Y} & & & Z \ar@{.>}[ul]^{p_Z} \\
}\end{aligned}$$]{}in which the spaces $Z$ and $\bar Z$ are push-outs in ${\mathbf{Top}}$ of $f$, $i$ and $\bar f$, $\bar i$ respectively; the map $p_Z$ exists and is unique because of the push-out property. Moreover, the charts of fibres of $Z$ are given as follows: $\bar Z$ is the union of the two subsets $\bar X$ and $\bar Y{\smallsetminus\/}\bar A$. For $b=\bar x \in \bar X\subset \bar Z$ there is a canonical composition of homeomorphisms $p_Z^{-1}(b) \approx p_X^{-1}(\bar x) \approx X_{\bar x}$ that yields the chart $$\Phi_{b}:p_Z^{-1}(b) \to X_{\bar x} = Z_b.$$ In fact, $i:A \to Y$ being a closed inclusion, for every $\bar a \in \bar A$ the fibre over $\bar a$ in $A$ is canonically isomorphic to the fibre over $\bar a$ in $Y$ and hence for every $\bar x \in X$ the fibre over $\bar x$ in $X$ is canonically isomorphic to the fibre over $\bar x$ in $Z$. Next, for $b=\bar y \in
\bar Y {\smallsetminus\/}\bar A \subset Z$ the homeomorphism is $$\Phi_b: p_Z^{-1}(b) \approx p_Y^{-1}(\bar y) \approx Y_{\bar y} = Z_b.$$ The ${{\bf\mathfrak{{F}}}}$-map $j: Y \to Z$ induced by the push-out is a closed inclusion, since $\bar j: \bar Y \to \bar Z$ is the inclusion of a closed subspace of $\bar Z$; furthermore, $j$ is an isomorphism on fibres and an inclusion, and therefore $Y = j^* Z$.
A point of this approach is to endow ${\mbox{${{\bf\mathfrak{{F}}}}$-${{{\mathbf{Top}}}}$}}$ with the structure of a cofibration category (or cylinder category), in order to have ${{\bf\mathfrak{{F}}}}$-complexes and stratifications. The dual concept, namely of ${{\bf\mathfrak{{F}}}}$-fibrations, was studied by May in [@may].
${{\bf\mathfrak{{F}}}}$-complexes {#section:cwcomplexes}
=================================
We recall that for an object $V$ in ${{\bf\mathfrak{{F}}}}$ the family ${F}(V)\to *$ with base space a singleton is termed a *${{\bf\mathfrak{{F}}}}$-point* and is also denoted by $V$. A disjoint union of ${{\bf\mathfrak{{F}}}}$-points is called a *${{\bf\mathfrak{{F}}}}$-set*. This is a ${{\bf\mathfrak{{F}}}}$-family for which the base space has the discrete topology. Let $D^n$ be the unit disc in ${\mathbb{R}}^n$ and $S^{n-1}$ its boundary with base point $*\in S^{n-1}$. The complement $e^n=D^n{\smallsetminus\/}S^{n-1}$ is the *open cell* in $D^n$. An ${{\bf\mathfrak{{F}}}}$-cell is a product family $V\times e^n\to e^n$ with $V\in {{\bf\mathfrak{{F}}}}$.
We say that a ${{\bf\mathfrak{{F}}}}$-family $X$ is obtained from a ${{\bf\mathfrak{{F}}}}$-family $D$ by *attaching* $n$-cells if a ${{\bf\mathfrak{{F}}}}$-set $Z$ together with a ${{\bf\mathfrak{{F}}}}$-map $f$ is given, such that the following diagram [$$\begin{aligned}\xymatrix{Z\times S^{n-1} \ar[r]^f \ar[d] & D \ar[d] \\
Z \times D^{n} \ar[r]^{\hspace{-24pt} {\Phi}} & X=D\cup_f(Z\times D^n) \\
}\end{aligned}$$]{}is a push-out in ${\mbox{${{\bf\mathfrak{{F}}}}$-${{{\mathbf{Top}}}}$}}$. The inclusion $Z\times S^{n-1} \to Z\times D^{n}$ is a closed inclusion, therefore the push-out exists and the induced map $D \to X$ is a closed inclusion and $X{\smallsetminus\/}D = Z\times e^n$ is a union of open ${{\bf\mathfrak{{F}}}}$-cells. If $Z$ is a ${{\bf\mathfrak{{F}}}}$-point then we say that $X$ is obtained from $D$ by attaching a ${{\bf\mathfrak{{F}}}}$-cell and ${\Phi}$ is the *characteristic map* of the ${{\bf\mathfrak{{F}}}}$-cell.
\[defi:CWcomplex\] A *relative ${{\bf\mathfrak{{F}}}}$-complex* $(X,D)$ is a family $X$ and a filtration $$D=X_{-1} \subset X_0
\subset X_1 \subset \dots \subset X_n \subset X_{n+1} \subset \dots \subset X$$ of ${{\bf\mathfrak{{F}}}}$-families $X_n$, $n\geq -1$, such that for every $n\geq 0$ the ${{\bf\mathfrak{{F}}}}$-family $X_n$ is obtained from $X_{n-1}$ by attaching $n$-cells and $$X = \lim_{n\geq 0} X_n.$$
The spaces $X_n$ are termed *$n$-skeleta* of $(X,D)$. If $D$ is empty we call $X$ a *${{\bf\mathfrak{{F}}}}$-complex*. Then $X$ is a union of ${{\bf\mathfrak{{F}}}}$-cells. It is not difficult to show that a relative ${{\bf\mathfrak{{F}}}}$-complex $(X,D)$ is Hausdorff and normal provided that $D$ is Hausdorff and normal. In \[defi:CWcomplex\] the limit of skeleta is the limit induced by the standard limit of topological spaces (tnd therefore the induced map $X \to \bar X$ is continuous).
\[ex:orbit\] Let $G$ be a compact Lie group and let ${{\bf\mathfrak{{F}}}}$ be the *category of orbits* of $G$, that is, ${{\bf\mathfrak{{F}}}}$ is the subcategory of ${\mathbf{Top}}$ consisting of spaces $G/H$, where $H$ is a closed subgroup of $G$, and $G$-equivariant maps $G/H \to G/H'$. Then each $G$-CW-complex (see [@tomdieck]) is a ${{\bf\mathfrak{{F}}}}$-complex.
Stratified bundles {#section:stratifiedbundles}
==================
Let $A$ be a closed subset of a space $X$. We say that $(X,A)$ is a *CW-pair* if there exists a homeomorphism $(X,A) \approx (X',A')$ of pairs where $X'$ is a CW-complex and $A'$ a subcomplex of $X'$. A *CW-space* is a space homeomorphic to a CW-complex.
\[defi:strati\] We call a space $X$ a *stratified* space if a filtration $$X_0 \subset X_1 \subset \dots
\subset X_n \subset \dots \subset \lim_{n\to \infty} X_n = X$$ is given and if for every $i\geq 1$ there is given a CW-pair $(M_i, A_i)$ and a map $h_i{\colon}A_i \to X_{i-1}$ with the following properties: The subspace $X_i$ is obtained by attaching $M_i$ to $X_{i-1}$ via the attaching map $h_i$, i.e. there is a push-out diagram: [$$\begin{aligned}\xymatrix{A_i \ar@{ >->}[r]
\ar[d]_{h_i}
& M_i \ar[d] \\
X_{i-1} \ar@{ >->}[r] & X_i.\\
}\end{aligned}$$]{}
Moreover $X_0$ is a CW-space. The complements $X_i{\smallsetminus\/}X_{i-1}$ are termed *strata*, while the filtration $\emptyset \subset X_0 \subset
X_1 \subset \dots \subset X_n\subset \dots$ is called *stratification* of $X$. The strata $X_i{\smallsetminus\/}X_{i-1}$ coincide with the complements $M_i{\smallsetminus\/}A_i$. We call the pairs $(M_i, A_i)$ and the space $X_0$ the *attached spaces* of $X$. If all attaching maps are cellular then $X$ is a CW-complex and $X_i$ is a subcomplex of $X$. In this case we say that the stratified space $X$ is a *stratified CW-complex*. In particular CW-complexes with the *skeletal* filtration are stratified CW-complexes. We say that a stratified space is *finite* if the number of non-empty strata is finite. We always assume that a stratified space which is *not finite is a stratified CW-complex*. The reason of this assumption is that theorems \[theo:bundle\], \[theo:principal\] and \[theo:principalbundle\] will be proved under the assumption that the base spaces are finite or arbitrary CW-complexes: with this choice of terminology the formulation of the hypothesis happens to be more compact. The cellular approximation theorem shows that a stratified space is homotopy equivalent to a stratified CW-complex. We point out that a stratified space $X$ is a Hausdorff and regular space, see Gray [@gray].
A map $$f{\colon}X \to X'$$ between stratified spaces is a filtration preserving map $f=\{f_n\}_{n\ge 0}$ together with commutative diagrams [$$\begin{aligned}\xymatrix{X_{i-1}
\ar[d]^{f_{i-1}}
&
\ar[l]
A_i
\ar@{ >->}[r]
\ar[d]
&
M_i
\ar[d]^{g_i}
\\
X'_{i-1}
&
A'_i
\ar[l]
\ar@{ >->}[r]
&
M_i'
\\
}\end{aligned}$$]{}such that $g_i\cup f_{i-1} = f_i$ for $i\geq 1$. A map is termed *stratum-preserving* if for every $i$ $$f_i (X_i {\smallsetminus\/}X_{i-1} )
\subset X'_i {\smallsetminus\/}X'_{i-1}.$$ Let ${\mathbf{Stra}}$ be the category of stratified spaces and stratum-preserving maps.
Consider a manifold with boundary $(M,\partial M)$, a manifold $N$, and a submersion $h\colon \partial M \to N$. The push-out of $h$ and the inclusion $\partial M \subset M$ [$$\begin{aligned}\xymatrix{
\partial M \ar@{ >->}[r]
\ar[d]_h & M
\ar[d]
\\
N \ar[r] &
M \cup_h N.\\
}\end{aligned}$$]{}yields a stratified space $X=M \cup_h N$ with stratification $X_0=N \subset X_1=X$. For example if a diffeomorphism $\partial M \cong Z \times P$ is given where $Z$ and $P$ are manifolds and if $h:\partial M \cong Z\times P \to Z$ is defined by the projection then $X=M\cup_hN$ is a *manifold with singularities*, see Rudyak [@rudyak], Baas [@baas], Botvinnik [@botvinnik], Sullivan [@sullivan], Vershinin [@vershinin]. Also the stratified manifolds (stratifolds) of Kreck [@kreck] are stratified spaces. Manifolds with singularities and stratifolds have “tangent bundles” which are stratified vector bundles. These are important examples of stratified fibre bundles introduced in the next definition.
Recall that ${{\bf\mathfrak{{F}}}}$ denotes a structure category together with a fibre functor ${F}{\colon}{{\bf\mathfrak{{F}}}}\to {\mathbf{Top}}$.
\[defi:stratifiedvectorbundle\] A *${{\bf\mathfrak{{F}}}}$-stratified fibre bundle* is a stratified space $\bar X$ together with a ${{\bf\mathfrak{{F}}}}$-family $$X \to \bar X$$ with the following properties. For $i\geq 1$ the restriction $X_i = X|\bar X_i = X_{i-1} \cup_{A_i} M_i$ is the push-out of ${{\bf\mathfrak{{F}}}}$-maps [$$\begin{aligned}\xymatrix{M_i
\ar[d]
&
\ar@{ >->}[l]
A_i
\ar[d]
\ar[r]^{h_i}
&
X_{i-1}
\ar[d]
\\
\bar M_i
&
\bar A_i
\ar@{ >->}[l]
\ar[r]^{\bar h_i}
&
\bar X_{i-1}
\\
}\end{aligned}$$]{}where $M_i \to \bar M_i$ is a ${{\bf\mathfrak{{F}}}}$-bundle and $A_i=M_i|\bar A_i$. Moreover, $X_0 \to \bar X_0$ is a ${{\bf\mathfrak{{F}}}}$-bundle and $X=\lim_{i\to \infty} X_i$. Hence the strata $X_i {\smallsetminus\/}X_{i-1} \to \bar X_i {\smallsetminus\/}\bar X_{i-1}$ are ${{\bf\mathfrak{{F}}}}$-bundles.
It is possible to show that a finite ${{\bf\mathfrak{{F}}}}$-stratified fibre bundle yields a *stratified system of fibrations* (according to the definition 8.2 at page 420 of F. Quinn in [@quinn2]) on $\bar X$. In fact, since $\bar X$ is a CW-complex and hence its strata are paracompact, the restrictions of $p_X {\colon}X \to \bar X$ to the strata are fibrations (theorem XX.4.2.(2) at page 405 of [@dugundji] – see also theorem 1.3.5 at page 29 and Ex. 1 at page 33 of [@fripic]). Furthermore, for every $n\geq 0$ the inclusion $X_n \subset X$ is a ${{\bf\mathfrak{{F}}}}$-cofibration and all the spaces are compactly generated, therefore it is possible to prove that $\bar X_n$ is a $p$-NDR subset of $\bar X$, in the notation of [@quinn2]. Thus a ${{\bf\mathfrak{{F}}}}$-striatified bundle $X$ yields in a natural way a stratified system of fibrations on $\bar X$. Actually, if the CW-structure of $\bar X$ is given by a triangulation or a simplicial structure, then $X\to \bar X$ yields what in [@quinn2; @farjon1; @farjon2] is termed a *simplicially stratified fibration*.
Starting from the papers [@quinn1; @hughes1; @hughes2], F. Quinn and B. Hughes developed a theory of homotopically stratified sets which goes in a different direction than this paper. In their approach isotopy Whitney lemmas and applications to surgery theory are central, while our motivations reside more in abstract homotopy theory and $K$-theory (see [@sm] and [@baues]). For a general and modern introduction to stratified spaces and stratified bundles, related to problems in surgery theory a common reference is [@hugwei].
By comparing definition \[defi:stratifiedvectorbundle\] and definition \[defi:CWcomplex\] it is easy to see that a ${{\bf\mathfrak{{F}}}}$-complex with the skeletal filtration is a ${{\bf\mathfrak{{F}}}}$-stratified bundle since the spaces $(V\times D^n,V\times S^{n-1})$ are trivial ${{\bf\mathfrak{{F}}}}$-bundles over the CW-pair $(D^n,S^{n-1})$. Moreover, if $X$ is a ${{\bf\mathfrak{{F}}}}$-stratified bundle such that all the attaching maps are cellular, then $X$ is a ${{\bf\mathfrak{{F}}}}$-complex. Here we use the bundle theorem in section \[sec:bundle\] below. In general, a ${{\bf\mathfrak{{F}}}}$-stratified bundle has the ${{\bf\mathfrak{{F}}}}$-homotopy type of a ${{\bf\mathfrak{{F}}}}$-complex (this is a consequence of a cellular approximation theorem for ${{\bf\mathfrak{{F}}}}$-complexes, see [@baues]).
\[ex:davis\] Let $G$ be a compact Lie group and $M$ a compact smooth $G$-manifold. Let ${{\bf\mathfrak{{F}}}}$ be the orbit category of $G$. Then the *augmented $\mathcal{G}$-normal system* associated to $M$, as defined by Davis in [@davis], yields naturally a ${{\bf\mathfrak{{F}}}}$-stratified bundle $M\to \bar M = M/G$ via the corresponding assembling functor. The strata are defined exactly as in the proof of theorem 4.9 of [@davis], by $M_n = M(n-1)$ and $A_n= \partial M_n$. Thus each open stratum contains the disjoint union of all the open strata given by the stratification of $M$ (or the corresponding stratification on $\bar M$) by normal orbit type of depth $n$ in the poset of the normal orbit types.
A *${{\bf\mathfrak{{F}}}}$-stratified map* $f{\colon}X \to X'$ between ${{\bf\mathfrak{{F}}}}$-stratified bundles is is given by sequences $\{f_n\}_{n\geq 0}$ and $\{g_n\}_{n\geq 1}$ of ${{\bf\mathfrak{{F}}}}$-maps such that, given the commutative diagrams [$$\begin{aligned}\xymatrix{X_{i-1}
\ar[d]^{f_{i-1}}
&
\ar[l]
A_i
\ar@{ >->}[r]
\ar[d]
&
M_i
\ar[d]^{g_i}
\\
X'_{i-1}
&
A'_i
\ar[l]
\ar@{ >->}[r]
&
M_i'
\\
}\end{aligned}$$]{}for every $i$, we have $g_i\cup f_{i-1} = f_i$ for $i\geq 1$. If for every $i$ the inclusion $f(X_i{\smallsetminus\/}X_{i-1}) \subset X'_i {\smallsetminus\/}X_{i-1}'$ holds, $f$ is termed *stratum-preserving*.
\[theo:pullback\] Let $\bar X$ and $\bar X'$ be finite stratified spaces with attached spaces which are locally finite and countable CW-complexes. Let $\bar f{\colon}\bar X \to \bar X'$ be a stratum-preserving map (in ${\mathbf{Stra}}$) and $X'\to\bar X'$ a ${{\bf\mathfrak{{F}}}}$-stratified bundle. Then the pull-back $\bar f^*X' \to \bar X$ is a ${{\bf\mathfrak{{F}}}}$-stratified bundle.
Let $n$ be the number of strata of $\bar X$ and $X_i = \bar f^* X_i'$ for $i=1,\dots n$. Clearly $\bar f_0^* X_0'$ is a ${{\bf\mathfrak{{F}}}}$-bundle on $\bar X_0$. Since $X'$ is a ${{\bf\mathfrak{{F}}}}$-stratified bundle there are ${{\bf\mathfrak{{F}}}}$-bundles $M_i'$ and ${{\bf\mathfrak{{F}}}}$-maps $h'_i{\colon}A'_i \subset M_i' \to X_{i-1}'$ such that $X_{i}' = X_{i-1}'\cup_{h_i'} M_i'$. Let the attached spaces of $\bar X$ be denoted by $(\bar M_i, \bar A_i)$ and $\bar h_i$ the attaching maps. The stratified map $\bar f$ yields maps $\bar g_i{\colon}\bar M_i \to \bar M_i'$ such that $\bar g_i \cup \bar f_{i-1} = \bar f_i$. For every $i\geq 1$ let $M_i$ be the pull-back $\bar g_i^* M_i' \to \bar M_i$. It is a ${{\bf\mathfrak{{F}}}}$-bundle. Let $A_i$ be its restriction to $\bar A_i$. By assumption $\bar M_i$ and $\bar M'_i$ are locally finite and countable; furthermore since $\bar f$ is stratum-preserving, for every $i$ we have $\bar f(\bar X_{i} {\smallsetminus\/}\bar X_{i-1})
\subset \bar X'_i {\smallsetminus\/}\bar X'_{i-1}$. Thus, by applying proposition \[propo:theo:pushout2\] below, the pull-back family $X_i$ is obtained by attaching the ${{\bf\mathfrak{{F}}}}$-bundle $M_i$ to $X_i$ via a ${{\bf\mathfrak{{F}}}}$-map $h_i{\colon}A_i\subset M_i \to X_i$, where $h_i$ is a suitable ${{\bf\mathfrak{{F}}}}$-map induced by the construction. This is true for $i=1,\dots, n$, hence $X_n = X$ is a ${{\bf\mathfrak{{F}}}}$-stratified bundle.
A similar theorem was proved by Davis [@davis] (theorems 1.1 and 1.3), in the case of the pull-back $f^*M$ of a smooth $G$-manifold $M$, where $f$ is a (weakly) stratified map and $G$ is a compact Lie group.
Let $({{\bf\mathfrak{{F}}}},{F})$ and $({{\bf\mathfrak{{F}}}}',{F}')$ be structure categories. Then we define the structure category $({{\bf\mathfrak{{F}}}}\times{{\bf\mathfrak{{F}}}}',{F}\times{F}')$ by the functor $${F}\times{F}'{\colon}{{\bf\mathfrak{{F}}}}\times{{\bf\mathfrak{{F}}}}' \to {\mathbf{Top}}.$$ Here ${{\bf\mathfrak{{F}}}}\times{{\bf\mathfrak{{F}}}}'$ is the product category consisting of pairs $(V,V')$ of objects $V\in {{\bf\mathfrak{{F}}}}$ and $V'\in{{\bf\mathfrak{{F}}}}'$ and pairs of morphisms. The functor ${F}\times{F}'$ carries $(V,V')$ to the product space ${F}(V)\times{F}'(V')$. For a ${{\bf\mathfrak{{F}}}}$-family $X$ and a ${{\bf\mathfrak{{F}}}}'$-family $X'$ with the base space $\bar X=B=\bar X'$ we define the *fiberwise product* $X\times_BX'$ by the pull-back diagram in ${\mathbf{Top}}$ [$$\begin{aligned}\xymatrix{X\times_BX' \ar[r]\ar[d] \ar@{.>}[rd]^{p} & X' \ar[d]
\\
X \ar[r] &
B \\
}\end{aligned}$$]{}Then it is clear that $X\times_BX'\to B$ is a $({{\bf\mathfrak{{F}}}}\times {{\bf\mathfrak{{F}}}}',{F}\times{F}')$-family. Moreover we get compatibility with stratifications as follows.
Let $X$ be a ${{\bf\mathfrak{{F}}}}$-stratified bundle and let $X'$ be a ${{\bf\mathfrak{{F}}}}'$-stratified bundle such that the stratified spaces $\bar X = B = \bar X'$ coincide. Moreover assume that $X$ and $X'$ are locally compact. Then $X\times_BX'\to B$ is a ${{\bf\mathfrak{{F}}}}\times {{\bf\mathfrak{{F}}}}'$-stratified bundle.
By construction $X$ is the quotient space of a projection $$q{\colon}X_0 \cup \coprod_{i\geq 0} M_i \to X,$$ and the same holds for $X'$, with a projection $q'$. By the Whitehead theorem (see [@dugundji], theorem XII.4.1 at page 262) the map $q\times q'$ is a quotient map onto $X\times X'$, since $X$ and $X'$ are locally compact. Now, $X \times_B X \subset X\times X'$ is a closed subspace, hence the restricted projection $$q\times q'{\colon}(q\times q')^{-1} X\times_B X \to X' \times_B X'$$ is a quotient map. We can consider now the union $$(X_0\times_{\bar X_0} X_0')\coprod_i (M_i\times_{\bar M_i} M_i')
\subset (q\times q')^{-1} X\times_B X \to X'.$$ Then it is possible to see that the projection $q\times q'$ yields a quotient map $$(X_0\times_{\bar X_0} X_0')\coprod_i (M_i\times_{\bar M_i} M_i' )
\to
X \times_B X',$$ which gives the ${{\bf\mathfrak{{F}}}}\times {{\bf\mathfrak{{F}}}}'$-stratification of $X\times_B X'$.
The bundle theorem {#sec:bundle}
==================
A *groupoid* is a category in which all morphisms are isomorphisms. If the structure category ${{\bf\mathfrak{{F}}}}$ is a groupoid then ${{\bf\mathfrak{{F}}}}$-complexes and ${{\bf\mathfrak{{F}}}}$-stratified bundles are actually ${{\bf\mathfrak{{F}}}}$-bundles. More precisely we show the following result which we could not find in the literature though special cases are well known like the clutching construction of Atiyah [@at], page 20. This can be briefly explained as follows: if $\bar X=\bar X_1 \cup \bar X_2$ and $\bar A = \bar X_1 \cap \bar X_2$, all the spaces are compact and $E_i \to \bar X_i$ are (vector) bundles with an isomorphism of the restrictions $\varphi{\colon}E_1|A \cong E_2|A$, then the clutching (or glueing) construction is a way to define the push-out of the two bundles $E_1\cup_\varphi E_2$ over the push-out space $\bar X_1 \cup_{\bar \varphi} \bar X_2$.
\[theo:bundle\] Let ${{\bf\mathfrak{{F}}}}$ be a structure category which is a groupoid. Then a ${{\bf\mathfrak{{F}}}}$-stratified bundle $X \to \bar X$ is a ${{\bf\mathfrak{{F}}}}$-bundle over $\bar X$. Conversely, given a stratified space $\bar X$ and a ${{\bf\mathfrak{{F}}}}$-bundle $X$ over $\bar X$ then $X$ is a ${{\bf\mathfrak{{F}}}}$-stratified bundle. In particular, a ${{\bf\mathfrak{{F}}}}$-complex $X\to\bar X$ is a ${{\bf\mathfrak{{F}}}}$-bundle over $\bar X$ and, given a CW-complex $\bar X$ and a ${{\bf\mathfrak{{F}}}}$-bundle $X$ over $\bar X$, then $X$ is ${{\bf\mathfrak{{F}}}}$-isomorphic to a ${{\bf\mathfrak{{F}}}}$-complex.
If $X$ is a ${{\bf\mathfrak{{F}}}}$-complex then the theorem is a consequence of corollary \[coro:clutch2\] and lemma \[lemma:cwpair\] below. Otherwise, assume that $X$ is a ${{\bf\mathfrak{{F}}}}$-stratified bundle. Since $X$ is obtained by attaching ${{\bf\mathfrak{{F}}}}$-bundles to ${{\bf\mathfrak{{F}}}}$-bundles via ${{\bf\mathfrak{{F}}}}$-maps, we can apply a finite number of times corollary \[coro:clutch2\] below and obtain the result. On the other hand assume that $\bar X$ is a stratified space and that $X\to \bar X$ is a ${{\bf\mathfrak{{F}}}}$-bundle. By applying a finite number of times lemma \[lemma:attaching\] it is possible to show that $X$ is a ${{\bf\mathfrak{{F}}}}$-stratified bundle.
\[ex:3\] Let ${{\bf\mathfrak{{F}}}}$ be the category consisting of finite sets $\{1,2,\dots,n\}$ and permutations of these sets. A ${{\bf\mathfrak{{F}}}}$-family $X$ is a finite-to-one map $X\to \bar X$. By theorem \[theo:bundle\], if $X$ is a ${{\bf\mathfrak{{F}}}}$-complex, then $p_X$ is locally trivial and hence a covering map.
\[lemma:clutch1\] Let ${{\bf\mathfrak{{F}}}}$ be a groupoid. Let $Y\to \bar Y$ be a ${{\bf\mathfrak{{F}}}}$-bundle, $\bar Z$ a set and $h{\colon}V\times \bar Z \times S^{n-1} \to Y$ a ${{\bf\mathfrak{{F}}}}$-map which is the attaching map of $X=Y\cup_h (V\times \bar Z \times D^n)$. Given a set $\bar U \subset \bar Y$ open in $\bar Y$ such that $Y|\bar U$ is trivial, there is an open set $\bar U'' \subset \bar X$ such that $\bar U'' \cap \bar Y = \bar U$ and such that $X|\bar U''$ is trivial. As a consequence, the push-out space $X=Y\cup_h (V\times \bar Z\times D^n)$ is a ${{\bf\mathfrak{{F}}}}$-bundle.
Let ${\Phi}$ denote the characteristic map ${\Phi}{\colon}V\times \bar Z\times D^n \to X$ and $\bar {\Phi}$, $\bar h$ the maps induced on the base spaces. The restriction of $V\times \bar Z \times S^{n-1}$ to $\bar h^{-1}\bar U$ is trivial, and with a suitable change of coordinates (which exists since ${\mathrm{Aut}}_{{\bf\mathfrak{{F}}}}(V)$ is a topological group) we can assume that the restriction $$h_0 {\colon}V \times \bar h^{-1}\bar U \to Y|\bar U \approx V\times \bar U$$ is of the form $1_V\times \bar h$. Now, there is an open set $\bar U' \subset \bar Z\times D^n$ with the property that $\bar U' \cap \bar Z\times S^{n-1} = \bar h^{-1}\bar U$. Since $\bar X = \bar Y \cup_{\bar h} (\bar Z \times D^n)$ the image $\bar U'' = \bar {\Phi}(\bar U') \cup \bar U$ is open in $\bar X$. Moreover, since $V$ is locally compact Hausdorff the following diagram is a push-out. [$$\begin{aligned}\xymatrix{V \times \bar h^{-1}\bar U
\ar[r]
\ar[d]
&
V \times \bar U \ar[d]
\\
V\times \bar U'
\ar[r]
&
V \times \bar U''\\
}\end{aligned}$$]{}But by assumption $X$ is the push-out $Y\cup_h(V\times \bar Z \times D^n)$, and the restriction of the push-out to the pre-images of $\bar U''$ yields a push-out [$$\begin{aligned}\xymatrix{V \times \bar h^{-1}\bar U
\ar[r]
\ar[d]
&
V \times \bar U
\ar[d]
\\
V\times \bar U'
\ar[r]
&
X|\bar U''.\\
}\end{aligned}$$]{}Therefore $X|\bar U'' \approx V \times \bar U''$ by an isomorphism which is an extension of the chosen isomorphism $Y|\bar U \approx V \times \bar U$. This implies that every point in $\bar Y$ has a neighborhood $\bar U''$ in $\bar X$ such that $X|\bar U''$ is trivial. On the other hand, if $x$ is a point in $\bar X{\smallsetminus\/}\bar Y$, then there is a neighborhood $\bar U''$ of $x$ contained in $\bar X {\smallsetminus\/}\bar Y$ and in this case $X|\bar U''$ is trivial because it is ${{\bf\mathfrak{{F}}}}$-homeomorphic to $ V\times \bar {\Phi}^{-1} \bar U'' \subset V\times \bar Z \times e^n$.
\[coro:clutch2\] Let $(X,Y)$ a relative ${{\bf\mathfrak{{F}}}}$-complex. Assume that ${{\bf\mathfrak{{F}}}}$ is a groupoid and that $Y \to \bar Y$ is a ${{\bf\mathfrak{{F}}}}$-bundle. Then $X\to \bar X$ is a ${{\bf\mathfrak{{F}}}}$-bundle.
Consider the cellular filtration $$Y=X_{-1} \subset X_0 \subset X_1 \subset \dots \subset X_n
\subset \dots \subset X=\lim_{n} X_n.$$ By applying inductively lemma \[lemma:clutch1\] we see that $X_n$ is a ${{\bf\mathfrak{{F}}}}$-bundle over $\bar X_n$ for every $n\geq 0$. Moreover, there is a sequence of sets $\bar U_n \subset \bar X_n$ open in $\bar X_n$ such that $\bar U_n \cap \bar X_{n-1} = \bar U_{n-1}$ and ${{\bf\mathfrak{{F}}}}$-isomorphisms $\alpha_n{\colon}X_n | \bar U_n \approx V \times \bar U_n$. Moreover $\alpha_n$ can be chosen to be extension of $\alpha_{n-1}$ (see the proof of \[lemma:clutch1\]). Since $V$ is locally compact Hausdorff we can take limits, and obtain a ${{\bf\mathfrak{{F}}}}$-homeomorphism $$X|\bar U \approx V \times \bar U,$$ where $\bar U = \lim_{n} \bar U_n$.
The principal bundle theorem {#section:principalbundle}
============================
It is well known that each fibre bundle $X\to \bar X$ with fibre $V$ and structure group $G$ yields the associated principal bundle $X_G\to \bar X$ with fibre $G$ such that $X_G$ is a right $G$-space for which there is an isomorphism of bundles. [$$\label{diag:5:1}\begin{aligned}\xymatrix{X \ar[rr]^{\cong} \ar[dr] & & X_G\times_G V \ar[dl]\\
& \bar X\\
}\end{aligned}$$]{}Moreover Steenrod [@steenrod] p. 39 points out that in case $G$ is a subgroup of the group of homeomorphisms of $V$ with the compact open topology then the principal bundle $$\label{eq:steenrod}
X_G = X^V$$ is the function space $X^V$ of “admissible” maps $V\to X$ with the compact open topology. The action of $g\in G$ on $\alpha\in X^V$ is given by composition $$\alpha \cdot g = \xymatrix@1{V\ar[r]^g & V \ar[r]^\alpha & X \\} \in X^V.$$ In this section we generalize these results for ${{\bf\mathfrak{{F}}}}$-complexes and ${{\bf\mathfrak{{F}}}}$-stratified bundles.
\[defi:phiV\] Given a faithful fibre functor ${F}{\colon}{{\bf\mathfrak{{F}}}}\to {\mathbf{Top}}$ and an object $V\in {{\bf\mathfrak{{F}}}}$ we obtain a new fibre functor $${F}^V{\colon}{{\bf\mathfrak{{F}}}}\to {\mathbf{Top}}$$ which carries an object $W$ in ${{\bf\mathfrak{{F}}}}$ to the space ${F}^V(W) = \hom_{{\bf\mathfrak{{F}}}}(V,W)$ with the compact open topology. The functor ${F}^V$ need not be faithful, even though ${F}$ is faithful.
Since for every $V$ the fibre ${F}(V)$ is Hausdorff, we see that for each $W\in {{\bf\mathfrak{{F}}}}$ the fibre ${F}^V(W)=W^V$ of ${F}^V$ is Hausdorff. Moreover, since $W$ and $V$ are $2^\circ$ countable, also $W^V$ is $2^\circ$ countable (see e.g. [@dugundji]). We will show that if $({{\bf\mathfrak{{F}}}},{F})$ is a NKC structure category then $W^V$ is locally compact so that ${F}^V$ fulfills assumption $(*)$.
The definition of a NKC structure category is the following. First, a family $\mathcal{K}$ of compact sets of $V$ is termed *generating* if for every ${{\bf\mathfrak{{F}}}}$-family $Y$ the subsets $$N_{K,U} = \{ f\in Y^V, f(K)\subset U\}$$ with $K\in \mathcal{K}$ and $U$ open in $Y$ yield a sub-basis for the topology of $Y^V$.
\[defi:NKC\] A structure category ${{\bf\mathfrak{{F}}}}$ with faithful fibre functor ${F}$ has the **NKC*-property* if for every $V\in {{\bf\mathfrak{{F}}}}$ there is a generating family of compact sets $\mathcal{K}$ such that for every $K\in \mathcal{K}$, every $W\in {{\bf\mathfrak{{F}}}}$ and every compact subset $C\subset W$ the subspace $$N_{K,C} \subset \hom_{{\bf\mathfrak{{F}}}}(V,W) = W^V$$ is compact. Examples are given in section \[section:NKC\] below.
\[rem:NKC\] Since the spaces $W$ in ${{\bf\mathfrak{{F}}}}$ are assumed locally compact this implies that for a NKC category the function spaces $W^V = \hom_{{\bf\mathfrak{{F}}}}(V,W)$ are locally compact. Furthermore, it is not difficult to see that a closed subcategory of a NKC category is also a NKC subcategory.
\[ex:1\] Let ${{\bf\mathfrak{{F}}}}$ denote the category consisting of a finite dimensional vector space $V$ and morphisms given by elements in a closed subgroup $G$ of $GL(V)$. The fibre functor is the embedding ${F}{\colon}{{\bf\mathfrak{{F}}}}\to {\mathbf{Top}}$. Then, by \[theo:bundle\], a ${{\bf\mathfrak{{F}}}}$-complex $X$ is a vector bundle with structure group $G$ and fibre $V$. We will see in theorem \[theo:NKC\] that $({{\bf\mathfrak{{F}}}},{F})$ has the NKC property. Therefore by theorem \[theo:principal\] the principal bundle $X^V$ is a $({{\bf\mathfrak{{F}}}},{F}^V)$-complex and hence a $({{\bf\mathfrak{{F}}}},{F})$-bundle by \[theo:bundle\]. This is the result of Steenrod in \[eq:steenrod\].
\[ex:2\] Let $G$ be a compact Lie group and ${{\bf\mathfrak{{F}}}}$ its orbit category (see \[ex:orbit\]). The fibre functor ${F}{\colon}{{\bf\mathfrak{{F}}}}\to {\mathbf{Top}}$ is the embedding. Then a smooth $G$-manifold $X$ has a structure of a ${{\bf\mathfrak{{F}}}}$-complex where $\bar X=X/G$ and $p_X {\colon}X \to \bar X$ is the projection onto the orbit space. For every closed subgroup $H\subset G$ the orbit $V=G/H$ is an object of ${{\bf\mathfrak{{F}}}}$ and the function space $X^V$ is homeomorphic to the fixed subspace $X^H \subset X$ via the evaluation map at $1\in G$. We will see as a consequence of proposition \[propo:theo:allcompact\] that the orbit category ${{\bf\mathfrak{{F}}}}$ with the embedding functor ${F}{\colon}{{\bf\mathfrak{{F}}}}\to {\mathbf{Top}}$ has the NKC property. Hence by theorem \[theo:principal\] the space $X^H$ is a ${{\bf\mathfrak{{F}}}}$-complex with fibre functor ${F}^{G/H}$.
We now define for every $({{\bf\mathfrak{{F}}}},{F})$-family $Y$ the associated $({{\bf\mathfrak{{F}}}},{F}^V)$-family $Y^V$ as follows. Let $Y^V$ denote the space of all ${{\bf\mathfrak{{F}}}}$-maps $V\to Y$ with the compact-open topology. Then we obtain the projection $$Y^V \to \bar Y$$ which sends a ${{\bf\mathfrak{{F}}}}$-map $a{\colon}V \to Y$ to the point $p_Ya(v_0)\in \bar Y$, where $v_0\in V$ is an arbitrary point of the fibre $V$. The pre-image of a point $b\in \bar Y$ under this projection is equal to ${F}\hom(V,Y_b) = Y_b^V$. Thus $Y^V \to \bar Y$ is a $({{\bf\mathfrak{{F}}}},{F}^V)$-family.
We recall some properties of the function spaces $X^V$ and $W^V=\hom_{{\bf\mathfrak{{F}}}}(V,W)$. By assumption $(*)$ the images under the fibre functor ${F}$ of all the objects in ${{\bf\mathfrak{{F}}}}$ are locally compact second-countable and Hausdorff. This implies that they are metrizable. Since every object $V$ is metrizable, the automorphism group ${\mathrm{Aut}}(V)$ is a topological group with the compact-open topology induced by the inclusion ${\mathrm{Aut}}(V) \subset \hom_{{\bf\mathfrak{{F}}}}(V,V)$. Given a ${{\bf\mathfrak{{F}}}}$-map $f{\colon}X \to Y$, let $f^V$ denote the map $f^V{\colon}X^V \to Y^V$ defined by composing with $f$ the ${{\bf\mathfrak{{F}}}}$-maps $V\to X$. Moreover for $\varphi{\colon}V \to W$ in ${{\bf\mathfrak{{F}}}}$ let $X^\varphi{\colon}X^W \to X^V$ be the map induced by $\varphi$. Let us recall that given a ${{\bf\mathfrak{{F}}}}$-map $f{\colon}Z\times V \to X$ the adjoint $\hat f$ of $f$ is the function $\hat f{\colon}Z\to X^V$ defined by $\hat f(z)(v) = f(z,v)$ for every $z\in Z$ and $v\in V$.
\[lemma:adjoint\] For every space $Z$ a ${{\bf\mathfrak{{F}}}}$-function $f{\colon}Z\times V \to X$ is continuous if and only if the adjoint $\hat f: Z \to X^V$ is continuous.
Because $V$ is locally compact and Hausdorff, a function $f{\colon}Z \times V \to X$ is continuous if and only if its adjoint $\hat f: Z \to \mathrm{Map}(V,X)$ is continuous, where $\mathrm{Map}(V,X)$ denotes the space of all (not necessarily controlled by ${{\bf\mathfrak{{F}}}}$) maps from $V$ to $X$ with the compact-open topology. But a function $f{\colon}Z\times V \to X$ is a ${{\bf\mathfrak{{F}}}}$-map if and only if its adjoint sends $Z$ into the subspace $X^V \subset \mathrm{Map}(V,X)$. This implies the lemma.
\[coro:adjoint\] The evaluation map $X^V\times V \to X$ which sends $(g,v)$ to $g(v)$ is continuous. Moreover, if $f{\colon}X\to Y$ is continuous then the induced function $f^V{\colon}X^V \to Y^V$ is continuous. Also, given $\varphi{\colon}V \to W$ the induced map $X^\varphi{\colon}X^W \to X^V$ is continuous.
The evaluation map is the adjoint of the identity of $X^V$, hence continuous by lemma \[lemma:adjoint\]. The evaluation at $v\in V$ is the restriction of the evaluation to the subspace $X^V\times\{v\}$ of $X^V\times V$ and hence continuous. The projection $p{\colon}X^V \to \bar X$ is the composition of the evaluation at any $v_0\in V$ with the projection $p_X$ and hence continuous. The induced function $f^V$ is the adjoint of the composition $X^V\times V \to X \to Y$ where the first arrow is the evaluation and the second is $f$. To see that $X^\varphi$ is continuous it is enough to see that it is the adjoint of the composition $X^W\times V \to X^W \times W \to X$, where the first arrow is $1\times \varphi$ while the second is the evaluation.
The next result is a crucial observation which shows that the function space $X^V$ of a stratified bundle $X$ is again a stratified bundle.
\[theo:principal\] Let ${{\bf\mathfrak{{F}}}}$ be a NKC category with fibre functor ${F}$. If $X$ is a $({{\bf\mathfrak{{F}}}},{F})$-stratified bundle and $V$ an object in ${{\bf\mathfrak{{F}}}}$ then the function space $X^V$ is also a $({{\bf\mathfrak{{F}}}},{F}^V)$-stratified bundle.
If $X$ is a $({{\bf\mathfrak{{F}}}},{F})$-complex then this follows from corollary \[coro:main\] below. In the general case one can apply a finite number of times corollary \[coro:nkc\] and lemma \[lemma:nnep2\].
Let ${\mathbf{Top}}_{\bar X}$ be the category of spaces over $\bar X$; objects are maps $X\to \bar X$. Given a ${{\bf\mathfrak{{F}}}}$-family $X$, the function spaces $X^V$ for every object $V$ in ${{\bf\mathfrak{{F}}}}$ yield a functor $X^\circ$ which carries $V$ to $X^\circ(V) = X^V$. Moreover, given a ${{\bf\mathfrak{{F}}}}$-map $f{\colon}X \to Y$ there is a natural transformation $f^\circ: X^\circ \to Y^\circ$ defined by $f^V{\colon}X^V \to Y^V$. Let ${\mbox{${{\bf\mathfrak{{F}}}}^{\mathrm{op}}$-${{{\mathbf{Diag}}}}$}}_{\bar X}$ denote the diagram category in which the objects are functors ${{\bf\mathfrak{{F}}}}^{\mathrm{op}}\to {\mathbf{Top}}_{\bar X}$ and the morphisms are the natural transformations between functors. The operator $(-)^\circ$ sends a ${{\bf\mathfrak{{F}}}}$-family $X$ to the ${{\bf\mathfrak{{F}}}}^{\mathrm{op}}$-diagram $X^\circ$ and a ${{\bf\mathfrak{{F}}}}$-map $f$ to the natural transformation $f^\circ$. Thus ${(-)}^\circ$ is a functor $${(-)}^\circ{\colon}{\mbox{${{\bf\mathfrak{{F}}}}$-${{{\mathbf{Top}}}}$}}_{\bar X} \to {\mbox{${{\bf\mathfrak{{F}}}}^{\mathrm{op}}$-${{{\mathbf{Diag}}}}$}}_{\bar X}.$$
The ${{\bf\mathfrak{{F}}}}$-diagram $X^\circ$ is the generalization of the concept of principal bundle. In fact, if $X\to \bar X$ is a fibre bundle with structure group $G$ and fibre $V$ and ${{\bf\mathfrak{{F}}}}=G \to {\mathbf{Top}}$ is given by the $G$-space $V$ then the diagram $X^\circ{\colon}{{\bf\mathfrak{{F}}}}^{\mathrm{op}}\to {\mathbf{Top}}$ coincides with the $G$-space $X_G$ given by the principal bundle $X_G\to\bar X$ associated to $X$, see \[eq:steenrod\].
For a topological enriched category ${{\bf\mathfrak{{F}}}}$ there is a naturally associated functor ${\widehat{\hom}}{\colon}{{\bf\mathfrak{{F}}}}\to {{\bf\mathfrak{{F}}}}^{\mathrm{op}}$-${{\mathbf{Diag}}}$ which sends an object $W$ of ${{\bf\mathfrak{{F}}}}$ to the ${{\bf\mathfrak{{F}}}}^{\mathrm{op}}$-diagram $\hom_{{\bf\mathfrak{{F}}}}(-,W) = W^\circ$. In case $X$ is a ${{\bf\mathfrak{{F}}}}$-stratified bundle and $({{\bf\mathfrak{{F}}}},{F})$ is a NKC structure category then by theorem \[theo:principal\] the function space $X^V$ is not only a space over $\bar X$ but a $({{\bf\mathfrak{{F}}}},{F}^V)$-stratified bundle. Moreover this holds for every object $V\in {{\bf\mathfrak{{F}}}}$ in a compatible way. Therefore the diagram $X^\circ$ given by stratified bundles $X^V$ leads to the concept of a *stratified bundle diagram*. For this consider the definition of a stratified bundle in \[defi:stratifiedvectorbundle\] and replace the category ${\mathbf{Top}}$ by the category of diagrams in ${\mathbf{Top}}$. This yields a notion of a stratified bundle diagram. If the stratification is given by the skeletal filtration of the base space $\bar X$ then a $({{\bf\mathfrak{{F}}}},{\widehat{\hom}})$-stratified bundle diagram is the same as a “free CW-complex” in the sense of Davis–Lück [@davislueck].
\[defi:principal\] If ${{\bf\mathfrak{{F}}}}$ is a topologically enriched category, then a $({{\bf\mathfrak{{F}}}},{\widehat{\hom}})$-stratified bundle diagram in the category ${{\bf\mathfrak{{F}}}}^{\mathrm{op}}$-${{\mathbf{Diag}}}_{\bar X}$ is termed *principal stratified bundle diagram*. This is a ${{\bf\mathfrak{{F}}}}^{\mathrm{op}}$-diagram of stratified ${{\bf\mathfrak{{F}}}}$-bundles over a stratified space $\bar X$. For example $X^\circ{\colon}{{\bf\mathfrak{{F}}}}^{\mathrm{op}}\to {\mathbf{Top}}_{\bar X}$ is such a principal stratified bundle diagram.
If ${{\bf\mathfrak{{F}}}}$ has a single object $V$ and $\hom_{{\bf\mathfrak{{F}}}}(V,V)$ is a topological group, then a $({{\bf\mathfrak{{F}}}},{\widehat{\hom}})$-stratified bundle diagram is nothing but a principal $G$-bundle with structure group $G=\hom_{{\bf\mathfrak{{F}}}}(V,V)$ over $\bar X$, see theorem \[theo:bundle\].
Now consider a principal stratified bundle diagram $P{\colon}{{\bf\mathfrak{{F}}}}^{{\mathrm{op}}} \to {\mathbf{Top}}_{\bar X}$ over $\bar{X}$ and a fibre functor ${F}{\colon}{{\bf\mathfrak{{F}}}}\to {\mathbf{Top}}$. We can build a $({{\bf\mathfrak{{F}}}},{F})$-stratified bundle $X\to \bar X$ in ${\mathbf{Top}}$ as the coend of the two functors $P$ and ${F}$: $$X = P\otimes_{{\bf\mathfrak{{F}}}}{F}= \int^{V\in {{\bf\mathfrak{{F}}}}} P(V)\times {F}(V).$$ We recall the coend construction: the space $P\otimes_{{{\bf\mathfrak{{F}}}}} {F}$ is the quotient space of the coproduct $$\coprod_{V\in {{\bf\mathfrak{{F}}}}} P(V)\times {F}(V)$$ with the identification $(P(\alpha)(x),y) \sim (x,{F}(\alpha)(y) )$ for every morphism $\alpha\in \hom_{{\bf\mathfrak{{F}}}}(V,W)$, $x\in P(W)$ and $y\in {F}(V)$. There is a well-defined map $X \to \bar X$ and it is not difficult to show:
\[lemma:coend\] $P\otimes_{{\bf\mathfrak{{F}}}}{F}$ is a $({{\bf\mathfrak{{F}}}},{F})$-family.
See e.g. [@huse], page 44, for the case of a groupoid ${{\bf\mathfrak{{F}}}}$. It suffices to show it in the case $P$ is a $({{\bf\mathfrak{{F}}}},{\widehat{\hom}})$-point, that is a diagram $P$ over a point. Since $P$ is principal, there is an object $W$ of ${{\bf\mathfrak{{F}}}}$ such that $P(V)=\hom_{{{\bf\mathfrak{{F}}}}}(V,W)$ for every $V\in {{\bf\mathfrak{{F}}}}$, or, equivalently, $P = W^\circ$. If $g{\colon}V \to V'$ is a morphism in ${{\bf\mathfrak{{F}}}}$, then $P(g){\colon}W^{V'} \to W^V$ is defined by $P(g)(\alpha) = \alpha g$. Thus the coend $W^\circ \otimes_{{\bf\mathfrak{{F}}}}{F}$ is defined as $$\left(\coprod_{V\in {{\bf\mathfrak{{F}}}}} W^V \times {F}(V) \right)/\sim,$$ where $(\alpha g, x) \sim (g, {F}(\alpha)(x) )$ for every $g\in \hom_{{\bf\mathfrak{{F}}}}(V,V')$, $\alpha\in W^V$ and $x\in {F}(V)$. Now it is easy to see that for every $W$ $$W^\circ \otimes_{{\bf\mathfrak{{F}}}}{F}= {F}(W)$$ and that given a morphism $\alpha^\circ {\colon}W^\circ\otimes_{{{\bf\mathfrak{{F}}}}} {F}\to
{W'}^\circ \otimes {F}$ the following diagram commutes: [$$\label{diag:nota}\begin{aligned}\xymatrix{W^\circ \otimes_{{\bf\mathfrak{{F}}}}{F}\ar[r]^{\alpha^\circ} \ar[d]^{=} &
{W'}^\circ \otimes_{{\bf\mathfrak{{F}}}}{F}\ar[d]^{=} \\
{F}(W) \ar[r]^{{F}(\alpha)} & {F}(W') \\
}\end{aligned}$$]{}
\[theo:principalbundle\] Let ${{\bf\mathfrak{{F}}}}$ be a structure category with a faithful fibre functor ${F}$ so that ${{\bf\mathfrak{{F}}}}$ is topological enriched by the compact open topology. Let $({{\bf\mathfrak{{F}}}},{F})$ be a NKC structure category and $X$ a $({{\bf\mathfrak{{F}}}},{F})$-stratified bundle. Then the associated diagram $X^\circ$ is a principal stratified bundle diagram with fibre functor ${\widehat{\hom}}$. Conversely, if $P$ is a principal $({{\bf\mathfrak{{F}}}},{\widehat{\hom}})$-stratified bundle diagram and ${F}$ a fibre functor, then the coend $P\otimes_{{\bf\mathfrak{{F}}}}{F}$ is a stratified bundle in ${\mathbf{Top}}$ with structure category ${{\bf\mathfrak{{F}}}}$ and fibre functor ${F}$. Moreover $$X^\circ \otimes_{{\bf\mathfrak{{F}}}}{F}= X.$$
This result generalizes the classical equation $$X_G\times_G V = X$$ for the principal bundle $X_G$ associated to $X$ in \[diag:5:1\]. This construction can be generalized. Let $\varphi{\colon}G \to H$ be a continuous homomorphism between topological groups and let $G$ ($H$) be a topological transformation group for $V$ ($W$). Then a *$(G,V)$-bundle* $X$ (i.e. a bundle with fibre $V$ and structure group $G$) yields the principal bundle $X_G$ which in turn yields the *$\varphi$-associated bundle* $$\varphi_\#(X) = X_G \times_G \varphi^*W \to \bar X.$$ Here $\varphi^*W$ is the $G$-space with the $G$-action induced by $\varphi$, that is $g\cdot w = \varphi(g)\cdot w$ for $g\in G$ and $w\in W$. The associated bundle $\varphi_\#(X)$ has fibre $W$ and structure group $H$. Hence $\varphi$ induces the functor
$$\label{eq:varphi}
\varphi_\#{\colon}{\mbox{$(G,V)$-${\mathbf{Bundles}}_{\bar X}$}} \to {\mbox{$(H,W)$-${\mathbf{Bundles}}_{\bar X}$}}.$$
We now generalize this functor for stratified bundles.
Let $({{\bf\mathfrak{{F}}}},{F})$ and $({{\bf\mathfrak{{G}}}},{G})$ be structure categories with faithful fibre functor ${F}$, ${G}$ so that ${{\bf\mathfrak{{F}}}}$ and ${{\bf\mathfrak{{G}}}}$ have the compact open topology. Let $$\varphi{\colon}{{\bf\mathfrak{{F}}}}\to {{\bf\mathfrak{{G}}}}$$ be a continuous functor. If $({{\bf\mathfrak{{F}}}},{F})$ is a NKC structure category and $X$ a $({{\bf\mathfrak{{F}}}},{F})$-stratified bundle then the principal stratified bundle diagram $X^\circ$ is defined and we obtain the *$\varphi$-associated stratified bundle* $$\varphi_\#(X) = X^\circ \otimes_{{\bf\mathfrak{{F}}}}(G\varphi) \to \bar X.$$ Here $\varphi_\#(X)$ is a $({{\bf\mathfrak{{G}}}},{G})$-stratified bundle. Hence $\varphi$ induces the functor $$\varphi_\#{\colon}({{\bf\mathfrak{{F}}}},{F})\mbox{-}{\mathbf{Stra}}_{\bar X} \to
({{\bf\mathfrak{{G}}}},{G})\mbox{-}{\mathbf{Stra}}_{\bar X}.$$
We have seen that, as a consequence of theorem \[theo:principal\], for each $V\in {{\bf\mathfrak{{F}}}}$ the function space $X^V$ is obtained by gluing $W^V$-bundles $M_i^V$ along with maps $h_i^V{\colon}A_i^V\subset M_i \to X_{i-1}^V$, for suitable objects $W\in {{\bf\mathfrak{{F}}}}$. In other words, the diagram $X^\circ$ is obtained by gluing the ${{\bf\mathfrak{{F}}}}$-diagrams $M_i^\circ$ (which are $({{\bf\mathfrak{{F}}}},{\widehat{\hom}})$-bundles) along with ${{\bf\mathfrak{{F}}}}$-diagram maps $h_i^\circ{\colon}A_i^V \to X_{i-1}^\circ$. Thus $X^\circ$ is a $({{\bf\mathfrak{{F}}}},{\widehat{\hom}})$-stratified bundle, that is a ${{\bf\mathfrak{{F}}}}$-stratified bundle with diagram-fibre functor ${\widehat{\hom}}$. By definition this is a principal stratified bundle diagram.
Conversely, let $P$ be a principal stratified bundle diagram and ${F}$ a fibre functor. It is not assumed that ${F}$ is faithful, but that the spaces ${F}V$ are second-countable locally compact Hausdorff spaces, see assumption $(*)$. Hence, if the $i$-skeleton of $P$ is obtained by attaching a $({{\bf\mathfrak{{F}}}},{\widehat{\hom}})$-bundle $\hat M_i$ to $P_{i-1}$ via a ${{\bf\mathfrak{{F}}}}$-diagram map $\hat h_i{\colon}\hat A_i\subset \hat M_i \to P_{i-1}$, i.e. $P_i = \hat M_{i} \cup_{\hat h_i} P_{i-1}$, the coends are corners of a push-out diagram in ${\mbox{${{\bf\mathfrak{{F}}}}$-${{{\mathbf{Top}}}}$}}$: [$$\begin{aligned}\xymatrix{\hat A_i\otimes_{{\bf\mathfrak{{F}}}}{F}\ar[r]^{\hat h_i\times_{{\bf\mathfrak{{F}}}}{F}} \ar[d]
&
P_{i-1}\otimes_{{\bf\mathfrak{{F}}}}{F}\ar[d] \\
\hat M_i\otimes_{{{\bf\mathfrak{{F}}}}} {F}\ar[r]
& P_i\otimes_{{{\bf\mathfrak{{F}}}}}{F}\\
}\end{aligned}$$]{}Thus $X$ has a filtration given by $X_i = P_i\otimes_{{\bf\mathfrak{{F}}}}{F}$, for all $i\geq 0$, and each $X_i$ is obtained by attaching $M_i = \hat M_i \otimes_{{\bf\mathfrak{{F}}}}{F}$ to $X_{i-1}$ via the map $$h_i= \hat h_i \otimes_{{\bf\mathfrak{{F}}}}{F}{\colon}A_i = \hat A_i\otimes_{{\bf\mathfrak{{F}}}}{F}\to
X_{i-1}.$$ Moreover, since $$\lim_{i\geq 0} \left( P_i\otimes_{{\bf\mathfrak{{F}}}}{F}\right) =
\left( \lim_{i\geq 0} P_i \right) \otimes_{{\bf\mathfrak{{F}}}}{F},$$ $X$ is equal to the colimit $X=\lim_{i\geq 0} X_i$. To show that it is a $({{\bf\mathfrak{{F}}}},{F})$-stratified bundle we need only to show that the attaching maps are ${{\bf\mathfrak{{F}}}}$-maps. But $\hat h_i$ are $({{\bf\mathfrak{{F}}}},{\widehat{\hom}})$-maps, and by diagram \[diag:nota\] the coends of such maps are $({{\bf\mathfrak{{F}}}},{F})$-maps. The proof is hence complete.
Pull-back of a push-out ${{\bf\mathfrak{{F}}}}$-family {#section:pullback}
======================================================
In this section we complete the proof of the pull-back theorem \[theo:pullback\].
Consider a ${{\bf\mathfrak{{F}}}}$-family $Y'$, a ${{\bf\mathfrak{{F}}}}$-CW-pair $(M',A')$ and a ${{\bf\mathfrak{{F}}}}$-map $h'{\colon}A' \to Y'$. Then we can glue $M'$ to $Y'$ and obtain a ${{\bf\mathfrak{{F}}}}$-family $X'$, as in the following diagram. [$$\label{diag:fri1}\begin{aligned}\xymatrix{A' \ar[rrr]^{h'}
\ar[ddd]^{j_{A'}}
\ar[rd]^{p_{A'}}
&
&
&
Y' \ar[ddd]^{j_{Y'}}
\ar[dl]^{p_{Y'}}
\\
&
\bar A'
\ar[r]^{\bar h'}
\ar[d]^{j_{\bar A'}}
&
\bar Y' \ar[d]^{j_{\bar Y'}}
\\
&
\bar M'
\ar[r]^{\bar {\Phi}'}
&
\bar X' \\
M' \ar[rrr]^{{\Phi}'} \ar[ur]^{p_{M'}}
& & &
X' \ar[ul]^{p_{X'}}\\
}\end{aligned}$$]{} Now consider a space $\bar Y$, a CW-pair $(\bar M, \bar A)$ and a map $\bar h{\colon}\bar A \to \bar Y$. Let $\bar X$ be the push-out space $\bar X = \bar Y \cup_{\bar h} \bar M$. Assume that maps $\tilde f_A$, $\tilde f$, $\bar f_{\bar Y}$ and $\bar f$ are given, such that the following diagram commutes. [$$\label{diag:fri2}\begin{aligned}\xymatrix{\bar A' \ar[rrr]^{\bar h'}
\ar[ddd]^{j_{A'}}
&
&
&
\bar Y'
\ar[ddd]^{j_{\bar Y'}}
\\
&
\bar A
\ar[ul]^{\tilde f_A}
\ar[r]^{\bar h}
\ar[d]^{j_{\bar A}}
&
\bar Y
\ar[d]^{j_{\bar Y}}
\ar[ur]^{\bar f_{\bar Y}}
\\
&
\bar M
\ar[r]^{\bar {\Phi}}
\ar[dl]^{\tilde f}
&
\bar X
\ar[dr]^{\bar f}
\\
\bar M'
\ar[rrr]_{\bar {\Phi}'}
&&&
\bar X'
\\
}\end{aligned}$$]{} Then the pull-back families $X=\bar f^* X'$, $Y=\bar f_{\bar Y}^* Y'$, $A= \tilde f_{A}^* A'$ and $M = \tilde f^*M'$ are defined. Moreover, we can define the maps $h=(\bar h,h'){\colon}A \to Y$, $j_Y=(j_{\bar Y},j_{Y'}){\colon}Y \to X$, $j_A=(j_{\bar A},j_{A'}){\colon}A \to M$ and ${\Phi}=(\bar {\Phi}, {\Phi}'){\colon}M \to X$. They fit into the following diagram. [$$\label{diag:fri3}\begin{aligned}\xymatrix{\bar A \times A'
&
\ar@{ >->}[l]
A \ar[r]^{h}
\ar[d]^{j_A}
&
Y
\ar@{ >->}[r]
\ar[d]^{j_Y}
&
\bar Y\times Y'
\\
\bar M \times M'
&
\ar@{ >->}[l]
M
\ar[r]^{{\Phi}}
&
X
\ar@{ >->}[r] &
\bar X \times X'.
}\end{aligned}$$]{} The aim of this section is to prove the following proposition.
\[propo:theo:pushout2\] Assume that $\bar M$ and $\bar M'$ are locally finite and countable CW-complexes and that $\bar f(\bar X {\smallsetminus\/}\bar Y) \subset \bar X' {\smallsetminus\/}\bar Y'$. Then the pull-back family $X=\bar f^*X'$ is obtained by attaching the pull-back ${{\bf\mathfrak{{F}}}}$-family $M=\tilde f^* M'$ to $Y=\bar f_{\bar Y}^*Y'$ via the ${{\bf\mathfrak{{F}}}}$-map $h{\colon}A\subset M \to Y$ induced by $(\bar h, h')$, i.e. the middle square in diagram \[diag:fri3\] is a push-out.
The proof of proposition \[propo:theo:pushout2\] will be the content of the rest of the section.
\[lemma:step1\] The maps in \[diag:fri3\] are well-defined and the diagram commutes.
Consider $(\bar a, a')\in A$. Then $\tilde f_A(\bar a) = p_{A'}(a')$, hence $$\bar f_{\bar Y} \bar h (\bar a) =
\bar h' \tilde f_A (\bar a) = \bar h'
p_{A'}(a')
=
p_{Y'} h'(a').$$ Hence $(\bar h(\bar a), h'(a') ) \in Y$. On the other hand, if $(\bar y, y' ) \in Y$, then $\bar f_{\bar Y}(\bar y) = p_{Y'}(y')$, hence $$\bar f j_{\bar Y} (\bar y) =
j_{\bar Y'} \bar f_{\bar Y} (\bar y) =
j_{\bar Y'} p_{Y'}(y') =
p_{X'} j_{Y'} (y').$$ That is, $j_Y( \bar y, y') = (j_{\bar Y} (\bar y), j_{Y'}(y') ) \in X$. The same argument can be applied literally to $j_A$ and ${\Phi}$. The diagram commutes since it is the restriction of a commutative diagram on the Cartesian products.
\[lemma:step2\] The maps $j_A{\colon}A \to M$ and $j_Y{\colon}Y \to X$ are a closed inclusions of ${{\bf\mathfrak{{F}}}}$-families.
Consider the following diagram. The left square is a pull-back (by definition) and the right square is a pull-back (the pair $(M',A')$ is a ${{\bf\mathfrak{{F}}}}$-CW-pair). [$$\label{diag:fri4}\begin{aligned}\xymatrix{A \ar[r]
\ar[d]
&
A'
\ar[r]\ar[d]
&
M' \ar[d]\\
\bar A \ar[r]
&
\bar A' \ar[r]
&
\bar M' \\
}\end{aligned}$$]{}Now consider this diagram. [$$\begin{aligned}\xymatrix{A \ar[r] \ar[d]
&
M \ar[r] \ar[d]
&
M' \ar[d]
\\
\bar A \ar[r]
&
\bar M \ar[r]
& \bar M'\\
}\end{aligned}$$]{}The composition of the two squares is equal to the composition of the two squares in diagram \[diag:fri4\], hence it is a pull-back. The right square is a pull-back by definition, hence the left square is a pull-back. (see e.g. exercise 8 of [@macl], page 72). Since $\bar A \subset \bar M$ is a closed inclusion, $j_A$ is a closed inclusion. The proof is the same for $j_Y$.
Now consider the maps $q={\Phi}\sqcup j_{Y}$, $\bar q = \bar {\Phi}\sqcup j_{\bar Y}$, $\bar q' = \bar {\Phi}' \sqcup j_{\bar Y'}$ and $q' = {\Phi}' \sqcup j_{Y'}$. They can be arranged in the following diagram. [$$\label{diag:fri5}\begin{aligned}\xymatrix{M\sqcup Y \ar[rrr] \ar[ddd]^q \ar[rd]
&&&
M' \sqcup Y' \ar[ddd]^{q'}
\ar[dl]
\\
&
\bar M \sqcup \bar Y
\ar[r]\ar[d]^{\bar q}
&
\bar M' \sqcup \bar Y' \ar[d]^{\bar q'}
\\
&
\bar X \ar[r] & \bar X' \\
X \ar[rrr] \ar[ur] &&& X' \ar[ul] \\
}\end{aligned}$$]{} By definition $\bar q$, $\bar q'$ and $\bar q'$ are quotient maps. We want to show that under suitable conditions $q$ is a quotient map. If this is the case, then proposition \[propo:theo:pushout2\] is proved, since $q$ is exactly the projection defining the topology of the push-out.
\[lemma:step4\] Assume that $\bar f(\bar X {\smallsetminus\/}\bar Y) \subset \bar X' {\smallsetminus\/}\bar Y'$. Then $q$ is onto, $q|Y$ and $q|(M{\smallsetminus\/}A)$ are mono.
Let $x=(\bar x, x') \in X$, so that then $\bar f ( \bar x ) = p_{X'} (x')$. If $\bar x \in \bar Y$, then $\bar f (\bar x) \in \bar Y'$, therefore $x' \in p_{X'}^{-1} \bar Y' = Y'$. Thus there is $y = (\bar y, y') \in Y $ such that $q(y) = x$. On the other hand, if $\bar x\not\in \bar Y$, since assumption $\bar f \bar x \not\in \bar Y'$, necessarily $x' \not\in Y'$ hence there is a unique $m'\in M'{\smallsetminus\/}A'$ such that $q'(m') = x'$. For the same reason there is a unique $\bar m\in \bar M{\smallsetminus\/}\bar A$ such that $\bar q(\bar m) = \bar x$. Now consider the since $\bar q'$ is mono in $\bar M' {\smallsetminus\/}A'$, the chain of equalities $$\bar q' \tilde f (\bar m)
=
\bar f \bar q(\bar m) = \bar f (\bar x)
= p_{X'}( x')
=
p_{X'} q' (m')
=
\bar q' p_{M'} (m')$$ implies that $\tilde f (\bar m) = p_{M'}(m')$, hence that $(\bar m, m') \in M$. We have shown that $q$ is onto. The restriction of $q$ to $Y$ is mono since $j_{\bar Y}
\times j_{Y'}$ is mono. To see that the restriction of $q$ to $M{\smallsetminus\/}A$ is mono, consider that if $(\bar m, m')\in M$ then $\bar m \not\in \bar A$ implies $m'\not \in A'$, hence $M {\smallsetminus\/}A \subset (\bar M{\smallsetminus\/}\bar A)\times (M'{\smallsetminus\/}A')$. Now, $\bar q$ and $q'$ are mono when restricted to $\bar M {\smallsetminus\/}\bar A$ and $M'{\smallsetminus\/}A'$ respectively, hence $q|(M{\smallsetminus\/}A)$ is mono.
From now on, we will assume that the condition of lemma \[lemma:step4\] is fulfilled.
\[lemma:step7\] The restriction $q|(M{\smallsetminus\/}A){\colon}M{\smallsetminus\/}A \to X{\smallsetminus\/}Y$ is a homeomorphism.
The map $q|(M{\smallsetminus\/}A)$ is bijective and continuous. We need to show that it is open. Consider an open set $U$ in $M{\smallsetminus\/}A$, and a point $x=(\bar x, x') \in U$. There are neighborhood $U_{\bar M} \subset \bar M {\smallsetminus\/}\bar A$ and $U_{M'} \subset M'{\smallsetminus\/}A'$ such that $$x \in ( U_{\bar M} \times U_{M'} ) \cap M \subset U.$$ Since $\bar q$ and $q'$ are homeomorphisms of $\bar M {\smallsetminus\/}\bar A$ and $M'{\smallsetminus\/}A'$ onto their images, the sets $U_{\bar X} = \bar q U_{\bar M}$ and $U_{X'} = q' U_{M'}$ are open subsets of $\bar X {\smallsetminus\/}\bar Y$ and $X' {\smallsetminus\/}Y'$. Hence $U_{\bar X}\times U_{X'} \cap X$ is open in $X$ and contained in $X{\smallsetminus\/}Y$. Moreover, since $q$ is mono in $M{\smallsetminus\/}A$, we have that $q \left[ ( U_{\bar M} \times U_{M'} ) \cap M \right]
= U_{\bar X} \times U_{X'} \cap X $, hence $U_{\bar X} \times U_{X'} \cap X \subset q(U)$. But $x$ is arbitrary, therefore $q(U)$ is open and so the map $q|M{\smallsetminus\/}A$ is open.
\[lemma:step6\] A subset $S\subset M\sqcup Y$ is saturated (i.e. $q^{-1}q(S) = S$ ) if and only if $S\cap A = h^{-1}(S\cap Y)$.
\[lemma:longnonce\] Assume that $\bar M$ and $M'$ are metrizable. If $U_{\bar M} \subset \bar M$ and $U_{A'} \subset A'$ are open sets with compact closure such that $$(\overline{U_{\bar M}} \times \overline{U_{A'}}) \cap M \subset q^{-1}U,$$ then there is an open subset $U_{M'} \subset M'$ with compact closure such that $$\begin{gathered}
U_{M'} \cap A' = U_{A'}\\
(\overline{U_{\bar M}} \times \overline{U_{M'}}) \cap M \subset q^{-1}U.\end{gathered}$$ In the same way, if $U_{\bar A} \subset \bar A$ and $U_{M'}\subset M'$ are open sets with compact closure such that $$(\overline{U_{\bar A}} \times \overline{U_{M'}}) \cap M \subset q^{-1}U,$$ then there is an open subset $U_{\bar M} \subset \bar M$ with compact closure such that $$\begin{gathered}
U_{\bar M} \cap \bar A = U_{\bar A}\\
(\overline{U_{\bar M}} \times \overline{U_{M'}}) \cap M \subset q^{-1}U.\end{gathered}$$
By assumption the set $q^{-1}U\cap M$ is open in $M\subset \bar M \times M'$, therefore there exists an open set
$\widetilde{q^{-1}U} \subset \bar M \times M'$ such that ${\widetilde{q^{-1}U}}\cap M = q^{-1}U \cap M$. We need to find the open set $U_{M'}\subset M'$ with the desired properties. Consider the following function $\eta{\colon}\bar M \times M' \to {\mathbb{R}}$ defined by $$\label{eq:function}
\eta(x) = d(x,{{\widetilde{q^{-1}U}}}^c) + d(x,M),$$ where ${\widetilde{q^{-1}U}}^c$ is the complement of ${\widetilde{q^{-1}U}}$ in $\bar M \times M'$ and $d$ is the distance given by a metric on $\bar M \times M'$. Consider a point $x$ in the compact closure $\overline{U_{\bar M}} \times \overline{U_{A'}}$. If $x\in M$ as well, then by the assumption $x\in {\widetilde{q^{-1}U}}$, and since ${\widetilde{q^{-1}U}}$ is open this implies $\eta(x)\geq d(x,{{\widetilde{q^{-1}U}}}^c)>0$. On the other hand, since $M$ is closed in $\bar M \times M'$, if $x\not\in M$ then $\eta(x)\geq d(x,M) >0$. Thus $\eta(x)>0$ for every $x\in (\overline{U_{\bar M}} \times \overline{U_{A'}} )$, and so there is an open set $U_{M'} \subset M'$ with compact closure such that $U_{M'} \cap A' = U_{A'}$ and $$x\in \overline{U_{\bar M}} \times \overline{U_{M'}} \implies \eta(x)>0.$$ This means that if $x\in (\overline{U_{\bar M}} \times \overline{U_{M'}}) \cap M$ then $d(x,{\widetilde{q^{-1}U}}^c)>0$, thus $x\in {\widetilde{q^{-1}U}}$, and this implies that $$(\overline{U_{\bar M}} \times \overline{U_{M'}}) \cap M \subset q^{-1}U.$$ For the second part of the lemma, the proof is exactly the same, it is only necessary to exchange the roles of $\bar A$ and $M'$.
\[lemma:a1\] Let $X$ be a ${{\bf\mathfrak{{F}}}}$-bundle over a locally finite and countable CW-complex $\bar X$. Then $X$ is metrizable. Moreover, every open subset $O\subset X$ is the union of an ascending sequence of open subsets $O_n \subset X$ with compact closure $\overline{O_n} \subset O_{n+1}$.
Since $\bar X$ is locally finite and countable, it is metrizable and $2^\circ$ countable. Since the fibres are $2^\circ$ countable by assumption, $X$ is $2^\circ$ countable, since it is obtained by attaching a countable number of ${{\bf\mathfrak{{F}}}}$-cells $V\times D^n$, which are $2^\circ$ countable. Now, $\bar X$ and the ${{\bf\mathfrak{{F}}}}$-cells $V\times D^n$ are completely regular, hence $X$ is regular; therefore by Urysohn metrization theorem $X$ is metrizable. Let $d$ denote its metric. Now consider an ascending chain of finite subcomplexes $\bar X_n \subset \bar X$ such that $\bar X_n$ is contained in the interior of $\bar X_{n+1}$ and $\bar X=\bigcup_n \bar X_n$ (such a sequence exists since $\bar X$ is locally finite and countable). Using the fact that the fibres are locally compact and second countable, it is possible to show that $X$ can be written as the union of an ascending sequence of subspaces $X_n \subset X_{n+1}\subset \dots$ where each $X_n$ is compact and is contained in the interior $\mathring{X}_{n+1}$. For every $n$ define the following open set $$O_n = \{
x\in X {\ \mathrm{:} \ }d(x,O^c)>\frac{1}{n}
\} \bigcap \mathring{X_n},$$ where $O^c$ denotes the complement of $O$. If $x$ is an element of the closure $\overline {O_n}$ (closure of $O_n$ in $X$), then $x \in O_{n+1} \bigcap X_n \subset X_n$, hence $\overline{O_n}$ is a closed subset of the compact space $X_n$, hence $\overline{O_n}$ is compact. Moreover, since $X_n \subset \mathring{X}_{n+1}$, we have that $\overline{O_n} \subset O_{n+1}$ as claimed. We need to show that $O = \bigcup_n O_n$: if $x\in O$, then $d(x,O^c)>0$, therefore there is $n_1\geq 1$ such that $d(x,O^c)>\frac{1}{n}$ for every $n\geq n_1$. Since $\bigcup_n X_n = X$, there is $n_2\geq n_1$ such that $x\in X_{n_2}\subset \mathring{X}_{n_2+1}$. But these conditions imply that $x\in O_{n}$ for $n=n_2+1$. Thus $O=\bigcup O_n$.
\[lemma:nonce\] Assume that $\bar M$ and $M'$ are metrizable. If $U_{\bar A} \subset \bar A$ and $U_{A'} \subset A'$ are open sets such that $$({U_{\bar A}} \times {U_{A'}}) \cap M \subset q^{-1}U,$$ then there are open sets $U_{\bar M} \subset \bar M$ and $U_{M'} \subset M'$ such that $$\begin{gathered}
U_{\bar M} \cap \bar A = U_{\bar A}\\
U_{M'} \cap A' = U_{A'}\\
(U_{\bar M} \times U_{M'}) \cap M \subset q^{-1}U.\end{gathered}$$
By lemma \[lemma:a1\], it is possible to find increasing sequences $U_{\bar A}^k \subset U_{\bar A}$ and $U_{A'}^k \subset U_{A'}$ with compact closures and such that $$\begin{gathered}
\bigcup_{k\geq 1} U_{\bar A}^k = U_{\bar A} \\
\bigcup_{k\geq 1} U_{A'}^k = U_{A'}.\end{gathered}$$ Now, by applying lemma \[lemma:longnonce\] twice it is possible to define compact subsets $U_{\bar M}^1$ and $U_{M'}^1$ such that $$\begin{gathered}
(\overline{U_{\bar M}^1} \times \overline{U_{M'}^1} ) \cap M \subset q^{-1}U.\end{gathered}$$ By induction, we will show that it is possible to find two sequences of open sets $U^k_{\bar M}$ and $U_{M'}^{k}$ with compact closures, with the property that $$\begin{gathered}
U^k_{\bar M} \cap \bar A = U_{\bar A}^k \\
U^{k}_{M'} \cap A' = U_{A'}^{k} \end{gathered}$$ for every $k$, $k' \geq 1$ and $$\label{eq:nonce}
( \overline{ U_{\bar M}^k} \times \overline{ U_{M'}^{k'}} ) \cap M \subset q^{-1}U$$ for every $k$, $k' \geq 1$. Assume that the sequences are defined for $j\leq k-1$. By applying $k-1$ times lemma \[lemma:longnonce\] and taking the intersection of the resulting open sets, we can show that there is $U_{\bar M}^k$ such that $U^k_{\bar M} \cap \bar A = U_{\bar A}^k$ and $$( \overline{ U_{\bar M}^k} \times \overline{ U_{M'}^{k'}} ) \cap M \subset q^{-1}U$$ for every $k'\leq k-1$. Now we can apply $k$ times lemma \[lemma:longnonce\] and take the intersection of the resulting open sets, to finally find and open set with compact closure $U_{M'}^k$ such that $U^k_{M'} \cap A' = U_{A'}^k$ and $$( \overline{ U_{\bar M}^j} \times \overline{ U_{M'}^{k}} ) \cap M \subset q^{-1}U$$ for every $j\leq k$. Thus equation \[eq:nonce\] holds for every $k$, $k'\geq 1$. But this implies that the open sets $$\begin{gathered}
U_{\bar M} = \bigcup_{k\geq 1} U^k_{\bar M} \\
U_{M'} = \bigcup_{k\geq 1} U^k_{M'}\end{gathered}$$ have the desired property.
\[lemma:step8\] If $\bar M$ and $\bar M' $ are locally finite and countable, then the map $q$ is a quotient map.
Consider a subset $U\subset X$ such that $q^{-1}U$ is open in $M\sqcup Y$. We want to show that $U$ is open. Let $x\in U {\smallsetminus\/}Y$. Since $q^{-1}U\cap (M{\smallsetminus\/}A)$ is open, by lemma \[lemma:step7\] there is an open neighborhood of $x$ in $X{\smallsetminus\/}Y$ contained in $U$. If $U\cap Y=\emptyset$ then we have proved that $U$ is open. Otherwise, let $x\in Y \cap U$. Since $q^{-1}U$ is open, $q^{-1}U \cap Y$ is open and contains $y\in q^{-1}(x)\cap Y$. Therefore there are two open sets $U_{\bar Y}$ and $U_{Y'}$ such that $$y \in (U_{\bar Y} \times U_{Y'} )\cap Y \subset q^{-1}Y \cap Y.$$ Let $$\begin{gathered}
U_{\bar A} = \bar h^{-1} U_{\bar Y}\\
U_{A'} = {h'}^{-1} U_{Y'}\end{gathered}$$ Since $\bar h$ and $h'$ are continuous, they are open subsets of $\bar A$ and $A'$ respectively. By lemma \[lemma:nonce\] there are open sets $U_{\bar M} \subset \bar M$ and $U_{M'} \subset M'$ such that $$\begin{gathered}
U_{\bar M} \cap \bar A = U_{\bar A}= \bar h^{-1} U_{\bar Y}\\
U_{M'} \cap A' = U_{A'}= {h'}^{-1} U_{Y'},\end{gathered}$$ and $$(U_{\bar M} \times U_{M'})\cap M \subset q^{-1}U.$$ Consider the sets $$\begin{gathered}
U_{\bar X} = \bar q ( U_{\bar M} ) \cup U_{\bar Y} \subset \bar X \\
U_{X'} = q' ( U_{M'} ) \cup U_{Y'} \subset X'.\end{gathered}$$ Since $\bar q$ and $q'$ are quotient maps (by definition), and since $U_{\bar M} \cup U_{\bar Y}$ and $U_{M'} \cup U_{Y'}$ are saturated with respect to the maps $\bar q$ and $q'$ (see lemma \[lemma:step6\]), $U_{\bar X}$ and $U_{X'}$ are open in $\bar X$ and $X'$. Therefore $U_x = U_{\bar X} \times U_{X'} \cap X$ is open in $X$ and contains $x$. It is left to show that $U_x \subset U$. Consider $t\in U_x\cap Y$. Then $q^{-1}t = h^{-1}t \sqcup \{t\} \subset M\sqcup Y$. But since $j_Y$ is mono, $t\in (U_{\bar Y} \times U{_Y'})\cap Y$ and therefore $t\in U$. Otherwise, if $t\in X{\smallsetminus\/}Y$ then $q^{-1}t = m$, unique point belonging to $(U_{\bar M} \times U_{M'}) \cap M \subset q^{-1}U$. Thus $q^{-1}t \subset q^{-1}U$ and therefore $t\in U$. We have proved that $U_x \subset U$ and the proof is now complete.
Since a pull-back of a ${{\bf\mathfrak{{F}}}}$-bundle is a ${{\bf\mathfrak{{F}}}}$-bundle, if $M'$ is a ${{\bf\mathfrak{{F}}}}$-bundle, then $M$ is a ${{\bf\mathfrak{{F}}}}$-bundle. Furthermore, by \[lemma:step2\], $A$ is the restriction of $M$ to $\bar A$. Thus, in this case, proposition \[propo:theo:pushout2\] implies that if $\bar M$ and $\bar M'$ are locally finite countable CW-complexes and $M'$ a ${{\bf\mathfrak{{F}}}}$-bundle, if $X'$ is obtained by attaching the ${{\bf\mathfrak{{F}}}}$-bundle $M'$ to a ${{\bf\mathfrak{{F}}}}$-family $Y'$, then the pull-back ${{\bf\mathfrak{{F}}}}$-family $\bar f^* X'$ is obtained by attaching the pull-back ${{\bf\mathfrak{{F}}}}$-bundle $M$ to the pull-back ${{\bf\mathfrak{{F}}}}$-family $\tilde f^* Y'$.
Push-out of pull-back ${{\bf\mathfrak{{F}}}}$-bundles {#section:pushout}
=====================================================
In this section we complete the proof of the bundle theorem \[theo:bundle\]. In case of the pull-back of ${{\bf\mathfrak{{F}}}}$-bundles, the following properties hold.
\[lemma:relativeMA\] Let $\bar Y$ be a space, $\bar h{\colon}S^{n-1} \to \bar Y$ a map and $\bar X= \bar Y\cup_{\bar h} D^n$. Let $\bar {\Phi}{\colon}D^n \to \bar X$ denote the characteristic map. If $X \to \bar X$ is a ${{\bf\mathfrak{{F}}}}$-bundle and $Y$ is the restriction of $X$ to $\bar Y$, then the following diagram is a push-out [$$\label{diag:here}\begin{aligned}\xymatrix{\bar h^*Y \ar[d] \ar[r] &
Y \ar[d]
\\
\bar {\Phi}^* X \ar[r] & X \\
}\end{aligned}$$]{}
Since the diagram commutes, one gets the map $t{\colon}P \to X$, where $P$ denotes the push-out space $\bar {\Phi}^*X \cup Y$. It is easy to see that $t$ is bijective and covers the identity of $\bar X$. Now we show that $t$ is a local homeomorphism when restricted to the space over $\bar U$, where $\bar U \subset \bar X$ is an open set such that $X|\bar U$ is trivial. In fact, in this case the diagram \[diag:here\] is reduced as [$$\begin{aligned}\xymatrix{\bar h^* (W\times(\bar U \cap \bar Y)) \ar[d] \ar[r] &
W \times (\bar U \cap \bar Y) \ar[d]
\\
W \times \bar {\Phi}^{-1} \bar U \ar[r] & W \times \bar U, \\
}\end{aligned}$$]{}which is a push-out since $W$ is locally compact Hausdorff and $\bar X = \bar Y \cup_{\bar h} D^n$. Thus $P|\bar U$ and $X|\bar U$ are homeomorphic. But $P|\bar U$ and $X|\bar U$, when $\bar U$ ranges over all the trivializing neighborhoods of $\bar X$, are open covers of $P$ and $X$ respectively, with $t(P|\bar U) = X|\bar U$ for each $\bar U$. Hence $t$ is an open map.
Since exactly the same argument can be applied to the attaching of a set $\bar Z$ of $n$-cells, we have the following generalization of lemma \[lemma:relativeMA\].
\[lemma:attaching\] Let $\bar Y$ be a space, $\bar Z$ a set and $\bar h{\colon}\bar Z \times S^{n-1} \to \bar Y$ an attaching map with $\bar X= \bar Y\cup_{\bar h} (\bar Z \times D^n)$. Let $\bar {\Phi}{\colon}\bar Z \times D^n \to \bar X$ denote the characteristic map. If $X \to \bar X$ is a ${{\bf\mathfrak{{F}}}}$-bundle and $Y$ is the restriction of $X$ to $\bar Y$ then the following diagram is a push-out [$$\begin{aligned}\xymatrix{\bar h^*Y \ar[d] \ar[r] &
Y \ar[d]
\\
\bar {\Phi}^* X \ar[r] & X. \\
}\end{aligned}$$]{}
We recall now the following important property. Let $M_0 \subset M_1 \subset \dots \subset M_n \subset \lim_{n\geq 0} M_n$ be a sequence of spaces, and $\displaystyle M = \lim_{n\to \infty} M_n$. Then for every locally compact Hausdorff space $W$ $$\label{eq:limitMn}
\lim_{n\to \infty} \left( W \times M_n \right) =
W \times \left( \lim_{n\to \infty} M_n \right) = W \times M.$$
\[lemma:cwpair\] Let $(\bar M, \bar A)$ be a CW-pair; let $M \to \bar M$ be a ${{\bf\mathfrak{{F}}}}$-bundle and $A$ the restriction of $M$ to $\bar A$. Then $(M,A)$ is a ${{\bf\mathfrak{{F}}}}$-CW-pair.
Consider the filtration on skeleta $$\bar A \subset \bar M_1 \subset \bar M_2 \subset \dots
\subset \bar M_n \subset \dots \subset \bar M.$$ It induces a filtration on restrictions of the bundles $M\to \bar M$ $$A \subset M_1 \subset M_2 \subset \dots
\subset M_n \subset \dots \subset M.$$ By lemma \[lemma:attaching\] for every $n\geq 0$ the diagram [$$\begin{aligned}\xymatrix{\bar h_n^*M_{n-1} \ar[d] \ar[r] &
M_{n-1} \ar[d]
\\
\bar {\Phi}_n^* M_n \ar[r] & M_n, \\
}\end{aligned}$$]{}is a push-out, where $\bar h_n{\colon}\bar Z_n \times S^{n-1} \to \bar M_{n-1}$ is the attaching map on $\bar M_{n-1}$ and $\bar {\Phi}_n$ is the corresponding characteristic map. Since $D^n$ is contractible, there exists a ${{\bf\mathfrak{{F}}}}$-isomorphism of (pair of ) bundles $$\label{eq:isomorph}
(\bar {\Phi}_n^* M_n, \bar h_n^* M_{n-1}) \cong
W \times ( \bar Z_n \times D^n, \bar Z_n \times S^{n-1})$$ for a suitable object $W\in {{\bf\mathfrak{{F}}}}$. This means that $M_n$ is obtained by attaching $W \times \bar Z_n \times D^n$ to $M_{n-1}$ with as attaching map the composition of the map induced by pull-back $\bar h^*$ and the isomorphism of \[eq:isomorph\]. Thus $(M_n,A)$ there are skeleta $M_n \subset M$ such that $$A \subset M_0 \subset M_1 \subset \dots M.$$ To show that $(M,A)$ is a relative ${{\bf\mathfrak{{F}}}}$-complex (and hence a ${{\bf\mathfrak{{F}}}}$-CW-pair) we need to show that $\displaystyle \lim_{n\to \infty} M_n = M$. Since there is a continuous bijection $t{\colon}\displaystyle \lim_{n\to \infty} M_n \to M$ covering the identity of $\bar M$, we need to show that $M$ has the topology of the limit. It suffices to restrict $M$ to the trivializing neighborhoods $\bar U\subset \bar M$, and this is true as a consequence of \[eq:limitMn\].
Lemma \[lemma:cwpair\] implies the following corollary, by taking $A=\emptyset$. Let $\bar M$ be a CW-complex; let $M \to \bar M$ be a ${{\bf\mathfrak{{F}}}}$-bundle. Then $M$ is a ${{\bf\mathfrak{{F}}}}$-complex. This is the second part of \[theo:bundle\].
Push-out of function spaces {#section:9}
===========================
We now start the actual proof of the principal bundle theorem, which is achieved in the next two sections.
Let $V$ be an object of ${{\bf\mathfrak{{F}}}}$. We recall that a family $\mathcal{K}$ of compact sets of $V$ is termed *generating* if for every ${{\bf\mathfrak{{F}}}}$-family $Y$ the subsets $N_{K,U}$ with $K\in \mathcal{K}$ and $U$ open in $Y$ yield a sub-basis for the topology of $Y^V$.
\[defi:nnep\] Let $Y\subset X$ a pair of ${{\bf\mathfrak{{F}}}}$-families and a closed inclusion. We say that $Y$ has the *$N$-neighborhood extension property in $X$* (NNEP) if for every $V\in {{\bf\mathfrak{{F}}}}$ there exists a generating family $\mathcal{K}$ of compact subsets of $V$ such that the following is true: let $U$ be open in $X^V$ and for $i=1,\dots, l$ let $K_i \subset V$ be compact sets in $\mathcal{K}$ and $U_i\subset Y$ be open such that $$\bigcap_{i=1}^l N_{K_i,U_i} \subset U \cap Y^V;$$ then there are $l$ open subsets $U'_i\subset X$ such that $U'_i \cap Y = U_i$ and $$\bigcap_{i=1}^l N_{K_i,U_i'} \subset U.$$
An important easy property of the sets $N_{K,U}$ is the following: if $U\subset Y$ and $h{\colon}A \to Y$ is a map, then $$N_{K,h^{-1}U} = {(h^V)}^{-1} N_{K,U}.$$
\[lemma:coprod\] Let $Z$ be a set and $(M_z,A_z)$ be an NNEP-pair for every $z\in Z$. Then the coproduct of inclusions $$\coprod_{z\in Z} A_z \subset \coprod_{z\in Z} M_z$$ yields a NNEP-pair.
\[lemma:nnep\] Assume that $A$ has the NNEP in $M$ and that $h{\colon}A \to Y$ is a ${{\bf\mathfrak{{F}}}}$-map to a ${{\bf\mathfrak{{F}}}}$-family $Y$. Then $Y$ has the NNEP in the push-out space $X=M\cup_h Y$.
By lemma \[lemma:pushout\] $Y\to X$ is a closed inclusion. Let ${\Phi}{\colon}M \to X$ denote the characteristic map of the push-out. Let $U\subset X^V$ be open, and let $K_i\subset V$ and $U_i \subset Y$ be compact subsets of a generating family $\mathcal{K}$ of $V$ and open subsets such that $$\bigcap_{i=1}^l N_{K_i,U_i} \subset U \cap Y^V$$ as in definition \[defi:nnep\]. The map ${\Phi}^V$ is continuous, hence $U''={({\Phi}^V)}^{-1} U \subset M^V$ is open in $M^V$. Moreover, the intersection $${(h^V)}^{-1}
\bigcap_{i=1}^l
N_{K_i,U_i} =
\bigcap_{i=1}^l
N_{K_i,h^{-1}U_i}$$ is contained in $U''$. Since $A$ has the NNEP in $X$, there are $l$ open sets $U_i''$ in $M$ such that $U_i'' \cap A = h^{-1}U_i$ and $$\bigcap_{i=1}^l N_{K_i,U''_i} \subset U''.$$ Now consider the subsets $U_i'=U_i \bigcup {\Phi}(U_i'') \subset X$. Since $U_i' \cap Y = U_i$, ${\Phi}^{-1}U'_i = U_i''$ and $X$ is the push-out of $M$ and $Y$, each $U_i'$ is open in $X$. Moreover, consider $$f\in \bigcap_{i=1}^l N_{K_i,U'_i}.$$ If $f\in Y^V$ then $$f\in \bigcap_{i=1}^l N_{K_i,U'_i} \cap Y^V =
\bigcap_{i=1}^l N_{K_i,U'_i\cap Y} =
\bigcap_{i=1}^l N_{K_i,U_i} \subset U\cap Y^V,$$ hence $f\in U$. On the other hand, if $f\not\in Y^V$, then there is a unique $f''{\colon}K \to M{\smallsetminus\/}A$ such that $f = {{\Phi}^V}(f'')$. Since for every $i$ $fK_i = hf''K_i \subset U'_i$, we have $$f''K_i \subset {\Phi}^{-1} U_i' = U''_i,$$ thus $$f'' \in \bigcap_{i=1}^l
N_{K_i,U''_i} \subset U''.$$ Hence $f= {\Phi}^V(f'') \in {\Phi}^V U'' = {\Phi}^V {({\Phi}^V)}^{-1} U = U$. Therefore $$\bigcap_{i=1}^l N_{K_i,U'_i} \subset U,$$ and the proof is complete.
\[lemma:nnep2\] Assume that $(M,A)$ is a NNEP-pair, $Y$ is a ${{\bf\mathfrak{{F}}}}$-family and $h{\colon}A \to Y$ is a ${{\bf\mathfrak{{F}}}}$-map. Let $X$ be the push-out $X=M\cup_h Y$. Then the following diagram [$$\begin{aligned}\xymatrix{A^V \ar[r]^{h^V} \ar@{ >->}[d]
&
Y^V \ar@{ >->}[d] \\
M^V \ar[r]^{{\Phi}^V}
&
X^V
}\end{aligned}$$]{}is a push-out.
By lemma \[lemma:nnep\] $(X,Y)$ is a NNEP-pair. Let $\mathcal{K}$ denote the generating family of compact sets of $V$. Let $P$ denote the push-out $M^V\cup_{h^V} Y^V$. By the push-out property there is a continuous map $P \to X^V$. It is easy to see that it is bijective, so that, by identifying $P$ and $X^V$, the only thing to prove is that if $U$ is an open in the push-out topology, then it is open in the compact-open topology. This is true if and only if for every $f_0\in U$ there is a subset $U'\subset X^V$ open in the compact-open topology such that $f_0\subset U' \subset U$.
Consider first the case $f_0\not\in Y^V$. Then there exists a unique $f_0'\in M^V {\smallsetminus\/}A^V$ such that ${\Phi}^V(f_0') = f_0$. Since $A^V$ is closed in $M^V$, $f_0'$ is contained in an open neighborhood $U_0'' \subset M^V$ such that $U_0'' \cap A^V = \emptyset$ and $U_0''$ is of type $$\bigcap_{i=1}^l N_{K_i,U_i''}$$ for some $K_i\subset V$ and some $U_i''$ in $M{\smallsetminus\/}A$. For each $i$ the image $U'_i={\Phi}U_i''$ is open in $X$, hence the set $$\bigcap_{i=1}^l N_{K_i,U_i'}$$ is open in $X^V$, contains $f_0$ and is contained in $U$.
Now consider $f_0\in Y^V$. The intersection $Y^V\cap U$ is open in $Y^V$ and hence there is a neighborhood of $f_0$ in $Y^V$ contained in $Y^V\cap U$. Because $Y^V$ has the compact-open topology this means that there are $l$ compact subsets $K_i\subset V$ and open sets $U_i\subset Y$ such that $$f_0 \in
\bigcap_{i=1}^l N_{K_i,U_i} \subset U,$$ where as above $N_{K_i,U_i}$ denotes the set of all the maps in $Y^V$ such that $fK_i\subset U_i$. Without loss of generality we can assume that $K_i$ belongs to the generating family $\mathcal{K}$ for every $i=1\dots l$.
Consider the sets $$U''=
{({\Phi}^V)}^{-1} U \subset M^V$$ and $$\bigcap_{i=1}^l N_{K_i,h^{-1}U_i} \subset A^V \cap U''.$$ Since $(M,A)$ is a NNEP-pair, by definition \[defi:nnep\] there are open sets $U_i''\subset M$ such that $U_i'' \cap A = h^{-1}U_i$ and $$\bigcap_{i=1}^l N_{K_i,U_i''} \subset U''.$$
Let $U_i'=U_i \bigcup {\Phi}U_i''$ for every $i=1$, $\dots$, $l$. As in the proof of lemma \[lemma:nnep\], they are $l$ open subsets of $X$ with the property that $U_i'\cap Y=U_i$, hence $f_0\in N_{K_i,U'_i}$ for every $i$. And, again as in the proof of lemma \[lemma:nnep\], $$U' = \bigcap_{i=1}^l N_{K_i,U_i'} \subset U.$$ This concludes the proof.
\[lemma:limit\] Consider a sequence of ${{\bf\mathfrak{{F}}}}$-families $$X_0 \subset X_1 \subset \dots \subset X_n \subset \dots \subset
X=\lim_{n\to \infty} X_n$$ such that for every $n\geq 1$ $(X_n,X_{n-1})$ is a NNEP-pair. Then $(X,X_0)$ is a NNEP-pair and $$X^V= \lim_{n\to \infty} X_n^V$$
Let $U\subset X^V$ be open, and $\bigcap_{i=1}^l N_{K_i,U_i^0} \subset U\cap X_0^V$ a subset of $X_0^V$, with $K_i \subset V$ compact containing bases and $U_i^0$ open in $X_0$. By induction, it is possible to define sequences of sets $U_i^n$ open in $X_n$ such that for every $\geq 1$ $$U_i^n \cap X_{n-1} = U_i^{n-1}$$ $$\bigcap_{i=1}^l N_{K_i,U_i^n} \subset X_n^V \cap U.$$ Now, because $X=
\lim_{n\to \infty}
X_n$, the sets $U_i^\infty = \bigcup_{n=0}^{\infty} U_i^n$ are open in $X$, hence $$\bigcap_{i=1}^l N_{K_i,U_i^\infty} \subset U.$$ is an open set. Moreover, by construction $U_i^\infty \cap X_0 = U_i^0$, therefore $X_0$ has the NNEP in $X$. The same construction shows that $X^V = \lim_{n\to \infty} X_n^V$.
NKC-categories {#section:10}
==============
The crucial property of a NKC-category as defined in \[defi:NKC\] is the next result.
\[propo:theo:tech\] Assume that ${{\bf\mathfrak{{F}}}}$ is NKC. Then for every $W\in {{\bf\mathfrak{{F}}}}$ the pair $(W\times D^n, W \times S^{n-1})$ is a NNEP-pair.
The proof of proposition \[propo:theo:tech\] will take the rest of the section.
\[lemma:product\] Let $R$ be a space and $V$ and $W$ objects of ${{\bf\mathfrak{{F}}}}$. Then $$(W\times R)^V = W^V \times R.$$
Let ${\mathrm{pr}}_1$ and ${\mathrm{pr}}_2$ denote the projections onto the first and the second factor of $W\times R$. The map ${\mathrm{pr}}_2\epsilon_0{\colon}(W\times R)^V \to R$ sending $f{\colon}V \to W\times R$ to ${\mathrm{pr}}_2 f(0) \in R$ is continuous. Furthermore, the map ${\mathrm{pr}}_1^V{\colon}(W\times R)^V \to W^V$ is continuous. Therefore the map $F=({\mathrm{pr}}_1^V, {\mathrm{pr}}_2\epsilon_0){\colon}(W\times R)^V \to W^V \times R$ is continuous. Consider now the function $G{\colon}W^V \times R \to (W\times R)^V$ defined by $G(f,r)(v) = (f(v),r)$ for every $v\in V$ and every $r\in R$. It is the adjoint of the map $V\times W^V \times R \to W \times R$ defined by $(v,f,r) \mapsto (f(v),r)$, which is continuous since the evaluation map is continuous. Therefore $G$ is continuous. It is readily seen that $GF=1$ and $FG=1$.
Let $D^n_\epsilon = \{ x\in D^n {\ \mathrm{:} \ }|x|>1-\epsilon \} $, for any $\epsilon$ small, and let $\rho_{\epsilon}{\colon}D^n_\epsilon \to S^{n-1}$ be the retraction defined by $x \mapsto \frac{x}{|x|}$. With an abuse of notation we use the same symbol for the induced retraction $\rho_\epsilon=1_{W} \times
\rho_\epsilon{\colon}W\times D_\epsilon^n \to W\times S^{n-1}$. For every $W\in {{\bf\mathfrak{{F}}}}$ there is an induced retraction $$\rho_\epsilon^V{\colon}(W\times D^n_\epsilon) ^V
\to
(W\times S^{n-1})^V.$$
\[lemma:epsilonk\] If $C$ is a compact in $(W\times S^{n-1})^V$, and $C$ is contained in an open set $A$ of $(W\times D^n)^V$, then there exists $\epsilon>0$ such that ${(\rho^{V}_\epsilon)}^{-1}C \subset A$.
If $\epsilon$ is small, we have $D^n_\epsilon = S^{n-1}\times
[0,\epsilon)$, and hence $ (W\times D^n_\epsilon) ^V = (W\times S^{n-1})^V \times [0,\epsilon)$ and $\rho^V_\epsilon$ is the projection onto the first factor. Each $x\in C$ is a point of $A$, therefore there exists an open neighborhood of $x$ of type $O_x\times [0,\epsilon_x)$ contained in $A$. Being $C$ compact there is a finite number of points $x$ such that $C$ is covered by $O_x\times [0,\epsilon_x)$. Therefore there exists a $\epsilon>0$ such that $\cup_{x} O_x \times [0,\epsilon)$ is an open subset of $A$ containing $C$ and thus ${(\rho^V_\epsilon)}^{-1} C \subset A$.
Now we can start the proof of proposition \[propo:theo:tech\]. Let $U$ be an open subset $U\subset (W\times D^n)^V$. Let $\mathcal{K}$ be the generating family of compact subsets of $V$ of the NKC property. Let $K_i\subset V$ be compact subsets in $\mathcal{K}$ and $U_i\subset W\times S^{n-1}$ open subsets, such that $$\label{eq:defiofN}
\bigcap_{i=1}^l N_{K_i,U_i} \subset U \cap (W\times S^{n-1})^V.$$ By lemma \[lemma:a1\] there is a sequence of open sets $A^k_i \subset W\times S^{n-1}$ with $k\geq 1$ such that, for every $k$, $A_i^k \subset A^{k+1}_i$, $$\label{eq:this}
\bigcup_{k\geq 1} A_i^k = U_i$$ and the closure $\bar A_i^k$ is a compact subset of $U_i$.
\[lemma:Ncompact\] For every $i=1,\dots, l$ the space $N_{K_i,\bar A_i^k}$ is compact.
It is closed, because it is equal to the intersection of all the spaces $N_{\{x\},\bar A_i^k}$, with $x\in K_i$, and they are closed because pre-images of the closed set $\bar A_i^k$ under the evaluation map $(W\times S^{n-1})^V\times \{x\} \to W\times S^{n-1}$, which is continuous because of lemma \[coro:adjoint\]. Now we prove that it is compact. Consider the projection $p{\colon}W\times S^{n-1} \to W$ onto the first factor. Since $\bar A_i^k$ is compact, its image $p\bar A_i^k$ is compact in $W$. Moreover, it is easy to see that $$p^V(N_{K_i,\bar A_i^k}) \subset
N_{K_i,p\bar A_i^k},$$ and hence that $$N_{K_i,\bar A_i^k} \subset
{(p^V)}^{-1}N_{K_i,p\bar A_i^k} = N_{K_i,p\bar A_i^k} \times S^{n-1}.$$ Now, since $K_i$ and $p\bar A_i^k$ are compact and ${{\bf\mathfrak{{F}}}}$ is an NKC-category, $N_{K_i,p\bar A_i^k}$ is compact. Therefore $N_{K_i,\bar A_i^k}$ is a closed subset of a compact space, and hence compact.
Consider for every $k\geq 1$ the set $$C_k = \bigcap_{i=1}^l N_{K_i,A^k_i} \subset (W\times S^{n-1})^V.$$ It is open, by definition of compact-open topology. Let $C_{\infty}$ denote the union $\bigcup_{k\geq 1} C_k $.
\[lemma:a3\] We have that $$C_\infty =
\bigcap_{i=1}^l N_{K_i,U_i} \subset (W\times S^{n-1})^V \cap U.$$
A ${{\bf\mathfrak{{F}}}}$-map $f\in (W \times S^{n-1})^V$ belongs to $C_{\infty}$ if and only if there is $k\geq 1$ such that $f\in C_k$, that is if and only if there is $k\geq 1$ such that $fK_i \subset A^k_i$ for every $i=1,\dots, l$. Of course this implies that $fK_i \subset U_i$ for every $i$. On the other hand, if $fK_i \subset U_i$, then by equation \[eq:this\] there is $k_i\geq 1$ such that $fK_i \subset A^k_i$ for every $k\geq k_i$. This implies that there is $k$ (take the maximum of all $k_i$) such that $fK_i \subset A^k_i$ for every $i=1,\dots, l$.
\[lemma:closure\] The closure $\bar C_k \subset (W\times S^{n-1})^{V}$ is compact and contained in $C_{\infty}$.
Because of its definition $\bar C_k \subset \cap_{i=1}^l \overline{N_{K_i,A_i^k}}$. Because for every $i=1,\dots, l $ the closure $\overline{N_{K_i,A_i^k}}$ is contained in $N_{K_i,\bar A^k_i}$, which is compact by lemma \[lemma:Ncompact\], we have that $\bar C_k$ is compact. Furthermore, by assumption $\bar A_i^k$ is contained in $U_i$, and hence $N_{K_i,\bar A_i^k}$ is contained in $N_{K_i, U_i}$. Therefore the closure of $C_k$ is contained in $C_{\infty}$.
Because of lemma \[lemma:closure\] and lemma \[lemma:a3\] the compact set $\bar C_k$ is contained in $U$ in $(W\times D^{n})^V$. By lemma \[lemma:epsilonk\] it is possible to find an $\epsilon_k>0$ such that $$\label{eq:contained2}
{(\rho_{\epsilon_k}^V)}^{-1} C_k \subset
U.$$ Without loss of generality we can assume that the sequence $\epsilon_k$, with $k\geq 1$, is decreasing.
Now consider for $i=1,\dots, l$ the following sets in $(W\times D^n)^V$: $$\label{eq:defi}
U_i' =
\bigcup_{k\geq 1}
\rho_{\epsilon_k}^{-1} A^k_i.$$ Since it is the union of open sets, for every $i=1,\dots, l$, the set $U'_i$ is open in $W\times D^n$. Hence the set $$U' = \bigcap_{i=1}^l N_{K_i,U_i'} \subset (W\times D^n)^V,$$ is open.
\[lemmaeq:4\] For every $i=1,\dots, l$, $U_i' \cap (W\times S^{n-1}) = U_i
$.
Because of \[eq:defi\], it suffices to prove that for every $i$ and every $k\geq 1$ $$\rho_{\epsilon_k}^{-1} A^k_i \cap (W\times S^{n-1}) \subset U_i,$$ and this is true because $\rho_{\epsilon_k}^{-1} A^k_i \cap (W\times S^{n-1})
= A^k_i$.
\[lemma:contained\] $U'\subset U$.
If $f\in U'\cap (W\times S^{n-1})^V$, then $fK_i \subset (W\times S^{n-1}) \cap U'_i$, and by lemma \[lemmaeq:4\] this implies $fK_i \subset U_i$. Thus, by \[eq:defiofN\], $f_0 \in U$. On the other hand, assume that $f\in (W\times e^n)^V$. For every $i=1,\dots, l$ the image of the compact $f(K_i)$ is in $U_i'$ and hence, because of \[eq:defi\], for every $i$ there exists an integer $k_i\geq 1$ (depending on $i$) such $f'(K_i) \subset \rho_{\epsilon_{k_i}}^{-1} A_i^{k_i}$. In particular, for every $i=1,\dots, l$, $$f'(K_i) \subset W\times D^n_{\epsilon_{k_i}},$$ and thus, because $f$ is a ${{\bf\mathfrak{{F}}}}$-map and $\epsilon_k$ is a decreasing sequence, $$f'(V) \subset W\times D^n_{\epsilon_{m}},$$ where $m$ denotes the maximum of $k_1,\dots, k_l$. Moreover, the sequence of spaces $A_i^k$ is increasing with $k\geq 1$, and therefore $A_i^{k_i} \subset A_i^m$ for every $i=1,\dots, l$. Thus for every $i=1,\dots, l$, $$f'(K_i) \subset \rho_{\epsilon_{m}}^{-1} A_i^{m}.$$ But this means that for every $i$ $$f \in N_{K_i,\rho_{\epsilon_m}^{-1}A_i^m} =
{(\rho_{\epsilon_m}^V )}^{-1} N_{K_i,A_i^m},$$ hence that $$f \in (\rho_{\epsilon_m}^V )^{-1} C_m.$$ Because of equation \[eq:contained2\] this implies that $f\in U$. As claimed, we have proved that $U' \subset U$.
This is the end of the proof of proposition \[propo:theo:tech\]. We can now draw the consequences which are needed.
\[lemma:step\] Let ${{\bf\mathfrak{{F}}}}$ be a NKC-category. Let $Y$ be a ${{\bf\mathfrak{{F}}}}$-family, $Z$ a ${{\bf\mathfrak{{F}}}}$-set and $h{\colon}Z\times S^{n-1} \to Y$ a ${{\bf\mathfrak{{F}}}}$-map, with $n\geq 1$. If $X$ is a family obtained as a push-out [$$\begin{aligned}\xymatrix{Z \times S^{n-1} \ar[r]^{\hspace{24pt} h} \ar@{ >->}[d]
& Y \ar@{>->}[d] \\
Z\times D^n \ar[r]^{\hspace{12pt} {\Phi}}
& X \\
}\end{aligned}$$]{}then the following diagram [$$\begin{aligned}\xymatrix{Z^V\times S^{n-1} \ar@{=}[r] \ar@{ >->}[d]&
(Z\times S^{n-1})^V \ar[r]^{\hspace{24pt} h^V} \ar@{ >->}[d]
& Y^V \ar[d] \\
Z^V \times D^n \ar@{=}[r] &
(Z\times D^n)^V \ar[r]^{\hspace{24pt} {\Phi}^V} & X^V\\
}\end{aligned}$$]{}is a push-out, and $Y$ has the $N$-neighborhood extension property in $X$.
By lemma \[lemma:coprod\] and proposition \[propo:theo:tech\] if $Z$ is a ${{\bf\mathfrak{{F}}}}$-set the pair $(Z\times D^n, Z \times S^{n-1})$ is a NNEP-pair. Hence, by lemma \[lemma:nnep\] and lemma \[lemma:nnep2\] $X^V$ is the push-out of $(Z\times D^n)^V$ and $Y^V$ and $Y$ has the $N$-neighborhood extension property in $X$.
\[coro:main\] Let $(X,D)$ be a relative ${{\bf\mathfrak{{F}}}}$-complex where ${{\bf\mathfrak{{F}}}}$ is a NKC-category. Let ${\Phi}: Z_n\times D^n \to X$ be its characteristic map of $n$-cells and let $h_n$ denote the $n$-attaching map of $X$, i.e. the restriction of ${\Phi}$ to $Z_n\times S^{n-1}$. For every $n\geq 1$ and every $V\in {{\bf\mathfrak{{F}}}}$ the diagram [$$\begin{aligned}\xymatrix{Z_n^V\times S^{n-1} \ar@{=}[r] \ar@{ >->}[d]&
(Z_n\times S^{n-1})^V \ar[r]^{\hspace{24pt} h_n^V} \ar@{ >->}[d]
& X_{n-1}^V \ar[d] \\
Z_n^V \times D^n \ar@{=}[r] &
(Z_n\times D^n)^V \ar[r]^{\hspace{24pt} {\Phi}^V} & X_n^V\\
}\end{aligned}$$]{}is a push-out (in ${\mathbf{Top}}$). Moreover, $$X^V =
\lim_{n\to \infty} X_n^V.$$
The first part is a direct consequence of lemma \[lemma:step\]. Then, by applying \[lemma:limit\] we obtain that $(X,X_n)$ is a NNEP-pair for every $n$ and that $X^V= \lim_{n\to \infty} X^V_n$.
\[coro:relat\] Let $(X,D)$ be a relative ${{\bf\mathfrak{{F}}}}$-complex (or a ${{\bf\mathfrak{{F}}}}$-CW-pair) with ${{\bf\mathfrak{{F}}}}$ NKC. Then $(X,D)$ is a NNEP-pair.
\[coro:nkc\] Let ${{\bf\mathfrak{{F}}}}$ be NKC. Let $(M,A)$ be a ${{\bf\mathfrak{{F}}}}$-CW-pair, $Y$ a ${{\bf\mathfrak{{F}}}}$-family and $h{\colon}A \to Y$ a ${{\bf\mathfrak{{F}}}}$-map. If $X$ is the push-out $X=M\cup_h Y$ then $(X,Y)$ is a NNEP-pair.
By corollary \[coro:relat\] $(M,A)$ is a NNEP-pair, hence by lemma \[lemma:nnep\] and lemma \[lemma:nnep2\] $(X,Y)$ is a NNEP-pair.
Examples of NKC categories {#section:NKC}
==========================
\[propo:theo:allcompact\] If the fibre functor ${F}$ is faithful and the Hom-sets $\hom_{{\bf\mathfrak{{F}}}}(V,W)$ are compact spaces then ${{\bf\mathfrak{{F}}}}$ is a NKC category.
The proof of the previous proposition is simple. The interesting fact is that there are NKC structure categories with non-compact hom-sets, as a consequence of the following proposition.
\[theo:NKC\] If ${{\bf\mathfrak{{F}}}}$ is a closed subcategory of the category ${{\mathbf{Vect}}}$ of finitely dimensional ${\mathbb{R}}$-vector spaces and linear maps, then ${{\bf\mathfrak{{F}}}}$ has the NKC property.
See [@sm], lemma (4.2).
This result yields the principal stratified bundle associated to a ${{\bf\mathfrak{{F}}}}$-stratified vector bundle.
[10]{}
. W. A. Benjamin, Inc., New York-Amsterdam, 1967.
On bordism theory of manifolds with singularities. (1973), 279–302 (1974).
. Springer-Verlag, Berlin, 1999.
${K}$-theory of stratified vector bundles, 2002. Max-Planck-Institut für Mathematik, Bonn. Preprint 2002-81 (http://www.mpim-bonn.mpg.de/)
. Cambridge University Press, Cambridge, 1992.
Spaces over a category and assembly maps in isomorphism conjectures in ${K}$- and ${L}$-theory. , 3 (1998), 201–252.
Smooth ${G}$-manifolds as collections of fiber bundles. , 2 (1978), 315–363.
. Allyn and Bacon Inc., Boston, Mass., 1966.
. Proc. Nat. Acad. Sci. U.S.A. 83 (1986), no. 15, 5364–5366.
. J. Amer. Math. Soc. 6 (1993), no. 2, 249–297.
. Cambridge Studies in Advanced Mathematics, 19. Cambridge University Press, Cambridge, 1990.
. Academic Press \[Harcourt Brace Jovanovich Publishers\], New York, 1975. An introduction to algebraic topology, Pure and Applied Mathematics, Vol. 64.
. Proc. Roy. Soc. Edinburgh Sect. A 129 (1999), no. 2, 351–384.
Stratifications of mapping cylinders. Special issue in memory of B. J. Ball. Topology Appl. 94 (1999), no. 1-3, 127–145.
. Surveys on surgery theory, Vol. 2, 319–352, Ann. of Math. Stud., 149, Princeton Univ. Press, Princeton, NJ, 2001.
. McGraw-Hill Book Co., New York, 1966.
Differentiable algebraic topology, 2002. Pre-print.
. Springer-Verlag, New York, 1971. Graduate Texts in Mathematics, Vol. 5.
. Mem. Amer. Math. Soc. 1 (1975), 1, no. 155, xiii+98 pp.
Invent. Math. 68 (1982), no. 3, 353–424.
. J. Amer. Math. Soc. 1 (1988), no. 2, 441–499.
. Springer-Verlag, Berlin, 1998. With a foreword by Haynes Miller.
. Princeton University Press, Princeton, N. J., 1951.
On the [H]{}auptvermutung for manifolds. (1967), 598–600.
. Walter de Gruyter & Co., Berlin, 1987.
. American Mathematical Society, Providence, RI, 1993. Translated from the Russian manuscript by M. Farber, Translation edited by David Louvish.
[^1]: Current address: Max-Planck-Institut für Mathematik, Vivatsgasse, 7 - 53111 Bonn – Germany, [*e-mail: [email protected]*]{}.
[^2]: Current address: Dipartimento di Matematica, del Politecnico di Milano, Piazza Leonardo da Vinci, 32, 20133 Milano – Italy. [*email: [email protected]*]{}.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We calculate the ferroelectric polarization dynamics induced by a femtosecond midinfrared pulse as measured in the recent experiment by R. Mankowsky et al., [Phys. Rev. Lett. 118, 197601 (2017)](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.118.197601). It is due to the nonlinear coupling of the excited infrared-active phonon with the ferroelectric mode or to the excitation of the ferroelectric mode itself depending on the pulse frequency. To begin with, we write the LiNbO$_3$ crystal symmetry invariant thermodynamic potential including electric field and nonlinear phonon coupling terms. We solve the equations of motion determined by this potential for phonon coordinates numerically in classical approximation. We explain the transient polarization reversal observed in the experiment by action of the depolarizing electric field which is due to bound charges at the polarization domain boundaries and give a reasonable estimate for its value. We argue that the polarization could be ultimately reversed when this field is screened.'
author:
- 'Veniamin A. Abalmasov'
title: Ultrafast reversal of the ferroelectric polarization by a midinfrared pulse
---
Introduction
============
Control over polarization is essential for many applications of ferroelectrics, from non-volatile memory storing [@scott2007] to the switchable surface chemistry and catalysis [@kakekhani2016]. The usual way of the polarization reversal by static or pulsed electric fields is limited in speed by hundreds of picoseconds [@li2004]. Several proposals have been made how to switch the polarization on a time scale of picoseconds by directly exciting the ferroelectric mode with ultra-short radiation pulses [@fahy1994; @qi2009; @herchig2014]. However, they still have not been realized in practice. Though, a 90 degrees polarization rotation in a part of domains of a multi-domain ferroelectric thin film of (Ba$_{0.8}$Sr$_{0.2}$)TiO$_3$ induced by a strong single-cycle terahertz pulse was claimed recently in [@grishunin2017]. At the same time, ultra-short (less than 20 ps) all-optical magnetic polarization control with low heat load in transparent ferromagnetic films has already been reported [@stupakiewicz2017].
Recently, it has been proposed to switch the ferroelectric polarization by resonantly exciting infrared-active phonon mode nonlinearly coupled to the ferroelectric mode [@subedi2015]. This approach, developed in last decades with an appearance of very high intensity lasers and called nonlinear phononics, has already proved to be successful in ultrafast lattice control [@foerst2011]. The follow-up experiment [@mankowsky2017] has indeed demonstrated a transient switching of the polarization in LiNbO$_3$ (LNO) crystal for laser pulse fluences larger than about 60 mJ/cm$^{2}$ (with the laser pulse duration of about 150 fs). One of possible explanations of the observed rapid polarization return to its initial state was the formation of uncompensated electric charges after polarization reversal in the irradiated part of the crystal [@mankowsky2017] which has not been taken into account theoretically [@fahy1994; @qi2009; @herchig2014; @subedi2015; @mankowsky2017].
In this paper we calculate the phonon modes dynamics in conditions of the experiment [@mankowsky2017]. We first argue that the equation of motion for the polarization is governed by the thermodynamic potential rather than a potential obtained from ab-initio calculations for the unrelaxed crystal which was used in [@subedi2015; @mankowsky2017]. This implies that nonlinear coupling terms must be invariant under symmetry transformations of the crystal parent group. We find biquadratic coupling constant values from the infrared phonons frequency shift at the ferroelectric phase transition which is known for LNO from ab-initio calculations. We solve numerically equations of motion for the coupled phonon modes and determine the depolarizing electric field value that corresponds better to the polarization dynamics observed in [@mankowsky2017]. We propose to screen this field in experiment by a metallic wire deposited on the crystal surface around the irradiated spot in order to get an ultimate polarization reversal. We also show that an ultimate reversal of polarization in conditions of the experiment [@mankowsky2017] is possible in our model when the ferroelectric mode is excited resonantly, though it demands very high pump fluences.
Theoretical approach {#sec:theor}
====================
In crystals, the movement of interacting with each other atoms near their equilibrium positions is a superposition of a complete set of normal modes $\{Q\}$ which are plain waves with definite frequencies and polarizations and which transform according to irreducible representations of the crystal symmetry group. According to the phenomenological Landau theory, the second order phase transition takes place when the coefficient of the quadratic term of one of the coordinates in the thermodynamic potential $F(\{Q\}, T)$ becomes negative below the critical temperature $T_c$. This leads to a non-zero thermal equilibrium value of this coordinate which is proportional to the spontaneous polarization in ferroelectrics [@landau2013].
It is believed that the thermodynamic potential determines the dynamics of the order parameter as well. In the case of a displacive structural phase transition the corresponding equation is [@ginzburg1980]: $$\begin{aligned}
\label{motion}
\ddot{Q} + \gamma \dot{Q} + \partial F(\{Q\}, T)/\partial Q = 0,\end{aligned}$$ where $\gamma$ is the damping constant. In the static case this equation reduces to the usual thermal equilibrium condition. In the low-frequency range $F(\{Q\}, T)$ is calculated for fixed generalized forces conjugate to all other generalized coordinates. Thus understood Eq. (\[motion\]), if valid up to optical phonon frequencies, expresses the essence of the so-called soft mode concept. Indeed, the square frequency in Eq. (\[motion\]) is equal to the inverse static susceptibility $\chi^{-1}(T) = \partial^2 F(\{Q\}, T)/\partial Q^2$ which becomes zero at $T = T_c$. Only for high-frequency coordinates $F(\{Q\}, T)$ should be calculated for fixed values of the slowly changing coordinates [@ginzburg1980].
We note that in [@subedi2015] the potential $V(\{Q\})$ which determines equations of motion similar to Eq. (\[motion\]) was calculated for PbTiO$_3$ crystal from first principles using density functional theory for fixed values (corresponding to the low-temperature ferroelectric phase) of all the rest of coordinates. For this reason the potential $V(\{Q\})$ was not symmetric in the ferroelectric mode coordinate $Q_P$ (though the symmetry was restored when other coordinates were relaxed for a given value of $Q_P$) [@subedi2015]. The thus obtained potential would be appropriate for a very fast dynamics of the ferroelectric mode. At the same time, in the experiment [@mankowsky2017] the characteristic time change of $Q_P$ is not smaller than its inverse frequency. Moreover, the signal of the second harmonics vanishes at some point. This implies that the crystal becomes centrosymmetric at this moment which is not possible when only the polarization vanishes but other coordinates are not relaxed with their values corresponding to the noncentrosymmetric ferroelectric phase.
LNO thermodynamic potential
===========================
In what follows we will consider for concreteness the LNO crystal and the experiment scheme and conditions as in [@mankowsky2017], see Fig. \[fig:scheme\]. Ferroelectric phase transition in LNO crystal occurs at the temperature about 1480 K from the paraelectric phase with symmetry $R\overline{3}c$ ($D^6_{3d}$) to the low-temperature ferroelectric phase with symmetry $R3c$ ($C^6_{3v}$) [@parlinski2000]. Two formula units in the unit cell implies 27 optical phonon modes, $4A_1 + 5A_2 + 9E$ in the ferroelectric phase and $A_{1g} + 2A_{1u} + 3A_{2g} + 3A_{2u} + 4E_g + 5E_u$ in the paraelectric phase. In the ferroelectric phase infrared-active polar optical phonon modes $A_1$(TO$_{1-4}$) have frequencies about 7.5, 8.1, 10 and 19 THz and $E$(TO$_{1-9}$) modes are with frequencies about 4.6, 7.0, 7.9, 9.7, 10.8, 11.1, 13.0, 17.3 and 19.8 THz [@kojima2016; @margueron2012]. In the paraelectric phase $A_1$(TO$_3$) has irreducible representation $A_{1g}$ while the others are $A_{2u}$ [@parlinski2000]. The mode $A_1$(TO$_1$) becomes softer approaching the phase transition [@ridah1997] and according to the first-principles calculations [@veithen2002] it has the strongest overlap (0.82) with the mode $A_{2u}$ which is unstable in the paraelectric phase and coincides with the atomic displacements during the phase transition. The mixing of $A_1$(TO$_1$) and $A_1$(TO$_2$) modes at temperatures between 400 and 600 K [@ridah1997] is probably the cause of the incomplete overlap in the first-principles calculations. We will label the ferroelectric soft mode as $Q_P$ and the other infrared-active modes $Q_{\text{IR}}$ as $Q_{A_{1g}}$, $Q_{A_{2u}}$, $(Q_{E_{g}, x}, Q_{E_{g}, y})$ and $(Q_{E_{u}, x}, Q_{E_{u}, y})$ according to their irreducible representation.
![Sketch of the bulk and domain polarization ${\bf P}_{b,d}$ and electric field $\bf{E}$ due to the bound charges in the sample when irradiated by a terahertz pulse; free charges on the surface are in circles (a) and the time-dependence of the terahertz radiation electric field (thick, red) and the corresponding infrared mode amplitude $Q_{A_1\text{(TO)}_4}$ (blue) nonlinearly coupled to the ferroelectric mode at a fluence of 95 mJ/cm$^{2}$ (b).[]{data-label="fig:scheme"}](fig1){width="0.9\columnwidth"}
During the ferroelectric phase transition the condensation of the soft mode occurs at the center of the Brillouin zone. The teraherz pulse as in the experiment [@mankowsky2017] also excites mostly long-wavelength phonons. So we will consider an homogeneous case and neglect the interaction of phonons with different wavelengths coming from nonlinear phonon coupling.
The thermodynamic potential density $F(\{Q\}, T)$ that we want to write should be invariant under transformations of the parent paraelectric phase crystal symmetry group. We will write it as a sum $$\begin{aligned}
\label{F}
F = F_0 + F_{E\text{-ph}} + F_{\text{ph-ph}}.\end{aligned}$$ The first part describes free phonons: $$\begin{aligned}
\label{F0}
F_0 = - \frac{\omega^2_P}{4} Q_P^2 + \frac{c_P}{4} Q_P^4 + \frac{\omega_{\text{IR}}^2}{2} Q_{\text{IR}}^2,\end{aligned}$$ where $\omega_{P, {\text{IR}}}$ are frequencies of the corresponding modes $Q_{P, {\text{IR}}}$. The coefficient $c_P$ determines the equilibrium value $Q_P^e$ through the equation $\partial F/\partial Q_P = 0$. In the absence of electric field and nonlinear phonon interactions this gives $c_P = \omega_P^2/2 (Q_P^e)^2$. Summation over all infrared-active modes $Q_{\text{IR}}$ in Eq. (\[F0\]) is assumed.
As a function of the ferroelectric mode amplitude $F_0(Q_P)$ was calculated ab-initio, for instance, in [@inbar1995] and when fitted to a forth order polynomial it provided the energy difference between the ground and the lowest excited state in agreement with experimental data.
The second part of the thermodynamic potential corresponds to the phonon-electric field coupling: $$\begin{aligned}
\label{F_E-ph}
F_{E\text{-ph}} = & - E_z (Z^*_P Q_P + Z^*_{A_{2u}} Q_{A_{2u}} + Z_{A_{1g}} Q_P Q_{A_{1g}}) \nonumber \\
& - \sum_{j=x,y} E_j (Z^*_{E_{u}} Q_{E_{u}, j} + Z_{E_{g}} Q_P Q_{E_{g}, j}),\end{aligned}$$ where $Z^*_{P, {\text{IR}}}$ are Born effective charges of the corresponding modes $Q_{P, {\text{IR}}}$, $Z_{A_{1g}, E_{g}}$ are coupling constants, $E_{x, y, z}$ are the electric field components.
We see from Eq. (\[F\_E-ph\]) that coupling of the electric field to symmetric phonon modes $A_{1g}$ and $E_{g}$ is possible only in ferroelectric phase, when $\langle Q_P \rangle \neq 0$. Though, even in ferroelectric phase this coupling is expected to be not large due to the small value of $\langle Q_P \rangle$. Indeed, the effective charge of $A_1$(TO$_3$), which has irreducible representation $A_{1g}$ in paraelectric phase, is very small [@kojima2016].
We will focus on a single phonon mode excitation and as a consequence only on two-phonon modes coupling, one of each is the ferroelectric soft mode $Q_P$ and the other is an infrared-active mode $Q_{\text{IR}}$ (see [@juraschek2017; @radaelli2018] for discussions on three phonon modes interaction). The leading coupling terms in phonon amplitudes (up to the forth order) are $$\begin{aligned}
\label{F_ph-ph}
& F_{\text{ph-ph}} = \sum_{i=1,2} a_{i} Q_P^2 Q_{A_{1g}}^i + \sum_{i=1,2,3} c_{i} Q_P^{i} Q_{A_{2u}}^{4-i} \nonumber \\
& + \sum_{\substack{i=g,u \\ j=x,y}} b_{i} Q_P^2 Q_{E_{i}, j}^2 + d\, Q_P Q_{E_{u}, y} (3 Q_{E_{u}, x}^2 - Q_{E_{u}, y}^2).\end{aligned}$$
Three-phonons interaction in Eq. (\[F\_ph-ph\]) with the coupling constant $a_1$ is not of much interest to us because it is difficult to excite $A_{1g}$ mode as we discussed above (since its effective charge is proportional to $Q_P$, this term effectively is of the forth order in phonon coordinates).
In what follows we apply the thermodynamic potential (\[F\]) through Eq. (\[motion\]) to describe the results of the experiment [@mankowsky2017] where the mode $A_1$(TO$_4$) was pumped resonantly by a high-intensity femtosecond electromagnetic pulse. For this purpose we precise in the next section the numerical values of parameters which enter Eqs. (\[motion\]) - (\[F\]).
Values of parameters
====================
The electric field in Eq. (\[F\_E-ph\]) has two constituents, $E_z = E_1 + E_2$, the both being directed along $z$-axis in the experiment [@mankowsky2017]. One is the driving midinfrared pulse electric field $E_1 (t) = E_0 \sin(\omega t) \exp(-4\ln2 \,t^2 / T^{2})$ with frequency $\omega$, Gaussian envelope of duration $T = 0.15$ ps and amplitude $E_0$ up to 25 MV/cm in the experiment [@mankowsky2017], see Fig. \[fig:scheme\](b). The other component is due to the depolarizing electric field $E_d$ in the reversed polarization domain created by the terahertz pump (Fig. \[fig:scheme\](a)). We suppose there is no screening of this field by free carriers on the time scale of the polarization reversal as in the experiment [@mankowsky2017]. So the resulting field is $E_2(t) = E_d (1 - Q_P(t)/Q_P^e)$.
We calculate the effective electric charge of a given optical mode from the experimental value of its oscillation strength [@hoegen2018]. Thus, we obtain $Z^*_P= 1.356$, $Z^*_{A_1(\text{TO}_2)} = 0.564$ and $Z^*_{A_1(\text{TO}_4)} = 1.404$ $e/\sqrt{\text{u}}$. The effective charge $Z^*_{A_1(\text{TO}_3)}$ is very small [@ridah1997] and we will neglect the dynamics of this mode.
The ions shifts between paraelectric and ferroelectric phase at room temperature according to [@lines2001] correspond to $Q_P^e = 2.9 \sqrt{\text{u}}$[Å]{} which agrees with the minimum position of the two-minimum energy surface calculated in [@inbar1995]. At the same time, according to the experimental data [@boysen1994] the Li and O atoms are shifted in the ferroelectric phase by about $\Delta z_{\text{Li}} = 0.460$ [Å]{} and $\Delta z_{\text{O}} = 0.270$ [Å]{} which gives the amplitude $Q_P^e = (\sum_i m_i \Delta z_i^2)^{1/2} = 3.1 \sqrt{\text{u}}$[Å]{}. The same value of $Q_P^e$, which we adopt in our calculations, follows from atomic displacements calculated ab-initio [@parlinski2000; @veithen2002; @friedrich2015]. $Q_P^e$ enters Eq. (\[motion\]) as an initial value of the $Q_P(t)$ coordinate. Initial values of other coordinates are obtained from the equilibrium condition $\partial F / \partial Q_{\text{IR}} = 0$ and they are not zero when the modes are nonlinearly coupled, see Fig. \[fig:scheme\](b). This also can explain a partial overlap of ferroelectric soft modes in para- and ferroelectric phases as it was discussed above.
We note that the spontaneous polarization $P_s = Z_P^*Q_P^e/v_0$, $v_0$ being the unit cell volume, calculated for $Q_P^e = 3.1 \sqrt{\text{u}}$[Å]{} is about $0.64$ C/m$^{2}$ and slightly lower than the experimental value about $P_s = 0.70$ C/m$^{2}$ [@veithen2002] which in its turn is attained for a rather large value $Q_P^e = 3.4 \sqrt{\text{u}}$[Å]{}. This slight discrepancy can also be attributed to the highly nonlinear evolution of the charges along the ferroelectric path of atomic displacements [@veithen2002].
We adopt the values for the damping constants in Eq. (\[motion\]) from [@hoegen2018] to be $\widetilde{\gamma}_P= 0.8$, $\widetilde{\gamma}_{A_1(\text{TO}_2)} = 0.6$ and $\widetilde{\gamma}_{A_1(\text{TO}_4)} = 1.0$ THz (with $\gamma = 2 \pi \widetilde{\gamma}$). The damping $\widetilde{\gamma}_P$ increases strongly with temperature and equals the soft mode frequency of 5 THz at about 1100 K [@ridah1997]. It is not clear whether this is due to the mode softening or just to the temperature dependence. We keep the damping $\widetilde{\gamma}_P$ constant in our calculations.
Finally, we note that the biquadratic phonon-phonon interaction with the coupling constant $c_2$ in Eq. (\[F\_ph-ph\]) does change not only the frequency of the soft mode $Q_P$ but the frequency of the $Q_{\text{IR}}$ as well, Eq. (\[F0\]). This allows, for instance, in some cases to reproduce the polarization temperature dependence from high-frequency modes temperature dependence [@krylov2013; @salje1997]. Usually the effect is not large. For the ferroelectric KDP crystal, however, the change is about ten percent for certain modes indicating both signs of the coupling constant [@brehat1987; @simon1988]. So we calculate the coupling constant as $c_{2} = (\omega_{\text{IR}}^2 - \Omega_{\text{IR}}^2)/2 (Q_P^e)^2$ where $\omega_{\text{IR}}$ and $\Omega_{\text{IR}}$ are the frequencies of the $Q_{\text{IR}}$ mode in the ferroelectric and paraelectric phases respectively, $Q_P^e$ is the thermal equilibrium value of $Q_P$ at room temperature. The positive sign of the coupling constant assures a single well potential $F(Q_P)$ for large values of $Q_{\text{IR}}^2$ and thus a possibility of the polarization reversal. We note that $\omega_{\text{IR}}$ does not change substantially in the ferroelectric phase of LNO up to about 1000 K [@johnston1968; @ridah1997] but this agrees with a small change in the polarization itself in this temperature range [@shostak2009].
Due to the very high temperature of the phase transition the values of $\Omega_{\text{IR}}$ for LNO crystal are available only from first-principles calculations. For $A_1$(TO$_4$) they vary from 15.6 [@parlinski2000] and 14.3 [@veithen2002] to 13.6 THz [@friedrich2015], for $A_1$(TO$_2$) from 3.5 [@parlinski2000] and 2.8 [@veithen2002] down to 0.9 THz [@friedrich2015]. For $A_1$(TO$_3$) in contrast the frequency is larger in the paraelectric phase varying from 12.1 [@parlinski2000; @veithen2002] to 10 THz [@friedrich2015] which indicates a possible negative value of the biquadratic phonon-phonon coupling constant. We adopt values $\Omega_{A_1(\text{TO}_4)} = 14$ THz and $\Omega_{A_1(\text{TO}_2)} = 3.5$ THz which correspond to $c_{2, A_1(\text{TO}_4)} = 34$ meV/u$^{2}$[Å]{}$^{4}$ and $c_{2, A_1(\text{TO}_2)} = 11$ meV/u$^{2}$[Å]{}$^{4}$. We note that our coupling constants appear to be an order of magnitude smaller than those obtained for quantum paraelectric crystals KTaO$_3$ [@subedi2017] and SrTiO$_3$ [@kozina2019] from DFT calculations. This might be due to the calculation procedure of the potential discussed in Sec. \[sec:theor\]. Indeed, a relaxed lattice has the lowest energy and, as a consequence, the nonlinear phonon coupling constants of our thermodynamic potential are smaller.
We keep coupling constants $c_1$ and $c_3$ in Eq. (\[F\_ph-ph\]) and the depolarizing electric field $E_d$ in Eq. (\[F\_E-ph\]) as fitting parameters in our calculations when compared to the experimental results [@mankowsky2017].
![Ferroelectric mode dynamics for the pump pulse fluences of 30, 60 and 95 mJ/cm$^{2}$ and frequency $\omega = 19$ THz when $c_{1, 3} = 0$ and $E_d = 0$ (dashed) and $c_1 = 10$, $c_3 = 17$ meV/u$^{2}$[Å]{}$^{4}$ for $A_1(\text{TO}_4)$ and $E_d = 2.8$ MV/cm (solid). Red dots are at minimums of $Q_P$ []{data-label="fig:QP"}](fig2){width="0.8\columnwidth"}
Calculation results
===================
We start with zero values of coupling constants $c_{1, 3}$ and the depolarizing electric field $E_d$ and calculate the phonon modes dynamics for three values of the infrared pulse fluence, see Fig. \[fig:QP\]. In the experiment [@mankowsky2017] for fluences above a threshold value of 60 mJ/cm$^{2}$ the second-harmonic intensity (and thus the ferroelectric mode $Q_P$) was observed to vanish completely. We see, however, only a reduction of it at this fluence. The dynamics becomes closer to the experiment if we put $c_{2, A_1(\text{TO}_4)} = 51$ meV/u$^{2}$[Å]{}$^{4}$ but this corresponds to the very low frequency $\Omega_{A_1(\text{TO}_4)} = 10.7$ THz expected in paraelectric phase. Interestingly, the situation can be improved if we take a positive value of the constant $c_{3, A_1(\text{TO}_4)} \gtrsim 13$ meV/u$^{2}$[Å]{}$^{4}$. It is rather unexpected because the force exerted by this coupling on $Q_P$ oscillates and changes its sign. At the same time, negative values of $c_{3, A_1(\text{TO}_4)}$ make the polarization reversal even harder. Finite values of $c_{1, A_1(\text{TO}_4)}$ do not influence much the dynamics. This can be easily understood because the latter coupling is cubic in $Q_{\text{IR}}$ while the former is linear and $Q_{\text{IR}} \ll Q_P$.
The reentrant behavior of the polarization for the largest fluence of 95 mJ/cm$^{2}$ available in [@mankowsky2017] appears for depolarizing electric fields larger than about $E_d \approx 2.5$ MV/cm. This value to be compared with the depolarizing field in a plate-like monodomain sample $E_{d} = P_s / (\varepsilon\varepsilon_0)$ which is about 26.4 MV/cm for $P_s= 0.70$ C/m$^{2}$ and the dielectric constant $\varepsilon_{33} = 30$ in LNO ($\varepsilon_0$ is the electric constant). The value of the depolarization factor $N \approx 0.1$ seems to be reasonable taken into account the oblong shape of the reversed polarization domain created (the pump penetration depth is about 3.2 $\mu$m [@mankowsky2017] and its spot size is about 65 $\mu$m [@hoegen2018], see Fig. \[fig:scheme\](a)).
Finally, we adopt the values $E_d = 2.8$ MV/cm and $c_1 = 10$ and $c_3 = 17$ meV/u$^{2}$[Å]{}$^{4}$ for $A_1(\text{TO}_4)$ mode, see Fig. \[fig:QP\]. For $A_1(\text{TO}_2)$ we keep these coupling constants zero since its dynamics (due to the small Born electric charge $Z^*_{A_1(\text{TO}_2)}$) does not influence visibly the dynamics of $Q_P$ even near the resonance which is close to the resonance of $Q_P$ itself.
![Minimum of $Q_P$ (normalized by $Q_P^e$) as a function of fluence for different frequencies of the pump pulse. At $\nu_{\text{pump}} = 7$ THz two lines correspond to opposite signs of the pump pulse electric field. Experimental data are from [@mankowsky2017][]{data-label="fig:minQP"}](fig3){width="0.8\columnwidth"}
![Susceptibility as a function of the pump frequency normalized to unity at 19 THz. Experimental data are from [@mankowsky2017][]{data-label="fig:suscept"}](fig4){width="0.8\columnwidth"}
We calculate the minimum value of $Q_P$ (normalized by $Q_P^e$) as a function of fluence for different frequencies of the pump pulse (with the same Gaussian envelope duration) and compare our results to the experiment [@mankowsky2017], see Fig. \[fig:minQP\]. The agreement is good enough. For the pump frequency of 7 THz, which is close to the ferroelectric mode resonance, the result depends slightly on the sign (phase) of the pulse oscillations due to comparable values of the pump duration and the period of oscillations. For this pump frequency we see an ultimate reversal of the polarization for fluences higher than about 40 mJ/cm$^{2}$. The reversal becomes even harder for larger values of the depolarizing field until it becomes impossible if $E_d \gtrsim 5$ MV/cm. This threshold corresponds to the coercive field value in our model, $E_{c} = P_s / (3\sqrt{3} \varepsilon\varepsilon_0)$ in the absence of electric field and nonlinear phonons coupling, at which the metastable state (local minimum) with opposite polarization disappears. We note, however, that it is an order of magnitude larger than experimental values of the coercive field in this crystal [@volk2008].
The dependence of the susceptibility $\text{min} Q_P / \text{fluence}$ calculated for small fluences as a function of the pump frequency reproduces the experimental results [@mankowsky2017] as well (Fig. \[fig:suscept\]). The width of peaks is determined mostly by the frequency width of the pump pulse.
Discussion
==========
In our calculations we see the oscillations of $Q_P$ with the ferroelectric mode frequency which are absent in the experiment [@mankowsky2017]. In the experiment these oscillations could be smeared for several reasons. First, initial values of velocities are not zero and those of coordinates are not at equilibrium but they are determined instead by the temperature and the coordinates wave functions. Second, the electric field amplitude of the midinfrared pulse which penetrates the crystal is not homogeneous and is determined by the Gaussian function perpendicular to its direction and a vanishing exponential deep into the crystal, $\exp(-z/z_{0})$ with $z_{0} \approx 3.5 \,\mu$m, while the second harmonic is generated at $z \lesssim 1 \,\mu$m [@mankowsky2017]. Finally, the nonlinear coupling to the phonon modes with nonzero wave numbers which we have not taken into account would probably also lead to the polarization oscillations smearing.
![Sketch of the screen (in golden) on the top of the crystal around the pump laser spot and the electric charges distribution (free charges are in circles)[]{data-label="fig:screen"}](fig5){width="0.98\columnwidth"}
High-intensity sources of far-infrared electromagnetic fields (more than 3.5 MV/cm) tunable between 4 and 18 THz were reported recently in [@liu2017]. These fluences, however, are not high enough to switch the polarization by the resonant pump of the ferroelectric mode (Fig. \[fig:minQP\]). A very high-intensity source of terahertz single-cylce pulse (up to 22 MV/cm) was reported in [@ovchinnikov2019] but its frequency is about 1 THz only. Thus, the resonant ferroelectric mode reversal in LNO seems to be unattainable at the moment.
Our calculations show, however, that the polarization can be ultimately switched in the absence of the depolarizing field (Fig. \[fig:QP\]). In order to screen this field a metallic wire could be used which is drawn in golden in Fig. \[fig:screen\]. The relaxation time of this screen $\tau = R C$ is determined by the resistance $R \approx \rho L / (h + d) \delta$ and capacitance $C \approx \varepsilon\varepsilon_0 h$ of the wire. The resistivity and skin depth for gold at frequency 10 THz are $\rho \sim 10 ^{-7} \,\Omega$ m and $\delta \approx 30$ nm. The metallic wire length, height and width are about $L \approx 400 \,\mu$m, $h \sim d \sim 10 \div 100 \,\mu$m. This yields the relaxation time smaller than $\tau \sim 0.1$ ps, which is enough to follow the polarization dynamics (Fig. \[fig:QP\]).
Recently, anharmonic oscillations of the lowest frequency $E$ mode in LNO were studied under resonant excitation by a single-cycle pulse with the electric field about 1 MV/cm [@dastrup2017]. At the same time, it would be interesting to see the dynamics of the ferroelectric mode $Q_P$ induced by the resonant excitation of the $E$ modes, Eqs. (\[F\_E-ph\])-(\[F\_ph-ph\]), due to their nonlinear coupling as it was described in this work for the $A$ modes excitation, especially keeping in mind orthogonal polarizations of the $A$ and $E$ modes. Calculated ab-initio $E$ modes frequencies [@parlinski2000; @friedrich2015] differ substantially in para- and ferroelectric phases meaning possible strong coupling constants of these modes to the ferroelectric mode.
Conclusion
==========
Our calculations of the ferroelectric mode dynamics in LNO, determined by the parent phase symmetry-invariant thermodynamic potential, reproduce well the transient reversal of polarization under high-frequency mode excitation reported in [@mankowsky2017] when the electric field of the bound charges is taken into account. The estimated strength of this field agrees with the polarization value in LNO and the expected depolarization factor of the transiently created polarization domain. We argue that the polarization could be ultimately reversed if the depolarizing field is screened, for example by the metallic wire on the top of the crystal around the pump laser spot. We preview that the same dynamics of the polarization could be probed by the resonant excitation of the $E$ modes which are orthogonal to the polarization.
Acknowledgments {#acknowledgments .unnumbered}
===============
I thank Roman Mankowsky and Alaska Subedi for useful discussions.
The study was carried out with the financial support of the Russian Foundation for Basic Research in the framework of the scientific project No. 18-02-00399.
[37]{}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 [****, ()](https://doi.org/10.1126/science.1129564) [****, ()](https://doi.org/https://doi.org/10.1016/j.susc.2015.10.055) [****, ()](https://doi.org/10.1063/1.1644917) [****, ()](https://doi.org/10.1103/PhysRevLett.73.1122) [****, ()](https://doi.org/10.1103/PhysRevLett.102.247603) [****, ()](https://doi.org/10.1063/1.4890128) [****, ()](https://doi.org/10.1038/s41598-017-00704-9) [****, ()](https://doi.org/10.1038/nature20807) [****, ()](https://doi.org/10.1103/PhysRevB.92.214303) [****, ()](https://doi.org/10.1038/nphys2055) [****, ()](https://doi.org/10.1103/PhysRevLett.118.197601) [**](https://books.google.ru/books?id=-I6GDAAAQBAJ), ed., Vol. (, ) [****, ()](https://doi.org/https://doi.org/10.1016/0370-1573(80)90117-9) [****, ()](https://doi.org/10.1103/PhysRevB.61.272) [****, ()](https://doi.org/10.7567/jjap.55.10tc02) [****, ()](https://doi.org/10.1063/1.4716001) [****, ()](https://doi.org/10.1103/PhysRevB.56.5967) [****, ()](https://doi.org/10.1103/PhysRevB.65.214302) [****, ()](https://doi.org/10.1080/00150199508221829) [****, ()](https://doi.org/10.1103/physrevlett.118.054101) [****, ()](https://doi.org/10.1103/PhysRevB.97.085145) [****, ()](https://doi.org/10.1038/nature25484) [**](https://books.google.ru/books?id=p6ruJH8C84kC), International series of monographs on physics (, ) [****, ()](https://doi.org/10.1107/S0108768193012820) [****, ()](https://doi.org/10.1088/0953-8984/27/38/385402) [****, ()](https://doi.org/10.1002/jrs.4263) [****, ()](https://doi.org/10.1080/01411599708228789) [****, ()](https://doi.org/10.1007/BF01012551) [****, ()](https://doi.org/10.1103/PhysRevB.37.1969) [****, ()](https://doi.org/10.1103/PhysRev.168.1045) [****, ()](https://doi.org/10.1134/S1063774509030195) [****, ()](https://doi.org/10.1103/physrevb.95.134113) [****, ()](https://doi.org/10.1038/s41567-018-0408-1) [**](https://doi.org/10.1007/978-3-540-70766-0) (, ) [****, ()](https://doi.org/10.1364/ol.42.000129) [****, ()](https://doi.org/10.1038/s41598-019-46284-8) [****, ()](https://doi.org/10.1063/1.4980112)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider the stable dependence of solutions to wave equations on metrics in $C^{1,1}$ class. The main result states that solutions depend uniformly continuously on the metric, when the Cauchy data is given in a range of Sobolev spaces. The proof is constructive and uses the wave packet approach to hyperbolic equations.'
address: 'Department of Mathematics and Statistics / RNI, University of Helsinki'
author:
- Mikko Salo
title: |
Stability for Solutions of Wave Equations\
with $C^{1,1}$ Coefficients
---
Introduction
============
We consider the wave equation in ${\mathbf{R}}_t \times {\mathbf{R}}^n_x$, $$\left\{ \begin{array}{rl}
(D_t^2 - A(x,D_x))u(t,x) &\!\!\!= F(t,x), \\[4pt]
u|_{t=0} &\!\!\!= f, \\[4pt]
\partial_t u|_{t=0} &\!\!\!= g.
\end{array} \right.$$ Here $A(x,D_x) = a^{ij}(x) D_{x_i} D_{x_j}$ is a uniformly elliptic operator, satisfying $a^{ij} = a^{ji}$, and $a^{ij}(x) \xi_i \xi_j \geq c {\lvert \xi \rvert}^2$ for $\xi \in {\mathbf{R}}^n$. We assume that the functions $a^{ij}$ are in the space $C^{1,1}({\mathbf{R}}^n)$, with norm ${\lVert a \rVert}_{C^{k-1,1}} = \sum_{{\lvert \alpha \rvert} \leq k} {\lVert \partial^{\alpha} a \rVert}_{L^{\infty}}$.
The question investigated in this article is the stable dependence of the solution $u$ on the metric $(a^{ij})$. Intuition for the problem can be obtained from the simplest possible case, namely the one-dimensional wave equation with constant sound speed $c$. Given $f \in L^2({\mathbf{R}})$, the equation $$\begin{aligned}
& (\partial_t^2 - c^2 \partial_x^2) u(t,x) = 0, & \\
& u(0) = f, \partial_t u(0) = 0 &\end{aligned}$$ has the solution $u(t,x) = \frac{1}{2} [ f(x-ct) + f(x+ct) ]$. Thus, the solution at time $t = 1$ is obtained by translating $f$ by $c$ units in the positive and negative directions. Since translation of $L^2$ functions is a uniformly continuous operation, we see that $u(1,\,\cdot\,)$ depends uniformly continuously in $L^2$ norm on the sound speed $c$. A stronger result may be obtained if the initial data is smoother: if $f \in H^1$, then ${\lVert f(\,\cdot\,-c) - f(\,\cdot\,-c') \rVert}_{L^2} \leq {\lVert \nabla f \rVert}_{L^2} {\lvert c-c' \rvert}$, and $u(1,\,\cdot\,)$ depends Lipschitz continuously in $L^2$ on the sound speed.
Our main result is the following theorem, which shows that uniformly continuous or Lipschitz dependence are valid also for general wave equations with $C^{1,1}$ metrics. In this introduction, we state the theorem only in the case where an initial velocity $g$ is present. In Section 7 we will give the straightforward extensions to cases where driving terms $F$ and initial positions $f$ are present, and also where $A(x,D_x)$ is replaced by a divergence form or Laplace-Beltrami operator.
\[thm:main\] Suppose $M$ is a large constant such that $$\begin{aligned}
& {\lVert a^{ij} \rVert}_{C^{1,1}} \leq M, \quad a^{ij} \xi_i \xi_j \geq M^{-2} {\lvert \xi \rvert}^2, & \label{a_assumptions} \\
& t \in [-M,M]. & \label{t_assumptions}\end{aligned}$$ If $-1 \leq \alpha \leq 2$, then for each $g \in H^{\alpha}$ there is a unique weak solution $u \in C([-M,M];H^{\alpha+1}) \cap C^1([-M,M];H^{\alpha})$ of the problem $$\label{main_cauchy}
\left\{ \begin{array}{rl}
(D_t^2 - A(x,D_x))u(t,x) &\!\!\!= 0, \\[4pt]
u|_{t=0} &\!\!\!= 0, \\[4pt]
\partial_t u|_{t=0} &\!\!\!= g.
\end{array} \right.$$ If $-1 \leq \alpha < 2$, and if $A = (a^{ij})$, $B = (b^{ij})$ satisfy and $u_A$, $u_B$ are the corresponding solutions, then for any $\varepsilon > 0$ there is $\delta > 0$ such that $$\label{uniform_stability}
{\lVert u_A(t) - u_B(t) \rVert}_{H^{\alpha+1}} < \varepsilon \ \text{ whenever }\ {\lVert A-B \rVert}_{C^{0,1}} < \delta.$$ Further, if $-1 \leq \alpha \leq 1$ and $g \in H^{\alpha+1}$, then $$\label{lipschitz_stability}
{\lVert u_A(t) - u_B(t) \rVert}_{H^{\alpha+1}} \leq C {\lVert A-B \rVert}_{C^{0,1}} {\lVert g \rVert}_{H^{\alpha+1}}$$ where $C$ depends only on $M$ and $n$.
Here, by a weak solution we mean a function $u \in C([-M,M];H^{\alpha+1}) \cap C^1([-M,M];H^{\alpha})$ which solves the equation in the sense of distributions, and satisfies the initial conditions in the vector-valued sense.
Existence and uniqueness of solutions to linear wave equations is of course classical and can be established via energy estimates under quite general conditions. We refer to [@stolk] for a comprehensive account and further references. A constructive method, valid in the setting of our theorem, for proving existence and uniqueness of weak solutions was introduced in [@smith1] using a wave packet approach.
Stability estimates such as the ones in Theorem \[thm:main\] are also classical, see [@stolk] and references given there. In Section 7 we give an easy argument which uses just the existence and uniqueness of solutions, without any precise knowledge about the solutions.
The novelty is that our proof of Theorem \[thm:main\] is constructive: an explicit expression for the solution is given, and the stability properties are deduced from that. The constructive method gives the same intuition to stability as in the $n=1$ case, namely that stability for solutions should be the same as for translating functions. We use the wave packet approach introduced in [@smith1], where the initial data is decomposed into wave packets, and an approximate solution to the equation is obtained by translating wave packets along the Hamilton flow. The stability is exactly governed by this translation.
The main motivation for this study comes from inverse problems in seismic imaging, where the analysis of solutions of wave equations has a key role. Many existing results, see [@dehoop_msri] for a survey, assume a linearization about a smooth sound speed $c_0$ and use microlocal analysis and calculus of Fourier integral operators (FIOs).
There has been recent interest, see [@dsu], in the practically more realistic case where $c_0$ is not smooth, and in this case few results are known. One reason is that there is no calculus of nonsmooth FIOs. However, solution operators for nonsmooth wave equations are understood quite well due to the wave packet approach of [@smith1], and a more precise analysis of these operators is expected to lead to new results. This work is an attempt in this direction. The related work [@dsu] discusses propagation of singularities (in terms of concentration of wave packets) and application of wave packets in numerical computations. We remark that wave packets are the same as curvelets [@candesdemanet], [@candesdonoho], which have been introduced in image processing as an efficient way of representing functions with singularities on smooth curves.
The proof of Theorem \[thm:main\] is based on constructing an explicit solution operator for the nonsmooth wave equation, following the method of [@smith1]. The idea is to localize the initial data to dyadic frequency shells, and write each localized piece as a superposition of wave packets at the given frequency. The action of the wave group on a wave packet is well approximated by translating the packet along the Hamilton flow, and this gives an approximate solution operator for frequency localized initial data. These are added up to obtain a parametrix for the full equation.
Actually, to handle the nonsmooth coefficients, at frequency level $2^k$ one truncates the coefficients to frequencies less than $2^{k/2}$ and uses the Hamilton flow for the truncated metric. The error terms resulting from this will add up to a bounded operator, and this error can be iterated away by solving a Volterra equation.
The precise construction of solution operator will be a combination of methods in [@smith1] and [@smith2] and it does not involve new ideas. However, in view of the stability result, we need to give the construction in detail to see how the operator depends on the metric. The main outline is the same as in [@smith1], which used a discrete wave packet frame. Some computations are simplified if one uses instead a continuous wave packet representation (i.e. the FBI transform, see [@delort]). This was used in [@tataru1], [@tataru2], [@tataru3] for wave packets based on the Gaussian. We will follow [@smith2] which used instead wave packets compactly supported in frequency, a property which will keep the dyadic frequency annuli separated.
The crux of the stability proof is Lemma \[flowstability\], which considers the stability of translating along Hamilton flow. Lipschitz stability for translation involves a loss of one derivative, and the main point in the proof is that there is a smooth deformation of two Hamilton flows obtained by deforming the corresponding metrics. For the full result, one needs to check that the several corrections required to obtain an exact solution operator, most importantly the Volterra iteration, do not affect the stability given by translation.
We first prove Lipschitz stability with a loss of one derivative, and the uniform continuity is an immediate consequence. Since there is a loss of one derivative arising from the flow, we can afford to lose one derivative in certain other estimates as well. In this way, we get stability in terms of the $C^{0,1}$ norm of the metric instead of $C^{1,1}$ norm.
The plan of the paper is as follows. Section 2 contains some basic facts about Hamilton flows, and Section 3 introduces the FBI transform. In Section 4 we outline the construction of the solution operator, and Section 5 contains the details. The stability result, Theorem \[thm:main\], is proved in Section 6. Section 7 discusses variations of the stability result.
[**Notation.**]{} We write $D_t = \frac{1}{i} \partial_t$ and $D_{x_j} = \frac{1}{i} \partial_{x_j}$. The gradient with respect to $x$ or $\xi$ is denoted by $d_x$ or $d_{\xi}$, and $D = \frac{1}{i} d$. Throughout the paper, $M$ will be a large constant so that and are satisfied. We write $a \lesssim b$ if $a \leq Cb$ where $C > 0$ depends only on $M$ and the dimension $n$. Also, $a \sim b$ means that $a \lesssim b$ and $b \lesssim a$. We write $L^p = L^p({\mathbf{R}}^n)$, similarly for the $L^2$ Sobolev spaces $H^{\alpha}$, the spaces $C^{0,1}$ and $C^{1,1}$, and the Schwartz space ${\mathscr{S}}$. The mixed norm spaces are denoted by $L^p_t H^{\alpha}_x = L^p([-M,M] ; H^{\alpha})$, similarly $C^j_t H^{\alpha}_x = C^j([-M,M] ; H^{\alpha})$.
Acknowledgements {#acknowledgements .unnumbered}
----------------
Research partly supported by the Academy of Finland. Part of this research was carried out during visits at MSRI and at the University of Washington, and I wish to express my gratitude to these institutions for their hospitality and support. I would like to thank Hart Smith for generous advice, and Gunther Uhlmann for suggesting the problem.
Hamilton flow
=============
We record for later use some elementary facts related to Hamilton flows. Let $\chi(\xi)$ be a smooth cutoff supported in the unit ball with $\chi = 1$ for ${\lvert \xi \rvert} \leq 1/2$. We define smooth approximations of the coefficients $a^{ij}$ by $$a^{ij}_k(x) = \chi(2^{-k/2}D_x) a^{ij}(x).$$ Then $a^{ij}_k$ is supported in frequency in $\{{\lvert \xi \rvert} \leq 2^{k/2}\}$, and satisfies derivative bounds $$\label{truncation_est}
{\lVert \partial^{\alpha} a^{ij}_k \rVert}_{L^{\infty}} \lesssim 2^{\frac{k}{2} \max(0,{\lvert \alpha \rvert}-2)}. \\[5pt]$$
Consider the Hamilton flow related to $D_t^2 - A_k(x,D_x)$, where $A_k(x,D_x) = a^{ij}_k(x) D_{x_i} D_{x_j}$. It will be useful to define this in terms of the functions $$p_k^{\pm}(x,\xi) = \pm \chi(2^{-k/2} D_x)\sqrt{A_k(x,\xi)}.$$ The Hamilton equations are given by $$\begin{aligned}
\dot{x}(t) &= d_{\xi} p(x(t),\xi(t)), \\
\dot{\xi}(t) &= -d_x p(x(t),\xi(t)),\end{aligned}$$ where $p = p_k^{\pm}$, $x = x_k^{\pm}$, and $\xi = \xi_k^{\pm}$. Here $(x,\xi) \mapsto (d_{\xi} p(x,\xi), -d_x p(x,\xi))$ is a smooth vector field in $T^* {\mathbf{R}}^n = {\mathbf{R}}^n_x \times ({\mathbf{R}}^n_{\xi} \smallsetminus \{0\})$, so given an initial condition $(x(0),\xi(0)) = (y,\eta) \in T^* {\mathbf{R}}^n$, the Hamilton equations have a solution $(x(t),\xi(t))$ depending smoothly on $t,y,\eta$ at least for small time.
It is well known that the solution $(x(t),\xi(t))$ exists for all time. For, if it exists for $t$ in an interval $I = (-t_0,t_0)$, then $\xi(t) \neq 0$ for $t \in I$, and $h(t) = {\lvert \xi(t) \rvert}^2$ satisfies by the homogeneity of $p$ $$\dot{h}(t) = 2 \xi(t) \cdot \dot{\xi}(t) = -2 \xi(t) \cdot d_x p(x(t),\xi(t)) = m(t) h(t)$$ where $m(t) = -2 \omega(t) \cdot d_x p(x(t),\omega(t))$ and $\omega(t) = \xi(t)/{\lvert \xi(t) \rvert}$. Solving the equation gives $h(t) = e^{\int_0^t m(s) \,ds} h(0)$. But ${\lvert m(t) \rvert} \leq 2 \sup_{x \in {\mathbf{R}}^n, {\lvert \omega \rvert} = 1} {\lvert d_x p(x,\omega) \rvert}$, and this argument shows that $$e^{-t_0 \sup {\lvert d_x p(x,\omega) \rvert}} {\lvert \xi(0) \rvert} \leq {\lvert \xi(t) \rvert} \leq e^{t_0 \sup {\lvert d_x p(x,\omega) \rvert}} {\lvert \xi(0) \rvert}$$ when $t \in I$. For $x(t)$ we have ${\lvert \dot{x}(t) \rvert} \leq \sup_{x \in {\mathbf{R}}^n, {\lvert \omega \rvert} = 1} {\lvert d_{\xi} p(x,\omega) \rvert}$ for $t \in I$, so ${\lvert x(t) - x(0) \rvert} \leq t_0 n^{1/2} \sup {\lvert d_{\xi} p(x,\omega) \rvert}$ for $t \in I$. Thus $(x(t),\xi(t))$ stays in a compact subset of $T^* {\mathbf{R}}^n$ when $t \in I$, and we may extend the solution past the endpoints of $I$ and to all time.
The following consequence will be used frequently: if $(x(t),\xi(t))$ satisfy the Hamilton equations and ${\lvert \xi(0) \rvert} \sim \lambda$, then $${\lvert \xi(t) \rvert} \sim \lambda$$ for ${\lvert t \rvert} \leq M$.
We write $\chi_{t,s} = \chi_{t,s}^{k,\pm}$ for the canonical transformation $(y,\eta) \mapsto (x(t),\xi(t))$, where $(x,\xi)$ solve the Hamilton equations with initial condition $(x(s),\xi(s)) = (y,\eta)$. Then $\chi_{t,s}$ is a symplectic diffeomorphism of $T^* {\mathbf{R}}^n$ with inverse $\chi_{s,t}$.
Wave packet representation
==========================
Let $g$ be a real, even Schwarz function in ${\mathbf{R}}^n$ with ${\lVert g \rVert}_{L^2} = (2\pi)^{-\frac{n}{2}}$, and assume $\hat{g}$ is supported in the unit ball. For $\lambda \geq 1$ and $y,x,\xi \in {\mathbf{R}}^n$ define $$\label{glambda_def}
g_{\lambda}(y ; x,\xi) = \lambda^{\frac{n}{4}} e^{i\xi \cdot (y-x)} g(\lambda^{\frac{1}{2}}(y-x)).$$ This is a wave packet at frequency level $\lambda$, centered in space at $x$ and in frequency at $\xi$. The Fourier transform is given by $$\label{glambdahat_def}
\hat{g}_{\lambda}(\eta ; x,\xi) = \lambda^{-\frac{n}{4}} e^{-i\eta \cdot x} \hat{g}(\lambda^{-\frac{1}{2}}(\eta-\xi)).$$ The FBI transform of a function $f \in {\mathscr{S}}({\mathbf{R}}^n)$ is given by $$T_{\lambda} f(x,\xi) = \int f(y) \overline{g_{\lambda}(y;x,\xi)} \,dy.$$ Suppose $\lambda \geq 2^6$. Then if $\hat{f}$ is supported in $\frac{1}{4} \lambda < {\lvert \xi \rvert} < \lambda$, then $T_{\lambda} f$ vanishes unless $\frac{1}{8} \lambda < {\lvert \xi \rvert} < 2 \lambda$. If $F \in {\mathscr{S}}({\mathbf{R}}^{2n}_{x,\xi})$, the adjoint has the form $$T_{\lambda}^* F(y) = \iint F(x,\xi) g_{\lambda}(y ; x,\xi) \,dx\,d\xi.$$ It follows that $T_{\lambda}^* T_{\lambda} = I$, and ${\lVert T_{\lambda} f \rVert}_{L^2({\mathbf{R}}^{2n}_{x,\xi})} = {\lVert f \rVert}_{L^2({\mathbf{R}}^n)}$.
The following result, stating the $L^2$ boundedness of FBI transform type operators, is from [@smith2 Lemma 3.1].
\[fbisobolevbounded\] Let $g_{x,\xi}$ be a family of Schwartz functions whose Schwartz seminorms are bounded uniformly in $x$ and $\xi$. Then the operator $T$, defined for Schwartz functions by $$Tf(x,\xi) = \int f(y) \overline{(g_{x,\xi})_{\lambda}(y; x,\xi)} \,dy$$ is bounded from $L^2({\mathbf{R}}^n)$ to $L^2({\mathbf{R}}^{2n})$, and the adjoint $T^*$ given by $$T^* F(y) = \iint F(x,\xi) (g_{x,\xi})_{\lambda}(y;x,\xi) \,dx\,d\xi$$ is bounded from $L^2({\mathbf{R}}^{2n})$ to $L^2({\mathbf{R}}^n)$. The norms of $T$ and $T^*$ are bounded by a sum of finitely many Schwartz seminorms of the $g_{x,\xi}$.
Outline of construction of solution operator {#sec:parametrixoutline}
============================================
We will outline the construction of a solution operator $S(t): g \mapsto u(t,\,\cdot\,)$ for the problem . More details are given in Section \[sec:parametrixdetails\].
Let us start with the standard Littlewood-Paley frequency localization. Let $\chi(\xi)$ be a smooth cutoff supported in the unit ball with $\chi = 1$ for ${\lvert \xi \rvert} \leq 1/2$, and take $\beta_k(D)$ to be a Littlewood-Paley partition of unity with $$\beta_0(\xi) + \sum_{k=1}^{\infty} \beta_k(\xi) = 1$$ where $\beta_0$ is supported in the unit ball, $\beta_1$ is supported in $\{ \frac{1}{2} < {\lvert \xi \rvert} < 2 \}$, and $\beta_k(\xi) = \beta_1(2^{-k+1} \xi)$.
We let $g_k = \beta_k(D)g$, and for each $k$ we want to find an approximate solution of $$\label{localizedwaveeq}
\left\{ \begin{array}{rl}
(D_t^2 - A_k(x,D_x))u_k &\!\!\!= 0, \\[4pt]
u_k(0) = 0,\ \partial_t u_k(0) &\!\!\!= g_k.
\end{array} \right.$$ This will be done by reducing matters to first order hyperbolic equations. Consider the pseudodifferential symbol $$p_k^{\pm}(x,\xi) = \pm \chi(2^{-k/2}D_x) \sqrt{A_k(x,\xi)}$$ so that $p_k^{\pm}$ is positively homogeneous of degree $1$ in $\xi$ and smooth when $\xi \neq 0$, and satisfies for ${\lvert \xi \rvert} = 1$ $${\lvert \partial_x^{\alpha} \partial_{\xi}^{\beta} p_k^{\pm}(x,\xi) \rvert} \lesssim 2^{\frac{k}{2} \max(0,{\lvert \alpha \rvert}-2)}.$$ Thus $p_k^{\pm} \beta_k(\xi) \in S^1_{1,1/2}$. The Fourier transform in $x$, $(p_k^{\pm}){\,\hat{\rule{0pt}{6pt}}\,}(\,\cdot\,,\xi)$, is supported in $\{ {\lvert \eta \rvert} \leq 2^{k/2}\}$. Since $$(p(x,D) f) {\,\hat{\rule{0pt}{6pt}}\,}(\eta) = (2\pi)^{-n} \int \hat{p}(\eta-\xi,\xi) \hat{f}(\xi) \,d\xi,$$ we see that $p_k^{\pm}(x,D) f$ will be supported in ${\lvert \eta \rvert} \sim 2^k$ if $\hat{f}$ is supported in ${\lvert \eta \rvert} \sim 2^k$. We will also use the symbols $$q_k^{\pm} = \chi(2^{-k/2}D_x) (1/p_k^{\pm})$$ which are homogeneous of degree $-1$ in $\xi$, and $q_k^{\pm} \beta_k(\xi) \in S^{-1}_{1,1/2}$.
The approximate solution operators related to are $$E_k^{\pm}(t) g = T_k^* U_k^{\pm}(t) T_k( \frac{i}{2} Q_k^{\pm} \beta_k(D)g)$$ where $T_k = T_{\lambda}$ with $\lambda = 2^k$, and $U_k^{\pm}(t)$ is translation for time $-t$ along the Hamilton flow of $p_k^{\pm}$, that is, $$U_k^{\pm}(t)F = F \circ \chi_{0,t}^{k,\pm}.$$ We are ready to write down the first approximate solution operator for . It will be given by $$\tilde{S}(t) g = t \sum_{k<k_0} g_k + \sum_{k \geq k_0} (u_k^+ + u_k^-)$$ where $k_0$ is a constant depending on $M$ which will be chosen later, and where $$u_k^{\pm} = E_k^{\pm}(t) g.$$
Since $Q_k^{\pm}$ is of order $-1$, it is not hard to see that $\tilde{S}(t)$ is an operator of order $-1$, i.e. it maps $H^{\alpha}({\mathbf{R}}^n)$ to $H^{\alpha+1}({\mathbf{R}}^n)$. The main point is that the operator $$\tilde{R}_k^{\pm}(t) = (D_t + P_k^{\pm}(y,D_y)) E_k^{\pm}(t),$$ which is a half-wave operator applied to $E_k^{\pm}(t)$, is of order $-1$. This implies that when one applies the wave operator $D_t^2 - A(x,D_x)$ to $\tilde{S}(t)$, the resulting operator will be of order $0$, which is one order lower than expected. Therefore we really have an approximate solution operator.
However, since $\partial_t \tilde{S}(t) g|_{t = 0}$ is only close to $g$ but not equal to $g$ in general, we need to correct the initial values of the operator. Thus instead of $\tilde{S}(t)$ we use $$\widehat{S}(t) = \tilde{S}(t)(I+K)^{-1}$$ where $K$ is given by $$K = \sum_{k \geq k_0} (i \tilde{R}_k^+(0) + i \tilde{R}_k^-(0) + \frac{1}{2} R_k^+ \beta_k(D) + \frac{1}{2} R_k^- \beta_k(D))$$ Here $R_k^{\pm} \beta_k(D)$ are order $-1$ operators given by $$P_k^{\pm} Q_k^{\pm} \beta_k(D) = (I + R_k^{\pm}) \beta_k(D).$$ If $k_0$ is chosen large enough then $K$ will be small on $H^{\alpha}({\mathbf{R}}^n)$, and $\widehat{S}(t)$ will be an operator of order $-1$ with $\widehat{S}(t)g |_{t=0} = 0$ and $\partial_t \widehat{S}(t)g |_{t=0} = g$. Thus $\widehat{S}(t)g$ will be an approximate solution of with correct Cauchy data.
It remains to show that one may iterate away the error and obtain an exact solution. To do this we seek a solution of of the form $$u(t,x) = \widehat{S}(t) g(x) + \int_0^t \widehat{S}(t,s) G(s,x) \,ds.$$ Here $\widehat{S}(t,s) = \widehat{S}(t-s)$ is the operator corresponding to $\widehat{S}(t)$ but where the initial surface is $\{t = s\}$ instead of $\{t = 0\}$. Then $\widehat{S}(t,s) g|_{t=s} = 0$ and $\partial_t \widehat{S}(t,s) g |_{t=s} = g$, and one has $$\partial_t^2 \Big( \int_0^t \widehat{S}(t,s) G(s,x) \,ds \Big) = G(t,x) + \int_0^t \partial_t^2 \widehat{S}(t,s) G(s,x) \,ds.$$ We obtain $$(D_t^2 - A(x,D_x)) u(t,x) = T(t)g(x) - G(t,x) + \int_0^t T(t,s) G(s,x) \,ds$$ where $T(t,s) = (D_t^2 - A(x,D_x)) \widehat{S}(t,s)$. As remarked above, we will show that $T(t,s)$ is an operator of order $0$, which in the present setting with a $C^{1,1}$ wave operator will mean that it is bounded on $H^{\alpha}({\mathbf{R}}^n)$ for $-1 \leq \alpha \leq 2$. Then $u$ will be a solution provided that $G(t,x) = V(T(t)g(x))$, where $G = VF$ solves the Volterra equation $$G(t,x) - \int_0^t T(t,s) G(s,x) \,ds = F(t,x).$$ Since $T(t,s)$ is bounded on $H^{\alpha}$ also $V$ is bounded on $L^{\infty}_t H^{\alpha}_x$, for $-1 \leq \alpha \leq 2$. Thus the full solution operator for will be $$S(t)g(x) = \widehat{S}(t)g(x) + \int_0^t \widehat{S}(t,s) V(T(s) g(x)) \,ds.$$ We will now give the details.
Details of construction of solution operator {#sec:parametrixdetails}
============================================
Let $(a^{ij})$ be a symmetric matrix of $C^{1,1}$ functions satisfying . We take $M$ large enough so that holds also for the truncated metrics $(a^{ij}_k)$, and we also assume . We will use all the notations in Section \[sec:parametrixoutline\]. Also, $k_0$ will be a sufficiently large integer, depending on $M$ and $n$.
We start by noting that if the Cauchy data is localized near frequency $2^k$, then the operators constructed in the preceding section preserve this localization.
Suppose the Fourier transform of $f$ is supported in ${\lvert \xi \rvert} \sim 2^k$. Then $T_k f$, $P_k^{\pm} f$, $Q_k^{\pm} f$ vanish unless ${\lvert \xi \rvert} \sim 2^k$, provided $k \geq k_0$. If $F(x,\xi)$ vanishes unless ${\lvert \xi \rvert} \sim 2^k$, then $T_k^* F$ and $U_k^{\pm}(t) F$ vanish unless ${\lvert \xi \rvert} \sim 2^k$, provided $k \geq k_0$.
Using this lemma, it will be enough to consider a fixed frequency and the result will follow by summing over dyadic annuli. Thus, let $\lambda = 2^k$, and write $p = p_k^{\pm}$, $U(t) = U_k^{\pm}(t)$, etc. The main idea is that the wave evolution of a wave packet at frequency $\lambda$ can be well approximated by transport along the Hamilton flow. The transport operator will be $L_{x,\xi} = L_k^{\pm}$, given by $$L_{x,\xi} = d_{\xi} p(x,\xi) \cdot D_x - d_x p(x,\xi) \cdot D_{\xi}.$$ It is easy to see that $U(t)F$ will solve the corresponding transport equation.
If $F \in {\mathscr{S}}({\mathbf{R}}^n_x \times {\mathbf{R}}^n_{\xi})$ then $U(t)F$ satisfies $$\left\{ \begin{array}{rl}
(D_t + L_{x,\xi})U(t)F &\!\!\!= 0, \\[4pt]
U(t)F|_{t=0} &\!\!\!= F
\end{array} \right.$$
Let now $f$ be a function localized near frequency $\lambda$. We write $f$ as a superposition of wave packets at frequency $\lambda$, $$f(y) = T_{\lambda}^* T_{\lambda} f(y) = \iint T_{\lambda} f(x,\xi) g_{\lambda}(y;x,\xi) \,dx\,d\xi.$$ To get an approximate solution $u$ to the half-wave equation $(D_t + P_y) u(t,y) = 0$ with $u(0,y) = f(y)$, where $P_y = p(y,D_y)$, we transport the wave packets for time $t$ along the Hamilton flow of $p$. Then $u$ is given by $$u(t,y) = \iint T_{\lambda} f(x,\xi) g_{\lambda}(y;\chi_{t,0}(x,\xi)) \,dx\,d\xi.$$ Using that $\chi_{t,0}$ is a symplectic map with inverse $\chi_{0,t}$, $u$ will be equal to $$u(t,y) = T_{\lambda}^* U(t) T_{\lambda} f.$$ To measure how far $u$ is from an exact solution, we compute $$\begin{gathered}
D_t u = T_{\lambda}^* D_t U(t) T_{\lambda} f = -T_{\lambda}^* L_{x,\xi} U(t) T_{\lambda} f \\
= \iint (U(t) T_{\lambda} f)(x,\xi) L_{x,\xi} g_{\lambda}(y;x,\xi) \,dx\,d\xi,\end{gathered}$$ the last equality by integration by parts. Thus $$\label{halfwaveu}
(D_t + P_y) u = \iint (U(t) T_{\lambda} f)(x,\xi) (P_y + L_{x,\xi}) g_{\lambda}(y;x,\xi) \,dx\,d\xi.$$ The following lemma, corresponding to [@smith2 Lemma 3.2], will be crucial.
\[wavepacketparameter\] Suppose ${\lvert \xi \rvert} \sim \lambda$. Then $$(P_y + L_{x,\xi}) g_{\lambda}(y ; x,\xi) = (g_{x,\xi})_{\lambda}(y ; x,\xi)$$ where $g_{x,\xi}$ is a family of functions whose Fourier transform is supported in a ball of radius $2$ and the Schwartz seminorms of $g_{x,\xi}$ are uniformly bounded in $x,\xi$. In fact $g_{x,\xi}(z) = m_{x,\xi}(z,D_z) g(z)$ where $$\label{mxxidef}
m_{x,\xi}(z,\zeta) = \int_0^1 (1-s) \partial_s^2 ( p(x + s \lambda^{-1/2} z, \xi + s \lambda^{1/2} \zeta) ) \,ds.$$ Here $m_{x,\xi}$ satisfies symbol estimates uniform in $x$ and $\xi$, $$\label{mxxiest}
{\lvert \partial_z^{\alpha} \partial_{\zeta}^{\beta} m_{x,\xi}(z,\zeta) \rvert} \lesssim {\langle z \rangle}^2, \qquad {\lvert \zeta \rvert} \leq 2.$$
We define $g_{x,\xi}$ by the relation $(g_{x,\xi})_{\lambda} = (P_y + L_{x,\xi}) g_{\lambda}$. Recalling the formula for $g_{\lambda}$, it follows that $$\begin{gathered}
(g_{x,\xi})_{\lambda} = (P_y - id_{\xi} p \cdot d_x + i d_x p \cdot d_{\xi}) g_{\lambda} = (P_y + id_{\xi} p \cdot d_y - d_x p \cdot (y-x)) g_{\lambda} \\
= (2\pi)^{-n} \int e^{iy \cdot \eta} (p(y,\eta) - d_{\xi} p(x,\xi) \cdot \eta - d_x p(x,\xi) \cdot (y-x)) (g_{\lambda}){\,\hat{\rule{0pt}{6pt}}\,}(\eta) \,d\eta \\
= (2\pi)^{-n} \int e^{iy \cdot \eta} \Big[ p(y,\eta) - p(x,\xi) - d_{\xi} p(x,\xi) \cdot (\eta-\xi) - d_x p(x,\xi) \cdot (y-x) \Big] (g_{\lambda}){\,\hat{\rule{0pt}{6pt}}\,}(\eta) \,d\eta\end{gathered}$$ since $\xi \cdot d_{\xi} p = p$ by homogeneity. By , we have $$\begin{gathered}
g_{x,\xi}(z) = (2\pi)^{-n} \int e^{iz \cdot \zeta} \Big[ p(x+\lambda^{-1/2}z,\xi+\lambda^{1/2}\zeta) - p(x,\xi) - d_x p(x,\xi) \cdot \lambda^{-1/2} z \\
- d_{\xi} p(x,\xi) \cdot \lambda^{1/2} \zeta \Big] \hat{g}(\zeta) \,d\zeta\end{gathered}$$ which shows that $(g_{x,\xi}) {\,\hat{\rule{0pt}{6pt}}\,}= 0$ outside a ball of radius $2$. From Taylor’s formula we see that the term in brackets is equal to $m_{x,\xi}(z,\zeta)$ given by . If $(\tilde{x}, \tilde{\xi}) = (x+s\lambda^{-1/2}z, \xi+s\lambda^{1/2}\zeta)$, we compute $$\label{mxxi_explicit}
\partial_s^2( p(\tilde{x}, \tilde{\xi}) ) = \sum_{j,k}[ \partial_{x_j x_k} p(\tilde{x}, \tilde{\xi}) \lambda^{-1} z_j z_k + \\
\partial_{x_j \xi_k} p(\tilde{x}, \tilde{\xi}) z_j \zeta_k + \partial_{\xi_j \xi_k} p(\tilde{x}, \tilde{\xi}) \lambda \zeta_j \zeta_k ]$$ which gives .
Now we can prove that the half-wave operator applied to the approximate solution operator gives an operator of order $0$.
\[halfwaveapproximate\] Suppose $\hat{f}$ is supported in ${\lvert \xi \rvert} \sim \lambda$. Then $${\lVert (D_t + P_y) T_{\lambda}^* U(t) T_{\lambda} f \rVert}_{L^2} \lesssim {\lVert f \rVert}_{L^2}.$$
If $\hat{f}$ is supported in ${\lvert \xi \rvert} \sim \lambda$, the same holds for $U(t) T_{\lambda} f$. The result follows from , Lemma \[wavepacketparameter\] and Lemma \[fbisobolevbounded\].
Next we apply a second order wave operator $D_t^2 - P_y^2$ to the approximate solution operator, and show that there is a loss of one derivative.
\[psdofullwaveapproximate\] Suppose $\hat{f}$ is supported in ${\lvert \xi \rvert} \sim \lambda$. Then $${\lVert (D_t^2 - P_y^2) T_{\lambda}^* U(t) T_{\lambda} f \rVert}_{L^2} \lesssim \lambda {\lVert f \rVert}_{L^2}.$$
We write $u = T_{\lambda}^* U(t) T_{\lambda} f$ and $$(D_t^2 - P_y^2) u = (D_t + P_y)^2 u - 2 P_y (D_t + P_y) u.$$ Since $u$ and $(D_t + P_y) u$ are supported in ${\lvert \xi \rvert} \sim \lambda$ and $p$ is of order $1$, we have $${\lVert P_y (D_t + P_y) u \rVert}_{L^2} \lesssim \lambda {\lVert f \rVert}_{L^2}$$ by Lemma \[halfwaveapproximate\].
Using and writing $(P_y + L_{x,\xi}) g_{\lambda} = (g_{x,\xi})_{\lambda}$, we get $$(D_t + P_y)^2 u = \iint (U(t) T_{\lambda} f)(x,\xi) (P_y + L_{x,\xi}) (g_{x,\xi})_{\lambda}(y;x,\xi) \,dx\,d\xi.$$ Since $$\begin{aligned}
D_x (g_{x,\xi})_{\lambda} &= -D_y (g_{x,\xi})_{\lambda} + (D_x g_{x,\xi})_{\lambda}, \\
D_{\xi} (g_{x,\xi})_{\lambda} &= (y-x) (g_{x,\xi})_{\lambda} + (D_{\xi} g_{x,\xi})_{\lambda},\end{aligned}$$ the argument in Lemma \[wavepacketparameter\] gives $$(P_y + L_{x,\xi}) (g_{x,\xi})_{\lambda} = (m_{x,\xi}(z,D_z) g_{x,\xi})_{\lambda} + (L_{x,\xi} m_{x,\xi}(z,D_z) g)_{\lambda}.$$ The Schwartz seminorms of $g_{x,\xi}$ are uniformly bounded in $x,\xi$, and shows that the same applies to the Schwartz seminorms of $m_{x,\xi}(z,D_z) g_{x,\xi}$. Lemma \[fbisobolevbounded\] gives $${\lVert \iint (U(t) T_{\lambda} f)(x,\xi) (m_{x,\xi}(z,D_z) g_{x,\xi})_{\lambda}(y;x,\xi) \,dx\,d\xi \rVert}_{L^2({\mathbf{R}}^n_y)} \lesssim {\lVert f \rVert}_{L^2}$$ It remains to study the symbol seminorms of $\tilde{m}_{x,\xi}(z,\zeta) = L_{x,\xi} m_{x,\xi}(z,\zeta)$. From we obtain $$\begin{gathered}
\tilde{m}_{x,\xi}(z,\zeta) = \int_0^1 (1-s) \partial_s^2 \Big[ d_{\xi} p(x,\xi) \cdot D_x p(x+s\lambda^{-1/2} z,\xi+s\lambda^{1/2} \zeta) \\
- d_x p(x,\xi) \cdot D_{\xi} p(x+s\lambda^{-1/2} z,\xi+s\lambda^{1/2} \zeta) \Big] \,ds. \label{mxxitilde_def}\end{gathered}$$ Since ${\lvert \xi \rvert} \sim \lambda$ a computation as in shows $$\label{mxxitildeest}
{\lvert \partial_z^{\alpha} \partial_{\zeta}^{\beta} \tilde{m}_{x,\xi}(z,\zeta) \rvert} \lesssim \lambda^{1/2} {\langle z \rangle}^2, \qquad {\lvert \zeta \rvert} \leq 2,$$ and Lemma \[fbisobolevbounded\] implies $${\lVert \iint (U(t) T_{\lambda} f)(x,\xi) (\tilde{m}_{x,\xi}(z,D_z) g)_{\lambda}(y;x,\xi) \,dx\,d\xi \rVert}_{L^2({\mathbf{R}}^n_y)} \lesssim \lambda^{1/2} {\lVert f \rVert}_{L^2}.$$ This concludes the proof.
We move on to the full parametrix, where all the frequencies are added up. The errors resulting from truncation are handled as in [@smith1 Theorem 4.5].
\[stildeestimates\] $\tilde{S}(t)$ is a bounded operator $H^{\alpha} \to H^{\alpha+1}$, for all $\alpha \in {\mathbf{R}}$. If $-1 \leq \alpha \leq 2$ then $(D_t^2 - A(x,D_x)) \tilde{S}(t)$ is a bounded operator $H^{\alpha} \to H^{\alpha}$.
Since $q_k^{\pm}$ is order $-1$ one has $E_k^{\pm}(t): H^{\alpha} \to H^{\alpha+1}$, and since $E_k^{\pm}(t) g$ is localized near ${\lvert \xi \rvert} \sim 2^k$ the sum converges in $H^{\alpha+1}$ and $\tilde{S}(t): H^{\alpha} \to H^{\alpha+1}$.
Let now $-1 \leq \alpha \leq 2$. We write $$\begin{gathered}
(D_t^2 - A(x,D_x)) \tilde{S}(t) g = -t \sum_{k < k_0} A(x,D_x) \beta_k(D) g \\
+ \sum_{k \geq k_0} (D_t^2 - A_k(x,D_x))(u_k^+ + u_k^-) + \sum_{k \geq k_0} (A_k(x,D_x) - A(x,D_x))(u_k^+ + u_k^-)\end{gathered}$$ and write the last expression as $\tilde{T}_1(t)g + \tilde{T}_2(t)g + \tilde{T}_3(t)g$. It is clear that ${\lVert \tilde{T}_1(t) g \rVert}_{H^{\alpha}} \lesssim {\lVert g \rVert}_{H^{\alpha}}$. For $\tilde{T}_2(t)$ we write $$D_t^2 - A_k(x,D_x) = D_t^2 - (P_k^{\pm})^2 + (P_k^{\pm})^2 - A_k(x,D_x).$$ One has $${\lVert \sum_{k \geq k_0} (D_t^2 - (P_k^+)^2) u_k^+ \rVert}_{H^{\alpha}} \lesssim {\lVert g \rVert}_{H^{\alpha}}$$ by Lemma \[psdofullwaveapproximate\]. From we obtain $$\label{pkak_estimate}
(\sqrt{A_k} - p_k^+)\beta_k(\xi) \in S^0_{1,1/2},$$ which shows that $(P_k^+)^2 - A_k(x,D_x)$ is of order $1$ and $${\lVert \sum_{k \geq k_0} ((P_k^+)^2 - A_k(x,D_x)) u_k^+ \rVert}_{H^{\alpha}} \lesssim {\lVert g \rVert}_{H^{\alpha}}.$$ Similar results hold for $u_k^-$.
For $\tilde{T}_3(t)$ we need to show that $$\Gamma: g \mapsto \sum_{k \geq k_0} (a_k(x) - a(x)) D_{x_i} D_{x_j} u_k^+$$ is bounded on $H^{\alpha}$, where $a_k = a^{ij}_k$ and $a = a^{ij}$. We will write $\beta_j(D) \Gamma = \sum_{k \geq k_0} \Gamma_{jk} D_{x_i x_j} u_k^+$ where $$\Gamma_{jk} = \tilde{\beta}_j(D) (a(x)-a_k(x)) \tilde{\beta}_k(D).$$ Here $\tilde{\beta}_k(\xi)$ are cutoffs to ${\lvert \xi \rvert} \sim 2^k$ which are $1$ on the frequency support of $u_k^+$. We let $l_0 = l_0(M)$ be an integer so that ${\mathrm{supp}}(\tilde{\beta}_k(\xi)) \subseteq \{2^{-l_0+k} \leq {\lvert \xi \rvert} \leq 2^{l_0+k} \}$.
By looking at the supports on the Fourier side, we get $$\Gamma_{jk} = \left\{ \begin{array}{ll}
\beta_j(D) (a-a_{2k-4l_0}) \tilde{\beta}_k(D), & j \leq k - 3l_0, \\
\beta_j(D) (a-a_{k}) \tilde{\beta}_k(D), & k - 3l_0 \leq j \leq k + 3l_0, \\
\beta_j(D) (a-a_{2j-4l_0}) \tilde{\beta}_k(D), & j \geq k + 3l_0.
\end{array} \right.$$ Using that ${\lVert a-a_k \rVert}_{L^{\infty}} \lesssim 2^{-k}$ we obtain $$\begin{gathered}
{\lVert \beta_j(D) \Gamma g \rVert}_{L^2}^2 \lesssim \sum_{k < j-3l_0} 2^{-4j} 2^{2k} {\lVert g_k \rVert}_{L^2}^2 + \sum_{{\lvert j-k \rvert} \leq 3l_0} 2^{-2j}2^{2k} {\lVert g_k \rVert}_{L^2}^2 \\
+ \sum_{k > j+3l_0} 2^{-4k} 2^{2k} {\lVert g_k \rVert}_{L^2}^2.\end{gathered}$$ It follows that $\Gamma$ is bounded on $H^{\alpha}$ with ${\lvert \alpha \rvert} \leq 1$. We compute $$D_{x_l} \Gamma g = \sum_{k \geq k_0} (D_l a_k) D_{x_i x_j} u_k^+ - D_l a \sum_{k \geq k_0} D_{x_i x_j} u_k^+ + \sum_{k \geq k_0} (a_k-a) D_{x_i x_j x_l} u_k^+$$ and note that this is bounded $H^{\alpha+1} \to H^{\alpha}$ if ${\lvert \alpha \rvert} \leq 1$, due to the $C^{1,1}$ regularity of $a$ and the argument above. This shows boundedness of $\Gamma$ for $-1 \leq \alpha \leq 2$.
In the next lemma we correct the value of $\partial_t \tilde{S}(t)$ at $t = 0$.
If $k_0$ is sufficiently large, then $$K = \sum_{k \geq k_0} (i \tilde{R}_k^+(0) + i \tilde{R}_k^-(0) + \frac{1}{2} R_k^+ \beta_k(D) + \frac{1}{2} R_k^- \beta_k(D))$$ has norm $\leq 1/2$ on $H^{\alpha}$. The operator $$\widehat{S}(t) = \tilde{S}(t) (I+K)^{-1}$$ will satisfy $\widehat{S}(t) g|_{t=0} = 0$ and $\partial_t \widehat{S}(t) g|_{t=0} = g$.
One has ${\lVert \tilde{R}_k^{\pm}(0) g \rVert}_{L^2} \lesssim 2^{-k} {\lVert g \rVert}_{L^2}$ by Lemma \[halfwaveapproximate\]. Also, $R_k^{\pm} \beta_k(D)$ is of order $-1$ by looking at the symbol expansion of $(P_k^{\pm} Q_k^{\pm} - I) \beta_k(D)$ and using . The norms do not depend on any previous value of $k_0$, and we may choose $k_0$ so large that $K$ will have norm $\leq 1/2$ on $H^{\alpha}$ for $\alpha \in [-\alpha_0,\alpha_0]$, for any $\alpha_0 > 0$. Thus $I+K$ will be invertible on these spaces and the norm of the inverse will be $\leq 2$.
Since $\tilde{S}(0)g = 0$, also $\widehat{S}(0)g = 0$. The derivative is $$\partial_t \tilde{S}(t) g = \sum_{k < k_0} g_k + i \sum_{k \geq k_0} (D_t u_k^+ + D_t u_k^-).$$ We write $D_t u_k^{\pm} = \tilde{R}_k^{\pm}(t) g - P_k^{\pm} u_k^{\pm}$. Since $u_k^{\pm}(0) = \frac{i}{2} Q_k^{\pm} g_k$, we get $$\partial_t \tilde{S}(t) g|_{t=0} = \sum_{k < k_0} g_k + \sum_{k \geq k_0} (i \tilde{R}_k^+(0) g + i \tilde{R}_k^-(0) g + \frac{1}{2} P_k^{+} Q_k^{+} g_k + \frac{1}{2} P_k^{-} Q_k^{-} g_k).$$ It follows that $\partial_t \tilde{S}(t) g|_{t=0} = (I+K)g$.
It now remains to show that one obtains a full solution operator from $\widehat{S}(t)$ by introducing a correction by solving a Volterra equation. The solution operator for the Volterra equation is the following.
\[lemma:volterra\] Suppose $T(t,s)$ is bounded on $H^{\alpha}$, with norm uniformly bounded by $C$ when $t,s \in [-M,M]$. There is a bounded map $V$ on $L^{\infty}_t H^{\alpha}_x$, with norm bounded by $e^{CM}$, such that for any $F(t,x) \in L^{\infty}_t H^{\alpha}_x$, $G = VF$ solves the equation $$G(t,x) - \int_0^t T(t,s)G(s,x) \,ds = F(t,x).$$
Define $(VF)(t,x)$ by $$F(t,x) + \sum_{j=1}^{\infty} \int_0^t \!\! \int_0^{s_1} \!\!\!\!\! \cdots \! \int_0^{s_{j-1}} \! T(t,s_1) T(s_1,s_2) \cdots T(s_{j-1},s_j) F(s_j,x) \,ds_j \cdots \,ds_1.$$ It is easy to check that the series converges in $L^{\infty}_t H^{\alpha}_x$ to a solution which satisfies the desired norm estimate.
As in the outline, we write $\widehat{S}(t,s) = \widehat{S}(t-s)$ as the operator corresponding to $\widehat{S}(t)$ but with initial surface $\{t = s\}$. If $s,t \in [-M,M]$ then $\widehat{S}(t,s)$ has the same properties as $\widehat{S}(t)$, except that $\widehat{S}(t,s) g|_{t=s} = 0$ and $\partial_t \widehat{S}(t,s)g |_{t=s} = g$. We let $$T(t,s) = (D_t^2 - A(x,D_x)) \widehat{S}(t,s), \qquad T(t) = T(t,0),$$ so $T(t,s)$ and $T(t)$ are bounded on $H^{\alpha}$ when $-1 \leq \alpha \leq 2$ by Lemma \[stildeestimates\]. We let $V$ be the solution operator for the Volterra equation corresponding to $T(t,s)$.
The argument in the outline shows that if $$S(t)g(x) = \widehat{S}(t)g(x) + \int_0^t \widehat{S}(t,s) V(T(s) g(x)) \,ds,$$ then $(D_t^2 - A(x,D_x))S(t)g = 0$ and $S(t)g|_{t=0} = 0$, $\partial_t S(t) g|_{t=0} = g$. One only needs to check that the time derivatives are justified, but this may be done as in [@smith1] and will not be needed for stability considerations. This ends the construction of the solution operator.
Stability {#sec:stability}
=========
We now proceed to prove the stability part of the result. Let $A(x) = (a^{ij}(x))$ and $B(x) = (b^{ij}(x))$ be two symmetric matrices satisfying , and take $M$ so large that the truncated metrics $(a^{ij}_k)$ and $(b^{ij}_k)$ also satisfy . Also assume that $t$ satisfies .
We write $P_A = P_k^{\pm}$ for the operator at frequency $\lambda = 2^k$ defined in terms of the metric $A$, and similarly $Q_A, R_A, \tilde{R}_A(t)$ etc. The following operators depend norm continuously on the metric.
\[psdostability\] If $\hat{f}$ is supported in ${\lvert \xi \rvert} \sim \lambda$, then $$\begin{aligned}
{\lVert (P_A - P_B)f \rVert}_{L^2} &\lesssim \lambda {\lVert A-B \rVert}_{L^{\infty}} {\lVert f \rVert}_{L^2}, \\
{\lVert (Q_A - Q_B)f \rVert}_{L^2} &\lesssim \lambda^{-1} {\lVert A-B \rVert}_{L^{\infty}} {\lVert f \rVert}_{L^2}, \\
{\lVert (R_A - R_B)f \rVert}_{L^2} &\lesssim \lambda^{-1} {\lVert A-B \rVert}_{L^{\infty}} {\lVert f \rVert}_{L^2}.\end{aligned}$$
Consider $h(x,\xi) = (\lambda^{n/2} \chi(\lambda^{1/2} \,\cdot\,) \ast [F(A_k(\,\cdot\,,\xi)) - F(B_k(\,\cdot\,,\xi))]) \tilde{\beta}_{\lambda}(\xi)$, where $\tilde{\beta}_{\lambda}$ is a cutoff to ${\lvert \xi \rvert} \sim \lambda$ and $F(t) = t^{1/2}$. We wish to show that $$\label{papb_est}
{\lvert \partial_x^{\alpha} \partial_{\xi}^{\beta} h(x,\xi) \rvert} \leq C_{M,\alpha,\beta} {\lVert A-B \rVert}_{L^{\infty}} \lambda^{1-{\lvert \beta \rvert} + \frac{1}{2} {\lvert \alpha \rvert}}.$$ This will show the estimate for $P_A - P_B$, and the estimate for $Q_A - Q_B$ follows from a similar result with $F(t) = t^{-1/2}$.
In $\partial_x^{\alpha} \partial_{\xi}^{\beta} h$ we let the $x$-derivatives hit the mollifier, which gives the desired growth. We may thus assume that $\alpha = 0$. Each $\xi$-derivative hitting $\tilde{\beta}_{\lambda}(\xi)$ gives $\lambda^{-1}$, so we only need to look at the case when the $\xi$-derivatives hit $F(A_k) - F(B_k)$. We write $$F(A_k) - F(B_k) = \int_0^1 F'(r A_k + (1-r) B_k) (A_k - B_k) \,dr.$$ The matrix $r A_k + (1-r) B_k$ satisfies . Consequently $${\lvert \partial_{\xi}^{\beta} [ (F(A_k) - F(B_k)) \tilde{\beta}_k(\xi) ] \rvert} \leq C_{M,\beta} {\lVert A-B \rVert}_{L^{\infty}} \lambda^{1-{\lvert \beta \rvert}}$$ as desired.
If $P, Q$ are pseudodifferential operators the symbol of $PQ$ is $$\sigma(PQ) = \sum_{{\lvert \alpha \rvert} < N} \frac{\partial_{\xi}^{\alpha} p D_x^{\alpha} q}{\alpha!} + \sum_{{\lvert \alpha+\beta \rvert}=2N} \int e^{-iy \cdot \eta} s_{\alpha \beta} (x,\xi,y,\eta) \,dy\,d\eta$$ where the last terms are oscillatory integrals, and $$s_{\alpha \beta} = \sum_{\gamma \leq \alpha,\beta} c_{\alpha \beta \gamma} \int_0^1 t^{{\lvert \alpha+\beta-2\gamma \rvert}} (1-t)^{2N-1} \partial_{\xi}^{\alpha+\beta-\gamma} p(x,\xi+t\eta) \partial_x^{\alpha+\beta-\gamma} q(x+ty,\xi) \,dt.$$ Note that ${\lvert \alpha+\beta-\gamma \rvert} \geq N$. We have $$(R_A - R_B) \beta_{\lambda}(D) = P_A \tilde{\beta}_{\lambda}(D) Q_A \beta_{\lambda}(D) - P_B \tilde{\beta}_{\lambda}(D) Q_B \beta_{\lambda}(D)$$ for a suitable $\tilde{\beta}_{\lambda}$. Suppressing the cutoffs, this has the symbol $$\begin{gathered}
\sigma(R_A - R_B) = p_A q_A - p_B q_B + \sum_{0 < {\lvert \alpha \rvert} < N} \frac{\partial_{\xi}^{\alpha} p_A D_x^{\alpha} q_A - \partial_{\xi}^{\alpha} p_B D_x^{\alpha} q_B}{\alpha!} \\
+ \sum_{{\lvert \alpha+\beta \rvert}=2N} \int e^{-iy \cdot \eta} (s^A_{\alpha \beta} (x,\xi,y,\eta) - s^B_{\alpha \beta} (x,\xi,y,\eta)) \,dy\,d\eta\end{gathered}$$ The principal symbol is $p_A q_A - p_B q_B = (1-\chi(2^{-k/2} D_x))(1/p_B - 1/p_A)$ since $q = \chi(2^{-k/2} D_x) (1/p)$, and the arguments above show that the corresponding operator has norm $\lesssim \lambda^{-1} {\lVert A-B \rVert}_{L^{\infty}}$ on $L^2$. It is easy to see that the terms with $0 < {\lvert \alpha \rvert} < N$ have the same bound. Finally, if ${\lvert \gamma \rvert} + {\lvert \delta \rvert} \leq 2n+1$ then $${\lvert \partial_y^{\gamma} \partial_{\eta}^{\delta} \partial_x^{\alpha'} \partial_{\xi}^{\beta'} (s^A_{\alpha \beta} - s^B_{\alpha \beta})(x,\xi,y,\eta) \rvert} \leq C \lambda^{-\frac{N}{2}+n+1-{\lvert \beta' \rvert}-\frac{1}{2}{\lvert \alpha' \rvert}} {\lVert A-B \rVert}_{L^{\infty}}.$$ Taking $N$ large enough and using standard estimates for oscillatory integrals gives the $L^2$ bound for $R_A - R_B$.
Let $g_{x,\xi}^A$ be the Schwartz functions in Lemma \[wavepacketparameter\], defined in terms of the metric $A$.
\[wavepacketparameterstability\] The Schwartz norms of $g_{x,\xi}^A - g_{x,\xi}^B$ are $\lesssim \lambda {\lVert A-B \rVert}_{L^{\infty}}$, uniformly in $x$ and $\xi$.
It is enough to show that $$\label{mxxidiff_est}
{\lvert \partial_z^{\alpha} \partial_{\zeta}^{\beta} (m_{x,\xi}^A - m_{x,\xi}^B)(z,\zeta) \rvert} \lesssim \lambda {\lVert A-B \rVert}_{L^{\infty}} {\langle z \rangle}^2, \quad {\lvert \zeta \rvert} \leq 2.$$ This follows from the expression for $m_{x,\xi}$, the computation , and the estimate .
It is also true that the Schwartz seminorms of $g_{x,\xi}^A - g_{x,\xi}^B$ are $\lesssim {\lVert A-B \rVert}_{C^{1,1}}$, which follows by replacing with the alternate estimate $${\lvert \partial_x^{\alpha} \partial_{\xi}^{\beta} h(x,\xi) \rvert} \leq C_{M,\alpha,\beta} {\lVert A-B \rVert}_{C^{1,1}} \lambda^{1-{\lvert \beta \rvert} + \frac{1}{2} \max(0,{\lvert \alpha \rvert}-2)}.$$ However, due to translation along Hamilton flow, there is a loss of one derivative in the stability estimate in any case. Therefore we can afford to lose one derivative in other estimates as well. This results in stability in terms of the $C^{0,1}$ norm of the metric instead of $C^{1,1}$.
${\lVert (\tilde{R}_A(0) - \tilde{R}_B(0))f \rVert}_{L^2} \lesssim {\lVert A-B \rVert}_{L^{\infty}} {\lVert f \rVert}_{L^2}$.
One has $$\begin{gathered}
(\tilde{R}_A(0) - \tilde{R}_B(0))f =\! \frac{i}{2} \Big( \iint (T_{\lambda}(Q_A-Q_B)\beta_{\lambda}(D)f)(x,\xi) (g_{x,\xi}^A)_{\lambda}(y;x,\xi) \,dx\,d\xi \\
+ \iint (T_{\lambda} Q_B \beta_{\lambda}(D)f)(x,\xi) (g_{x,\xi}^A-g_{x,\xi}^B)_{\lambda}(y;x,\xi) \,dx\,d\xi \Big).\end{gathered}$$ The result follows from Lemmas \[fbisobolevbounded\], \[wavepacketparameter\], \[psdostability\] and \[wavepacketparameterstability\].
From the preceding results, and from the factorization $$(I-K_A)^{-1} - (I-K_B)^{-1} = (I-K_A)^{-1}(K_B-K_A)(I-K_B)^{-1},$$ we see that $(I+K)^{-1}$, the operator which corrects the initial values, depends norm continuously on the metric: $$\label{boundarycorrectionstability}
{\lVert (I+K_A)^{-1} - (I+K_B)^{-1} \rVert}_{H^{\alpha} \to H^{\alpha}} \lesssim {\lVert A-B \rVert}_{L^{\infty}}.$$
Next we consider the stability of the flow operator. Here we will lose one derivative to get Lipschitz stability.
\[flowstability\] If $\hat{f}$ vanishes unless ${\lvert \xi \rvert} \sim \lambda$, then $${\lVert T_{\lambda}^* (U_A - U_B) T_{\lambda} f \rVert}_{L^2} \lesssim \lambda {\lVert A-B \rVert}_{C^{0,1}} {\lVert f \rVert}_{L^2}.$$
Fix $t$ and write $\chi_A = \chi_{0,t}^A$, $\chi_B = \chi_{0,t}^B$. One has $$T_{\lambda}^* (U_A - U_B) T_{\lambda} f(y) = \iint [T_{\lambda} f(\chi_A(x,\xi)) - T_{\lambda} f(\chi_B(x,\xi))] g_{\lambda}(y;x,\xi) \,dx\,d\xi.$$ We let $C_r(x,\xi) = r A_{\lambda}(x,\xi) + (1-r) B_{\lambda}(x,\xi)$, and let $\Phi_r = \chi_{C_r} = (x_r,\xi_r)$ be the flow corresponding to the metric $C_r$. Then $(\Phi_r)_{r \in [0,1]}$ is a smooth family of symplectic diffeomorphisms of $T^* {\mathbf{R}}^n$, and we have $$T_{\lambda} f(\chi_A(x,\xi)) - T_{\lambda} f(\chi_B(x,\xi)) = \int_0^1 (d_{x,\xi} T_{\lambda} f)(\Phi_r(x,\xi)) \cdot \partial_r \Phi_r(x,\xi) \,dr.$$ Let $h(s) = h(s,r,x,\xi) = (x_r(s), \xi_r(s)/\lambda)$ where $(x_r(0),\xi_r(0)/\lambda) = (x,\xi/\lambda)$. Differentiating the Hamilton equations for $(x_r,\xi_r/\lambda)$ with respect to $r$, and using $$\partial_r p_{C_r} = \chi(\lambda^{-1/2} D_x) \frac{A_{\lambda}(x,\xi)-B_{\lambda}(x,\xi)}{2\sqrt{C_r(x,\xi)}},$$ we obtain $${\lvert (\partial_r h){\,\dot{\rule{0pt}{6pt}}\,}(s) \rvert} \lesssim {\lvert \partial_r h(s) \rvert} + {\lVert A-B \rVert}_{C^{0,1}}.$$ Since $\partial_r h(0) = (0,0)$, Gronwall’s inequality shows ${\lvert \partial_r h(s) \rvert} \lesssim {\lVert A-B \rVert}_{C^{0,1}}$ for ${\lvert s \rvert} \leq M$. This implies $$\begin{aligned}
{\lvert \partial_r x_r \rvert} &\lesssim {\lVert A-B \rVert}_{C^{0,1}}, \\
{\lvert \partial_r \xi_r \rvert} &\lesssim \lambda {\lVert A-B \rVert}_{C^{0,1}}.\end{aligned}$$ Using that $\partial_{x_j} T_{\lambda} f = T_{\lambda}(\partial_j f)$ and $\partial_{\xi_j} T_{\lambda} f = \lambda^{-1/2} \tilde{T}_{j,\lambda} f$ where $\tilde{T}_{j,\lambda} f(x,\xi) = (f, (iz_j g)_{\lambda}(\,\cdot\,;x,\xi))$, we get $$\begin{gathered}
\label{uaub_difference}
T_{\lambda}^* (U_A - U_B) T_{\lambda} f(y) = \int_0^1 \iint \Big[ (\partial_r x_r)_j(x,\xi) T_{\lambda} (\partial_j f)(\Phi_r(x,\xi)) g_{\lambda}(y;x,\xi) \\
+ (\partial_r \xi_r)_j(x,\xi) \lambda^{-1/2} \tilde{T}_{j,\lambda} f(\Phi_r(x,\xi)) g_{\lambda}(y;x,\xi) \Big] \,dx\,d\xi\,dr.\end{gathered}$$ Lemmas \[fbisobolevbounded\] and \[wavepacketparameter\] imply the desired estimate.
\[localizederrorstability\] Suppose $\hat{f}$ is supported in ${\lvert \xi \rvert} \sim \lambda$. Write $M_A f = (D_t + P_y^A) T_{\lambda}^* U_A T_{\lambda} f$ and $N_A f = (D_t^2 - (P_y^A)^2) T_{\lambda}^* U_A T_{\lambda} f$. Then $$\begin{aligned}
{\lVert (M_A - M_B) f \rVert}_{L^2} &\lesssim \lambda {\lVert A-B \rVert}_{C^{0,1}} {\lVert f \rVert}_{L^2}, \label{halfwaveerrorstability} \\
{\lVert (N_A - N_B) f \rVert}_{L^2} &\lesssim \lambda^2 {\lVert A-B \rVert}_{C^{0,1}} {\lVert f \rVert}_{L^2}. \label{fullwaveerrorstability}\end{aligned}$$
From we have $$\begin{gathered}
(M_A - M_B) f = \iint (U_A - U_B) T_{\lambda} f(x,\xi) (g_{x,\xi}^A)(y;x,\xi) \,dx\,d\xi \\
+ \iint U_B T_{\lambda} f(x,\xi) (g_{x,\xi}^A - g_{x,\xi}^B)_{\lambda}(y;x,\xi) \,dx\,d\xi\end{gathered}$$ where $g_{x,\xi}^A$ is as in Lemma \[wavepacketparameter\]. That lemma, the argument in Lemma \[flowstability\], and Lemma \[wavepacketparameterstability\] give . For the other estimate we write $$D_t^2 - P_y^2 = (D_t + P_y)(D_t + P_y) - 2P_y(D_t + P_y)$$ which gives $$\begin{gathered}
(N_A - N_B) f = [(D_t + P_y^A) M_A - (D_t + P_y^B) M_B] f \\
- 2 (P_y^A - P_y^B) M_A f - 2 P_y^B (M_A - M_B) f.\end{gathered}$$ The last two terms have $L^2$ norms $\lesssim \lambda^2 {\lVert A-B \rVert}_{C^{0,1}} {\lVert f \rVert}_{L^2}$ by Lemma \[psdostability\] and . The first term has the form $$\begin{gathered}
[(D_t + P_y^A) M_A - (D_t + P_y^B) M_B] f =\!\! \iint (U_A - U_B) T_{\lambda} f(x,\xi) (\tilde{g}_{x,\xi}^A)(y;x,\xi) \,dx\,d\xi \\
+ \iint U_B T_{\lambda} f(x,\xi) (\tilde{g}_{x,\xi}^A - \tilde{g}_{x,\xi}^B)_{\lambda}(y;x,\xi) \,dx\,d\xi\end{gathered}$$ where $\tilde{g}_{x,\xi}^A = m_{x,\xi}^A(z,D_z) g_{x,\xi}^A + \tilde{m}_{x,\xi}^A(z,D_z) g$, using the notation in Lemma \[psdofullwaveapproximate\]. The Schwartz seminorms of $\tilde{g}_{x,\xi}^A$ are $\lesssim \lambda^{1/2}$ by , and those of $\tilde{g}_{x,\xi}^A - \tilde{g}_{x,\xi}^B$ are $\lesssim \lambda^2 {\lVert A-B \rVert}_{L^{\infty}}$ by , , . Now follows in the same way as .
We proceed to prove stability results for the operators where all the frequencies are summed up.
\[lemma:shatstability\] ${\lVert (\widehat{S}_A(t)-\widehat{S}_B(t)) g \rVert}_{H^{\alpha}} \lesssim {\lVert A-B \rVert}_{C^{0,1}} {\lVert g \rVert}_{H^{\alpha}}$.
Because of it is enough to prove the estimate for $\tilde{S}_A(t)-\tilde{S}_B(t)$. But we have $$(\tilde{S}_A(t)-\tilde{S}_B(t))g = \sum_{k \geq k_0} ( (E_{k,A}^+(t) - E_{k,B}^+(t))g + (E_{k,A}^-(t) - E_{k,B}^-(t))g )$$ where $$\begin{gathered}
(E_{k,A}^{\pm}(t) - E_{k,B}^{\pm}(t))g = T_k^* [U_{k,A}^{\pm}(t) - U_{k,B}^{\pm}(t)] T_k(\frac{i}{2} Q_{k,A}^{\pm} \beta_k(D) g) \\
+ T_k^* U_{k,B}^{\pm}(t) T_k(\frac{i}{2} [Q_{k,A}^{\pm} - Q_{k,B}^{\pm}] \beta_k(D) g).\end{gathered}$$ By Lemmas \[psdostability\] and \[flowstability\] one gets $$\label{estability}
{\lVert (E_{k,A}^{\pm}(t) - E_{k,B}^{\pm}(t))g \rVert}_{L^2} \lesssim {\lVert A-B \rVert}_{C^{0,1}} {\lVert g \rVert}_{L^2}.$$ We sum up these estimates and use frequency localization to obtain $${\lVert (\tilde{S}_A(t)-\tilde{S}_B(t))g \rVert}_{H^{\alpha}} \lesssim {\lVert A-B \rVert}_{C^{0,1}} {\lVert g \rVert}_{H^{\alpha}}.$$
The next lemma considers the error $T_A(t) = (D_t^2 - A(x,D_x)) \widehat{S}_A(t)$.
\[lemma:tstability\] ${\lVert (T_A(t) - T_B(t)) g \rVert}_{H^{\alpha}} \lesssim {\lVert A-B \rVert}_{C^{0,1}} {\lVert g \rVert}_{H^{\alpha+1}}$ when ${\lvert \alpha \rvert} \leq 1$.
As above, it is enough to consider the operator $$\begin{gathered}
\tilde{T}_A(t)g = (D_t^2 - A(x,D_x)) \tilde{S}_A(t)g = -t \sum_{k < k_0} A(x,D_x) g_k \\
+ \sum_{k \geq k_0} (D_t^2 - A_k(x,D_x))(E_{k,A}^+(t)g + E_{k,A}^-(t)g) \\
+ \sum_{k \geq k_0} (A_k(x,D_x) - A(x,D_x))(E_{k,A}^+(t)g + E_{k,A}^-(t)g).\end{gathered}$$ We write the last three terms as $\tilde{T}_{A,j}(t) g$ for $j = 1,2,3$. Clearly $${\lVert (\tilde{T}_{A,1}(t) - \tilde{T}_{B,1}(t)) g \rVert}_{H^{\alpha}} \lesssim {\lVert A-B \rVert}_{C^{0,1}} {\lVert g \rVert}_{H^{\alpha}}.$$ For the second term we use that $$(D_t^2 - A_k(x,D_x)) E_{k,A}^{\pm}(t) g = (D_t^2 - (P_{k,A}^{\pm})^2) E_{k,A}^{\pm}(t) g + C_{k,A}^{\pm} E_{k,A}^{\pm}(t) g$$ where $C_{k,A}^{\pm} = (P_{k,A}^{\pm})^2 - A_k(x,D_x)$. The first term is just $N_A (\frac{i}{2} Q_{k,A}^{\pm} g_k)$, and $${\lVert N_A(\frac{i}{2}Q_{k,A}^{\pm} g_k) - N_B(\frac{i}{2}Q_{k,B}^{\pm}g_k) \rVert}_{L^2} \lesssim 2^k {\lVert A-B \rVert}_{C^{0,1}} {\lVert g \rVert}_{L^2}$$ by and Lemma \[psdostability\]. Writing $$C_{k,A}^{\pm} - C_{k,B}^{\pm} = (P_{k,A}^{\pm})^2 - (P_{k,B}^{\pm})^2 - (A_k(x,D_x) - B_k(x,D_x)),$$ and using the argument in the end of Lemma \[psdostability\], we see that $${\lVert (C_{k,A}^{\pm} - C_{k,B}^{\pm}) \tilde{\beta}_k(D) g \rVert}_{L^2} \lesssim 2^{2k} {\lVert A-B \rVert}_{L^{\infty}} {\lVert g \rVert}_{L^2}$$ when $\tilde{\beta}_k(\xi)$ is a cutoff to ${\lvert \xi \rvert} \sim 2^k$. Lemma \[localizederrorstability\] shows $${\lVert (\tilde{T}_{A,2}(t) - \tilde{T}_{B,2}(t)) g \rVert}_{H^{\alpha}} \lesssim {\lVert A-B \rVert}_{C^{0,1}} {\lVert g \rVert}_{H^{\alpha+1}}.$$ For the last term we write $$\begin{gathered}
(\tilde{T}_{A,3}(t)-\tilde{T}_{B,3}(t))g = \sum_{\pm} \sum_{k \geq k_0} (A_k(x,D_x) - A(x,D_x))((E_{k,A}^{\pm}(t)-E_{k,B}^{\pm}(t))g) \\
+ \sum_{\pm} \sum_{k \geq k_0} (A_k(x,D_x) - A(x,D_x) - (B_k(x,D_x) - B(x,D_x))) E_{k,B}^{\pm}(t) g\end{gathered}$$ The discussion in Lemma \[stildeestimates\] and give the bound for the first term. The second term is handled as in Lemma \[stildeestimates\], except that we use $${\lVert a-a_k(x) - (b-b_k(x)) \rVert}_{L^{\infty}} \leq {\lVert A-B \rVert}_{C^{0,1}} 2^{-k/2}.$$ One first gets the bound for $-1 \leq \alpha \leq 0$ and then for ${\lvert \alpha \rvert} \leq 1$ by computing the derivative. The result follows.
We note that the last estimate holds for the operators $T(t,s)$ with uniform constants when $t,s \in [-M,M]$. The final estimate we need is for the Volterra solution operator.
\[lemma:volterrastability\] ${\lVert (V_A - V_B) F \rVert}_{L^{\infty}_t H^{\alpha}_x} \lesssim {\lVert A-B \rVert}_{C^{0,1}} {\lVert F \rVert}_{L^{\infty}_t H^{\alpha+1}_x}$ when ${\lvert \alpha \rvert} \leq 1$.
Recalling the definition of $V$ from Lemma \[lemma:volterra\], we need to estimate the $L^{\infty}_t H^{\alpha}_x$ norm of terms of the form $$\begin{gathered}
I(t,x) = \int_0^t \!\! \int_0^{s_1} \!\!\!\!\! \cdots \! \int_0^{s_{j-1}} \! T_B(t,s_1) \cdots T_B(s_{l-1},s_l) (T_A-T_B)(s_l,s_{l+1}) \\
T_A(s_{l+1},s_{l+2}) \cdots T_A(s_{j-1},s_j) F(s_j,x) \,ds_j\cdots\,ds_1\end{gathered}$$ Choose $C = C(M)$ such that for $t,s \in [-M,M]$, $$\begin{aligned}
& {\lVert T_A(t,s) g \rVert}_{H^{\alpha}} + {\lVert T_B(t,s) g \rVert}_{H^{\alpha}} \leq C {\lVert g \rVert}_{H^{\alpha}}, \quad -1 \leq \alpha \leq 2, & \\
& {\lVert (T_A-T_B)(t,s) g \rVert}_{H^{\alpha}} \leq C {\lVert A-B \rVert}_{C^{0,1}} {\lVert g \rVert}_{H^{\alpha+1}}, \quad -1 \leq \alpha \leq 1. & \end{aligned}$$ Using the $H^{\alpha}$ estimate for each $T_B$, the $H^{\alpha+1} \to H^{\alpha}$ estimate for $T_A-T_B$, and then the $H^{\alpha+1}$ estimate for each $T_A$ gives $${\lVert I(t,\,\cdot\,) \rVert}_{H^{\alpha}} \leq \frac{C^j t^j}{j!} {\lVert A-B \rVert}_{C^{0,1}} {\lVert F \rVert}_{L^{\infty}_t H^{\alpha+1}_x}.$$ There are $j$ terms of the form $I(t,x)$ at level $j$. It follows that $${\lVert (V_A - V_B) F \rVert}_{L^{\infty}_t H^{\alpha}_x} \leq \Big( \sum_{j=1}^{\infty} \frac{j (CM)^j}{j!} \Big) {\lVert A-B \rVert}_{C^{0,1}} {\lVert F \rVert}_{L^{\infty}_t H^{\alpha+1}_x}.$$
For the full solution operator we have $$\begin{gathered}
(S_A - S_B)(t)g = (\widehat{S}_A - \widehat{S}_B)(t)g + \int_0^t (\widehat{S}_A - \widehat{S}_B)(t,s) V_A(T_A(s)g(x)) \,ds \\
+ \int_0^t \widehat{S}_B(t,s) (V_A - V_B) (T_A(s)g(x)) \,ds \\
+ \int_0^t \widehat{S}_B(t,s) V_B( (T_A-T_B)(s)g(x) ) \,ds.\end{gathered}$$ Consequently, for $g \in H^{\alpha}$ with $-1 \leq \alpha \leq 1$, $${\lVert (S_A - S_B)(t)g \rVert}_{H^{\alpha+1}} \lesssim {\lVert A-B \rVert}_{C^{0,1}} {\lVert g \rVert}_{H^{\alpha+1}}$$ by Lemmas \[lemma:shatstability\], \[lemma:tstability\] and \[lemma:volterrastability\]. This shows .
To prove , take $g \in H^{\alpha}$ with $-1 \leq \alpha < 2$, and write $g = g_s + g_r$ with $g_s \in H^2$ and ${\lVert g_{r} \rVert}_{H^{\alpha}}$ small. Using the triangle inequality and , we get $$\begin{aligned}
{\lVert (S_A(t)-S_B(t))g_s \rVert}_{H^3} &\lesssim {\lVert g_s \rVert}_{H^2}, \\
{\lVert (S_A(t)-S_B(t))g_s \rVert}_{H^2} &\lesssim {\lVert A-B \rVert}_{C^{0,1}} {\lVert g_s \rVert}_{H^2},\end{aligned}$$ and by interpolation ${\lVert (S_A(t)-S_B(t))g_s \rVert}_{H^{3-\kappa}} \lesssim {\lVert A-B \rVert}_{C^{0,1}}^{\kappa} {\lVert g_s \rVert}_{H^2}$. Thus, given $\varepsilon > 0$, by choosing ${\lVert g_r \rVert}_{H^{\alpha}}$ and $\delta > 0$ small enough we obtain $$\begin{gathered}
{\lVert (S_A(t)-S_B(t))g \rVert}_{H^{\alpha+1}} \leq {\lVert (S_A(t)-S_B(t))g_s \rVert}_{H^{\alpha+1}} \\
+ {\lVert S_A(t)g_r \rVert}_{H^{\alpha+1}} + {\lVert S_B(t)g_r \rVert}_{H^{\alpha+1}} < \varepsilon\end{gathered}$$ whenever ${\lVert A-B \rVert}_{C^{0,1}} < \delta$. This ends the proof of Theorem \[thm:main\].
Extensions
==========
Here we give some extensions of Theorem \[thm:main\], following Section 4 of [@smith1]. The solution operators in each case are obtained from $\widehat{S}(t)$ similarly as in [@smith1], and the stability results follow from the arguments given in Section \[sec:stability\]. Therefore we will omit the proofs in this section. We emphasize that for each theorem below, the stability proof is constructive.
\[thm:main\_full\] Assume – , and let $-1 \leq \alpha \leq 2$. Suppose that $f \in H^{\alpha+1}$, $g \in H^{\alpha}$, and $F \in L^1_t H^{\alpha}_x$. Then there is a unique solution in $C^0_t H^{\alpha+1}_x \cap C^1_t H^{\alpha}_x$ for the problem $$\left\{ \begin{array}{rl}
(D_t^2 - A(x,D_x))u(t,x) &\!\!\!= F(t,x), \\[4pt]
u|_{t=0} &\!\!\!= f, \\[4pt]
\partial_t u|_{t=0} &\!\!\!= g.
\end{array} \right.$$ The solution satisfies $$\label{fullcauchy_normestimate}
{\lVert u(t) \rVert}_{H^{\alpha+1}} \lesssim {\lVert f \rVert}_{H^{\alpha+1}} + {\lVert g \rVert}_{H^{\alpha}} + {\lVert F \rVert}_{L^1_t H^{\alpha}_x}.$$ Also, let $A = (a^{ij})$, $B = (b^{ij})$ satisfy , and let $u_A$, $u_B$ be the corresponding solutions. If $-1 \leq \alpha < 2$, then for any $\varepsilon > 0$ there is $\delta > 0$ such that $$\label{uniform_stability_full}
{\lVert u_A(t) - u_B(t) \rVert}_{H^{\alpha+1}} < \varepsilon \ \text{ whenever }\ {\lVert A-B \rVert}_{C^{0,1}} < \delta.$$ Further, if $-1 \leq \alpha \leq 1$ and $f \in H^{\alpha+2}, g \in H^{\alpha+1}, F \in L^1 H^{\alpha+1}_x$, then $$\label{lipschitz_stability_full}
{\lVert u_A(t) - u_B(t) \rVert}_{H^{\alpha+1}} \lesssim {\lVert A-B \rVert}_{C^{0,1}} ({\lVert f \rVert}_{H^{\alpha+2}} + {\lVert g \rVert}_{H^{\alpha+1}} + {\lVert F \rVert}_{L^1_t H^{\alpha+1}_x}).$$
Here we show how the existence and uniqueness part of Theorem \[thm:main\_full\] can be used to prove stability of the map $A \mapsto u(t,\,\cdot\,)$ in Theorem \[thm:main\]. It is enough to prove for $-1 \leq \alpha \leq 1$, and the uniform continuity will follow as in Theorem \[thm:main\]. Let $A, B$ be two metrics satisfying , and let $g \in H^{\alpha}$. We denote by $u_A$, $u_B$ the solutions to with data $g$, and write $v = u_A - u_B$. Then $v$ satisfies $$\left\{ \begin{array}{rl}
(D_t^2 - A(x,D_x))v &\!\!\!= (A(x,D_x)-B(x,D_x)) u_B, \\[4pt]
v(0) = \partial_t v(0) &\!\!\!= 0,
\end{array} \right.$$ which implies $${\lVert v(t) \rVert}_{H^{\alpha+1}} \lesssim {\lVert A-B \rVert}_{C^{0,1}} (\sup_{{\lvert t \rvert} \leq M} {\lVert u_B(t) \rVert}_{H^{\alpha+2}}) \lesssim {\lVert A-B \rVert}_{C^{0,1}} {\lVert g \rVert}_{H^{\alpha+1}}.$$ This is the required bound.
Now we consider the wave equation with divergence form operator $$A^D(x,D_x) u = D_{x_i}(a^{ij}(x) D_{x_j} u).$$
\[thm:divergence\_full\] Assume – , and let $-2 \leq \alpha \leq 1$. Suppose that $f \in H^{\alpha+1}$, $g \in H^{\alpha}$, and $F \in L^1_t H^{\alpha}_x$. Then there is a unique solution in $C^0_t H^{\alpha+1}_x \cap C^1_t H^{\alpha}_x$ for the problem $$\left\{ \begin{array}{rl}
(D_t^2 - A^D(x,D_x))u(t,x) &\!\!\!= F(t,x), \\[4pt]
u|_{t=0} &\!\!\!= f, \\[4pt]
\partial_t u|_{t=0} &\!\!\!= g.
\end{array} \right.$$ The solution satisfies . Also, solutions satisfy and , if the ranges for $\alpha$ are replaced by $-2 \leq \alpha < 1$ and $-2 \leq \alpha \leq 0$, respectively.
Finally, we consider the Laplace-Beltrami type operator $$A^L(x,D_x) u = \rho(x)^{-1} D_{x_i}(\rho(x) a^{ij}(x) D_{x_j} u)$$ where $\rho = (\det\,(a^{ij}))^{1/2}$.
\[thm:laplacebeltrami\_full\] Assume – , and let $-1 \leq \alpha \leq 1$. Suppose that $f \in H^{\alpha+1}$, $g \in H^{\alpha}$, and $F \in L^1_t H^{\alpha}_x$. Then there is a unique solution in $C^0_t H^{\alpha+1}_x \cap C^1_t H^{\alpha}_x$ for the problem $$\left\{ \begin{array}{rl}
(D_t^2 - A^L(x,D_x))u(t,x) &\!\!\!= F(t,x), \\[4pt]
u|_{t=0} &\!\!\!= f, \\[4pt]
\partial_t u|_{t=0} &\!\!\!= g.
\end{array} \right.$$ The solution satisfies . Also, solutions satisfy and , if the ranges for $\alpha$ are replaced by $-1 \leq \alpha < 1$ and $-1 \leq \alpha \leq 0$, respectively.
[10]{}
E. J. Cand[è]{}s and L. Demanet, *[The curvelet representation of wave propagators is optimally sparse]{}*, Comm. Pure Appl. Math. **58** (2005), 1472–1528.
E. J. Cand[è]{}s and D. L. Donoho, *[New tight frames of curvelets and optimal representations of objects with piecewise $C^2$ singularities]{}*, Comm. Pure Appl. Math. **57** (2004), 219–266.
M. V. [de Hoop]{}, *Microlocal analysis of seismic inverse scattering*, [Inside out. Inverse problems and applications]{} (G. Uhlmann, ed.), Cambridge University Press, Cambridge, 2003, pp. 219–296.
M. V. [de Hoop]{}, H. Smith, and G. Uhlmann, Preprint.
J.-M. Delort, *[F.B.I. transformation. Second microlocalization and semilinear caustics]{}*, Lecture notes in mathematics, vol. 1522, Springer-Verlag, Berlin, 1992.
H. Smith, *[A parametrix construction for wave equations with $C^{1,1}$ coefficients]{}*, Ann. Inst. Fourier Grenoble **48** (1998), no. 3, 797–835.
[to3em]{}, *[Spectral cluster estimates for $C^{1,1}$ metrics]{}*, Amer. J. Math. **128** (2006), no. 5, 1069–1103.
C. C. Stolk, *On the modeling and inversion of seismic data*, Ph.D. thesis, Utrecht University, 2000, 139pp.
D. Tataru, *[Strichartz estimates for operators with nonsmooth coefficients and the nonlinear wave equation]{}*, Amer. J. Math. **122** (2000), no. 2, 349–376.
[to3em]{}, *[Strichartz estimates for second order hyperbolic operators with nonsmooth coefficients. II]{}*, Amer. J. Math. **123** (2001), no. 3, 385–423.
[to3em]{}, *[Strichartz estimates for second order hyperbolic operators with nonsmooth coefficients. III]{}*, J. Amer. Math. Soc. **15** (2002), 419–442.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The dynamics of quantized vortices in weakly interacting superfluids are often modeled by a nonlinear Schrödinger equation. In contrast, we show that quantized vortices in fact obey a non-Hamiltonian evolution equation, which enhances dispersion along the vortex line while introducing a gain mechanism. This allows the vortex medium to support a helical shock front propagating ahead of a dissipative soliton. This dynamic relaxes localized curvature events into Kelvin wave packets. Consequently, a beyond local induction model provides a pathway for decay in low-temperature quantum turbulence.'
author:
- 'Scott A. Strong'
- 'Lincoln D. Carr'
title: 'Non-Hamiltonian Dynamics of Quantized Vortices in Bose-Einstein Condensates'
---
Quantized vortices are slender, non-diffusive regions of low density about which the superfluid bulk circulates at strengths defined by multiples of Plank’s constant scaled by a characteristic mass \[\]. The three-dimensional Gross-Pitaevskii equation of mean-field theory restricts this density depletion to a formally one-dimensional subregion, of the otherwise irrotational gas. This vortex defect compromises the connectedness of the condensate in accordance with Helmholtz theorem. If vorticity acts as a primitive source of a fluid flow, then quantum liquids are a setting where a complete theory of the topological hydrodynamics is most likely. Such a theory will resolve mean-field physics at the scale of the vortex core with that of the disordered arrangement of persistent and stable vortex structures with homogeneous circulation characterizing quantum turbulence. This quantum tangle supports various cascade processes which transfer energy between the spatial scales \[\]. Unique to quantum fluids is the Kelvin wave cascade which relaxes high curvature cusps to small wavelength helical excitations along the vortex. These waves transport turbulent energy to the boundaries \[\]. In this Letter, we derive a relatively simple, but non-Hamiltonian evolution of the geometric properties of a vortex defect under *local induction models*. This evolution approximates dynamics at intermediate scales and predicts that the vortex medium transforms an initially localized curvature soliton into a helical shock wave where a dissipative soliton travels behind helical excitations. Gain mechanisms introduced by the non-Hamiltonian structure results in an increase of average curvature and signifies the emergence of the small-scale structures necessary for acoustic emission.
The Biot-Savart integral provides a representation of the velocity field by “un-curling” the vorticity field. Locally induced evolutions are given by asymptotic approximations to regularizations of this singular integral. There are several regularization techniques available and all yield a lowest order *local induction approximation*, which states that the binormal component of the velocity field is proportional to local curvature \[\]. Applying Hasimoto’s transform to the local induction approximation defines a cubic focusing nonlinear Schrödinger evolution of the curvature and torsion of the vortex \[\]. This theory predicts a bright curvature soliton defining a traveling kink on the vortex line. We call the vortex state, corresponding to this integrable dynamic of the geometric properties, a *Hasimoto vortex soliton*, [see Fig. (\[fig:schematics\]).]{} We study breakdown in the integrable theory through a non-Hamiltonian description which introduces a curvature gain/loss mechanism while enhancing dispersion on the vortex medium. This model evolves the Hasimoto vortex soliton towards a log-normal type distribution, which we call a *cascade soliton*. Represented on the vortex line, the soliton kink decreases its amplitude while the curvature of the straight background increases ahead of the disturbance. Our non-Hamiltonian dynamic predicts the existence of two qualitatively different dynamical regimes, which are confirmed via simulation. For small perturbations, the kink maintains its structure for longer times, a robustness that is indicative of a dissipative soliton. Increasing dispersion erodes the previously robust kink into a packet of Kelvin waves generating a profile similar to a dispersive shock wave. This Letter introduces a fully nonlinear integro-differential equation and uses it to approximate the non-Hamiltonian dynamics about the bright soliton fixed point of the integrable theory. Prior to simulations, we consider the dynamics predicted by the emergent gain/loss mechanism in conjunction with the changes to plane wave dispersion to understand the short-time behavior of a soliton.
[Vortex filament methods]{} simulate a quantum fluid by evolving its vortical skeleton according to the Biot-Savart integral and are significantly more efficient than mean-field methods for vortex dominated flows \[\]. Recent simulations by Salman demonstrate that evolutions given by the mean-field, Biot-Savart and induction models are consistent up to the point where the smallest length scales dominate the physics, a known limitation of filament methods \[\]. At the same time, Bustamante and Nazarenko have shown that the Biot-Savart integral manifests from Gross-Pitaevskii mean–field dynamics and provides a self-consistent regularization procedure \[\]. This allows us to derive a non-Hamiltonian evolution consistent with a locally induced flow generated by a region of vortex whose arc-length is on the order of the condensate healing length. Our prediction of helical waves generated from a localized curvature event is consistent with current models of energy transfer in the highly quantum turbulent regime \[\].
[As shown in Fig. (\[fig:schematics\]),]{} we define $\vec{\gamma}\in\mathbb{R}^{3}$ as the set of points in three-space [corresponding to a Hasimoto vortex soliton, i.e., a bright curvature soliton shown in the inset of Fig. (\[fig:schematics\]).]{} The [vortex is]{} parameterized by an arc-length, $s$, and changes with time, $t$, so that $\vec{\gamma}=\vec{\gamma}(s,t)$. In our previous work we derive, from the Biot-Savart integral, an exact representation of the velocity field induced by a plane circular arc of vorticity \[\]. The corresponding asymptotic representation, taken to lowest order in curvature, is equivalent to the local induction approximation. Specifically, the non-circulatory and non-axial velocity field neighboring an arbitrary reference point on the vortex line generates a flow satisfying the vector evolution law, $$\begin{aligned}
\label{eqn:BNF}
\frac{\partial \vec{\gamma}}{\partial t} = \alpha \left( \frac{\partial \vec{\gamma}}{\partial s} \times \frac{\partial^{2} \vec{\gamma}}{\partial s^{2}}\right),\end{aligned}$$ where the cross-product $\vec{\gamma}_{s}\times\vec{\gamma}_{ss}=\kappa \hat{\bm{b}}$ is defined using the binormal vector of the Frenet-Serret frame and $\kappa=\kappa(s,t)$ is the local curvature. The Hasimoto vortex soliton is a prediction of the local induction approximation, given by Eq. (\[eqn:BNF\]) where $\alpha=1$, and forms a propagating curvature disturbance, $\kappa(s,t) = 2 \operatorname{sech}(s-t)$, with constant torsion, $\tau=1$. To understand the fragility of the integrable model, and its transition to a cascade soliton, we consider the case of non-constant $\alpha$.
![Hasimoto’s Vortex Soliton. We depict the Frenet frame, $\hat{\textbf{t}}, \hat{\textbf{n}}, \hat{\textbf{b}}$, for a hyperbolic secant (inset) bright soliton, $\kappa, \tau=1$, from the integrable theory, $2i\psi_{t} + 2\psi_{ss}+|\psi|^{2}\psi=0$, corresponding to Hasimoto’s map applied to the local induction approximation, $\vec{\gamma}_{t} = \kappa \hat{\textbf{b}}$. []{data-label="fig:schematics"}](Fig1.pdf){width="50.00000%"}
The condensate healing length, $\xi$, defines the vortex core size and its product with the characteristic curvature, $\kappa$, yields the small parameter $\epsilon = \xi \kappa \ll 1$. The healing length, in ratio with the characteristic system size, $d$, defines the parameter $\delta = d/\xi$ and is large when the characteristic system size is taken to be the condensate width. The proportionality constant in Eq. (\[eqn:BNF\]) is a function of the dimensionless parameters, $\alpha = \alpha(\delta, \epsilon)$, and has a tidy representation given by the matched asymptotic expansion \[\], $$\begin{aligned}
\label{eqn:SV}
\alpha(\delta,\epsilon) = \frac{\Gamma}{4\pi} \left[\ln\left(\frac{1}{\delta}\right) +\ln\left(\sqrt{1+ \frac{\epsilon^{2} \delta^2}{8}}\right)\right],\end{aligned}$$ where $\Gamma$ measures the strength of the condensate circulation about the vortex line. The local induction approximation retains only the curvature-independent logarithmic singularity and is defined as Eq. (\[eqn:BNF\]), where $\alpha=\alpha(\delta,\epsilon=0)$. If we assume that $\epsilon$ is not formally zero, then the scale separation $\epsilon \ll \epsilon \delta \ll \delta$ defines a regime where the linearization of Eq. (\[eqn:SV\]), i.e. the local induction approximation, is accurate. [If a parameterization is given, then the characteristic length, $d$, can be associated with the domain of Biot-Savart integration. Parameterizing an arbitrary vortex element by a plane circular arc gives a wide range for the non-dimensional parameter, $ \delta \in (0.3416293, 100)$. The lower bound of this interval is given by the Bustamante-Nazarenko regularization, while the upper bound corresponds to a vortex ring with a radius on the order of tens of micrometers. While this approach can maintain the previous scale separation, it also permits the study of flows induced by arcs of vorticity with small central angle such that $\epsilon \delta=O(\epsilon)$. We specifically consider motion induced by filament elements whose length is near the new self-consistent cutoff \[\].]{}
Hasimoto’s transformation rotates the Frenet basis into $\mathbb{C}^{3}$ where the [geometric wave function]{}, $ \psi(s,t) = \kappa(s,t)~ \mbox{Exp}\left[{i \int^{s}_{0}ds' \tau(s',t)}\right]$, carries the curvature and torsion variables of a vortex line satisfying local induction approximation \[\]. This transformation is robust and can be used to map more general flows, often leading to complicated integro-differential equations \[,\]. Restricting ourselves to binormal flows defined by Eq. (\[eqn:BNF\]) and applying the asymptotic representations of $\alpha$ we recast the corresponding integro-differential equation as a fully nonlinear differential equation. [The predicted dynamical state is a cascade soliton, which is a solitary wave accompanied by helical excitations, and is depicted in Fig. (\[fig:schematics2\]).]{}
![Non-Hamiltonian Cascade Soliton. An initial Hasimoto vortex soliton experiences dispersion producing helical waves propagating ahead of the soliton kink. The non-Hamiltonian gain mechanism supports both the kink and the helical excitations as the localized curvature event is transitioned to a cascade process. []{data-label="fig:schematics2"}](Fig2.pdf){width="50.00000%"}
Measuring time in units defined by the ratio of characteristic vortex with curvature, fluid circulation, and the first term in the expansion of $\alpha$, $\tilde{t}=4\pi\tilde{\gamma}/(\tilde{\kappa} \Gamma a_{0}(\delta))$, gives $\alpha(\delta,0)=1$. Application of the Hasimoto transformation to the general case, $\alpha=\alpha(\delta, \epsilon)$ yields a nonlinear integro-differential evolution of the geometric variables, $$\begin{aligned}
\label{eqn:GHNLS}
&i \psi_{t} + \left[\alpha\psi\right]_{ss} +\frac{\alpha}{2} |\psi|^{2}\psi +\frac{\psi}{2} \int^{s} |\psi|^{2} \alpha_{s'}\, ds'=0.
\end{aligned}$$ This evolution maintains a Schrödinger structure, however, it is difficult to derive useful information from it. The small parameter $\epsilon \ll 1$ provides an expansion of $\alpha$ in powers of $\kappa$. For Eq. (\[eqn:SV\]), we find that $\alpha(-\kappa)=\alpha(\kappa)$. [After truncating quartic and higher terms we arrive at the]{} simpler evolution, $$\begin{aligned}
\label{eqn:GHNLSPDE1}
&i \psi_{t} + \psi_{ss} + \frac{1}{2}|\psi|^{2} \psi + \lambda \left( \left[|\psi|^2 \psi\right]_{ss}+ \frac{3}{4}|\psi|^{4}\psi \right) = 0,\end{aligned}$$ which we call the *non-Hamiltonian vortex cascade equation* (NVC). The [correction parameter]{}, $\lambda$, depends on our dimensionless constants and is the ratio of the second and first coefficients in the expansion of $\alpha$. In the local induction approximation $\lambda=0$ and we have an integrable theory. If $\lambda \neq 0$, then integrability is compromised so severely that nearly all underlying symmetries are broken. With the exception of arc-length, a quantity conserved by the binormal flow itself, nothing typical, [like energy or momentum]{}, is conserved.
To understand this loss of mathematical structure we inspect the the fully nonlinear term $\left[|\psi|^{2} \psi\right]_{ss}$. One can show that there does not exist a functional whose variational derivative satisfies the necessary self-adjoint conditions and therefore the evolution cannot be written as an infinite dimensional Hamiltonian system \[,\]. The system is invariant with respect to arbitrary time and space translations. However as the system is not Hamiltonian, Noether’s theorem does not apply and our continuous symmetries need not correspond to conserved densities. Application of the SYM symmetry software package \[\] to Eq. (\[eqn:GHNLSPDE1\]) found no additional continuous symmetries. Additionally, a Mathematica package that symbolically calculates conservation laws found no low-order conserved densities \[\]. While the system does possess discrete parity and time symmetries, we consider non-symmetric, time-irreversible dynamics of the nonlinear evolution.
Without $\left[|\psi|^{2} \psi\right]_{ss}$, Eq. (\[eqn:GHNLSPDE1\]) is a complex quintic Ginzburg-Landau equation used in the study of dissipative solitons \[\]. Our real coefficients imply a Hamiltonian structure and a nonlinear gain/loss mechanism must enter through other means. Madelung’s transformation decomposes Schrödinger evolutions into real and imaginary parts \[\]. [Transforming Eq. (\[eqn:GHNLSPDE1\]) yields a system of first-order evolutions on the bending density, $\rho=\kappa^{2}=|\psi|^{2}$ and torsion. From this system we find that the total bending across the vortex, $$\begin{aligned}
\label{eqn:L2}
\frac{d}{dt}\int_{a}^{b} \kappa^{2} \, ds &=-2\lambda \int_{a}^{b} \rho_{s} \rho \tau \, ds, \end{aligned}$$ is no longer conserved in time. Specifically, the bending dynamics arise from the correction term and are, in part, determined by the helicity density, $\rho \tau$. This density corresponds to the momentum density in the condensate picture, and is also not conserved by the NCV for superpositions of helical modes. This non-Hamiltonian geometric gain/loss mechanism provides a pathway for dissipative soliton dynamics. For example, if the torsion is positive, then bending energy grows/decays over regions where curvature is decreasing/increasing. Simulations indicate that this feature is robust against distortions manifesting from plane wave dispersion.]{}
A plane wave solution of the form $\psi = A e^{i(k s- \omega t)}$ defines a single mode helix. According to the NVC plane waves obey the corrected nonlinear dispersion relation \[\], $$\begin{aligned}
\label{eqn:Dispersion}
\omega(k,A,\lambda) = k^{2}(1+ \lambda A^{2}) - \frac{A^{2}}{2} - \frac{3 \lambda}{4} A^{4}.
\end{aligned}$$ The initial state of the Hasimoto soliton is given by $\psi = 2 \mbox{sech}(s) e^{is}$, which defines a narrow-band curvature packet, with over $99\%$ of its total [initial]{} bending captured between wave numbers $k\in[0,5]$. Relating the wave amplitude, $A$, to wavenumber via Fourier transform allows us to plot the group velocity for the initial data, which is given in Fig. (\[fig:dispersion\]). These data show that increasing $\lambda$ enhances the propagation speed of long wavelength curvature modes. Enhancing dispersion of these modes causes the curvature function to distort. [Simulations indicate that the peak jettisons curvature, which causes the first moment of the distribution to propagate faster, distorting the distribution into a log-normal form.]{} Additionally, the simulations depict a localized curvature peak that stays discernible under considerable dispersion because of the support provided by the emergent gain/loss mechanism. If total bending were conserved, then the dynamic would cause the peak amplitude to erode completely into the vortex.
![Dispersion of Hasimoto’s Soliton. The non-Hamiltonian evolution enhances to group velocity of the long wavelength modes in an initial soliton state. The speed of the curvature peak and first moment (inset) are increasing functions of $\lambda$ until roughly $\lambda=0.19$ where a majority of shape defining Fourier modes leave an eroding peak to bolster the first moment. []{data-label="fig:dispersion"}](Fig3.pdf){width="50.00000%"}
Our non-Hamiltonian vortex dynamics can be simulated through the vector evolution equation, Eq. (\[eqn:BNF\]), on a mesh of points representing the vortex in $\mathbb{R}^{3}$, or by scalar evolution of the geometric variables through the NVC. In this way the Hasimoto transformation effectively separates the geometric evolution from its Frenet-Serret representation and is significantly more efficient if vortex visualization is not required. In addition to the heightened regularity requirements on the geometric variables for the Hasimoto transformation, it should be noted that the boundary conditions for each problem are physically different. Specifically, for an arc-length variable $s\in[a,b]$, Dirichlet boundary conditions, $\psi(a,t)=\psi(b,t)=0$, allow for freely moving endpoints with zero curvature. In contrast, Dirichlet conditions on the vector evolution fix the endpoints, $\vec{\gamma}(a,t)=\vec{\gamma}_{a}$ and $\vec{\gamma}(b,t)=\vec{\gamma}_{b}$ where $\vec{\gamma}_{a},\vec{\gamma}_{b}\in\mathbb{R}^{3}$. To suppresses any differences manifesting from the endpoint behavior, we simulate Hasimoto’s soliton under both the vector and scalar evolutions on an arc-length domain an order of magnitude larger than the characteristic width of this curvature disturbance . [The total bending was calculated for both methods and found to have less than 1% squared relative error in the curvature across the lifetime of the simulation.]{} Our focus is on simulations where $\lambda = 0.17$, which displays features types of non-Hamiltonian waves, [and corresponds to flows generated by arcs whose length is near the regularization limit.]{}
[The NVC predicts]{} two qualitatively different dynamical regimes, one characterized [the cascade soliton]{} and the other by the strong dispersion of this dissipative soliton. Figure (\[fig:17Density\]) plots curvature as a density for $\lambda=0.17$ and we see that the maximum amplitude is strongly localized while the bending energy disperses in the direction of peak propagation. In Fig. (\[fig:17Density\]a) we plot the maximum value, first moment (red) and wave front (yellow), i.e., the furthest point ahead of the peak were $\kappa\approx 2\%$ of the initial peak value are fitted to linear models with square residuals, $0.99999$, $0.999198$, $0.998574$, respectively. Our simulations also show the presence of breathing oscillations, seen by the dark side bands to the peak appearing twice. In Fig. (\[fig:17Density\]c) the middle of a breath can be seen as pinch in the curvature function occurring at $t=15$ and is plotted along with curvatures at $t=0,5,10,25$ in the co–moving frame. Simulations omitting the $\left[|\psi|^{2} \psi\right]_{ss}$ term exhibit a similar dynamic and indicates that breathing is, in part, a consequence of the quintic nonlinearity. The completion of two breaths was corroborated with a power–spectrum analysis of the time data. In addition to this breathing, the gain/loss mechanism creates an asymmetry in the curvature profile that when coupled to dispersion leads to a trailing helical wake of low wavenumber curvature modes. The dissipative[/cascade]{} soliton corresponds to the peak amplitude following the helical excitations. Additional simulations show that for $\lambda < 0.17$ we see a similar dynamic, but the peak is strongly maintained and less curvature is dispersed. Also, for $\lambda > 0.17$ the peak speed begins to decrease as low wavenumber modes shift the curvature distribution to a log-normal form. While the curvature peak is discernible, it is difficult to spot immersed in a sea of Kelvin waves. Additionally, the breathing is abated on the simulated time scale of $40$ seconds. [If a condensate is punctured by a vortex defect with circulation $997\times 10^{-4} \textrm{$\mu$m} \cdot \textrm{cm}$, length $100\textrm{$\mu$m}$ and core size of $0.67\textrm{$\mu$m}$, then for an vortex element whose radius of curvature is $12\textrm{$\mu$m}$, a reasonable size for a vortex ring, Eq. (\[eqn:SV\]) yields a characteristic time scale on the order of milliseconds, which is with the range of times considered in reconnection studies \[-\].]{}
![Cascade Dynamics. (a) For moderate dispersion the cascade soliton has a well defined peak as it breaths and disperses. The clearly defined peak is tracked along with the first moment (red) and wave front (yellow) on a density plot of curvature. (b) Additionally, we depict vortex configurations at $t=5$ (orange), $t=15$ (green) which illustrates the breathing dynamic. (c) Lastly, a sequence of curvatures at five-second intervals are plotted against the initial state, adjusted for translations, and show a clear asymmetry in the profile in addition to the breathing. []{data-label="fig:17Density"}](Fig4.pdf){width=".5\textwidth"}
Understanding that the Biot-Savart integral is a manifestation of the mean-field Gross-Pitaevskii dynamics of the condensate, our beyond local induction model self-consistently describes the dynamics of isolated quantized vortices whose flow is induced at length scales nearing the healing length. While this non-integrable and computationally inexpensive result can be easily added to current filament models, it is also a useful symbolic tool for investigation of post reconnection dynamics \[\]. One expects that such events are the driving mechanism of highly localized curvature distributions which are now experimentally realizable for condensates with a few vortex defects \[\]. Perhaps through minimally defected flows we can gain greater insight into Onsager’s conjectured mechanism of anomalous dissipation which asserts that weak solutions of inviscid fluid dynamics are not necessarily conservative and that the geometry itself is capable of relaxation by radiating turbulent energy toward the fluid boundaries \[\].
In conclusion, we derive a non-Hamiltonian evolution for the curvature and torsion of a quantized vortex that breaks the integrability of the local induction approximation and introduces a helical shock wave on the vortex medium. Such a dynamic is necessary to support the cascade process associated with low temperature quantum turbulence. The shock consists of a leading packet of Kelvin waves dispersed from a dissipative vortex soliton, i.e., a non-Hamiltonian cascade soliton.
The authors acknowledge support from the US National Science Foundation under grant numbers PHY-1306638, PHY-1207881, PHY-1520915, OAC-1740130, and the US Air Force Office of Scientific Research grant number FA9550-14-1-0287. This work was performed in part at the Aspen Center for Physics, which is supported by the US National Science Foundation grant PHY-1607611.
[29]{}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 in [**](\doibase 10.1007/978-3-642-37569-9{_}13), (, , ) pp. [****, ()](\doibase 10.1073/pnas.1400033111) [****, ()](\doibase 10.1103/PhysRevLett.86.1410) [****, ()](\doibase 10.1016/0169-5983(96)82495-6) [****, ()](\doibase 10.1017/S0022112072002307) [****, ()](\doibase 10.1007/s10909-017-1789-8) [****, ()](\doibase 10.1103/PhysRevLett.111.165301) [****, ()](\doibase 10.1103/PhysRevE.92.053019) [****, ()](\doibase 10.1073/pnas.1312544110) [**** ()](\doibase 10.1063/1.3696689) [****, ()](\doibase 10.1088/0953-8984/13/12/201) [****, ()](\doibase 10.1017/S0022112091001143) [**](http://www.cambridge.org/gb/academic/subjects/mathematics/fluid-dynamics-and-solid-mechanics/vorticity-and-incompressible-flow?format=PB&isbn=9780521639484#DhouIglYjTlGZwDF.97), (, ) [**](https://link.springer.com/book/10.1007/978-1-4684-0274-2), (, ) [**](http://www.springer.com/us/book/9783540698432), (, , ) in @noop [**](http://www.internationalmathematicasymposium.org/IMS2006/IMS2006_CD/articles/Dimas.pdf) () [****, ()](\doibase 10.1016/j.jsc.2011.08.014) [**](http://www.springer.com/us/book/9783540233732), (, , ) [****, ()](\doibase 10.1007/BF01504657) [****, ()](\doibase 10.1137/0147063). @noop [****, ()](\doibase 10.1103/PhysRevLett.86.3080) [****, ()](\doibase
10.1103/PhysRevLett.115.170402) [****, ()](\doibase 10.1103/PhysRevLett.85.2857) [****, ()](\doibase 10.1088/0953-4075/40/24/002) [****, ()](\doibase https://doi.org/10.1016/j.physd.2009.03.006) [****, ()](\doibase 10.1073/pnas.1312535111) [****, ()](\doibase 10.1103/PhysRevX.7.021031) [****, ()](\doibase 10.1103/RevModPhys.78.87)
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
\
\
\
\
\
---
*Keywords:*
*2010 MSC:*
PREPRINT
{width="12cm"}
Acknowledgements
================
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A Chandra X-ray observation has detected an unresolved source at the center of the supernova remnant Kes 79. The best single-model fit to the source spectrum is a blackbody with an X-ray luminosity $L_{X}$(0.3-8.0 keV)=$7\times 10^{33}$ ergs s$^{-1}$. There is no evidence for a surrounding pulsar wind nebula. There are no cataloged counterparts at other wavelengths, but the absorption is high. The source properties are similar to the central source in Cas A even though the Kes 79 remnant is considerably older.'
author:
- 'F.D. Seward, P.O. Slane, R.K. Smith'
- 'M. Sun'
title: A Compact Central Object in the Supernova Remnant Kes 79
---
Introduction
============
Source 79 in the radio catalog of Kesteven (1968) lies directly in the Galactic Plane, $33^{\circ}$ NE of the Galactic Center. It is a moderately large supernova remnant (SNR) and is sometimes called G33.6+0.1. Radio observations (Velusamy, Becker & Seward 1991) show an outer shell, $11^{\prime}$ in diameter, which is approximately circular over the SW half of the remnant but with large indentations in the N and E boundaries. The brightest part of the radio remnant is an interior region with shell-like form. The southern part of this “inner shell” has the highest surface brightness in the remnant. The distance has been determined by neutral H absorption measurements to be $10\pm2$ kpc by Frail & Clifton (1989).
The first X-ray detection of Kes 79 was made using the Einstein Observatory (Seaquist & Gilmore 1982; Velusamy, Becker & Seward 1991) and showed amorphous structure with a bright center. This inspired a ROSAT observation (Seward & Velusamy 1995) to search for Crab-like structure in the interior. The results, however, showed no indication of a central pulsar or pulsar-wind nebula. The southern arc of the inner ring was found to be bright in X-rays and faint emission was observed from the outer shell, particularly close to the eastern indentation. Assuming a thermal spectrum, analysis of the ROSAT data indicated an age of $6-12\times 10^{3}$ years, an X-ray luminosity of $\sim 10^{36}$ erg s$^{-1}$, and an energy release of $5\times 10^{50}$ ergs in the SN explosion. Subsequent observations with ASCA (Sun & Wang 2000) showed that the spectrum was indeed thermal with strong lines from Mg, Si and S. The global spectrum was fit well by a single NEI model. Seward & Velusamy (1995) speculated that this remnant might be younger and closer than believed and the result of a Type Ia supernova. As evidence they cited: the circular shell, the fact that the absorption was the same as that in the path to W44 ($1^{\circ}$ distant in the plane of the sky and only 3 kpc distant from the sun), and the lack of an observable pulsar. The Chandra observation described here shows that this is not the case.
In this paper we report the detection of a point-like source at the center of the remnant, which is likely to be a neutron star created in the SN explosion. We discuss the spectrum, the apparent absence of any surrounding synchrotron emission, and briefly compare the source with similar objects. Discussion of the SNR shell is defered to a subsequent publication.
Chandra Observation
===================
Chandra observed Kes 79 on 31 July 2001. The observation was undertaken to better determine the shell-like structure and to measure spectra from different regions. An exposure of 30 ks was obtained and there were no “flares” from particle-induced background. The remnant was centered in the ACIS-I array and consequently spread over the four, $8^{\prime}$ square chips comprising the detector. The telescope was dithered and images are exposure corrected so the gaps between chips do not appear in the images. Because of the CTI degradation of the detector (Chandra X-ray Center 2001), the spectral resolution is better along the NE and SW edges of the remnant than at the center. We took this into account using appropriate tasks from the CIAO software, version 2.1, which was used for data analysis.
{width="0.95\linewidth"}
Figure 1 shows the Chandra image. The data have been adaptively smoothed (with the CSMOOTH algorithm) and show the bright inner shell of the remnant. Emission from the outer shell is weak and is hard to see (except in the E indentation) in this figure. Five unresolved sources are easily seen, including 3 within the shell. The brightest, by an order of magnitude, is the source at the middle which is the subject of this paper. The sources are labeled 1-5 in Figure 1, and Table 1 lists their properties. The May 2002 Chandra aspect solution was used. For sources within $2^{\prime}$ of the telescope axis the 90% certainty radius in the Chandra position is $0.6^{\prime \prime}$. Positions were compared with the GSC 2.2 Catalog (STScI 2001). Sources 3 ($7^{\prime}$ off-axis) and 4 ($3^{\prime}$ off-axis) are probably stars; the X-ray spectra are soft and the X-ray source positions are $0.9^{\prime \prime}$ and $0.2^{\prime \prime}$ from optical counterparts, which shows the accuracy of the aspect determination for these parts of the field.
Hardness ratios are given for eventual comparison with 2 surveys of serendipitous Chandra sources, extragalactic (Green et al 2002) and Galactic (Grindlay et al. 2002). The hardness ratio, HR = H-S/H+S where S is the number of counts from 0.3 to 2.5 keV and H is the number of counts from 2.5 to 8 keV. There is no counterpart brighter than R magnitude 19 for the central source (1) or for sources 2 and 5. Because the spectra of sources 2 and 5 are hard, they are probably background AGN.
The radial profile of the central source is consistent with the expected Chandra point-spread-function. Any extent is $<1.0^{\prime
\prime}$. Thinking this likely to be an isolated neutron star, we searched for time variability and pulsations. The light curve is constant within statistics - $2\sigma$ fluctuations $\leq$ 20% on an hourly basis. The search for periodicity was limited to periods longer than 6.4 s because the ACIS instrument, in normal mode, integrates for 3.2 s. An FFT analysis showed no coherent power significantly above the noise, giving an upper limit of $\approx30$% for pulsed power in this range.
{width="0.95\linewidth"}
Since many young pulsars are embedded in diffuse synchrotron radiation from a surrounding PWN, we searched for such a PWN at high energies. In the range 5-8 keV, the central source is still visible but diffuse emission is absent, both from the bright inner shell and from the vicinity of the central source. In the range 3-5 keV, shown in Figure 2, the inner ring is discernable and there is a small feature adjacent to the central source and extending SW. The spectrum of emission from this SW feature shows Mg and Si lines, indicating that much of the emission is thermal. Since there are only 300 counts from this feature, model parameters cannot be accurately determined from spectral fits. By subtracting the thermal spectrum observed elsewhere in the remnant we estimate that up to about 1/3 of the emission from this $\sim 0.6^{\prime} \times 0.6^{\prime}$ feature could be non-thermal. Assuming a power-law spectrum with photon index = 2 (like the Crab Nebula), the upper limit to the luminosity of a PWN at this location is calculated to be $1.5 \times 10^{33}$ ergs s$^{-1}$. This upper limit is about 1/4 the luminosity of the point-like source. We note that a PWN is usually more luminous than the pulsar itself \[20:1 for the Crab Pulsar (Toor & Seward 1977), 1.3:1 for the Vela Pulsar (Helfand et al 2001)\]
The spectrum of the central source is shown in Figure 3 and Table 2 lists the results of spectral fits. 723 counts were extracted from a circle of radius $4^{\prime\prime}$, grouped over 9 ACIS energy channels, and were restricted to the energy band 0.8-4.7 keV. The SHERPA software was used for spectral fitting. A power-law gives a fit which, although it produces an acceptable $\chi^2$ = 1.2, is obviously too weak in the range 2-4 keV and too strong in the range 4-6 keV. Furthermore, the photon index, $\gamma = 4.2$, is higher than observed for most cosmic sources and the value for absorption is high. The ASCA measured $N_{H}$ was $(1.75\pm.07)\times 10^{22}$ atoms cm$^{-2}$ and spectral fits to the bright thermal emission in our Chandra data yield $N_{H}
\approx 1.8\times 10^{22}$ atoms cm$^{-2}$. The central source spectrum is soft, softer than expected from an AGN ($\alpha
\approx 1.7$) or from a PWN ($\alpha \approx 1.5 - 2.5$)(e.g. Slane et al 2000).
{width="0.95\linewidth"}
A single blackbody spectrum gives a good fit, shown in Figure 3, but with $N_{H}$ a bit lower than expected. It is also not difficult to achieve good fits with 2-component spectra. The soft component can be either blackbody or power law and is not well constrained. A power law does not work well as the hard component and, if used as the soft component, only makes a small contribution to the emission. If we require that $N_{H}=1.8\times 10^{22}$ atoms cm$^{-2}$, the absorption observed by Chandra for the diffuse part of the remnant, then a 2-component spectrum produces a better fit than a single blackbody spectrum but the relative contribution of the soft component is small. Values of parameters are given in Table 2.
Chandra spectra of various parts of the diffuse remnant are all thermal. The Mg, Si, and S lines detected by ASCA are prominent and there is not much variation from region to region. These data and their interpretation will be the subject of a second paper (Sun et al. 2002).
------------------- ---------------- ----------------------------------------------------- -------- -------------------------- ------------------- ---------------
ACIS Hardness Optical
Source RA Dec Counts Ratio R Mag Counterpart
1, Central Source 18h 52m 38.56s +00$^{\circ}$ $40^{\prime}$ $19.84^{\prime \prime}$ 723 $-0.49\pm0.04$ $>18.5$ None
2 18h 52m 41.35s +00$^{\circ}$ $37^{\prime}$ $42.5^{\prime 60 $\phantom{-}0.37\pm0.14$ $>18.5$ None
\prime}$
3, Star? 18h 52m 41.66s +00$^{\circ}$ $47^{\prime}$ $01.4^{\prime 61 $-0.39\pm0.19$ $\phantom{>}16.0$ N020120163554
\prime}$
4, Star? 18h 52m 45.38s $+00^{\circ}$ $37^{\prime}$ $34.1^{\prime 47 $-0.83\pm0.19$ $\phantom{>}16.3$ N020120158710
\prime}$
5 18h 53m 02.87s $+00^{\circ}$ $44^{\prime}$ $22.9^{\prime 68 $\phantom{-}0.70\pm0.14$ $>18.5$ None
\prime}$
------------------- ---------------- ----------------------------------------------------- -------- -------------------------- ------------------- ---------------
The radio pulsar, B1849-00, proposed to be a high-velocity object associated with Kes 79 (Han 1997), is located just outside the remnant $3^{\prime}$ south of the rim. Frail & Clifton (1989), however, obtained 21 cm absorption data which showed PSR 1849-00 to be considerably further away than Kes 79. The present discovery of a different object at the center of Kes 79, verifies their conclusion that PSR 1849-00 is not associated with Kes 79.
As a matter of interest, PSR 1849-00 was in the field of view of the ACIS-I detector during our observation and was not detected. An upper limit on the flux is $1 \times 10^{-14}$ ergs cm$^{-2}$ s$^{-1}$. If the pulsar were 15 kpc distant, and the transmission of the ISM is $t_{ism}$, then $L_x$ would be $2.5 \times 10^{32}$ $t_{ism}$$^{-1}$ erg s$^{-1}$. Since Ė is $\sim 4 \times 10^{32}$ erg s$^{-1}$ for this pulsar, $L_x$ is expected to be well below this upper limit.
Discussion
==========
The blackbody spectrum strongly suggests that the central point source in Kes 79 is not a foreground star or background AGN, as are the other unresolved sources in Figure 1. The central location and similarity to other recently-discovered objects (see list 2 below) indicate that this is probably a neutron star — the remnant of the core of the star which exploded to produce Kes 79. It was not detected in previous X-ray observations because the counting rate of the central source is only $\sim 10^{-2}$ that of the entire remnant in the 1-10 keV energy band. The luminosity at 10 kpc distance, in the band 0.3-8 keV, is $7\times 10^{33}$ erg s$^{-1}$, about 4 times the luminosity as that of the central object in Cas A (Chakrabarty et al. 2001). The blackbody spectrum and lack of a PWN favor thermal emission originating from a small region, perhaps on the surface of the star.
-------------- ----------------------- ------------------------------------ ---------------
Reduced
Model Spectral Index $N_{H}$(10$^{22}$ atoms cm$^{-2}$) $\chi^{2}$
Power Law $\alpha = 4.2\pm0.25$ $2.7\pm0.2$ 1.19 (25 dof)
Black Body $kT=0.48\pm0.025$ $1.45\pm0.17$ 0.92 (25 dof)
2 Black Body $kT=0.18\pm0.06$ 1.8 (fixed) 0.94 (24 dof)
$kT=0.47\pm0.02$
-------------- ----------------------- ------------------------------------ ---------------
Central objects in SNRs were well-understood when the only two known were the Crab and Vela pulsars. For each of these objects, the pulsar characteristic age, P/2P, and the age of the surrounding remnant are about the same (950 years for the Crab and $\sim 10^4$ years for the Vela Remnant), most of the energy radiated from the vicinity of the pulsar is nonthermal, and spin-down energy is adequate to power all emission from the pulsar and PWN. Now that more objects have been discovered, the situation is complex.
Some putative neutron stars within remnants show no sign of non-thermal emission. There is no PWN, the X-ray spectrum is more blackbody than power law, and no radio or gamma-ray pulsations have been observed. These have been called “radio-quiet” pulsars and/or CCOs (Compact Central Objects). Such objects are found in Cas A (Murray et al. 2001; Chakrabarty et al. 2001), Pup A (Zavlin et al. 1999), PKS 1209-51/52 (Zavlin et al. 2000; Pavlov et al. 2002) and G347.3-0.5 (Slane et al. 1999). Properties of these objects are reviewed by Pavlov, Sanwal, Garmire, and Zavlin (2002). The Kes 79 source appears to be of this type. All have luminosities, $L_{X}$ between $10^{33}$ and $10^{34}$ erg s$^{-1}$. There is only one convincing case for pulsed emission; the central source in PKS 1209-51/52 has a period of 0.424 s, a sinusoidal pulse shape, and a pulsed fraction of $\approx10$% (Zavlin et al. 2000; Pavlov et al. 2002).
We note that Kes 79 was searched for radio pulsations by Gorham et al. (1996) at 430 and 1420 MHz. No pulsations were observed above a level of 0.7 mJy at 1420 MHz. (This is rather weak support of the “radio-quiet” classification since a large fraction of pulsars discovered in modern surveys, (e.g. Manchester et al 2001) have lower fluxes than this limit.)
The X-ray spectra of all these objects are close to blackbody spectra. The classical blackbody model, however, predicts a temperature which is too high and a surface area which is too small when compared with generally accepted models of neutron stars. For example, Chakrabarty et al. (2001) have fit the spectrum of the Cas A object and find a temperature of 0.49 keV and star radius of 0.52 km; compared with the 8-16 km radius expected. The radiation source could be a single hot spot on the star surface but this is hard to reconcile with the observed lack of pulsations ($\leq25\%$ pulsed) determined by Murray et al. (2001).
A light-element atmosphere will reduce the derived temperature and increase the derived radius. Lloyd, Hernquist & Heyl (2002) have calculated emergent spectra and find that the actual temperature, $T_{eff}$, is always less than the temperature, $T_{bb}$, derived by fitting a blackbody spectrum to the observed data. Chakrabarty et al. (2000) find that a H atmosphere model applied to the Cas A source yields a temperature of 0.26 kev and a radius of 2.2 km. They conclude that existing surface-radiation models do not explain the Cas A object and that accretion models also fail to account for the lack of an optical counterpart. The more luminous Kes 79 object, if a classical blackbody, would require a radius of only 1.0 km to achieve the observed luminosity at the measured temperature of $T_{bb}$ = 0.48 kev. Application of the light atmosphere model of Lloyd, Hernquist and Heyl predicts that $T_{eff}$ is a factor of 1.8 lower than $T_{bb}$, or 0.27 keV. This lower temperature would increase the required radius to 3.2 km, still short of the 8-16 km expected. Since the theoretical models used so far have been simple, perhaps adjustments might be made in the model atmosphere to achieve a fit with a standard radius. Obviously better models are needed.
Among the 5 Central Compact Objects (CCOs) mentioned, The most luminous is in the largest remnant, G347.3-0.5. The source in Kes 79 is the second most luminous. Kes 79 is about the same size as PKS 1209-51/52 but in a denser environment, so is probably older. Thus the central source in Kes 79 may also be the second oldest specimen of the “radio-quiet” isolated pulsar group. This agrees exactly with the conclusion of Pavlov, Sandwal, & Garmire (2002) that the older CCOs are more luminous. If these central objects are similar, they are certainly not cooling rapidly and the apparent increase in emitting area with age is a mystery.
Certainly, the source in Kes 79 is worthy of more study. It happens to be in a remnant which has a well defined outer shell and, interpreting this as a shock, one can derive information about the SN explosion which produced it. Future work should include sensitive searches for radio and/or X-ray pulsations and, although none is expected, a sensitive search for an optical counterpart.
This work was supported by NASA Grant GO1-2067X.\
Chakrabarty, D., Pivovaroff, M,J., Hernquist, L.E., Heyl, J.S. & Narayan, R. 2001, ApJ 548, 800
Chandra X-ray Center, 2001, Proposers Observatory Guide Rev. 4.0, TD 403.00.004
Frail, D.A. & Clifton T.R. 1989, ApJ 336, 854.
Green P., et al. 2002, in preparation
Grindlay, J., et al. 2002, in preparation
Gorham, P.W., Ray, P.S., Anderson, S.B., Kulkarni, S.R. & Prince, T.A. 1996, ApJ 458, 257
Han, J.L. 1997, A&A 318, 485
Helfand, D.J., Gotthelf, E.V., & Halpern, J.P. 2001, ApJ 556, 380
Space Telescope Science Institute, Guide Star Catalog version 2.2, 2001, http://www-gsss.stsci.edu/gsc/gsc2/GSC2home.htm
Kesteven, M.J.L. 1968, Aust. J. Phys. 21, 369.
Lloyd, D.A., Hernquist, L. & Heyl, J.S. 2002, in preparation Manchester, R.N., et al 2001, MNRAS 328, 17
Murray, S.S., Ransom, S.M., Juda, M., Hwang, U. & Holt, S.S. 2001, Astro-ph 0106516
Pavlov, G.G., Sanwal, D., Garmire, G.P. & Zavlin, V.E. 2002, Astro-ph/0112322, to appear in “Neutron Stars and Supernova Remnants” eds. P.O. Slane & B.M. Gaensler
Pavlov, G.G., Zavlin, V.E., Sanwal, D. & Trumper, J. 2002, ApJ 569, L95
Seaquist, E.R. & Gilmore, W.S. 1982, AJ 87, 378.
Seward, F.D. & Velusamy, T. 1995, ApJ 439, 715.
Slane, P.O., Gaensler, B.M., Dame, T., Hughes, J.P., Plucinsky, P.P. & Green, A. 1999, ApJ 525, 357
Slane, P.O., Chen, Y., Schulz, N., Seward, F., Hughes, J.P., & Gaensler, B. 2000, ApJ 533, L29
Sun, M., Seward, F.D., Slane, P.O. & Smith, R.K. 2002, In preparation.
Sun, M. & Wang, Z-R. 2000, Adv. Space Res. 25, No. 3/4, 549
Toor, A., & Seward, F.D. 1977, ApJ 216, 560.
Velusami, T., Becker, R.H. & Seward, F.D. 1991, AJ 102, 676.
Zavlin, V., Pavlov, G., Sanewal, D. & Truemper, J. 2000, ApJ 540, L25.
Zavlin, V., Truemper, J. & Pavlov, G. 1999, ApJ 525, 959.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We derive stationary solutions to the two-dimensional hyperbolic discrete nonlinear Schrödinger (HDNLS) equation by starting from the anti-continuum limit and extending solutions to include nearest-neighbor interactions in the coupling parameter. We use pseudo-arclength continuation to capture the relevant branches of solutions and explore their corresponding stability and dynamical properties (i.e., their fate when unstable). We focus on nine primary types of solutions: single site, double site in- and out-of-phase, squares with four sites in-phase and out-of phase in each of the vertical and horizontal directions, four sites out-of-phase arranged in a line horizontally, and two additional solutions having respectively six and eight nonzero sites. The chosen configurations are found to merge into four distinct bifurcation events. We unveil the nature of the bifurcation phenomena and identify the critical points associated with these states and also explore the consequences of the termination of the branches on the dynamical phenomenology of the model.'
author:
- 'J. D’Ambroise'
- 'P.G. Kevrekidis'
title: 2D solutions of the hyperbolic discrete nonlinear Schrödinger equation
---
Introduction
============
The hyperbolic nonlinear Schr[ö]{}dinger equation is a model of increasing interest both in applied mathematics and in theoretical/experimental physics [@SuSu; @CoDiTri; @CoTri; @GHS1; @GHS2; @GHS3] as it arises in a diverse host of physical applications. Among others, one can mention as specific examples deep water waves [@abseg; @zakh] and cyclotron waves in plasmas [@sen; @myra], although the equation has been also quite popular in nonlinear optics. Within the latter, the examination of normally dispersive (quasi-discrete) optical waveguide arrays [@Drou; @Lah] has offered a framework for the study of optical pulses. Additionally, the nonlinear, experimentally accessible X-wave structures [@trillo1; @trillo2] (but also more elaborate states including dark-bright [@hay] or vortex-bright [@efr] solitary waves) have motivated its theoretical and numerical study. More recent efforts have also seen the development of methods based on hyperbolic coordinates to study the standing waves of the HNLS [@zeng], a consideration of its universal asymptotic regime for a wide range of initial conditions [@rumanov], as well as the analysis of its profile decomposition in different mass-critical and supercritical cases [@dodson]; see also references therein.
While these extensive studies have addressed numerous aspects of the continuum HNLS model, we are not aware of any efforts considering the (genuinely) discrete aspects of the model, the so-called hyperbolic discrete nonlinear Schr[ö]{}dinger or HDNLS equation. This is an interesting endeavor on a number of counts. On the one hand, the elliptic variant of the discrete problem is quite well understood (see, e.g., the monograph [@dnls]), hence, it is conceivable that some of the corresponding analytical and numerical techniques may be adapted to the present setting. In fact, there exists a so-called staggering transformation $u_{n,m}=(-1)^n w_{n,m}$ (where $w_{n,m}$ is the solution of the elliptic problem) which can convert the former to the latter. However, we will not use this approach here, given that this transformation becomes singular in the continuum limit. Instead, we note that the phenomenology of the HDNLS model is of interest not only given its consideration as a numerical scheme for the continuum HNLS, but also because some of the applications may bear a(n at least partially) discrete character, as discussed, e.g., in [@Drou; @Lah].
There are additional characteristics that add to the appeal of the HDNLS model. For instance, in the so-called anti-continuum (AC) limit of vanishing coupling between adjacent nodes, [*any*]{} stationary configuration is “permissible” as is known also for the elliptic case [@dnls]. However, in the continuum limit, on the other hand, the work of [@GHS3] established that there are no nontrivial standing wave solutions. This implies that [*all*]{} the solutions initiated at the AC limit must terminate at some point prior to reaching the continuum i.e., at some finite value of the coupling strength. This is fundamentally different from the standard elliptic DNLS case, where solitary waves, and even vortical solutions may persist in the continuum limit [@SuSu]. Thus, it is of interest to explore the bifurcations through which these branches terminate and to classify the dynamical behavior of the model prior to, as well as past the corresponding critical points. It is the aim of the present work to address a number of these issues for some of the most fundamental (one- and few-site) configurations of the HDNLS model.
Our presentation will be structured as follows: In section II, we will explore theoretical aspects of the existence (via solvability conditions) and stability theory (linearizing around the equilibrium configurations and exploring the corresponding spectrum). Then, numerical computations will be used in section III to corroborate the analytical existence/stability results and direct numerical simulations will help us determine the fate of such waveforms when unstable (or when they may not exist closer to the continuum limit). Finally, in section IV, we will summarize our findings and present our conclusions, as well as a number of challenges towards future work.
Model
=====
We consider the HDNLS equation for $u_{n,m}(z)$ as follows $$i \frac{du_{n,m}}{dz}+\epsilon \Delta_H u_{n,m} +|u_{n,m}|^2 u_{n,m}=0 \label{dyneq}$$ where $\Delta_H u_{n,m} = u_{n,m+1}+u_{n,m-1}-u_{n+1,m}-u_{n-1,m}$ stands for the hyperbolic operator, i.e., a discretization of $u_{xx}-u_{yy}$ with unit spacing, while $\epsilon$ is the nearest neighbor coupling parameter. In the context of this being a(n isotropic) discrete approximation to the continuum problem, one should think of $\epsilon=1/\Delta x^2$, where $\Delta x$ is the spacing between adjacent lattice nodes in both directions. The indexing $n$ represents the discrete vertical direction and $m$ the horizontal one. Setting $u_{n,m}(z)=\phi^{(\epsilon)}_{n,m}e^{i\mu z}$ we obtain the stationary equation $F^{(\epsilon)}_{n,m}=0$ for $$F^{(\epsilon)}_{n,m}(\phi) \stackrel{def.}{=} (\mu-|\phi^{(\epsilon)}_{n,m}|^2)\phi^{(\epsilon)}_{n,m}-\epsilon \Delta_H \phi^{(\epsilon)}_{n,m}.
\label{stateq}$$ We can then seek standing wave solutions with frequency $\mu$, by solving the algebraic set of Eqs. (\[stateq\]).
Existence of Solutions
----------------------
In the $\epsilon = 0$ anti-continuum limit, the values of $\phi^{(0)}_{n,m}$ for each site $\{n,m\}$ can be chosen independently from each other since the nearest-neighbor coupling parameter is zero. Localized solutions are thus found by specifying $\phi^{(0)}_{n,m}=0$ for most sites $\{n,m\}$. For a few nonzero sites we set $\phi^{(0)}_{n,m} = e^{i\theta_{n,m}}$ with $\mu = 1$ and $\theta_{n,m}\in\{ 0, \pi\}$. Table I lists some possible solutions for $\phi^{(0)}_{n,m}$ and a naming convention for each example configuration.
We will use the following general notation. Let $\mathcal{S}=\{ \phi^{(0)}_{n_1,m_1}, \phi^{(0)}_{n_2,m_2}, ..., \phi^{(0)}_{n_d,m_d}\}$ for $d\in\mathds{N}$ represent an enumeration of the nonzero sites of the initial $\epsilon = 0$ configuration, and let $\vec{\theta}={\rm arg}(\mathcal{S})\in [0,2\pi]^d$ represent a vector whose elements are the arguments of elements of $\mathcal{S}$. For simplicity we enumerate the nonzero sites in a natural way with the top-most left nonzero site corresponding to the first index. Notice that the configurations listed in Table I are not necessarily closed loops, but when they are we enumerate from the top left then counterclockwise. It will be convenient to denote $\vec{\delta}_L$ as a vector whose components are either $0$ or $1$ corresponding to whether the left neighbor (when considered on the full two-dimensional grid) $\phi^{(0)}_{n_j,m_j-1}$ of each element of $\mathcal{S}$ is zero or nonzero. Similarly define $\vec{\delta}_R, \vec{\delta}_T, \vec{\delta}_B$ corresponding to whether the right ($\phi^{(0)}_{n,m+1}$), top ($\phi^{(0)}_{n-1,m}$), and bottom ($\phi^{(0)}_{n+1,m}$) neighbors of each element of $\mathcal{S}$ are zero or nonzero on the 2D grid. Finally, let $\vec{\theta}_L, \vec{\theta}_R, \vec{\theta}_T, \vec{\theta}_B\in [0,2\pi]^d$ denote the arguments of the corresponding nonzero nearest neighbors to each element of $\mathcal{S}$ (with the subscripts having the same neighbor designation as above). Note that since $\mathcal{S}$ contains all of the nonzero elements of $\phi^{(0)}_{n,m}$, the vectors $\vec{\theta}_\star$ are permutations of $\vec{\theta}$.
[cc]{}\
&\
\
1s &
+
---
: Solutions $\phi^{(0)}_{n,m}$ to the stationary equation (\[stateq\]) for $\mu = 1$ are listed with corresponding naming convention for each type. Nonzero sites of the configuration are shown with the values $\phi^{(0)}_{n,m} = \pm 1$ denoted as plus or minus. All other sites are zero. Solutions are grouped by branch number, i.e. according to which merge after continuing in $\epsilon $.
\
\
2i-horz &
[|>p[0.25cm]{}|>p[0.25cm]{}|]{}+ & +
\
\
4o-vert &
[|>p[0.25cm]{}|>p[0.25cm]{}|]{}+ & + – & –
[cc]{}\
&\
\
2o-horz &
[|>p[0.25cm]{}|>p[0.25cm]{}|]{}+ & –
\
\
4o-line &
[|>p[0.25cm]{}|>p[0.25cm]{}|>p[0.25cm]{}|>p[0.25cm]{}|]{}+ & + & – & –
\
\
\
\
[cc]{}\
&\
\
4i-sqr &
[|>p[0.25cm]{}|>p[0.25cm]{}|]{}+ & + + & +
\
\
8s &
[|>p[0.25cm]{}|>p[0.25cm]{}|]{}– & – + & + + & + – & –
[cc]{}\
&\
\
4o-horz &
[|>p[0.25cm]{}|>p[0.25cm]{}|]{}– & + – & +
\
\
6s &
[|>p[0.25cm]{}|>p[0.25cm]{}|]{} + & – & + – & + + &
For $\epsilon > 0$ real-valued solutions $\phi^{(\epsilon )}_{n,m}$ are computed from $\phi^{(0)}_{n,m}\in\mathds{R}$ via continuation in the coupling parameter $\epsilon$. Such solutions satisfying the limit $\displaystyle\lim_{\epsilon\rightarrow 0} \phi^{(\epsilon)}_{n,m} = \phi^{(0)}_{n,m}$ are unique and guaranteed to exist for $\epsilon$ in some a neighborhood $I_0=(-\epsilon_0,\epsilon_0)$ by an application of the implicit function theorem. From the stationary equation $F^{(\epsilon)}_{n,m}=0$ one can directly compute the solvability condition ${\rm Im}\left(\phi^{(\epsilon)}_{n,m}\overline{F}^{(\epsilon)}_{n,m}(\phi)\right) = 0$ where overline represents the complex conjugate. That is, solutions $\phi^{(\epsilon)}_{n,m}$ are also roots of $\vec{g}=[g_{n,m}]$ for elements defined as $g_{n,m}=$ $$\epsilon {\rm Im}\left( \ \phi_{n,m}\left( \overline{\phi}_{n,m+1}+\overline{\phi}_{n,m-1}-\overline{\phi}_{n+1,m}-\overline{\phi}_{n-1,m}\right) \right).
\nonumber
\label{gdef}$$ Considered as a(n implicit) function of $\vec{\theta}$ and $\epsilon$, the vector function $\vec{g}$ can be expanded in Taylor series that is convergent on the interval $I_0$ [@dnls; @ChowHale; @GS]. That is, $$\vec{g}(\vec{\theta},\epsilon) = \displaystyle\sum_{k=1}^\infty \epsilon^k \vec{g}^{(k)}(\vec{\theta}) \mbox{ where } \vec{g}^{(k)}(\vec{\theta}) = \frac{1}{k!}\partial^k_\epsilon \vec{g}(\vec{\theta},0).\label{taylor}$$
Since the initial configuration at $\epsilon = 0$ exhibits a gauge invariance $\vec{\theta} \rightarrow \vec{\theta} + \theta_0$ for $\theta_0\in\mathds{R}$ this gives a one parameter family of roots of $\vec{g}$ for any fixed $\epsilon\in I_0$. This implies that if the first order Jacobian matrix $J=\partial \vec{g}^{(1)}/\partial \vec\theta$ has a simple zero eigenvalue, there exists a unique (modulo gauge transformation) analytic continuation of the limiting solution $\phi^{(0)}$ into the domain $I_0$ [@dnls; @ChowHale; @GS]. Having provided the conditions for the existence of the different branches of solutions, we now turn to their corresponding spectral stability analysis.
Spectral Stability
------------------
For each example solution $\phi^{(\epsilon)}_{n,m}$ in Table I the stability is monitored for each fixed $\epsilon>0$ via the linearization ansatz $$\label{pert}
u = e^{i\mu z}\left( \phi^{(\epsilon)}_{n,m} + \delta \left[ a_{n,m}e^{\nu z} + b^*_{n,m}e^{\nu^*z}\right] \right)$$ which yields the order $\delta$ linear system $$\left[ \begin{array}{cc} M_1 & M_2 \\ -M_2^* & -M_1^* \end{array} \right] \left[ \begin{array}{c} a\\ b \end{array} \right] = -i \nu \left[ \begin{array}{c} a\\ b \end{array} \right]\label{mat}$$ where $M_1 = \epsilon \Delta - \mu + 2|\phi^{(\epsilon)}|^2$ and $M_2 = \left(\phi^{(\epsilon)}\right)^2$. Thus max(Re$(\nu)$) $> 0$ corresponds to instability, yielding the relevant instability growth rate, while max(Re$(\nu)$) = 0 corresponds to (neutral) stability. Note that $\psi = [a \ b]^T$ represents a column vector of length $2N^2$, where $N\times N$ is the two-dimensional grid size.
In the numerical computations that follow, we identify the relevant solutions via fixed point iterations and subsequently solve numerically the matrix eigenvalue problem of Eq. (\[mat\]) to determine their stability. However, it is particularly useful to have some theoretical prediction/expectation about which configurations should be expected to be stable and which ones should not. To that effect, we adapt the methodology summarized in [@dnls] (based on earlier works such as [@pkf1; @pkf2]). This allows us to connect the stability of the few-site configurations with the Jacobian of the solvability conditions, as follows.
From equations (\[taylor\]) and using the notation of Section A we may write the bifurcation function $\vec{g}^{(1)} $ as follows: $$\begin{aligned}
\vec{g}^{(1)}(\vec{\theta}) &=& \vec{\delta}_{L}\sin(\vec{\theta} - \vec{\theta}_L)+ \vec{\delta}_{R}\sin(\vec{\theta} - \vec{\theta}_R)\\
&& - \vec{\delta}_{T}\sin(\vec{\theta} - \vec{\theta}_T) - \vec{\delta}_{B}\sin(\vec{\theta}_j - \vec{\theta}_B)\nonumber,\end{aligned}$$ where we intend the equation to be considered element-wise in each of the excited sites.
Thus the first order Jacobian matrix $J=\partial \vec{g}^{(1)}/\partial \vec\theta$ has entries that can be computed manually given any example solution. The diagonal vector of the matrix $J$ is $ \delta_{L}\cos({\theta} - {\theta}_{L})+\delta_{R}\cos({\theta} - {\theta}_{R}) - \delta_{T}\cos({\theta} - {\theta}_{T}) - \delta_{B}\cos({\theta} - {\theta}_{B})$. Non-zero off-diagonal entries are of the form $\pm\cos(\theta_j-\theta_{\star,j})$ where the index of the nonzero entry is the index of nonzero entries of $\delta_\star$ for each of $\star = L,R,T,B$ with the plus sign corresponding to $T,B$ and the minus sign corresponding to $L,R$. The eigenvalues $\lambda_i$ of $J$ are then connected to the full stability problem via the relation $\displaystyle\lim_{\epsilon\rightarrow 0} \nu_i^2/\epsilon = 2\lambda_i$, with the relevant proof going through in a same way as with the elliptic case of [@dnls].
[ccc]{} & &\
\
1s & $\left( 0 \right)$ & $\{0\}$\
\
2i-horz & $ \left(\begin{array}{cc} 1 & -1 \\ -1 & 1 \end{array}\right) $&$ \{2,0\}$\
\
4o-vert & $ \left(\begin{array}{cccc} 2 & -1 & 0 & -1 \\ -1 & 2 & -1 & 0\\ 0 & -1 & 2 & -1\\ -1& 0 & -1 & 2\end{array}\right)$ &$\{4,2,2,0\}$\
\
2o-horz & $ \left(\begin{array}{cc} 1 & -1 \\ -1 & 1 \end{array}\right) $&$ \{-2,0\}$\
\
4o-line & $ \left(\begin{array}{cccc} 1 & -1 & 0 & 0 \\ -1 & 0 & 1 & 0\\ 0 & 1 & 0 & -1\\ 0& 0 & -1 & 1\end{array}\right)$ &$\{2,\pm\sqrt{2},0\}$\
\
4i-sqr & $ \left(\begin{array}{cccc} 0 & 1 & 0 & -1 \\ 1 & 0 & -1 & 0\\ 0 & -1 & 0 & 1\\ -1& 0 & 1 & 0\end{array}\right)$ &$\{\pm 2,0,0\}$\
\
8s & $\left(\begin{array}{cccccccc} 2& -1& 0 & 0& 0& 0 &0 & -1\\ -1 & 1 & 1 & 0 & 0 & 0 & -1 & 0 \\ 0& 1 & 1 & -1 & 0 & -1 & 0 & 0\\
0& 0& -1 & 2 & -1 & 0 & 0 & 0 \\ 0& 0&0&-1 &2 & -1 & 0 &0 \\ 0 & 0& -1 & 0 & -1 & 1 & 1 & 0 \\ 0 & -1 & 0 & 0 & 0 & 1 & 1 & -1\\ -1 & 0& 0& 0& 0& 0& -1 & 2 \end{array}\right) $& $\begin{array}{c}\{4, 2 \pm \sqrt{2}, \\ \ \ \ \ \ \ 2, 2, \pm \sqrt{2}, 0 \}\end{array}$\
\
4o-horz & $ \left(\begin{array}{cccc} -2 & 1 & 0 & 1 \\ 1 & -2 & 1 & 0\\ 0 & 1 & -2 & 1\\ 1& 0 & 1 & -2\end{array}\right)$ &$\{-4,-2,-2,0\}$\
\
6s & $\left(\begin{array}{cccccc} 1 & -1 & 0& 0& 0& 0 \\ -1 & -1 & 1 & 0 & 1 & 0 \\ 0 & 1 & -2 & 1 & 0 & 0\\ 0 & 0 & 1 & -2 & 1 & 0 \\ 0 & 1 & 0 & 1 & -1 & -1 \\ 0& 0& 0& 0& -1 & 1 \end{array}\right)$ &$\begin{array}{c} \{ -3.68133, -1.64207, \\ \ \ \ \ \ \ \pm\sqrt{3}, 1.3234, 0\} \end{array}$
The Jacobian matrix for each example configuration that we consider is listed in Table II. Based on the eigenvalues $\{\lambda_i\}$ listed in Table II the 1s, 2o-horz, and 4o-horz configurations are found to be stable for very small $\epsilon$. For such small $\epsilon$ we find one unstable direction for configurations 2i-horz and 4i-sqr; two for 4o-line and 6s; three for 4o-vert; and six for 8s. [Note that adjacent in-phase excitations along the horizontal direction such as 2i-horz, 4i-sqr, and 8s lead to instability, as well as out-of-phase excitations in the vertical direction such as 4o-vert. ]{} For more complex configurations (like 6s or 8s), whether or not they will bear an instability depends on whether they include such unstable “base ingredients” i.e., any in phase pair along the horizontal (as is the case for 8s) or out of phase pair along the vertical (as is the case for 6s). Having the analytical predictions of Table II at hand, we now turn to a numerical exploration of the corresponding (potential) instabilities.
Numerical Results: Existence, Stability and Dynamics
====================================================
In the Appendix we provide a short pseudocode algorithm for the arclength continuation procedure that we utilize in order to identify the relevant branches of solutions numerically. The power of the resulting solutions $P(\epsilon) = \sum_{n,m} |\phi^{(\epsilon)}_{n,m}|^2$ is plotted as a function of $\epsilon > 0$ in Figure \[pows\]. The figure shows that the power curves merge into four bifurcation “events”. The branch labels are indicated in the caption of Figure \[pows\] as ordered from lowest to highest power.
In Figures \[newt1\]-\[newt4\] the sample solutions $\phi^{(\epsilon)}_{n,m}$ on the left columns show typical branch members while the result of the two-dimensional continuation procedure over $\epsilon$ is shown on the right panel through the unstable eigendirection growth rates Re$(\nu)$. Comparing Table I to the left columns of Figures \[newt1\]-\[newt4\] we find that the extended solutions originate from the $\epsilon = 0$ solutions in the following manner as the nearest neighbor interaction is turned on for $\epsilon > 0$. Generally sites to the left and right of the initial configuration become nonzero with the same sign as the initial configuration, and sites neighboring the initial configuration vertically become nonzero with opposite sign as the initial configuration. Signs of the sites then alternate vertically and stay the same horizontally in a manner that respects the signs of the initial configuration as the footprint continues to expand for increasing $\epsilon > 0$.
Solutions initiated at $\epsilon = 0$ from the configurations types of 1s, 2i-horz, and 4o-vert merge into a single branch we denote as Branch 1. These three solution types merge at $\epsilon\approx 0.242$ and the left column of Figure \[newt1\] shows example solutions for $\epsilon = 0.2414$ on Branch 1. In the right column of Figure \[newt1\] we plot the nonzero real parts of eigenvalues $\nu$ as computed from equation (\[mat\]). Note that the eigenvalue plots denote real eigenvalues with an “x“ mark and nonreal complex eigenvalues with an ”o" mark. For very small $\epsilon$ the prediction $\nu \approx \sqrt{2\lambda \epsilon}$ from Section B is plotted in a red dashed line based on the values of $\lambda$ in Table II. Clearly, the one unstable eigenvalue of the 2i-horz configuration and the two unstable eigenvalues of the 4o-vert configuration are well captured for small $\epsilon$. As $\epsilon$ increases, however, these real eigenvalue pairs appear to turn around towards $\nu=0$ and tend the origin of the spectral plane as the bifurcation point is approached.
One can follow the eigenvalue diagrams in Figure \[newt1\] with the following description of the change of the eigenvalue types as a function of $\epsilon$. Starting with the 1s configuration, near the merging point there is one nonzero real pair of eigenvalues within the range $0.201\leq \epsilon \leq 0.241$. [ Near the merge point this one real pair of eigenvalues moves towards the origin. Additionally two pairs of eigenvalues on the imaginary axis tend toward the origin (these are not reflected in Figure \[newt1\]).]{} The 2i-horz has one nonzero real pair for $0<\epsilon\leq 0.211$ and two real pairs for $0.212\leq \epsilon \leq 0.241$. [ Near the merge point the two real pairs and one additional imaginary pair approach the origin.]{} The 4o-vert has initially a total of three nonzero pairs (one pair coinciding for a total of two distinct) within the range $0<\epsilon\leq 0.205$, then within the range $0.206 \leq \epsilon \leq 0.229$ two real pairs, and the remaining interval $0.230\leq \epsilon \leq 0.241$ again three nonzero pairs (one pair coinciding, two distinct). [ Nearest to the merge point the three real eigenvalues decrease in amplitude towards the origin.]{}
Solutions initiated at $\epsilon = 0$ from the configurations types 2o-horz and 4o-line merge at $\epsilon\approx 0.226$ as Branch 2. Figure \[newt2\] shows example solutions for $\epsilon = 0.2258$ in the left column. Note that according to Table II the 2o-horz type is initially (i.e., for small $\epsilon$) stable and the 4o-line type initially has two unstable directions for small $\epsilon$. The predictions $\nu\approx\sqrt{2\lambda\epsilon}$ according to Table II are plotted as red dotted lines in the right column of Figure \[newt2\]. For small $\epsilon$, the two unstable eigendirections of 4o-line are well captured. As $\epsilon$ is increased, in the right column one can follow the eigenvalue changes over $\epsilon$. The 2o-horz type has one quartet with nonzero real part within the range $0.083\leq \epsilon \leq 0.225$ and then within the range $0.210\leq \epsilon \leq 0.225$ there is an additional real pair. The 4o-line type has two real pairs for $0<\epsilon \leq 0.225$ and an additional quartet for $0.101\leq \epsilon\leq 0.225$. Near the merge point for Branch 2 the magnitude of the real parts of the quartet eigenvalues from both the 2o-horz and the 4o-line approach approximately $0.05$ as $\epsilon\rightarrow 0.226$. The merge point also has one real pair with magnitude approximately $0.38$. Notice that the 2o-horz type has the real pair [*increasing*]{} toward this value while the 4o-line type has its largest real pair [*decreasing*]{} toward this value (the smaller real pair for 4o-line type goes to zero). Moreover, to confirm the saddle-center nature of this bifurcation, the 2nd real pair of the 4o-line branch decreases towards the origin $\nu=0$ as the bifurcation point is approached, while the 2o-horz branch has an imaginary eigenvalue pair (not shown here) tending to collide with this real pair (of 4o-line) at the origin.
Solutions initiated from 4i-sqr and 8s merge at $\epsilon\approx 0.210$ as Branch 3 and Figure \[newt3\] shows example solutions for $\epsilon = 0.2101$. Note that according to Table II the 4i-sqr type has one unstable direction and 8s has six unstable directions for small $\epsilon$. These predictions are plotted as red dotted lines in the right column of Figure \[newt3\], again in good agreement with the numerical results at least for small values of $\epsilon$, before turning around towards $\nu=0$, which in this case too happens around $\epsilon=0.1$. The eigenvalue types change over $\epsilon$ as follows. The type 4i-sqr has a real pair for $0<\epsilon\leq 0.210$, a second real pair for the interval $0.056\leq\epsilon\leq 0.210$ and a third real pair for the interval $0.203\leq\epsilon\leq 0.210$. Note also that a quartet appears in the interval $0.088\leq\epsilon\leq 0.210$. The type 8s has six real pairs for most of the $\epsilon$ range with two persisting for the smaller interval $0<\epsilon\leq 0.199$ and four persisting (two overlapping) for the whole interval $0<\epsilon\leq 0.210$. Additionally a quartet exists for $0.108\leq \epsilon\leq 0.210$. On both types 4i-sqr and 8s the largest real pair approaches a magnitude of approximately $0.73$ as $\epsilon\rightarrow 0.2101$ and the complex quartet has the magnitude of its real part approaching 0.02. On both types 4i-sqr and 8s, focusing on the smaller real pairs of eigenvalues, they approach a magnitude of either $0.24$ or $0.31$. Notice that the 8s configuration has at this point two extra real pairs (the overlapping pair) unaccounted for thus far – they approach zero as the merge point nears. These are the eigenvalues responsible for this saddle-center bifurcation, while the 4i-sqr branch has two corresponding pairs tending to $\nu=0$ from the imaginary side.
Solutions initiated from 4o-horz and 6s merge at $\epsilon\approx 0.251$ as Branch 4 in the final example among our saddle-center bifurcations. Figure \[newt4\] shows example solutions for $\epsilon = 0.2509$. Table II predicts that 4o-horz is initially stable and 6s initially has two unstable directions. The 4o-horz type has one complex quartet for $0.059\leq \epsilon \leq 0.082$ and three such for $0.083\leq \epsilon\leq 0.096$ (two coinciding), while two coinciding ones remain for $0.097\leq \epsilon\leq 0.185$. [Near $\epsilon = 0.185$ these complex eigenvalues rapidly return to the imaginary axis.]{} A real pair of eigenvalues exists close to the merging point in the interval $0.237\leq\epsilon \leq 0.251$ and a second real pair appears very near that merge point for $0.248\leq\epsilon\leq 0.251$. The 6s type has two real pairs of eigenvalues for $0 < \epsilon \leq 0.251$ and an additional quartet for $0.061\leq \epsilon\leq 0.091$ and a total of three quartets for $0.092\leq\epsilon\leq 0.102$; then, it has a total of two quartets for $0.103\leq \epsilon \leq 0.197$ and one quartet for $0.198\leq \epsilon\leq 0.217$. [Near $\epsilon = 0.217$ the non-real quartets eigenvalues rapidly return to the imaginary axis.]{} Near the merging point for Branch 4 the magnitudes of the two real pairs of eigenvalues [*decrease*]{} towards the limiting magnitude value of $0.02$ and $0.23$ as $\epsilon\rightarrow 0.251$ for the 6s configuration while the two real pairs of eigenvalues on the 4o-horz side of the branch [*increase*]{} towards those same magnitude values as $\epsilon\rightarrow 0.251$. Thus, in the vicinity of this point, the two configurations collide and merge through the associated turning point of this final saddle-center bifurcation.
![The power $P(\epsilon) = \sum_{n,m} |\phi^{(\epsilon)}_{n,m}|^2$ of solutions to the stationary equation (\[stateq\]), obtained by arclength continuation, plotted as a function of $\epsilon>0$. The continuation is initiated at $\epsilon = 0$ with the various solutions listed in Table I. Underneath each branch segment is a label representing the $\epsilon=0$ configuration from which the solution is continued. Dashed lines represent unstable solutions and solid ones represent stable solutions. Branch 1 (lowest power) exists up to the value of $\epsilon \approx 0.242$ where solutions originating from types 1s, 2i-horz, and 4o-vert meet. Branch 2 (lower middle) exists up to $\epsilon\approx 0.226$ at which point the solutions originating from types 2o-horz and 4o-line meet. Branch 3 (upper middle) exists up to $\epsilon\approx 0.210$ where types 4i-sqr and 8s meet. Branch 4 (highest power) exists up to $\epsilon \approx 0.251$ where types 4o-horz and 6s meet. []{data-label="pows"}](pows.png){width="\columnwidth"}
![Each configuration plotted in the left column shows a Branch 1 (real-valued) solution that was obtained by continuation of the coupling parameter to the value of $\epsilon= 0.2414$ where the original configuration at $\epsilon = 0$ is 1s (top), 2i-horz (middle), or 4o-vert (bottom). In the right column the set of values {Re$(\nu) > 0$} is plotted for the corresponding whole branch segment versus $\epsilon$ where similarly the original configuration at $\epsilon = 0$ is 1s (top), 2i-horz (middle), or 4o-vert (bottom). For the right column, circles mark the values for which Im$(\nu)>0$ (i.e. $\nu$ is complex non-real and existing as quartets in the complex plane) and x’s mark the values for which $\nu\in\mathds{R}$ (i.e. $\nu$ is real and existing as pairs on the real axis of the complex plane). []{data-label="newt1"}](newt1.png){width="\columnwidth"}
![Plots are similar to Figure \[newt1\] but here for Branch 2 with the left column corresponding to $\epsilon=0.2258$ with initial $\epsilon=0$ configurations here as 2o-horz (top row) and 4o-line (bottom row). []{data-label="newt2"}](newt2.png){width="\columnwidth"}
![Plots are similar to Figure \[newt1\] but here for Branch 3 with the left column corresponding to $\epsilon=0.2101$ with initial $\epsilon=0$ configurations here as 4i-sqr (top row) and 8s (bottom row). []{data-label="newt3"}](newt3.png){width="\columnwidth"}
![Plots are similar to Figure \[newt1\] but here for Branch 4 with the left column corresponding to $\epsilon=0.2509$ with initial $\epsilon=0$ configurations here as 4o-horz (top row) and 6s (bottom row). []{data-label="newt4"}](newt4.png){width="\columnwidth"}
In Table III the dynamical fate of some case example solutions is listed together with the type of the perturbing eigenvectors. The 1s type solution at $\epsilon=0.2250$ evolves towards a single expanding mass marked as 1m in the table. Figure \[rk1\] shows the original solution at $z=0$ in the top right panel, the evolved solution at a later $z$ value in the bottom right panel, and the corresponding maximal eigenvalue and eigenvector in the top left and bottom left respectively. The result of the evolution in the bottom right clearly illustrates the dispersive nature of the temporal dynamics.
Table III shows that at $\epsilon = 0.2096$ the solution of type 2o-horz has two different fates depending on whether one perturbs in the eigendirection corresponding to the maximal real eigenvalue (marked as the Im$(\nu)=0$ column of the table) versus the other eigendirection corresponding to the maximal complex eigenvalue (marked as the Im$(\nu)\neq 0$ column of the table). In the former case the 2o-horz type solution evolves towards an expanding mass with two “blobs" moving outwards along the horizontal direction; this is marked as 2m in the table and shown in the middle right panel of Figure \[rk2\] with the corresponding eigenvector shown in the middle left panel. In the latter case such expansion is not symmetric – this is marked as 1-2m in the table and shown in the bottom right panel of Figure \[rk2\] with the corresponding eigenvector shown in the bottom left panel. Apparently here the perturbation added on top of the initial 2o-horz configuration breaks its symmetry, leading to the asymmetric evolution of the bottom right of Fig. \[rk2\].
For other configurations such as 4o-line, 4i-sqr and 8s, according to Table III, their instabilities typically led to a single-site resulting evolution for the values of $\epsilon$-considered (for which the single site configuration was dynamically stable). [In the cases of 2i-horz and 4o-vert for $\epsilon=0.225$, the configurations approach a transient 1s state, vibrating near a 1s type solution with a pulsating core, since the stationary 1s configuration is unstable for this value of $\epsilon$.]{} On the other hand, at $\epsilon = 0.2400$ the 4o-horz and 6s configurations evolve with mass expanding mostly towards the four corners, as is demonstrated in Figure \[rk4\]. Indeed, all other solutions shown in Table III revert towards the 1s type, i.e., disperse mass while transforming to a single site excited configuration. We additionally tested solutions on Branches 1-4 when propagated according to equation (\[dyneq\]) with $\epsilon$ beyond the bifurcation points such as $\epsilon = 0.28, 0.3$ to find that all tested standing wave solutions disperse for such higher $\epsilon$ values, closer to the continuum limit. This is in line with the expectation that no coherent structure exists in the vicinity of the continuum limit. Yet, our quantitative analysis illustrates that dispersion dominates already for rather weak couplings i.e., $\epsilon > 0.25$. Whether a discrete analogue of self-similar dynamics arises for this interval (corresponding to the continuum observations of [@rumanov]) is an interesting open question for future study.
![The top left panel shows the values $\{\nu\}$, defined according to equations (\[pert\]) and (\[mat\]), plotted in the complex plane, for the corresponding stationary solution $\phi^{(\epsilon)}_{n,m}$ that is plotted in the top right panel. The solution at $z=0$ is of type 1s for $\epsilon = 0.2250$ and it is real-valued. The eigenvector $[a \ b]^T$ corresponding to the positive real $\nu$ value is used to perturb the solution according to equation (\[pert\]) and $|a+b^*|$ is plotted in the bottom left panel. After propagating according to equation (\[dyneq\]) the result $|\phi^{(\epsilon)}_{n,m}(z)|$ is plotted in the bottom right for $z=50$. []{data-label="rk1"}](RKbranch1.png){width="\columnwidth"}
![Plots are similar to Figure \[rk1\] where the top right panel shows the real-valued solution of type 2o-horz at $z=0$ for $\epsilon = 0.2096$. The middle left panel shows $|a+b^*|$ for $[a \ b]^T$ the eigenvector corresponding to the eigenvalue with maximum real part among the quartet of eigenvalues (with Im$(\nu)\neq 0$). Perturbing according to (\[pert\]) with this eigenvector results in solution $|\phi^{(\epsilon)}_{n,m}(z)|$ in the middle right panel plotted at $z=218$. The bottom row of panels is similar but for eigenvector corresponding to the real eigenvalue with maximum real part among the real eigenvalues (with Im$(\nu)=0$). Perturbing according to (\[pert\]) with this eigenvector results in solution $|\phi^{(\epsilon)}_{n,m}(z)|$ in the bottom right panel plotted at $z=122$. []{data-label="rk2"}](RKbranch2.png){width="\columnwidth"}
![Plots are similar to Figure \[rk1\] where the top right panel shows the real-valued solution of type 4o-horz at $z=0$ for $\epsilon = 0.2400$. The middle left panel shows $|a+b^*|$ for $[a \ b]^T$ the eigenvector corresponding to the eigenvalue with maximum real part. Perturbing according to (\[pert\]) with this eigenvector results in the solution $|\phi^{(\epsilon)}_{n,m}(z)|$ plotted in the bottom right panel at $z=116$. []{data-label="rk4"}](RKbranch4.png){width="\columnwidth"}
\[my-label\]
[ >p[1.5cm]{} | >p[1.5cm]{} | >p[1.5cm]{} | >p[1.5cm]{} | >p[1.5cm]{} | ]{} & $Im(\nu) = 0 $ & $Im(\nu) \neq 0 $ & $Im(\nu) = 0 $ & $Im(\nu) \neq 0$ & & & (stable) & – & 1m &– & 1s & – & 1s-trans & – & 1s &– & 1s-trans & – & & & – & 1s &2m & 1-2m & 1s &1s & 1s-trans & 1s-trans & & & 1s &1s & & & 1s & 1s & & & & &– &1s & 4m & – & 1s & 1s & 4m & –
Conclusions & Future Challenges
===============================
In the present work, we have explored some of the fundamental solutions of the hyperbolic discrete nonlinear Schr[ö]{}dinger model. We have initiated our search for such waveforms at the convenient anti-continuum limit and have used continuation in the coupling parameter for some of the most prototypical ones, most notably one-, two- and four-site ones, with some exceptions of involving six- and eight-site ones, when these were participating in bifurcations involving the lower number of site branches. We have adapted the solvability condition methodology of the elliptic case to this hyperbolic one and have accordingly derived existence conditions and predictions for the eigenvalues of the linearization of such few-site configurations. Subsequently, we obtained the states via fixed point iterations and examined the validity of the analytics as a function of the coupling strength $\epsilon$. It was generally found that the eigenvalue predictions worked well in the vicinity of the anti-continuum limit. However, at larger values of the coupling (typically of $\approx 0.2$), the eigenvalues were found in many configurations to “turn around” and either meet up with a merging segment of the branch or return to the origin leading to a set of bifurcation patterns that were elucidated herein, some in fact involving more than 2 configurations (as was the case with the branches 1s, 2i, 4o-vert). This aligns itself with our expectation that [*all*]{} standing wave solutions disappear in the continuum limit [@GHS3]. Sufficiently beyond these critical bifurcation thresholds (all of which satisfied $0.2 \leq \epsilon_{cr}
\leq 0.25$ for the examples considered), the fate of standing wave-like initial conditions was also examined and it was found that they disperse, forming one or more dispersing “blobs”, depending on the form of the initial condition. Interestingly, this type of fate (of dispersion into one or multiple blobs) could arise for select examples before the bifurcation critical points, as elaborated in Table III. Nevertheless, in numerous cases of the latter scenario, the configurations just rearranged themselves towards eventually reaching a single site state ($1s$).
Naturally, the present work paves the way for the numerous intriguing questions both at the theoretical and at the numerical level. A difficult set of questions concerns the phenomenology around $\epsilon_{cr}$. Our analysis enables an understanding for small $\epsilon$; is there, however, a way to capture the “turning” of the eigenvalues or the emergence of these bifurcations around these critical values of $\epsilon$ ? Beyond these critical values, does one encounter a discrete variant of the universal regimes presented in [@rumanov] and if so is there a way to analyze such phenomenology at the discrete level ? Finally, extending considerations to the 3-dimensional setting with two directions bearing the same sign of the dispersion (diffraction, at the discrete level) and one the opposite would be a possibility of interest in its own right. Some of the questions are presently under consideration and will be reported in future publications.
[*Acknowledgements.*]{} The authors acknowledge early efforts in this direction by Dr. Kai Li. P.G.K. is also grateful to Profs. M.J. Ablowitz and Boris A. Malomed for illuminating discussions on the subject. This material is based upon work supported by the National Science Foundation under Grant No. DMS-1809074 (P.G.K.).
Appendix {#appendix .unnumbered}
========
In Algorithm I an initial solution $\phi^{(0)}_{n,m}$ for $\epsilon_0 = 0$ from Table I is assumed to be represented as a column vector of length $N^2$, where $N\times N$ is the size of the two-dimensional grid. The function $F(\phi, \epsilon) = F_\epsilon(\phi)$ is defined according to equation (\[stateq\]) and also outputs a column vector of length $N^2$. The constant values of the change in arclength parameter $ds$, the maximum $\epsilon$ value $\epsilon_{max}$, and the tolerance are assumed to be pre-set.
$\epsilon_0 = 0$; $\phi_0 = \phi^{(0)}$; $v = $ nullspace$\left(\left[\frac{\partial F}{ \partial \phi}(\phi_0,\epsilon_0) \ \ \frac{\partial F}{\partial \epsilon}(\phi_0);\right]\right)$; $v = v/$norm$(v)$; $\phi = \phi_0; \epsilon = \epsilon_0;$ $D = [\phi; \epsilon;] - [\phi_0; \epsilon_0;];$ $G = [F(\phi,\epsilon); \ D\cdot v-ds;]; $ $M = \left[\frac{\partial F}{ \partial \phi}(\phi_0,\epsilon) \ \ \frac{\partial F}{\partial \epsilon}(\phi); \ v^T; \right]$ $corr = M\backslash G$; $[\phi; \epsilon;] = [\phi; \epsilon;] -$ corr; $v = \left[\frac{\partial F}{ \partial \phi}(\phi_0,\epsilon) \ \ \frac{\partial F}{\partial \epsilon}(\phi); \ v^T; \right] \backslash \left[ zeros(N^2,1); 1;\right] $; $v = v/$norm$(v)$; $\phi_0 = \phi; \epsilon_0 = \epsilon;$
[999]{}
C. Sulem and P.L. Sulem, [*Nonlinear Schrödinger Equations: Self-Focusing And Wave Collapse,*]{} Applied Mathematical Sciences [**139**]{}, Springer \[1999\].
C. Conti, P. Di Trapani and S. Trillo, in [*Self-Focusing: Past and Present - Fundamentals and Prospects*]{}, Topics in Applied Physics [**114**]{}, p. 439, Springer-Verlag, Berlin (2009).
C. Conti and S. Trillo, in [*Localized Waves*]{}, H.E. Hernandez-Figueroa, M. Zamboni-Rached and E. Recomi (Eds.), p. 243 (2007).
J.M. Ghidaglia and J.C. Saut, Nonlinerity [**3**]{}, 475 (1990).
J.M. Ghidaglia and J.C. Saut, J. Nonlinear Sci. [**3**]{}, 169 (1993).
J.M. Ghidaglia and J.C. Saut, J. Nonlinear Sci. [**6**]{} 139 (1996).
M. Ablowitz, H. Segur. J. Fluid Mech., [**92**]{}, 691 –715, (1979).
V. Zakharov, E. Kuznetsov. Phys. Uspekhi [**55**]{}, 535 (2012).
N. Pereira, A. Sen, A. Bers. Phys. Fluids [**21**]{}, 117 (1978).
J. Myra, C. Liu. Phys. Fluids [**23**]{}, 2258 (1980).
S. Droulias, K. Hizanidis, J. Meier, and D.N Christodoulides, Optics Express [**13**]{} 1827 (2005).
Y. Lahini, E. Frumker, Y. Silberberg, S. Droulias, K. Hizanidis , R. Morandotti and D. Christodoulides, Phys. Rev. Lett. [**98**]{}, 023901 (2007).
C. Conti, S. Trillo, P. Di Trapani, A. Piskarkas, O. Jedrkiewicz and J. Trull, Phys. Rev. Lett. [**90**]{}, 170406 (2003).
P. Di Trapani, G. Valiulis, A. Piskarkas, O. Jedrkiewiecz, J. Trull, C. Conti and S. Trillo, Phys. Rev. Lett. [**91**]{}, 093904 (2003).
K. Hayata and M. Koshiba, Phys. Rev. E [**48**]{}, 2312 (1993).
N. Efremidis, K. Hizanidis, B. Malomed, and P. Di Trapani Phys. Rev. Lett. [**98**]{} 113901 (2007).
P. Kevrekidis, A.R. Nahmod, and C. Zeng, Nonlinearity [**24**]{}, 1523 (2011).
M.J. Ablowitz, Y.-P. Ma, and I. Rumanov, SIAM J. Appl. Math. [**77**]{}, 1248 (2017).
B. Dodson, J.L. Marzuola, B. Pausader, D. Spirn, arXiv:1708.08014.
P.G. Kevrekidis, [*The discrete nonlinear Schr[ö]{}dinger equation: Mathematical Analysis, Numerical Computation and Physical Perspectives*]{}, Springer-Verlag (Heidelberg, 2009).
S. N. Chow and J. K. Hale, [*Methods of Bifurcation Theory*]{}, Springer-Verlag (Heidelberg, 1982).
M. Golubitsky and D.G. Schaeffer, [*Singularities and Groups in Bifurcation Theory. Vol 1*]{}, Springer-Verlag (New York, 1985).
D.E. Pelinovsky, P.G. Kevrekidis, and D.J. Frantzeskakis, Physica D **212**, 1 (2005).
D. E. Pelinovsky, P. G. Kevrekidis, and D. J. Frantzeskakis, Physica D **212**, 20 (2005).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Jason Portenoy
- Jessica Hullman
- 'Jevin D. West'
bibliography:
- 'authorvis.bib'
date: 'October, 2016'
title: Leveraging Citation Networks to Visualize Scholarly Influence Over Time
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The status of the theoretical uncertainties for LEP 1 observables associated with the corresponding comparison among different codes is briefly reviewed.'
address:
- 'Dipartimento di Fisica Teorica, Università di Torino'
- 'INFN, Sezione di Torino'
author:
- Giampiero Passarino
title: 'The ${Z}$ Line-shape in the Standard Model[^1]DFTT/G-94-1'
---
—————————-Latex File————————————–
The achieved experimental accuracy at LEP 1 requires a detailed comparison of the theoretical predictions from different codes with the final goal of an estimate of the theoretical error. This is presently under investigation by the LEP 1 Precision Calculations Working Group and here a short report will be given on the status of the ${Z}$ parameters.
The comparison will proceed in three different phases by considering the results of five different codes, MIZA [@miza], LEPTOP [@leptop], TOPAZ0 [@topaz0], WOH [@miza] and ZFITTER [@zfitter]. In phase 1 one has taken into account the predictions for the so called pseudo-observables, namely ${\sin^2\theta_{e}}, {\sin^2\theta_{b}}, \dots$, while in a second step the ${Z}$ parameters have been compared, i.e. ${\Gamma_{Z}},{\Gamma_{h}},{\Gamma_{e}}, \dots$ or $R,{r_{b}}, \dots$ or ${A_{FB}}^0(l), \dots$. The final step in this project will require a comparison at the level of realistic observables, like $\sigma(e^+e^- \to {\overline f}f)$ and $A_{FB}(e^+e^- \to {\overline f}f)$, including a realistic set-up and $f = e$. At present the first two phases have been completed. As a result of this work we have realized the fact that there are two complementary objectives. In one case the theoretical uncertainty for a given observable $O$ is estimated as
$$\Delta O = \frac{1}{2} (\max_{i=Codes} - \min_{i=Codes})\,O_i$$
On the other end we can just derive additional informations from each single code by adopting different options which in turns are related to different implementations of higher orders radiative corrections. In the end we propose to compare bands of predictions instead of lines of predictions, as a function of ${m_{top}}$ or of any other unknown parameter of the standard model. In the following we would like to present a short list of Options with some words of comments.
The scale of $\alpha$ in the final state QED corrections. This is the well known factor $1 + \frac{3}{4}Q_f^2\frac{\alpha}{\pi}$. Actually this should not be an Option at all since the correct scale is $\alpha({m_{Z}})$ [@cos], i.e. the correction factor is
$$1 + \frac{\alpha({m_{Z}})}{\pi}Q_f^2[\frac{3}{4}
- \frac{1}{4}\frac{\alpha_s({m_{Z}})}{\pi}] + {\cal O}(\alpha^2)$$
The scale in the vertex EW corrections. We simply mention the three possibilities, namely $\alpha(0)$, $\alpha({m_{Z}}) \equiv G_F$ or $G_F$ for the leading (sub-leading) ${m_{top}}$ corrections and $\alpha(0)$ otherwise. The scale in the FTJR [@ftjr] corrections to ${Z}\to {\overline b}b$. At the moment the common choice is ${\alpha_{s}}({m_{top}})$, i.e.
$$\begin{aligned}
\Gamma_{Z{\overline b}b} &=& \Gamma_0\,\{ 1 + \frac{{\alpha_{s}}({m_{Z}})}{\pi} \\
&-& \delta [ 1 + (3 - \pi^2)\,\frac{{\alpha_{s}}({m_{top}})}{\pi}]\}\end{aligned}$$
but one could also isolate the gluon radiation as
$$\begin{aligned}
\Gamma_{Z{\overline b}b} &=& \Gamma_0\{1 + \frac{{\alpha_{s}}({m_{Z}})}{\pi}\} \\
&\times& \{1 - \delta[1 - \pi^2\,\frac{{\alpha_{s}}({m_{top}})}{\pi}]\}\end{aligned}$$
and the QCD factor $1 + {\alpha_{s}}({m_{Z}})/\pi$ is not included for the asymmetry $A_{FB}(b)$ and for $\sin^2\theta_b$. The aimed accuracy requires that ${m_{b}}\neq 0$ also in one loop diagrams. At the moment what it is used is
$$\Gamma_{Z{\overline b}b} = {{G_F{m_{Z}}^3}\over {8\sqrt{2}\,\pi}}\,\rho\,
\{ g_V^2 + (1 - 6\,{{{m_{b}}^2}\over {{m_{Z}}^2}})\,g_A^2\}$$
However vertex corrections, included in $g_{V,A}$ are usually computed for ${m_{b}}= 0$ and this is not fully consistent even if in the missing contributions the leading terms ${\cal O}({m_{b}}^2{m_{top}}^2)$ are absent. Moreover it is an open question what to use for ${m_{b}}$ in this case, the pole mass $m_b = 4.7\,$ GeV or the running one ${\overline m}_b({m_{Z}})$?
The physical Higgs contribution to the correction factor $\Delta\rho$ is not ultra-violet finite and only through the $\overline{MS}$ prescription we have a finite $\Delta\rho_H(\overline{MS})$. After that we are left with the option of a re-summation of such contribution, which makes the result slightly scale dependent. Otherwise we can just decide that all bosonic corrections are expanded to first order.
The scale of ${\alpha_{s}}$ in the ${\cal O}(\alpha{\alpha_{s}})$ corrections to the vector boson self-energies [@aas]. This question is particularly relevant in view of a correct treatment of the ${\overline t}t$ thresholds [@ttbth] where it has been recently suggested [@sv] that the non-perturbative effects can be numerically recovered by allowing a relatively small scale in the perturbative expansion, i.e. ${\alpha_{s}}(0.154{m_{top}})$. The singlet QCD contribution which is simple and unambiguous for the hadronic width but which becomes ambiguous, starting at ${\cal O}({\alpha_{s}}^2)$, for individual ${\overline q}q$ channels. Actually some sort of agreement has been recently reached on these matter but we want to summarize the roots of the problem [@qcd]. From a pragmatic point of view there is a hierarchical description where
$$\Gamma_{{\overline q}q} = \Gamma({Z}\to {\overline q}q(g) + {\overline q}' q')$$
for all $q'$ such that $m_{q'} < m_q$. On the other end we could have a democratic description where the final states ${\overline q}q + {\overline q}' q'$ are assigned for $\frac{1}{2}$ to $\Gamma_{{\overline q}q}$ and for the other $\frac{1}{2}$ to $\Gamma_{{\overline q}' q'}$. The two descriptions agree fortunately for the leading terms. In general one could also decide that such final states should not be assigned to any specific channels in such a way that
$$\begin{aligned}
{\Gamma_{h}}&\neq& \sum_q\,\Gamma_{{\overline q}q} = \sum_q\,\Gamma({Z}\to {\overline q}q) \\
&+& \sum_{q,q'}\,\Gamma({Z}\to{\overline q}q {\overline q}' q') + \dots\end{aligned}$$
In particular there is an ${\cal O}({\alpha_{s}}^3)$ contribution to $\Gamma_V^S$ which cannot be assigned to any specific flavour.
Perhaps the most important effect is related to the missing EW higher orders. Typically let $g_{V,A}$ be the vector or axial-vector coupling of the ${Z}$ to fermions. We write
$$g_i = g^0_i + \frac{\alpha}{\pi}\, g^1_i + {\cal O}(\alpha^2)$$
where $i= V,A$ and where $g^0_i$ is including the re-summation of universal terms. In computing partial widths or deconvoluted asymmetries different codes adopt different choices, i.e.
$$\begin{aligned}
g_i^2 &=& (g^0_i + \frac{\alpha}{\pi}\, g^1_i)^2 \\
{} &=& (g_i^0)^2 + 2\,\frac{\alpha}{\pi}\,g_i^0g_i^1\end{aligned}$$
where in the first case we square numerically and in the second one we expand consistently in perturbation theory. This two Options lead to sizeable effects if we consider for instance ${A_{FB}}^0$.
Another Option which should be taken into account is relative to the factorization vs non-factorization of final state QCD corrections. For a given channel one can write
$$\Gamma_{{\overline q}q} = {{G_F{m_{Z}}^3}\over {8\sqrt{2}\,\pi}}\,\rho\,
( g_V^2 + g_A^2)\,( 1 + \delta_{QCD})$$
where $g_V,g_A$ include non-universal vertex corrections. Notice that for $b$-quarks, in order to avoid double-counting, the FTJR term must not include $1+{\alpha_{s}}/\pi$. However the complete answer at ${\cal O}(\alpha{\alpha_{s}})$ is not known and strictly speaking the QCD correction should only multiply the universal terms absorbed into $g_V$ and $g_A$. Thus one can also adopt a non-factorized width both for $b$ and light quarks. We mention also that the double scale in the FTJR term, while certainly gauge-invariant for the leading ${m_{top}}$ corrections, remains questionable for the sub-leading and constant ones. Indeed, going back to the ${\cal O}(\alpha{\alpha_{s}})$ corrections, we recall that the full result, even including ${m_{b}}\neq 0$, is available
$$\alpha{\alpha_{s}}\,{m_{top}}^2 [ 1 + \frac{K}{{m_{top}}^2} + {\cal O}(\frac{{m_{Z}}^2}{{m_{top}}^4})]$$
while for vertex corrections to ${Z}\to {\overline q}q$ the following results are known: for $q\neq b$ the ${\cal O}(\alpha{\alpha_{s}}\,const.)$ is missing, for $q = b$ the leading $\alpha{\alpha_{s}}{m_{top}}^2$ corrections is the well known FTJR term while all the sub-leading (log and non-log) terms are missing.
The present choice for the comparisons is $1/\alpha({m_{Z}})|_{light} =
128.87 \pm 0.12$ [@priv] but unfortunately the full updated analysis, including the behavior of $\alpha(p^2)$, has not yet been published.
The running of ${m_{c}}$ has been included in the numerical work but it should be noticed that the use of
$$\begin{aligned}
{\overline m}({m_{c}}) &=& {m_{c}}\,[ 1 - \frac{4}{3}\,x({m_{c}}) \\
&+& (\frac{16}{9} - K_c)\,x^2({m_{c}})]\end{aligned}$$
where $x = \frac {{\alpha_{s}}}{\pi}$ gives unreliable results due to the low scale needed in ${\alpha_{s}}$. We suggest therefore to use
$${\overline m}({m_{c}}) = {{{m_{c}}}\over {1 + \frac{4}{3}\,x({m_{c}}) + K_c\,x^2({m_{c}})}}$$
where ${m_{c}}= 1.5\,$GeV is the pole mass.
In the light of the fact that ${m_{top}}\approx 1.9\,{m_{Z}}$ it looks opportune to rise the question whether or not the full two-loop standard model predictions are needed and requested. There are two possible answers, namely the present experimental accuracy plus the uncertainties connected with QED strongly support the idea that we don’t need a full two-loop calculation. Actually there are discrepancies among various ${\cal O}(\alpha^2)$ QED Bhabha generators that should be solved before devoting any attempt towards the two-loop electroweak effects. To the contrary of that a certain requirement of internal consistency, especially illustrated by our choice of options, would suggest that we indeed need them.
In conclusion we can affirm that by comparing the available semi-analytical codes which are based, among other things, on different renormalization schemes (${\overline{MS}}$ or on-shell) or the same code with different Options for radiative corrections we derive the result that
$$\Delta_{th}O \ll \Delta_{exp}O$$
where $O$ is any pseudo-observable or ${Z}$ parameter with the inclusion of QCD corrections. Certainly the remaining differences must be investigated both for internal consistency and for aesthetical reasons. Hopefully new calculations will contribute in a near future to make some of the Options obsolete and to unify the various treatments of radiative corrections. On the other hand a certain tendency to create a common default, by adopting a common procedure in front of alternative possibilities, should be taken with the due caution.
For cross sections and asymmetry with a realistic set-up the work is in progress among those groups with libraries which are also QED-dressers (BHM, TOPAZ0 and ZFITTER), but the greatest effort must be spent in order to reduce the theoretical error. Certainly the first step will be a detailed comparison for the forward-backward asymmetry ${A_{FB}}$.
Finally a very small sample of the various comparisons is shown in Figures 1-5. In order to give an idea of the situation we have decided to plot in Figure 1 the various predictions for ${\Gamma_{e}}$ as a function of ${m_{top}}$. The same information is displayed in Figure 2 for $\Gamma_b$. For a better understanding of the differences among the four codes we present in Figure 3-4) the absolute deviation of the four codes for $\sin^2\theta(e)$ from its average. In Figure 5 we present the theoretical uncertainty for the ratio $R$ as estimated by TOPAZ0 by switching on and off the various options that we have discussed above.
I would like to thank Oreste Nicrosini, Guido Montagna and Fulvio Piccinini for the continuous collaboration. I am also grateful to Dima Bardin and Manel Martinez for stimulating discussions. Finally I would like to thank Tord Riemann for the invitation and the very pleasant atmosphere at this Conference.
[9]{} G. Burgers, W. Hollik and M. Martinez, program BHM; W.Hollik, Fortschr. Phys. 38 (1990) 3, 165; M.Consoli, W.Hollik and F.Jegerlehner: Proceedings of the Workshop on Z physics at LEP I, CERN Report 89-08 Vol.I,7; G.Burgers, F.Jegerlehner, B.Kniehl and J.H.K[ü]{}hn: the same proceedings, CERN Report 89-08 Vol.I,55.
V. A. Novikov, L. B. Okun, A. N. Rozanov and M. I. Vysotsky, CERN preprint CERN-TH.7217/94.
G. Montagna, O. Nicrosini, G. Passarino, F. Piccinini and R. Pittau, Nucl. Phys. B401(1993)3. G. Montagna, O. Nicrosini, G. Passarino, F. Piccinini and R. Pittau, Comput. Phys. Commun. 76(1993)328.
D. Bardin et al., program ZFITTER 4.0; Nucl. Phys. B351 (1991) 1; Z. Phys. C44 (1989) 493; Phys. Lett. B255 (1991) 290.
A. L. Kataev, Phys. Lett. B287()209.
J. Fleischer, O. V. Tarasov, F. Jegerlehner and P. Raczka, Phys. Lett. B293 (1992) 437.
B. A. Kniehl, Nucl.Phys. B347(1990)86; A. Djouadi, Nuovo Cim. 100A(1988)357.
V. S. Fadin and V. A. Khoze, JETP Lett. 46(1987)525; B. A. Kniehl and A. Sirlin, Phys. Rev. D47(1993)883; F. J. Yndurain, Madrid preprint FTUAM 38/93; B. A. Kniehl and A. Sirlin, DESY-preprint DESY 93-194.
B. H. Smith and M. B. Voloshin, UMN-TH-1241/94, TPI-MINN-94/5-T.
K. G. Chetyrkin, Phys. Lett. B307(1993)169; K. G. Chetyrkin and A. Kwiatkowski, Phys. Lett. B305 (1993)285; S. A. Larin, T. van Ritbergen and J. A. M. Vermaseren, Nikhef Preprint NIKHEF-H/93-26; K. G. Chetyrkin and J. H. Kühn, Phys. Lett. B308 (1993)127.
F. Jegerlehner, private communication.
[^1]: Talk given at the 1994 Zeuthen Workshop on Elementary Particle Theory
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a natural and low-complexity technique for achieving the capacity of the Gaussian relay network in the high SNR regime. Specifically, we propose the use of end-to-end structured lattice codes with the amplify-and-forward strategy, where the source uses a nested lattice code to encode the messages and the destination decodes the messages by lattice decoding. All intermediate relays simply amplify and forward the received signals over the network to the destination. We show that the end-to-end lattice-coded amplify-and-forward scheme approaches the capacity of the layered Gaussian relay network in the high SNR regime. Next, we extend our scheme to non-layered Gaussian relay networks under the amplify-and-forward scheme, which can be viewed as a Gaussian intersymbol interference (ISI) channel. Compared with other schemes, our approach is significantly simpler and requires only the end-to-end design of the lattice precoding and decoding. It does not require any knowledge of the network topology or the individual channel gains.'
author:
- 'Yun Xu[^1]'
- 'Edmund Yeh[^2]'
- 'Muriel Médard[^3]'
bibliography:
- 'IEEE-lattice.bib'
title: 'Approaching Gaussian Relay Network Capacity in the High SNR Regime: End-to-End Lattice Codes'
---
Introduction
============
Finding the capacity of Gaussian relay networks with one source, one destination, and a set of relays, has been a long-standing open problem in network information theory. The relay channel was first investigated in the seminal work of Cover and EI Gamal [@cov79]. More recently, Kramer [*et al.*]{} considered transmission techniques for larger Gaussian relay networks, [*e.g.*]{}, the amplify-and-forward, decode-and-forward, and compressed-and-forward schemes [@kra05]. Avestimeher [*et al.*]{} presented a deterministic model for Gaussian relay networks and proposed the quantize-map-and-forward strategy [@tse11]. It has been shown that the quantize-map-and-forward strategy can achieve a rate within a constant number of bits from the information-theoretic cut-set bound for Gaussian relay networks, which is independent of the channel gains and the operating SNR [@tse11]. In [@dig12], Ozgur and Diggavi incorporated lattice codes, lattice quantization, and lattice-to-lattice mapping into the quantize-map-and-forward scheme. It was shown that the lattice-based quantize-map-and-forward scheme can still achieve the capacity of Gaussian relay networks within a constant gap. While offering strong performance in terms of achievable rates, the schemes presented in [@tse11; @dig12], involve considerable operational complexity at intermediate relays.
As pointed out in [@med12], in wireless communication settings, signals simultaneously transmitted from different sources add, leading to the receiver obtaining a superposition of these signals, scaled by the channel gains. Since the relays are not interested in the messages sent by the source, they do not necessarily have to decode, compress or quantize the messages. Since the relays have already observed the sum of the signals, in some settings a natural strategy would be to simply amplify and forward without explicitly dealing with the noise.
A multihop amplify-and-forward scheme with random encoding and decoding was proposed in [@med12]. In this strategy, the message sent by the source is propagated over many intermediate nodes (relays) and possibly over multiple hops. All relays exploit the interference and forward the received signals over the network to the destination. It has been shown in [@med12] that the achievable rate of the multihop amplify-and-forward scheme approaches the capacity of the Gaussian relay network when the SNR at the destination is sufficiently high. In this paper, we propose the use of end-to-end structured lattice codes with the amplify-and-forward strategy, where the source uses a nested lattice code to encode the messages and the destination decodes the messages by lattice decoding. All intermediate relays simply amplify and forward the received signals over the network to the destination. Relative to the random coding approach of [@med12], the use of structured lattice codes significantly reduces system complexity by making possible computationally tractable encoding and decoding. Furthermore, the use of end-to-end lattice codes implies that we require little information concerning the network topology, or individual channel gains. Instead, we require only the end-to-end channel response, which can be obtained by using probing signals.
We show that the end-to-end lattice-coded amplify-and-forward scheme approaches the capacity of the layered Gaussian relay network under the high-SNR condition presented in [@med12]. This result is facilitated by the key observation that a Gaussian layered relay network under amplify-and-forward is equivalent to a point-to-point Gaussian channel. Next, we extend our scheme to non-layered Gaussian relay networks under the amplify-and-forward scheme, which can crucially be viewed as a Gaussian intersymbol interference (ISI) channel. Our lattice-coded amplify-and-forward scheme is simpler than the lattice-based quantize-map-and-forward scheme proposed in [@dig12], since it does not require lattice quantization and lattice-to-lattice mapping at relay nodes. For layered networks, only end-to-end design of the precoding and decoding with nested lattice codes is required. Thus, the end-to-end lattice-coded amplify-and-forward scheme is a natural and low-complexity technique for achieving the capacity of the Gaussian relay network in the high SNR regime.
The nested lattice code was originally proposed by Erez [*et al.*]{} in [@zam02], [@zam04] and [zam05-2]{}. In [@zam04], it was shown that the nested lattice code with lattice decoding can achieve the capacity of the additive white Gaussian noise (AWGN) channel at any SNR. Erez [*et al.*]{} showed that the power-constrained AWGN channel can be transformed into the modulo-lattice additive noise (MLAN) channel by minimum mean-square error (MMSE) scaling along with dithering. The capacity of the MLAN channel, achieved by uniform inputs, becomes the capacity of the AWGN channel in the limit of large lattice dimension.
In [@zam02] and [@zam05-2], Erez [*et al.*]{} extended their techniques to the AWGN dirty-paper channel. They obtained the achievable rate at any SNR by incorporating MMSE scaling. It was shown that with an appropriate choice of the lattice, the achievable rate approaches the capacity of the AWGN dirty-paper channel as the dimension of the lattice goes to infinity. These results provide an information-theoretic framework used in [@zam02] to study precoding for the Gaussian ISI channel. Erez [*et al.*]{} showed that when combined with the techniques of interleaving/deinterleaving and water-filling, nested lattice precoding and decoding can achieve the capacity of the Gaussian ISI channel [zam02]{}.
The remainder of this paper is organized as follows. The network model is presented in Section II. The fundamental properties of lattices and nested lattice codes are summarized in Section III. The main result on the lattice-coded multihop amplify-and-forward strategy is presented in Section IV. Section V extends the analysis to non-layered networks. Section VI concludes the paper.
Network Model
=============
Layered Network
---------------
We first focus on the layered network in which each path from the source to the destination has the same number of hops. We denote the layer $l$ by $%
\mathcal{L}_{l}$, $l=0,1,\ldots ,L$. Assume that the source $s$ is located at layer $\mathcal{L}_{0}$, and the destination $d$ at layer $\mathcal{L}%
_{L} $. We denote the number of relays at layer $\mathcal{L}_{l}$ by $n_{l}$, and thus $\sum_{l=1}^{L-1}n_{l}=N$. In a layered network, the input-output relationship is simple due to the fact that all copies of a source message transmitted on different paths arrive at the destination simultaneously. An example of a layered network is shown in Figure $\ref{fig-LN}$.
![Example of a Layered Network[]{data-label="fig-LN"}](fig-LN){height="2in"}
Now consider a layered Gaussian relay network consisting of a single source $%
s$, a single destination $d$, and a set of $N$ relays. The communication link from node $i$ to node $j$ has a nonnegative real channel gain, represented by $h_{ij}\in \mathbb{R}_{+}$. The channel output at node $j\neq
s$ is $$y_{j}=\sum_{i\in \mathcal{N}\left( j\right) }h_{ij}x_{i}+z_{j}. \label{WGN}$$where $x_{i}$ is the channel input at node $i$, $z_{j}$ is the real Gaussian noise with zero mean and unit variance, and $\mathcal{N}\left( j\right) $ denotes the set of nodes that can transmit to node $j$ with a direct link, i.e., $\mathcal{N}\left( j\right) =\left\{ i:h_{ij}>0\right\} $. Note that the links are assumed to be directed so that $i\in \mathcal{N}\left(
j\right) $ does not imply $j\in \mathcal{N}\left( i\right) $. We assume that there is an average power constraint at each node:$$E\left[ X_{i}^{2}\right] \leq P_{i} \label{PC}.$$The source $s$ wishes to send a message from a message set $\mathcal{W}%
=\left\{ 1,\ldots ,2^{nR}\right\} $ to the destination $d$ with transmission rate $R$. The encoding function at the source is given by $%
X_{s}^{n}=f\left( W\right) ,W\in \mathcal{W}$, and a decoding function at the destination $d$ is given by $\hat{W}=g\left( Y_{d}^{n}\right) $. A $%
\left( R,n\right) $ code consists of a message set $\mathcal{W}$, an encoding function at the source, and a decoding function at the destination. The average error probability of the $\left( R,n\right) $ code is given by $%
P_{e}=\Pr [\hat{W}\neq W]$. A rate $R$ is said to be achievable if for any $%
\epsilon >0$, there exists a $\left( R,n\right) $ code such that $P_{e}\leq
\epsilon $ for a sufficiently large $n$.
High SNR Regime
---------------
As in [@med12], we are interested in the scenario in which all relays forward the data with large enough power to guarantee that the total propagated noise at the destination by multihop amplify-and-forward is low enough. Assume that each node $i$ transmits with the average power $P_{i}$ given by $\left( \ref{PC}\right) $. The power received at relay $j\in
\mathcal{L}_{l},l=1,\ldots ,L-1$, is then determined by$$P_{R,j}=\left( \sum_{i\in \mathcal{L}_{l-1}}h_{ij}\sqrt{P_{i}}\right) ^{2},\
\ \ j\in \mathcal{L}_{l}$$ and the power received at the destination $d$ is given by $$P_{d}=\left( \sum_{i\in \mathcal{L}_{L-1}}h_{id}\sqrt{P_{i}}\right) ^{2}.
\label{Pd}$$
As in [@med12], we consider a high SNR regime where for some small $\delta >0$, the transmit powers of the relays satisfy $$\min_{j\in \mathcal{L}_{l}}P_{R,j}\geq \frac{1}{\delta },\text{ \ \ }%
l=1,\ldots ,L-1. \label{HSNR}$$
We then assume that $P_{d}$ remains a constant as $\delta \rightarrow 0$, so that the Multiple-Access Channel (MAC) at the destination is a bottleneck for the data transmission.[^4] The MAC cut-set bound is given by $$C_{MAC}=\frac{1}{2}\log \left( 1+P_{d}\right). \label{CMAC}$$
Lattices and Nested Lattice Codes
=================================
In this section, we briefly review some basic properties of lattices and nested lattice codes. A more extensive discussion can be found in references such as [@zam02] and [@zam04]. A lattice $\Lambda $ is a discrete subgroup of the Euclidean space $\mathbb{R}^{n}$. If $\lambda _{1}$ and $\lambda _{2}$ are two elements of a lattice $\Lambda $, then the sum $\lambda
_{1}+\lambda _{2}$ and the additive inverse $-\lambda _{1}$ are also elements of $\Lambda $. Any lattice can be written in terms of its generating matrix $G$:$$\Lambda =\left\{ \lambda =G\boldsymbol{x}:\boldsymbol{x}\in \mathbb{Z}%
^{n}\right\}.$$ We can then define the nearest neighbor quantizer associated with $\Lambda $ by$$Q_{\Lambda }\left( \boldsymbol{x}\right) =\func{argmin}_{\lambda \in \Lambda
}\left\Vert \boldsymbol{x}-\lambda \right\Vert.$$The Voronoi region of a lattice point $\lambda \in \Lambda $ is the set of all points that quantize to it. The fundamental Voronoi region $\mathcal{V}$ is the set of all points that quantize to the origin, i.e., $$\mathcal{V}=\{\boldsymbol{x}:Q_{\Lambda }\left( \boldsymbol{x}\right) =%
\boldsymbol{0}\}.$$Define the modulo-$\Lambda $ operation corresponding to $\mathcal{V}$ as $$\boldsymbol{x}\text{ mod }\Lambda =\boldsymbol{x}-Q_{\Lambda }\left(
\boldsymbol{x}\right).$$
The second moment of a lattice $\Lambda $ is defined by$$\sigma _{\Lambda }^{2}=\frac{1}{nVol\left( \mathcal{V}\right) }\int_{%
\mathcal{V}}\left\Vert \boldsymbol{x}\right\Vert ^{2}d\boldsymbol{x},$$and the normalized second moment of a lattice $\Lambda $ is defined by $$G\left( \Lambda \right) =\frac{1}{n\left[ Vol\left( \mathcal{V}\right) %
\right] ^{1+2/n}}\int_{\mathcal{V}}\left\Vert \boldsymbol{x}\right\Vert ^{2}d%
\boldsymbol{x},$$where $Vol\left( \mathcal{V}\right) $ is the volume of $\mathcal{V}$.
Two lattices $\Lambda _{1}$ and $\Lambda _{2}$ are said to be nested if $%
\Lambda _{1}\subseteq \Lambda _{2}$, where $\Lambda _{1}$ is called the coarse lattice and $\Lambda _{2}$ the fine lattice. Denote by $\mathcal{V}%
_{1}$ and $\mathcal{V}_{2}$ the fundamental Voronoi regions of $\Lambda _{1}$ and $\Lambda _{2}$, respectively. The coding rate is defined by $$R=\frac{1}{n}\log \left[ \frac{Vol\left( \mathcal{V}_{1}\right) }{Vol\left(
\mathcal{V}_{2}\right) }\right].$$The points in the set $$\mathcal{C}=\Lambda _{2}\cap \mathcal{V}_{1}$$are called the coset leaders of $\Lambda _{1}$ relative to $\Lambda _{2}$. For each $\boldsymbol{c}\in \mathcal{C}$, the shifted coarse lattice $%
\Lambda _{1,c}=\boldsymbol{c}+\Lambda _{1}$ is called a coset of $\Lambda
_{1}$ relative to $\Lambda _{2}$.
The use of high-dimensional nested lattice code is justified by the existence of asymptotically good lattices. We consider two types of goodness as introduced in [@zam02] and [@zam04].
\(1) *Good for AWGN Channel Coding:* For any $\epsilon >0$ and sufficiently large $n$, there exists an $n$-dimensional lattice $\Lambda $ with the volume of the fundamental Voronoi region $Vol\left( \mathcal{V}%
\right) <2^{n\left[ h\left( Z\right) +\epsilon \right] }$, where $Z$ is Gaussian noise with variance $\sigma _{Z}^{2}$, and $h\left( Z\right) =%
\frac{1}{2}\log \left( 2\pi e\sigma _{Z}^{2}\right) $ is the differential entropy of $Z$, such that$$P_{e}=\Pr \left[ Z\notin \mathcal{V}\right] <\epsilon.$$
*(2) Good for Source Coding:* For any $\epsilon >0$ and sufficiently large $n$, there exists an $n$-dimensional lattice $\Lambda $ whose normalized second moment $G\left( \Lambda \right) $ satisfies$$\log \left( 2\pi eG\left( \Lambda \right) \right) <\epsilon.$$
Lattice-Coded Multihop Amplify-and-Forward
==========================================
In [@med12], it is shown that the simple multihop amplify-and-forward scheme with random coding can approximately achieve the unicast capacity of a Gaussian relay network in the high SNR regime. In this paper, we propose the use of structured nested lattice codes in conjunction with the multihop amplify-and-forward scheme. Choose a pair of nested lattices $\left( \Lambda _{1},\Lambda _{2}\right) $, $\Lambda
_{1}\subseteq \Lambda _{2}$, with the coding rate $$R=\frac{1}{n}\log \left[ \frac{Vol\left( \mathcal{V}_{1}\right) }{Vol\left(
\mathcal{V}_{2}\right) }\right] \geq R_{LAF},$$where $R_{LAF}$ is the rate achieved by the lattice-coded amplify-and-forward scheme defined by $\left( \ref{AFR}\right) $. We choose the coarse lattice $\Lambda _{1}$ to be good for source coding, with the second moment $\sigma _{\Lambda _{1}}^{2}=P_{s}$, where $P_{s}$ is the average power constraint of the source node $s$. We choose the fine lattice $%
\Lambda _{2}$ to be good for AWGN channel coding. Let $Q_{\Lambda _{2}}$ denote the nearest neighbor quantizer of the fine lattice $\Lambda _{2}$. We apply the scheme proposed in [@zam02], [@zam04] and [@zam05-2] to the layered Gaussian relay network as follows:
\(1) *Source:* the source $s$ maps the message $W\in \mathcal{W}$ uniformly at random to a coset leader of $\Lambda _{1}$ relative to $\Lambda _{2}$:$$\boldsymbol{t}\in \mathcal{C}=\Lambda _{2}\cap \mathcal{V}_{1}.$$Then a random dither vector $\boldsymbol{u}$ is generated uniformly over $%
\mathcal{V}_{1}$, i.e., $\boldsymbol{u}\sim Unif\left( \mathcal{V}%
_{1}\right) $. Given the message $W$, the source encoder sends$$\boldsymbol{x}_{s}=[\boldsymbol{t}+\boldsymbol{u}]\text{ mod }\Lambda _{1}.
\label{SEC}$$
\(2) *Relay:* each relay $i\in \mathcal{L}_{l},l=1,\ldots ,L-1$, performs the multihop amplify-and-forward scheme$$\boldsymbol{x}_{l,i}=\beta _{i}\boldsymbol{y}_{l,i},\text{ \ \ \ }i\in
\mathcal{L}_{l} \label{AF}$$where the amplification gain is chosen as$$\beta _{i}=\frac{\sqrt{P_{i}}}{\sqrt{\left( 1+\delta \right) P_{R,i}}},\text{
\ \ }i\in \mathcal{L}_{l}. \label{AmGain}$$In [@med12], it is shown that the power constraint $\left( \ref{PC}\right) $ at each node $i$ is satisfied by choosing the amplification gain in $\left( \ref%
{AmGain}\right) $. Also shown in [@med12] are the following two results on the propagated noise.
[@med12] At any node $i\in \mathcal{L}_{l}$, the noise propagated from all nodes in layer $\mathcal{L}_{l-k},k=1,\ldots ,l-1$, via the multihop amplify-and-forward scheme in the high SNR regime has the power$$P_{z,i}^{l-k}\leq \frac{\delta P_{R,i}}{\left( 1+\delta \right) ^{k}}.$$
\[TPN\] [@med12] The total noise propagated to the destination $d \in \mathcal{L}_{L}$ has the power $$P_{z,d}=\sum_{k=1}^{L-1}P_{z,d}^{L-k}=\delta P_{d}\sum_{k=1}^{L-1}\frac{1}{%
\left( 1+\delta \right) ^{k}}\leq L\delta P_{d}. \label{TPNP}$$
\(3) *Destination:* the destination computes $$\boldsymbol{\hat{y}}_{d}=Q_{\Lambda _{2}}\left( \alpha \boldsymbol{y}_{d}+%
\boldsymbol{u}\right) \text{ mod }\Lambda _{1},$$where $$\alpha =\frac{\gamma }{1+\gamma },\text{ \ \ }\gamma =\frac{P_{d}}{\left(
1+\delta \right) ^{L-1}P_{z,d}}.$$ The following theorem is the main result for the lattice-coded amplify-and-forward scheme.
In a layered relay network $\left( \ref{WGN}\right) $ in the high SNR regime defined by $\left( \ref{HSNR}\right) $, the lattice-coded multihop amplify-and-forward achieves the rate$$R_{LAF}\geq \frac{1}{2}\log \left[ 1+\frac{1}{\left( 1+\delta \right) ^{L-1}}%
\frac{P_{d}}{1+L\delta P_{d}}\right]. \label{AFR}$$
As in [@med12], if the amplification gain at each relay is chosen as $%
\left( \ref{AmGain}\right) $, the received signal at the destination can be written as$$y_{d}=\hat{h}_{d}x_{s}+\hat{z}_{d}+z_{d}, \label{GC}$$where $\hat{z}_{d}$ the total propagated noise, $z_{d}$ is the noise at the destination, and $$\hat{h}_{d}=\frac{\sqrt{P_{d}}}{\sqrt{P_{s}\left( 1+\delta \right) ^{L-1}}}.
\label{hd}$$By $\left( \ref{GC}\right) $ and $\left( \ref{hd}\right) $, the received signal power at the destination is$$\hat{P}_{d}=\frac{P_{d}}{\left( 1+\delta \right) ^{L-1}}.$$By Corollary $\ref{TPN}$, the power of the total propagated noise $\hat{z}%
_{d}$ is$$P_{z,d}=\delta P_{d}\sum_{k=1}^{L-1}\frac{1}{\left( 1+\delta \right) ^{k}}%
\leq L\delta P_{d}.$$Therefore the SNR at the destination satisfies$$SNR\geq \frac{1}{\left( 1+\delta \right) ^{L-1}}\frac{P_{d}}{1+L\delta P_{d}}.
\label{GSNR}$$In other words, the received signal $y_{d}$ can be viewed as the output of the AWGN channel characterized by $\left( \ref{GC}\right) $ with the SNR given by $\left( \ref{GSNR}\right) $. The capacity, or equivalently the achievable rate via amplify-and-forward, is then given by $$\begin{aligned}
R_{LAF} &=&\frac{1}{2}\log \left( 1+SNR\right) \\
&\geq &\frac{1}{2}\log \left[ 1+\frac{1}{\left( 1+\delta \right) ^{L-1}}%
\frac{P_{d}}{1+L\delta P_{d}}\right].\end{aligned}$$It is shown in [@zam02] and [@zam04] that if we choose the coarse lattice to be good for source coding and the fine lattice $\Lambda _{2}$ to be good for AWGN channel coding, nested lattice codes can achieve the capacity of the AWGN Gaussian channel when the dimension, or equivalently, the length of the codewords $n$, tends to infinity. Hence, the above lattice-coded amplify-and-forward scheme can achieve the rate $R_{LAF}$.
Note that as $\delta \rightarrow 0$, the rate achieved by the lattice-coded multihop amplify-and-forward scheme in $\left( \ref{AFR}\right) $ approaches the MAC cut-set bound $\left( \ref{CMAC}\right) $, and the unicast capacity of the Gaussian relay network.
Extensions
==========
Non-layered Networks
--------------------
In layered networks, each path from the source to the destination has the same number of hops, so that all copies of the source message transmitted on different paths arrive at the destination with the same delay. In non-layered networks, however, copies of the source message may arrive at the destination with different delays through different paths. Assume that the number of hops (length) in the longest path is $L\geq 1$. We can then classify all paths from the source to the destination according to the path length$$P_{l}=\left\{ \text{paths of length }l\right\}$$Assume that the number of paths of length $l$ is $K_{l},l=1,\ldots ,L$. As shown in [@med12], the received signal at the destination $d$ at time $t$ is then given by$$\begin{aligned}
y_{d}\left( t\right) &=&h_{0}x_{s}\left( t\right) +\sum_{j\in
P_{1}}h_{j,1}x_{s}\left( t-1\right) +\ldots \\
&&+\sum_{j\in P_{L}}h_{j,L}x_{s}\left( t-L\right) +z_{e}\left( t\right),
\label{eq:ISI}\end{aligned}$$where $h_{0}$ is the channel gain on the direct link from the source to the destination, and $h_{j,l}$ is the equivalent channel gain of path $j$ in the set $P_{l}$. Note that $h_{j,l}$ depends on the network topology, and contains the accumulated channel gains and amplification gains on the source-destination path $j$. Finally, $z_{e}\left( t\right) $ denotes the total propagated noise at the destination.
From , we see that under the amplify-and-forward scheme, the non-layered Gaussian relay network is equivalent to a Gaussian ISI channel: $$y_{d}\left( t\right) =\sum_{l=0}^{L}h_{l}x_{s}\left( t-l\right) +z_{e}\left(
t\right), \label{GISI}$$where $x_{s}\left( t-l\right) ,l=0,\ldots ,L$, are the inputs to the Gaussian ISI channel, $h_{l}=$ $\sum_{j\in P_{l}}h_{j,l}$ represents the ISI coefficient, and $y_{d}\left( t\right) $ stands for the received samples. The additive Gaussian noise is denoted by $z_{e}\left( t\right) $.
We now focus on the feedforward MMSE decision feedback (MMSE-DFE) equalizing filter for the Gaussian ISI channel [@for95], [@for95-2]. The output of the MMSE-DFE feedforward filter can be written as $$r\left( t\right) =x_{s}\left( t\right) +s\left( t\right) +n\left( t\right),
\label{DP}$$where $s\left( t\right) =\sum_{l=1}^{L}\hat{h}_{l}x_{s}\left( t-l\right)$ is the post-cursor intersymbol interference, and $\hat{h}_{l}$ and $n\left(
t\right) $ represent the ISI coefficients and the sampled noise at the output of the MMSE-DFE feedforward filter, respectively.
The SNR associated with the MMSE-DFE filter is defined by [@for95]$$SNR_{MMSE-DFE}=\frac{E\left[ X_{s}^{2}\left( t\right) \right] }{E\left[
N^{2}\left( t\right) \right] },$$and the capacity of the Gaussian ISI channel is given by [@for95-2]$$C_{ISI}=\frac{1}{2}\log \left( 1+SNR_{MMSE-DFE}\right). \label{CISI}$$
If the encoder knows the entire post-cursor intersymbol interference vector before transmission, as mentioned in [@zam02], the input-output relationship given by $\left( \ref{DP}\right) $ can be viewed as the Gaussian dirty-paper channel whose capacity is given by $\left( \ref{CISI}%
\right) $. Based on that observation, [@zam02] proposed a coding strategy for the Gaussian ISI channel, in which the MMSE-DFE feedback equalizing filter is replaced by nested lattice precoding, as described in the dirty paper case. In the interleaver, the messages are encoded row by row and are transmitted column by column. When a message which comprises the $j$th row of the interleaver is to be encoded, the post-cursor interfering symbols belong to the codewords for messages which have been already encoded, similar to the dirty paper scenario [@zam02]. In [@zam02], it is shown that nested lattice precoding with interleaving/deinterleaving and waterfilling can achieve the capacity of the Gaussian ISI channel given by $%
\left( \ref{CISI}\right) $. Waterfilling is required because the sampled noise $n(t)$ in may not be white Gaussian. If we incorporate nested lattice precoding into the multihop amplify-and-forward scheme, the achievable rate can thus be obtained as the capacity of the corresponding Gaussian ISI channel. The lattice-coded multihop amplify-and-forward scheme is shown in Figure $\ref{fig-ISI}$.
Unlike the case for layered networks, in order to implement nested lattice precoding with interleaving for non-layered Gaussian wireless relay networks, it is necessary to know the channel gains as manifested in the “ISI coefficients." In the absence of such knowledge, techniques such as blind equalization may be needed to preserve the performance of our scheme.
If the encoder does not know the entire post-cursor ISI vector before transmission, we have to use the original MMSE-DFE technique proposed in [@for95] with nested lattice encoding and decoding. In that case, however, the decoded messages must be fed back. Thus, any decoding error will affect the performance of the MMSE-DFE feedback filter, and hence the subsequent decoding process. This may lead to the decoding error propagating over multiple symbols.
![Lattice-Coded Multihop Amplify-and-Forward Scheme[]{data-label="fig-ISI"}](fig-ISI){height="3.5in"}
Conclusion
==========
In this paper, we considered an end-to-end lattice-coded multihop amplify-and-forward strategy for Gaussian wireless relay networks in the high SNR regime. When the power received at all relays are large enough, our strategy performs well for both layered and non-layered Gaussian relay networks. In the worst case, the bottleneck of the multihop transmission is at the multi-access channel (MAC) at the destination. We showed that our strategy approaches the MAC cut-set bound as the received powers at the relays increase. The lattice-coded multihop amplify-and-forward scheme is simpler than the decode-and-forward scheme and the quantize-map-and-forward scheme. Our scheme requires only end-to-end design: lattice precoding at the source and decoding at the destination. It does not require any knowledge of the network topology or the individual channel gains.
[^1]: Department of Electrical Engineering, Yale University, New Haven, CT 06511. Email: [email protected]
[^2]: Department of Electrical and Computer Engineering, Northeastern University, Boston, MA 02115. Email: [email protected]
[^3]: Department of Electrical Engineering and Computer Science, MIT, Cambridge, MA 02139. Email: [email protected]
[^4]: Note that for network capacity, the worst case occurs when the bottleneck is at the MAC at the destination. In this case, the noise is propagated over more hops than in any other case.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We perform a microscropic analysis of how the constraints imposed by conservation laws affect $q=0$ Pomeranchuk instabilities in a Fermi liquid. The conventional view is that these instabilities are determined by the static interaction between low-energy quasiparticles near the Fermi surface, in the limit of vanishing momentum transfer $q$. The condition for a Pomeranchuk instability is set by $F\cs_l =-1$, where $F\cs_l$ (a Landau parameter) is a properly normalized partial component of the anti-symmetrized static interaction $F(k,k+q; p,p-q)$ in a charge (c) or spin (s) sub-channel with angular momentum $l$. However, it is known that conservation laws for total spin and charge prevent Pomeranchuk instabilities for $l=1$ spin- and charge- current order parameters. Our study aims to understand whether this holds only for these special forms of $l=1$ order parameters, or is a more generic result. To this end we perform a diagrammatic analysis of spin and charge susceptibilities for charge and spin density order parameters, as well as perturbative calculations to second order in the Hubbard $U$. We argue that for $l=1$ spin-current and charge-current order parameters, certain vertex functions, which are determined by high-energy fermions, vanish at $F\cs_{l=1}=-1$, preventing a Pomeranchuk instability from taking place. For an order parameter with a generic $l=1$ form-factor, the vertex function is not expressed in terms of $F\cs_{l=1}$, and a Pomeranchuk instability does occur when $F\cs_1=-1$. We argue that for other values of $l$, a Pomeranchuk instability occurs at $F\cs_{l} =-1$ for an order parameter with any form-factor'
author:
- 'Yi-Ming Wu'
- Avraham Klein
- 'Andrey V. Chubukov'
bibliography:
- 'ref.bib'
title: The Conditions for $l=1$ Pomeranchuk Instability in a Fermi Liquid
---
Introduction {#sec:introduction}
============
This paper is devoted to the analysis of subtle effects associated with a Pomeranchuk instability in a Fermi liquid (FL) due to the interplay with conservation laws. A system of interacting fermions is called a Fermi liquid if its properties differ from those of free fermions in a quantitative, but not qualitative manner [@AGD; @Lifshitz1980; @Shankar1994]. Specifically, the distribution function $n_k$ undergoes a finite jump at the Fermi momentum, $k_F$, with some jump magnitude $Z < 1$; the velocity $v^*_F$ of fermionic excitations near the Fermi surface (FS) remains finite; and the lifetime of fermionic excitations near a FS is parametrically larger than the energy counted from the Fermi level, i.e., fermions infinitesimally close to the FS can be viewed as infinitely long lived. These three features form the basis for the description of low-energy fermionic states in terms of quasiparticles, whose distribution function at $T=0$ is a step function. The validity of FL postulates has been verified in microscopic calculations for realistic interaction potentials and was found to hold at small/moderate couplings in dimensions $d >1$.
Stronger interactions can destroy a FL. In general, such destruction can occur in two ways. One option is the transformation of a metal into a Mott insulator, once the interaction $U$ becomes comparable to a fermionic bandwidth $W$. This instability involves fermions located everywhere in the Brillouin zone. Another option is an instability driven by fermions only very near the FS, such as superconductivity and $q=0$ instabilities in a particle-hole channel, often called Pomeranchuk instabilities. The latter leads to either phase separation, or ferromagnetism, or a deformation of a FS and the development of a particle-hole order with non-zero angular momentum (see e.g. Refs ). A Pomeranchuk instability in a given channel occurs when the corresponding interaction exceeds $1/N_F$, where $N_F$ is the density of states at the FS. When $W N_F \gg 1$, a Pomeranchuk instability occurs well inside the metallic regime.
A Pomeranchuk instability is generally expressed as a condition on a Landau parameter. For a rotationally-invariant and SU(2) spin-invariant FL, an anti-symmetrized static interaction between fermions at the FS and at strictly zero momentum transfer, $\Gamma^\omega (k,k;p,p)$ can be separated into spin and charge components, and each can be further decomposed into sub-components with different angular momenta $l$. Landau parameters are properly normalized dimensionless sub-components $F_{l}^{c(s)}$, where $c(s)$ selects charge (spin) channel, and $l=0,1,2,...$ [@Landau1; @Landau2; @AGD; @Lifshitz1980]. Pomeranchuk argued in his original paper [@Pomeranchuk] that a static susceptibility $\chi^{c(s)}_{l}$ scales as $1/(1 + F^{c(s)}_l)$ and diverges when the corresponding $F^{c(s)}_l = -1$. The divergence signals an instability towards a $q=0$ density-wave order with angular momentum $l$.
The $1/(1 + F^{c(s)}_l)$ form of the susceptibility can be reproduced diagrammatically by summing up particle-hole bubbles of free fermions within RPA. The momentum/frequency integration within each bubble is confined to the FS, hence the dimensionless interaction between the bubbles is exactly $F^{c(s)}_l$. The RPA series are geometric, hence $\chi^{c(s)}_{l, RPA} = \chi_{l,0} /(1 + F^{c(s)}_l)$, where $ \chi_{l,0}$ is a free-fermion susceptibility. This agrees with the exact forms of the susceptibilities of the $l=0$ order parameters, which correspond to the total charge and the total spin: $$\chi^{c(s)}_{l=0} = \chi_{l=0,0} \frac{m^*/m}{1 + F^{c(s)}_{l=0}},
\label{n1}$$ where $m^* = k_F/v^*_F$. It is tempting to assume that RPA works for a generic order parameter with angular momentum $l$. However, corrections to RPA are of order one when $F^{c(s)}_l = O(1)$, and it is a’priori unclear whether in the generic case the full susceptibility has the same functional form as $\chi^{c(s)}_{l, RPA}$.
One can actually go beyond RPA and obtain the exact expression for a static $\chi^{c(s)}_l$ for a generic order parameter $$\begin{aligned}
\label{eq:rho-def}
\r^{c}_l (\q) &= \sum_{\k,\alpha} \lambda^c_l (k) c^{\dagger}_{\k-\q/2,\alpha}c_{\k+\q/2,\alpha}, \\
{\bf \r}^{s}_l (\q) &= \sum_{\k,\alpha\beta} \lambda^s_l (k) c^{\dagger}_{\k-\q/2,\alpha} {\bf \sigma}^{\alpha\beta}c_{\k+\q/2,\beta},
\end{aligned}$$ with any $l$ and any form-factor $\lambda^{c(s)}_l (k)$. The exact formula was originally obtained by Leggett [@Leggett1965], based on earlier work by Eliashberg [@Eliashberg1962] (we present the diagrammatic derivation in Sec. \[sec:2\]). It reads $$\chi_{l}^{c(s)}= \left(\Lambda^{c(s)}_l Z\right)^2 \chi^{c(s)}_{l,qp}+\chi^{c(s)}_{l,inc}.
\label{eq:full}$$ Here $$\chi^{c(s)}_{l,qp} = \frac{m^*}{m} \chi^{c(s)}_{l,RPA}
\label{eq:full_1}$$ is the RPA result with an extra factor of $m^*/m$, often called the quasiparticle contribution. Other terms in Eq. describe two contributions that incorporate effects beyond RPA. First, $\chi^{c(s)}_{l,qp}$ gets multiplied by $(\Lambda^{c(s)}_l Z)^2$, where $Z$ is the quasiparticle residue and $\Lambda^{c(s)}_l$ accounts for the renormalization of the vertex containing the form-factor. Second, there is an extra term $\chi^{c(s)}_{l,inc}$. These additional terms come from processes in which at least one fermion is located away from the FS.
Although the $Z$-factor itself comes from fermions away from the FS, its presence in (\[eq:full\]) can be easily understood because the fermionic propagator near the FS is $$G (\omega, k) \approx Z G_{qp} (\omega,k) = \frac{Z}{\omega - v^*_F (k-k_F) + i \delta {\text sgn} \omega},
\label{G}$$ hence there is a $Z^2$ factor in each bubble ($Z^2$ also appears in the normalization of $F^{c(s)}_l$, see below). As long as $Z$ is a number $0<Z<1$, it alone does not change the functional form of $\chi^{c(s)}_{l}$ compared to the RPA result. The other two terms are potentially more relevant. First, the vertex function $\Lambda^{c(s)}_l$ may cancel the $1 + F^{c(s)}_l$ term and, second, either $\Lambda^{c(s)}_l$ or $\chi^{c(s)}_{l,inc}$ may diverge on their own and give rise to a Pomeranchuk-type instability not associated with $F^{c(s)}_l = -1$ (and with the fermions on the FS).
Some of these issues were addressed by Landau and Pitaevskii (see e.g. Ref. ) and by Leggett (Ref. ) in the early years of FL theory, by invoking (i) conservation laws and the corresponding Ward identities [@AGD; @Lifshitz1980] and (ii) the continuity equation and the longitudinal sum rule [@Leggett1965; @Ehrenreich1967]. For a conserved order parameter, conservation laws require that the full susceptibility coincides with the coherent term, i.e., the corresponding $\Lambda Z = 1$ and $\chi_{inc} =0$. The $l=0$ charge and spin Pomeranchuk order parameters $\r^c_{l=0}$ and ${\bf \r}^s_{l=0}$ with a [*constant*]{} form-factor are conserved quantities, hence the corresponding $\chi^{c(s)}_{l=0} = \chi^{c(s)}_{l=0,qp}$, as in Eq. (\[n1\]). This is fully consistent with RPA. For $l=1$ charge- or spin-current order parameters with $\lambda^{c(s)}_{l=1} (k) = {\bf k}$ the continuity equation imposes the relation $$Z\Lambda^{c(s)}_{l=1} = \frac{m}{m^*} \left(1 + F^{c(s)}_{l=1}\right)
\label{n2}$$ such that $(Z\Lambda^{c(s)}_{l=1})^2 \chi^{c(s)}_{l,qp} \propto \left(1 + F^{c(s)}_{l=1}\right)$ vanishes at $F^{c(s)}_{l=1}=-1$ instead of diverging. In addition, the longitudinal sum rule yields $$\label{n22}
\chi^{c(s)}_{l=1} = \chi_{1,0},$$ i.e. all interaction-induced renormalizations of $\chi^{c(s)}_{l=1}$ cancel out, implying, $$\chi^{c(s)}_{l=1,inc} = \chi_{l=1,0} \left(1 - \frac{m}{m^*} \left(1 + F^{c(s)}_{l=1}\right)\right)
\label{n3}$$ We emphasize that this holds even in the presence of a lattice potential $V(r)$. For a Galilean-invariant FL, $m^*/m$ by itself is expressed via Landau parameters as $m^*/m = 1 + F^c_{1}$. Then, for spin-current susceptibility, $Z\Lambda^{s}_{l=1} =(1+F^s_1)/(1+ F^c_1)$ and $\chi^{s}_{l=1,inc}= \chi_{l=1,0} (F^c_1 -F^s_1)/(1+ F^c_1)$, while for charge-current susceptibility, $Z\Lambda^{c}_{l=1} =1$ and $\chi^{c}_{l=1,inc}=0$. This last result is consistent with the fact that for a Galilean-invariant FL, charge-current coincides with the momentum and is a conserved quantity.
Eqs. + represent a qualitative breakdown of RPA for spin and charge-current order parameters. Within RPA, $\Lambda Z = 1$, and so to reproduce Eqs. + one must require $m^*/m = (1+ F_1^c)^{-1} = (1+F_1^s)^{-1}$. Such behavior comes about naturally if one assumes [@Quintanilla] that the dressed interaction remains a function of ${\bf k} - {\bf p}$, i.e., $\Gamma^\omega(k, k;p, p) = U_{eff} (|{\bf k} - {\bf p}|)$. In this situation the charge component of $\Gamma^\omega(k, k;p, p)$ is $(U(0) - U(k-p)/2$ and the spin component is $- U(k-p)/2$. Then $F_l^c = F_l^s$, for all $l > 0$, including $l=1$. However, in fact, at order $U^2$ and higher, the interaction gets renormalized in both particle-hole and particle-particle channels, and the renormalized interaction between fermions on the FS depends on both ${\bf k} - {\bf p}$ and ${\bf k} + {\bf p}$. The terms which depend on ${\bf k} - {\bf p}$ and on ${\bf k} + {\bf p}$ behave differently under antisymmetrization, and, as the consequence, spin and charge components of $\Gamma^\omega(k, k;p, p)$ are generally not equivalent for any $l$. In this situation, Eqs. and are obeyed not because of some some special relation between Landau parameters, but rather because $\Lambda Z$ is expressed via the particular combination of Landau parameters, such that for $l=1$ spin current, $\Lambda^s_{l=1} Z$ cancels out $1+F^s_1$.
This issue has been recently re-analyzed by Kiselev et al[@Kiselev2017]. They discussed how the absence of $l=1$ Pomeranchuk instability for the spin-current order parameter places additional constraints on spontaneous generation of spin-orbit coupling[@Wu1; @Wu2], often associated with $l=1$ spin Pomeranchuk order. Kiselev et al also derived a general formula for the susceptibility of a current of a conserved order parameter.
The purpose of the current work is three-fold. First, we provide a transparent diagrammatic derivation of Eq. (\[eq:full\]) and extend it to the case when both $q$ and $\Omega$ are small, but the ratio $v_F q/\Omega$ is arbitrary. Second, we analyze Eqs. (\[n2\]) and (\[n3\]) from a microscopic perspective, and identify what relates the contributions to the susceptibility from fermions near the FS, which determine Landau parameters, and fermions away from the FS, which determine $Z$, $\Lambda^{c(s)}_{l}$, and $\chi^{c(s)}_{l,inc}$ (and $m^*/m$ in the absence of Galilean invariance). Lastly, we investigate how generic is the statement about the absence of Pomeranchuk instabilities for $l=1$ order parameters, and what happens for other $l$.
To derive Eqs. (\[eq:full\]) and (\[eq:full\_1\]) diagrammatically, we use the expansion in the number of fermionic loops, and at each loop order separate the contributions from fermions at the FS and away from it. The contributions away from the FS can be computed by setting $q$ to zero, while for the contributions from the vicinity of the FS one needs to keep ${\bf q}$ small but finite, because each bubble contribution to $\chi^{c(s)}_{l,qp}$ comes from the tiny range near the FS where the poles of the two Green’s function in a bubble are in different half-planes of frequency. We then re-arrange the perturbation series and evaluate partial contributions to the susceptibility with $M =0,1,2$ etc. cross-sections in which the contribution comes from the FS. Summing up terms with all $M$ we reproduce Eqs. (\[eq:full\]) and (\[eq:full\_1\]). We then extend the analysis and consider the dynamical susceptibility $\chi^{c(s)}_{l} ({\bf q}, \Omega)$ in the limit when both $|{\bf q}|$ and $\Omega$ are small, but the ratio $v^*_F |{\bf q}|/\Omega$ is arbitrary. We show that in an arbitrary FL, the form of the dynamical susceptibility is rather complex, except for special cases when $v^*_F |{\bf q}|/\Omega$ is either small or large, or $v^*_F |{\bf q}|/\Omega$ is arbitrary, but only a few Landau parameters are not small.
In order to understand Eqs. (\[n2\]) and (\[n3\]) for current order parameters, we explicitly compute low-energy and high-energy components of charge and spin susceptibilities for $l=1$ for the 2D Hubbard model, to second order in Hubbard $U$. At this order the dressed interaction between fermions becomes dynamical, and both low-energy and high-energy contributions are non-zero. For simplicity, in this calculation we neglect the lattice potential, i.e., consider a Galilean-invariant system. We show that there exists a particular identity on the sum of dynamical polarization bubbles in particle-hole and particle-particle channels, which relates the contribution to this sum coming from fermions at the FS and the one from fermions away from the FS. We use this identity to prove diagrammatically Eqs. (\[n2\]) and (\[n3\]). We do the same computation for $l=0$ and verify that for conserved spin and charge order parameters $\chi^{c(s)}_{l,inc} =0$ and $\Lambda^{c(s)}_l =1/Z$, i.e., $\chi^{c(s)}_{l} = \chi^{c(s)}_{l,qp}$.
We next consider $l=2$ and investigate an argument by Kiselev et al [@Kiselev2017] that for certain order parameters with $l=2$, spin and charge susceptibilities again do not diverge when the corresponding $F^{c(s)}_{l=2} =-1$ because $(1 + F^{c(s)}_{l=2})^{-1}$ in $\chi^{c(s)}_{l=2,qp}$ is canceled out by $Z \Lambda^{c(s)}_{l=2}$. Our perturbative results do not support this claim. We argue that $Z\Lambda^{c(s)}_{l=2}$ cannot be expressed solely in terms of Landau parameters. Of particular interest here is the $l=2$ charge order parameter in a Galilean-invariant case. It is tempting to view this order parameter as a current of conserved $l=1$ total momentum, and relate the corresponding $Z \Lambda^c_{l=2}$ to Landau parameters via the continuity equation. However, we show that the current operator for momentum cannot be expressed solely in terms of bilinear combination of fermions, and contains an interaction-induced four-fermion term. As a result, the $l=2$ charge susceptibility is only a portion of the full current-current correlator, and as such is not determined by the continuity equation.
To understand how generic is the statement about the absence of Pomeranchuk instabilities for $l=1$, we notice that there exists an infinite set of Pomeranchuk order parameters in any channel, including $l=0$. These order parameters contain form-factors $\lambda^{c(s)}_l (k)$, which are obtained by multiplying the base form factor (a constant for $l=0$, ${\bf k}$ for $l=1$, etc), by an arbitrary function $f_l(|{\bf k}|)$. When $f_l (|{\bf k}|)$ is not a constant, it changes the contribution to susceptibility from fermions away from the FS compared to that from fermions at the FS. We argue that the identity, which allowed us to express $Z\Lambda^{c(s)}_{l=1}$ in terms of Landau parameters and cancel $1/(1+ F^{c(s)}_{l=1})$, does not hold if $f_l (|{\bf k}|)$ is not a constant. As a result, $Z\Lambda^{s}_{l=1}$ no longer vanishes when $F^{s}_{l=1} =-1$. We show this non-cancellation explicitly to second order in the Hubbard $U$ by comparing susceptibilities for order parameters with form-factors ${\bf k}$ and $k_F \hat k$, where $\hat k$ is a unit vector directed along ${\bf k}$. We further argue that for a non-constant $f_l(|{\bf k}|)$, the incoherent contribution to susceptibility cannot be expressed via Landau parameters for any $l$, including $l=0$, and can potentially diverge on its own, even if the corresponding Landau parameter is still larger than $-1$. This opens up a possibility for an instability of a FL, not associated with the singularity in the coherent part of the susceptibility.
The paper is organized as follows. In the next Section we review the diagrammatic formulation of FL theory and present our diagrammatic derivation of Eq. (\[eq:full\]) for the static susceptibility $\chi^{c(s)}_l$ and its extension to finite $v_F^*|\q|/\Omega$. In Sec. \[sec:3\] we discuss the forms of $l=1$ susceptibilities for the currents of conserved fermionic charge and spin, and also discuss the relation between the $l=2$ charge order parameter (bilinear in fermions) and the current of a total fermionic momentum. In Sec. \[sec:4\] we present the results of numerical and analytical calculations to second order in the Hubbard $U$. Our key emphasis here is to understand why contributions to the $l=1$ susceptibilities from fermions at the FS and away from it are related. In Sec. \[sec\_new\] we discuss the implications for susceptibilities of order parameters which contain an additional dependence on $k$ beyond symmetry related overall factors. We summarize our results in Sec. \[sec:6\]
FL theory. Diagrammatic approach {#sec:2}
================================
In this section we briefly review the diagrammatic approach to a FL and present the diagrammatic derivation of Eq. (\[eq:full\]). We also obtain a more general expression for $\chi^{c(s)}_{l} (\q,\Omega)$ when both $\q$ and $\Omega$ are small, but the ratio $v^*_F q/\Omega$ is arbitrary. The full formula is $$\chi_{l}^{c(s)} (\q,\Omega) = \left(\Lambda^{c(s)}_l Z\right)^2 \chi^{c(s)}_{l,qp} (\q,\Omega) +\chi^{c(s)}_{l,inc} \label{eq:full_full}$$ For arbitrary $v^*_F q/\Omega$, $\chi^{c(s)}_{l,qp} (\q,\Omega)$ is a complex function of all Landau parameters. For definiteness and to make computational steps less involved, we consider two-dimensional (2D) Galilean-invariant systems.
Our goal is to distinguish between high-energy and low-energy contributions to the susceptibility and relate $\chi^{c(s)}_{l,qp}$, $\chi^{c(s)}_{l,inc}$, $Z$, and $\Lambda^{c(s)}_{l}$ to particular sets of diagrams. We express different contributions to the susceptibility via the vertex function $\Gamma^\omega (k,p)$. Here and below $k$ denotes a fermionic 3-vector, $k = (\k, \omega_k)$ and $q=(\q,\Omega)$ denotes a bosonic 3-vector. We show that $\chi^{c(s)}_{l,qp}$ is expressed via $\Gamma^\omega (k,p)$ in which both $k$ and $p$ are on the FS, $\chi^{c(s)}_{l,inc}$ is expressed via $\Gamma^\omega (k,p)$ in which both $k$ and $p$ are away from the FS, and $\Lambda^{c(s)}_{l}$ is expressed via $\Gamma^\omega (k,p)$ in which $k$ is on the FS and $p$ is away from it, or vice versa. We combine our analysis with the Landau-Pitaevskii equations [@AGD; @Lifshitz1980] which relate the inverse quasiparticle residue $1/Z$ to $\Gamma^\omega (k,p)$ in which $k$ is on the FS and $p$ is away from it, similarly to $\Lambda^{c(s)}_{l}$. The contributions from away from the FS are insensitive to the ratio $\Omega/|\q|$ and can be computed at $\q =0$ and $\Omega =0$. The quasiparticle part $\chi^{c(s)}_{l,qp} (q)$ depends on how the limit $\q,\Omega \to 0$ is taken.
Perturbation theory
-------------------
The free fermion Hamiltonian is $$\begin{aligned}
H_{kin} &= \int dr \sum_\alpha c^\dagger_\alpha (r) \left(\frac{- \nabla^2}{2m} -\mu\right)c_\alpha (r) \nn \\
&= \sum_{\k \alpha} \xi_{\k}
c^\dagger_{k, \alpha} c_{k, \alpha}
\label{n_1}
\end{aligned}$$ where $\xi_{\k} = \k^2/(2m) - \mu$, and we set $\hbar=1$ throughout this paper. The corresponding free-fermion Green’s function is $$G_0(\tk) = \frac{1}{\omega_k - \xi_{\k} + i \delta_\omega}
\label{n_2}$$ where $\delta_\omega = \delta \mbox{sgn} \omega$ and $\delta = 0^+$.
\[fig:1\] {width="3cm"}
The free-fermion susceptibility $\chi^{c(s)}_{l,0} (\tq)$ is diagrammatically represented as the bubble made out of two fermionic propagators (Fig \[fig:1\]) with form-factors $\lambda^{c(s)}_l$ in the vertices: $$\chi^{c(s)}_{l,0} (\tq)=-2 \int\frac{d^3\tk}{(2\pi)^3} \left(\lambda^{c(s)}_l ({\mathbf}{k}) \right)^2 G_0 (\tk + \frac{\tq}{2}) G_0 (\tk-\frac{\tq}{2}),
\label{n_3_1}$$ where the factor 2 comes from spin summation. In 2D, $$\label{eq:form-factor-def}
\lambda^{c(s)}_l ({\mathbf}{k})= \cos{l\phi_{k}} |\k|^l\times f^{c(s)}_l(|\k|)$$ where $\phi_k$ is the angle between ${\bf k}$ and $\bf q$. One may verify that the frequency integral in (\[n\_3\_1\]) is non-zero only if $\xi_{\k+\q/2}$ and $\xi_{\k-\q/2}$ have opposite signs, i.e., it comes from the tiny range near the FS of width $O(q)$. In explicit form we have, after integrating over frequency $$\chi^{c(s)}_{l,0} (\tq)=-2\int\frac{d^2k}{(2\pi)^2}\frac{n_F(\xi_{\k-\frac{\q}{2}})-n_F(\xi_{\k+\frac{\q}{2}})}{\Omega+\xi_{\k-\frac{\q}{2}}-\xi_{\k+\frac{\q}{2}}+i\delta_{\Omega}} \left(\lambda^{c(s)}_l ({\mathbf}{k})\right)^2
\label{n_3}$$ where $n_F(\xi)=\Theta(-\xi)$ is a unit step function in zero temperature limit. In the case of vanishingly small $|\q|$ one can integrate over $k$ and obtain, $$\begin{aligned}
&&\chi_{l,0}^{\cs} (\tq)= - \frac{m}{\pi}
\left(k_F^lf^{c(s)}_l (k_F)\right)^2
\nonumber \\
&& \int \frac{d\phi_{k}}{2\pi} (\cos{l \phi_k})^2 \frac{ v_F |\q| \cos\phi_k}{\Omega - v_F |\q| \cos\phi_k + i \delta_\Omega}
\label{n_4}\end{aligned}$$ In the static limit $\Omega =0, {\bf q} \to 0$ we have $$\begin{aligned}
&&\chi^{c(s)}_{l=0,0} = \frac{m}{\pi} \left(f^{c(s)}_{l=0} (k_F)\right)^2 \nonumber \\
&& \chi^{c(s)}_{l,0} = \frac{m}{2\pi}\left(k_F^lf^{c(s)}_{l} (k_F)\right)^2, l>0
\label{n_14}
\end{aligned}$$ We next include the interaction term $$\begin{aligned}
&& H_{int} = \frac{1}{2} \int d{\bf r} d{\bf r'} \sum_{\alpha,\beta}
c^\dagger_\alpha (r) c_\alpha (r)
U\left(|{\bf r}-{\bf r}'|\right)
c^\dagger_{\beta} (r') c_\beta (r') \nonumber \\
&& = \frac{1}{2V}
\sum U (|\q|)
c^\dagger_{\k+\q/2,\alpha} c^\dagger_{\p-\q/2,\beta} c_{\p+\q/2,\delta} c_{\k-\q/2,\gamma}
\delta_{\alpha \gamma} \delta_{\beta \delta}, \nonumber \\
\label{n_5}
\end{aligned}$$ where the summation is over all momenta and all spin indices.
### First order in $U(|\q|)$ {#sec:first-order-uq-1}
![Corrections to $\chi_l\cs$ to first order in $U$.[]{data-label="fig:chi-1st-order"}](1order.pdf){width="7cm"}
To first order in $U({\bf q})$, there are three interaction-induced corrections to the bubble diagram for the susceptibility. They are shown in Fig. \[fig:chi-1st-order\]. Diagram \[fig:chi-1st-order\]a represents a self-energy correction. The self-energy is purely static (because $U(|\q|)$ is static) and gives rise to mass renormalization $m^*/m = 1 - (1/v_F) d \Sigma/d|\k|$. One can easily verify that the integral for $\Sigma (k)$ for $\k$ near the FS is determined by $\q$ connecting points on the FS. A simple calculation yields $$\frac{m^*}{m} = 1 - \frac{m}{2\pi} \int \frac{d \theta}{2\pi} U \left(2k_F \left|\sin\frac{\theta}{2}\right|\right) \cos{\theta}
\label{n_6}$$ and $$\chi^{c(s)}_{l,2a} (\tq) = \left(\frac{m^*}{m}-1\right) \chi^{c(s)}_{l,0} (\tq)
\label{n_7}$$ Diagram \[fig:chi-1st-order\]b contains two cross-sections with internal $\tk$ and $\tp$. Because the interaction $U(\k-\p)$ is static, in each cross-section the frequency integral is again non-zero only if the dispersions have opposite signs. The result is that the integration is again confined to a narrow region near the FS. Evaluating frequency and momentum integrals, we obtain
\^[c(s)]{}\_[l,2b]{} () = 1 2 ()\^2 (k\_F\^lf\^[c(s)]{}\_l (k\_F))\^2 U(2 k\_F | |)\
\[n\_8\]
In the static limit $\Omega=0$, $\q \to 0$, && \^[c(s)]{}\_[l,2b]{} = 1 2 ()\^2 (k\_F\^lf\^[c(s)]{}\_l (k\_F))\^2\
&& U(2 k\_F ||) Finally, diagram $2c$ contains $U(0)$ and is non-zero only for charge susceptibility at $l=0$. It gives $$\chi^{c}_{l=0,2c} (\tq) = U(0) \left(\chi^{c}_{l=0,0} (\tq)\right)^2
\left(f^c_{l=0} (k_F)\right)^{-2}
\label{n_9}$$ The sum of the three diagrams can be cast into a known FL form by re-expressing the results in terms of the Landau function $F_{\alpha\beta,\gamma\delta} (\k,\p) = (Z^2 m^*/\pi) \Gamma^\omega_{\alpha\beta,\gamma\delta} (k,k;p,p)$, where $\Gamma^\omega_{\alpha\beta,\gamma\delta} (k,k;p,p)$ is the fully renormalized antisymmetrized static interaction between fermions on the FS, taken in the limit of zero momentum transfer. The antisymmetrized interaction to first order in $U$ is shown graphically in Fig. \[fig:int-1st-order\]. To this order, $Z^2 m^*/\pi = m/\pi$. Combining the diagrams from this figure, we obtain $$\begin{aligned}
&&F_{\alpha \beta, \gamma \delta} (\k,\p) = \frac{m}{\pi} \left[U(0)\delta_{\alpha \gamma} \delta_{\beta \delta} - U(\k-\p) \delta_{\alpha \delta} \delta_{\beta \gamma}\right] \label{n_10} \\
&&= \frac{m}{\pi} \left[\left(U(0) - \frac{1}{2} U(\k-\p) \right) \delta_{\alpha \gamma} \delta_{\beta \delta} - \frac{1}{2} U(\k-\p) {\bf \sigma}_{\alpha \gamma} {\bf \sigma}_{\beta \delta}\right] \nonumber
\end{aligned}$$ The two terms in the last line in (\[n\_10\]) are charge and spin components of the Landau function $F_{\alpha \beta, \gamma \delta} (\k,\p) = F^c (\k,\p) \delta_{\alpha \gamma} \delta_{\beta \delta} + F^s (\k,\p) {\bf \sigma}_{\alpha \gamma} {\bf \sigma}_{\beta \delta}$. Each component can be further expanded in partial harmonics with different $l$ as $$F^{c(s)} (\k,\p) = F^{c(s)}_0 + 2 \sum_{l>0} F^{c(s)}_l \cos{l \phi},
\label{n_11}$$ where $\phi = \phi_k - \phi_p$ is the angle between $\k$ and $\p$ ($|\k|=|\p| = k_F$). Using this expansion, one may easily check that the sum of zero-order and first-order contributions to the static susceptibility can be cast into $$\begin{aligned}
\chi^{c(s)}_l
&= \chi_{l,0}\left(1 + F^c_{l=1} - F^{c(s)}_{l}\right) \nn\\
&\approx \chi_{l,0} \left(1 + F^c_{l=1}\right) \left(1 - F^{c(s)}_l\right)
\label{n_12}
\end{aligned}$$ This formula is valid for all $l$, including $l=0$. Eq. trivially fulfils the constraints of Eqs and for the simple reason that to this order, $F_l^c = F_l^s$ for all $l > 0$.
![The vertex $\Gamma^\omega_{\alpha\beta,\gamma\delta}(\k,\k;\p,\p)$ to first order in $U$.[]{data-label="fig:int-1st-order"}](Gamma1storder.pdf){width="7cm"}
### Higher orders in $U(q)$, static limit {#sec:higher-orders-uq}
We now move to higher orders in $U$, still considering the static limit $\Omega =0, \q \to 0$. Within RPA, higher-order diagrams are treated as series of ladder graphs ($l>0$) or ladder and bubble graphs ($l=0$), Each element of the ladder/bubble series contains the product of two fermionic Green’s functions, dressed by static self-energy. The two Green’s functions have the same frequency and their momenta differ by ${\bf q}$. Within this approximation, a non-zero contribution to susceptibility from each cross-section comes from the states very near the FS, where the poles in the two fermionic Green’s functions, viewed as functions of frequency, are shifted in different directions from the real frequency axis. A simple analysis shows that the series is geometric and its sum yields $$\chi^{c(s)}_{l, RPA} = \chi_{l,0} \frac{1 + F^c_{l=1}}{1 + F^{c(s)}_l}
\label{n_16}$$ The RPA susceptibility obviously diverges when $F^{c(s)}_l = -1$, except for the special case of $F_l\cs = F_1^c$, as occurs e.g. for $l=1$ if we require that the interaction is purely static, see the previous section and our comments in the Introduction.
We next go beyond RPA. A diagram for $\chi_l\cs$ at any loop order is represented by a series of ladder segments separated by interactions. In each of these ladders there is an integration over both high-energy and low-energy frequencies and momenta. To obtain $\chi^{c(s)}_{l}$, we follow earlier diagrammatic studies [@Eliashberg1962; @Finkelstein2010; @Chubukov2014a] and and re-arrange perturbation series by assembling contributions to $\chi^{c(s)}_{l}$ from diagrams with a given number $M$ of ladder segments with poles shifted into different directions from the real frequency axis, and then sum up contributions from the sub-sets with different $M =0,1,2$, etc.
We start with $M=0$. The corresponding contributions to the susceptibility contain products of $G^2 (\k, \omega_k)$. Taken alone, each such term will vanish after integration over frequency. The total $M=0$ contribution then vanishes to first order in $U(\q)$ because the static interaction does not affect the frequency integration. However, at second and higher orders in $U(\q)$, the interaction gets screened by particle-hole bubbles and becomes a dynamical one. An example of second-order susceptibility diagram with screened interaction inserted into the bubble is shown in Fig. \[fig:bubble-screening\].
![Example of a higher order contribution to $\chi_l\cs$. At this order, the static interaction acquires dynamics due to particle-hole screening. The diagram’s computation is split into three (see Sec. \[sec:higher-orders-uq\]). It belongs to the $M=0$ sector when both bubbles are evaluated away from the FS, to $M=1$ when one is evaluated on the FS and one away from it, and to $M=2$ when both are evaluated at the FS.[]{data-label="fig:bubble-screening"}](fig4){width="0.3\hsize"}
This screened dynamical interaction contains a Landau damping term, which is non-analytic in both half-planes of complex frequency. As a result, the product of $G^2 (\k, \omega_k)$ and the dressed interaction at order $U^2$ and higher has both a double pole and a branch cut. A pole can be avoided by closing the integration contour in the appropriate frequency half-plane, but the branch cut is unavoidable, and its presence renders the frequency integral finite. Since there is no splitting, relevant fermionic $\omega_k$ and ${\bf k}$ are not confined to the FS and are generally of order $E_F$ (or bandwidth). Fermions at such high energies have a finite damping, i.e., are not fully coherent quasiparticles. By this reason, the $M=0$ contribution to $\chi^{c(s)}_{l}$ is labeled as an incoherent one, $ \chi^{c(s)}_{l, M=0} = \chi^{c(s)}_{l,inc}$ (although at small $U$ fermions with energies of order $E_F$ are still mostly coherent).
![The ladder series of diagrams for the static susceptibility $\chi_l\cs$. The exact $\chi_l$ is represented as a series $M=0,1,2,\ldots$ of bubbles comprised of Green’s functions with poles on opposite halves of the complex frequency plane, i.e. whose contributions are computed close to the FS.[]{data-label="fig:bubble-series"}](Msum){width="\hsize"}
We next move to the $M=1$ sector. Here we select the subset of diagrams with one cross-section, in which we pick up the contribution from $G({\bf k}, \omega_k) G({\bf k}+{\bf q}, \omega_k)$ from the range where the poles in the two Green’s functions are in different half-planes of complex frequency. The sum of such diagrams can be graphically represented by the skeleton diagram in Fig. \[fig:bubble-series\] labeled $M=1$. The internal part of this diagram gives $Z^2 (m^*/m) \chi_{l,0} (\tq)$, where $\chi_{l,0} (\tq)$ is given by (\[n\_4\]). The side vertices contain $\Lambda_1\cs\lambda_l\cs(k_F)$, i.e. the product of the bare form-factor (which we already incorporated into $\chi_{l,0} (\tq)$), and the contributions from all other cross-sections, in which $G({\bf k}, \omega_k) G({\bf k}+{\bf q}, \omega_k)$ is approximated by $G^2({\bf k}, \omega_k)$. These contributions would vanish if we used a static $U(|\q|)$ for the interaction, but again become non-zero once we include dynamical screening at order $U^2$ and higher. Similarly to the $M=0$ sector, the difference $\Lambda_l^{c(s)} -1$ is determined by fermions with energies of order $E_F$. Note, however, that in the $M=0$ sector, all internal energies are of order $E_F$. In the $M=1$ sector, internal energies for the vertices $\Lambda_l^{c(s)}$ are of order $E_F$, but external $\omega_k$ are infinitesimally small, and external ${\bf k}$ are on the FS. Overall, the contribution to the static susceptibility from the $M=1$ sector is $$\chi^{c(s)}_{l, M=1} = \left(Z \Lambda^{c(s)}_l\right)^2 \frac{m^*}{m} \chi_{l,0}\cs
\label{n5}$$ Sectors with $M=2$, $M=3$ are the subsets of diagrams with $2, 3,\ldots$ cross-sections in which we split the poles of the Green’s functions with equal frequencies and momenta separated by ${\bf q}$. In the cross-sections in between the selected ones $G({\bf k}, \omega_k) G({\bf k}+{\bf q}, \omega_k)$ is again approximated by $G^2({\bf k}, \omega_k)$. The contribution from the $M=2$ sector is represented by the skeleton diagram in Fig. \[fig:bubble-series\] labeled $M=2$. It contains fully dressed side vertices $\Lambda^{c(s)}_l$ and a fully dressed anti-symmetrized static interaction between fermions on the FS. One can easily verify that this interaction appears with the prefactor $Z^2 (m^*/m)$, i.e., the extra factor in the $M =2$ sector compared to $M=1$ is the product of $\chi_{l,0}$ and the corresponding component of the Landau function. Using (\[n\_11\]) we then obtain $$\chi^{c(s)}_{l, M=1}+
\chi^{c(s)}_{l, M=2} = \left(Z \Lambda^{c(s)}_l\right)^2 \frac{m^*}{m} \chi_{l,0}\cs \left(1 - F^{c(s)}_l\right)
\label{n6}$$ (the minus sign comes from the number of fermion bubbles.) A simple bookkeeping analysis shows that contributions from sectors with larger $M$ form a geometric series, which transform $1 - F^{c(s)}_l$ into $1/(1 + F^{c(s)}_l)$. Collecting all contributions, we reproduce Eq. (\[eq:full\]).
### The susceptibility $\chi^{c(s)}_l ({\bf q}, \Omega)$ at finite $\Omega/v^*_F|{\bf q}|$. {#sec:susc-chics_l-bf}
We now extend the analysis to the case when both transferred momentum ${\bf q}$ and transferred frequency $\Omega$ are vanishingly small, but the ratio $\Omega/v^*_F |{\bf q}|$ is finite. The computational steps are the same as for static susceptibility. The contribution to $\chi^{c(s)}_l (\tq)$ from the $M=0$ sector and the vertex function $\Lambda^{c(s)}_l$ do not depend on the ratio of $\Omega/(v^*_F |{\bf q}|)$ and remain the same as in the static case. However, the integrand in the expression for $\chi_{l,0} (\tq)$, Eq. (\[n\_4\]), now contains a non-trivial angular dependence via $ v_F|\q|\cos\phi_k/(\Omega - v_F|\q|\cos\phi_k + i \delta_\Omega)$. This makes the computation of series with $M=1,2, \ldots$ more involved.
Consider first the limit $\Omega \ll v_F |\q|$. For even $l$, the free-fermion susceptibility is $$\begin{aligned}
&& \chi^{c(s)}_{l,0}(q) =
\frac{m}{\alpha_l\pi} \left(k^l_F f^{c(s)}_l (k_F)\right)^2 \left(1 + \alpha_l \frac{i\Omega}{v_F |\q|} \right) \nonumber \\
&& =
\chi^{c(s)}_{l,0}\left(1 +
\alpha_l
\frac{i\Omega}{v_F |\q|} \right)\end{aligned}$$ where $\alpha_l=1$ if $l=0$ and $\alpha_l=2$ if $l = 2 m$, $m >0$. For odd $l$, the expansion in $\Omega$ starts with $\Omega^2$. The total contribution from the $M=1$ sector still is proportional to $\chi_{l,0}$: $$\begin{aligned}
&& \chi^{c(s)}_{l,M =1} (\tq) \approx
\frac{m^*}{\alpha_l\pi} \left(Z \Lambda^{c(s)}_l\right)^2 \left(k^l_F f^{c(s)}_l (k_F)\right)^2 \nonumber \\
&& \left(1 + \frac{m^*}{m} \alpha_l\frac{i \Omega}{v_F |{\bf q}|} \right) \nonumber \\
&& = \left(Z \Lambda^{c(s)}_l\right)^2 \frac{m^*}{m}\chi^{c(s)}_{l,0}\left(1 + \frac{m^*}{m}
\alpha_l
\frac{i \Omega}{v_F |{\bf q}|}\right)
\label{n12}\end{aligned}$$ In the contribution from the $M=2$ sector, the $i\Omega/v^*_F |{\bf q}|$ term can be taken from the cross-section on the right or on the left. This gives a combinatoric factor of $2$. Then $$\begin{aligned}
&& \chi^{c(s)}_{l,M =1}(q)+\chi^{c(s)}_{l,M =2} (\tq) \approx
\left(Z \Lambda^{c(s)}_l\right)^2 \frac{m^*}{m} \chi_{l,0}\cs \nonumber \\
&& \left(1 - F^{c(s)}_l + (1 - 2F^{c(s)}_l) \frac{m^*}{m}\alpha_l\frac{i \Omega}{v_F |{\bf q}|}
\right)
\label{n14}\end{aligned}$$ For the contribution from the $M=3$ sector the same reasoning yields the combinatoric factor of 3 and so on. Using $$1 - 2F^{c(s)}_l + 3 \left(F^{c(s)}_l\right) + ... = \frac{1}{(1 + F^{c(s)}_l)^2}$$ we obtain $$\chi^{c(s)}_{l} (\tq) = \left(Z \Lambda^{c(s)}_l\right)^2 \chi^{c(s)}_{l,qp} (\tq) + \chi^{c(s)}_{l,inc}
\label{n16}$$ where to order $\Omega/|\q|$, for even $l$, $$\chi^{c(s)}_{l,qp} (\tq) =
\chi_{l,0}\cs
\left(\frac{m^*/m}{(1 + F^{c(s)}_l)} +
\alpha_l
\frac{i \Omega}{v_F |{\bf q}|}
\left(\frac{m^*/m}{1 + F^{c(s)}_l} \right)^2 \right)
\label{n15}$$ For $l=0$ this result has been obtained before [@Nozieres1999].
In the opposite limit $\Omega \gg v^*_F |{\bf q}|$ we have $$\begin{aligned}
\chi_{l,0} (\tq) \approx - \frac{m}{\pi}
\left(\frac{v_F |{\bf q}|}{\Omega}\right)^2 \left(k^l_F f_l (k_F)\right)^2
\nn\\
\int \frac{d\phi_{k}}{2\pi} (\cos{l \phi_k})^2 (\cos{\phi_k})^2
\label{n7}\end{aligned}$$ The presence of $|\q|^2/\Omega^2$ in the susceptibility for $l=0$ is a natural consequence of the fact that the total fermionic charge and spin are conserved quantities, i.e., they don’t change when we probe the system at different times. For free fermions, this holds for all $l$ because all partial fermionic densities at a given direction of ${\bf k}$ are separately conserved, hence $\chi_{l,0} ({\bf q} =0, \Omega)$ must vanish for an any angle-dependent form-factor. The contribution from the $M=1$ sector is,
\^[c(s)]{}\_[l,M =1]{} () -& (Z \^[c(s)]{}\_l)\^2 ()\^2 (k\^l\_F f\_l (k\_F))\^2\
& ()\^2 ()\^2. \[n8\]
The overall $m/m^*$ factor is due to one $m^*/m$ factor from the integration over momentum and an $(m/m^*)^2$ from the expansion to second order in $v^*_F |\q|/\Omega$. From the $M=2$ sector we have, at order $|\q|^2/\Omega^2$
$$\chi^{c(s)}_{l, M=2} = -
\left(\frac{v_F |{\bf q}|}{\Omega}\right)^2 \left(\frac{m}{\pi}\right)
\left(Z \Lambda^{c(s)}_l\right)^2 \frac{m}{m^*} \left(k^l_F f^{c(s)}_l (k_F)\right)^2
\int \int \frac{d \phi_k}{2\pi} \frac{d \phi_p}{2\pi} (\cos{l \phi_k}) (\cos{l \phi_p}) (\cos{\phi_k}) (\cos{\phi_p})
F^{c(s)} (\phi_k - \phi_p)
\label{n9}$$
Substituting $F^{c(s)}$ from Eq. (\[n\_11\]), we obtain $$\begin{aligned}
&&\chi^{c(s)}_{l=0, M=2} = -\frac{1}{2}\left(\frac{v_F |{\bf q}|}{\Omega}\right)^2 \left(\frac{m}{\pi}\right) \left(Z \Lambda^{c(s)}_l\right)^2 \left(f^{c(s)}_{l=0} (k_F)\right)^2 \frac{m}{m^*}
F^{c(s)}_1 \nonumber \\
&&\chi^{c(s)}_{l=1, M=2} =
-\frac 1 8
\left(\frac{v_F |{\bf q}|}{\Omega}\right)^2 \left(\frac{m}{\pi}\right)
\left(Z \Lambda^{c(s)}_l\right)^2 \left(k_F f^{c(s)}_{l=1} (k_F)\right)^2 \frac{m}{m^*}
\left(2 F_0\cs + F_2\cs\right) \nonumber \\
&&\chi^{c(s)}_{l>1, M=2} = -\frac{1}{8} \left(\frac{v_F |{\bf q}|}{\Omega}\right)^2 \left(\frac{m}{\pi}\right)
\left(Z \Lambda^{c(s)}_l\right)^2 \left(k^l_F f^{c(s)}_{l} (k_F)\right)^2 \frac{m}{m^*} \left( F^{c(s)}_{l-1} + F^{c(s)}_{l+1}\right) \nonumber \\
\label{n10}
\end{aligned}$$ The contribution from the sectors with $M >2$ contains higher power of $|\q|/\Omega$. Hence, to order $|\q|^2/\Omega^2$, the full result for the dynamical susceptibility is $$\begin{aligned}
&&\chi^{c(s)}_{l=0} (\tq) =
-\frac{1}{2}
\left(\frac{v_F |{\bf q}|}{\Omega}\right)^2 \chi^{c(s)}_{l=0,0} \left(Z \Lambda^{c(s)}_l\right)^2 \frac{m}{m^*} \left(1 + F^{c(s)}_1\right)
+ \chi^{c(s)}_{l=0,inc}
\nonumber \\
&&\chi^{c(s)}_{l=1} (\tq) =
-\frac{3}{4}
\left(\frac{v_F |{\bf q}|}{\Omega}\right)^2 \chi^{c(s)}_{l=1,0} \left(Z \Lambda^{c(s)}_l\right)^2 \frac{m}{m^*} \left(1 +
\frac{2}{3} F^{c(s)}_0 + \frac{1}{3} F^{c(s)}_2\right)
+ \chi^{c(s)}_{l=1,inc}
\nonumber \\
&&\chi^{c(s)}_{l>1} (\tq) =
-\frac{1}{2}
\left(\frac{v_F |{\bf q}|}{\Omega}\right)^2 \chi^{c(s)}_{l,0} (\tq) \left(Z \Lambda^{c(s)}_l\right)^2 \frac{m}{m^*} \left(1 + \frac{1}{2} \left( F^{c(s)}_{l-1} + F^{c(s)}_{l+1}\right)\right)
+ \chi^{c(s)}_{l,inc}
\nonumber \\
\label{n11}
\end{aligned}$$
For $l=0$ this result has been obtained in Ref. .
For a generic $\Omega/v_F |{\bf q}|$, the full expression for $\chi^{c(s)}_l (\tq)$ is rather involved for all $l$, including $l=0$. As an illustration, consider the seemingly simplest case $l =0$ and set $f_0 (|k|) =1$ (i.e., consider susceptibilities for spin and charge order parameters). Due to spin/charge conservation $Z \Lambda^{c(s)}_{l=0} =1$ and $\chi^{c(s)}_{l=0,inc} =0$, so $\chi^{c(s)}_{l=0} (\tq) = \chi^{c(s)}_{l=0,qp} (\tq)$.
The full dynamical $\chi^{c(s)}_{l=0,qp} (\tq)$ is given by series of bubbles, each is determined by fermions in the vicinity of the FS. The integration over frequency and over fermionic dispersion can be performed independently in each bubble, but angular integration is, in general, rather involved, because the interaction between the bubbles with internal momenta ${\bf k}$ and ${\bf p}$ is expressed via the Landau function $ F^{c(s)} (\k,\p)$, Eq. (\[n\_11\]), and the latter dependens on $\phi = \phi_k - \phi_p$. It is sufficient to analyze the first few orders in the expansion in powers of $ F^{c(s)} (\k,\p)$ to understand that the full result is $$\chi^{c(s)}_{l=0,qp} (\tq) = \frac{m^*}{\pi} \frac{{\bar \chi} (\tq)}{1 + F^{c(s)}_{l=0} {\bar \chi} (\tq)}
\label{c1}$$ where ${\bar \chi} (\tq)$ is given by series of terms
$$\begin{aligned}
{\bar \chi} (\tq) &=& K_0 - 2 \sum_{n,m >0} F^{c(s)}_{n} K_n K_m \times \nonumber \\
&& \left[\delta_{n,m} - \sum_{m_1 >0} Q_{n,m_1} F^{c(s)}_{m_1}\left[\delta_{m_1,m} - \sum_{m_2 >0} Q_{m_1,m_2} F^{c(s)}_{m_2}
\left(\delta_{m_2,m} - ...\right) \right] \right]
\label{c2}
\end{aligned}$$
where $\delta_{n,m}$ is Kroneker symbol and
\[c3\] $$\label{eq:Qdef}
Q_{n,m} = K_{n+m} + K_{n-m}.$$ Here $$\begin{aligned}
K_n (\tq) &= -\int \frac{d \theta}{2\pi} \cos{n \theta} \frac{v^*_F |\q| \cos {\theta}}{\Omega - v^*_F |\q| \cos{\theta} + i \delta_{\Omega}} \nonumber \\
&= \delta_{n,0} - \frac{\alpha}{\sqrt{\alpha^2-1+i\delta}}(\alpha-\sqrt{\alpha^2-1})^{|n|},
\end{aligned}$$ and $\alpha = \Omega/v_F^*|\q|$.
In explicit form $$\begin{aligned}
&& K_0 (\tq) = 1 - \frac{\Omega}{\sqrt{\Omega^2 - (v^*_F |\q|)^2 + i \delta}} \nonumber \\
&& K_1 (\tq)= \frac{\Omega}{v^*_F |\q|} \left( 1- \frac{\Omega}{\sqrt{\Omega^2 - (v^*_F |\q|)^2 + i \delta}}\right)\nonumber \\
&& K_2 (\tq) =2 \left(\frac{\Omega}{v^*_F |\q|}\right)^2 \nonumber \\
&&
+ \frac{\Omega}{\sqrt{\Omega^2 - (v^*_F |\q|)^2 + i \delta}} \left(1 - 2\left(\frac{\Omega}{v^*_F |\q|}\right)^2\right)
\label{c8}\end{aligned}$$
Eq. (\[c2\]) can be equivalently re-expressed as $${\bar \chi} (\tq) = K_0 - 2 \sum_{n,m >0} F^{c(s)}_{n} K_n K_m S^m_n
\label{c4}$$ where $S^m_n$ is the solution of the matrix equation $$S^m_n + \sum_{m_1 >0} Q_{n,m_1} F^{c(s)}_{m_1} S^m_{m_1} = \delta_{n,m}
\label{c5}$$ In the static limit $K_0 =1$, $K_{n>0} =0$. Then ${\bar \chi} (\tq) =1$, and Eq. (\[c1\]) reduces to Eq. (\[n1\]) for the static susceptibility. For a generic $\Omega/v^*_F |{\bf q}|$ a closed-form expression for $\chi^{c(s)}_{l=0,qp} (\tq)$ can be obtained if only a few Landau parameters are sizable, e.g., if we assume that $|F_l| \ll |F_0|,|F_1|$ for all $l > 1$. In this situation, only one term in each sum in (\[c4\]) and (\[c5\]) survives, and these two equations simplify to $${\bar \chi} (\tq) = K_0 - 2 F^{c(s)}_{1} K^2_1 S^1_1
\label{c6}$$ and $$S^1_1 \left(1 + Q_{1,1} \Gamma^{c(s)}_{1}\right) = 1
\label{c7}$$ Using $Q_{1,1} = K_0 + K_2$ we find $S^1_1 = 1/(1 + ( K_0 + K_2) F^{c(s)}_{1})$. Substituting this into (\[c6\]) and then substituting (\[c6\]) into (\[c1\]), we obtain $$\label{n18}
\chi^{c(s)}_{l=0,qp} (\tq) = \frac{m^*}{\pi} \frac{K_0 - \frac{2 F^{c(s)}_1 K^2_1}{1 + F^{c(s)}_1 \left(K_0 + K_2 \right)}}{1 + F^{c(s)}_0 K_0 - \frac{2 F^{c(s)}_0 F^{c(s)}_1 K^2_1}{1 + F^{c(s)}_1 \left(K_0+ K_2\right)}}$$ The same result has been obtained previously [@MaslovBoltzmann] using a Boltzmann equation approach. At $\Omega/v_F^* |{\bf q}| \gg 1$, we have $K_0 (\tq) \approx -(1/2) (v_F |{\bf q}|/\Omega)^2$, $K^2_1 (\tq) \approx (1/4) (v_F |{\bf q}|/\Omega)^2$, $K_2 (\tq) \approx -(3/8) (v_F |{\bf q}|/\Omega)^2$. Substituting into (\[n18\]) we obtain $\chi^{c(s)}_{l=0} (\tq) = -(1/2) (v_F |{\bf q}|/\Omega)^2 (1 + F^{c(s)}_1)$, as in Eq. (\[n11\]).
Susceptibilities of the currents of conserved order parameters {#sec:3}
==============================================================
In this section we discuss the relationship between order parameters associated with conserved “charges” (to be distinguished from the specific electric charge) and their currents. We review the derivation of the continuity equation for susceptibilities of these order parameters (Refs. and show that this equation explicitly connects high energy properties of a FL, namely $\chi_{l,inc}\cs,\Lambda_l\cs,Z$, with low-energy properties, namely $\chi_{l,qp}$. We discuss the implications for the $l=0,1$ channels and obtain Eqs. -. Finally we discuss the implications of the continuity equation for the $l\geq 1,2$ channels. Our focus here is to identify the constraints placed by the conservation law on high- and low- energy FL properties. We will then analyze these constraints microscopically in Sec. \[sec:4\].
The continuity equation for charge and current susceptibilities {#sec:cont-equat-susc}
---------------------------------------------------------------
A conserved “charge” is an operator $\hat \rho(\q,t)$ that commutes with the Hamiltonian at $\q = 0$, so that it does not evolve in the Heisenberg picture, $$\label{eq:conservation-law}
\frac{\partial\r}{\partial t} = \frac{1}{i}[\r,H] = 0.$$ Examples of such charges are the number (or electric charge) and spin density in the model of Sec. \[sec:2\]: $\r^{c}_{l=0}$ and ${\bf \r}^s_{l=0}$ from Eq. (\[eq:rho-def\]) with constant form-factors. The continuity equation for a conserved charge $\r$ can be derived in the Heisenberg picture: $$\label{eq:drho-dt}
\frac{\partial\r(\q,t)}{\partial t} = \frac{1}{i} [\r(\q,t), H=H_{kin}+H_{int}] \equiv -
i
\q \cdot \J.$$ The continuity equation relates the susceptibilities of order parameters associated with $\hat\rho$ and $\hat\J$, $$\begin{aligned}
\label{eq:susc-defs}
\chi_\rho &= \langle[\r(\q,t),\r(-\q,t')]\rangle \\
\chi_J &= \langle[(\J)^{i}(\q,t),(\J)^{j}(-\q,t')]\rangle\end{aligned}$$ Taking the derivative $\partial_t\partial_{t'}\chi_\rho$ and transforming to the frequency domain we obtain $$\label{eq:cont-susc}
\Omega^2 \chi_\rho(\tq) =
\sum_{m,n}q_m \left[\chi^{mn}_J(\tq) -\chi^{mn}_J(\q,0) \right] q_n.$$ Here, the sum is over spatial indices $m,n=\{x,y\}$. Equivalently we may write, $$\label{eq:cont-susc-par}
(\Omega/q)^2\chi_\rho(\tq) =
\chi_J^\parallel(\tq) - \chi_J^\parallel(\q,0).$$ Here we have defined the longitudinal component of the susceptibility $\hat q \cdot \chi_J \cdot \hat q$. Note, that the RHS of Eqs. + includes only the time dependent part of $\chi_J$. This is an automatic consequence of taking the time derivative of $\chi_\rho$ and going to the Fourier domain.
Let’s assume that both $\r$ and $J$ are expressed via bilinear combinations of fermions with some given $l$. We then can use Eq. and write $$\begin{aligned}
\label{eq:full-rho}
\chi_\rho (q,\Omega) = (\Lambda_\rho Z)^2\chi_{\rho,qp} (q,\Omega) + \chi_{\rho,inc},\\
\chi_J (q, \Omega) = (\Lambda_J Z)^2\chi_{J,qp} (q,\Omega) + \chi_{J,inc}, \label{eq:full-j}\end{aligned}$$ Combining these expressions and Eq. we express the current susceptibility via the susceptibility of a conserved charge.
Implication of conservation laws for the susceptibilities {#sec:impl-cons-laws}
---------------------------------------------------------
For a conserved charge yields $$\label{eq:chi-rho-vanish}
\chi_\rho(\q=0,\Omega) = 0.$$ We also we recall that the coherent part of $\chi_\rho$, which corresponds to the $M=1,2,\ldots$ diagrams of Fig. \[fig:bubble-series\], vanishes at $\q = 0$. Thus, Eq. also implies $$\label{eq:chi-rho-inc-zero}
\chi_{\rho,inc} = 0.$$ Finally, the relation $\Lambda_\rho Z =1$ follows from the fact that $\Lambda_\rho$ and $1/Z$ are identically expressed via the vertex $\Gamma^\omega$,
![Relation between a 3-leg vertex $\Lambda$ and a 4-leg vertex $\Gamma$, for a conserved charge density.[]{data-label="fig:3-4vertex"}](3-4vertex.pdf){width="7cm"}
$$\begin{aligned}
\label{eq:xxxx1}
&\Lambda_\rho = 1 - \frac{i}{2 k_F}\sum_{\alpha\beta}\int \frac{d^3 k}{(2\pi)^3} \Gamma^\omega_{\alpha\beta,\alpha\beta}(k_F\hat p, \k)(G_k^2)^\omega \frac{\lambda_\rho (k)}{\lambda_\rho (k_F)} \\
& \frac{1}{Z} = 1 - \frac{i}{2 k_F}\sum_{\alpha\beta}\int \frac{d^3 k}{(2\pi)^3} \Gamma^\omega_{\alpha\beta,\alpha\beta}(k_F\hat p, \k)(G_k^2)^\omega \frac{\lambda_\rho (k)}{\lambda_\rho (k_F)}
\label{eq:xxxx2}
\end{aligned}$$
where $(G_q^2)^\omega = \mbox{lim}_{\Omega\to 0} G(\q,\omega)G(\q,\omega+\Omega)
= G^2(\q,\omega)$ is the regular part of the product of two Green’s functions, For the vertex, Eq. (\[eq:xxxx1\]) follows from Fig. \[fig:3-4vertex\] (and is valid for a conserved “charge” in both charge and spin channels, while for $1/Z$ the relation is the Ward identity for a conserved charge with form-factor $\lambda_\rho (k)$. We recall that $\Lambda_\rho$ is defined without the factor $\lambda_\rho (k_F)$.
We plug these results into Eq. , take the limit $\Omega \gg v_F q \to 0$, and obtain, $$\label{eq:chi-j-determined}
(\Lambda_JZ)^2\chi_{J,qp}(\q \to 0,0) =
-
\frac{\Omega^2}{q^2} \chi_{\rho,qp}(\frac{q}{\Omega} \to 0).$$ We showed in Sec. \[sec:2\] that for any $l$, $\chi^{c(s)}_{l, qp}(\frac{q}{\Omega} \to 0)$ scales as $q^2/\Omega^2$, and the prefactor is expressed in terms of Landau parameters and is not singular. Assuming that this holds for the conserved charge, we find that $(\Lambda_JZ)^2\chi_{J,qp}(\q \to 0,0)$ remains finite when Landau parameters change and pass through $-1$. Eq. then implies that there is no Pomeranchuk instability in the $J$ channel. It also explicitly connects $\Lambda_\rho,\Lambda_J, Z, m^*/m$ and $\chi_{\rho,qp},\chi_{J,qp}$ via Eqs. +. This is the essence of our argument that the continuity equation implies constraints that connect low- and high- energy properties of the FL.
For the specific case of spin and charge density order parameters, one can easily verify that $\r^c (q)$ and $\r^s (q)$ commute with $H_{int}$ so the current density is bilinear in the creation and annihilation operators: $$\begin{aligned}
\label{eq:j-def-bilin}
\J^{c}(\q,t) &= \frac{1}{m}\sum_{\k,\alpha}\k c^\dagger_{\k-\q/2,\alpha}c_{\k+\q/2,\alpha}, \\
\J^{s}_m(\q,t) &= \frac{1}{m}\sum_{\k,\alpha\beta}\sigma^{\alpha\beta}_m\k c^\dagger_{\k-\q/2,\alpha}c_{\k+\q/2,\beta}.\end{aligned}$$ In this case, the susceptibilities of $\r\cs,\J\cs$ correspond precisely to $\chi_{l=0}$ and $\chi_{l=1}$: $$\label{eq:rho-j-susc-l}
\chi\cs_\rho = \chi_{l=0} \cs,\quad \chi\cs_J = \chi_{l=1} \cs.$$ Eq. then implies $$\label{eq:cont-susc-l01}
(\Omega/q)^2\chi_{l=0}\cs(\tq) =
\frac{1}{m^2} \hat q\cdot \left[\chi\cs_{l=1}(\tq) -\chi\cs_{l=1}(\q,0) \right] \cdot \hat q.$$
Taking the $\Omega \gg v^*_F q$ limit, we obtain $$\label{eq:chi-0-coh-expansion}
\chi_{l=0}\cs
=
-
\chi_{l=0,0}\frac{v_F}{v_F^*}\left(\frac{v_F^*|\q|}{\Omega}\right)^2(1+F_1\cs) + O(|\q|^4/\Omega^4)$$ Plugging the result into Eq. yields, $$\label{eq:charge-current-coh-equal}
(\Lambda^{c,s}_{l=1}Z)^2\frac{v_F}{v_F^*}{}\frac{1}{1+F_1^{c,s}} = \frac{v_F^*}{v_F}(1+F_1\cs),$$ i.e., $$\label{eq:lambda-1-z-constraint}
\Lambda_{l=1}\cs Z =\frac{v_F^*}{v_F}(1+F_1\cs),$$ which is Eq. .
For the currents of conserved charge and spin there exists another constraint imposed by the longitudinal sum rule[@Ehrenreich1967; @Leggett1965]: $$\label{eq:chi-j-sum-rule}
\chi_J^\parallel(\q,0) = n/m$$ where $n$ is the number density. The longitudinal sum rule is analogous to the longitudinal f-sum rule for the imaginary part of the inverse dielectric function [@Mahan] and can be derived from the gauge-invariance of the electromagnetic field [@Ehrenreich1967]. It is exact for a system where the electric current is proportional to the momentum density (with or without Galilean invariance), which is the case for any model of the form of Eqs. , with or without external potential $V(r)$. In effective low energy models (e.g. on a lattice), it is only approximately correct[@Kiselev2017]. In either case, its implication is that the total $\chi_J$ is also finite.
Conservation of momentum and $l=2$ susceptibility {#sec:cons-moment-l=2}
-------------------------------------------------
Finally, we address the issue of the implication of the continuity equation for momentum in a Galilean invariant system. In this section we will refer to the momentum density by the symbol $\rho \equiv \rho_i$, and to the energy tensor by $J \equiv J_{ij}$ where $i,j$ denote spatial indices.
In Sec. \[sec:cont-equat-susc\] we did not specify the nature of charge density and current. Thus, eq. is equally valid for the momentum densities and currents, the only change being that $\chi_\rho = \chi_\rho^{ij}(\q,\Omega)$ is a rank-2 symmetric tensor, and so is $(\chi_J^\parallel)^{ij} = (\hat q \cdot \chi_{J}(\q,\Omega)\cdot \hat q)^{ij}$. In the same manner, all arguments relating high frequency behavior of $\chi_\rho$ with the static behavior of $\chi_J$ go through, leading to Eq. . Thus $(\Lambda_J Z)^2\chi_{J,qp}$ is fully determined by $\chi_\rho$ and furthermore is always finite.
However, we now demonstrate that $J_{ij}$ cannot, in general, be expressed as a bilinear operator in $c^\dagger,c$. As a result, $\chi_J^\parallel$ does not have a simple relationship with $\chi_l$, e.g. with $\chi_{l=2}$. To see this, it is enough to examine the Hubbard model, i.e. take $U(|\q|) = U$ in Eq. . The current operator Eq. has the following equation of motion, $$\begin{aligned}
\label{eq:dJ-dt}
\frac{\partial\hat\rho(\q,t)}{\partial t} = -i\q \cdot\hat J\end{aligned}$$ where $$\begin{aligned}
\label{eq:J-momentum-def}
\hat J = \hat{J}_{kin} + \hat{J}_{int},\end{aligned}$$ with $$\label{eq:j2-defs}
\q \cdot \hat J_{kin} = \left[\rho,H_{free}\right],\quad \q \cdot \hat J_{int} = \left[\rho, H_{int}\right]$$ which gives, $$\begin{aligned}
\label{eq:j-kin-def}
\hat J^{ij}_{kin} &= \frac{1}{m^2}\sum_{\k}k_ik_jc^\dagger_{\k-\q/2}c_{\k+\q/2},\\
\hat J^{ij}_{int} &= \delta_{ij}\frac{U}{m^2}\sum_{\k}n(\k)n(\q-\k),\end{aligned}$$ where $n(\k) = \sum_{\p}c^\dagger_{\p-\k/2}c_{\p+\k/2}$. If we had had $\hat J_{int} = 0$, then indeed Eq. could be used to constrain the $l=0,l=2$ channels, both of which appear in $\hat J_{kin}$. However, as it is, while Eq. does constraint $\chi_J$ to be finite, by itself it does not constrain any specific $l$ channels.
Perturbative calculations for the Hubbard model: charge-current ad spin-current order parameters. {#sec:4}
=================================================================================================
In this section we perform perturbative analysis of Eq. for $l=1$ and Eq. . We have three goals in our calculation: the first is to show how one can derive the continuity equation diagrammatically, the second is to verify the relations between $\Lambda^{c(s)}_{l=1} Z, \chi^{c(s)}_{l=1,inc}$ and $F^{c(s)}_{l=1}$, Eqs. and , in direct expansion in the interaction, and the third goal is to clarify the origin of the relation between high- and low- energy contributions to Eqs. and .
We proceed in three steps. First, we derive Eq. diagrammatically to first-order in $U(\q)$. We will see that although there are no dynamical corrections to this order (i.e. $Z,\Lambda^{c(s)}_l = 1$), nevertheless self-energy corrections are crucial, indicating one should go beyond RPA. Then, we perform a combined analytical and numerical analysis of $\chi^{c(s)}_{l=1}$ at order $U^2$ for the Hubbard model, and explicitly verify Eqs. , . Going to to second order in $U$ is essential, because only at this order do contributions away from the FS begin to accumulate, see Sec. \[sec:higher-orders-uq\]. Finally, we demonstrate that the high-energy contributions to $\chi\cs_{l=1}$ can be re-expressed as low-energy ones, due to a special property of the sum of particle-hole and particle-particle bubbles.
Diagrammatic derivation of the continuity equation {#sec:first-order-uq}
--------------------------------------------------
In this subsection we show how Eq. (\[eq:cont-susc-l01\]) can be reproduced in a diagrammatic calculation. Already at this order we will see that one needs to treat self-energy and vertex corrections on equal footings because the continuity equation emerges due to a particular cancellations between these two types of corrections.
To begin with, we re-write Eq. for free-fermion susceptibility for a current order parameter with $\lambda^{c(s)}_{l=1} (k) = {\bf k}\cdot \hat q$ as $$\begin{aligned}
\label{eq:chi1-zero-order}
q^2\chi_{l=1,0}\cs(q)
=-2\int\frac{d^2k}{(2\pi)^2}\frac{n_F(\xi_{\k-\frac{\q}{2}})-n_F(\xi_{\k+\frac{\q}{2}})}{\Omega-\frac{1}{m}\k\cdot\q+i\delta_{\Omega}}(\k\cdot\q)^2 \end{aligned}$$ Here and later on we omit the $\parallel$ symbol for clarity. We then rewrite the form factor as: $$\label{eq:add-sub}
(\k\cdot\q)^2=(\k\cdot\q+m\Omega)(\k\cdot\q-m\Omega)+m^2\Omega^2$$ and obtain $$\begin{aligned}
q^2\chi\cs_{l=1,0}(q) &= 2m^2\int\frac{d^2k}{(2\pi)^2}[n_F(\xi_{\k-\frac{\q}{2}})-n_F(\xi_{\k+\frac{\q}{2}})] \nn\\
&\qquad \times \left[-\Omega -\frac{1}{m}\k\cdot \q + \frac{ \Omega^2 }{\Omega-\frac{1}{m}\k\cdot\q+i\delta_\Omega}\right]
\label{ll2}\end{aligned}$$ The $\Omega$ term vanishes after integration over $k$. The other two terms are easily identified as $q^2\chi_{l=1,0}\cs(\q,0)$ and $\Omega^2\chi_{l=0,0}\cs(q)$, so that: $$\label{eq:chi-zero-order-cont}
\frac{q^2}{m^2}\left(\chi\cs_{l=1,0}(q) - \chi\cs_{l=1,0}(\q,0)\right) = \Omega^2\chi\cs_{l=0,0}(q).$$
We now use the same tactics for first order corrections to $\chi\cs_{l=1}$. The corresponding diagrams are given in Fig. \[fig:chi-1st-order\]. Diagram \[fig:chi-1st-order\]c, the RPA correction, gives
\[eq:chi-bubble-u1\] \_[l=1,2c]{}= \^2\
U()
+ By making use of $$\label{eq:n64}
\frac{\k\cdot\q}{\Omega-\frac{1}{m}\k\cdot\q+i\delta_{\Omega}}=\frac{m\Omega}{\Omega-\frac{1}{m}\k\cdot\q+i\delta_{\Omega}}-m$$ we find $$\label{eq:delta-chi-0-cont}
\frac{q^2}{m^2}
\chi_{l=1,2c}\cs(\tq) = \Omega^2 \chi_{l=0,2c}\cs(\tq)$$ Note that there is no need to subtract the static part because $\chi_{l=1,2c}\cs({\bf q}, 0)$ vanishes.
For the remaining two diagrams in Fig. \[fig:chi-1st-order\] we obtain
$$\begin{aligned}
\label{eq:RPA-MT-corrections}
q^2 \chi\cs_{l=1,2a}(\tq) &= -2\int\frac{d^2k}{(2\pi)^2}\frac{d^2p}{(2\pi)^2}\frac{[n_F(\xi_{\p-\frac{\q}{2}})-n_F(\xi_{\p+\frac{\q}{2}})] [n_F(\xi_{\k-\frac{\q}{2}})-n_F(\xi_{\k+\frac{\q}{2}})]}{(\Omega-\frac{1}{m}\k\cdot\q+i\delta_{\Omega})^2}U(|\p-\k|) (\k \cdot \q)^2\\
\label{eq:2b-finite-q}
q^2 \chi\cs_{l=1,2b}(\tq) &= 2\int\frac{d^2k}{(2\pi)^2}\frac{d^2p}{(2\pi)^2}\frac{[n_F(\xi_{\p-\frac{\q}{2}})-n_F(\xi_{\p+\frac{\q}{2}})][n_F(\xi_{\k-\frac{\q}{2}})-
n_F(\xi_{\k+\frac{\q}{2}})]}{(\Omega-\frac{1}{m}\p\cdot\q+i\delta_{\Omega})(\Omega-\frac{1}{m}\k\cdot\q+i\delta_{\Omega})}U(\p-\k) (\k \cdot \q) (\p \cdot \q)
\end{aligned}$$
Applying again we find, $$\frac{q^2}{m^2}(\chi_{l=1,2a}\cs(\tq)+\chi_{l=1,2b}\cs(\tq)) = \Omega^2 (\chi_{l=0,2a}\cs(\tq)+\chi_{l=0,2b}\cs(\tq))
\label{ll1}$$ The static part of the sum of the two contributions cancel out. Eqs. and verify Eq. to order $U$.
We emphasize that $\chi\cs_{l=1,2a}(\tq)$ and $\chi\cs_{l=1,2b}(\tq)$, when taken separately, do not satisfy the continuity equation , and only the sum of the two terms obeys (\[ll1\]). This is an indication that, within diagrammatics, the continuity equation emerges due to fine cancellations between self-energy and vertex corrections, and one should go beyond RPA at each order in $U$ to reproduce it.
{width="17cm"}
Evaluation of $\chi_{l=1}\cs$ to order $U^2$ {#sec:eval-chi_l-1cs-order}
--------------------------------------------
We now present the results of explicit calculations of the static susceptibilities to order $U^2$. We identify contributions to $\chi^{c(s)}_{l=1,inc}$ and $\left(\Lambda^{c(s)}_{l=1} Z\right)^2$, and $\chi^{c(s)}_{l,qp}$ from each diagram, and compute them by a combination of analytical and numerical methods. We also independently compute the vertex renormalization $\Lambda\cs_{l=1}$ to order $U^2$.
There are nine different diagrams for the current susceptibility to second order in $U(q)$, see Fig. \[fig:2nd-order\]. To simplify the numerics, we approximate $U(q)$ by a constant $U$ i.e., consider $U^2$ renormalizations in the Hubbard model. For a constant $U$, Landau parameters $F^{c(s)}_{l}$ also only emerge at order $U^2$, i.e., the incoherent part of the susceptibility, vertex renormalizion, renormalization of the quasiparticle $Z$, and Landau parameters are all of order $U^2$. We make use of previously known results [@Chubukov] $$\label{eq:mmstar-z}
F_1^c=-F_1^s=\frac{m^2U^2}{8\pi^2},\qquad Z = 1 - 1.39 \frac{m^2U^2}{8\pi^2},$$ and $$\label{eq:mstar-m}
\frac{m^*}{m} = 1 + F_1^c = 1 + \frac{m^2U^2}{8\pi^2}$$ which holds for a Galilean-invariant system [@AGD; @Lifshitz1980]. The order $U^2$ is the first one in perturbative expansion at which differences between the charge and spin channels emerge, in the form of the Aslamazov-Larkin (AL) diagrams, Figs. \[fig:2nd-order\]c,d. The AL diagrams contribute in the charge channel and vanish in the spin channel, as can be seen from direct spin summation.
Consider the charge channel first. It is straightforward to identify the diagrams in Figs. \[fig:2nd-order\], which give equal contributions, up to overall factor. One can easily verify that $\chi_{6a} = -2 \chi_{6g},~ \chi_{6d} = -2 \chi_{6b}$, and $\chi_{6e} = -\chi_{6f}$. In addition, using the relation $$\label{eq:g-relation}
\int d\omega_{p} (G_{\tp-\frac{\tq}{2}}^0)^3G^0_{\tp+\frac{\tq}{2}}=-\frac{1}{2}\int d\omega_p (G^0_{\tp-\frac{\tq}{2}}G^0_{\tp+\frac{\tq}{2}})^2,$$ we find $\chi_{6h} = -\chi_{6i}$. In Eq. and throughout this section we denote $G_0(k) \equiv G^0_k$ for compactness. Summing up the contributions to the charge-current susceptibility, we obtain at order $U^2$, \^c\_[l=1]{} = \_[6a]{} + \_[6c]{} + \_[6d]{} \[ll3\] A similar consideration for the spin susceptibility yields \^s\_[l=1]{} = (\_[6a]{} - \_[6d]{}) \[ll4\] In explicit form
$$\label{eq:all-diagrams}
\begin{aligned}
\chi_{6a}&=
8 U^2 \int \frac{d^3k d^3k' d^3p}{(2\pi)^9} (\p\cdot\hat{q})^2 (G^0_{\tp-\frac{\tq}{2}})^2G^0_{\tp+\frac{\tq}{2}}G^0_{\tp-\tk}G^0_{\tk'-\tk}G^0_{\tk'-\frac{\tq}{2}},\\
\chi_{6c}&=
4 U^2 \int \frac{d^3k d^3k' d^3p}{(2\pi)^9} (\p\cdot\hat{q}) (\k'\cdot\hat{q})
G^0_{\tp-\frac{\tq}{2}}G^0_{\tp+\frac{\tq}{2}}G^0_{\tk'-\frac{\tq}{2}}G^0_{\tk'+\frac{\tq}{2}}G^0_{\tp-\tk}G^0_{\tk'-\tk},\\
\chi_{6d}&=
4 U^2 \int \frac{d^3k d^3k' d^3p}{(2\pi)^9} (\p\cdot\hat{q}) (\k'\cdot\hat{q}) G^0_{\tp-\frac{\tq}{2}}G^0_{\tp+\frac{\tq}{2}}G^0_{\tk'-\frac{\tq}{2}}G^0_{\tk'+\frac{\tq}{2}}G^0_{\tp-\tk}G^0_{\tk'+\tk}.
\end{aligned}$$
We set $\Omega=0$ and take ${\bf q}$ to be small but finite. After integration over frequency, we split each diagram into three parts: “high”, “middle”, and “low” (which we label “H”, “M”, and “L”), depending on whether zero, one, or two internal fermionic momenta are confined to the FS, e.g. $\chi_{6a}=\chi^{H}_{6a}+\chi^M_{6a}+\chi^L_{6a}$. In this computational scheme, AL diagrams contain “H”, “M”, and “L” parts, while the diagram with self-energy renormalization contains “H” and “M” parts. In explicit form we have
\[eq:all\] $$\begin{aligned}
\chi^H_{6a}&= -8 U^2 \int \frac{d^2k d^2k' d^2p}{(2\pi)^6} (\p\cdot\hat{q})^2
\frac{[n_F(\xi_{{\mathbf}{k'}})-n_F(\xi_{{\mathbf}{k'-k}})] [n_F(\xi _{{\mathbf}{p}})-n_F(\xi _{{\mathbf}{p-k}})][n_B(\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'-k}})-n_B(\xi _{{\mathbf}{p}}-\xi _{{\mathbf}{p-k}})]}{(\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'-k}}-\xi _{{\mathbf}{p}}+\xi _{{\mathbf}{p-k}})^3}\\
\chi^H_{6c}&= +8 U^2 \int \frac{d^2k d^2k' d^2p}{(2\pi)^6} (\p\cdot\hat{q}) (\k'\cdot\hat{q})
\frac{[n_F(\xi_{{\mathbf}{k'}})-n_F(\xi_{{\mathbf}{k'-k}})] [n_F(\xi _{{\mathbf}{p}})-n_F(\xi _{{\mathbf}{p-k}})][n_B(\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'-k}})-n_B(\xi _{{\mathbf}{p}}-\xi _{{\mathbf}{p-k}})]}{(\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'-k}}-\xi _{{\mathbf}{p}}+\xi _{{\mathbf}{p-k}})^3}\\
\chi^H_{6d}&= -8 U^2 \int \frac{d^2k d^2k' d^2p}{(2\pi)^6} (\p\cdot\hat{q}) (\k'\cdot\hat{q}) \frac{[n_F(\xi_{{\mathbf}{k'}})-n_F(\xi_{{\mathbf}{k'+k}})] [n_F(\xi _{{\mathbf}{p}})-n_F(\xi _{{\mathbf}{p-k}})][n_B(-\xi _{{\mathbf}{k'}}+\xi _{{\mathbf}{k'+k}})-n_B(\xi _{{\mathbf}{p}}-\xi _{{\mathbf}{p-k}})]}{(\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'+k}}+\xi _{{\mathbf}{p}}-\xi _{{\mathbf}{p-k}})^3} \\
\chi^M_{6a}&= -4 U^2 \int \frac{d^2k d^2k' d^2p}{(2\pi)^6} (\p\cdot\hat{q})^2
\left(1+\frac{|{\mathbf}{k'}|\cos\phi_{{\mathbf}{k'}}}{|{\mathbf}{p}|\cos\phi_{{\mathbf}{p}}}\right)n'_F(\xi_{{\mathbf}{p}})\frac{[n_F(\xi_{{\mathbf}{k'}})-n_F(\xi_{{\mathbf}{k'-k}})] [n_B(\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'-k}})-n_B(\xi _{{\mathbf}{p}}-\xi _{{\mathbf}{p-k}})]}{(\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'-k}}-\xi _{{\mathbf}{p}}+\xi _{{\mathbf}{p-k}})^2}\\
\chi^M_{6c}&= +8 U^2 \int \frac{d^2k d^2k' d^2p}{(2\pi)^6} (\p\cdot\hat{q}) (\k'\cdot\hat{q})
n'_F(\xi_{{\mathbf}{p}})\frac{[n_F(\xi_{{\mathbf}{k'}})-n_F(\xi_{{\mathbf}{k'-k}})] [n_B(\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'-k}})-n_B(\xi _{{\mathbf}{p}}-\xi _{{\mathbf}{p-k}})]}{(\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'-k}}-\xi _{{\mathbf}{p}}+\xi _{{\mathbf}{p-k}})^2}\\
\chi^M_{6d}&= +8 U^2 \int \frac{d^2k d^2k' d^2p}{(2\pi)^6} (\p\cdot\hat{q}) (\k'\cdot\hat{q})
n'_F(\xi_{{\mathbf}{p}})\frac{[n_F(\xi_{{\mathbf}{k'}})-n_F(\xi_{{\mathbf}{k'+k}})] [n_B(-\xi _{{\mathbf}{k'}}+\xi _{{\mathbf}{k'+k}})-n_B(\xi _{{\mathbf}{p}}-\xi _{{\mathbf}{p-k}})]}{(\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'+k}}+\xi _{{\mathbf}{p}}-\xi _{{\mathbf}{p-k}})^2}\\
\chi^L_{6a}&= +4 U^2 \int \frac{d^2k d^2k' d^2p}{(2\pi)^6} (\p\cdot\hat{q})^2
\frac{|{\mathbf}{k'}|\cos\phi_{{\mathbf}{k'}}}{|{\mathbf}{p}|\cos\phi_{{\mathbf}{p}}}n'_F(\xi_{{\mathbf}{p}})n'_F(\xi_{{\mathbf}{k'}})\frac{n_B(\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'-k}})-n_B(\xi _{{\mathbf}{p}}-\xi _{{\mathbf}{p-k}})}{\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'-k}}-\xi _{{\mathbf}{p}}+\xi _{{\mathbf}{p-k}}}\\
\chi^L_{6c}&= -4 U^2 \int \frac{d^2k d^2k' d^2p}{(2\pi)^6} (\p\cdot\hat{q}) (\k'\cdot\hat{q})
n'_F(\xi_{{\mathbf}{p}})n'_F(\xi_{{\mathbf}{k'}})\frac{n_B(\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'-k}})-n_B(\xi _{{\mathbf}{p}}-\xi _{{\mathbf}{p-k}})}{\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'-k}}-\xi _{{\mathbf}{p}}+\xi _{{\mathbf}{p-k}}}\\
\chi^L_{2d}&= -4 U^2 \int \frac{d^2k d^2k' d^2p}{(2\pi)^6} (\p\cdot\hat{q}) (\k'\cdot\hat{q})
n'_F(\xi_{{\mathbf}{p}})n'_F(\xi_{{\mathbf}{k'}})\frac{n_B(-\xi _{{\mathbf}{k'}}+\xi _{{\mathbf}{k'+k}})-n_B(\xi _{{\mathbf}{p}}-\xi _{{\mathbf}{p-k}})}{\xi _{{\mathbf}{k'}}-\xi _{{\mathbf}{k'+k}}+\xi _{{\mathbf}{p}}-\xi _{{\mathbf}{p-k}}}
\end{aligned}$$
Here $n_B(\xi) = -\Theta(-\xi)$ at $T=0$.
The “H” contributions can be evaluated by just setting $\Omega=0$ and $\q=0$ in Eq., e.g. $$\chi^H_{6a}=8 U^2 \int \frac{d^3k d^3k' d^3p}{(2\pi)^9} {{\mathbf}{p}}^2
(G^0_{\tp})^3G^0_{\tp-\tk}G^0_{\tk'-\tk}G^0_{\tk'}$$ The sum of “H” parts is then the incoherent part of the susceptibility $$\delta \chi_{l=1}^{c(s),H} = \chi_{l=1,inc}\cs$$ The “M” and “L” parts determine $$\begin{aligned}
\label{lambdaz}
&& \delta \chi_{l=1}^{c(s),M} = \chi_{l=1,0}\cs \left( \frac{m^*}{m} \left(\Lambda^{c(s)}_{l=1} Z\right)^2 -1\right) \nonumber \\
&& \delta \chi_{l=1}^{c(s),L} = - \chi^{c(s)}_{l=1,0} F^{c(s)}_{l=1}
\end{aligned}$$ The “L” part can be computed analytically and yields $$\label{eq:bl-al1l-res}
\chi^L_{6a} = \chi^L_{6c} = 0,$$ and $$~\chi^L_{6d} = -\frac{m^2U^2}{4\pi^2}\chi_{l=1,0}$$ where $\chi_{l,0}$ is a free-fermion susceptibility, given by . Using (\[ll3\]), (\[ll4\]) and (\[eq:mmstar-z\]), we find $\delta \chi_{l=1}^{c(s),L} = - \chi^{c(s)}_{l=1,0} F^{c(s)}_{l=1}$ as in (\[lambdaz\]). The “M” and “H” terms in Eq. are high dimensional principal value integrals, which we evaluate numerically. Details of our numerics can be found in the Appendix.
According to Eqs. and , the total “H” contributions to charge-current susceptibility, $\delta \chi^c_{l=1,H}$ should vanish, while other contributions should obey, to order $U^2$, $$\begin{aligned}
\label{eq:deviations-cs}
&&\delta \chi^c_{l=1,M} = \left(\frac{m}{m^*} (1+ F^c_{1})^2 -1\right) \chi_{l=1,0} \approx F^c_{1} \chi_{l=1,0} \nonumber \\
&&\delta \chi^s_{l=1,M} = \left(\frac{m}{m^*} (1+ F^s_1)^2 -1\right) \chi_{l=1,0} \approx (2 F^s_{1} - F^c_1) \chi_{l=1,0} \nonumber \\
&&\delta \chi^s_{l=1,H} = \left(\frac{m^*}{m} -1 - F^s_{l}\right) \chi_{l=1,0} \approx (F^c_{1} - F^s_1) \chi_{l=1,0}\end{aligned}$$ Using Eq. (\[eq:mmstar-z\]) for $F^{c(s)}_1$ and $\chi_{l=1,0} = mk^2_F/(2\pi)$ (recall that for spin and charge currents $f^{c(s)}_{l=1} (k_F) =1$), we obtain $$\label{f1}
\begin{aligned}
&\delta \chi^c_{l=1,M} = \chib,\qquad \delta \chi^c_{l=1,H} =0, \\
&\delta\chi^s_{l=1,M} = -3\chib, \quad\delta \chi^s_{l=1,H} =2\chib,
\end{aligned}$$ where $\chib = m^3U^2 k^2_F/16\pi^3$. In Table \[tab:delta-chi\] we list $\delta \chi^{c(s)}_{l=1,H}$, $\delta \chi^{c(s)}_{l=1,M}$ and $\delta\chi_{l=1,L}\cs$ in units of $\chib$.
![ The two AL vertex correction diagrams for three-leg vertex. []{data-label="fig:3vertex"}](triplevertex.pdf){width="8cm"}
![Diagrams for $\Gamma^{\omega}$ to second order, for a constant $U$. []{data-label="fig:4vertex"}](vertex.pdf){width="8cm"}
We also computed $\Lambda^{c(s)}_{l=1}$ independently, by collecting vertex correction diagrams, keeping external particles at the FS. Applying the same tactics as before, i.e., identifying equivalent contributions to reduce the number of diagrams, we find that &&\^[c]{}\_[l=1]{} = 1+ \_[7a]{} + \_[7b]{}\
&& \^[s]{}\_[l=1]{} = 1 - \_[7b]{} where $\Lambda_{7a}$ and $\Lambda_{7b}$ are two vertex corrections in Fig.\[fig:3vertex\]. In explicit form $$\label{eq:3vertex}
\begin{aligned}
\Lambda_{7a}&=-\frac{2U^2}{k_F} \int \frac{d^3k d^3p}{(2\pi)^6}(\p\cdot\hat{q})(G^0_{\tp})^2G^0_{\tp-\tk}G^0_{k_F\hat n-\tk}\\
\Lambda_{7b}&=-\frac{2U^2}{k_F}\int \frac{d^3k d^3p}{(2\pi)^6}(\p\cdot\hat{q})(G^0_{\tp})^2G^0_{\tp-\tk}G^0_{k_F\hat n+\tk}
\end{aligned}$$ where $ \hat n$ is a unit vector. We evaluated the integrals in the RHS of numerically and the results are presented in Table \[tab:delta-lambda\]. From (\[eq:deviations-cs\]) and (\[eq:mmstar-z\]) we expect $$\label{eq:delta-gamma-z-expected}
\Lambda^c_{l=1} - 1 =1.39 \overline{\Lambda},\quad \Lambda^s_{l=1} - 1 = -0.61\overline{\Lambda}$$ where $\overline{\Lambda} = \frac{m^2U^2}{8\pi^2}$. We see that these relations are satisfied, as they should be.
channel $\delta\chi_{l=1,L}$ $\delta\chi_{l=1,M}$ $\delta\chi_{l=1,H}$ $\delta\chi_{l=1}$
--------- ---------------------- ---------------------- ---------------------- --------------------------
charge $-1$ $+0.99 \pm 0.02$ $+0.01 \pm 0.04$ $\mathbf{0.01 \pm 0.04}$
spin $+1$ $-2.97 \pm 0.02$ $+1.98 \pm 0.02$ $\mathbf{0.01 \pm 0.02}$
channel from Eq. from $\delta\chi_{l=1,M}$
--------- --------------------- ---------------------------
charge $ +1.39 \pm 0.02$ $+1.38 \pm 0.01$
spin $-0.604 \pm 0.008 $ $-0.60 \pm 0.01$
Microscopic explanation for the absence of $l=1$ Pomerachuk instabilities {#sec:micr-expl-absence}
-------------------------------------------------------------------------
We now present microscopic arguments as to why $\Lambda Z$ and $\chi_{l=1, inc}$, for charge-current and spin-current susceptibilities are expressed via Landau parameters. We will analyze Eq. for the spin channel, where $\Lambda Z = 1 + F^s_1-F^c_1$.
The quasiparticle residue $Z$ can be expressed via $\Gamma^\omega_{\alpha \beta, \alpha \beta}$ using a Ward identity for any conserved “charge” [@AGD; @Lifshitz1980; @Peskin1995]. For our purpose it is best to use the Ward identity associated with conservation of total momentum (recall that we consider a Galilean invariant system). Substituting ${\lambda_\rho (k} = {\bf k}$ into (\[eq:xxxx2\]) we obtain $$\label{eq:Z-GI-FL}
\frac{1}{Z} = 1
- \frac{i}{2 k_F}\sum_{\alpha\beta}\int \frac{d^3q}{(2\pi)^3} \Gamma^\omega_{\alpha\beta,\alpha\beta}(k_F\hat p, \q)(G_q^2)^\omega \hat p \cdot \q$$ The renormalization of the spin-current vertex can be written as $$\begin{aligned}
\label{eq:delta-lambda-1s-FL}
&\Lambda_{l=1}^s \sigma^z_{\beta\beta} = \sigma^z_{\beta\beta} - \nn \\
&\qquad\qquad\frac{i}{k_F}\sum_{\alpha}
\int \frac{d^3q}{(2\pi)^3}
\Gamma^\omega_{\alpha\beta,\alpha\beta}(k_F\hat p, \q)(G_q^2)^\omega \hat p \cdot \q \sigma^z_{\alpha\alpha}\end{aligned}$$ where $\hat p \cdot \q$ is now simply the form-factor for the current. The vertex function $\Gamma^\omega$ to order $U^2$ is given by the diagrams in Fig. \[fig:4vertex\]. In explicit form
$$\begin{aligned}
\label{eq:gamma-w-u2}
\Gamma^\omega_{\alpha\beta,\gamma\delta}(p=(k_F\hat p, 0),q) &= \frac{1}{2}\delta_{\alpha\gamma}\delta_{\beta\delta}\left[U + i U^2 \int \frac{d^3k}{(2\pi)^3}\left(2 G_k G_{q-p + k} + G_k G_{q+p-k}\right)\right] \nn \\
&\qquad\qquad -\frac{1}{2}\sigma_{\alpha\gamma}\cdot\sigma_{\beta\delta}\left[U + i U^2 \int \frac{d^3k}{(2\pi)^3} G_k G_{q+p-k}\right].\end{aligned}$$
Summing up contributions from both $Z$ and $\Gamma$ we obtain, to order $U^2$, $$\begin{aligned}
\label{eq:delta-lambdaZ-u2}
&\Lambda^s_{l=1} Z = 1 - \nn \\
&\qquad\quad\frac{U^2}{k_F} \int \frac{d^3kd^3q}{(2\pi)^6}\left(G_k G_{q-p + k} + G_k G_{q+p-k}\right)\hat p \cdot \q (G_q^2)^\omega\end{aligned}$$ As written, the integral in the RHS of Eq. is not confined to the FS. However, the sum can be re-expressed as an integral over the FS. The reason for this is the identity [@Chubukov],
\[eq:ph-pp-ident\] (G\_k G\_[q-p + k]{} + G\_k G\_[q+p-k]{})(G\_[q+p\_]{} - G\_q)=0,
where $p_\epsilon = \epsilon(k_F\hat p,\Omega)$ and $\epsilon \to 0$. This identity can be proven by a simple relabeling of indices on the p-h bubble. Choosing $\Omega = 0$ and expanding to first order in $\epsilon$, we obtain $$\int \frac{d^3kd^3q}{(2\pi)^6}\left(G_k G_{q-p + k} + G_k G_{q+p-k}\right)(G_q^2)^k\hat p \cdot \q=0,
\label{eq:ph-pp-connect}$$ where $(G_q^2)^k= \mbox{lim}_{\k\to 0}G(\q+\k,\Omega)G(\q,\Omega)$. This $(G_q^2)^k$ has a regular piece, equal to $(G_q^2)^\omega$, and an extra piece which comes from the FS. Using the known relation [@AGD] $$(G_q^2)^k = (G_q^2)^\omega - \frac{2 \pi i Z^2}{v_F^*}
\delta(\omega)\delta(|\q|-k_F),
\label{ll7}$$ substituting into (\[eq:ph-pp-connect\]), and using Eq. (\[eq:gamma-w-u2\]) to extract the Landau parameters, we obtain
\[llla\] (G\_k G\_[q-p + k]{} + G\_k G\_[q+p-k]{})(G\_q\^2)\^p = ( F\^c () - F\^s () ) = F\^c\_[l=1]{} - F\^s\_[l=1]{}
Substituting into (\[eq:delta-lambdaZ-u2\]), we recover Eq. .
We emphasize that only the product $ \Lambda^s_{l=1} Z$ is expressed via the integral over the FS. Taken separately, $ \Lambda^s_{l=1}$ and $Z$ are determined by integrals which are not confined to the FS. We also note that the same Eq. (\[llla\]) allows one to express the effective mass, computed to order $U^2$ in a direct perturbation theory, as the integral over the FS in Eq. (\[eq:mstar-m\]) (see Ref. [@Chubukov] for details).
Arbitrary form-factor $\lambda^{c(s)}_{l=1} (k)$ and other values of $l$ {#sec_new}
=========================================================================
channel $\delta\chi_{l=1,L}$ $\delta\chi_{l=1,M}$ $\delta\chi_{l=1,H}$ $\delta\chi_{l=1}$
--------- ---------------------- ---------------------- ---------------------- --------------------
charge $-1$ $+0.36 \pm 0.04$ $+0.12 \pm 0.04$ $-0.52\pm 0.06$
spin $+1$ $-0.86 \pm 0.02$ $+1.50 \pm 0.04$ $-0.36 \pm 0.04$
channel from Eq. from $\delta\chi_{l=1,M}$
--------- -------------------- ---------------------------
charge $+1.07 \pm 0.01$ $+1.07 \pm 0.01$
spin $-0.548 \pm 0.008$ $-0.538 \pm 0.008$
The purpose of this final section is to clarify how generic are the constraints imposed by Eqs. , , which prevent a Pomeranchuk instability for charge-current and spin-current order parameters. In this section we first study the case of an order parameter $\rho\cs_{l=1}$ with form factor $\lambda_{l=1} (k) = \k f\cs_{l=1} (|k|)$ for which $f_{l=1}(|\k|) \neq 1$. We argue that in this case there is no relation $\Lambda\cs_{l}Z \propto (1 + F\cs_1)$ and therefore a Pomeranchuk instability does occur when $F_l\cs = -1$.
The argument is quite straightforward – $\rho\cs_{l=1}$ with $f_{l=1}(|\k|) \neq 1$ is not a current of a conserved quantity, hence it is not related by a continuity equation to a quantity, such as a conserved charge, whose susceptibility is expressed in terms of Landau parameters. Rather, it has two pieces and is of the form, $$\label{eq:chi-fl-def}
\chi_{l=1}\cs = \tilde\chi_{l=1}\cs + \delta\chi\cs_{l=1},$$ where $\tilde\chi_{l=1}\cs$ is finite and can be expressed in terms of Landau parameters at $\Omega/v_F |\q| \to 0$, but $\delta\chi_{l=1}\cs$ cannot. As a result,while $\tilde\chi_{l=1}\cs$ remains finite when $F_{l=1}\cs = -1$, $\delta\chi_{l=1}\cs$ diverges, signaling a Pomeranchuk instabilitiy.
An indication of this appears already at first order in $U$. To see this, we evaluate the diagrams of Fig. \[fig:chi-1st-order\] in Sec. \[sec:first-order-uq\] for the more general case $f_l(|\k|) \neq 1$. Then we find the contribution of diagrams \[fig:chi-1st-order\]a,b is:
\[eq:gen-chi-l1-1st-order\] (\_[l=1,2a]{}+\_[l=1,2b]{}) = (\_[l=0,2a]{}+\_[l=0,2b]{}) + \_[l=1]{},
Here, $\tilde\chi_{l=0}\cs$ is the susceptibility of a channel with $l=0$ symmetry, but with $f_{l=1}(|\k|)$ in the form-factor, and
$$\begin{aligned}
\label{eq:chi-fl-1st}
\delta\chi_{l=1}\cs &= 2\int\frac{d^2k d^2p}{(2\pi)^2}\frac{\left[n_F\left(\xi_{\p-\q/2}\right)-n_F\left(\xi_{\p+\q/2}\right)\right]\left[n_F\left(\xi_{\k-\q/2}\right)-n_F\left(\xi_{\k+\q/2}\right)\right]}{(m^{-1} q)^2} \nn \\
&\qquad U(\p-\k)\left[f_{l=1}(|\k|)f_{l=1}(|\p|) - f_{l=1}^2(|\k|)\right]\left(1 - \frac{2\Omega}{\Omega - m^{-1}\k\cdot\q}\right),\end{aligned}$$
is an additional term which is exactly zero for $f_l = 1$. The results to order $U$ are somewhat special because each of the three terms in Eq. has an additional $q^2$ factor in the $\Omega/q \to 0$ limit.
channel $\delta\chi_{l=2,L}$ $\delta\chi_{l=2,M}$ $\delta\chi_{l=2,H}$ $\delta\chi_{l=2}$
---------------------------------------- ---------------------- ---------------------- ---------------------- ----------------------
charge, $f_{l=2}=1$ $+\frac{1}{2}$ $-1.48 \pm 0.02$ $+6.40 \pm 0.04$ $\bf +4.92 \pm 0.04$
spin, $f_{l=2}=1$ $-\frac{1}{2}$ $-1.50 \pm 0.02$ $+1.32 \pm 0.04$ $\bf -0.18 \pm 0.04$
charge, $f_{l=2}=\frac{k_F^2}{|\k|^2}$ $+\frac{1}{2}$ $-1.46 \pm 0.02$ $+1.18 \pm 0.04$ $\bf -0.28 \pm 0.04$
spin, $f_{l=2}=\frac{k_F^2}{|\k|^2}$ $-\frac{1}{2}$ $-1.66 \pm 0.02$ $+0.70 \pm 0.04$ $\bf -0.96 \pm 0.04$
Nevertheless, the appearance of $ \delta\chi_{l=1}$ already at this order indicates that $\delta\chi\cs_{l=1}$ is not expressed via $\delta\chi\cs_{l=0}$, taken in the $q/\Omega \to 0$ limit, as it was the case for a current of a conserved order parameter.
To see explicitly that for $f_l \neq 1$ Eqs. and are no longer valid we perform the same calculations as in Sec. \[sec:micr-expl-absence\] for $f^{c,s}_{l=1} (|k|) \neq \text{constant}$. For definiteness, we consider $f\cs_{l=1}(|\k|) = k_F/|\k|$, i.e., $\lambda\cs_{l=1}(\k) = k_F\hat\k \cdot \hat \q$. The cancellation between different diagrams for susceptibility still holds, and the results for $\delta \chi^s_{l=1}$ and $\delta \chi^s_{l=1}$ to order $U^2$ are still given by Eqs (\[ll3\]) and (\[ll4\]), and the contribution from each diagram can again be split into “H”, “M”, and “L” parts. However, now each contribution has to be computed with different prefactors. This does not affect the “L” contribution as, by construction, $f\cs_{l=1} (k_F) =1$, but the modification of $f\cs_{l=1} (k)$ does affect “M” and “H” contributions. In Table \[tab:delta-chi\] we present the results for “H”, “M”, and “L” contributions to $\delta \chi^c_{l=1}$ and $\delta \chi^s_{l=1}$ in units of $\chib$. We also computed $\Lambda\cs_{l=1}$ by evaluating the renormalization of the three-leg vertex. We show the results in Table \[tab:delta-lambda-nl\], again in units of $\overline{\Lambda}$. We see that neither the constraints on the components of the susceptibilities, Eq. (\[f1\]), nor the conditions on $\Lambda\cs_{l=1}$, Eq. \[eq:delta-gamma-z-expected\], are obeyed. Therefore, $\Lambda\cs_{l=1}$ does not scale with $(1 + F\cs_1)$ and does not cancel $1/(1 + F\cs_1)$ in the quasiparticle part of the susceptibility. Since there is no cancellation of the diverging part, a Pomeranchuk instability does occur when $F_1\cs = -1$ for any order parameter with $f\cs_{l=1} \neq 1$.
We also explicitly calculated “L”, “M”, and “H” contributions to susceptibility in $l=2$ with $f\cs_{l=2} (k) =1$ and $f_{l=2}=\frac{k_F^2}{|\k|^2}$. For $l=2$, $F^c_2 = - F_2^s = \chib/2$, such that the low-energy contributions to the $l=2$ charge and spin susceptibilities are $\delta \chi^c_{l=2,L} = - \delta \chi^s_{l=2,L} = -\chib'/2$, where $\chib' = \chib k_F^2 = m^3U^2 k^4_F/16\pi^3$. We show the results in Table \[tab:l2-data\]. We didn’t find any relation between “M” and “H” contributions to both spin and charge susceptibilities and $1 + F\cs_{l=2}$. In particular, we checked the expressions for $l=2$ case presented in Ref. and did not reproduce them. This can be also seen by comparing the results in Ref. with our expressions for susceptibility to first order in momentum-dependent $U(q)$, Eq. (\[n\_12\]).
Summary {#sec:6}
=======
In this paper we studied the constraints placed by conservation laws on Pomeranchuk transitions, particularly the role of the continuity equation and longitudinal sum rule. This issue has been previously considered by Leggett [@Leggett1965] back in 1965, and was re-analyzed recently by Kiselev et al[@Kiselev2017]. The continuity equation and the sum rule reveal interesting properties of susceptibilities of currents of conserved total charge and spin. Namely, high energy features of a system, such as $\Lambda_{l=1}\cs Z$, and the incoherent piece of the susceptibility, $\chi\cs_{l=1,inc}$, can be expressed in terms of the Landau parameters $F_l\cs$, which describe the interaction between fermions on the FS. In particular, $\Lambda\cs_{l=1} Z$ scales as $(1+ F\cs_1)$ and vanishes at $F\cs_1 =-1$, when the quasiparticle contribution to susceptibility diverges as $1/(1+ F\cs_1)$. The vanishing of $\Lambda\cs_{l=1} Z$ cancels out the divergence, and, as a result, the system does not undergo a p-wave Pomeranchuk instability. Our aim was to verify this in diagrammatic perturbation theory, present a microscopic explanation why high-energy and low-energy contributions to susceptibility are related, and check how general such constraints are.
We showed that the constraints work only for $l=1$ and for the specific $l=1$ order parameter with form-factor $\lambda\cs_{l=1} (k) = {\bf k}$. Such an order parameter describes currents of the fermionic number and spin - both of which are conserved quantities. For any form factor with $l=1$ symmetry, but different functional behavior, $\lambda\cs_{l=1}(\k) = f\cs_{l=1}(|\k|) \k $ with $f(|\k|) \neq 1$, high-energy and low-energy contributions to the susceptibility are not correlated. The same is true for other values of $l$. As a result, the susceptibility for any other order parameter with either $l=1$ or other $l$ diverges when $F\cs_l =-1$, i.e., the Pomeranchuk instability does occur.
Acknowledgements
================
We thank J. Schmalian, P. Woelfle, and particularly D. Maslov for valuable discussions. The work was supported by NSF DMR-1523036. AVC is thankful to KITP at UCSB where part of the work was done. KITP is supported by NSF grant PHY-1125915.
Appendix: Details of the numerical evaluation {#sec:numerical_evaluation_details .unnumbered}
=============================================
Form factor $\frac{1}{2}\chi_{l,6a}^H$ $\chi_{l,6c}^H$ $\frac{1}{2}\chi_{l,6d}^H$ $\frac{1}{2}\chi_{l,6a}^M$ $\chi_{l,6c}^M$ $\frac{1}{2}\chi_{l,6d}^M$
-------------------------------------------------------- ---------------------------- ------------------ ---------------------------- ---------------------------- ----------------- ----------------------------
$\lambda_{p}=|{\mathbf}{p}|\cos\phi_{{\mathbf}{p}}$ $0.613\pm0.008$ $-0.226\pm0.017$ $-0.379\pm0.005$ $-0.879\pm0.008$ $0.767\pm0.006$ $ 0.608\pm0.005$
$\lambda_{p}=k_F\cos\phi_{{\mathbf}{p}}$ $0.500\pm0.013$ $-0.193\pm0.008$ $-0.250\pm0.008$ $-0.879\pm0.008$ $0.508\pm0.013$ $ 0.548\pm0.005$
$\lambda_{p}=|{\mathbf}{p}|^2\cos2\phi_{{\mathbf}{p}}$ $1.956\pm0.015$ $-0.052\pm0.008$ $1.292\pm0.014$ $-0.879\pm0.008$ $0.266\pm0.006$ $ -0.379\pm0.006$
$\lambda_{p}=k_F^2\cos2\phi_{{\mathbf}{p}}$ $0.500\pm0.013$ $-0.058\pm0.006$ $0.147\pm0.012$ $-0.879\pm0.008$ $0.202\pm0.008$ $ -0.303\pm0.005$
For our numerical evaluation of the high energy and middle energy contributions to the second order diagrams, Eqs. and we used Mathematica 11.1.1 with the built-in algorithm *NIntegrate*, using the Monte Carlo integration strategy. In our evaluation of diagrams we used polar coordinates and cut off the momentum at $15k_F$, e.g. $\{|\p|,0,15k_F\}$. The UV divergence in Eqs. is avoided by the symmetry factor $\cos l\phi$ and this $15k_F$ truncation is large enough to obtain our results accurately. Since only the angle differences of three the momenta($\p$, $\k$ and $\k'$) enter our integrals, one can integrate out one of these three angles by hand to achieve higher accuracy.
In evaluations of the high energy contributions($\chi_{l,6a}^H$, $\chi_{l,6c}^H$ and $\chi_{l,6d}^H$), the integral region $\{|\p|,0,15k_F\}\times\{|\k|,0,15k_F\}\times\{|\k'|,0,15k_F\}$ is divided into 8 parts: every dimension of momentum is divided into $(0, 3k_F)$ and $(3k_F, 15k_F)$, e.g. $\{|\p|,0,15k_F\}=\{|\p|,0,3k_F\}+\{|\p|,3k_F,15k_F\}$ . In evaluations of mixed energy contributions($\chi_{l,6a}^M$, $\chi_{l,6c}^M$ and $\chi_{l,6d}^M$), the integral region $\{|\k|,0,15k_F\}\times\{|\k'|,0,15k_F\}$ is divided into 9 parts instead. Each momentum dimension is divided as $(0, 3k_F)$, $(3k_F, 6k_F)$ and $(3k_F, 15k_F)$. Every subregion was sampled using a maximum of $10^9$ points. We evaluated each subregion 10 times to ensure the convergence of the numerical sums. The various $\chi_l\cs,\Lambda_l\cs Z$ we needed are readily found from the numerical expressions for the “H” and “M” diagrams as detailed in the text. The deviation of these 10 evaluations are the basis for computing the error brackets of Tables \[tab:delta-chi\]-\[tab:l2-data\].
As a check of the reliability of our numerical scheme we computed the quasiparticle residue $Z$, which is known to be $Z=1-1.39\frac{m^2U^2}{8\pi^2}$ (see text). Our calculation for $Z$ is based on Pitaevskii-Landau relations, Eq. of the text and $$\label{eq:PL}
\frac{1}{Z} = 1 - \frac{i}{2}\sum_{\alpha\beta}\int\frac{d^3q}{(2\pi)^3}\Gamma^{\omega}_{\alpha\beta,\beta\alpha}(k_F\hat p,\q)(G^2_{q})^\omega$$ Eq. and Eq. must give the same result. Numerically we found, $$\begin{aligned}
Z=1-(1.389\pm0.045)\frac{m^2U^2}{8\pi^2}\quad\text{based on Equation~\eqref{eq:PL}},\\
Z=1-(1.390\pm0.028)\frac{m^2U^2}{8\pi^2}\quad\text{based on Equation~\eqref{eq:Z-GI-FL}},\\
\end{aligned}$$ which gives us confidence our integrals are accurate.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Periodic-finite-type shifts (PFT’s) form a class of sofic shifts that strictly contains the class of shifts of finite type (SFT’s). In this paper, we investigate how the notion of “period” inherent in the definition of a PFT causes it to differ from an SFT, and how the period influences the properties of a PFT.'
author:
-
title: 'On the Period of a Periodic-Finite-Type Shift'
---
Introduction\[intro\]
=====================
Shifts of finite type (SFT’s) are objects of fundamental importance in symbolic dynamics and the theory of constrained coding [@ML]. A well-known example of an SFT would be the $(d,k)$ run-length limited ($(d,k)$-RLL) shift, where the number of 0’s between successive 1’s is at least $d$ and at most $k$. Constrained codes based on these $(d,k)$-RLL shifts are used in most storage media such as magnetic tapes, CD’s and DVD’s.
A generalization of SFT’s was introduced by Moision and Siegel [@MS2] who were interested in examining the properties of distance-enhancing constrained codes, in which the appearance of certain words is forbidden in a periodic manner. This new class of shifts, called periodic-finite-type shifts (PFT’s), contains the class of SFT’s and some other interesting classes of shifts, such as constrained systems with unconstrained positions [@BCF],[@SMNW], and shifts arising from the time-varying maximum transition run constraint [@PM]. The class of PFT’s is in turn properly contained within the class of sofic shifts [@MS], a fact we discuss in more detail in Section \[background\].
The properties of SFT’s are now quite well understood (cf. [@ML]), but the same cannot be said for PFT’s. The study of PFT’s has primarily focused on finding efficient algorithms for constructing their presentations [@BCF], [@MS], [@P1]. The difference between the definitions of SFT’s and PFT’s is quite small. An SFT is defined as a set of bi-infinite sequences (over some alphabet) that do not contain as subwords any word from a certain finite set. Thus, an SFT is defined by forbidding the appearance of finitely many words at any position of a bi-infinite sequence. A PFT is also defined by forbidding the appearance of finitely many words, except that these words are only forbidden to appear at positions of a bi-infinite sequence that are indexed by certain pre-defined periodic integer sequences; see Section \[background\] for a formal definition. This paper aims to initiate a study of how the “period” inherent in the definition of a PFT influences its properties.
After a review of relevant definitions and background in Section \[background\], we will see in Section \[period\_influence\] that given an SFT $\cY$, we can associate with it a PFT $\cX$ in such a way that it is only the period that differentiates $\cX$ from $\cY$. We then seek to understand how the period determines the properties of the PFT $\cX$ by means of a comparative study of $\cX$ and $\cY$. We investigate a different aspect of periods in Section \[periods\], where we study the influence of the period of a PFT $\cX$ on the periods of periodic sequences in $\cX$, and on the periods of graphical presentations of $\cX$.
Basic Background on SFT’s and PFT’s\[background\]
=================================================
We begin with a review of basic background, based on material from [@ML] and [@MS]. Let $\S$ be a finite set of symbols; we call $\S$ an *alphabet*. We always assume that $|\S|=q\geq 2$ since $q=1$ gives us a trivial case. Let $\w=\ldots w_{-1}w_{0}w_{1}\ldots$ be a bi-infinite sequence over $\S$. A word (finite-length sequence) $u\in \S^n$ (for some integer $n$) is said to be a *subword* of $\w$, denoted by $u\prec \w$, if $u=w_iw_{i+1}\ldots w_{i+n-1}$ for some integer $i$. If we want to emphasize the fact that $u$ is a subword of $\w$ starting at the index $i$, (*i.e.*, $u=w_iw_{i+1}\ldots w_{i+n-1}$), we write $u\prec_i \w$. By convention, we assume that the empty word $\epsilon \in \S^0$ is a subword of any bi-infinite sequence. Also, we define $\sigma$ to be the shift map, that is, $\sigma(\w)=\ldots w^*_{-1}w^*_{0}w^*_{1}\ldots$ is the bi-infinite sequence satisfying $w^{*}_{i}=w_{i+1}$ for all $i$.
Given a labeled directed graph $\cG$, where labels come from $\S$, let $S(\cG)$ be the set of bi-infinite sequences which are generated by reading off labels along bi-infinite paths in $\cG$. A *sofic shift* $\cS$ is a set of bi-infinite sequences such that $\cS=S(\cG)$ for some labeled directed graph $\cG$. In this case, we say that $\cS$ is *presented by* $\cG$, or that $\cG$ is a *presentation* of $\cS$. It is well known that every sofic shift has a *deterministic* presentation, *i.e.*, a presentation such that outgoing edges from the same state (vertex) are labeled distinctly. For a sofic shift $\cS$, $\cB_n(\cS)$ denotes the set of words $u\in \S^n$ satisfying $u\prec \w$ for some bi-infinite sequence $\w$ in $\cS$, and $\cB(\cS)=\cup_{n\geq 0} \cB_n(\cS)$. A sofic shift $\cS$ is *irreducible* if there is an irreducible (*i.e.*, strongly connected) presentation of $\cS$, or equivalently, for every ordered pair of words $u$ and $v$ in $\cB(\cS)$, there exists a word $z\in \cB(\cS)$ such that $uzv \in \cB(\cS)$.
A *shift of finite type* (SFT) $\cY_{\cF'}$, with a finite set of forbidden words (a forbidden set) $\cF'$, is the set of all bi-infinite sequences $\w=\cdots w_{-1}w_{0}w_{1}\cdots$ over $\S$ such that $\w$ contains no word $f' \in \cF'$ as a subword. That is, the finite number of words $f'$ in $\cF'$ are not in $\cB(\cY_{\cF'})$. A *periodic-finite-type shift*, which we abbreviate as *PFT*, is characterized by an ordered list of finite sets $\cF=(\cF^{(0)},\cF^{(1)}, \ldots, \cF^{(T-1)})$ and a *period* $T$. The PFT $\cX_{\{\cF,T\}}$ is defined as the set of all bi-infinite sequences $\w$ over $\S$ such that for some integer $r\in \{0,1,\ldots, T-1\}$, the $r$-shifted sequence $\sigma ^{r}(\w)$ of $\w$ satisfies $u\prec _i\sigma^{r}(\w)$ $\Longrightarrow $ $u\not \in \cF^{(i\ \mbox{mod}\ T)}$ for every integer $i$. For simplicity, we say that a word $f$ is in $\cF$ (symbolically, $f\in \cF$) if $f\in \cF^{(j)}$ for some $j$. Since the appearance of words $f\in \cF$ is forbidden in a periodic manner, note that $f$ can be in $\cB(\cX_{\{\cF,T\}})$. Also, observe that a PFT $\cX_{\{\cF,T\}}$ satisfying $\cF^{(0)}=\cF^{(1)}= \cdots =\cF^{(T-1)}$ is simply the SFT $\cY_{\cF'}$ with $\cF'=\cF^{(0)}$. Thus, SFT’s are special cases of PFT’s. We call a PFT *proper* when it cannot be represented as an SFT.
Any SFT can be considered to be an SFT in which every forbidden word has the same length. More precisely, given an SFT $\cY=\cY_{\cF^*}$, find the longest forbidden word in $\cF^*$ and say it has length $\ell$. Set $\cF'=\{f'\in \S^{\ell}: \mbox{$f'$ has some $f^*\in \cF^*$
as a prefix}\}$. Then, $\cY_{\cF^*}=\cY_{\cF'}$, and each word in $\cF'$ has the same length, $\ell$. Furthermore, we can also assume that $\cB_{\ell}(\cY)=\S^{\ell}\setminus \cF'$ since if not (that is, if $\cB_{\ell}(\cY)\subsetneq \S^{\ell}\setminus \cF'$), every word in $(\S^{\ell}\setminus \cF')\setminus \cB_{\ell}(\cY)$ can be added to $\cF'$, without affecting $\cY$ in any way.
Correspondingly, every PFT $\cX$ has a representation of the form $\cX_{\{\cF,T\}}$ such that $\cF^{(j)} = \emptyset$ for $1 \leq j \leq T-1$, and every word in $\cF^{(0)}$ has the same length. An arbitrary representation $\cX_{\{\cF,T\}}$ can be converted to one in the above form as follows. If $f\in \cF^{(j)}$ for some $1\leq j\leq T-1$, list out all words with length $j+|f|$ whose suffix is $f$, add them to $\cF^{(0)}$, and delete $f$ from $\cF^{(j)}$. Continue this process until $\cF^{(1)}=\cdots=\cF^{(T-1)}=\emptyset $. Then, apply the method described above for SFT’s to make every word in $\cF^{(0)}$ have the same length.
It is known that PFT’s belong to the class of sofic shifts.\
All periodic-finite-type shifts $\cX$ are sofic shifts. That is, for any PFT $\cX$, there is a presentation $\cG$ of $\cX$.
Moision and Siegel proved the theorem by giving an algorithm that, given a PFT $\cX$, generates a presentation, $\cG_\cX$, of $\cX$. We call the presentation $\cG_{\cX}$ the *MS presentation* of $\cX$. The *MS algorithm*, given a PFT $\cX$ as input, runs as follows.
1. Represent $\cX$ in the form $\cX_{\{\cF,T\}}$, such that every word in $\cF$ has the same length $\ell$ and belongs to $\cF^{(0)}$.
2. Prepare $T$ copies of $\S^{\ell}$ and name them $\cV^{(0)}, \cV^{(1)}, \ldots, \cV^{(T-1)}$.
3. Consider the words in $\cV^{(0)}, \cV^{(1)}, \ldots, \cV^{(T-1)}$ as states. Draw an edge labeled $a\in \S$ from $u=u_1u_2\cdots u_{\ell}\in \cV^{(j)}$ to $v=v_1v_2\cdots v_{\ell}\in \cV^{(j+1 \mod{T})}$ if and only if $u_2\cdots u_{\ell}=v_1\cdots v_{\ell-1}$ and $v_{\ell}=a$.
4. Remove states corresponding to words in $\cF^{(0)}$ from $\cV^{(0)}$, together with their incoming and outgoing edges. Call this labeled directed graph $\cG'$.
5. If there is a state in $\cG'$ having only incoming edges or only outgoing edges, remove the state from $\cG'$ as well as its incoming or outgoing edges. Continue this process until we cannot find such a state. The resulting graph $\cG_{\cX}$ is a presentation of $\cX$.
It is evident that the MS presentation of a PFT is always deterministic. Also, for a path $\alpha $ in $\cG_{\cX}$ with length $|\alpha|\geq \ell$, $\alpha$ terminates at some state that is a copy of $u=u_1u_2\ldots u_{\ell}$ iff the length-$\ell$ suffix of the word generated by $\alpha$ is equal to $u$. \[terminal\_rem\]
Influence of the Period $T$ on a PFT\[period\_influence\]
=========================================================
From this point on, whenever we consider an SFT $\cY_{\cF'}$ in this paper, we will implicitly assume that every forbidden word in $\cF'$ has the same length $\ell$, and that $\cB_{\ell}(\cY)=\S^{\ell}\setminus \cF'$. As we observed in the previous section, there is no loss of generality in doing so. Given an SFT $\cY_{\cF'}$, consider the PFT $\cX =\cX_{\{\cF,T\}}$ in which $$\cF=(\cF^{(0)}, \cF^{(1)}, \ldots, \cF^{(T-1)})=
(\cF',\emptyset ,\ldots, \emptyset ).$$ While $\cY_{\cF'} \subseteq \cX_{\{\cF,T\}}$, equality does not hold in general. Note that it is only the influence of the period $T$ that causes the shifts $\cX = \cX_{\{\cF,T\}}$ and $\cY = \cY_{\cF'}$ to differ. So, a comparative study of $\cX$ and $\cY$ is a useful means of understanding how the period $T$ determines the properties of the PFT $\cX$. In this section, we present a sampling of results that illustrate how properties of the SFT $\cY$ can affect those of the PFT $\cX$.
The following result, which shows that the irreducibility of $\cY$ has a significant effect on the irreducibility of $\cX$, may be considered typical of the comparative study proposed above.\
Suppose that $\cY=\cY_{\cF'}$ is an irreducible SFT. Let $\cX=\cX_{\{\cF, T\}}$ be the PFT satisfying $$\cF=(\cF^{(0)}, \cF^{(1)}, \ldots, \cF^{(T-1)})=
(\cF',\emptyset , \ldots,\emptyset ).$$ If there exists a periodic bi-infinite sequence $\y$ in $\cY$ with a period $p$ satisfying $p \equiv 1 \pmod{T}$, then the MS presentation, $\cG_{\cX}$, of $\cX$ is irreducible as a graph. That is, $\cX$ is irreducible. \[irred.thm\]
\
*Proof*: Throughout this proof, for a path $\eta$ in a graph, let $s(\eta)$ and $t(\eta)$ be the starting state and the terminal state, respectively, of $\eta$ in the graph. Also, for a state $v=v_1v_2\ldots v_{\ell}$ in $\cG_{\cX}$, $v\in \cV^{(j)}$ is denoted by $v^{(j)}$ for $0\leq j\leq T-1$.
Let $\cG'$ be the graph defined in Step 4 of the MS algorithm. Consider the subgraph $\cH$ of $\cG'$ that is induced by the states in $\S^{\ell}\setminus \cF'$. Since $\S^{\ell}\setminus \cF'=\cB_{\ell}(\cY)$, all states in $\cH$ have incoming edges and outgoing edges. Hence, $\cH$ is a subgraph of $\cG_{\cX}$.
Key points of the proof are the following.\
*Claim 1*: $\cH$ is a presentation of $\cY$.\
*Claim 2*: $\cH$ is irreducible as a graph if there exists a periodic bi-infinite sequence $\y$ in $\cY$ with a period $p$ satisfying $p \equiv 1 \pmod{T}$.
Once these claims are proved, it is straightforward to check that the MS presentation $\cG_{\cX}$ of $\cX$ is irreducible. Note that the graph $\cG'$ is obtained from $\cH$ by adding words in $\cF^{(0)}$ to $\cV^{(1)}, \cV^{(2)}, \ldots, \cV^{(T-1)}$ and corresponding incoming and outgoing edges. Observe that (by Step 5 of the MS algorithm) a word $f' \in \cF^{(0)}$ is a state in $\cG_{\cX}$ if and only if there exist paths $\rho _1$, $\rho _2$ in $\cG'$ satisfying $s(\rho _1)=f'$, $t(\rho _1)\in \S^{\ell} \setminus \cF'$ and $s(\rho _2)\in \S^{\ell}\setminus \cF'$, $t(\rho _2)=f'$. Since $\cH$ is irreducible, $\cG_{\cX}$ is irreducible as well.\
: We need to show that $S(\cH)\subseteq \cY$ and $\cY \subseteq S(\cH)$. It is clear that $S(\cH)\subseteq \cY$ since, by Remark \[terminal\_rem\], there is no path in $\cH$ which generates words in $\cF'$.
Conversely, take an arbitrary bi-infinite sequence $\x=\ldots x_{-1}x_0x_1\ldots \in \cY$. Since $f'\not\prec \x$ for every forbidden word $f'\in \cF'$, we see that for any integer $i$, the states corresponding to $x_{i-\ell+1}x_{i-\ell+2}\ldots x_{i}$ are in $\cH$. Therefore, there exists an edge labeled $x_{i+1}$ from $x_{i-\ell+1}x_{i-\ell+2}\ldots x_{i}\in\cV^{(j)}$ to $x_{i-\ell+2}\ldots x_{i}x_{i+1}\in \cV^{(j+1 \mod T)}$ for all integers $i$ and $0\leq j\leq T-1$. Hence, $\x\in S(\cH)$, that is, $\cY \subseteq S(\cH)$.\
: A periodic bi-infinite sequence $\y\in \cY$ with period $p \equiv 1 \pmod{T}$ can be written as $\y=(y_1y_2\ldots y_n)^{\infty}$, for some $y_1y_2\ldots y_n\in \S^n$, where $n$ is some multiple of $p$ satisfying $n \equiv 1 \pmod{T}$ and $n\geq \ell$.
As $\y\in \cY$, $y_{n-\ell+1}\ldots y_ny_1y_2\ldots y_n \in \cB(\cY)$. Thus, for every $i \in \{0,1,\ldots, T-1\}$, there exists a path $\alpha$ in $\cH$ satisfying $s(\alpha)=z^{(i)}=y_{n-\ell+1}\ldots y_n$ and generating $y_1y_2\ldots y_n$. Observe that $t(\alpha)$ is also $z^{(i')}=y_{n-\ell+1}\ldots y_n$ for some $i' \in \{0,1,\ldots, T-1\}$. However, since $|y_1y_2\ldots y_n| = n\equiv 1 \pmod{T}$, we have $i' = i+1 \mod{T}$. This automatically implies that for the word $z=y_{n-\ell+1}\ldots y_n$ in $\cB(\cY)$, there is a path $\beta_{jk}$ in $\cH$ such that $s(\beta_{jk})=z^{(j)}$ and $t(\beta_{jk})=z^{(k)}$ for any ordered pair $(j,k)$, where $0\leq j,k\leq T-1$.
Now take an arbitrary pair of states $u^{(r)}$ and $v^{(s)}$ in $\cH$. Since $\cY$ is irreducible, there exist words $w'$ and $w^{*}$ in $\cB(\cY)$ so that $uw'z$ and $zw^{*}v$ are in $\cB(\cY)$. Thus, there exists a path $\gamma $ generating $w'z$ such that $s(\gamma )=u^{(r)}$ and $t(\gamma )=z^{(j)}$ for some $0\leq j\leq T-1$, and a path $\delta $ generating $w^{*}v$ such that $s(\delta )=z^{(k)}$ for some $0\leq k\leq T-1$ and $t(\delta )=v^{(s)}$. As there is a path $\beta_{jk}$ from $z^{(j)}$ to $z^{(k)}$ from the argument above, we have a path $\gamma\beta_{jk}\delta$ starting from $u^{(r)}$ and terminating at $v^{(s)}$. Hence, the presentation $\cH$ is irreducible as a graph.\
From Theorem \[irred.thm\], we can obtain the following corollary.\
Let $\cY=\cY_{\cF'}$ be an irreducible SFT such that $|\cF'|<|\S|$. Then for all $T\geq 1$, the PFT $\cX=\cX_{\{\cF, T\}}$ with $$\cF=(\cF^{(0)}, \cF^{(1)}, \ldots, \cF^{(T-1)})=
(\cF',\emptyset , \ldots,\emptyset )$$ is irreducible. \[uni-for.cor\]
\
*Proof*: Since $|\cF'|<|\S|$, there is a symbol $a\in \S$ which is not used as the first symbol of any word in $\cF'$. Hence, the bi-infinite sequence $\a=a^{\infty}$ is in $\cY$. As $\a$ has period $1$, we have from Theorem \[irred.thm\] that $\cX$ is irreducible.\
The proof of Theorem \[irred.thm\] shows that the SFT $\cY=\cY_{\cF'}$ has a presentation $\cH$ that is a subgraph of the MS presentation $\cG_{\cX}$ of $\cX=\cX_{\{\cF,T\}}$, where $\cF=(\cF',\emptyset , \ldots,\emptyset )$. This fact may allow us to compare some of the invariants associated with the two shifts $\cY$ and $\cX$, for example, their entropies and their zeta functions (see [@ML Chapters 4 and 6]). The entropy (or the Shannon capacity) $h(\cS)$ of a sofic shift $\cS$ can be computed from a deterministic presentation $\cG$ of $\cS$ as follows: $h(\cS)=\log_2\lambda$, where $\lambda$ is the largest eigenvalue of the adjacency matrix $A_{\cG}$ of $\cG$. Equivalently, $\lambda$ is the largest root of the characteristic polynomial $\chi_{A_{\cG}}(t)=\det(tI-A_{\cG})$ of $A_{\cG}$ (see, *e.g.*, [@ML Chapter 4]).
Returning to the shifts $\cX$ and $\cY$ as above, since $\cH$ is a subgraph of $\cG_{\cX}$, it may be possible to express the characteristic polynomial of $A_{\cG_{\cX}}$ in terms of the characteristic polynomial of $A_{\cH}$. This would allow us to compare the entropies of $\cX$ and $\cY$. However, this seems to be hard to do in general. We have a partial result in the special case when $\cY=\cY_{\cF'}$ with $|\cF'|=1$, and $\cX=\cX_{\{\cF,2\}}$, as we describe next.
Recall that $|\S|=q$. Now suppose that $\cY=\cY_{\cF'}$ is an SFT with the set $\cF'$ consisting of a single forbidden word $f'$, and $\cX=\cX_{\{\cF, 2\}}$ is the PFT with period 2 and $\cF=(\cF^{(0)}, \cF^{(1)})=(\{f'\},\emptyset)$. Also, let $A_{\cG_{\cX}}$ be the adjacency matrix of the MS presentation $\cG_{\cX}$ of $\cX$, and let $A_{\cH}$ be that of the subgraph $\cH$ of $\cG_{\cX}$ induced by the states in $\S^{\ell}\setminus \{f'\}$. Observe that the matrix $A_{\cG_{\cX}}$ is a $(2q^{\ell}-1)\times (2q^{\ell}-1)$ 0-1 matrix. Without loss of generality, for $A_{\cG_{\cX}}$, we can assume the following.
- The first $q^{\ell}-1$ rows and columns correspond to states in $\cV^{(0)}$, and the last $q^{\ell}$ rows and columns correspond to those in $\cV^{(1)}$.
- Assign $f'\in \cV^{(1)}$ to the $(2q^{\ell}-1)$-th row and column, and arrange the first row so that the $(1,2q^{\ell}-1)$-th entry of $A_{\cG_{\cX}}$ is 1.
- Let $u\in \cV^{(1)}$ be such that the longest proper suffix of $u$ is equal to that of $f'$. Assign this $u$ to the $q^{\ell}$-th row and column so that the $q^{\ell}$-th row and the $(2q^{\ell}-1)$-th row are the same.
For a matrix $M$, set $M^{(i,j)}$ to be the submatrix of $M$ obtained by deleting its $i$-th row and $j$-th column. Then, observe that $A_{\cG_{\cX}}^{(2q^{\ell}-1,2q^{\ell}-1)}=A_{\cH}$. In this case, by applying elementary row operations to the matrix $N=tI-A_{\cG_{\cX}}$, we have $$\chi_{A_{\cG_{\cX}}}(t)=\det(N)=
\begin{vmatrix}
B & \c \\
\d & t
\end{vmatrix},$$ where $B$ is a $(2q^{\ell}-2)\times (2q^{\ell}-2)$ matrix satisfying $\det(B)=\chi_{A_{\cH}}(t)$, $\c$ is the $(2q^\ell-2) \times 1$ column vector $[-1\ 0\ \ldots\ 0]^T$, and $\d \in \{-1,0\}^{2q^\ell-2}$. Using the form given in $(1)$ for $\det(N)$, we can derive the following theorem. The complete proof will be published in the full version of this paper.
Let $\cY=\cY_{\cF'}$ and $\cX=\cX_{\{\cF,2\}}$ be the SFT and PFT described above, respectively. Then, the characteristic polynomial $\chi _{A_{\cG_{\cX}}}(t)$ of the adjacency matrix $A_{\cG_{\cX}}$ is given by $$\chi _{A_{\cG_{\cX}}}(t) =
t (\chi_{A_{\cH}}(t)+(-1)^{q^{\ell}}\det(B^{(1,q^{\ell})})).$$ \[char\_poly\_thm\]
\
Periods in PFT’s\[periods\]
===========================
The period $T$ involved in the description of a PFT is not the only notion of “period” that can be associated with the shift. For any shift $\cX$, we can always define its *sequential period*, $T_{seq}^{(\cX)}$, to be the smallest period of any periodic bi-infinite sequence in $\cX$. Furthermore, if $\cX$ is an irreducible sofic shift, we can define a “graphical period” for it as follows. Let $\cG$ be a presentation of $\cX$ with state set $\cV(\cG)=\{V_1, \ldots, V_r\}$. For each $V_i \in \cV(\cG)$, define $\per(V_i)$ to be the greatest common divisor (gcd) of the lengths of paths (cycles) in $\cG$ that begin and end at $V_i$, and further define $\per(\cG) = \gcd(\per(V_1),\ldots,\per(V_r))$. It is well known that when $\cG$ is irreducible, $\per(V_i) = \per(V_j)$ for each pair of states $V_i,V_j \in \cV(\cG)$, and hence $\per(\cG) = \per(V)$ for any $V \in \cV(\cG)$. The *graphical period*, $T_{graph}^{(\cX)}$, of an irreducible sofic shift $\cX$ is defined to be the least $\per(\cG)$ of any irreducible presentation $\cG$ of $\cX$.
Given a PFT $\cX$, define its *descriptive period*, $T_{desc}^{(\cX)}$, to be the smallest integer among all $T^*$ such that $\cX=\cX_{\{\cF^*,T^*\}}$ for some $\cF^*$. In this section, we determine what influence, if any, the descriptive period of a PFT has on its sequential and graphical periods.
Let $\cX=\cX_{\{\cF,T\}}$ be an irreducible PFT, and let $\cG$ be an irreducible presentation of $\cX$. Proposition 1 of [@MS] says that if $\cX$ is proper, then $\gcd(\per(\cG),T)\not=1$. Using that proposition, we can obtain the following result, which shows that a proper PFT $\cX$ can have $T_{desc}^{(\cX)}$ arbitrarily larger than $T_{seq}^{(\cX)}$.\
Suppose that $\cY=\cY_{\cF'}$ is an irreducible SFT, such that the bi-infinite sequence $a^{\infty}\in \cY$ for some $a \in \S$. Let $\cX=\cX_{\{\cF, T\}}$ be the PFT satisfying $$\cF=(\cF^{(0)}, \cF^{(1)}, \ldots, \cF^{(T-1)})=
(\cF',\emptyset , \ldots,\emptyset ).$$ Then, $a^{\infty} \in \cX$, so $T_{seq}^{(\cX)}=1$. Furthermore, if $\cX$ is a proper PFT and $T$ is prime, we have $T_{desc}^{(\cX)}=T$. \[desc\_pro\]
\
*Proof*: Since $\cY \subseteq \cX$, it is clear that $a^{\infty} \in \cX$, and hence, $T_{seq}^{(\cX)}=1$. Now, let $\cX = \cX_{\{\cF, T\}}$ be a proper PFT with $T$ prime. First observe that the MS presentation $\cG_{\cX}$ of $\cX$ is irreducible since the bi-infinite sequence $\a=a^{\infty }$ is in $\cY$ and $\a$ has period 1. Also, note that $\per(\cG_{\cX})$ must be $kT$ for some $k\geq 1$ from the construction of $\cG_{\cX}$. However, if we consider the period of the states $a^{\ell}$ in $\cG_{\cX}$, it is $T$. Thus, $\per(\cG_{\cX})=T$ by the irreducibility of $\cG_{\cX}$. Since $\cX$ is proper, we have from Proposition 1 of [@MS] that $\gcd(\per(\cG_{\cX}), T^*)\not=1$ for all $T^*$ satisfying $\cX=\cX_{\{\cF^*, T^* \}}$. As $T$ is prime, $\gcd(\per(\cG_{\cX}), T')=\gcd(T, T')=1$ for all $T'<T$. Therefore, $T$ is the descriptive period of $\cX$.\
For example, consider an SFT $\cY=\cY_{\cF'}$ with a forbidden set $\cF'=\{b^2\}$ for some $b\in \S$. Then, $\cY$ is irreducible, and $a^{\infty}\in \cY$ for any $a\in \S\setminus \{b\}$. In this case, for a PFT $\cX=\cX_{\{\cF,T\}}$ with $T$ prime, such that $\cF=(\{b^2\},\emptyset , \ldots,\emptyset )$, it may be verified that $\cX$ is proper, and hence, $T=T_{desc}^{(\cX)}$.
Conversely, $T_{seq}^{(\cX)}$ can be arbitrarily larger than $T_{desc}^{(\cX)}$ for proper PFT’s $\cX$. We present such an example next.
Set $\S=\{0,1\}$. We define a sliding-block map $\psi$ as follows: for a non-empty word $u=u_1u_2\ldots u_n\in \S^n$, (resp. a bi-infinite sequence $\w=\ldots w_{-1}w_0w_1\ldots$ over $\S$), define $\psi(u)=u^{*}_1u^{*}_2 \ldots u^{*}_{n-1}$, where $u^{*}_i=u_i+u_{i+1}\! \pmod{2}$ for $1\leq i \leq n-1$ (resp. $\psi(\w)=\ldots w^*_{-1}w^*_0w^*_1\ldots$, where $w^{*}_i=w_i+w_{i+1}\! \pmod{2}$ for each $i$). By convention, $\psi(u)=\epsilon $ when $u\in \S^1$. For $k\geq 1$, consider the PFT $\cX_k=\cX_{\{\cF_{k}, 2\}}$ with $\cF_k=(\cF_{k}^{(0)},\cF_{k}^{(1)})$, defined as follows.
- $\cF_k^{(1)}=\emptyset $ for all $k\geq 1$.
- $\cF_1^{(0)}=\{0\}$, and for $k\geq 2$, we set $\cF_{k}^{(0)}=\psi^{-1}(\cF_{k-1}^{(0)})$. That is, $\cF_{k}^{(0)}$ is the inverse image of $\cF_{k-1}^{(0)}$ under $\psi$.
It is easy to see that for each $k\geq 1$, every word $f \in \cF_k^{(0)}$ has length $|f|=k$, and in particular, we have $0^k \in \cF_k^{(0)}$. Moreover, as $\psi$ is a two-to-one mapping, we have $|\cF_k^{(0)}|=2^{k-1}$. The following proposition contains another useful observation concerning $\psi$. We omit the straightforward proof by induction.
For a binary word $u = u_1 u_2 \ldots u_r$ of length $r > m$, let $u_1^*u_2^*\ldots u_{r-m}^* = \psi^m(u)$. If $m = 2^j$ for some $j \geq 0$, then $u^*_i=u_{i}+u_{i+2^j} \! \pmod{2}$ for $1\leq i\leq r-m$. Furthermore, if $m = 2^j - 1$ for some $j \geq 0$, then $u^*_i=u_{i}+u_{i+1}+\cdots +u_{i+2^j-1} \! \pmod{2}$ for $1\leq i\leq r-m$. \[forbid\_pro\]
The corollary below simply follows from the fact that for any $f \in \cF_k^{(0)}$, we must have $\psi^{k-1}(f) = 0$.\
If $z \in \S^{2^j}$ (for some $j \geq 0$) has an odd number of 1’s, then $z \notin \cF_{2^j}^{(0)}$. \[odd1\_cor\]
We next record some important facts about the PFT’s $\cX_k$.\
For $k \geq 1$, the following statements hold: (a) $\cX_{k+1}=\psi^{-1}(\cX_{k})$; (b) $\cX_k$ is irreducible iff $1\leq k\leq 6$; and (c) $\cX_k$ is a proper PFT. \[Xk\_prop\]
\
*Proof*: Statement (a) follows straightforwardly from the definition of the PFT’s $\cX_k$.
For (b), first note that $\cX_k$ is irreducible for $1\leq k\leq 6$ since its MS presentation may be verified to be irreducible as a graph. When $k=7$, it can be shown that $\cX_k$ is not irreducible, which implies that $\cX_k$ is not irreducible when $k\geq 7$ by (a).
To prove (c), suppose to the contrary that $\cX_k$ is not a proper PFT for some $k \geq 1$. Then, $\cX_k=\cY$ for some SFT $\cY = \cY_{\cF'}$, where every forbidden word in $\cF'$ has the same length, $\ell$. Pick a $j \geq 0$ such that $2^j \geq k$, and set $r=2^j-k$. By (a) above, $\cX_{2^j} = \psi^{-r}(\cX_k) = \psi^{-r}(\cY)$. Note that $\psi^{-r}(\cY)$ is also an SFT, with forbidden set $\psi^{-r}(\cF')$. All words in $\psi^{-r}(\cF')$ have length $\ell' = \ell + r$.
For the PFT $\cX_{2^j}$, observe that the bi-infinite sequence $\w=(0^{2^j-1}1)^{\infty}0^{2^j}(10^{2^j-1})^{\infty}$ is in $\cX_{2^j}$ as $\w$ contains a word in $\cF^{(0)}_{2^j}$ (*i.e.*, $0^{2^j}$) only once, by Corollary \[odd1\_cor\]. Therefore, every subword of $\w$ is in $\cB(\cX_{2^j}) = \cB(\psi^{-r}(\cY))$.
Now, consider the bi-infinite sequence $$\w'=(0^{2^j-1}1)^{\infty}0^{2^j} (10^{2^j-1})^{2\ell'+1}1
0^{2^j} (10^{2^j-1})^{\infty}.$$ Note that every length-$\ell'$ subword of $\w'$ is also a subword of $\w$, and hence, is in $\cB(\psi^{-r}(\cY))$. This implies that $\w' \in \psi^{-r}(\cY)$. For the two distinct indices $m,n$ $(m<n)$ such that $0^{2^j} \prec_m \w'$ and $0^{2^j} \prec_n \w'$, we have $n-m= 2^j(2\ell'+2)+1$, so that $m \not\equiv n\! \pmod{2}$. But, since $0^{2^j} \in \cF_{2^j}^{(0)}$, this implies that $\w' \not \in \cX_{2^j}$, which is a contradiction.\
Statement (c) of Proposition \[Xk\_prop\] implies that $T_{desc}^{(\cX_k)} = 2$ for all $k \geq 1$. In contrast, the following theorem shows that $T_{seq}^{(\cX_k)}$ grows arbitrarily large as $k \rightarrow \infty$.\
For any $j\geq 0$ and $2^j+1\leq k\leq 2^{j+1}$, the periods of periodic sequences in $\cX_k$ must be multiples of $2^{j+1}$. \[mult.thm\]
To prove Theorem \[mult.thm\], we need the next three lemmas. We omit the simple proof of the first lemma.\
If $\x \in \{0,1\}^\Z$ is a periodic sequence, then so is $\psi(\x)$. Furthermore, any period of $\x$ is also a period of $\psi(\x)$. \[periodic\_lem\]
For any $j\geq 0$, $\cF^{(0)}_{2^j+1}=\{f^*f^*_1 :
f^*=f^*_1f^*_2\ldots f^*_{2^j}\in \S^{2^j}\}$. \[forbidden\_lem\]
\
*Proof*: Recall that for a word $f\in \cF^{(0)}_{2^j+1}$, $\psi^{2^j}(f)=0$. Since Proposition \[forbid\_pro\] shows that $\psi^{2^j}(f)=f_{1}+f_{2^j+1} \! \pmod{2}$, we have $f_1=f_{2^j+1}$. Noting that $|\cF^{(0)}_{2^j+1}|=2^{2^j}=|\S^{2^j}|$, we thus have $\cF^{(0)}_{2^j+1}=\{f^*f^*_1 :
f^*=f^*_1f^*_2\ldots f^*_{2^j}\in \S^{2^j}\}$.
For $j\geq 0$, there is no periodic sequence $\x$ in $\cX_{2^j+1}$ whose period is $(2t+1)2^{j}$ for some $t\geq 0$. \[small\_per\_lem\]
\
*Proof*: We deal with $j=0$ first. Note that $\cF_2^{(0)} = \{00,11\}$. So, if $\cX_2$ has a periodic bi-infinite sequence $\w = (w_1 w_2 \ldots w_m)^\infty$ with an odd period $m$, then $00 \not\prec w_1w_2\ldots w_m$, $11 \not\prec w_1w_2\ldots w_m$, and $w_1 \neq w_m$. But there is no word $w_1w_2\ldots w_m \in \S^m$ that satisfies these conditions.
Now, consider $j \geq 1$. Assume, to the contrary, that there exists a periodic sequence $\x=\ldots x_{-1}x_0x_1\ldots \in \cX_{2^j+1}$ whose period is $(2t+1)2^{j}$ for some $t\geq 0$. Then, $\x$ is of the form $(x_0x_1\ldots x_{(2t+1)2^j-1})^{\infty}$. Without loss of generality, we may assume that for every even integer $i$, $u\prec _i \x$ implies $u \not \in \cF^{(0)}_{2^j+1}$. Then, for each integer $m$, $x_{m2^j}x_{m2^j +1} \ldots x_{(m+1)2^j} \notin
\cF^{(0)}_{2^j+1}$. So, by Lemma \[forbidden\_lem\], we have $x_{m2^j}\neq x_{(m+1)2^j}$. This implies that $x_0=x_{(2t)2^{j}}$ as $|\S|=2$. But then, $x_{(2t)2^j} \ldots x_{(2t+1)2^j-1}x_0 \in \cF^{(0)}_{2^j+1}$, which is a contradiction.\
We are now in a position to prove Theorem \[mult.thm\].\
*Proof of Theorem \[mult.thm\]*: To prove the theorem, it is enough to show that for $j \geq 0$, the periods of periodic sequences in $\cX_{2^j+1}$ must be multiples of $2^{j+1}$. It then follows, by Lemma \[periodic\_lem\], that the same also applies to periodic sequences in $\cX_k$, for $2^{j}+1 < k \leq 2^{j+1}$.
When $j=0$, the required statement clearly holds by Lemma \[small\_per\_lem\]. So, suppose that the statement is true for some $j \geq 0$, so that periodic sequences in $\cX_{2^{j+1}}$ have only multiples of $2^{j+1}$ as periods. Therefore, by Lemma \[periodic\_lem\], periodic sequences in $\cX_{2^{j+1}+1}$ also can only have multiples of $2^{j+1}$ as periods. However, by Lemma \[small\_per\_lem\], no periodic sequence in $\cX_{2^{j+1}+1}$ can have an odd multiple of $2^{j+1}$ as a period. Hence, all periodic sequences in $\cX_{2^{j+1}+1}$ have periods that are multiples of $2^{j+2}$. The theorem follows by induction.\
Theorem \[mult.thm\] shows that for $2^j+1 \leq k \leq 2^{j+1}$, we have $T_{seq}^{(\cX_k)} \geq 2^{j+1}$. In fact, this holds with equality.\
$T_{seq}^{(\cX_1)} = 1$, and for $k \geq 2$, if $j \geq 0$ is such that $2^j+1 \leq k \leq 2^{j+1}$, then $T_{seq}^{(\cX_k)} = 2^{j+1}$. \[Tseq\_cor\]
\
*Proof*: When $k=1$, $T_{seq}^{(\cX_1)}=1$ as $1^{\infty}\in \cX_1$. So let $k \geq 2$, and let $j \geq 0$ be such that $2^j+1\leq k\leq 2^{j+1}$. We only need to show that $T_{seq}^{(\cX_k)} \leq 2^{j+1}$. The bi-infinite sequence $\w=(0^{2^{j+1}-1}1)^{\infty}$ is in $\cX_{2^{j+1}}$ since, by Corollary \[odd1\_cor\], $\w$ contains no word in $\cF_{2^{j+1}}^{(0)}$ as a subword. Since $\w$ has period $2^{j+1}$, by Lemma \[periodic\_lem\], $\w'= \psi^{2^{j+1}-k}(\w) \in \cX_k$ has period $2^{j+1}$ as well. Thus, $T_{seq}^{(\cX_k)}\leq 2^{j+1}$. Theorem \[mult.thm\] also implies the following corollary.\
$T_{graph}^{(\cX_k)} \geq T_{seq}^{(\cX_k)}$ holds when $1\leq k\leq 6$. \[gra\_seq\_cor\]
\
*Proof*: Since $\cX_1$ is proper, $T_{graph}^{(\cX_1)}\geq 2$ by Proposition 1 in [@MS]. Thus, $T_{graph}^{(\cX_1)} > T_{seq}^{(\cX_1)} = 1$.
So, let $k\geq 2$ and suppose $2^j+1\leq k\leq 2^{j+1}$ for some $j\geq 0$. By Corollary \[Tseq\_cor\], we have $T_{seq}^{(\cX_k)} = 2^{j+1}$. On the other hand, for any irreducible presentation $\cG$ of $\cX_k$, we have $\per(\cG)\geq 2^{j+1}$. Indeed, for each vertex $V$ in $\cG$, we have $\per(V)$ being a multiple of $2^{j+1}$; otherwise we would have a contradiction of Theorem \[mult.thm\]. Hence, $T_{graph}^{(\cX_k)}\geq 2^{j+1}=T_{seq}^{(\cX_k)}$ as required.\
Corollary \[Tseq\_cor\] shows that $T_{seq}^{(\cX_k)}$ grows arbitrarily large as $k \rightarrow \infty$, while $T_{desc}^{(\cX_k)} = 2$ for all $k$. It also follows from Corollary \[gra\_seq\_cor\] that $T_{graph}^{(\cX_k)}$ is strictly larger than $T_{desc}^{(\cX_k)}$ when $3\leq k\leq 6$. Equality can hold in Corollary \[gra\_seq\_cor\] — for example, when $k=2$. Indeed, $\cX_2$ is proper, and its MS presentation, $\cG_{\cX_2}$, is irreducible, with $\per(\cG_{\cX_2})=2$, so that $T_{graph}^{(\cX_2)}=2$. From Corollary \[Tseq\_cor\], we also have $T_{seq}^{(\cX_2)}=2$. Thus, $\cX_2$ is an example of a proper PFT $\cX$ in which $T_{seq}^{(\cX)}=T_{graph}^{(\cX)}=T_{desc}^{(\cX)}$ holds.
Thus, to summarize, there appears to be no relationship between the descriptive period of a PFT and its sequential period, as we have examples where each of these can be arbitrarily larger than the other. We have also found that, for a PFT $\cX$, $T_{graph}^{(\cX)}$ can be larger than $T_{desc}^{(\cX)}$. However, we believe that the reverse cannot hold; in fact, we conjecture that $T_{desc}^{(\cX)}$ divides $T_{graph}^{(\cX)}$ for any PFT $\cX$.
Finally, we note that we also have examples of proper PFT’s $\cX$ where $T_{seq}^{(\cX)}$ is arbitrarily larger than $T_{graph}^{(\cX)}$. We omit the proof due to space constraints.
Set $\S=\{0,1\}$ and $k\geq 2$, and let $\cP$ denote the set of all periodic bi-infinite sequences over $\S$ with period $k!$. Consider the PFT $\cX = \cX_{\{\cF,2\}}$ with $\cF=(\cF^{(0)},\emptyset)$, such that $\cF^{(0)}=\{ w\in \Sigma^{2k!}\ : \exists \, \x \in \cP
\text{ such that } w \prec \x\}$. The following statements hold: (a) $\cX$ is proper; (b) $\cX$ is irreducible; and (c) $T_{seq}^{(\cX)}\geq k+1$ and $T_{graph}^{(\cX)}=2$. \[seq>>graph\_thm\]
[1]{}
M.-P. B[é]{}al, M. Crochemore and G. Fici, “Presentations of constrained systems with unconstrained positions,” *IEEE Trans. Inf. Theory*, vol. 51, pp. 1891–1900, May 2005.
D. Lind and B.H. Marcus, *An Introduction to Symbolic Dynamics and Coding,* Cambridge University Press, 1995.
B.E. Moision and P.H. Siegel, “Periodic-finite-type shift spaces,” preprint.
B.E. Moision and P.H. Siegel, “Periodic-finite-type shift spaces,” *Proc. ISIT 2001*, Washington DC, June 24–29, 2001, p. 65.
D.P.B. Chaves and C. Pimentel, “An algorithm for finding the Shannon cover of a periodic shift of finite type,” preprint.
T.L. Poo and B.H. Marcus, “Time-varying maximum transition run constraints,” *IEEE Trans. Inf.. Theory*, vol. 52, pp. 4464–4480, Oct. 2006.
J.C. de Souza, B.H. Marcus, R. New and B.A. Wilson, “Constrained systems with unconstrained positions,” *IEEE Trans. Inf. Theory*, vol. 48, pp. 866–879, April 2002.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Predicted earlier, microcracks on the crystal surface of both finely dispersed $LiYF_4$ powders and single crystals of the Van Vleck paramagnet $%
LiTmF_4$ were detected by using the NMR Cryoporometry and Atomic-Force Microscopy technique.
author:
- |
R.Yu. Abdulsabirov${}^1$, A.A. Bukharaev${}^2$, M.R. Zhdanov${}^1$,\
R.Sh. Zhdanov${}^1$,A.V. Klochkov${}^1$,S.L. Korableva${}^1$,\
V.V. Naletov$%
{}^1$, N.I. Nurgazizov${}^2$, M.S. Tagirov${}^1$, D.A. Tayurskii${}^1$\
${}^1$[*Kazan State University, Kazan, 420008, Russia*]{}\
${}^2$[*Kazan Physical-Technical Institute, 420029, Kazan, Russia*]{}
date: 'PACS[*: 61.16.Ch, 67.55.-s, 68.35.Bs, 76.60.-k*]{}'
title: Detection of nanocracks on double fluoride rare earth crystal surface
---
In paper [@1], in order to explain the specific features of the liquid $%
^3He$ magnetic relaxation in its contact with single crystals of the dielectric Van Vleck paramagnet $LiTmF_4$ and its diamagnetic analog - $%
LiYF_4$, we suggested the following hypothesis: On the surface of these crystals, microcracks exist with an average size near 10 nm. In these microcracks the character of diffusion motion of $^3He$ atoms was supposed to change drastically. In such a bounded geometry, the atoms of $^3He$ determine, in fact, the kinetics of magnetic relaxation of the liquid $^3He$ which is in contact with a solid body substrate. In this work we present our results on the detection of the microcracks, their size being near 30nm (called ”nanocracks”). This was obtained as the result of the two independent methods - the NMR Cryoporometry and the Atomic-Force Microscopy (AFM).
For investigation by means of the NMR Cryoporometry, the three finely dispersed LiYF4 samples with typical sizes: 700 $nm$ (sample-I), 900 $nm$ (sample-II), and 4000 $nm$ (sample-III) were prepared. The technique of powder preparation can be found in [@2]. All samples were immersed in their own containers, the packing factor being about 0.5. The prepared samples can be considered as a porous body containing a three-dimensional net of channels of various dimensions and forms, i.e., in fact, it can be considered as a single pore with a very complicated geometry. For analysis purposes such a space is usually partitioned into a set of pores of various sizes, the pores being assumed to be connect to each other. In this situation, by the term ”pore size” the least distance between opposite walls of a pore is understood. The distribution of the pores with respect to their sizes is usually determined by means of the gas absorption and mercury porometry methods. However, both the methods require long time measurements, rather complicated procedures, and precise experimental equipment. On the other hand, it is well known that the physical properties of a liquid bounded by microscopic pores strongly differ from those of a bulk liquid (see, for example, [@3],[@4]). In part, the melting temperature can lower due to increasing contribution from the relative part of the surface free energy ( for example, the melting temperature of $H_2O$ may occur 60 $K$ lower in a media with average pores about 2 $nm$, see [@5]). This phenomena lies in the base of pores distribution measuring by means of the NMR of water protons. The part of the liquid phase and the corresponding size of pores are determined by the intensity of the pulse-NMR of protons with relatively greater spin-spin relaxation time $T_2$. Obtained in this manner, the dependence of the intensity of the NMR line on the temperature then can be transformed into a curve of pores distribution with respect to their size with help of the probe curve ”change of melting temperature- pore size” (see [@6]).
If we start with the assumption that powder particles are solid spheres with diameter 1000 $nm$, then, for a hexagonal dense packing, our estimates show that the pores whose size is less than 10 $nm$ constitute a part which is less than 1% of the whole quantity. Therefore, we can assert[* a priori*]{} that greater values of the specific weight will demonstrate the presence of developed microrelief on the surface of crystal particles.
The curves of the pore size distribution were obtained by using the data of the impulse NMR of the distilled water protons on the frequencies 20 $MHz$ (a home-made spectrometer) and 80 $MHz$ (Bruker NMR spectrometer W80) for all three samples. This data is shown on the Fig. 1a. Obviously, if the nanocracks have no place, the graph distribution has to have a functional dependence of the type $D^\alpha $, where $D$ stands for the characteristic size of pores, and the exponent $\alpha $ exceeds 1. Our experimental data show that the nanocracks are evidently present on the surface of particles. As one can see on Fig 1b, the distribution has two maximums which correspond to the specific porous sizes of 3-5nm and 15-20nm, their part occupying near 20% of the whole hollow volume of the sample.
To visualize the microrelief of a particle surface, we carried out investigations of the particles of the sample-III by means of the Atomic-Force Microscope P4-SPM-MDT (produced by NT-MDT, Zelenograd, RF). The construction of the microscope allows to assert the presence of the nanocracks with a size greater than 30 $nm$. The AFM-images of various magnifications of a particle surface can be seen on Fig. 2. It is evidently seen that the surface possesses sufficiently developed microrelief.
One may suppose that the nanocracks appear due to the technique of powder preparation, namely, in view of the mechanical grinding. In this connection the investigation of the surface of a single crystal sample seems to be of interest. In view of our further investigations of the magnetic coupling between the liquid $^3He$ and a solid-state substrate, we applied the AFM-investigation to a single crystal of $LiTmF_4$ . The surfaces (100), (110), and (001) of the single crystal were mechanically polished in a certain direction with respect to crystallographic axes (the abrasive for coarse polishing had a grain 1030, while the finishing was made with the GOI-paste with particles smaller than 0.5). One of the AFM-pictures of the single crystal surface is shown on the Fig 3. Our experimental results do not show a correlation between the nanocracks’ character and crystallographic planes. Obviously, the obtained surface microrelief cannot be conditioned as the only result of a mechanical polishing in a certain direction. Thus, investigations detect the presence of the nanocracks on the surface of both particles and single crystals. The most probable reason for the nanocracks’ appearance is formation of local mechanical stresses in a sample sawing by a diamond-charged saw, which then are discharged due to surface reforming.
We should note that the crystal structure of double fluorides of rare earth has no cleavage planes. The AFM experiments carried out with single crystal $%
LiF$ and $CaF_2$ samples which have evident cleavage planes (100) and (111), respectively, showed the absence of nanocracks on the surfaces of a ”fresh” cleavage.
The investigations carried out result in the established presence of nanocracks, predicted earlier by the authors (see [@1]), and their distribution by size on $LiYF_4$ and $LiTmF_4$ crystal surface. This seems to be of importance for both the detailed investigation of the nature of magnetic coupling between liquid $^3He$ and a solid-state substrate, on one hand, and the practical realization of a method of dynamic nuclei polarization of liquid $^3He$ with the use of the dielectric Van Vleck paramagnets. This technique was suggested by the authors in [@7]. From our results it follows that the surface of particles of finely dispersed powders is deformed in a high degree. The latter is in good correspondence with our estimates of the deformed layer part in systems of this kind (see [@8]).
We wish to express gratitude to Prof. for constant interest to our investigations, to Prof. T.Ando (Kanazawa University, Japan) for his preliminary AFM NMR investigations and to P.P Chernov for his support in the NMR investigations.
The work was supported by Russian Foundation for Basic Research (the projects 96-02-16323 and 97-02-16470).
[1]{} V.V.Naletov, M.S.Tagirov, D.A.Tayurskii, and M.A.Teplov, [*JETP*]{}[** 81**]{}, 311 (1995).
A.V.Klochkov, V.V.Naletov, M.S.Tagirov, D.A.Tayurskii, M.A.Teplov, V.N.Efimov, and G.V.Mamin, [*JETP Lett*]{}. [**62**]{}, 585 (1995).
K.A.Jackson, B.Chalmers,[* J. Appl. Phys*]{}. [**29**]{}, 1178 (1958).
J.Warnock, D.D.Awschalom, M.W.Shafer, [*Phys.Rev.Lett*]{}. [**57**]{}, 1753 (1986).
Blachere G.R., Young G.E., [*Journal of American Ceramic Society*]{} [**55**]{}, n.6, 306 (1972).
J.H.Strange and J.B.W.Webber, [*Appl. Magn. Res.*]{} [**12**]{}, 231 (1997).
M.S.Tagirov and D.A.Tayurskii, [*JETP Lett*]{}. [**61**]{}, 672 (1995).
A.V.Klochkov, V.V.Naletov, I.R.Mukhamedshin, M.S.Tagirov, D.A.Tayurskii and H.Suzuki ,[* JETP Lett*]{}. [**66**]{}, 266 (1997).
[**Figure capture**]{}
Fig.1. a) The relative volume of nanocracks in dependence on their size for three samples, obtained by NMR Cryoporometry. b) The distribution of nanocracks with respect to their size.
Fig.2. a), b) The picture of the surface of a particle of sample III, obtained by AFM; one division of the graduation scale is 100 $nm$.
Fig.3. The picture of the surface (110) of a single crystal $LiTmF_4$, obtained by AFM. Polishing was done along the $X$- axis, scanned along the $%
X $-axis. The graduation on both the axes $X$ and $Y$ is 1000 $nm$, on the axis $Z$ -100 $nm$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We find a Lax pair equation corresponding to the Connes-Kreimer Birkhoff factorization of the character group of a Hopf algebra. This flow preserves the locality of counterterms. In particular, we obtain a flow for the character given by Feynman rules, and relate this flow to the Renormalization Group Flow.'
address:
- ' Institute of Mathematics of the Romanian Academy, PO Box 1-764, 014700 Bucharest, Romania and Max-Planck-Institut f" ur Mathematik, P.O. Box 7280, D-53072 Bonn, Germany. `[email protected]`'
- 'Department of Mathematics and Statistics, Boston University, Boston, MA 02215, USA. `[email protected]`'
author:
- Gabriel Bădiţoiu
- Steven Rosenberg
bibliography:
- 'paper-new2.bib'
date: '. Erwin Schrödinger Institut preprint number 2144. AMS classification: 81T15,17B80. Keywords: Renormalization, Lax pair equations, Hopf algebras. '
title: Feynman diagrams and Lax pair equations
---
Introduction
============
In the theory of integrable systems, many classical mechanical systems are described by a Lax pair equation associated to a coadjoint orbit of a semisimple Lie group, for example via the Adler-Kostant-Symes theorem [@adler]. Solutions are given by a Birkhoff factorization on the group, and in some cases, this technique extends to loop group formulations of physically interesting systems such as the Toda lattice [@guest; @sts]. By the work of Connes-Kreimer [@ck1], there is a Birkhoff factorization of characters on general Hopf algebras, in particular on the Kreimer Hopf algebra of 1PI Feynman diagrams. In this paper, we reverse the usual procedure in integrable systems: we construct a Lax pair equation $\frac{d L}{dt}=[L,M]$ on the Lie algebra of infinitesimal characters of the Hopf algebra whose solution is given precisely by the Connes-Kreimer Birkhoff factorization (Theorem \[t:8.2\]). The Lax pair equation is nontrivial in the sense that it is not an infinitesimal inner automorphism. The main technical issue, that the Lie algebra of infinitesimal characters is not semisimple, is overcome by passing to the double Lie algebra with the simplest possible Lie algebra structure. In particular, the Lax pair equation induces a flow for the character given by Feynman rules in dimensional regularization. This flow has the physical significance that it preserves locality, the independence of the character’s counterterm on the mass parameter.
In §§1-4, we introduce a method to produce a Lax pair on any Lie algebra from equations of motion on the double Lie algebra. In §\[hopf\], we apply this method to the particular case of the Lie algebra of infinitesimal characters of a Hopf algebra, and prove Theorem \[t:8.2\].
The Renormalization Group Flow (RGF) usually considered in quantum field theory is a flow on the character group $G_{\mathcal A}$, while the Lax pair flow is on the corresponding Lie algebra ${\mathfrak g}_{\mathcal A}$ of infinitesimal characters. There are various bijections from ${\mathfrak g}_{\mathcal A}$ to $G_{\mathcal A}$, and via these bijections we can compare the Lax pair flow to the RGF. These flows are not the same, so we study how physically significant quantities behave under the Lax pair flow. In §6, we derive an equation for the flow of the $\beta$-function of characters $\varphi_t\in G_{\mathcal A}$ associated to the Lax pair flow via the exponential map $\exp: \mathfrak
g_{\mathcal A}\to G_{\mathcal A}$ (Corollary \[cor:beta\]). In §7, we first show that the Lax pair flow is trivial on primitives in the Hopf algebra. We then use Manchon’s bijection [@man] $\tilde R^{-1}:\mathfrak
g_{\mathcal A}\to G_{\mathcal A}$ to prove various locality results (Theorems \[t:7.9\], \[t:7.14\]). The $\beta$-function flow defined via $\tilde R^{-1}$ itself satisfies a Lax pair equation (Theorem \[t:corr\]). Thus $\tilde R^{-1}$ is much better behaved than the exponential map. In §8, we work out several examples of this theory, and in particular keep track of the leading log terms.
An alternative algebraic geometric approach to Lax pair equations is to apply spectral curve techniques to linearize the flow on the Jacobian of the spectral curve. Unfortunately, in the worked example of §\[worked-example\], the spectral curve is reducible, and the only invariants we find are trivial. We hope to find examples with nontrivial invariants in the future.
We would like to thank Dirk Kreimer for suggesting we investigate the connection between the Connes-Kreimer factorization and integrable systems, and Dominique Manchon for helpful conversations.
The double Lie algebra and its associated Lie Group
===================================================
There is a well known method to associate a Lax pair equation to a Casimir element on the dual $\mathfrak g^*$ of a semisimple Lie algebra $\mathfrak g$ [@sts]. The semisimplicity is used to produce an $\mathrm{Ad}$-invariant, symmetric, non-degenerate bilinear form on $\mathfrak g$, allowing an identification of $\mathfrak g$ with $\mathfrak g^*$. For a general Lie algebra $\mathfrak g$, there may be no such bilinear form. To produce a Lax pair, we need to extend $\mathfrak g$ to a larger Lie algebra with the desired bilinear form. We do this by constructing a Lie bialgebra structure on $\mathfrak g$, whose definition we now recall (see e.g. [@ksch]).
\[bialg\] A Lie bialgebra is a Lie algebra $(\mathfrak g, [\cdot ,\cdot ])$ with a linear map $\gamma:\mathfrak g\to\mathfrak g\otimes\mathfrak g$ such that
- $^t\gamma:\mathfrak g^*\otimes\mathfrak g^*\to\mathfrak g^*$ defines a Lie bracket on $\mathfrak g^*$,
- $\gamma$ is a $1$-cocycle of $\mathfrak g$, i.e. $$\mathrm{ad}^{(2)}_x(\gamma(y))-\mathrm{ad}^{(2)}_y(\gamma(x))-\gamma([x,y])=0,$$ where $\mathrm{ad}^{(2)}_x:\mathfrak g\otimes\mathfrak
g\to\mathfrak g\otimes\mathfrak g$ is given by $\mathrm{ad}^{(2)}_x(y\otimes z)=\mathrm{ad}_x(y)\otimes z+y\otimes
\mathrm{ad}_x(z)
= [x,y]\otimes z + y \otimes [x,z]$.
A Lie bialgebra $(\mathfrak g,[\cdot,\cdot ],\gamma)$ induces an Lie algebra structure on the [*double Lie algebra*]{} $\mathfrak g\oplus\mathfrak g^*$ by $$[X,Y]_{\mathfrak g\oplus\mathfrak g^*}=[X,Y],$$ $$[X^*,Y^*]_{\mathfrak g\oplus\mathfrak g^*}= {}^t\gamma(X\otimes Y),$$ $$[X,Y^*]=\mathrm{ad}^*_X(Y^*),$$ for $X$, $Y\in\mathfrak g$ and $X^*$, $Y^*\in\mathfrak g^*$, where $\mathrm{ad}^*$ is the coadjoint representation given by $\mathrm{ad}^*_X(Y^*)(Z)=-Y^*(\mathrm{ad}_X(Z))$ for $Z\in\mathfrak g$.
Since it is difficult to construct explicitly the Lie group associated to the Lie algebra $\mathfrak g\oplus\mathfrak g^*$, we will choose the trivial Lie bialgebra given by the cocycle $\gamma=0$ and denote by $\delta=\mathfrak g\oplus \mathfrak g^*$ the associated Lie algebra. Let $\{Y_i,
i= 1,\ldots ,l\}$ be a basis of $\mathfrak g$, with dual basis $\{Y^*_i\}$. The Lie bracket $[ \cdot,\cdot ]_\delta$ on $\delta$ is given by $$[Y_i,Y_j]_\delta=[Y_i,Y_j],\ [Y_i^*,Y_j^*]_\delta=0,\
[Y_i,Y_j^*]_\delta=-\sum_k c^j_{ik}Y^*_k,$$ where the $c^j_{ik}$ are the structure constants: $[Y_i,Y_j]=\sum_kc^k_{ij}Y_k$. The Lie group naturally associated to $\delta$ is given by the following proposition.
\[prop22\] Let $G$ be the simply connected Lie group with Lie algebra $\mathfrak{g}$ and let $\theta:G\times \mathfrak g^*\to \mathfrak g^*$ be the coadjoint representation $\theta(g,X)=\mathrm{Ad}^*_{G}(g)(X)$. Then the Lie algebra of the semi-direct product $\tilde G =
G\ltimes_\theta \mathfrak g^*$ is the double Lie algebra $\delta$.
The Lie group law on the semi-direct product $\tilde
G$ is given by $$(g,X)\cdot (g',X')=(gg',X+\theta(g,X')).$$ Let $\tilde{\mathfrak g}$ be the Lie algebra of $\tilde G$. Then the bracket on $\tilde{\mathfrak g}$ is given by $$[X,Y^*]_{\tilde{ \mathfrak g}}
=d\theta(X,Y^*), \ \ [X,Y]_{\tilde {\mathfrak
g}}=[X,Y],\ \ [X^*,Y^*]_{\tilde {\mathfrak g}}=0,$$ for left-invariant vector fields $X$, $Y$ of $G$ and $X^*, Y^*\in\mathfrak g^*$. We have $d\theta(X,Y^*)=d\mathrm{Ad}^*_{G}(X)(Y^*)=[X,Y^*]_\delta$ since $d\mathrm{Ad}_{G}=\mathrm{ad}_{\mathfrak g}$.
The main point of this construction is existence of a good bilinear form on the double.
The natural pairing $\langle\cdot, \cdot\rangle:\delta\otimes\delta\to{\mathbb C}$ given by $$\langle (a, b^*), (c,d^*)\rangle = d^*(a) + b^*(c), \ \ \ a,c\in\mathfrak{g},\ \
b^*, d^*\in\mathfrak{g^*},$$ is an $\mathrm{Ad}$-invariant symmetric non-degenerate bilinear form on the Lie algebra $\delta$.
By [@ksch], this bilinear form is ad-invariant. Since $\tilde G$ is simply connected, the Ad-invariance follows. As an explicit example, we have $$\mathrm{Ad}_{\tilde G}((g,0))(Y_i,0) = (\mathrm{Ad}_G(g)(Y_i),0),\ \
\text{and}\ \ \mathrm{Ad}_{\tilde
G}((g,0))(0,Y_j^*) = (0, \mathrm{Ad}_{G}^*(g)(Y_j^*)),$$ from which the invariance under $\Ad_{\tilde G}(g,0)$ follows.
The loop algebra of a Lie algebra
=================================
Following [@adler], we consider the loop algebra $$L\delta=\{L(\lambda)=\sum\limits_{j=M}^N \lambda^jL_j \ |
\ M,N\in \mathbb Z, L_j\in\delta\}.$$ The natural Lie bracket on $L\delta $ is given by $$\left[\sum \lambda^iL_i,\sum \lambda^j L_j'\right]=
\sum\limits_k \lambda^k\sum\limits_{i+j=k}[L_i,L_j'].$$ Set $$\begin{aligned}
L\delta _+ &=&
\{L(\lambda)= \sum\limits_{j=0}^N
\lambda^jL_j \ | \ N\in\mathbb Z^+\cup \{0\}, L_j\in\delta\}\\
L\delta _-&=&
\{L(\lambda)=\sum\limits_{j=-M}^{-1} \lambda^jL_j \ | \ M\in\mathbb Z^+,
L_j\in\delta\}.\end{aligned}$$ Let $P_+:L\delta \to L\delta _+$ and $P_-:L\delta \to L\delta _-$ be the natural projections and set $R=P_+-P_-$.
The natural pairing $\langle\cdot,\cdot\rangle$ on $\delta$ yields an $\mathrm{Ad}$-invariant, symmetric, non-degenerate pairing on $L\delta $ by setting $$\left\langle\sum\limits_{i=M}^N \lambda^iL_i ,\sum\limits_{j=M'}^{N'}
\lambda^jL'_j \right\rangle=
\sum\limits_{i+j=-1}\langle L_i,L_j'\rangle.$$
For our choice of basis $\{Y_i\}$ of $\mathfrak g$, we get an isomorphism $$\label{I}I: L(\delta ^*)\to L\delta$$ with $$I\left(\sum L^j_iY_j\lambda^i\right)=\sum
L^j_iY^*_j\lambda^{-1-i}.$$ We will need the following lemmas.
[@adler] We have the following natural identifications: $$L\delta _+=L(\delta^*)_- \ \mathrm{and\ } L\delta _-=L(\delta^*)_+.$$
[@sts Lem. 4.1]\[lem1\] Let $\varphi$ be an $\mathrm{Ad}$-invariant polynomial on $\delta$. Then $$\varphi_{m,n}[L(\lambda)]=
\mathrm{Res}_{\lambda=0}(\lambda^{-n}\varphi(\lambda^mL(\lambda)))$$ is an $\mathrm{Ad}$-invariant polynomial on $L\delta $ for $m, n\in\mathbb
Z.$
As a double Lie algebra, $\delta$ has an Ad-invariant polynomial, the quadratic polynomial $$\psi(Y)=\langle Y,Y\rangle$$ associated to the natural pairing. Let $Y_{l+i}=Y^*_i$ for $i\in\{1,\ldots ,l =\mathrm{dim}({\mathfrak g})
\}$, so elements of $L\delta$ can be written $L(\lambda)=\sum\limits_{j=1}^{2l} \sum\limits_{i=-M}^N
L_i^jY_j\lambda^i$. Then the Ad-invariant polynomials $$\label{psimn1}
\psi_{m,n}(L(\lambda))=\mathrm{Res}_{\lambda=0}(\lambda^{-n}
\psi(\lambda^mL(\lambda))),$$ defined as in Lemma \[lem1\] are given by $$\label{psimn2}
\psi_{m,n}(L(\lambda))=2\sum
\limits_{j=1}^l\sum\limits_{i+k-n+2m=-1} L_i^jL_k^{j+l}.$$ Note that powers of $\psi$ are also $\mathrm{Ad}$-invariant polynomials on $\delta $, so $$\label{psimnk}
\psi^k_{m,n}(L(\lambda))=\mathrm{Res}_{\lambda=0}(\lambda^{-n}\psi^k(\lambda^mL(\lambda)))$$ are $\mathrm{Ad}$-invariant polynomials on $L\delta $. It would be interesting to classify all Ad-invariant polynomials on $L\delta$ in general.
The Lax pair equation {#s:7}
=====================
Let $P_+$, $P_-$ be endomorphisms of a Lie algebra $\mathfrak h$ and set $R =
P_+-P_-.$ Assume that $$[X,Y]_R=[P_+X,P_+Y]-[P_-X,P_-Y]$$ is a Lie bracket on $\mathfrak h$. From [@sts Theorem 2.1], the equations of motion induced by a Casimir (i.e. Ad-invariant) function $\varphi$ on ${\mathfrak h}^*$ are given by $$\label{e:5.1}
\frac{dL}{dt}=-\mathrm{ad}^*_{\mathfrak h}M\cdot L,$$ for $L\in\mathfrak h^*,$ where $ M=\frac{1}{2}R(d\varphi(L))\in\mathfrak h.$
Now we take $\mathfrak h=(L\delta)^*=L(\delta^*)$, with $\delta$ a finite dimensional Lie algebra and with the understanding that $(L\delta)^*$ is the graded dual with respect to the standard $\Z$-grading on $L\delta.$ Let $P_\pm$ be the projections of $L\delta^*$ onto $L\delta^*_\pm$. After identifying $L\delta ^*=L\delta $ and $\mathrm{ad}^*=-\mathrm{ad}$ via the map $I$ in (\[I\]), the equations of motion (\[e:5.1\]) can be written in Lax pair form $$\label{e:5.2}
\frac{{d}L}{{d}t}=[M,L],$$ where $ M=\frac{1}{2}R(I(d\varphi(L(\lambda))))\in L\delta,$ and $\varphi$ is a Casimir function on $L\delta^* = L\delta$ [@sts Theorem 2.1]. Finding a solution for (\[e:5.2\]) reduces to the Riemann-Hilbert (or Birkhoff) factorization problem. The following theorem is a corollary of [@adler Theorem 4.37] [@sts Theorem 2.2].
\[t:7.4\] Let $\varphi$ be a Casimir function on $L\delta$ and set $X=I(d\varphi(L(\lambda)))\in L\delta$, for $L(\lambda) = L(0)(\lambda)\in L\delta$. Let $g_{\pm}(t)$ be the smooth curves in $L\tilde G$ which solve the factorization problem $$\exp(-tX)=g_-(t)^{-1}g_+(t),$$ with $g_{\pm}(0)=e$, and with $g_+(t) = g_+(t)(\lambda)$ holomorphic in $\lambda\in {\mathbb C}$ and $g_-(t)$ a polynomial in $1/\lambda$ with no constant term. Let $M=\frac{1}{2}R(I(d\varphi(L(\lambda))))\in L\delta$. Then the integral curve $L(t)$ of the Lax pair equation $$\frac{d L}{d t}=[L, M]$$ is given by $$\label{quick}
L(t)=\mathrm{Ad}_{L\tilde G}g_{\pm}(t)\cdot L(0).$$
This Lax pair equation projects to a Lax pair equation on the loop algebra of the original Lie algebra $\mathfrak g.$ Let $\pi_1$ be either the projection of $\tilde G$ onto $G$ or its differential from $\delta$ onto $\mathfrak g$. This extends to a projection of $L\delta$ onto $L\mathfrak g$. The projection of (\[e:5.2\]) onto $L\mathfrak g$ is $$\label{e:pi1}
\frac{d(\pi_1( L(t)))}{dt}=[\pi_1(L),\pi_1(M)],$$ since $\pi_1 = d\pi_1 $ commutes with the bracket. Thus the equations of motion (\[e:5.2\]) induce a Lax pair equation on $L\mathfrak g$, although this is not the equations of motion for a Casimir on $L \mathfrak g.$
\[t:4.2\] The Lax pair equation of Theorem \[t:7.4\] projects to a Lax pair equation on $L\mathfrak g.$
The content of this theorem is that a Lax pair equation on the Lie algebra of a semi-direct product $G\ltimes G'$ evolves on an adjoint orbit, and the projection onto $\mathfrak g$ evolves on an adjoint orbit and is still in Lax pair form. Lax pair equations often appear as equations of motion for some Hamiltonian, but the projection may not be the equations of motion for any function on the smaller Lie algebra. We thank B. Khesin for this observation.
When $\psi_{m,n}$ is the Casimir function on $L\delta$ given by (\[psimn1\]), $X$ can be written nicely in terms of $L(\lambda)$.
\[p:5.3\] Let $X=I(d \psi_{m,n}(L(\lambda)))$. Then $$\label{e:5.3}
X=2\lambda^{-n+2m} L(\lambda).$$
Write $
L(\lambda) =\sum\limits_{i,j}L_i^j\lambda^{i}Y_j.$ By formula (\[psimn2\]), we have $$\label{e:partialpsimn}
\frac{\partial \psi_{m,n}}{\partial L_p^t}= \left\{
\begin{array}{cc}
2L_{n-1-2m-p}^{t+l}, \ \ \ \text{if} \ \ t\leq l\\
2L_{n-1-2m-p}^{t-l}, \ \ \ \text{if} \ \ t>l.\\
\end{array}
\right.$$ Therefore $$\begin{aligned}
\nonumber X&=&I(d\psi_{m,n}(L(\lambda)))
=\sum\limits_{p,t}\frac{\partial \psi_{m,n}}{\partial L_p^t}\lambda^{-1-p}Y^*_t\\
\nonumber &=&
2\lambda^{-n+2m}\sum\limits_p\left(
\sum\limits_{t=1}^{l}L^{t+l}_{n-1-2m-p}Y_{t+l}\lambda^{n-1-2m-p}
+\sum\limits_{t=l+1}^{2l}
L^{t-l}_{n-1-2m-p}Y_{t-l}\lambda^{n-1-2m-p}\right)\\
\nonumber &=&2 \lambda^{-n+2m}L(\lambda).
\end{aligned}$$
The main theorem for Hopf algebras {#hopf}
==================================
In this section we give formulas for the Birkhoff decomposition of a loop in the Lie group of characters of a Hopf algebra and produce the Lax pair equations associated to the Birkhoff decomposition. We present two approaches, both motivated by the Connes-Kreimer Hopf algebra of 1PI Feynman graphs. First, in analogy to truncating Feynman integral calculations at a certain loop level, we truncate a (possibly infinitely generated) Hopf algebra to a finitely generated Hopf algebra, and solve Lax pair equations on the finite dimensional piece (Theorem \[t:8.2fg\]). We also discuss the compatibility of solutions related to different truncations. Second, we solve a Lax pair equation associated to the full Hopf algebra, but for a restricted family of Casimirs (Theorem \[t:8.2\]).
Let ${\mathcal H}=({\mathcal H}, 1, \mu,\De,\ep,S)$ be a graded connected Hopf algebra over ${\mathbb C}$. Let $\mathcal A$ be a unital commutative algebra with unit $1_\mathcal A$. Unless stated otherwise, $\mathcal{A}$ will be the algebra of Laurent series; the only other occurrence in this paper is ${\mathcal A} = \CC.$
The [**character group**]{} $G_\mathcal A$ of the Hopf algebra ${\mathcal H}$ is the set of algebra morphisms $\phi:{\mathcal H}\to\mathcal{A}$ with $\phi(1)=1_\mathcal{A}.$ The group law is given by the convolution product $$(\psi_1\star\psi_2)(h)=\langle \psi_1\otimes\psi_2,\De h\rangle;$$ the unit element is $\ep$.
An $\mathcal A$-valued [**infinitesimal character**]{} of a Hopf algebra ${\mathcal H}$ is a ${\mathbb C}$-linear map $Z:{\mathcal H}\to\mathcal{A}$ satisfying $$\langle Z,hk\rangle =\langle Z, h\rangle \varepsilon(k)+\varepsilon(h)
\langle Z,k\rangle.$$ The set of infinitesimal characters is denoted by $\mathfrak{g}_\mathcal{A}$ and is endowed with a Lie algebra bracket: $$[Z,Z']=Z\star Z'-Z'\star Z,\ \ \mathrm{for\ }Z,\ Z'\in\mathfrak{g}_\mathcal{A},$$ where $\langle Z\star Z',h\rangle=\langle Z\otimes Z',\Delta(h)\rangle$. Notice that $Z(1)=0$.
For a finitely generated Hopf algebra, $G_\CC$ is a Lie group with Lie algebra ${\mathfrak g}_\CC$, and for any Hopf algebra and any ${\mathcal A}$, the same is true at least formally.
We recall that $\delta=\mathfrak{g}_{\mathcal{{\mathbb
C}}}\oplus\mathfrak{g}_{\mathcal{{\mathbb C}}}^{*}$ is the double of $\mathfrak{g}_{\mathbb C}$ and the $\mathfrak{g}_{\mathbb C}^{*}$ is the graded dual of $\mathfrak{g}_{\mathbb C}$. We consider the algebra $\Omega\delta = \delta\otimes\mathcal{A}$ of formal Laurent series with values in $\delta$ $$\Omega\delta=\{L(\lambda)=\sum\limits_{j=-N}^\infty \lambda^jL_j \ |
\ L_j\in\delta, N\in \Z\}.$$ The natural Lie bracket on $\Omega\delta $ is $$\left[\sum \lambda^iL_i,\sum \lambda^j L_j'\right]=
\sum\limits_k \lambda^k\sum\limits_{i+j=k}[L_i,L_j'].$$ Set $$\begin{aligned}
\Omega\delta _+ &=&
\{L(\lambda)= \sum\limits_{j=0}^\infty
\lambda^jL_j \ | \ L_j\in\delta\}\\
\Omega\delta _- &=& \{L(\lambda)=\sum\limits_{j=-N}^{-1}
\lambda^jL_j \ | \ L_j\in\delta, N\in \Z^+\}.\end{aligned}$$
Recall that for any Lie group $K$, a loop $L(\lambda)$ with values in $K$ has a Birkhoff decomposition if $L(\lambda) = L(\lambda)
_-^{-1}L(\lambda)_+$ with $L(\lambda)_-^{-1}$ holomorphic in $\lambda^{-1} \in \mathbb P^1 -\{0\}$ and $L(\lambda)_+$ holomorphic in $\lambda \in\mathbb P^1 - \{\infty\}.$ In the next lemma, $\tilde G$ refers to $G\ltimes_\theta \mathfrak g^*$ as in Prop. \[prop22\].
We prove the existence of a Birkhoff decomposition for any element $(g,\alpha)\in \Omega\tilde G$.
\[t:omegatildeg\] Every $(g,\alpha)\in\Omega\tilde
G=G_\mathcal{A}\ltimes_{Ad^*_{G_{\mathcal A}}}
\mathfrak{g}_{\mathcal{A}}^{*}
$ has a Birkhoff decomposition $(g,\alpha)=(g_-,\alpha_-)^{-1}(g_+,\alpha_+)$ with $(g_+,\alpha_+)$ holomorphic in $\lambda$ and $(g_-,\alpha_-)$ a polynomial in $\lambda^{-1}$ without constant term.
We recall that $(g_1,\alpha_1)(g_2,\alpha_2)=
(g_1g_2,\alpha_1+\mathrm{Ad}^*(g_1)(\alpha_2)).$ Thus\
$
(g,\alpha)=(g_-,\alpha_-)^{-1}(g_+,\alpha_+)
\text{ if and only if }
g=g_-^{-1}g_+
\text{ and }
\alpha=\mathrm{Ad}^*(g_-^{-1})(-\alpha_-+\alpha_+).
$ Let $g=g_-^{-1}g_+$ be the Birkhoff decomposition of $g$ in $ G_\mathcal{A}$ given in [@ck1; @egk; @man]. Set $\alpha_+=P_+(\mathrm{Ad}^*(g_-)(\alpha))$ and $\alpha_-=-P_-(\mathrm{Ad}^*(g_-)(\alpha))$, where $P_+$ and $P_-$ are the holomorphic and pole part, respectively. Then for this choice of $\alpha_+$ and $\alpha_-$, we have $(g,\alpha)=(g_-,\alpha_-)^{-1}(g_+,\alpha_+)$. Note that the Birkhoff decomposition is unique.
For a finitely generated Hopf algebra, we can apply Theorems \[t:7.4\], \[t:4.2\] to produce a Lax pair equation on $L\delta$ and on the loop space of infinitesimal characters $L\mathfrak{g}$. However, the common Hopf algebras of 1PI Feynman diagrams and rooted trees are not finitely generated.
As we now explain, we can truncate the Hopf algebra to a finitely generated Hopf algebra, and use the Birkhoff decomposition to solve a Lax pair equation on the infinitesimal character group of the truncation. A graded Hopf algebra $\mathcal H=\oplus_{n\in\mathbb N} \mathcal H_n$ is said to be of [**finite type**]{} if each homogeneous component $\mathcal H_n$ is a finite dimensional vector space. Let $\mathcal B=\{T_i\}_{i\in\mathbb N}$ be a minimal set of homogeneous generators of the Hopf algebra $H$ such that $\deg(T_i)\leq\deg(T_j)$ if $i<j$ and such that $T_0=1$. For $i>0$, we define the $\mathbb{C}$-valued infinitesimal character $Z_i$ on generators by $Z_i(T_j)=\delta_{ij}.$ The Lie algebra of infinitesimal characters $\mathfrak g$ is a graded Lie algebra generated by $\{Z_i\}_{i>0}$. Let $\mathfrak{g}^{(k)}$ be the vector space generated by $\{Z_i\ | \ \deg(T_i)\leq k\}$. We define $\deg(Z_i)=\deg(T_i)$ and set $$[Z_i,Z_j]_{\mathfrak{g}^{(k)}}=\left\{
\begin{array}{cc}
[Z_i,Z_j] & \text{if }\deg(Z_i)+\deg(Z_j)\leq k \\
0 & \text{if }\deg(Z_i)+\deg(Z_j)>k
\end{array}\right.$$ We identify $\varphi\in G_{\mathbb C}$ with $\{\varphi(T_i)\}\in\mathbb C^{\mathbb N}$ and on $\mathbb
C^{\mathbb N}$ we set a group law given by $\{\varphi_1(T_i)\}\oplus\{\varphi_2(T_i)\}=\{(\varphi_1\star\varphi_2)(T_i)\}$. $G^{(k)}=\{ \{\varphi(T_i)\}_{\{i\, |\, \deg(T_i)\leq
k\}}\ | \ \varphi\in G_{\mathbb C}\}$ is a finite dimensional Lie subgroup of $G_{\mathbb C}=(\mathbb C^{\mathbb N},\oplus)$ and the Lie algebra of $G^{(k)}$ is $\mathfrak{g}^{(k)}$. There is no loss of information under this identification, as $\varphi(T_iT_j)=\varphi(T_i)\varphi(T_j)$.
Let $\delta^{(k)}$ be the double Lie algebra of $\mathfrak{g}^{(k)}$ and let $\tilde G^{(k)}$ be the simply connected Lie group with $\mathrm{Lie}(\tilde G^{(k)})=\delta^{(k)}$ as in Proposition \[prop22\]. The following theorem is a restatement of Theorem \[t:7.4\] in our new stage.
\[t:8.2fg\] Let $\mathcal H=\oplus_{n}\mathcal
H_n$ be a graded connected Hopf algebra of finite type, and let $\psi:L\delta^{(k)}\to{\mathbb C}$ be a Casimir function (e.g. $\psi(L ) = \psi_{m,n}(L(\lambda))=\mathrm{Res}_{\lambda=0}
(\lambda^m\psi(\lambda^n L(\lambda)))$ with $\psi:\delta^{(k)}\times\delta^{(k)}\to{\mathbb C}$ the natural paring of $\delta^{(k)}$). Set $
X=I(d\psi(L_0 ))$ for $L_0 \in L\delta^{(k)}$. Then the solution in $L\delta^{(k)}$ of $$\label{8:1fg}
\frac{dL}{dt}=[L,M]_{L\delta^{(k)}}, \ \ \ M=\frac{1}{2}
R(I(d\psi(L)))$$ with initial condition $L(0)=L_0$ is given by $$\label{e:8.3fg}
L(t)=\mathrm{Ad}_{L\tilde G^{(k)}}g_{\pm}(t)\cdot L_0,$$ where $\exp(-tX)$ has the Connes-Kreimer Birkhoff factorization\
$\exp(-tX)=g_-(t)^{-1}g_+(t)$.
\(i) If $L_0\in L\delta$, there exists $k\in\mathbb N$ such that $L_0\in L\delta^{(k)}$. Indeed $L_0\in L\delta$ is generated over $\mathbb C[\lambda,\lambda^{-1}]$ by a finite number of $\{Z_i\}$, and we can choose $k\geq\max\{\deg(Z_i)\}$.
\(ii) While the Hopf algebra of rooted trees and the Connes-Kreimer Hopf algebra of 1PI Feynman diagrams satisfy the hypothesis of Theorem \[t:8.2fg\], the Feynman rules character does not lie in $L\tilde G$, as explained below.
In the next sections, we will investigate the relationship between the Lax pair flow $L(t)$ and the Renormalization Group Equation. In preparation, we project from $L\delta^{(k)}$ to $L\mathfrak{g}^{(k)}$ via $\pi_1$ as in §4.
\[tc:8.2fg\] Let $\psi $ be a Casimir function on $L\delta^{(k)}$. Set $L_0\in L\mathfrak{g}^{(k)}\subset L\delta^{(k)}$, $X=\pi_1(I(d\psi(L_0)))$. Then the solution of the following equation in $L\mathfrak{g}^{(k)}$ $$\label{c:e:8.1}
\frac{dL}{dt}=[L,M_1]_{L\mathfrak{g}^{(k)}}, \ \ \ M_1=\pi_1(\frac{1}{2}
R(I(d\psi (L ))))$$ with initial condition $L(0)=L_0$ is given by $$\label{c:e:8.3}
L(t)=\mathrm{Ad}_{L G^{(k)}}g_{\pm}(t)\cdot L_0,$$ where $\exp(-tX)$ has the Connes-Kreimer Birkhoff factorization in $L\mathfrak{g}^{(k)}$ $$\exp(-t X)=g_-(t)^{-1}g_+(t).$$
\[lastrem\] (i) For Feynman graphs, this truncation corresponds to halting calculations after a certain loop level. From our point of view, this truncation is somewhat crude. $\mathfrak g^{(k)}$ is not a subalgebra of $\mathfrak g$, and if $k < \ell$, $\mathfrak g^{(k)}$ is not a subalgebra of $\mathfrak g^{(\ell)}$. Although the Casimirs $\psi_{m,n}$ and the exponential map restrict well from $\mathfrak g$ to $\mathfrak g^{(k)}$, the Birkhoff decomposition $\exp(-tX)$ of $X\in
L{\mathfrak g}^{(k)}$ is very different from the Birkhoff decompositions in $L\mathfrak g, L\mathfrak g^{(\ell)}$. In fact, if $g\in G^{(k)}$ has Birkhoff decomposition $g = g_-^{-1}g_+$ in $G$, there does not seem to be $f(k)\in \mathbb N$ such that $g_\pm\in G^{(f(k))}.$ Nevertheless, in the last section we will follow standard procedure and present calculations of truncated Hopf algebras.
(ii)It would interesting to know, especially for the Hopf algebras of Feynman graphs or rooted trees, whether there exists a larger connected graded Hopf algebra ${\mathcal H}'$ containing ${\mathcal H}$ such that the associated infinitesimal Lie algebra $\mathrm{Lie}(G'_\mathbb{C})$ is the double $\delta$. This would provide a Lax pair equation associated to an equation of motion on the infinitesimal Lie algebra of ${\mathcal H}'$. The most natural candidate, the Drinfeld double $\mathcal D({\mathcal H})$ of ${\mathcal H}$, does not work since the dimension of the Lie algebra associated to $\mathcal D({\mathcal H})$ is larger than the dimension of $\delta$.
In [@ck1], Connes and Kreimer give a Birkhoff decomposition for the character group of the Hopf algebra of 1PI graphs, and in particular for the Feynman rules character $\varphi(\lambda)$ given by minimal subtraction and dimensional regularization. The truncation process treated above does not handle the Feynman rules character, as the Feynman rules character and the toy model character of the Hopf algebra of rooted trees considered in §8 are not polynomials in $\lambda,\lambda^{-1}$, but Laurent series in $\lambda$. Thus Corollary \[tc:8.2fg\] does not apply, as in our notation $\log(\varphi(\lambda)) \in
\Omega\mathfrak{g} \setminus L\mathfrak{g}$. This and Remark \[lastrem\](i) force us to consider a direct approach in $\Omega\mathfrak{g}$ as in next theorem. However, we cannot expect that the Lax pair equation is associated to any Hamiltonian equation, and we replace Casimirs with Ad-covariant functions.
[@suris] Let $G$ be a Lie group with Lie algebra $\mathfrak{g}$. A map $f:\mathfrak{g}\to \mathfrak{g}$ is $\mathrm{Ad}$-[*covariant*]{} if $\mathrm{Ad}(g)(f(L))=f(\mathrm{Ad}(g)(L))$ for all $g\in G$, $L\in\mathfrak{g}$.
\[t:8.2\] Let ${\mathcal H}$ be a connected graded commutative Hopf algebra with $\mathfrak{g}_{\mathcal A}$ the associated Lie algebra of infinitesimal characters with values in Laurent series. Let $f:\mathfrak{g}_{\mathcal A}\to\mathfrak{g}_{\mathcal A}$ be an $\mathrm{Ad}$-covariant map. Let $L_0\in\mathfrak{g}_{\mathcal A}$ satisfy $[f(L_0),L_0]=0$. Set $X=f(L_0)$. Then the solution of $$\label{8:1}
\frac{dL}{dt}=[L,M], \ \ \ M=\frac{1}{2}
R(f(L))$$ with initial condition $L(0)=L_0$ is given by $$\label{e:8.3}
L(t)=\mathrm{Ad}_{G}g_{\pm}(t)\cdot L_0,$$ where $\exp(-tX)$ has the Connes-Kreimer Birkhoff factorization\
$\exp(-tX)=g_-(t)^{-1}g_+(t)$.
The proof is similar to [@sts Theorem 2.2]. First notice that $$\begin{aligned}
\frac{d}{dt}\left({\Ad}(g_-(t)^{-1}g_+(t))\cdot
L_0\right) &=&
\frac{d}{dt}(\exp(-tX)L_0\exp(tX))\\
&=& -\exp(-tX)XL_0\exp(tX)+\exp(-tX)L_0X\exp(tX)\\
&=& \exp(-tX)[X,L_0]\exp(tX)=0,\end{aligned}$$ which implies ${\Ad}(g_-(t)^{-1}g_+(t))\cdot
L_0=L_0$ and ${\Ad}(g_-(t) )\cdot L_0={\Ad}(g_+(t))\cdot L_0$. Set $L(t)={\Ad}(g_\pm(t) )\cdot L_0 = g_\pm(t)L_0g_\pm(t)^{-1}$. As usual, $$\frac{dL}{dt}
= \left[\frac{d g_\pm(t)}{dt}g_\pm(t)^{-1},L(t)\right],$$ so $$\frac{dL}{dt}=\frac{1}{2}\left[\frac{d g_+(t)}{dt}g_+(t)^{-1}+\frac{d
g_-(t)}{dt}g_-(t)^{-1},L(t)\right].$$ The Birkhoff factorization $g_+(t)=g_-(t)\exp(-tX)$ gives $$\frac{dg_+(t)}{dt}
=\frac{dg_-(t)}{dt}\exp(-tX)+g_-(t)(-X)\exp(-tX),$$ and so $$\frac{dg_+(t)}{dt}g_+(t)^{-1}
=\frac{dg_-(t)}{dt}g_-(t)^{-1}+g_-(t)(-X)g_-(t)^{-1}.$$ Thus $$\begin{aligned}
2M &=& R(f(L(t)))=R(f({\Ad}( g_-(t))\cdot L_0))=R({\Ad}( g_-(t))\cdot f(L_0))\\
&=& R({\Ad}( g_-(t))\cdot X)) =
-R(\frac{dg_+(t)}{dt}g_+(t)^{-1})
+R(\frac{dg_-(t)}{dt}g_-(t)^{-1})\\
&= & -\frac{dg_+(t)}{dt}g_+(t)^{-1}
-\frac{dg_-(t)}{dt}g_-(t)^{-1}.\end{aligned}$$ Here we use $(\frac{dg_\pm(t)}{dt}g_\pm(t)^{-1})(x)\in\mathcal
A_\pm$ for $x\in {\mathcal H}$. Thus $
\frac{dL}{dt}= [L,M].$
If $f:\mathfrak{g}_\mathcal{A}\to\mathfrak{g}_\mathcal{A}$ is given by $f(L)=2\lambda^{-n+2m}L$, then $f$ is $\mathrm{Ad}$-covariant and $[f(L_0),L_0]=[2\lambda^{-n+2m}L_0,L_0]=0$.
\[tc:8.2\] Let ${\mathcal H}$ be a connected graded commutative Hopf algebra with $\mathfrak{g}_{\mathcal A}$ the Lie algebra of infinitesimal characters with values in Laurent series. Pick $L_0\in\mathfrak{g}_{\mathcal A}$ and set $X=2\lambda^{-n+2m}L_0$. Then the solution of $$\label{8:1gen}
\frac{dL}{dt}=[L,M], \ \ \ M=
R(\lambda^{-n+2m} L)$$ with initial condition $L(0)=L_0$ is given by $$\label{e:8.3gen}
L(t)=\mathrm{Ad}_{ G_\mathcal{A}}g_{\pm}(t)\cdot L_0,$$ where $\exp(-tX)$ has the Connes-Kreimer Birkhoff factorization\
$\exp(-tX)=g_-(t)^{-1}g_+(t)$.
\[FRrmk\] Let $\varphi$ be the Feynman rules character. We can find the Birkhoff factorization of $\varphi$ itself within this framework by adjusting the initial condition. Namely, set $L_0(\lambda)=\frac{1}{2}\lambda^{n-2m}\exp^{-1}(\varphi(\lambda)).$ Then $\exp(X) = \varphi$ by Prop. \[p:5.3\], so the solution of involves the Birkhoff factorization $ \varphi =
g_-(-1)^{-1}g_+(-1)$. Namely, we have $$L (-1 )= \frac{\lambda^{n-2m}}{2}
\mathrm{Ad}_{ G_\mathcal{A}}g_\pm (-1 )
\exp^{-1}( \varphi).$$
The Connes-Kreimer $\beta$-function
===================================
The flow of characters usually considered in quantum field theory is the renormalization group flow (RGF). In contrast, the Lax pair flow lives on the Lie algebra of the character group. Since the $\beta$-function of the RGF is an element of the Lie algebra of the $\CC$-valued characters, it is natural to examine the relationship between the Lax pair equations and the $\beta$-function. In this section, we continue to work in the general setup of Hopf algebras and character groups.
Here we consider two flows for the $\beta$-function. First, we extend the (scalar) beta function of a local character $\varphi$ (see (\[locality\])) to an infinitesimal character $\tilde \beta_\varphi$ (Lemma \[l:holo\]). This “beta character” has already appeared in the literature: $\tilde\beta_\varphi = \lambda\tilde R(\varphi)$, in the language of [@man] explained below (Lemma \[lemma:6.6\]), but it seems worth highlighting. For certain Casimirs, we show that the beta character is a fixed point of the Lax pair flow (Theorem \[t:4.5.3\]).
It is more important and more difficult to consider the flow of the $\beta$-function itself. Namely, given a character $\varphi$, we can set $L_0 =
\log(\varphi)$ and study the $\beta$-functions of the characters $\varphi(s)
= \exp(L(s)).$ In Theorem \[t:flow\], we give a differential equation for $\beta_{\varphi(s)}.$
To define the beta character, we recall material from [@ck2; @ef-manchon; @man]. Throughout this section, $\mathcal A$ denotes the algebra of Laurent series.
Let $\mathcal H=\bigoplus\limits_{n}\mathcal H_n$ be a connected graded Hopf algebra. Let $Y$ be the biderivation on $\mathcal H$ given on homogeneous elements by $$Y:\mathcal H_n\to \mathcal H_n,\ \ \ \ \ Y(x)=nx\ \ \ \text{for }
x \in\mathcal H_n.$$
\[def:6.1\] [@man] We define the bijection $\tilde R:G_{\mathcal A}\to\mathfrak g_{\mathcal A}$ by $$\tilde R(\varphi)=\varphi^{-1}\star(\varphi\circ Y).$$
Consider the semidirect product Lie algebra $\tilde{\mathfrak g}_{\mathcal A}
=\mathfrak g_{\mathcal A}\rtimes{\mathbb C}\cdot Z_0,$ where $Z_0$ acts via $[Z_0,X]=X\circ Y$ for $X\in\mathfrak g_{\mathcal A}$. Let $\{\theta_t\}_{t\in\mathbb C}$ be the one-parameter group of automorphisms of $\mathcal
H$ given by $$\theta_t(x)=e^{nt}x, \text{ for }x\in\mathcal H_n.$$ Then $\varphi\ \
\mapsto
\varphi\circ\theta_t$ is an automorphism of $G_{\mathcal A}$. Let $\tilde{G}_{\mathcal A}$ be the semidirect product $$\tilde{G}_{\mathcal A}=G_{\mathcal A}\rtimes\mathbb C ,$$ with the action of $\mathbb C $ on $G_{\mathcal A}$ given by $\varphi\cdot t=\varphi\circ\theta_t$. $G_{\mathcal A}$ has Lie algebra $\mathfrak g_{\mathcal A}$.
We now define a second action of $\mathbb C$ on $G_{\mathcal A}$. For $t\in\mathbb C$ and $\varphi\in G_{\mathcal A}$ we define $\varphi^t(x)$ on an homogeneous element $x\in \mathcal H$ by $$\varphi^t(x)(\lambda)=e^{t\lambda |x|} \varphi(x)(\lambda),$$ for any $\lambda\in\mathbb C$, where $|x|$ is the degree of $x$.
Let $$\label{locality}
G^{\Phi}_{\mathcal A}=\{\varphi\in G_{\mathcal A}\ \big|\ \frac{ \ \ d}{dt}
(\varphi^t)_-=0\},$$ be the set of characters with the negative part of the Birkhoff decomposition independent of $t$. Elements of $ G^{\Phi}_{\mathcal A}$ are called [*local characters*]{}.
The dimensional regularized Feynman rule character $\varphi $ is local. Referring to [@ck2; @ef-manchon], the physical meaning of locality is that the counterterm $\varphi_-$ does not depend on the mass parameter $\mu$: $\frac{\partial\varphi_- }{\partial\mu}=0$.
Let $\varphi\in G_{\mathcal A}^{\Phi}$. Then the limit $$F_\varphi(t)=\lim\limits_{\lambda\to 0}\varphi ^{-1}(\lambda)\star
\varphi^t (\lambda)$$ exists and is a one-parameter subgroup in $G_{\mathcal A}\cap G_{\mathbb C}$ of scalar valued characters of $\mathcal H$.
Notice that $(\varphi ^{-1}(\lambda)\star \varphi^t (\lambda))(\Gamma)\in\mathcal A_+$ as $$\varphi ^{-1}(\lambda)\star \varphi^t
(\lambda)=\varphi_+^{-1}\star\varphi_-\star
(\varphi^t)_-^{-1}\star(\varphi^t)_+=
\varphi_+^{-1}\star(\varphi^t)_+.$$
For $\varphi\in G^\Phi_{\mathcal A}$, the $\beta$-function of $\varphi$ is defined to be $\beta_\varphi=-(\mathrm{Res}(\varphi_-))\circ Y)$.
We have [@ck2] $$\beta_\varphi={d\over dt}\Big|_{t=0} F_{\varphi_-^{-1}}(t),$$ where $F_{\varphi_-^{-1}}$, the one-parameter subgroup associated to $\varphi_-^{-1}$, also belongs to $G^\Phi_\mathcal{A}$.
To relate the $\beta$-function $\beta_\varphi\in {\mathfrak g}_\CC$ to our Lax pair equations, which live on ${\mathfrak g}_{\mathcal A}$, we can either consider ${\mathfrak g}_\CC$ as a subset of ${\mathfrak g}_{\mathcal A}$, or we can extend $\beta_\varphi$ to an element of ${\mathfrak g}_{\mathcal A}.$ Since ${\mathfrak g}_\CC$ is not preserved under the Lax pair flow, we take the second approach.
For $\varphi\in G_{\mathcal A}^{\Phi}$, $x\in H$, set $$\tilde{\beta}_\varphi(x)(\lambda) ={ d\over dt}\Big|_{t=0}
(\varphi^{-1}\star\varphi^t)(x)(\lambda).$$
The following lemma establishes that $\tilde\beta$ is an infinitesimal character.
\[l:holo\] Let $\varphi\in G_{\mathcal A}^{\Phi}$.
i\) $\tilde{\beta}_\varphi $ is an infinitesimal character in $\mathfrak g_{\mathcal A}$.\
ii) $\tilde{\beta}_\varphi $ is holomorphic (i.e. $\tilde{\beta}_\varphi(x)\in\mathcal A_+ $ for any $x$).
i\) For two homogeneous elements $x,y\in\mathcal H$, we have: $$\varphi^t(xy)=e^{t|xy|\lambda}\varphi(xy)=
e^{t|x|\lambda}\varphi(x)e^{t|y|\lambda}\varphi(y)=\varphi^t(x)\varphi^t(y).$$ Therefore $\varphi\star\varphi^t\in G_{\mathcal A}$. Since $\varphi^{-1}\star\varphi^0=e$ we get $${d\over dt}\Big|_{t=0} \varphi^{-1}\star\varphi^t\in\mathfrak
g_{\mathcal A}.$$
ii\) Since ${d\over dt}(\varphi^t)_-=0$, we get $$\tilde{\beta}_\varphi=(\varphi_+)^{-1}\star\varphi_-\star((\varphi^t)_-)^{-1}\star(\varphi^t)_+
=(\varphi_+)^{-1}\star(\varphi^t)_+.$$ Then $$\tilde{\beta}_\varphi(x)=(\varphi_+)^{-1}(x')(\varphi^t)_+(x'')
=(\varphi_+)(S(x'))(\varphi^t)_+(x'')$$ Therefore $\tilde{\beta}_\varphi(x)\in\mathcal A_+$.
\[lemma:6.6\] If $\varphi\in G^\Phi_{\mathcal{A}}$ then
\(i) $\tilde\beta_\varphi=\lambda\tilde R(\varphi)$,
\(ii) $\beta_\varphi=\mathrm{Ad}(\varphi_+(0))(\tilde\beta_\varphi\big|_{\lambda=0})$,
\(iii) $\tilde{\beta}_{\varphi_-}(x)(\lambda = 0) =-\beta_\varphi(x).$
\(i) For $\Delta(x)=x'\otimes x''$, we have $$\tilde{\beta}_\varphi(x)(\lambda) ={ d\over dt}\Big|_{t=0}
(\varphi^{-1}\star\varphi^t)(x)(\lambda)=
\varphi^{-1}(x'){ d\over dt}\Big|_{t=0}(\varphi^t)(x'')$$ $$=
\varphi^{-1}(x')\lambda\cdot
\mathrm{deg}(x'')\varphi(x'')=\lambda\varphi^{-1}(x')\varphi\circ
Y(x'')=\lambda(\varphi^{-1}\star(\varphi\circ Y))(x)=\lambda\tilde
R(\varphi)(x)
.$$
\(ii) The cocycle property of $\tilde R$ [@ef-manchon], $\tilde
R(\phi_1\star\phi_2)=\tilde R(\phi_2)+\phi_2^{-1}\star\tilde
R(\phi_1)\star\phi_2$, implies that $$\begin{aligned}
\label{e:e:6.6}
\lambda\tilde R(\varphi)=\lambda\tilde
R(\varphi_-^{-1}\star\varphi_+)=\lambda\tilde
R(\varphi_+)+\varphi_+^{-1}\star\lambda\tilde
R(\varphi_-^{-1})\star\varphi_+.\end{aligned}$$ Since $\tilde
R(\varphi_+)=\varphi_+^{-1}\star(\varphi_+\circ Y)$ is always holomorphic and since $\lambda\tilde
R(\varphi_-^{-1})=\mathrm{Res}(\varphi_-^{-1})\circ
Y=-\mathrm{Res}(\varphi_-)\circ
Y=\beta$ by [@man Theorem IV.4.4], when we evaluate (\[e:e:6.6\]) at $\lambda=0$ we get $\tilde\beta(\varphi)\big|_{\lambda=0}=\mathrm{Ad}(\varphi_+^{-1}(0))\beta$.
\(iii) The Birkhoff decomposition of $\varphi_-=(\varphi_-)_-^{-1}\star(\varphi_-)_+$ is given by $(\varphi_-)_-=\varphi_-^{-1}$ and $(\varphi_-)_+=\varepsilon$. By definition, $\beta_{\varphi_-}=-\mathrm{Res}((\varphi_-)_-)\circ Y
=-\mathrm{Res}(\varphi_-^{-1})\circ Y
=\mathrm{Res}(\varphi_- )\circ Y=-\beta_\varphi
$. Applying (ii) to $\varphi_-$, we get $$-\beta_{\varphi} = \beta_{\varphi_-}
=\mathrm{Ad}(\varepsilon\big|_{\lambda=0})
(\tilde\beta_{\varphi_-}\big|_{\lambda=0})
=\tilde\beta_{\varphi_-}\big|_{\lambda=0}.$$
If $\varphi\in G_{\mathcal A}^\Phi$, the Lax pair equation in Corollary \[tc:8.2\] for $L_0=\tilde{\beta}_\varphi $ is $$\begin{aligned}
\label{p:beta-flow-h}
{d\over ds}\tilde{\beta}_\varphi(s)=[\tilde{\beta}_\varphi(s),M],\end{aligned}$$ where $M=R(\lambda^{-n+2m}\tilde{\beta}_\varphi(s))$ and the solution is given by $$\begin{aligned}
\label{p:beta-flow}
\tilde{\beta}_\varphi(s)=\mathrm{Ad}(g_+(s))\tilde{\beta}_\varphi(0)\end{aligned}$$ for $g_\pm(s)$ given by the Birkhoff decomposition $\exp(s\lambda^{-n+2m}\tilde{\beta}_\varphi)=g^{-1}_-(s)\star
g_+(s).$
The next theorem shows that the $\beta$-function is a fixed point of the Lax pair flow for certain Casimirs. Of course this is not the same as having the $\beta$-function a fixed point of the RGF.
\[t:4.5.3\] $\tilde{\beta}_\varphi(s)$ and therefore $\beta_\varphi(s)=\mathrm{Ad}(\varphi_+(s)\Bigl|_{\lambda=0}
)(\tilde{\beta}_\varphi(s)\Big|_{\lambda=0})$ are constant under the Lax flow if $-n+2m\geq 0$.
We drop $s$ from the notation. If $-n+2m\geq 0$ then $$M=R( \lambda^{-n+2m}\tilde{\beta}_\varphi)=
\lambda^{-n+2m}\tilde{\beta}_\varphi,$$ since $\tilde{\beta}_\varphi$ is holomorphic by Lemma \[l:holo\] and Theorem \[p:holo\]. (The proof of this Theorem is independent of this section.) So the Lax pair equation becomes $${d\over ds}\tilde{\beta}_\varphi
=[ \tilde\beta_\varphi, \lambda^{-n+2m}\tilde{\beta}_\varphi]
=
\lambda^{-n+2m}[\tilde{\beta}_\varphi, \tilde{\beta}_\varphi]=0.$$
Now we consider the more interesting case of the flow $\beta_{\psi(s)}$ of the $\beta$-function of exponentiated infinitesimal characters. We first establish some simple properties of $\varphi^t$.
Let $\varphi\in G_{\mathcal A}$.
\(i) $(\varphi\star\psi)^t=\varphi^t\star\psi^t$,
\(ii) $(\varphi^{-1})^t=(\varphi^t)^{-1}$,
We have $$\begin{aligned}
(\varphi\star\psi)^t(x)&=&e^{t|x|\lambda}(\varphi\star\psi)
(x)=\sum\limits_{(x)}e^{t|x|\lambda}\varphi(x')\psi(x'') \\
&=&\sum\limits_{(x)}e^{t(|x'|+|x''|)\lambda}\varphi(x')\psi(x'')
=e^{t|x'|\lambda}\varphi(x')\ e^{t|x''|\lambda}\psi(x'')
=\varphi^t(x')\psi^t(x'')\\
&=&(\varphi^t\star\psi^t)(x).
\end{aligned}$$ Therefore $$\varphi^t\star(\varphi^{-1})^t
=(\varphi\star\varphi^{-1})^t=\varepsilon^t=\varepsilon=\varphi^t\star(\varphi^t)^{-1},$$so $(\varphi^{-1})^t=(\varphi^t)^{-1}$.
The exponential map $\exp:{\mathfrak g}_{\mathcal A}\to G_{\mathcal A}$ is a bijection. Therefore, we can transfer the Lax pair flow on ${\mathfrak g}_{\mathcal
A}$ to a flow on $G_{\mathcal A}$, and study the associated flow of beta characters.
\[t:flow\] Let $\varphi\in G^\Phi_\mathcal A$. Let $$\dot\psi(s)=[\psi(s),M]$$ be the Lax pair from Theorem \[t:8.2\] with $\psi(0)=\psi = \log(\varphi)$. Let $\varphi(s)=\exp(\psi(s))$. For $$\tilde{\beta}_{\varphi(s)}=
{d\over dt}\Big|_{t=0}\varphi(s)^{-1}\star(\varphi(s))^t,$$ we have $$\begin{aligned}
\label{e:t:phi}
{d\over ds}
\tilde{\beta}_{\varphi(s)}&=&
[\tilde{\beta}_{\varphi(s)},\varphi^{-1}(s)\star
d\exp[\log\varphi(s),M]]\\
&&\qquad +\lambda(\varphi^{-1}(s)\star d\exp[\log\varphi(s),M])\circ Y.\nonumber\end{aligned}$$
Omitting some stars, we have $$\begin{aligned}
\label{e:beta1}
{d\over ds}
\tilde{\beta}_{\varphi(s)}(x)
\nonumber &=&
{d\over ds} {d\over dt}\Big|_{t=0}(\varphi^{-1}(s)\star\varphi^t(s))(x)\\
\nonumber &=&
{d\over dt}\Big|_{t=0}( -\varphi^{-1}(s) {d\over ds}\exp\psi(s)\varphi^{-1}(s)\varphi^t(s)
+\varphi^{-1}(s) ({d\over ds}\exp\psi(s))^t)(x)\\
\nonumber&=&
{d\over dt}\Big|_{t=0}( -\varphi^{-1}(s)
d\exp\dot\psi(s)\varphi^{-1}(s)\varphi^t(s)
+\varphi^{-1}(s) (d\exp\dot\psi(s))^t)(x)\\
\nonumber &=&
( -\varphi^{-1}(s)
d\exp[\psi(s),M]\tilde{\beta}_{\varphi(s)}
+ {d\over dt}\Big|_{t=0}\varphi^{-1}(s)(d\exp[\psi(s),M])^t)(x).\\\end{aligned}$$ The last term in is $$\begin{aligned}
\label{e:beta2}
\nonumber
\lefteqn{\frac{d}{dt}\Big|_{t=0}\left(\varphi^{-1}(s)(d\exp[\psi(s),M])^t\right)(x)}\\
\nonumber &=&
{d\over
dt}\Big|_{t=0}\left(\varphi^{-1}(s)\varphi(s)^t
(\varphi(s)^t)^{-1}(d\exp[\psi(s),M])^t\right)(x)\\
\nonumber &=&
{d\over dt}\Big|_{t=0}\left(\varphi^{-1}(s)\varphi(s)^t\right)\
\left((\varphi(s)^t)^{-1}(d\exp[\psi(s),M])^t\right)\Big|_{t=0}(x)\\
\nonumber && +\left(\varphi^{-1}(s)\varphi(s)^t\right)\Big|_{t=0}\ \
{d\over dt}\Big|_{t=0}\left((\varphi(s)^{-1})^{t}(d\exp[\psi(s),M])^t\right)(x)\\
\nonumber &=&
\left(\tilde{\beta}_{
\varphi(s)}\star(\varphi(s)^{-1}d\exp[\psi(s),M])\right)(x)
+{d\over dt}\Big|_{t=0}\left((\varphi(s)^{-1}(d\exp[\psi(s),M]))^t\right)(x)\\
\nonumber &=&
\left(\tilde{\beta}_{
\varphi(s)}\star(\varphi(s)^{-1}d\exp[\psi(s),M])\right)(x)
+|x|\lambda\left(\varphi(s)^{-1}(d\exp[\psi(s),M])\right)(x)\\
\nonumber &=&
\left(\tilde{\beta}_{
\varphi(s)}\star(\varphi(s)^{-1}d\exp[\psi(s),M])\right)(x)
+\left(\lambda(\varphi(s)^{-1}(d\exp[\psi(s),M]))\circ Y\right)(x)\nonumber\end{aligned}$$ Substituting back into we get .
\[cor:beta\] Let $\varphi\in G^\Phi_\mathcal A$. Let $\dot\psi(s)=[M,\psi(s)]$ be the Lax pair from Theorem \[t:8.2\] with $\psi(0)=\psi = \log (\varphi)$. Let $\varphi(s)=\exp(\psi(s))$ and assume that $\varphi(s)\in
G^\Phi_{\mathcal A}$ for all $s$. Then $$\begin{aligned}
{d\over ds}
\beta_{\varphi(s)} &=&
\mathrm{Ad}(\varphi(s)_+(0))\left([\tilde{\beta}_{\varphi(s)},\varphi^{-1}(s)\star
d\exp[\log(\varphi(s)),M]]_+ \Big|_{\lambda=0}\right.\\
&&\qquad \left. +
\mathrm{Res}\;\left((\varphi^{-1}(s)\star
d\exp[\log(\varphi(s)),M])\circ Y\right)\right).\end{aligned}$$
This follows from the previous Theorem and Lemma \[lemma:6.6\].
In general, $\varphi\in G^\Phi_{\mathcal A}$ does not imply $\varphi(s)= \exp(\psi(s))\in G^\Phi_{\mathcal
A}$ for all $s$ (see Theorem \[t:h3\]). A simple example with $\varphi(s)\in
G^\Phi_{\mathcal A}$ is given by a holomorphic $\varphi$ (i.e $\varphi(x)\in\mathcal A_+$) with $-n+2m=0$. Indeed $(\varphi^t)_-=\varepsilon$ as $\varphi^t$ is holomorphic, so $(\varphi^t)_-$ does not depend on $t$. From the Taylor series of the exponential, $\exp(-s\log(\varphi))$ has only a holomorphic part so $g_-(s)=\varepsilon$. Therefore the solutions $\psi(s)$ of the Lax pair equation are constant, so $\varphi(s)=\varphi(0)\in
G^\Phi_{\mathcal A}$.
The Lax pair flow and the renormalization group flow {#s:rge}
====================================================
The Lax pair flow lives on the Lie algebra $\mathfrak g_{\mathcal A}$ of infinitesimal characters, while the beta character flow is on the Lie group $G_{\mathcal A}$ of characters. Theorem \[t:flow\] and Corollary \[cor:beta\] show that under the exponential map $\exp:{\mathfrak
g}_{\mathcal A}\to G_{\mathcal A}$, the corresponding flow of beta characters and $\beta$-functions are not in Lax pair form. The main point of this section is that the bijection $\tilde R^{-1}
:{\mathfrak g}_{\mathcal A}\to G_{\mathcal A}$ of [@man] is much better behaved: under $\tilde R^{-1}$, local characters remain local under the Lax pair flow (Theorem \[t:7.9\]), and the beta characters and the $\beta$-functions satisfy Lax pair equations (Theorems \[t:corr\]). In contrast, we give a rooted trees example of the nonlocality of the Lax pair flow of characters using the exponential map.
The pole order under the Lax pair flow
--------------------------------------
To begin, we investigate the dependence of the pole order of the Lax pair flow $L(t)$ on the pole order of the initial condition $L_0$ and the Casimir function (e.g. the functions $\psi_{m,n}$). In the rooted trees case, the computations are considerably simplified using the normal coordinates of [@chr], which we refer to for details.
Let $H$ be the Hopf algebra of rooted trees and $\mathcal{T}$ the set of trees. We choose a order on $\mathcal{T}=\{t_i\}_{i\in\mathbb{N}}$, such that $deg(t_i)\leq
deg(t_j)$ for any $i<j$ and such that $h(t_i)\geq h(t_j)$ for any trees $t_i,t_j$ with $deg(t_i)=deg(t_j)$ and $i<j$. Here $deg(t)$ is the number of vertices of $t$, and $h(t)$ is the height of the tree $t$, the length of the path from the root to the deepest node in the tree. For example, we can choose $$t_0=1_\mathcal{T}, \;\;\;\;
t_1= \ta1, \;\;\;\;
t_2= \tb2, \;\;\;\;
t_3= \tc3, \;\;\;\;
t_4=\td31, \;\;\;\;
t_5=\te4, \;\;\;\;
t_6=\tf41, \;\;\;\;
t_7=\thj44, \;\;\;\;
t_8=\th43$$
We recall that the $\exp:\mathfrak g_{\mathcal A}\to G_\mathcal{A}$ is bijective with inverse $\log:G_\mathcal{A}\to\mathfrak g_{\mathcal A}$ given by $$\log(\varphi)=\sum_{k=1}^\infty
(-1)^{k-1}\frac{(\varphi-\varepsilon)^{k}}{k}.$$ Set $f_0= 1_\mathcal{T}$ and let $\{f_i\}_{i\in\mathbb{N^*}}$ be the normal coordinates, i.e. $f_i$ is the forest in $H$ satisfying $$\log(\varphi)(t_i)=(\varphi-\varepsilon)(f_i),$$ for every character $\varphi$. For example, $$f_1=\ta1\; , \;\;\;\; f_2=\tb2-\frac{1}{2}\ta1\ta1 \; ,\;\;\;\;
f_3=\tc3-\ta1\tb2+\frac{1}{3}\ta1^3\; ,\;\;\;\;
f_4=\td31-\ta1\tb2+\frac{1}{6} \ta1^3\;.$$ $$f_5=\te4-\ta1\tc3-\frac{1}{2}\tb2^2+\ta1^2\tb2-\frac{1}{4}\ta1^4\; ,\;\;\;\;
f_8=\th43-\frac{3}{2}\ta1\td31+\frac{1}{2}\ta1^2\tb2
\;.$$
For a ladder tree $t$, the forest $f$ given by $\log(\varphi)(t )=(\varphi-\varepsilon)(f )$ for every character $\varphi$, is a primitive element of the Hopf algebra $H$.
We identify $\varphi\in G_\mathcal{A}$ with $\{\varphi(f_i)\}_{i\in\mathbb{N}}\in\mathcal{A}^\mathbb{N}$ and call the $\varphi(f_i)$ the $i$-component of $\varphi$. Since $\varphi(f_0)=1$ for all $\varphi$, we drop the $0$-component. We use Sweedler’s notation for the reduced coproduct $\tilde\Delta(x)=x'\otimes x''$, where $\tilde\Delta(x)=\Delta(x)-x\otimes 1_{\mathcal{T}}- 1_{\mathcal{T}}\otimes
x$. Notice that $\deg(x')+\deg(x'')=\deg(x)$ and $1\leq\deg(x'),\ \deg(x'')<\deg(x).$ For $x\neq 1_{\mathcal{T}}$ and $\tilde\Delta(x')=(x')'\otimes (x')''$, we have $$\begin{aligned}
((\varphi_1\varphi_2)\varphi_1^{-1})(x)
&=& \langle
(\varphi_1\varphi_2)\otimes\varphi_1^{-1},
x\otimes1_{\mathcal{T}}
+1_{\mathcal{T}}\otimes x
+ x'\otimes x''\rangle\\
&=& (\varphi_1\varphi_2)(x)+\varphi_1^{-1}(x)
+(\varphi_1\varphi_2)(x')\varphi_1^{-1}(x'')\\
&=& \varphi_1(x)+\varphi_2(x)
+\varphi_1(x')\varphi_2(x'')
+\varphi_1^{-1} (x)+\big(\varphi_1(x')+\varphi_2(x')\\
&&\qquad
+\varphi_1((x')')\varphi_2((x')'')\big)\varphi_1^{-1}(x'')\\
&=& \varphi_2(x)
+\varphi_1(x')\varphi_2(x'')
+ \big(\varphi_1(x)+\varphi_1^{-1}(x)
+\varphi_1(x')\varphi_1^{-1}(x'')\big)\\
&&\qquad
+\varphi_2(x' )\varphi_1^{-1}(x'')
+\varphi_1((x')')\varphi_2((x')'') \varphi_1^{-1}(x'')\\
&=& \varphi_2(x)
+\varphi_1(x')\varphi_2(x'')
+\varphi_2(x' )\varphi_1^{-1}(x'')
+\varphi_1((x')')\varphi_2((x')'') \varphi_1^{-1}(x'')\end{aligned}$$ Differentiating with respect to $\varphi_2$ and setting $L=\dot\varphi_2$ gives the adjoint representation: $$\begin{aligned}
\label{e:adjoint-repres}
\mathrm{Ad}(\varphi_1)(L)(x)&=&L(x)
+\varphi_1(x')L(x'')
+L(x' )\varphi_1(S x'')\\
&&\qquad +\varphi_1((x')')L((x')'') \varphi_1(Sx''),\nonumber\end{aligned}$$ where $S$ is the antipode of the Hopf algebra.
\[p:holo\] i) If the initial condition $L_0\in\mathfrak{g}_{\mathcal{A}}$ is holomorphic in $\lambda$, then the solution $L(t)={\Ad}(g_+(t))L_0$ of the Lax pair equation is holomorphic in $\lambda$.\
ii) If $L_0\in\mathfrak{g}_\mathcal{A}$ has a pole of order $n$, then $L(t)={\Ad}(g_+(t))L_0$ has a pole of order at most $n$.
By , we have $$\begin{aligned}
\label{e:adjoint-repres2}
\mathrm{Ad}(g_+(t))(L_0)(x)&=& L_0(x)
+g_+(t)(x')L_0(x'')
+L_0(x')g_+(t)(Sx'')\\
&&\qquad +g_+(t)((x')')L_0((x')'') g_+(t)(Sx'').\nonumber\end{aligned}$$ Notice that $g_+(t)(x)$ is holomorphic for $x\in H$. If $L_0$ is holomorphic, then every term of the right hand side of is holomorphic, so $\mathrm{Ad}(g_+(t))(L_0)$ is holomorphic. Since multiplication with a holomorphic series cannot increase the pole order, $L(t)$ cannot have a pole order greater than the pole order of $L_0$.
We can also use normal coordinates to measure the nontriviality of the Lax pair flow.
\[t:ladder\] If $f_i$ is a primitive element of $H$ (e.g. $f_i$ corresponds to a ladder tree), then the $i$-component $L(t)(f_i)$ of the Lax pair flow is does not depend on $t$.
It is shown in [@chr] that $\varphi^{-1}(f_i)=-\varphi(f_i),$ for every character $\varphi$ and $i\geq 1$. Since $f_i$ is a primitive element, the inner automorphism $C_g:G_{\mathcal{A}}\to G_{\mathcal{A}}$, $C_g(h)=ghg^{-1}$, satisfies $C_g(h)(f_i)=g(f_i)+h(f_i)-g(f_i)=h(f_i).$ Therefore $(Ad(g)L_0)(f_i)=(L_0)_i,$ where $(L_0)_i$ is the i-component of $L_0$.
Thus everything of interest in the Lax pair flow occurs off the primitives, e.g. the normal coordinate $f_4$ corresponding to $\td31$ is the first component in the Hopf algebra on which the Lax pair is nonconstant.
The Lax pair flow, the RGE flow, and locality
---------------------------------------------
We now investigate whether the Lax pair flow can ever be identified with the RGF. Some identification is necessary: the RGF $(\varphi^{t})_+(\lambda=0)$ lives in the Lie group of characters $G_{\mathbb C}$, while the Lax pair flow $L(t)$ lives in a Lie algebra $\mathfrak g_\mathcal{A}$ To match these flows, we can transfer the Lax pair flow to the Lie group level using either of the maps $\tilde
R^{-1}$ and $\exp$, namely by defining $$\label{phi-chi}
\varphi_t=\tilde R^{-1}(L(t))\ \ \ \text{and} \ \ \ \chi_t=\exp(L(t))$$ and then setting $\lambda =0.$
The most naive hope would be that $\varphi_t$ or $\chi_t$ coincide with the RGF $\varphi^t$, perhaps after a rescaling of the parameter $t$. We shall see that this fails even in the trivial case. As usual, we take $\mathcal A$ to be the algebra of Laurent series.
On a commutative, cocommutative, graded connected Hopf algebra $\mathcal H$, $\varphi_t\not=\varphi^t$ and $\chi_t\not=\varphi^t$.
If $\tilde R(\varphi^t)=L(t)$ then by Definition \[def:6.1\] $$(\varphi^t\circ Y)(\Gamma)=(\varphi^t\star L(t))(\Gamma)$$ for every $\Gamma\in \mathcal H$. For a primitive homogenous element $\Gamma\in {\mathcal H}_n$ we get $$\label{7.4a}
|\Gamma|e^{|\Gamma|t\lambda}\varphi(\Gamma)
=e^{|\Gamma|t\lambda}\varphi(\Gamma)L(t)(1)
+e^{0|\Gamma|\lambda}\varphi(\Gamma)L(t)(\Gamma).$$ Therefore $$L(t)(\Gamma)=|\Gamma|e^{|\Gamma|t\lambda}\varphi(\Gamma).$$ Since ${\mathcal H}$ is cocommutative, its Lie bracket is abelian. Thus the left hand side of (\[7.4a\]) is constant in $t$, while the right hand side is not.
The same argument works for $\chi_t$ on $\mathcal H$.
In a positive direction, we will show that locality of characters is preserved under the Lax pair flow, using the identification given by $\tilde R.$ This indicates that $\tilde R$ is more useful than the exponential map.
Recall from [@man Theorem IV.4.1] that $\lambda\tilde R:G_\mathcal{A}\to\mathfrak{g}_{\mathcal{A}}$ restricts to a bijection from $G^\Phi_\mathcal{A}$ to $\mathfrak{g}_{\mathcal{A}+}$, where $\mathfrak{g}_{\mathcal{A}+}$ is the set infinitesimal characters on $\mathcal H$ with values in $\mathcal
A_+$. In this sense, $\lambda \tilde R$ is better behaved than $\tilde R$, as the following locality result shows.
\[t:7.9tau\] For a local character $\varphi\in G^\Phi_\mathcal{A}$, let $L(t)$ be the solution of the Lax pair equation (\[8:1\]) with initial condition $L_0=\lambda\tilde R(\varphi)$ and any $\mathrm{Ad}$-covariant function $f$. Let $\tau_t$ be the flow of characters given by $$\tau_t=(\lambda\tilde R)^{-1}(L(t)).$$ Then $\tau_t$ is a local character for all $t$.
By [@ef-manchon], $$\tilde R(\varphi\star\xi)=\tilde R(\xi)+\xi^{*-1}\star\tilde R(\varphi)\star
\xi.$$ Taking $\xi=g_+(t)^{-1}$ and multiplying by $\lambda$, we get $$\lambda\tilde R(\varphi\star g_+(t)^{-1})=\lambda\tilde R(g_+(t)^{-1})+
g_+(t)\star\lambda\tilde R(\varphi)\star
g_+(t)^{-1}.$$ Since $\varphi\in G^\Phi_\mathcal{A}$ and $g(t)_+^{-1}$ is an element in $G_\mathcal{A}$ without polar part, by [@man Lemma IV.4.3.], $\varphi\star g(t)_+^{*-1}$ is local. Thus $\lambda\tilde R(\varphi\star g_+(t)^{-1})\in{\mathcal{A}_+}$. We have $g(t)_+^{-1}\in G^\Phi_\mathcal{A}$, simply because $g(t)_+^{-1}$ does not have a polar part, so $\lambda\tilde R (g(t)_+^{-1})$ is holomorphic. It follows that $$\tau_t=(\lambda\tilde R)^{-1}({\Ad}(g_+(t))L_0)=(\lambda\tilde R)^{-1}(
g_+(t)\star L_0\star g_+(t)^{-1})\in G^\Phi_\mathcal{A}$$
We can now show that locality of characters is preserved under the Lax pair flow via the $\tilde R$ identification.
\[t:7.9\] For a local character $\varphi\in G^\Phi_\mathcal{A}$, let $L(t)$ be the solution of the Lax pair equation for any $\mathrm{Ad}$-covariant function $f$, with the initial condition $L_0=\tilde R(\varphi)$. Let $\varphi_t$ be the flow given by $$\varphi_t=\tilde R^{-1}(L(t)).$$ Then $\varphi_t$ is a local character for all $t$.
We show that the flow $\tau_t$ constructed in the previous Proposition with the initial condition $L_0=\lambda\tilde
R(\varphi)$, for the $\mathrm{Ad}$-covariant function $h:\mathfrak{g}_\mathcal{A}\to\mathfrak{g}_\mathcal{A}$ given by $h(L)=f(\lambda^{-1} L)$, is equal to the flow $\varphi_t$ constructed for the $\mathrm{Ad}$-covariant function $f$, with the initial condition $L_0=\tilde
R(\varphi)$. We have $$\tau_t=(\lambda\tilde R)^{-1}\left(g_+(t)\star\lambda\tilde
R(\varphi)\star g_+^{-1}(t)\right)=\tilde R^{-1}\left(g_+(t)\star \tilde
R(\varphi)\star g_+^{-1}(t)\right),$$ where $g_+(t)$ is given by the Birkhoff decomposition of $$\exp(-t f( \tilde R(\varphi)))= \exp(-t h(\lambda\tilde
R(\varphi))).$$ Therefore the two $g_+(t)$ involved in the definitions of $\varphi_t$ and $\tau_t$ coincide, so $\varphi_t=\tau_t$.
In contrast to Theorem \[t:flow\], it is immediate that the flow of beta characters associated to $\tilde R$ is in Lax pair form.
\[l:rminusoneversion\] For a local character $\varphi\in G^\Phi_\mathcal{A}$, let $\varphi_t$ be the flow from Theorem 7.5. Then $$\begin{aligned}
\label{e:rminusoneversion}
\frac{d\tilde\beta_{\varphi_t}}{dt}=[\tilde\beta_{\varphi_t},M],
$$
By Lemma 6.7, we get $\tilde\beta_{\varphi_t}=\lambda\tilde R(\varphi_t)=\lambda L(t)$. Then $$\frac{d\tilde\beta_{\varphi_t}}{dt}= \frac{d(\lambda\tilde
R(\varphi_t))}{dt}=
\frac{\lambda d
L(t)}{dt}=\lambda[L(t),M]=[\tilde\beta_{\varphi_t},M].$$
The corresponding $\beta$-functions also satisfy a Lax pair equation.
\[t:corr\] For a local character $\varphi\in G^\Phi_\mathcal{A}$, let $L(t)$ be the Lax pair flow of Corollary 5.10 with initial condition $L_0=\tilde R(\varphi)$. Let $\varphi_t=\tilde R^{-1}(L(t))$. Then
- for $-n+2m\geq 1$, $\varphi_t = \varphi$ and hence $\beta_{\varphi_t}=\beta_{\varphi}$ for all $t$.
- for $-n+2m\leq 0$, $\beta_{\varphi_t}\in
\mathfrak{g}_{\mathbb C}$ satisfies $$\frac{d\beta_{\varphi_t}}{dt}=\big[\beta_{\varphi_t},
-\frac{d((\varphi_t)_+(0))}{dt}((\varphi_t)_+(0))^{-1} +
2\mathrm{Ad}((\varphi_t)_+(0))\big(\mathrm{Res}(\lambda^{-n+2m-2}\tilde\beta_{\varphi_t})\big)
\big].$$
By Theorem 7.5, $\varphi_t$ are local characters, so by [@man Theorem IV.4.], $\tilde\beta_{\varphi_t}=\lambda
L(t)=\lambda \tilde R (\varphi_t)$ is holomorphic.
\(i) If $-n+2m\geq 1$, then $\lambda^{-n+2m}L(t)$ is holomorphic, which implies $$M=R(\lambda^{-n+2m}L(t))=\lambda^{-n+2m}L(t)=
\lambda^{-n+2m-1}\tilde\beta_{\varphi_t}.$$ $L(t)$ satisfies the Lax pair equation $$\frac{dL}{dt}=[L,M]=[L,\lambda^{-n+2m}L]=\lambda^{-n+2m}[L,L]=0.$$ Thus $L(t)=L_0$ for all $t$, which gives $\varphi_t=\tilde
R^{-1}(L(t))=\tilde R^{-1}(L_0)=\varphi$ for all $t$.
\(ii) For $-n+2m\leq 0$, we have $$\begin{aligned}
M &=&
R(\lambda^{-n+2m}L(t))=\lambda^{-n+2m}L(t)-2P_-(\lambda^{-n+2m}L(t))\\
&=&
\lambda^{-n+2m-1}\tilde\beta_{\varphi_t}-2P_-(\lambda^{-n+2m-1}\tilde\beta_{\varphi_t}).\end{aligned}$$ (\[e:rminusoneversion\]) becomes $$\frac{d\tilde\beta_{\varphi_t}}{dt}=
-2[\tilde\beta_{\varphi_t},P_-(\lambda^{-n+2m-1}\tilde\beta_{\varphi_t})]
=-2[P_+(\lambda^{-n+2m-1}\tilde\beta_{\varphi_t}),\lambda^{n-2m+1}
P_-(\lambda^{-n+2m-1}\tilde\beta_{\varphi_t})]
.$$ Expand $\tilde\beta_{\varphi_t}$ as $$\tilde\beta_{\varphi_t}=\sum_{k=0}^{\infty} \tilde\beta_k(t)\lambda^k.$$ Then $$\frac{d\tilde\beta_{\varphi_t}}{dt}=-2\left[\sum_{k=n-2m+1}^\infty
\tilde\beta_{k}(t)\lambda^{k-n+2m-1},\sum_{j=0}^{n-2m}\tilde\beta_j(t)\lambda^j\right],$$ and evaluating at $\lambda=0$ gives $$\begin{aligned}
\frac{d\tilde\beta_0(t)}{dt}=-2[\tilde\beta_{n-2m+1}(t),\tilde\beta_0(t)]
=2[\tilde\beta_0(t),\tilde\beta_{n-2m+1}(t)].\end{aligned}$$ Using the facts that Ad$(g)$ is a Lie algebra homomorphism and $(d/dt){\rm Ad}(g(t))X$\
$ = [(dg(t)/dt)g^{-1}(t), {\rm Ad}(g(t))X]$ (see the proof of Theorem \[t:7.9tau\]), we get $$\begin{aligned}
\frac{d\beta_{\varphi_t}}{dt}&=&
\frac{d}{dt}\left(\mathrm{Ad}(\varphi_t)( \tilde\beta_0(t)) \right)\nonumber\\
& =&
\left[\frac{d((\varphi_t)_+(0))}{dt}((\varphi_t)_+(0))^{-1},
\mathrm{Ad}(\varphi_t)(\tilde\beta_0(t))\right]
+
\mathrm{Ad}((\varphi_t)_+(0))\left(\frac{d\tilde\beta_0(t)}{dt}\right)\\
&=&
\left[\frac{d((\varphi_t)_+(0))}{dt}((\varphi_t)_+(0))^{-1},\beta_{\varphi_t}\right]
+2\left[\mathrm{Ad}((\varphi_t)_+(0))\tilde\beta_0(t),
\mathrm{Ad}((\varphi_t)_+(0))\tilde\beta_{n-2m+1}(t)\right]\nonumber\\
&=&\left[\beta_{\varphi_t},-\frac{d((\varphi_t)_+(0))}{dt}((\varphi_t)_+(0))^{-1
}+2\mathrm{Ad}((\varphi_t)_+(0))\left(\mathrm{Res}(\lambda^{-n+2m-2}\tilde\beta_{\varphi_t})
\right)\right],\nonumber \end{aligned}$$ since $\tilde\beta_{n-2m+1}(t)=\mathrm{Res}(\lambda^{-n+2m-2}\tilde\beta_{\varphi_t})$.
Local characters satisfy the abstract Renormalized Group Equation [@0609035], which we now recall. For a local character $\varphi\in G^\Phi_{\mathcal A}$, the renormalized character is defined by $\varphi_{\mathrm{ren}}(t)=(\varphi^t)_+(\lambda=0)$.
\[aRGE\] For $\varphi\in G^\Phi_{\mathcal A}$, the renormalized character $\varphi_{\mathrm{ren}}$ satisfies the abstract Renormalized Group Equation: $$\frac{\partial}{\partial t}
\varphi_{\mathrm{ren}}(t)=\beta_\varphi\star\varphi_{\mathrm{ren}}(t).$$
Here our parameter $t$ corresponds to $e^t$ in [@0609035].
In light of Theorem \[t:7.9\], we can ask for the relation between $(\varphi_t)_{\mathrm{ren}}(s)$ and $\varphi_{\mathrm{ren}}(s)$ corresponding to $\varphi_t$ and $\varphi$. In §9, we consider a toy model character on a Hopf algebra of rooted trees and show that these renormalized characters differ.
We can also show that for certain initial conditions, the flow $\tau_t$ is constant.
If $\varphi\in G^\Phi_\mathcal{A}$ and $\varphi_+=\varepsilon$ (i.e. $\varphi$ has only a pole part), then the flow $\tau_t$ of Proposition \[t:7.9tau\] for the $\mathrm{Ad}$-covariant function $f(L)=\lambda^{-n+2m}L$ has $\tau_t=\varphi$ for all $t$.
If we show that either $g_\pm(t)=\varepsilon$, then $$\tau_t=(\lambda\tilde R)^{-1}\big(g_\pm(t)\star\lambda\tilde R
(\varphi)\star g_\pm(t)^{-1}\big)=(\lambda\tilde R)^{-1}\big(\varepsilon\star\lambda\tilde R
(\varphi)\star\varepsilon^{-1}\big)=\varphi.$$ $g_\pm(t)$ are given by the Birkhoff decomposition of $$\begin{aligned}
\label{e:expforX}
g(t)=\exp(-2t\lambda^{-n+2m}L_0)=\sum_{k=0}^\infty
\frac{(-2t\lambda^{-n+2m})^k L_0^{k}}{k!},\end{aligned}$$ where $L_0=\lambda\tilde R(\varphi)\in\mathfrak{g}_{\mathbb C}$ [@man Theorem IV.4.4]. We split the problem into two cases depending on the sign of $-m+2n$. If $-m+2n\geq 0$, then $g(t)(x)\in\mathcal A_+$ for any $x$, which implies $g_-(t)=\varepsilon$. Similarly, if $-m+2n< 0$, then $g(t)(x)\in\mathcal A_-$ for any $x$, which implies $g_+(t)=\varepsilon$. Notice that the right hand side of (\[e:expforX\]) is a finite sum, namely up to $k=\mathrm{deg}(x)$ when evaluated on $x\in \mathcal H$.
We next study the locality of the flow $\chi_t$ defined in (\[phi-chi\]) for the usual Lax pair flow $L(t).$ . Thus for an initial character $\varphi$ and $L_0 = \log(\varphi)$, $$\chi_t =\exp(g_+(t)\star L_0\star g_+(t)^{-1})=g_+(t)\star
\exp(L_0)\star g_+(t)^{-1}=g_+(t)\star \varphi\star g_+(t)^{-1},$$ with $\chi_0=\varphi$. As before, in normal coordinates $\chi_t$ is trivial on primitives.
\[l:constant\] If $\varphi\in G^\Phi_{\mathcal A} $ and $f_i$ is a primitive element, then $ \chi_t(f_i)$ does not depend on $t$.
$\chi_t(f_i)=g_+(t)(f_i)+\varphi(f_i)-g_+(t)(f_i)=\varphi(f_i)$.
We now present some calculations showing the interplay between the Lax pair flow and locality.
For the first example, we construct a nontrivial Hopf subalgebra on which $\chi_t$ is local. Let $\mathcal H^{2}$ be the Hopf subalgebra generated by the following trees $$t_0=1_\mathcal{T}, \;\;\;\;
t_1= \ta1, \;\;\;\;
t_2= \tb2, \;\;\;\;
t_4=\td31,$$ together with any set of ladder trees. The first normal coordinate of $\chi_t$ to depend on $t$ is $f_4$, corresponding to $\td31$. Let $G^2_\mathcal{A}$ be the group of characters associated to the $\mathcal H^2$.
\[t:7.9a\] For $\varphi\in G^{2\, \Phi}_{\mathcal A}$, let $\chi_t$ be the flow of characters on $\mathcal H^2$ given by $$\chi_t=\exp(L(t)),$$ where $L(t)$ is the solution of the Lax pair equation (\[8:1\]) for any $\mathrm{Ad}$-covariant function with the initial condition $L_0=\log(\varphi)$. Then $\chi_t$ is local for all $t$.
Let $\pi$ denote the projection to the pole part of a Laurent series. By [@chr], $\tilde\Delta(f_4)=f_1\otimes
f_2-f_2\otimes f_1$, so $$(\chi_t^s)_-(f_4)=-\pi\big(\chi_t^s(f_4)+(\chi_t^s)_-(f_1)\chi_t^s(f_2)
-(\chi_t^s)_-(f_2)\chi_t^s(f_1)\big).$$ Subtracting from this equation the corresponding equation for $t=0$, and remembering that $\chi_t(f_1)$ and $\chi_t(f_2)$ do not depend on $t$ by Lemma \[l:constant\], we get $$(\chi_t^s)_-(f_4)=(\varphi^s)_-(f_4)-\pi(\chi_t^s(f_4)-\varphi^s(f_4)).$$ We have $$\begin{aligned}
\label{e:pole}
\pi(\chi_t^s(f_4)-\chi^s(f_4))
&=& \pi
\big(
e^{3\lambda s}
(
-2g_+(t)(f_2)\varphi(f_1)
+2g_+(t)(f_1)\varphi(f_2)
)
\big)\\
&=& \pi
\big(
- 2g_+(t)(f_2)\pi(e^{3\lambda s}\varphi(f_1))
+ 2g_+(t)(f_1)\pi(e^{3\lambda s}\varphi(f_2))
\big).\end{aligned}$$ Since $\varphi\in G^{2\, \Phi}_\mathcal{A}$, both $\pi(e^{s\lambda}\varphi(f_1))=-(\varphi^s)_-(f_1)$ and $\pi(e^{2s\lambda}\varphi(f_2))=-(\varphi^s)_-(f_2)$ are independent of $s$. By rescaling $s$, $\pi(e^{3s\lambda}\varphi(f_1))$ and $\pi(e^{3s\lambda}\varphi(f_2))$ are independent of $s$. Therefore $\pi(\chi_t^s(f_4)-\varphi^s(f_4))$ is independent of $s$, which finishes the proof.
We can apply the previous proposition to the Hopf subalgebra of Feynman diagram generated by the empty graph and the graphs $$\begin{aligned}
A_1= \p\ ,\ A_2= \pdp\ ,\ A_3= \pdpdp\ ,\
A_4= \pddpp\ ,\ A_5= \pdpdpdp\end{aligned}$$ and with $\varphi$ the Feynman rules character. The characters $\chi_t$ restricted to this Hopf algebra are all local.
To investigate how $\chi_t$ fails to be local on a non-ladder tree with a larger number of vertices, we consider the Hopf subalgebra $\mathcal H^3$ generated by $$t_0=1_\mathcal{T}, \;\;\;\;
t_1= \ta1, \;\;\;\;
t_2= \tb2, \;\;\;\;
t_3=\tc3, \;\;\;\;
t_4=\td31, \;\;\;\;
t_6=\tf41, \;\;\;\;
t_7=\thj44, \;\;\;\;
t_8=\th43$$ together with any set of ladder trees. Let $f_i$ be the corresponding normal coordinates.
The next lemma gives the pole order of a local character $\varphi\in G^\Phi_{\mathcal A}$ on primitives.
\[lemma:7.2\] If $\varphi$ is a local character and $f_i$ is primitive, then both $\varphi(f_i)$ and $L_0(f_i)=\log(\varphi)(f_i)$ have a pole of order at most one.
If ${\rm deg}(f_i)=d$ and $\varphi=\sum_{k=-n}^\infty \varphi_k\lambda^k$ is the Laurent expansion of $\varphi$, then $$(\varphi^s)_-(f_i)=-\pi(e^{sd\lambda}\varphi(f_i))=
-\pi(
\varphi_{-n}(f_i)\lambda^{-n}+(\varphi_{-n}(f_i)sd+\varphi_{-n+1}(f_i))\lambda^{-n+1}
+o(\lambda^{-n+2}))$$ If $\varphi$ has a pole, then $\varphi^s_- = \varphi_-$ implies $-n+1=0$.
\[t:h3\] Let $\varphi$ be a local character on $\mathcal H^3$ and let $\chi_t$ be the flow of characters given by $$\chi_t=\exp(L(t)),$$ where $L(t)$ is the solution of the Lax pair equation (\[8:1\]) with the initial condition $L_0=\log(\varphi)$. Then $\chi_t$ is local on $\mathcal H^3$ for all $t$ if and only if either $$\begin{aligned}
\label{e:condition:chi}
\varphi_-(f_1)=0\ \ \text{or} \ \
3\varphi_-(f_1)(\varphi_+(f_2)\big|_{\lambda=0})=\varphi_-(f_2)(\varphi_+(f_1)\big|_{\lambda=0}).\end{aligned}$$
The point is that (\[e:condition:chi\]) is unlikely to hold.
By [@chr], $$\tilde\Delta(f_8)=\frac{3}{2}f_1\otimes f_4-\frac{3}{2}f_4\otimes f_1
-\frac{1}{2}f_1\otimes f_1 f_2-\frac{1}{2}f_1f_2\otimes f_1
+\frac{1}{2}f_1f_1\otimes f_2+\frac{1}{2}f_2\otimes f_1f_1.$$
Since $(\chi^s_t)_-(f_4)$ does not depend on $s$ (Prop. \[t:7.9a\]) and since $\chi_t(f_1)$ and $\chi_t(f_2)$ do not depend on $t$, after cancelations of terms involving only the primitives $f_1$ and $f_2$, we get $$\begin{aligned}
\label{e:tf8}
\lefteqn{
(\chi^s_t)_-(f_8)-(\varphi^s)_-(f_8)}\\
&=&
-\pi
\Big(\chi^s_t(f_8)-\varphi^s(f_8)\\
&&\qquad +\frac{3}{2}\varphi_-(f_1)(\chi^s_t(f_4)-\varphi^s(f_4))
-\frac{3}{2}\varphi^s(f_1)((\chi_t)_-(f_4)-\varphi_-(f_4))
\Big)\nonumber\end{aligned}$$
By Lemma \[lemma:7.2\], $\varphi(f_1)$ and $\varphi(f_2)$ have poles of order at most one. Set $$\varphi(f_1)=\sum_{k=-1}^\infty a_k\lambda^k\ \ \ \text{and}\ \ \
\varphi(f_2)=\sum_{k=-1}^\infty b_k\lambda^k.$$
From the proof of Proposition \[t:7.9a\], $$\begin{aligned}
\label{e:chi1}
\chi_t^s(f_4)-\varphi^s(f_4))
=
e^{3\lambda s}
(
-2g_\pm(t)(f_2)\varphi(f_1)
+2g_\pm(t)(f_1)\varphi(f_2)
)\end{aligned}$$ We have $g_-(t)(f_1)=\exp(-2t\lambda^{-n+2m}L_0)_-(f_1)=-\pi(-2t\varphi(f_1))=2ta_{-1}\lambda^{-1}$, and\
$g_+(t)(f_1)=-2t(a_0+a_1\lambda+o(\lambda^2))$. Similarly $g_-(t)(f_2)=2tb_{-1}\lambda^{-1}$. (\[e:chi1\]) becomes $\chi_t^s(f_4)-\varphi^s(f_4))=4te^{3\lambda
s}\big((-b_{-1}a_0+a_{-1}b_0)\lambda^{-1}+o(\lambda^0)\big)$, which implies $$\begin{aligned}
\lefteqn{
\pi\big(\frac{3}{2}\varphi_-(f_1)(\chi^s_t(f_4)-\varphi^s(f_4))
-\frac{3}{2}\varphi^s(f_1)((\chi_t)_-(f_4)-\varphi_-(f_4))\big)}\\
&=&
\frac{3}{2}\pi\big((-a_{-1}\lambda^{-1}e^{3s\lambda}(4t((-b_{-1}a_0+a_{-1}b_0)
\lambda^{-1}+o(\lambda^0)))\big)\\
&&\qquad
-\frac{3}{2}\pi\big(e^{s\lambda}(a_{-1}\lambda^{-1}+o(\lambda^0))(4t)(-1)(-b_{-1}a_0
+a_{-1}b_0)\lambda^{-1}\big)\\
&=& -12sta_{-1}(-b_{-1}a_0+a_{-1}b_0)\lambda^{-1}+P(\lambda^{-1}),\end{aligned}$$ where $P(\lambda^{-1})$ is some polynomial in $\lambda^{-1}$ independent of $s$. We get $$\begin{aligned}
\label{e:tf8s}
\lefteqn{ \pi
\Big(\chi^s_t(f_8)-\varphi^s(f_8)
\Big)}\\
& =&
\pi
\big(
e^{4s\lambda}
(
( -3g_+(t)(f_1)g_+(t)(f_2)-3g_+(t)(f_3)
)\varphi(f_1)
)
\big)\\
&&\qquad +3\pi
\big(e^{4s\lambda}
( g_+(t)(f_1) )^2\varphi(f_2)
\big)
+\pi
\big(3e^{4s\lambda}
g_+(t)(f_1)\varphi(f_4)
\big)\nonumber\end{aligned}$$ Since $\varphi(f_1)$ and $\varphi(f_2)$ have poles of order at most one, the first two terms of the right hand side of (\[e:tf8s\]) do not depend on $s$. Since $\varphi\in G^{3\, \Phi}_\mathcal{A}$, $$\varphi^s_-(f_4)=-\pi\big(e^{3s\lambda}\varphi(f_4)+\varphi_-(f_1)e^{2s\lambda}\varphi(f_2)
-\varphi_-(f_2)e^{s\lambda}\varphi(f_1))\big)$$ is independent of $s$, and so $\pi(e^{3s\lambda}\varphi(f_4)-a_{-1}b_{-1}s\lambda^{-1})$ is also independent of $s$. By rescaling $s$, $\pi(3e^{4s\lambda}\varphi(f_4)-4a_{-1}b_{-1}s\lambda^{-1})$ does not depend on $s$. In conclusion, the terms independent of $s$ in $-(\chi^s_t)_-(f_8)+(\varphi^s)_-(f_8)$ are $$-12sta_{-1}(-b_{-1}a_0+a_{-1}b_0)\lambda^{-1} +
4a_{-1}b_{-1}s\lambda^{-1}(-2ta_0).$$ Therefore $\chi^s_t(f_8)$ is independent of $s$ if and only if either $a_{-1}=0$ or $3a_{-1}b_0=a_0b_{-1}$. Similar computations hold for the normal coordinates $f_6$ and $f_7$.
The choice $-n+2m=0$ in the Proposition is just for the sake of concreteness. A more detailed analysis reveals the following:
- If $-n+2m\geq 1$ then $\chi_t$ is local on $\mathcal H^3$ without any additional conditions. Indeed, in this case $L_0(f_1)$ and $L_0(f_2)$ are holomorphic and thus $g_-(t)(f_1)=g_-(f_2)=0$. which implies that $\chi^s_t(f_4)-L_0(f_4)=0$ for every $t$. By (\[e:tf8\]) we get that $(\chi_t^s)_-(f_8)$ does not depend on $s$. Similar statements hold for $f_6$ and $f_7$.
- If $-n+2m=-1$, the situation is similar to Proposition \[t:h3\], namely $\chi_t$ is local on $\mathcal H^3$ if and only if $$\varphi_-(f_1)(\varphi_+(f_2)\big|_{\lambda=0})=\varphi_-(f_2)(\varphi_+(f_1)\big|_{\lambda=0})$$ and either $$\varphi_-(f_1)=0\ \ \text{or} \ \
3\varphi_-(f_1)(\frac{\partial\varphi_+(f_2)}{\partial\lambda}\big|_{\lambda=0})
=\varphi_-(f_2)(\frac{\partial\varphi_+(f_1)}{\partial\lambda}\big|_{\lambda=0}).$$
For $-n+2m\in\Z^+$, the flows $\chi_t$ and $L(t)$ gain locality, in the sense that they become constant on larger Hopf subalgebras as $-n+2m$ increases. Indeed, $\chi_t$ and $L(t)$ are constant on the Hopf algebra generated by the primitives (e.g. the normal coordinates associated to ladder trees. In contrast, if we decrease $-n+2m<0$, we preserve locality only when an increasing number of conditions are fulfilled.
The Lax pair flow of the $\beta$-function
-----------------------------------------
Recall from §6 that the beta characters for the exponentiated Lax pair flow $\exp(L(t))$ do not themselves satisfy a Lax pair equation. In the next theorem, we reverse this procedure by taking a Lax pair flow $L(t)$ starting at the $\beta$-function of a character, and then producing characters $\xi_t$ whose $\beta$-functions are $L(t)\big|_{\lambda = 0}.$
\[t:7.14\] Let $\varphi\in G^\Phi_{\mathcal A}$ and let $L(t)$ be the flow given by Theorem \[t:8.2\] with the initial condition $L_0=\beta_\varphi.$ Let $\xi_t=(\lambda \tilde{R})^{-1}(L(t)\big|_{\lambda=0})$. Then $\xi_t$ is local for all $t$. The $\beta$-function of $\xi_t$ satisfies $$\beta_{\xi_t} = L(t)\big|_{\lambda =0}.$$ Moreover, $\xi(0)= \varphi_-$.
Since $L_0=\beta_\varphi$ is scalar valued, by Theorem \[p:holo\] $L(t)=\mathrm{Ad}(g_+(t))(L_0)$ is holomorphic. Therefore. $L(t)$ can be evaluated at $\lambda=0$. $L(t)\big|_{\lambda=0}$ is also scalar valued, so by [@man Theorem IV.4.4], $\xi_t\in G^\Phi_{\mathcal A-}$, the set of local characters taking values in $\lambda^{-1}\mathbb C[\lambda^{-1}]$. In particular, $\xi_t$ is local. By Lemma \[lemma:6.6\], $\tilde
\beta_{\xi_t} = (\lambda \tilde R)(\xi_t) = L_t\big|_{\lambda = 0}$, so $\tilde\beta_{\xi_t}$ must be constant in $\lambda.$ This implies $$\beta_{\xi_t} = \tilde\beta_{\xi_t}\big|_{\lambda =0} = L(t)\big|_{\lambda =0}.$$ It follows from the Connes-Kreimer scattering formula [@ck2; @man] that $\xi(0)=\varphi_-$.
Worked examples {#worked-example}
================
In this section we give some explicit computations on two Hopf algebras which illustrate results in previous sections. We first consider the Hopf algebra $\mathcal H^1$ generated by the following trees: $$t_0=1_\mathcal{T}, \;\;\;\;
t_1= \ta1, \;\;\;\;
t_2= \tb2, \;\;\;\;
t_4=\td31, \;\;\;\;
t_8=\th43,$$ and the regularized toy model character $\varphi=\varphi(q,\mu,\lambda)$ (see [@krei99; @krde99]) given on trees by $$\varphi(T)(q,\mu,\lambda)=(q/\mu)^{-\lambda\mathrm{deg(T)}}\prod_{v}
B_{w(T_v)}(\lambda).$$ Here the product is taken over all vertices $v$ of the tree $T$, $w(T_v)$ is the number of vertices of the subtree $T_v$ of $T$ which has $v$ as a root, and $B_j(\lambda)=B(j\lambda,1-j\lambda)$ for $j\in\mathbb N^*$, with $B$ the Euler beta function. Referring to [@ef-manchon], $q$ is interpreted as a dimensional external parameter, and $\mu$ is the ’t Hooft mass. $\varphi$ has enough similarity of realistic QFT calculations to be worth considering [@bk99; @krde99; @krei99; @krei00]. Set $b=q/\mu$ and $a=\log(b)$. Thus terms in $a$ (or $\log(q^2/\mu^2) = 2a$ as in e.g. [@ky]) are the leading log terms in the various expansions. We have $\varphi(\ta1)=b^{-\lambda}B_1(\lambda)$, $\varphi(\tb2)=b^{-2\lambda}B_2(\lambda)B_1(\lambda)$, $\varphi(\td31)=b^{-3\lambda}B_3(\lambda)B_1(\lambda)^2$, $\varphi(\th43)=b^{-4\lambda}B_4(\lambda)B_1(\lambda)^3$, etc. In the normal coordinates $f_i$, the Laurent series of $\varphi(f_i)$ (cf. [@chr]) are given by $$\varphi(f_1)=\frac{1}{\lambda}-a+o(\lambda),\ \ \ \ \
\varphi(f_2)=\frac{\pi^2}{4}+o(\lambda),
$$ $$\varphi(f_4)=\frac{7\pi^2}{36 \lambda} -\frac{7\pi^2a}{12} +
o(\lambda), \ \ \
\varphi(f_8)=\frac{\pi^2}{12 \lambda^2}
-\frac{\pi^2a}{3\lambda} + o(\lambda^0).$$ The character $\varphi$ is local, and the Lax pair flow on $\mathcal H^1$ $\varphi_t=\tilde R^{-1}(L(t))$ as in Corollary \[tc:8.2\], with $L(t)=Ad(g_\pm(t))L_0$, $-m+2n=0$, and initial condition $L_0=\tilde R(\varphi)$ is given by $$\varphi_t(f_1)=\varphi(f_1),\ \ \ \ \varphi_t(f_2)=\varphi(f_2),\ \
\ \varphi_t(f_4)=\frac{\pi^2 (7 + 24 t)}{36 \lambda}
-\frac{\pi^2}{12} (7 + 16 t)a + o(\lambda),$$ $$\begin{aligned}
\varphi_t(f_8)&=&\frac{\pi^2 (1 + 6 t)}{12 \lambda^2} - \frac{\pi^2(2 + 15 t + 9
t^2) a}{6 \lambda}\\
&&\qquad + \frac{\pi^2}{144} (\pi^2 (83 + 288 t + 126 t^2)
+ 12 (8 + 51 t + 27 t^2) a^2)+ o(\lambda).\end{aligned}$$ The flow $L(t)$ has poles of order at most one: $$L(t)(f_1)=\frac{1}{\lambda}-a+o(\lambda),\ \ \
L(t)(f_2)=\frac{\pi^2}{2}+o(\lambda),$$ $$L(t)(f_4)=
\frac{\frac{\pi^2}{3} + 2 \pi^2 t}
{\lambda}
+ (-\pi^2 - 4 \pi^2 t) a +o(\lambda),$$ $$L(t)(f_8)=\frac{-2 (\pi^2 t (2 + 3 t)a)}{\lambda}
+\frac{\pi^2}{6}
(\pi^2 (7 + 37 t + 21 t^2) + 6 t (8 + 9 t) a^2)+o(\lambda).$$ This confirms that $\lambda L(t)=(\lambda\tilde
R)(\varphi_t)=\tilde\beta_{\varphi_t}$ is holomorphic, which implies that $\varphi_t$ is local on $\mathcal H^1$. It can be explicitly checked that $(\varphi_t^s)_-$ does not depend on $s$: $$(\varphi_t^s)_-(f_1)=-\frac{1}{z}, \ \ \ \ \ \
(\varphi_t^s)_-(f_2)=0, \ \ \ \ \ \ \
(\varphi_t^s)_-(f_4)=\frac{\pi^2(1-12t)}{18\lambda},$$ $$(\varphi_t^s)_-(f_8)=-\frac{\pi^2(1-12t)}{24 \lambda^2}+\frac{3\pi^2t(1+t)a}{2\lambda}.$$ The Connes-Kreimer $\beta$-functions $\beta_{\varphi_t}=-{\rm Res}(\varphi_t)_-\circ Y$ are: $$\beta_{\varphi_t}(f_1)=1,\ \ \ \beta_{\varphi_t}(f_2)=0,\ \ \
\beta_{\varphi_t}(f_4)=-\frac{\pi^2}{6} + 2 \pi^2 t, \ \ \ \
\beta_{\varphi_t}(f_8)=-6 \pi^2 t (1 + t)a.$$ The associated RGFs $(\varphi_t)_{\mathrm{ren}}(s)=(\varphi_t^s)_+\big|_{\lambda=0}$, which all satisfy the abstract RGE, are $$(\varphi_t)_{ren}(s)(f_1)=(s - a), \ \ \ \ \ (\varphi_t)_{ren}(s)(f_2)=\frac{\pi^2}{4},$$ $$(\varphi_t)_{ren}(s)(f_4)=
\frac{\pi^2}{12}(s + 24 s t - (1 + 16 t)a),$$ $$(\varphi_t)_{ren}(s)(f_8)=\frac{\pi^2}{96}
(12 s^2 + \pi^2 (11 + 136 t + 84 t^2) - 24 s (1 + 20 t + 24 t^2)a + 12 (1 + 22 t + 18 t^2)
a^2).$$
Thus in this model there is a polynomial dependence in $t$ of the leading log terms in each of $L(t), \beta_{\varphi_t}$ and $(\varphi_t)_{\rm ren}$, although the first diagram with a nonzero leading log term differs. By the recursion formula in [@ky (26)], the next to ... leading log terms in the Green’s functions will then also depend polynomially on $t$. We conjecture that this polynomial dependence extends to the Feynman rules character on the full Hopf algebra.
We emphasize that the Renormalized Group Flows $(\varphi_t)_{\mathrm{ren}}$ and $\varphi_{\mathrm{ren}}(s)=
(\varphi_t)_{\mathrm{ren}}(s)\big|_{t=0}$ of the characters $\varphi_t$ and $\varphi$ are different. While $\varphi_t(f_1),
\varphi_t(f_2)$ are independent of $t$, we have $$\label{8:1new}
\varphi_t(f_3)=\varphi(f_3)-4t(\varphi_-(f_1)\varphi(f_2)-\varphi_-(f_2)\varphi(f_1).$$ We have used $\varphi_t=\tilde
R^{-1}(\mathrm{Ad}(g_-(t))\tilde R(\varphi))$ for this calculation, since it is easier to extract the pole part of a Laurent series than the holomorphic part, but we could also use $\varphi_t=\tilde
R^{-1}(\mathrm{Ad}(g_+(t))\tilde R(\varphi))$. In this case, we get $$\label{8:2}
\varphi_t(f_3)=\varphi(f_3)-4t(\varphi_+(f_1)\varphi(f_2)-\varphi_+(f_2)\varphi(f_1).$$ As a check, we verify that (\[8:1new\]) and (\[8:2\]) are equal. Let $\pi$ denote the projection onto the pole part of a Laurent series. Then $$\begin{aligned}
\varphi_+(f_1)\varphi(f_2)-\varphi_+(f_2)\varphi(f_1)
&=& (\varphi(f_1)-\pi(\varphi(f_1))\varphi(f_2)
(\varphi(f_2)-\pi(\varphi(f_2))\varphi(f_1)\\
&=& -\pi(\varphi(f_1)\varphi(f_2) -(-\pi(\varphi(f_2)))\varphi(f_1)\\
&=&\varphi_-(f_1)\varphi(f_2)-\varphi_-(f_2)\varphi(f_1).\end{aligned}$$
The computations for the character $\chi_t$ in (\[phi-chi\]) associated to the toy model character $\varphi$ with $-n+2m=0$ give $$(\chi_t^s)_-(f_1)=-\frac{1}{\lambda},\ \ \ \ \ (\chi_t^s)_-(f_1)=0, \
\ \ \ \
(\chi_t^s)_-(f_4)=-\frac{-\frac{\pi^2}{18}+\pi^2t}{\lambda},$$ $$(\chi_t^s)_-(f_8)=\frac{\pi^2}{24\lambda^2}+\frac{\pi^2t(18s+(5+18t)a)}{6\lambda}.$$ In agreement with Theorem \[t:7.9a\], $(\chi_t^s)_-$ is independent of $s$ when evaluated on $f_1$, $f_2$ and $f_4$. However, $(\chi_t^s)_-(f_8)$ depends on $s$, so $\chi_t$ is not local. We confirm that the necessary condition (\[e:condition:chi\]) for locality in Theorem \[t:h3\] does not hold. Indeed, $\varphi_-(f_1)=1/z\not= 0$ and $3(-1/z)(\pi^2)\not=0\cdot a$. The $\beta$-function on $f_1$, $f_2$, $f_4$ is given by $$\beta_{\chi_t}(f_1)=1, \ \ \ \ \ \beta_{\chi_t}(f_2)=0,\ \ \
\ \ \
\beta_{\chi_t}(f_4)=-\frac{\pi^2}{6}+3\pi^2t.$$ The renormalized character $(\chi_t)_{\mathrm{ren}}(s)$ is given by $$(\chi_t)_{\mathrm{ren}}(s)(f_1)=s-a, \ \ \
(\chi_t)_{\mathrm{ren}}(s)(f_2)=\pi^2/4, \ \ \
(\chi_t)_{\mathrm{ren}}(s)(f_4)=\frac{\pi^2}{12}(s+36st-(1+24t)a),$$ and satisfies the abstract RGE.
Let $\mathcal H^2$ be the Hopf subalgebra generated by the trees $$t_0=1_\mathcal{T}, \;\;\;\;
t_1= \ta1, \;\;\;\;
t_2= \tb2, \;\;\;\;
t_3= \tc3, \;\;\;\;
t_4=\td31, \;\;\;\;
t_5=\te4
.
$$ For $T\in \{t_1,\ldots , t_{5}\}$, let $Z_T$ be the corresponding infinitesimal character. The Lie algebra $\mathfrak g_2$ of scalar valued infinitesimal characters of $\mathcal H^2$ is generated by $Z_{t_1},\ldots,\ Z_{t_{5}}$. Let $G_1$ be the scalar valued character group of $\mathcal H^2$, and let $G_0$ be the semi-direct product $G_1\rtimes{\mathbb C}$ given by $$(g,t)\cdot (g',t')=(g\cdot\theta_t( g'),t+t'),$$ where $\theta_t(g)(T)=e^{t\mathrm{deg}(T)}g(T)$ homogenous $T$. Define a new variable $Z_0$ with $[Z_0,Z_{t_i}] = \mathrm{deg}(t_i)Z_{t_i},$ so formally $Z_0=\frac{d}{d\theta}$. The Lie algebra $\mathfrak g_0$ of $G_0$ is generated by $Z_0,Z_{t_1},\ldots,\ Z_{t_{5}}$.
The conditions a) and b) in Definition \[bialg\] of a Lie bialgebra can be written in a basis as a system of quadratic equations. We can solve this system explicitly, e.g. via Mathematica. It turns out that there are $43$ families of Lie bialgebra structures $\gamma$ on $\mathfrak g_0$. In more detail, the system of quadratic equations involves 90 variables. Mathematica gives 1 solution with 82 linear relations (and so 8 degrees of freedom), 7 solutions with 83 linear relations, 16 solutions with 84 linear relations, 13 solutions with 85 linear relations, 5 solutions with 86 linear relations, and 1 solution with 87 linear relations.
To any Lax equation with a spectral parameter, one can associate a spectral curve and study its algebro-geometric properties (see [@sts]). In our case, we consider the adjoint representation $\mathrm{ad}:\delta\to
{\mathfrak gl}(\delta)$ and the induced adjoint representation of the loop algebra. The spectral curve is given by the characteristic equation of $\mathrm{ad}(L\lambda)$: $\Gamma_0=\{ (\lambda,\nu)\in{\mathbb C}-\{0\}\times{\mathbb C}\ | \
\det(\mathrm{ad}(L(\lambda)-\nu {\mathrm{ Id} }))=0\}$.
The theory of the spectral curve and its Jacobian usually assumes that the spectral curve is irreducible. For all 43 families of Lie bialgebra structures on $\delta$, on the associated Lie algebra $\mathrm{ad}(\delta)$ all eigenvalues of the characteristic equation are zero, and the zero eigenspace is nine dimensional. The spectral curve itself is the union of degree one curves. Thus each irreducible component has a trivial Jacobian, and the spectral curve theory breaks down. The integrability of these Lax pair equation remains open for future investigations.
Gabriel Baditoiu would like to thank the Max-Planck-Institute for Mathematics, Bonn and the Erwin Schrödinger International Institute for Mathematical Physics for the hospitality. Steven Rosenberg would also like to thank ESI and the Australian National University.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The dispute on whether the three-dimensional (3D) incompressible Euler equations develop an infinitely large vorticity in a finite time (blowup) keeps increasing due to ambiguous results from state-of-the-art direct numerical simulations (DNS), while the available simplified models fail to explain the intrinsic complexity and variety of observed structures. Here, we propose a new model formally identical to the Euler equations, by imitating the calculus on a 3D logarithmic lattice. This model clarifies the present controversy at the scales of existing DNS and provides the unambiguous evidence of the following transition to the blowup, explained as a chaotic attractor in a renormalized system. The chaotic attractor spans over the anomalously large six-decade interval of spatial scales. For the original Euler system, our results suggest that the existing DNS strategies at the resolution accessible now (and presumably rather long into the future) are unsuitable, by far, for the blowup analysis, and establish new fundamental requirements for the approach to this long-standing problem.'
author:
- 'Ciro S. Campolina'
- 'Alexei A. Mailybaev'
bibliography:
- 'refs.bib'
title: Chaotic Blowup in the 3D Incompressible Euler Equations on a Logarithmic Lattice
---
The existence of blowup (a singularity developing in a finite time from smooth initial data) in incompressible ideal flow is a long-standing open problem in physics and mathematics. Such blowup is anticipated by Kolmogorov’s theory of developed turbulence [@frisch1999turbulence], predicting that the vorticity field diverges at small scales as $\delta\omega \sim \ell^{-2/3}$, while the time of the energy transfer between the integral and viscous scales remains finite in the inviscid limit. In this context, the blowup would reveal an efficient mechanism of energy transfer to small scales. Similar open problems on finite-time singularities, which are fundamental for the understanding of physical behavior, exist across many other fields such as natural convection [@majda2002vorticity], geostrophic motion [@pedlosky2013geophysical; @constantin1994formation], magnetohydrodynamics [@biskamp1997nonlinear], plasma physics [@glassey1986singularity; @andreasson2011einstein] and, of course, general relativity [@choquet2009general].
Besides purely mathematical studies, e.g., [@beale1984remarks; @chae2008incompressible; @tao2016finite], a crucial role in the blowup analysis is given to direct numerical simulations (DNS). The chase after numerical evidence of blowup in the 3D incompressible Euler equations has a long history [@gibbon2008three]. Most early numerical studies were in favor of blowup, e.g., [@pumir1992finite; @kerr1993evidence; @grauer1998adaptive]. But the increase of resolution owing to more powerful computers showed that the growth of small-scale structures may be depleted at smaller scales, even though it was demonstrating initially the blowup tendency [@hou2007computing; @grafke2008numerical; @hou2009blow]. It is fair to say that, now, there is a lack of consensus even on the more probable answer (existence or not) to the blowup problem. Blowup remains an active area of numerical research [@kerr2013bounds; @brenner2016potential; @larios2018computational], but computational limitations are still the major obstacle. See also [@luo2013potentially; @elgindi2018finite] for the blowup at a physical boundary, which is a related but different problem.
Numerical limitations of the DNS can be overcome using simplified models [@uhlig1997singularities; @dombre1998intermittency; @mailybaev2012renormalization], which were developed in lower spatial dimensions [@constantin1985simple; @okamoto2008generalization] or by exploring the cascade ideas in so-called shell models [@gledzer1973system; @ohkitani1989temporal; @l1998improved]. The reduced wave vector set approximation (REWA) model introduced in [@eggers1991does; @grossmann1996developed] restricted the Euler or Navier-Stokes dynamics to a self-similar set of wave vectors. Despite being rather successful in the study of turbulence [@grossmann1992intermittency; @biferale2003shell; @bohr2005dynamical], these models fall short of reproducing basic features of full DNS for the blowup phenomenon.
Here, we resolve this problem with a new model that demonstrates qualitative agreement with the existing DNS and permits a highly reliable blowup analysis. The model is formulated in a form identical to the original Euler equations, but with the algebraic structure defined on the 3D logarithmic lattice. We show that the blowup in this model is associated with a chaotic attractor of a renormalized system, in accordance with some earlier theoretical conjectures [@pomeau2005unfinished; @greene2000stability; @mailybaev2012c; @de2017chaotic]; one can also make an interesting connection with the chaotic Belinskii-Khalatnikov-Lifshitz singularity in general relativity [@belinskii1970oscillatory; @choquet2009general]. A distinctive property of the attractor is its anomalous multiscale structure, which explains the diversity of the existing DNS results, discloses fundamental limitations of current strategies, and provides new guidelines for the original blowup problem.
{width="\textwidth"}
<span style="font-variant:small-caps;">Model.</span> Consider the set $\mathbb{\Lambda} = \{ \pm \lambda^{n},\,n\in\mathbb{Z}\}$ of positive and negative integer powers of a fixed real number $\lambda>1$. Then wave vectors $\mathbf{k} = (k_1,k_2,k_3) \in \mathbb{\Lambda}^3$ define a logarithmic lattice in 3D Fourier space. We retain three independent spatial directions, unlike shell or REWA models [@biferale2003shell; @eggers1991does] featuring a fixed number of wave vectors per spherical shell. In analogy to the convolution operation, we define $$(u \ast v)(\mathbf{k}) = \sum_{\substack{\mathbf{p},\mathbf{q} \in \mathbb{\Lambda}^3\\[2pt] \mathbf{p} + \mathbf{q} = \mathbf{k}}} u(\mathbf{p})v(\mathbf{q})
\label{convolution}$$ for complex-valued functions $u(\mathbf{k})$ and $v(\mathbf{k})$. Since the sum is restricted to exact triads on the lattice $\mathbf{k} \in \mathbb{\Lambda}^3$, operation (\[convolution\]) is nontrivial only for specific values of $\lambda$. We will consider the golden mean, $\lambda = (1+\sqrt{5})/2$, which also appeared in a similar context for shell models [@l1999hamiltonian; @gurcan2017nested]. In this case the sum in contains 216 distinct terms originating from the equality $\lambda^{n-1}+\lambda^n = \lambda^{n+1}$ and coupling the wave numbers that differ by $\lambda$ or $\lambda^2$ in each spatial direction. Note that Eq. (\[convolution\]) can be seen as a projection of the convolution to the nodes of the 3D logarithmic lattice, which keeps the middle-range interactions. One may expect that the long-range interactions are less important for the blowup problem than, e.g., for the developed turbulence, because very small scales are weakly perturbed in smooth initial conditions.
Just like the classical convolution, operation is bilinear, commutative and satisfies the Leibniz rule $$\partial_j(u \ast v) = \partial_j u \ast v + u \ast \partial_j v \quad \text{for} \quad j=1,2,3,
\label{leibniz}$$ where derivatives are given by the Fourier factors, $\partial_j u(\mathbf{k}) = ik_j u(\mathbf{k})$; here $i$ is the imaginary unit. However, operation is not associative, $(u \ast v) \ast w \ne u \ast (v \ast w)$. Nonetheless, it possesses the weaker property $$\langle u \ast v, w \rangle = \langle u , v \ast w\rangle,
\label{associativity}$$ where $\langle u, v \rangle = \sum_{\mathbf{k} \in \mathbb{\Lambda}^3 } u(\mathbf{k})v^*(\mathbf{k})$ is the scalar product.
In our simplified model, we represent the velocity field as a function $\mathbf{u}(\mathbf{k},t) = (u_1,u_2,u_3) \in \mathbb{C}^3$ of the wave vector $\mathbf{k} \in \mathbb{\Lambda}^3$ and time $t \in \mathbb{R}$. Thus, at each lattice point, $\mathbf{u}$ stands for the corresponding velocity in Fourier space. Similarly, we define the scalar function $p(\mathbf{k})$ representing the pressure. All functions are supposed to satisfy the reality condition: $u_i(-\mathbf{k}) = u_i^*(\mathbf{k})$. For the governing equations, we use the exact form of 3D incompressible Euler equations $$\partial_t u_i + u_j \ast \partial_j u_i = -\partial_i p, \quad \partial_j u_j=0,
\label{euler}$$ which are now considered on the logarithmic lattice; here and below repeated indices imply the summation.
The proposed model retains most of the properties of the continuous Euler equations, which rely only upon the structure of the equations and elementary operations such as (\[leibniz\]) and (\[associativity\]). These include the basic symmetries: scaling (in a discrete form $\mathbf{k} \mapsto \lambda\mathbf{k}$), isotropy (reduced to the discrete group $\mathsf{O_h}$ of cube isometries [@landau2013quantum], Sec. 93), and spatial translations (given in Fourier representation by $\mathbf{u} \mapsto e^{-i\mathbf{a}\cdot\mathbf{k}}\mathbf{u}$). The system conserves energy $E = \frac{1}{2} \langle u_j, u_j \rangle$ and helicity $H = \langle u_j ,\omega_j \rangle$, where $\pmb{\omega} = \nabla \times \mathbf{u}$ is the vorticity. It also has an infinite number of invariants, which can be interpreted as Kelvin’s circulation theorem; see the Supplemental Material (SM) [@SupMat]. Proofs of all these properties are identical to the continuous case.
<span style="font-variant:small-caps;">Simulations.</span> For numerical simulations, we used the Euler equations in vorticity formulation $$\partial_t \omega_i + u_j \ast \partial_j \omega_i - \omega_j \ast \partial_j u_i = 0,
\label{vorticity}$$ where $\mathbf{u} = \mathrm{rot}^{-1}\pmb{\omega} = i\mathbf{k}\times \pmb{\omega}/|\mathbf{k}|^2$. Aiming for the blowup study, we consider initial conditions limited to large scales, $\lambda \le |k_{1,2,3}| \le \lambda^3$; see SM [@SupMat] for an explicit form of the initial conditions. Equations (\[vorticity\]) are integrated with double-precision using the fourth-order Runge-Kutta-Fehlberg adaptive scheme. The local error, relative to $\omega_{\max}(t)$, was kept below $10^{-10}$. The number of nodes $n = 1,\ldots,N$ was increased dynamically during the simulation in order to avoid the error due to truncation at small scales: the truncation error was kept below $10^{-20}$ for the enstrophy $\Omega = \frac{1}{2}\langle \omega_j , \omega_j \rangle$; see SM [@SupMat] for more details. Together, this provided the remarkably high accuracy of numerical results. We stopped the simulation with $N = 80$, thus, covering the scale range of $\lambda^N \sim 10^{17}$ with the total of $13180$ time steps. The energy was conserved at all times with the relative error below $3.8\times 10^{-10}$.
{width="\textwidth"}
Figures \[fig1\](a) and \[fig1\](b) analyze the temporal evolution of the maximum vorticity $\omega_{\max}(t) = \max_{\mathbf{k}\in\mathbb{\Lambda}^3}|\pmb{\omega}(\mathbf{k},t)|$ and the corresponding wave number $k_{\max}(t) = |\mathbf{k}|$. The Beale-Kato-Majda theorem [@beale1984remarks] (whose proof for our model is identical to the continuous case) states that the blowup of the solution at finite time $t_b$ requires that the integral $\int_0^t \omega_{\max}(t)dt$ diverges as $t \to t_b$. In particular, this implies that the growth of maximum vorticity must be at least as fast as $\omega_{\max}(t) \gtrsim (t_b-t)^{-1}$. This dependence is readily confirmed in Fig. \[fig1\](a) providing the blowup time $t_b = 15.870 \pm 0.001$. Furthermore, Fig. \[fig1\](b) tracks the dependence $\omega_{\max}(t) \sim (t_b-t)^{-1}$ in logarithmic coordinates up to the values $\omega_{\max} \sim 10^{5}$. The same figure demonstrates the power-law dependence $k_{\max}(t) \sim (t_b-t)^{-\gamma}$ with the exponent $\gamma = 2.70\pm0.01$, simulated up to extremely small physical scales, $\ell \sim 1/k_{\max} \sim 10^{-15}$. Finally, Fig. \[fig1\](c) shows the development of the power law $E(k) \propto k^{-\xi}$ in the energy spectrum as $t \to t_b$. The exponent can be obtained with the dimensional argument $E \propto \omega_{\max}^2/k_{\max}^3$, which yields $\xi = 3-2/\gamma \approx 2.26$.
<span style="font-variant:small-caps;">Chaotic blowup.</span> The observed scaling agrees with the Leray-type [@leray1934mouvement] self-similar blowup solution $\pmb{\omega}_L(\mathbf{k},t)$ defined as $$\pmb{\omega}_L(\mathbf{k},t) = (t_b-t)^{-1}\mathbf{W}[(t_b-t)^{\gamma}\mathbf{k}].
\label{Leray}$$ Such a solution, however, cannot describe the blowup in Fig. \[fig1\], where the maximum vorticity and the corresponding scale $\ell \sim 1/k_{\max}$ have the power-law behavior only in average, with persistent irregular oscillations.
In order to understand the nonstationary blowup dynamics, we perform the change of coordinates $$\begin{aligned}
\widetilde{\pmb{\omega}} &= (t_b-t)\pmb{\omega}, & \eta &= \log|\mathbf{k}|, \nonumber \\
\mathbf{o} &= \mathbf{k}/|\mathbf{k}|, & \tau &= -\log(t_b-t).
\label{Renorm}\end{aligned}$$ This change of coordinates applies similarly in Fourier space $\mathbb{R}^3$ and in our 3D lattice $\mathbb{\Lambda}^3$. The Euler equations (\[vorticity\]) in renormalized coordinates take the form $$\partial_\tau \widetilde{\pmb{\omega}} =
G[\widetilde{\pmb{\omega}}],
\label{eqRen}$$ where the $i$th component of the nonlinear operator $G[\widetilde{\pmb{\omega}}]$ is $$\label{RHSrenorm}
(G[\widetilde{\pmb{\omega}}])_i = -\widetilde{\omega}_i -\widetilde{u}_j \ast \widetilde{\partial}_j\widetilde{\omega}_i + \widetilde{\omega}_j \ast \widetilde{\partial}_j \widetilde{u}_i, \quad \widetilde{\partial}_j = io_j;$$ see SM [@SupMat] for derivations. The choice of variables (\[Renorm\]) is motivated by the scaling invariance: the operator $G[\widetilde{\pmb{\omega}}]$ is homogeneous (invariant to translations) with respect to $\tau$ and $\eta$, which correspond to temporal and spatial scaling, respectively. In our model, the scaling invariance is represented by the shifts of $\eta$ with integer multiples of $\log \lambda$. These properties allow studying the blowup as an attractor of system (\[eqRen\]); see, e.g., [@eggers2009role; @mailybaev2012renormalization]. For example, the self-similar blowup solution (\[Leray\]) corresponds to the traveling wave $\widetilde{\pmb{\omega}} = \mathbf{W}(e^{\eta-\gamma\tau}\mathbf{o})$, which has a stationary profile in the comoving reference frame $\eta' = \eta-\gamma\tau$. In the limit $\eta \sim \gamma \tau \to \infty$, the original variables (\[Renorm\]) yield the blowup dynamics: $|\pmb{\omega}| \to \infty$ and $\ell \sim 1/|\mathbf{k}| \to 0$ as $t \to t_b$. Such a blowup is robust to small perturbations if the traveling wave is an attractor in system (\[eqRen\]).
![Evolution of a small perturbation of vorticity, $\max_{\mathbf{k}}|\delta \widetilde{\pmb{\omega}}|$, in renormalized variables. Solutions deviate exponentially with the Lyapunov exponent $\lambda_{\max} \approx 9.18$.[]{data-label="fig3"}](fig3.pdf){width="0.81\columnwidth"}
Irregular evolution observed in Fig. \[fig1\] suggests that the attractor of system (\[eqRen\]) cannot be a traveling wave. We will now argue that the attractor in the renormalized system represents a chaotic wave moving with the average speed $\gamma$. Figure \[fig2\] (see also the Supplemental video [@SupMat] for the 3D picture) shows absolute values of the third component $\widetilde{\omega}_3$ as functions of two wave numbers $k_1 > 0$ and $k_2 > 0$ for four different values of $\tau$; here the third wave vector component is constant and chosen at the node nearest to $k_3 = e^{\gamma\tau + 6} \propto (t_b-t)^{-\gamma}$. This figure presented in log scale demonstrates a wave moving with constant speed in average $\eta \sim \gamma \tau$, but not preserving exactly the spatial vorticity distribution. In order to confirm that the wave is chaotic, we computed the largest Lyapunov exponent $\lambda_{\max} = 9.18 \pm 0.07$ in Fig. \[fig3\]; here we added a tiny perturbation to the original solution at $\tau = 1.7$, when the attractor is already fully established, and observed the exponential deviation of the solutions $\max_{\mathbf{k}}|\delta \widetilde{\pmb{\omega}}(\tau)| \propto e^{\lambda_{\max}\tau}$ in renormalized time $\tau$. In the original variables, this yields the rapid power-law growth $$\max_{\mathbf{k}}|\delta \pmb{\omega}(t)| \propto (t_b-t)^{-\zeta},\quad
\zeta = \lambda_{\max}+1 \approx 10.18.
\label{lyapunov}$$
The striking property of the chaotic attractor is that it restores the isotropy in the statistical sense, even though the solution at each particular moment is essentially anisotropic, in similarity to the recovery of isotropy in the Navier-Stokes turbulence [@frisch1999turbulence; @biferale2005anisotropy]. This property is confirmed in Fig. \[fig4\] presenting the averages of renormalized vorticity components $|\tilde\omega_i|$, considered in the comoving reference frame $\eta'= \eta-\gamma\tau$. The isotropy, as well as other statistical properties, are expected to be established very rapidly in realistic conditions, e.g., in the presence of microscopic fluctuations, because of the very large Lyapunov exponent; see Eq. (\[lyapunov\]). This resembles closely a similar effect in developed turbulence [@ruelle1979microscopic].
![Statistical isotropy: Left panel shows the $\tau$ average of $|\widetilde{\omega}_3|$ from Fig. \[fig2\] in a comoving reference frame $\eta' = \eta-\gamma\tau$. Right panel shows analogous result for the average of $|\widetilde{\omega}_2|$ on plane $(\widetilde{k}_1,\widetilde{k}_3)$. Planes of the two figures are related by the $90^\circ$ rotation about the $\widetilde{k}_3$ axis. Similar results are obtained for other elements of the rotation symmetry group $\mathsf{O_h}$.[]{data-label="fig4"}](fig4.pdf){width="\columnwidth"}
<span style="font-variant:small-caps;">Relation to existing DNS.</span> As one can infer from Figs. \[fig2\] and \[fig4\], the chaotic attractor has the span of about six decades of spatial scales. This property imposes fundamental limitations on the numerical resources necessary for the observation of blowup, assuming that the dynamics in the continuous 3D Euler equations can be qualitatively similar to our model. The approximate time limit, which would be accessible for the state-of-the-art DNS with the $8192^3$ grid [@grafke2008numerical; @hou2009blow; @kerr2013bounds] can be estimated in our model as $t_{\textrm{DNS}} \approx 9$ or $\tau_{\textrm{DNS}} \approx -1.9$ for the renormalized time; see Fig. \[fig2\] (left panel). At this instant, the chaotic attractor is still at its infant formation stage and, hence, the dynamics is essentially transient. The increase of the vorticity from $\omega_{\max}(0) = 0.91$ to $\omega_{\max}(t_{\textrm{DNS}}) = 1.89$ and of the enstrophy from $\Omega(0) = 27.2$ to $\Omega(t_{\textrm{DNS}}) = 1.92 \times 10^3$ is moderate, which is also common for the DNS. Moreover, Fig. \[fig5\] shows that the growth of enstrophy and vorticity for $t \lesssim t_{\textrm{DNS}}$ is not faster than double exponential in agreement with [@hou2007computing; @hou2009blow; @kerr2013bounds]. The chaotic blowup behavior offers a diversity of flow structures as it is indeed observed for different initial conditions [@gibbon2008three]; some DNS showed the incipient development of power-law energy spectra [@agafontsev2015], in qualitative agreement with Fig. \[fig1\](c).
At the time $t_{\textrm{DNS}}$, the wave vector at the vorticity maximum is equal to $\mathbf{k}_{\max} = (\lambda^6,-\lambda^3,\lambda^{10}) \approx (18.9,-4.2,123)$. Its third component is much larger than the other two. This has a similarity with DNS, which typically demonstrate depleting of vorticity growth within quasi-2D (thin in one and extended in the other two directions) vorticity structures [@brachet1992numerical; @frisch2003singularities; @agafontsev2017asymptotic]. Such dominance of one scale over the others by 1 or 2 orders of magnitude persists for larger times in our model.
![Evolution of $\log\log\Omega$ and $\log\log\,(50\omega_{\max})$ for the enstrophy and maximum vorticity; the factor $50$ is used to avoid complex values of the logarithm. The dash-dotted line indicates the blowup time. The vertical solid line estimates the limit $t_{\textrm{DNS}} \approx 9$ that would be accessible for the state-of-the-art DNS with the grid $8192^3$. Until $t_{\textrm{DNS}}$, both $\Omega$ and $\omega_{\max}$ demonstrate the growth not greater than double exponential.[]{data-label="fig5"}](fig5.pdf){width="0.88\columnwidth"}
<span style="font-variant:small-caps;">Conclusions.</span> We propose an explanation for the existing controversy in the blowup problem for incompressible 3D Euler equations. This is accomplished using a new model, which is formally identical to the incompressible Euler equations and defined on the 3D logarithmic grid with proper algebraic operations. Such a model retains most symmetries of the original system along with intrinsic invariants (energy, helicity, circulation, etc.), but permits simulations in extremely large interval of scales.
We show that our model has the non-self-similar blowup, which is explained as a chaotic attractor in renormalized equations. Our results demonstrate that the blowup has enormously higher complexity than anticipated before: its “core” extends to six decades of spatial scales. This suggests that modern DNS of the original continuous model are unsuitable, by far, for the blowup observation; still, the blowup may be accessible to experimental measurements. Since the attractor is chaotic, blowup cannot be probed by the study of local structures.
Our approach to the blowup phenomenon is not limited to the Euler equations, but is ready-to-use for analogous studies in other fields such as natural convection, geostrophic motion, magnetohydrodynamics, and plasma physics.
The authors are grateful to Luca Biferale, Gregory Eyink, Uriel Frisch, Simon Thalabard and Dmitry Agafontsev for most helpful discussions. The work was supported by the CNPq Grant No. 302351/2015-9 and the RFBR Grant No. 17-01-00622.
Supplemental material
=====================
Conservation laws
-----------------
Conservation of quadratic invariants follows from the fact that the product contains only the exact triples of wave vectors. Taking the energy as an example, let us show how the proof can be written using the basic operations defined on the 3D logarithmic lattice, following the standard approach of fluid dynamics. Using the Euler equations , we obtain $$\begin{array}{rcl}
\displaystyle
\frac{dE}{dt} &=&
\displaystyle
\frac{d}{dt} \left( \frac{1}{2} \langle u_i,u_i \rangle \right) = \langle u_i,\partial_t u_i \rangle \\[12pt]
&=& -\,\langle u_i,\partial_i p \rangle -\langle u_i,u_j \ast \partial_j u_i \rangle.
\end{array}$$ The pressure term vanishes owing to the incompressibility condition as $$\label{byparts}
\langle u_i,\partial_i p \rangle = -\langle \partial_i u_i,p \rangle = 0,$$ where the first relation represents the derivation by parts on the 3D lattice. In the inertial term, using commutativity of the product and the properties and , one obtains $$\begin{aligned}
\langle u_i,u_j \ast \partial_j u_i \rangle
= \langle u_i \ast \partial_j u_i, u_j \rangle = \frac{1}{2} \langle \partial_j (u_i \ast u_i),u_j \rangle.\end{aligned}$$ After integration by parts, analogous to , this term vanishes due to the incompressibility condition.
Conservation of helicity can be proved following a similar line of derivations. For the Beale-Kato-Majda theorem, one has to define the functional spaces and the corresponding inequalities; technical details of this functional analysis on the 3D logarithmic lattice will be given elsewhere.
Furthermore, one can make sense of Kelvin’s circulation theorem in system . It is related to the conservation of cross-correlation $\Gamma = \langle u_j,h_j \rangle$ for an arbitrary “frozen-into-fluid” divergence-free field $\mathbf{h}(\mathbf{k},t) = (h_1,h_2,h_3)$ satisfying the equations [@moffatt1969degree] $$\partial_t h_i + u_j \ast \partial_j h_i - h_j \ast \partial_j u_i = 0, \quad \partial_j h_j = 0.
\label{frozen_field}$$ In the continuous formulation, the circulation around a closed material contour $\mathbf{C}(s,t)$ in physical space ($s$ is the arc length parameter) is given by the cross-correlation $\Gamma$ with the field $\mathbf{h}(\mathbf{x},t) = \oint \frac{\partial\mathbf{X}}{\partial s}\,\delta^3(\mathbf{x}-\mathbf{C}(s,t))\,ds$, where $\delta^3$ is the 3D Dirac delta-function; see, e.g. [@zakharov1997hamiltonian; @majda2002vorticity]. Therefore, $\Gamma$ represents the generalized circulation in Kelvin’s theorem. Its conservation yields the infinite number of circulation invariants in our model: the cross-correlation $\Gamma$ is conserved for any solution of system .
Note that zero wave number can also be considered in the model by adding it into the set $\mathbb{\Lambda}$. This will further increase the number of terms in the sum . Note also that the number of degrees of freedom in our model is substantially smaller than in the original Euler system. On one hand, this is an important advantage of the model that allows numerical simulations for extremely large range of scales. On the other hand, one should be careful when using this model in situations where thermalization at small scales may play a role, e.g., [@cichowlas2005effective; @bowman2006links].
Initial conditions
------------------
Initial conditions used in numerical simulations are given below in terms of velocities. Nonzero components are limited to large scales $\lambda \le |k_{1,2,3}| \le \lambda^3$ and taken in the form $$u_j(\mathbf{k}) = \frac{|\epsilon_{jmn}|}{2}k_m k_n e^{i\theta_j(\mathbf{k})-|\mathbf{k}|}, \quad \text{for} \quad j=1,2.$$ Here $\epsilon_{jmn}$ is the Levi-Civita permutation symbol and the phases $\theta_j$ are given by $$\begin{array}{rcl}
\theta_j(\mathbf{k}) & = & \mathrm{sgn}(k_1)\alpha_j + \mathrm{sgn}(k_2)\beta_j
+\mathrm{sgn}(k_3)\delta_j
\\[3pt]
&& +\,\mathrm{sgn}(k_1k_2k_3)\gamma_j
\end{array}$$ with the constants $(\alpha_1,\beta_1,\delta_1,\gamma_1) = (1,-7,13,-3)/4$ and $(\alpha_2,\beta_2,\delta_2,\gamma_2) = (-1,-3,11,7)/4$. The third component of velocity is uniquely defined by the incompressibility condition. Several tests were also performed with random initial conditions limited to large scales. In all the test, we observed the same chaotic attractor of the renormalized system and, therefore, the same (universal) asymptotic form of the chaotic blowup.
Adaptive scheme
---------------
Since only a finite number of modes can be simulated, the infinite-dimensional nature of the problem was tracked very accurately by using the following adaptive scheme in the simulation. At each time step, we computed the enstrophy of the modes with the wave numbers $|\mathbf{k}| \ge K_{\max}/\lambda$, where $K_{\max}$ is the largest wavenumber in each direction of the lattice. This quantity estimates the enstrophy error due to mode truncation, and it was kept extremely small, below $10^{-20}$, during the whole simulation. Every time the threshold of $10^{-20}$ was reached we increased the number of nodes in each direction by five, i.e., multiplying $K_{\max}$ by $\lambda^5$.
Renormalized Euler equations
----------------------------
With the renormalized variables , it is convenient to define new differentiation operators as the Fourier factors $\widetilde{\partial}_j = io_j$, where $\mathbf{o} = (o_1,o_2,o_3) = \mathbf{k}/|\mathbf{k}|$ and $i$ is the imaginary unit. Thus, derivatives in the original and in the renormalized variables are related as $\partial_j = e^{\eta}\widetilde{\partial}_j$. Also, the renormalized velocity can be defined as $\widetilde{\mathbf{u}} = (t_b-t)|\mathbf{k}|\mathbf{u}$, which is related to the renormalized vorticity as $$\widetilde{\mathbf{u}} = i\mathbf{o} \times \widetilde{\pmb{\omega}}.
\label{renorm_u}$$
Using relations and , the vorticity equation , after dropping the common factor $e^{2\tau}$, becomes $$\partial_{\tau}\widetilde{\omega}_i + \widetilde{\omega}_i
+ \widetilde{u}_j \ast \widetilde{\partial}_j\widetilde{\omega}_i - \widetilde{\omega}_j \ast \widetilde{\partial}_j \widetilde{u}_i = 0.$$ This equation has the form –. Since $\tau$ and $\eta$ do not appear explicitly in , the renormalized system (8) is translation invariant with respect to these two variables.
Note that the existence of a chaotic wave traveling with constant mean velocity in the renormalized system yields the power law $\omega_{\max}(t) \sim (t_b-t)^{-1}$ observed in Fig. \[fig1\](b). This follows from the transformation , similarly to the Leray-type solution . In fact, existence of a chaotic or regular wave with a positive speed $\gamma$ as an attractor in the renormalized system is a sufficient condition for the finite-time blowup.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Geometric, topological and graph theory modeling and analysis of biomolecules are of essential importance in the conceptualization of molecular structure, function, dynamics, and transport. On the one hand, geometric modeling provides molecular surface and structural representation, and offers the basis for molecular visualization, which is crucial for the understanding of molecular structure and interactions. On the other hand, it bridges the gap between molecular structural data and theoretical/mathematical models. Topological analysis and modeling give rise to atomic critical points and connectivity, and shed light on the intrinsic topological invariants such as independent components (atoms), rings (pockets) and cavities. Graph theory analyzes biomolecular interactions and reveals biomolecular structure-function relationship. In this paper, we review certain geometric, topological and graph theory apparatuses for biomolecular data modeling and analysis. These apparatuses are categorized into discrete and continuous ones. For discrete approaches, graph theory, Gaussian network model, anisotropic network model, normal mode analysis, quasi-harmonic analysis, flexibility and rigidity index, molecular nonlinear dynamics, spectral graph theory, and persistent homology are discussed. For continuous mathematical tools, we present discrete to continuum mapping, high dimensional persistent homology, biomolecular geometric modeling, differential geometry theory of surfaces, curvature evaluation, variational derivation of minimal molecular surfaces, atoms in molecule theory and quantum chemical topology. Four new approaches, including analytical minimal molecular surface, Hessian matrix eigenvalue map, curvature map and virtual particle model, are introduced for the first time to bridge the gaps in biomolecular modeling and analysis. Emphasis is given to the connections of existing biophysical models/methods to mathematical theories, such as graph theory, Morse theory, Poicaré-Hopf theorem, differential geometry, differential topology, algebraic topology and geometric topology. Potential new directions and standing open problems are briefly discussed.'
author:
- |
Kelin Xia$^1$ [^1] and Guo-Wei Wei$^{2,3}$ [^2]\
$^1$Division of Mathematical Sciences, School of Physical and Mathematical Sciences,\
Nanyang Technological University, Singapore 637371\
$^2$Department of Mathematics\
Michigan State University, MI 48824, USA\
$^3$Department of Biochemistry & Molecular Biology\
Michigan State University, MI 48824, USA\
title: 'A review of geometric, topological and graph theory apparatuses for the modeling and analysis of biomolecular data '
---
Key words: Molecular bioscience, Molecular biophysics, Graph theory, Graph spectral theory, Graph Laplacian, Differential geometry, Differential topology Persistent homology, Morse theory, Conley index, Poincaré-Hopf index, Laplace-Beltrami operator, Dynamical system.
[ ]{}
Introduction
============
Life science is regarded as the last forefront in natural science and the 21$^{\rm st}$ century will be the century of biological sciences. Molecular biology is the foundation of biological sciences and molecular mechanism, which is governed by all the valid mechanics, including quantum mechanics when it is relevant, and is the ultimate truth of life science. One trend of biological sciences in the 21$^{\rm st}$ century is that many traditional disciplines, such as epidemiology, neuroscience, zoology, physiology and population biology, are transforming from macroscopic and phenomenological to molecular-based sciences. Another trend is that biological sciences in the 21$^{\rm st}$ century are transforming from qualitative and descriptive to quantitative and predictive, as many other disciplines in natural science have done in the past. Such a transformation creates unprecedented opportunities for mathematically driven advances in life science [@Wei:2016].
Biomolecules, such as proteins and the nucleic acids, including DNA and RNA, are essential for all known forms of life, such as animals, fungi, protists, archaea, bacteria and plants. Indeed, proteins perform a vast variety of biological functions, including membrane channel transport, signal transduction, organism structure supporting, enzymatic catalysis for transcription and the cell cycle, and immune agents. In contrast, nucleic acids function in association with proteins and are essential players in encoding, transmitting and expressing genetic information, which is stored through nucleic acid sequences, i.e., DNA or RNA molecules and transmitted via transcription and translation processes. The understanding of biomolecular structure, function, dynamics and transport is a fundamental issue in molecular biology and biophysics. A traditional dogma is that sequence determines structure, while structure determines function [@Anfinsen:1973]. This, however, has been undermined by the fact that many intrinsically disordered proteins can also be functional [@Onuchic:1997; @White:1999; @Schroder:2005; @Chiti:2006]. Disordered proteins are associated with sporadic neurodegenerative diseases, including Alzheimer’s disease, Parkinson’s disease and mad cow disease [@Chiti:2006; @Uversky:2008]. Randomness in disordered proteins is a consequence of protein flexibility, which is an intrinsic protein function. In general, the understanding of protein structure-function relationship is also crucial for shedding light on protein specification, protein-protein interactions, protein-drug binding that are essential to drug design and discovery, and improving human health and wellbeing.
Much of the present understanding of biomolecular structures and functions and their relationship come from experimental data that are collected from a number of means, such as macromolecular X-ray crystallography, nuclear magnetic resonance (NMR), cryo-electron microscopy (cryo-EM), electron paramagnetic resonance (EPR), multiangle light scattering, confocal laser-scanning microscopy, scanning capacitance microscopy, small angle scattering, ultra fast laser spectroscopy, etc. The major players for single macromolecules are X-ray crystallography and NMR. For example, advanced X-ray crystallography technology is able to offer decisive structural information at Armstrong and sub-Armstrong resolutions, while an important advantage of NMR experiments is that they are able to provide biomolecular structural information under physiological conditions. The continuously effort in the past few decades has made X-ray crystallography and NMR technologically relatively well developed, except for their use in special circumstances, such as the study of membrane proteins. However, theses approaches are not directly suitable for proteasomes, subcellular structures, organelles, cells and tissues, whose study has become increasingly popular in structural biology. Currently, a unique experimental tool for imaging subcellular structures, organelles, multiprotein complexes and even cells and tissues is cryo-EM [@Volkmann:2010].
The rapid advances of experimental technology in the past few decades have led to the accumulation of vast amount of three-dimensional (3D) biomolecular structural data. The [Protein Data Bank (PDB)](http://www.rcsb.org/pdb/home/home.do) has collected more than one hundred twenty thousands of 3D biomolecular structures. Biomolecular geometric information holds the key to our understanding of biomolecular structure, function, and dynamics. It has wide spread applications in virtual screening, computer-aid drug design, binding pocket descriptor, quantitative structure activity relationship, protein design, RNA design, molecular machine design, etc. In general, the availability of biomolecular structural data has paved the way for the transition from the traditional qualitative description to quantitative analysis and prediction in biological sciences. An essential ingredient of quantitative biology is geometric, topological and graph theory modeling, analysis and computation. Aided by increasingly powerful high performance computers, geometric, topological and graph theory modeling, analysis and computation have become indispensable apparatuses not only for the visualization of biological data, but also filling the gap between biological data and mathematical models of biological systems [@ZYu:2008; @ZYu:2008b; @XFeng:2012a; @XFeng:2013b; @KLXia:2014a; @MXChen:2012; @Rocchia:2002; @PMach:2011; @XShi:2011; @JLi:2013; @Decherchi:2013; @LiLin:2014].
One of the simplest molecular geometric models, or molecular structural models, is the space-filling Corey-Pauling-Koltun (CPK) theory, which represents an atom by a solid sphere with a van der Waals (VDW) radius [@Koltun:1965]. The outer boundary of CPK model gives rise to the van der Waals (vdW) surface, which is composed of piece-wise unburied sphere surfaces. Solvent accessible surface (SAS) and solvent-excluded surface (SES) have also been introduced to create smooth molecular surfaces by rolling a probe molecule over the vdW surface [@Richards:1977; @Connolly85]. SESs have widely been applied to protein folding [@Spolar], protein surface topography [@Kuhn], protein-protein interactions [@Crowley], DNA binding and bending [@Dragan], macromolecular docking [@Jackson], enzyme catalysis [@LiCata], drug classification [@Bergstrom], and solvation energies [@Raschke]. The SES model also plays a crucial role in implicit solvent models [@Baker:2005; @DuanChen:2011a], molecular dynamics simulations [@Geng:2011] and ion channel transports [@DuanChen:2011a; @QZheng:2011a; @QZheng:2011b]. Computationally, efficient algorithms for computing SES are developed or introduced, such as alpha-shapes [@WYChen:2010] and marching tetrahedra [@SLChan:1998]. A popular software for the Lagrangian representation of SESs, called MSMS, has been developed [@Sanner:1996]. Recently, a software package, called Eulerian solvent excluded surface (ESES), for the Eulerian representation of SESs [@ESES:2015], has also been developed. However SAS and SES are still not differentiable and have geometric singularities, i.e., cusps and tips. To construct smooth surface representation of macromolecules, Gaussian surface (GS) has been proposed to represent each atom by a C$^{\infty}$ Gaussian function, while accounting their overlapping properties [@Grant:1995]. Differential geometry theory of surfaces provides a natural approach to describe biomolecular surfaces and boundaries. Utilizing the Euler-Lagrange variation, a differential geometry based surface model, the minimal molecular surface (MMS), has been introduced for biomolecular geometric modeling [@Bates:2006; @Bates:2008]. Differential geometry based variational approach has been widely applied to biophysical modeling of solvation [@Wei:2009; @ZhanChen:2010a; @ZhanChen:2010b; @BaoWang:2015a], ion channel [@Wei:2012; @DuanChen:2012a; @DuanChen:2012b] and multiscale analysis [@Wei:2009], in conjugation with other physical models, such as electrostatics, elasticity and molecular mechanics [@Wei:2013]. Geometry modeling and annotation of biomolecular surfaces together with physical features, such as electrostatics and lipophilicity, provide some of the best predictions of biomolecular solvation free energies [@BaoWang:2016FFTS; @BaoWang:2016HPK], protein-drug binding affinities [@BaoWang:2016FFTB], protein mutation energy changes [@ZXCang:2016a] and protein-protein interaction hot spots [@Darnell:2008; @Demerdash:2009].
Theoretical modeling of the structure-function relationship of biomolecules is usually based on fundamental laws of physics, i.e., quantum mechanics (QM), molecular mechanism (MM), continuum mechanics, statistical mechanics, thermodynamics, etc. QM methods are indispensable for chemical reactions and protein degradations [@Warshel:1976; @Cui:2002; @YZhang:2009a]. Molecular dynamics (MD) [@McCammon:1977] is a powerful tool for the understanding of the biomolecular conformational landscapes and elucidating collective motion and fluctuation. Currently, MD is a main workhorse in molecular biophysics for biomolecular modeling and simulation. However, all-electron or all-atom representations and long-time integrations lead to such an excessively large number of degrees of freedom that their application to real-time large-scale dynamics of large proteins or multiprotein complexes becomes prohibitively expensive. For instance, current computer simulations of protein folding take many months to come up with a very poor copy of what Nature administers perfectly within a tiny fraction of a second. Therefore, in the past few decades, many graph theory based biomolecular models, including normal mode analysis (NMA) [@Go:1983; @Tasumi:1982; @Brooks:1983; @Levitt:1985], elastic network model (ENM) [@Bahar:1997; @Bahar:1998; @Atilgan:2001; @Hinsen:1998; @Tama:2001; @LiGH:2002] become very popular for understanding protein flexibility and long time dynamics. In these models, the diagonalization of the interaction matrix or Hamiltonian of a protein is a required procedure to obtain protein eigenmodes and associated eigenvalues. The low order eigenmodes can be interpreted as the slow motions of the protein around the equilibrium state and the Moore-Penrose pseudo-inverse matrix can be used to predict the protein thermal factors, or B-factors. However, NMA interaction potentials are quite complicated. Tirion simplified its complexity by retaining only the harmonic potential for elasticity, which is the dominant term in the MD Hamiltonian [@Tirion:1996]. Network theory [@Flory:1976] has had a considerable impact in protein flexibility analysis. The combination of elasticity and coarse-grained network gives rise to elastic network model (ENM) [@Hinsen:1998]. In this spirit, Gaussian network model (GNM) [@Bahar:1997; @Bahar:1998; @QCui:2010] and anisotropic network model (ANM) [@Atilgan:2001] have been proposed. Yang et al. [@LWYang:2008] have shown that the GNM is about one order more efficient than most other flexibility approaches. The above graph theory based methods have been improved in a number of aspects, including crystal periodicity and cofactor corrections [@Kundu:2002; @Kondrashov:2007; @Hinsen:2008; @GSong:2007], and density - cluster rotational - translational blocking [@Demerdash:2012]. These approaches have many applications in biophysics, including stability analysis [@Livesay:2004], molecular docking [@Gerek:2010], and viral capsid analysis [@Rader:2005; @Tama:2005]. In particular, based on spectral graph theory that the behavior of the second eigenmode can be used for clustering, these methods have been utilized to unveil the molecular mechanism of the protein domain motions of hemoglobin [@CXu:2003], F1 ATPase [@WZheng:2003; @QCui:2004], chaperonin GroEL [@Keskin:2002; @WZheng:2007] and the ribosome [@Tama:2003; @YWang:2004]. The reader is referred to reviews for more details [@JMa:2005; @LWYang:2008; @Skjaven:2009; @QCui:2010].
Note that ENM type of methods is still too expensive for analyzing subcellular organelles and multiportein complexes, such as HIV and Zika virus, and molecular motors, due to their matrix decomposition procedure which is of the order of ${\cal O}(N^3)$ in computational complexity, where $N$ is the number of network nodes, or protein atoms. An interesting and important mathematical issue is how to reduce the computational complexity of ENM, GNM and ANM for handling excessively large biomolecules. Flexibility-rigidity index (FRI) has been developed as a more accurate and efficient method for biomolecular graph analysis [@KLXia:2013d; @Opron:2014]. In particular, aided with a cell lists algorithm [@Allen:1987], the fast FRI (fFRI) is about ten percent more accurate than GNM on a test set of 364 proteins and is orders of magnitude faster than GNM on a set of 44 proteins, due to its ${\cal O}(N)$ computational complexity. It has been demonstrated that fFRI is able to predict the B-factors of an entire HIV virus capsid with 313,236 residues in less than 30 seconds on a single-core processor, which would require GNM more than 120 years to accomplish if its computer memory were not a problem [@Opron:2014].
Topological analysis of molecules has become very popular since the introduction of the theory of atoms in molecules (AIM) for molecular electron density data by Bader and coworkers [@Bader:1985; @Bader:1990]. AIM was proposed to quantitatively define the atomic bonds and interatomic surfaces (IASs) by employing a topology based partition of electron density. It has two main strands: the scalar field topology of molecular electron density maps and the scalar field topology of the local Laplacian of electron density [@Popelier:2000; @Popelier:2005]. The former characterizes chemical bonds and atoms, and the latter provides a new procedure to analyze electron pair localization. Electron localization function (ELF) [@Silvi:1994] was proposed for the study of electron pairing. ELF utilizes the gradient vector field topology to partition the electron density map into topological basins. In fact, a general theory called quantum chemical topology (QCT) [@Popelier:2005] has been developed for the topological analysis of electron density functions. Apart form the above mentioned AIM and ELF, QCT also includes the electrostatic potential [@Leboeuf:1999], electron localizability indicator (ELI) [@Kohout:2004], localized orbital locator (LOL) [@Schmider:2000], the virial field [@Keith:1996], the magnetically induced current distribution [@Keith:1993], the total energy (catchment regions) [@Mezey:1981] and the intracule density [@Cioslowski:1999]. QCT has proved to be very effective in analyzing interactions between atoms in molecular systems, particularly the covalent interactions and chemical structure of small molecular systems. Many software packages have been developed for QCT analysis [@Biegler:2002; @Henkelman:2006].
The essential idea behind QCT is the scalar field topology analysis [@Beketayev:2011; @Gunther:2014], which includes several major components such as critical points (CPs) and their classification, zero-flux interface (interatomic interface), gradient vector field, etc. In fact, when vector field topology is applied to the gradient of a scalar function, it coincides with scalar field topology. Mathematically, this topological analysis is also known as the Morse theory, which describes the topological structure of a closed manifold by means of a nondegenerate gradient vector field. Morse theory is a powerful tool for studying the topology of molecular structural data through critical points of a Morse function. A well-defined Morse function needs to be differentiable and its CPs are isolated and non-degenerated. It can be noticed that all the above-mentioned scalar fields in QCT are Morse functions and more can be proposed as long as they satisfy the Morse function constraints. In the Morse theory, critical points are classified into minima, maxima, and saddle points based on their indices. In AIM, the three types of CPs are associated with chemical meanings. A maximal CP is called a nucleic critical point (NCP). A minimal CP is related with cage critical point (CCP). Finally, saddle points can be further classified into two types, i.e., bond critical points (BCPs) and ring critical points (RCPs). Current research issues in QCT include how to reduce the computational complexity and extend this approach for biomolecules [@Gillet:2012; @Gunther:2014]. Additionally, its connection to scalar field topology and vector field topology needs to be further clarified so that related mathematical theories, including Poincaré - Hopf theorem [@Poincare:1890], Conley index theory [@Conley:1978], Floer homology, etc., and algorithms developed in computer science can be better applied to molecular sciences.
In additional to differential topology, algebraic topology, specifically, persistent homology, has drawn much attention in recent years. Persistent homology has been developed as a new multiscale representation of topological features. The 0th dimensional version was originally introduced for computer vision applications under the name “size function" [@Fro90; @Frosini:1999] and the idea was also studied by Robins [@Robins:1999]. Persistent homology theory was formulated, together with an algorithm given, by Edelsbrunner et al. [@Edelsbrunner:2002], and a more general theory was developed by Zomorodian and Carlsson [@Zomorodian:2005]. There has since been significant theoretical development [@BH11; @CEH07; @CEH09; @CEHM09; @CCG09; @CGOS11; @Carlsson:2009theory; @CSM09; @SMV11; @zigzag], as well as various computational algorithms [@OS13; @DFW14; @Mischaikow:2013; @javaPlex; @Perseus; @Dipha]. Often, persistent homology can be visualized through barcodes [@CZOG05; @Ghrist:2008], in which various horizontal line segments or bars are the homology generators that survive over filtration scales. Persistence diagrams are another equivalent representation [@edelsbrunner:2010]. [ Computational homology and persistent homology have been applied to a variety of domains, including image analysis [@Carlsson:2008; @Pachauri:2011; @Singh:2008; @Bendich:2010; @Frosini:2013], chaotic dynamics verification [@Mischaikow:1999; @kaczynski:mischaikow:mrozek:04], sensor network [@Silva:2005], complex network [@LeeH:2012; @Horak:2009], data analysis [@Carlsson:2009; @Niyogi:2011; @BeiWang:2011; @Rieck:2012; @XuLiu:2012], shape recognition [@DiFabio:2011; @AEHW06] and computational biology [@Kasson:2007; @Gameiro:2014; @Dabaghian:2012; @Perea:2015a; @Perea:2015b].]{} Compared with traditional computational topology [@Krishnamoorthy:2007; @YaoY:2009; @ChangHW:2013] and/or computational homology, persistent homology inherently has an additional dimension, the filtration parameter, which can be utilized to embed some crucial geometric or quantitative information into topological invariants. The importance of retaining geometric information in topological analysis has been recognized [@Biasotti:2008], and topology has been advocated as a new approach for tackling big datasets [@BVP15; @BHPP14; @Fujishiro:2000; @Carlsson:2009; @Ghrist:2008]. Most recently, persistent homology has been developed as a powerful tool for analyzing biomolecular topological fingerprints [@KLXia:2014c; @KLXia:2015d; @KLXia:2015e], quantitative fullerene stability analysis [@KLXia:2015a], topological transition in protein folding [@KLXia:2015c], cryo-EM structure determination [@KLXia:2015b], and in conjugation with machine learning for protein classification [@ZXCang:2015] and protein-ligand/drug binding affinity prediction [@ZXCang:2016b]. Differential geometry based topological persistence [@BaoWang:2016a] and multidimensional persistence [@KLXia:2015c] have also been developed for biomolecules analysis and modeling. It is worthy to mention that persistent topology along is able to outperform all the eminent methods in computational biophysics for the blind binding affinity prediction of protein-ligand complexes from massive data sets [@ZXCang:2016b].
The objective of this paper is threefold. First, the main objective is to provide a review of some widely used geometric, topological and graph theory apparatuses for the modeling and analysis of biomolecular data. We keep our description concise, elementary and accessible to upper level undergraduate students in mathematics and most researchers in computational biophysics. We point out some open problems and potential topics in our discussions. Our goal is to provide a reference for mathematicians who are interested in mathematical molecular bioscience and biophysics (MMBB), an emergent field in mathematics [@Wei:2016], and for biophysicists and theoreticians who are interested in mathematical foundations of many theoretical approaches in molecular biology and biophysics. Obviously, our topic selection is limited by our knowledge, experience and understanding, and for the same reason, we might have missed many important results and references on the selected topics as well. Additionally, inspired by the success of QCT and persistent homology, the density filtration for Hessian matrix eigenvalue maps and molecular curvature maps has been introduced. Both maps are constructed from molecular rigidity density obtained via a discrete to continuum mapping (DCM) technique, which transfers atomic information in a molecule to atomic density distribution, a continuous scalar function. In this approach, a series of isosurfaces are generated and systematically studied for eigenvalue and curvature maps. Geometric and topological (Geo-Topo) fingerprints are identified to characterize unique patterns within eigenvalue and curvature maps, specifically, the maps of three eigenvalues derived from local Hessian matrix at each location of the rigidity density and the maps of Gaussian, mean, maximal and minimal curvatures computed everywhere of the rigidity density. Topological properties of eigenvalue and curvature maps are classified by their critical points. The evolution of isosurfaces during the filtration process is found to be well characterized by CPs. Different behaviors are found in different types of maps. Persistent homology analysis is also employed for eigenvalue map analysis to reveal intrinsic topological invariants of three Hessian matrix eigenvalues. Finally, a new minimal molecular surface, called analytical minimal molecular surface (AMMS) via the zero-value isosurface of the mean curvature map, has been introduced. It is found that this new surface definition can capture the topological property, such as the inner bond information. It also offers an efficient geometric modeling of biomolecules.
The rest of this paper is organized as following: Section \[sec:discrete\] is devoted to a review of some discrete mathematical apparatuses, namely, graph theory and persistent homology, for the analysis and modeling of biomolecular data. More specifically, we illustrate the applications of graph theory, Gaussian network model, anisotropic network model, normal mode analysis, flexibility and rigidity index, spectral graph theory, and persistent homology to biomolecular data analysis, such as protein B-factor prediction, domain separation, anisotropic motion, topological fingerprints, etc. The review of some continuous geometric and topological apparatuses for scalar field topology and geometry are given in Section \[sec:continuous\]. We examine the basic concepts in differential geometry, biomolecular surfaces, curvature analysis, and theory of atoms in molecules. Discrete to continuum mapping and two algorithms for curvature evaluation are discussed. Further, we introduce two new approaches, i.e., Hessian matrix eigenvalue maps and curvature maps, for geometric-topological fingerprint analysis of biomolecular data. The relation between scalar field geometry and topological CPs are discussed in detail. A new analytical minimal molecular surface is also introduced. Virtual particle model is proposed to analyze the anisotropic motions of continuous scalar fields, such as cryo-EM maps. Finally, persistent homology analysis for eigenvalue scalar field is discussed. This paper ends with some concluding remarks.
Discrete apparatuses for biomolecules {#sec:discrete}
=====================================
One of the major challenges in the biological sciences is the prediction of protein functions from protein structures. One function prediction is about protein flexibility, which strongly correlates with biomolecular enzymatic activities, such as allosteric transition, ligand binding and catalysis, as well as protein stiffness and rigidity for structural supporting. For instance, in enzymatic processes, protein flexibility enhances protein-protein interactions, which in turn reduce the activation energy barrier. Additionally, protein flexibility and motion amplify the probability of barrier crossing in enzymatic chemical reactions. Therefore, the investigation of protein flexibility at a variety of energy spectra and time scales is vital to the understanding and prediction of other protein functions. Currently, the most important technique for protein flexibility analysis is X-ray crystallography. Among more than one hundred twenty thousand structures in the protein data bank (PDB), more than eighty percent structures are collected by X-ray crystallography. The Debye-Waller factor, or B-factor, can be directly computed from X-ray diffraction or other diffraction data. In the PDB, biomolecular structures are recorded in terms of (discrete) atomic types, atomic positions, occupation numbers, and B-factors. Although atomic B-factors are directly associated with atomic flexibility, they can be influenced by variations in atomic diffraction cross sections and chemical stability during the diffraction data collection. Therefore, only the B-factors for specific types of atoms, say C$_\alpha$ atoms, can be directly interpreted as their relative flexibility without corrections. Another important method for accessing protein flexibility is nuclear magnetic resonance (NMR) which often provides structural flexibility information under physiological conditions. NMR spectroscopy allows the characterization of protein flexibility in diverse spatial dimensions and a large range of time scales. About six percent of structures in the PDB are determined by electron microscopy (EM) which does not directly offer the flexibility information at present. Therefore, it is important to have mathematical or biophysical methods to predict their flexibility.
Graph theory related methodologies {#sec:graph}
----------------------------------
With the development of experimental tools, vast amount of data for biomolecular structures and interaction networks are available and provide us with unprecedented opportunities in mathematical modeling. The graph theory and network models have been widely used in the study of biomolecular structures and interactions and found many applications in drug design, protein function analysis, gene identification, RNA structure representation, etc. [@DasGupta2016; @Gan:2004rag; @Fera:2004rag] Generally speaking, biomolecular graph and network models can be divided into two major categories, namely, abstract graph/network models, which include biomolecular interaction-network models, and geometric graph/network models, where the distance geometry plays an important role.
The geometric graph models or biomolecular structure graph models construct unique graphs based on biomolecular 3D structural data. The graph theory is then employed to analyze biomolecular properties in four major aspects: flexibility and rigidity analysis, protein mode analysis, protein domain decomposition and biomolecular nonlinear dynamics. Many other network based approaches, including GNM [@Bahar:1997; @Bahar:1998] and ANM [@Atilgan:2001], have been developed for protein flexibility analysis. More recently, FRI has been proposed as a matrix-decomposition-free method for flexibility analysis, [@KLXia:2013d; @Opron:2014]. The fundamental assumptions of the FRI method are as follows. Protein functions, such as flexibility, rigidity, and energy, are fully determined by the structure of the protein and its environment, and the protein structure is in turn determined by the relevant interactions. Therefore, whenever a native protein structure is available, there is no need to analyze protein flexibility and rigidity by tracing back to the protein interaction Hamiltonian. Consequently, the FRI bypasses the ${\cal O}(N^3)$ matrix diagonalization. In fact, FRI does not even require the 3D geometric information of the protein structure. It assesses graphic connectivity of the protein distance geometry and analyzes the geometric compactness of the protein structure. It can be regarded as a kernel generalization of the local density model [@Halle:2002; @DWLi:2009; @CPLin:2008]. Another very important application of biomolecular structure graph model is the protein mode analysis. As stated above, the low order eigenmodes provide information of the protein dynamics at equilibrium state. Normal mode analysis (NMA) plays important roles in mode analysis. However, its potential function involves too many interactions and it is very inefficient for large biomolecular systems. Anisotropic network model (ANM) has dramatically reduced the complexity of the potential function by representing the biological macromolecule as an elastic mass-and-spring network. In the network each node is a C$_\alpha$ atom of the associated residue and springs represent the interactions between the nodes. The overall potential is the sum of harmonic potentials between interacting nodes. To describe the internal motions of the spring connecting two atoms, there is only one degree of freedom. Qualitatively, this corresponds to the compression and expansion of the spring in a direction given by the locations of the two atoms. In other words, ANM is an extension of the GNM to three coordinates per atom, thus accounting for directionality.
The biomolecular structure graph models can also be used in protein domain decomposition. The biomolecular structural domains are stable and compact units of the structure that can fold independent of the rest of the protein and perform a specific function. A domain usually contains a hydrophobic core and a protein is usually formed by the combination of two or several domains. There are many methods that decompose a protein structure into domains [@alexandrov:2003; @Guo:2003; @holm:1996; @SKundu:2004; @murzin:1995; @orengo:1997; @veretnik:2007]. Some of them are done manually through structure visualization. One of them is the structural classification of proteins (SCOP) database, where data are largely manually classified into protein structural domains based on similarities of their structures and amino acid sequences. With the surge of protein structure data, efficient and robust computational algorithms are developed. They have demonstrated a high level of consistency and robustness in the process of partitioning a structure into domains. Graph theory is also used in RNA structure analysis, particularly in RNA motif representation and RNA classification[@Gan:2004rag; @Fera:2004rag; @Kim2004:candidates]. Spectral graph theory is widely used for clustering. The essential idea is to study and explore graphs through the eigenvalues and eigenvectors of matrices naturally associated with these graphs.
Molecular nonlinear dynamics (MND) models can be naturally derived from biomolecular graph models [@KLXia:2014b]. Essentially, each node in the graph is an atom and represented by a nonlinear oscillator. These nonlinear oscillators are connected through the graph connectivity. In this manner, one can study protein structure and function through the nonlinear dynamics theory widely used in chaos, synchronization, stability, pattern formation, etc.
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Illustration of a graph. The associate adjacent, weight and Laplacian matrix can be found in Eq. (\[eq:graph\_matrix\]). []{data-label="fig:GRAPH"}](GRAPH.png "fig:"){width="30.00000%"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
### Elementary graph theory {#sec:Graph}
Graph theory deals with a set of discrete vertices (or atoms) and their connectivity (or bonds). Normally, an undirected graph $G$ can be denoted as a pair $G(V,E)$, where $V= \{v_i;i=1,2,...,N \}$ denotes its set of $N$ vertices (or protein atoms), $N=|V|$. Here $E=\{e_i=(v_{i_1},v_{i_2});1\leq i_1 \leq N, 1\leq i_2 \leq N \}$ denotes its set of edges, which can be understood as certain covalent or noncovalent bonds among atoms in a molecule. Each edge in $E$ is an unordered pair of vertices, with the edge connecting distinct vertices $v_{i_1}$ and $v_{i_2}$ written as $e_i=(v_{i_1},v_{i_2})$. Then the adjacency matrix $A$ of $G$ is given by [@ChungOverview; @Mohar:1991laplacian; @Mohar:1997some; @von:2007tutorial] $$\begin{aligned}
\label{eq:couple_matrix28}
A_{ij}=\begin{cases} \begin{array}{ll}
1 & (v_i,v_j) \in E\\
0 & (v_i,v_j) \not \in E.\\
\end{array}
\end{cases}\end{aligned}$$ The degree of a vertex $v_i$ is defined as $d_i=\sum_{i \neq j}^N A_{ij}$, which is the total number of edges that are connected to node $v_i$. The degree matrix $D$ can be defined as $$\begin{aligned}
\label{eq:couple_matrix27}
D_{ij}=\begin{cases} \begin{array}{ll}
\sum_{i \neq j}^N A_{ij} & i=j\\
0 & i \neq j.
\end{array}
\end{cases}\end{aligned}$$ With these two matrices, one can defined Laplacian matrix as $L=D-A$. The Laplacian matrix is also known as admittance matrix, Kirchhoff matrix or discrete Laplacian. It is widely used to represent a graph. More specifically, it can be expressed as, $$\begin{aligned}
\label{eq:couple_matrix26}
L_{ij}=\begin{cases} \begin{array}{ll}
-1 & i \neq j~{\rm and} ~ (v_i,v_j) \in E\\
-\sum_{i \neq j}^N L_{ij} &i=j\\
0 &{ \rm otherwise.}
\end{array}
\end{cases}\end{aligned}$$ For example, the adjacency, degree and Kirchhoff matrices for the graph in Fig. \[fig:GRAPH\] are, respectively [$$\begin{aligned}
\label{eq:graph_matrix}
A=\left( \begin{array}{llllllll}
0 &1 &0 &0 &0 &0 &0 &1 \\
1 &0 &1 &0 &0 &0 &0 &0 \\
0 &1 &0 &1 &0 &0 &1 &0 \\
0 &0 &1 &0 &1 &0 &1 &0 \\
0 &0 &0 &1 &0 &1 &0 &0 \\
0 &0 &0 &0 &1 &0 &1 &0 \\
0 &0 &1 &1 &0 &1 &0 &1 \\
1 &0 &0 &0 &0 &0 &1 &0
\end{array}
\right),
D=\left( \begin{array}{llllllll}
2 &0 &0 &0 &0 &0 &0 &0 \\
0 &2 &0 &0 &0 &0 &0 &0 \\
0 &0 &3 &0 &0 &0 &0 &0 \\
0 &0 &0 &3 &0 &0 &0 &0 \\
0 &0 &0 &0 &2 &0 &0 &0 \\
0 &0 &0 &0 &0 &2 &0 &0 \\
0 &0 &0 &0 &0 &0 &4 &0 \\
0 &0 &0 &0 &0 &0 &0 &2
\end{array}
\right),
L=\left( \begin{array}{llllllll}
2 &$ -1$ &0 &0 &0 &0 &0 &$-1$ \\
$ -1$ &2 &$-1$ &0 &0 &0 &0 &0 \\
0 &$ -1$ &3 &$-1$ &0 &0 &$-1$ &0 \\
0 &0 &$-1$ &3 &$-1$ &0 &$-1$ &0 \\
0 &0 &0 &$-1$ &2 &-1 &0 &0 \\
0 &0 &0 &0 &$-1$ &2 &$-1$ &0 \\
0 &0 &$-1$ &$-1$ &0 &$-1$ &4 &$-1$ \\
$-1$ &0 &0 &0 &0 &0 &$-1$ &2
\end{array}
\right)\end{aligned}$$ ]{} The Laplacian matrix has several basic properties. It is a symmetric and semi-positive definite. The rank of the Laplacian matrix is $N-N_0$ with $N_0$ the number of connected components. Its second smallest eigenvalue is known as the algebraic connectivity (or Fiedler value)[@ChungOverview; @von:2007tutorial].
More generally, one can assign weights to edges to construct a weighted graph $G(V,E,W)$. Here $G(V,E)$ is the associated unweighted graph, and $W=\{w_{ij}; 1\leq i \leq N, 1\leq j \leq N, w_{ij}\geq 0\}$ is the weighted adjacent matrix. The weight is also known as pairwise distance or pairwise affinity. The new degree of vertex $v_i$ is $d_i=\sum_{j=1}^N w_{ij}$. The weighted degree matrix $D$ and weighted Laplacian matrix $L$ can be defined accordingly.
Normally, a graph structure is not given. Instead, one may have the information of nodes and general weight functions. In this case there are several general ways to construct a graph[@von:2007tutorial]:
1. $\epsilon$-neighborhood graph: connect all points whose pairwise distances are smaller than $\epsilon$;
2. $k$-nearest neighbor graph: connect vertex $v_i$ with vertex $v_j$, if $v_j$ is among the $k$-nearest neighbors of $v_i$; and
3. fully connected graph: connect all points with positive similarity with each other.
In biomolecular structure graph models, coordinates for atoms in molecules are available. Therefore, distances and distance-based functions can be used to construct structure graphs. The simplest way is to use a cutoff distance $r_c$ and build up edges between atoms or residues within the cutoff distance only. This approach has been used in GNM, which is an important tool for the study of protein flexibility and rigidity.
To unify the notation, in the following discussion, one can consider an $N$-particle representation of a biomolecule. Here a particle can be an ordinary atom in a full atomic representation or a C$_\alpha$ atom in a coarse-grained representation. One can denote $\{ {\bf r}_{i}| {\bf r}_{i}\in \mathbb{R}^{3}, i=1,2,\cdots, N\}$ the coordinates of these particles and $r_{ij}=\|{\bf r}_i-{\bf r}_j\|$ the Euclidean space distance between $i$th and $j$th particles. More specifically, the coordinate is a position vector ${\bf r}_i=(x_i, y_i, z_i)$.
### Gaussian network model (GNM) {#sec:GNM}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Illustration of a weighted graph Laplacian matrix for HIV capsid protein 1E6J. Left: six subdomains. Right: correlation map for residue C$_{\alpha}$ atoms indicating domain separations.[]{data-label="fig:correlation_matrix"}](correlation_matrix.png "fig:"){width="60.00000%"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Gaussian network model (GNM) [@Bahar:1997; @Bahar:1998; @QCui:2010; @LiGH:2002; @Yang:2006] can be viewed as a special graph model using the Kirchhoff matrix. It was proposed for biomolecular flexibility and long-time scale dynamics analysis, particularly, the prediction of the Debye-Waller factor or B-factor. Experimentally, B-factor is an indication of the relative thermal fluctuations of different parts of a structure. Atoms with small B-factors belong to a part of the structure that is very rigid. Atoms with large B-factors generally belong to part of the structure that is very flexible. The B-factor information can be found in the structural data downloaded from the Protein Data Bank (PDB).
As stated above, the graph or network in GNM is constructed by using a cutoff distance $r_c$. If the distance between two atoms are less than the cutoff distance, an edge is formed between them. Otherwise, no edge is built. The corresponding discrete Laplacian matrix describes the relative connectivity within a protein structure, and thus, it is also called a connectivity matrix. $$\begin{aligned}
\label{eq:couple_matrix25}
L_{ij}=\begin{cases} \begin{array}{ll}
-1 & i \neq j~{\rm and} ~ r_{ij} \leq r_c\\
-\sum_{i \neq j}^N L_{ij} &i=j\\
0 &{ \rm otherwise}
\end{array}.
\end{cases}\end{aligned}$$
In a nutshell, the GNM prediction of the $i$th B-factor of the biomolecule can be expressed as [@Bahar:1997; @Bahar:1998; @JKPark:2013] $$\begin{aligned}
\label{eqn:GNM}
B_i^{\rm GNM}=a \left(L^{-1} \right)_{ii}, \forall i=1,2,\cdots, N,\end{aligned}$$ where $a$ is a fitting parameter that can be related to the thermal energy and $\left(L^{-1} \right)_{ii}$ is the $i$th diagonal element of the Moore-Penrose pseudo-inverse of graph Laplacian matrix $L$. More specifically, $\left(L^{-1} \right)_{ii}=\sum_{k=2}^N \lambda_k^{-1}\left[{\bf q}_k {\bf q}_k^T \right]_{ii}$, where $T$ denotes the transpose and $\lambda_k$ and ${\bf q}_k$ are the $k$th eigenvalue and eigenvector of $\Gamma$, respectively. The summation omits the first eignmode whose eigenvalue is zero.
### Anisotropic network model (ANM) {#sec:ANM}
In Gaussian network model [@Bahar:1997; @Bahar:1998; @QCui:2010; @LiGH:2002; @Yang:2006], only the distance information is used with no consideration about the anisotropic properties in different directions. It should be noticed that in GNM, the Kirchhoff matrix is of the dimension $N*N$ with $N$ being the total number of atoms. In order to introduce the anisotropic information, one has to discriminate the distance between atoms in three different directions. To this end, at each label of $ij$, a local $3*3$ Hessian matrix is constructed [@Atilgan:2001] $$\begin{aligned}
\label{eq:multi-kirchoff1}
H_{ij} = -\frac{1}{r_{ij}^2}\left[ \begin{array}{ccc}
(x_j-x_i)(x_j-x_i) &(x_j-x_i)(y_j-y_i) &(x_j-x_i)(z_j-z_i)\\
(y_j-y_i)(x_j-x_i) &(y_j-y_i)(y_j-y_i) &(y_j-y_i)(z_j-z_i)\\
(z_j-z_i)(x_j-x_i) &(z_j-z_i)(y_j-y_i) &(z_j-z_i)(z_j-z_i)
\end{array}\right] ~ \forall ~ i \neq j ~{\rm and} ~ r_{ij}\leq r_c.
\end{aligned}$$ As the same in the GNM, the diagonal part is the negative summation of the off diagonal elements: $$\begin{aligned}
\label{eq:multi-kirchoff1_diagonal}
H_{ii} = -\sum_{i\neq j} H_{ij}.
\end{aligned}$$ This approach, called anisotropic network model (ANM), can be used to generate the anisotropic motion of biomolecules. It is noticed that the dimension of the Hessian matrix is no longer $N*N$, instead it is $3N * 3N$. The dimension of an eigenvector is $3N$. Therefore, for each atom, one now has a vector associated with it, which gives an direction in the ${\mathbb R}^3$. The norm of this vector gives a relative amplitude. This eigenvector is also called eigenmode. It describes the relative motion of the protein near its equilibrium state.
#### Generalized GNM and generalized ANM
In both Gaussian network model and anisotropic network model, a cutoff distance is used to construct their connectivity matrices, i.e., Laplacian matrix and Hessian matrix, respectively. However, physically, the correlation between any two particles normally decays with respect to distance. To account for this effect, a correlation function $\Phi(r_{ij}; \eta_{ij}) $ is introduced. In general, it is a real-valued monotonically decreasing radial basis function satisfying [@KLXia:2013d; @Opron:2014], $$\begin{aligned}
\label{eq:couple_matrix1-1}
\Phi( r_{ij};\eta_{ii})&=&1 \\ \label{eq:couple_matrix1-12}
\Phi( r_{ij};\eta_{ij})&=&0 \quad {\rm as }\quad r_{ij} \rightarrow\infty.\end{aligned}$$ In this function, the parameter $\eta_{ij}$ is a characteristic distance between particles $v_i$ and $v_j$. It can also be simplified to atomic parameter $\eta_{j}$, which depends only on the atomic type. In coarse-grained models, only $C_{\alpha}$ atom is considered. Therefore, one can further simplify it to a constant $\eta$. This parameter can also be viewed as a resolution parameter.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Protein domain separation of protein 3PGK C$_\alpha$ atoms using a new spectral clustering method, gGNM. The separation is carried out with the eigenvector corresponding to the second lowest eigenvalue.[]{data-label="fig:cp1"}](3PGK.png "fig:"){width="40.00000%"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Delta sequences of the positive type discussed in an earlier work [@GWei:2000] are all good choices. For example, one can use generalized exponential functions $$\begin{aligned}
\label{eq:couple_matrix1}
\Phi(r_{ij};\eta_{ij}) = e^{-\left(r_{ij}/\eta_{ij}\right)^\kappa}, \quad \kappa >0\end{aligned}$$ or generalized Lorentz functions $$\begin{aligned}
\label{eq:couple_matrix24}
\Phi(r_{ij};\eta_{ij}) = \frac{1}{1+ \left( r_{ij}/\eta_{ij}\right)^{\upsilon}}, \quad \upsilon >0.
\end{aligned}$$ Using these correlation functions, one can obtain a weighted graph representation or weighted graph Laplacian as [@KLXia:2015f], $$\begin{aligned}
\label{eq:couple_matrix23}
L_{ij}=\begin{cases} \begin{array}{ll}
-\Phi(r_{ij};\eta_{ij}) & i \neq j\\
- \sum_{i \neq j}^N L_{ij} &i=j\\
\end{array}.
\end{cases}\end{aligned}$$ It is found that the weighted graph can deliver a better prediction of B-factors. This weighted graph approach is called the generalized GNM (gGNM). Figure \[fig:cp1\] shows the protein domain separation obtained with gGNM.
The local Hessian matrix in Eq. (\[eq:multi-kirchoff1\]) can also be generated to consider the distance effect to obtain a generalized form [@KLXia:2015f], $$\begin{aligned}
\label{eq:multi-kirchoff12}
H_{ij} = -\frac{\Phi( r_{ij};\eta_{ij})}{r_{ij}^2}\left[ \begin{array}{ccc}
(x_j-x_i)(x_j-x_i) &(x_j-x_i)(y_j-y_i) &(x_j-x_i)(z_j-z_i)\\
(y_j-y_i)(x_j-x_i) &(y_j-y_i)(y_j-y_i) &(y_j-y_i)(z_j-z_i)\\
(z_j-z_i)(x_j-x_i) &(z_j-z_i)(y_j-y_i) &(z_j-z_i)(z_j-z_i)
\end{array}\right] ~ \forall ~ i \neq j.
\end{aligned}$$ Again the diagonal part is the negative summation of the off-diagonal elements the same as Eq. (\[eq:multi-kirchoff1\_diagonal\]). Note that Hinsen [@Hinsen:1998] has proposed a special case: $\Phi( r_{ij};\eta_{ij} )= e^{-\left(\frac{r_{ij}}{\eta }\right)^2}$, where $\eta$ is a constant. It was shown that gGNM and generalized anisotropic network model (gANM) outperform the original GNM and ANM respectively in B-factor predictions [@KLXia:2015f]. Figure \[fig:2ABH\] illustrates an eigenmode for protein 2ABH obtained with gANM. There will be a continuous interest in design new and optimal graph theory approaches for biomolecular analysis.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Illustration of eigenmode for protein 2ABH. The eigenmode can be used to describe the biomolecular dynamics near equilibrium state. The eigenmode is generated by the generalized anisotropic normal model method.[]{data-label="fig:2ABH"}](2ABH.png "fig:"){width="30.00000%"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
### Normal mode analysis (NMA) and quasi-harmonic analysis
Normal mode analysis (NMA) is one of the major tools for the study of biomolecular motions [@Go:1983; @Levitt:1983normal; @Brooks:1983harmonic; @Lopez:2014normal; @Hayward:2008normal]. It is found that protein normal modes with the largest fluctuation or the lowest frequency are functionally relevant. Mathematically, NMA has its root in harmonic analysis. It assumes that conformational energy surface at an energy minimum can be approximated by some harmonic functions.
In normal mode analysis, one usually needs the atomic coordinates and a force field describing the interactions between constituent atoms. Typically, there are three major steps in applying NMA [@Hayward:2008normal]. Firstly, one needs to carry out molecular dynamics simulations to minimize the conformational potential energy to obtain an equilibrium state. Secondly, one needs to calculate the second derivatives of the potential energy to construct the Hessian matrix. Finally, one needs to perform the eigenvalue decomposition of the Hessian matrix.
Originally, NMA uses exactly the same force fields as used in molecular dynamics simulations. Due to the computational inefficiency, elastic network models (ENMs) was proposed. Generally speaking, ENM is just the NMA with a simplified force field and associated coarse-grained representation. It has two major advantages. Firstly, there is no need for energy minimization as the distances of all of the elastic connections are taken to be at their minimal energy lengths. Secondly, due to the coarse-grained representation, the eigenvalue decomposition is much efficient. Even through ENM is a much simplified model, it is found that ENM is able to reproduce the NMA results with a respectable degree of similarity.
#### Standard NMA
For a mechanical system consisting of $N$ atoms ${\bf r}=({r}_1,{ r}_2,\cdots, { r}_{3N})$, its Hamiltonian $\mathcal{H}({\bf r})$ is given by the sum of kinetic energy $\mathcal{K}({\bf r})$ and potential energy $\mathcal{U}({\bf r})$: $$\begin{aligned}
\mathcal{H}({\bf r})=\mathcal{K}({\bf r})+\mathcal{U}({\bf r}).\end{aligned}$$
If the structure has an equilibrium conformation ${\bf r}^0=({r}_1^0,{ r}_2^0,...,{ r}_{3N}^0)$, one can have the Taylor expansion of the potential energy $$\begin{aligned}
\mathcal{U}({\bf r}) \approx \mathcal{U}({\bf r}^0)+ \sum_{i}^{3N}\frac{\partial \mathcal{U}}{\partial r_i} \Big| _{\bf r=r^0}(r_i-r_i^0)+ \frac{1}{2}\sum_{i}^{3N}\sum_{j}^{3N}\frac{\partial^2 \mathcal{U}}{\partial r_i \partial r_j} \Big|_{\bf r=r^0}(r_i-r_i^0)(r_j-r_j^0)+\cdots.\end{aligned}$$ Since the biomolecular system achieves a minimum of the energy at the equilibrium conformation ${\bf r}^0$, the related derivative functions $\frac{\partial \mathcal{U}}{\partial r_i}\Big| _{\bf r=r^0}$ vanishes. If one uses the mass-weighted coordinates $X_i=m_i^{\frac{1}{2}} (r_i-r_i^0)$, the potential function becomes $$\begin{aligned}
&& \mathcal{U}({\bf r}) \approx \frac{1}{2}\sum_{i}^{3N}\sum_{j}^{3N}\frac{\partial^2 \mathcal{U}}{\partial r_i \partial r_j} \Big|_{\bf r=r^0}(r_i-r_i^0)(r_j-r_j^0) \\ \nonumber
&& \qquad \approx \frac{1}{2}\sum_{i}^{3N}\sum_{j}^{3N}\frac{\partial^2 \mathcal{U}}{\partial r_i \partial r_j} \Big|_{\bf r=r^0}(r_i-r_i^0)(r_j-r_j^0).\end{aligned}$$ The related kinetic energy is $$\begin{aligned}
\mathcal{K}({\bf r})=\frac{1}{2}\sum_{i}^{3N}m_i\left(\frac{dr_i}{dt}\right)^2=\frac{1}{2}\sum_{i}^{3N}\dot{X}_i^2.\end{aligned}$$ The Hamiltonian is $$\begin{aligned}
\mathcal{H}({\bf r}) \approx \frac{1}{2}\sum_{i}^{3N}\sum_{j}^{3N} \dot{X}_i^2+\frac{1}{2}\sum_{i}^{3N}\sum_{j}^{3N}\frac{\partial^2\mathcal{ U}}{\partial X_i \partial X_j} \Big|_{ X=X^0}(X_i-X_i^0)(X_j-X_j^0),\end{aligned}$$ where $\dot X$ indicates the derivative of $ X$ with respect to time.
It can be seen that the oscillatory motions in this system are coupled and thus the movement of one atom depends on that of others. However, one can decompose the motion into independent harmonic oscillators with an appropriate normal mode coordinates. This is done by the eigenvalue decomposition of the Hessian matrix $H=Q\Lambda Q^T$. Here $H$ is obtained from the second order derivative of the potential function. The matrix $Q=\{{ {\bf q}_1,{\bf q}_2,...,{\bf q}_{3N}} \}$ contains the eigenvectors and $Q^T Q=I$. The diagonal matrix $\Lambda$ contains the corresponding eigenvalues. The mass-weighted Cartesian and normal mode coordinates are linearly related by $X=QY$. Finally, the Hamiltonian is expressed in the form, $$\begin{aligned}
\mathcal{H} \approx \frac{1}{2}\sum_{i}^{3N} \dot{Y}_i^2+\frac{1}{2}\sum_{i}^{3N}\lambda_i {Y}_i^2.\end{aligned}$$
#### Essential dynamics and quasi-harmonic analysis
Due to the complexity of biomolecular systems, it is notoriously difficult to carry out molecular dynamics simulations over the relevant biological time scales. However, it has been found that the vast majority of protein dynamics can be described by a surprisingly low number of collective degrees of freedom. In this manner, a principal components analysis (PCA) is often employed to analyze the simulation results[@Garcia:1992large; @Kitao:1991effects]. Mathematically, similar to NMA, PCA also employs the eigenvalue decomposition as it assume that that the major collective modes of fluctuation dominate the functional dynamics. In contrast to NMA, PCA of a molecular dynamics simulation trajectory does not rest on the assumption of a harmonic potential. Modes in PCA are usually sorted according to variance rather than frequency. As the collective motion is highly related to biomolecular functions, the dynamics in the low-dimensional subspace spanned by these modes was termed “essential dynamics” [@Amadei:1993essential]. A major advantage of PCA is that individual modes can be visualized and studied separately.
PCA on the mass weighted MD trajectory is also called quasi-harmonic analysis [@Brooks:1995harmonic], which typically consists of three steps [@Hayward:2008normal]. Firstly, one can superimpose all biomolecular configurations from the simulation trajectory to remove the internal rotation and translation. Secondly, one can perform an average over the regularized trajectory to construct a covariance matrix. Thirdly, an eigenvalue decomposition is employed on the covariance matrix. The original trajectory can then be analyzed in terms of principal components.
For a $3N$-dimensional vector trajectory $\bf {r}(t)$, the correlation between atomic motions can be expressed in the covariance matrix $C$: $$\begin{aligned}
C={\rm cov}({\bf r})=<\left({\bf r}(t)-<{\bf r}(t)>\right)\cdot\left({\bf r}(t)-<{\bf r}(t)>\right)>\end{aligned}$$ where ${\rm cov}$ is the statistical covariance and $< >$ denote the average over time. The correlation matrix is symmetric and can be diagonalized by an orthogonal transformation, $$\begin{aligned}
C=Q'\Lambda (Q')^T\end{aligned}$$ with $Q'=\{{\bf q}'_1,{\bf q}'_2,...,{\bf q}'_{3N} \}$ being eigenvectors. The original configurations can be projected into principal components ${\bf q}_i$, i.e., ${q}'_i(t)=({\bf r}(t)-<{\bf r}(t)>)\cdot {\bf q}_i$. For visualization, one can transform principal components into the Cartesian coordinates: $ {\bf r}'(t)={q}'_i(t){\bf q}_i+<{\bf r}(t)>$.
### Flexibility rigidity index (FRI) {#sec:FRI}
Due to the involved matrix diagonalization, the computational complexity of GNM is of the order of ${\cal O}(N^3)$, which is intractable for large biomolecules, such as viruses and subcellular organelles. Therefore, it is both important and desirable to have a method whose computational complexity scales as ${\cal O}(N^2)$ or better, as ${\cal O}(N)$. This order reduction is a standard mathematical issue and is mathematically challenging. However, by examining Eq. (\[eqn:GNM\]), one notices that what is used in the GNM theoretical prediction is the diagonal elements of the inverse of the graph Laplacian matrix. Mathematically, a good approximation is given by the inverse of the diagonal elements of the graph Laplacian matrix, providing that the matrix is diagonally dominant. Flexibility rigidity index (FRI) [@KLXia:2013d; @Opron:2014] is such a method and has several major characteristics. FRI provides a more straightforward and computationally-efficient way to predict B-factors. A major advantage of the FRI method is that it does not resort to mode decomposition and its computational complexity can be reduced to ${\cal O}(N)$ by means of the cell lists algorithm used in fast FRI (fFRI) [@Opron:2014]. The fundamental assumptions of the FRI method are as follows. Protein functions, such as flexibility, rigidity, and energy, are fully determined by the structure of the protein and its environment, and the protein structure is in turn determined by the relevant interactions. Therefore, whenever the protein structural data is available, there is no need to analyze protein flexibility and rigidity by tracing back to the protein interaction Hamiltonian. Consequently, the FRI bypasses the ${\cal O}(N^3)$ matrix diagonalization.
In a nutshell, the FRI prediction of the $i$th B-factor of the biomolecule can be given by [@KLXia:2013d; @Opron:2014] $$\begin{aligned}
\label{eqn:FRI}
B_i^{\rm FRI}=a \frac{1}{\sum_{j,j\neq i}^N w_j\Phi(r_{ij};\eta_{ij})} + b, \forall i=1,2,\cdots, N,\end{aligned}$$ where $a$ and $b$ are fitting parameters, $f_i=\frac{1}{\sum_{j,j\neq i}^N w_j\Phi(r_{ij};\eta_{ij})}$ is the $i$th flexibility index and $$\begin{aligned}
\label{rigidity1}
\mu_i=\sum_{j,j\neq i}^N w_j\Phi(r_{ij};\eta_{ij})\end{aligned}$$ is the $i$th rigidity index. Here, $w_j$ is an atomic number depended weight function that can be set to $w_j=1$ and $\eta_{ij}=\eta$ for a C$_{\alpha}$ network. The correlation function $\Phi( r_{ij};\eta)$ can be chosen from any monotonically decreasing function satisfying Eqs. (\[eq:couple\_matrix1\]) and (\[eq:couple\_matrix2\]). FRI was shown to outperform GNM and ANM in B-factor predictions based on hundreds of biomolecules [@KLXia:2013d; @Opron:2014].
#### Multiscale FRI {#sec:MFRI}
Biomolecules are inherently multiscale in nature due to their multiscale interactions. For example, proteins involve covalent bonds, hydrogen bonds, van der Walls bonds, electrostatic interactions, dipolar and quadrupole interactions, hydrophobic interactions, domain interactions, and protein-protein interactions. Therefore, their thermal motions are influenced by the multiscale interactions among their particles. Multiscale FRI (mFRI) was proposed to capture biomolecular multiscale behavior [@Opron:2015a]. Essential idea is to build multiscale kernels, i.e., kernels parametrized at multiple scales. Multiscale flexibility index can be expressed as $$\begin{aligned}
\label{eq:flexibility3}
f^{n}_i = \frac{1}{\sum_{j=1}^N w^{n}_{j} \Phi^{n}( \|{\bf r}_i - {\bf r}_j \|;\eta^{n} )},
\end{aligned}$$ where $w^{n}_{j}$, $\Phi^{n}( \|{\bf r}_i - {\bf r}_j \|;\eta^{n}) $ and $\eta^{n}$ are the corresponding quantities associated with the $n$th kernel. Then, one organizes these kernels in a multi-parameters minimization procedure $$\begin{aligned}
\label{eq:regression2}
{\rm Min}_{a^{n},b} \left\{ \sum_i \left| \sum_{n}a^n f^{n}_i + b-B^e_i\right|^2\right\}\end{aligned}$$ where $\{B^e_i\}$ are the experimental B-factors. In principle, all parameters can be optimized. For simplicity and computational efficiency, one only needs to determine $\{a^n\}$ and $b$ in the above minimization process. For each kernel $\Phi^n$, $w^n_j$ and $\eta^n_j$ will be selected according to the type of particles.
Specifically, for a simple C$_\alpha$ network (graph), one can set $w^n_j=1$ and choose a single kernel function parametrized at different scales. The predicted B-factors can be expressed as $$\begin{aligned}
\label{eq:flexibility4}
B^{\rm mFRI}_i = b+ \sum_{n=1}\frac{a^n}{\sum_{j=1}^N \Phi( \|{\bf r}_i - {\bf r}_j \|;\eta^{n} )}.
\end{aligned}$$ The difference between Eqs. (\[eq:flexibility3\]) and (\[eq:flexibility4\]) is that, in Eqs. (\[eq:flexibility3\]), both the kernel and the scale can be changed for different $n$. In contrast, in Eq. (\[eq:flexibility4\]), only the scale is changed. One can use a given kernel, such as $$\begin{aligned}
\label{eq:couple_matrixn}
\Phi(\|{\bf r} - {\bf r}_j \|;\eta^n) = \frac{1}{1+ \left( \|{\bf r} - {\bf r}_j \|/\eta^n\right)^{3}},
\end{aligned}$$ to achieve good multiscale predictions. It was demonstrated that mFRI is about 20% more accurate than GNM in the B-factor predictions [@Opron:2015a]. Parameters learned from mFRI were incorporated in GNM and ANM to create multiscale GNM (mGNM) and multiscale ANM (mANM) [@Opron:2015a].
#### Consistency between GNM and FRI {#sec:MFRI2}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![A comparison of B-factor prediction of protein 1CLL by various models, including flexibility rigidity index (FRI), Gaussian network model (GNM), generalized GNM (gGNM) and multiscale FRI (mFRI). Experimental results (Exp.) are given as a reference.[]{data-label="fig:bfactor"}](bfactor.png "fig:"){width="70.00000%"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
To further explore the relation between GNM and FRI, let us examine the parameter limits of generalized exponential functions (\[eq:couple\_matrix1\]) and generalized Lorentz functions (\[eq:couple\_matrix2\]) $$\begin{aligned}
\label{eq:Asmpt1}
e^{-\left( r_{ij} /\eta\right)^\kappa} \rightarrow \Phi( r_{ij};r_c) & {\rm as } & \kappa\rightarrow\infty\\ \label{eq:Asmpt2}
\frac{1}{1+ \left( r_{ij} /\eta\right)^{\upsilon}} \rightarrow \Phi( r_{ij};r_c) & {\rm as } &\upsilon\rightarrow\infty,\end{aligned}$$ where $r_c=\eta$ and $\Phi( r_{ij};r_c) $ is the ideal low-pass filter (ILF) used in the GNM Kirchhoff matrix $$\begin{aligned}
\label{eqn:IdealLF}
\Phi( r_{ij};r_c) = \begin{cases}\begin{array}{ll}
1, & r_{ij} \leq r_c \\
0, & r_{ij} > r_c \\
\end{array}
\end{cases}.\end{aligned}$$ Relations (\[eq:Asmpt1\]) and (\[eq:Asmpt2\]) unequivocally connect FRI correlation functions to the GNM Kirchhoff matrix.
It has been observed that GNM-ILF and FRI-ILF provide essentially identical predictions when the cutoff distance is equal to or larger than 20Å[@KLXia:2015f]. This phenomenon indicates that when the cutoff is sufficiently large, the diagonal elements of the gGNM inverse matrix and the direct inverse of the diagonal elements of the FRI correlation matrix become linearly strongly dependent. To understand this dependence at a large cutoff distance, an extreme case is considered when the cutoff distance is equal to or even larger than the protein size so that all the particles within the network are fully connected. In this situation, one can analytically calculate $i$th diagonal element of the GNM inverse matrix $$\begin{aligned}
\label{eqn:proprotion0}
\left(L (\Phi( r_{ij};r_c\rightarrow\infty)) \right)_{ii} = \frac{N-1}{N^2},\end{aligned}$$ and the FRI inverse of the $i$th diagonal element $$\begin{aligned}
\label{eqn:proprotion01}
\frac{1}{\sum_{j,j\neq i}^N \Phi(r_{ij};r_c\rightarrow\infty)}= \frac{1}{N-1}.\end{aligned}$$ These two expressions have the same asymptotic behavior as $N\rightarrow\infty$, which explains numerical results. However, mathematically, it is still an open problem to estimate the error bound between gGNM and FRI methods, i.e., the difference between the direct inverse of a diagonal element of a given weighted graph Laplacian matrix and the diagonal element of its matrix inverse.
A comparison of the performances of GNM, FRI, gGNM and mFRI is illustrated in Fig. \[fig:bfactor\]. The hinge around 75th residue is well captured by mFRI. Indeed, mFRI has the best accuracy in B-factor prediction based on the test of 364 proteins [@Opron:2015a]. The possible application to FRI in other systems, such as social networks, genetic networks, cellular networks and tissue networks, is still an open problem.
#### Anisotropic FRI
The anisotropic FRI (aFRI) has been proposed for mode analysis [@Opron:2014]. In this model, depending on one’s interest, the size of the Hessian matrix can vary from $3\times 3$ for a completely local aFRI to $3N\times 3N$ for a completely global aFRI. To construct such a Hessian matrix, one can partition all $N$ atoms in a molecule into a total of $M$ clusters $\{c_1,c_2,\dots,c_M\}$. Each cluster $c_k$ with $k=1,\dots,M$ has $N_k$ atoms so that $N=\sum_{k=1}^{M} N_k$. For convenience, one can denote $$\begin{aligned}
\label{eq:Anisorigidity1}
\Phi^{ij}_{uv} = \frac{\partial}{\partial u_i} \frac{\partial}{\partial v_j} \Phi( \|{\bf r}_i - {\bf r}_j \|; \eta_{ij} ), \quad u,v= x, y, z; i,j =1,2,\cdots,N.\end{aligned}$$ Note that for each given $ij$, one can define $\Phi^{ij}=\left( \Phi^{ij}_{uv} \right)$ as a local anisotropic matrix $$\label{eq:afri_local_Hessian}
\Phi^{ij}=\left(
\begin{array}{ccc}
\Phi^{ij}_{xx} & \Phi^{ij}_{xy}& \Phi^{ij}_{xz}\\
\Phi^{ij}_{yx} & \Phi^{ij}_{yy}& \Phi^{ij}_{yz}\\
\Phi^{ij}_{zx} & \Phi^{ij}_{zy}& \Phi^{ij}_{zz}
\end{array}
\right).$$
In the anisotropic flexibility and rigidity (aFRI) approach, a flexibility Hessian matrix ${\bf F}^{1}(c_k)$ for cluster $c_k$ is defined by $$\begin{aligned}
\label{eq:Anisoflexibility}
{\bf F}^{1}_{ij}(c_k) =& - \frac{1}{w_{j}} {\rm adj}(\Phi^{ij}), &\quad i,j \in c_k; i\neq j; u,v= x, y, z \\ \label{eq:Anisoflexibilityy3}
{\bf F}^{1}_{ii}(c_k)=& \sum_{j=1}^N \frac{1}{w_{j}} {\rm adj}(\Phi^{ij}), &\quad i \in c_k; u,v= x, y, z \\ \label{eq:Anisoflexibility4}
{\bf F}^{1}_{ij}(c_k)=& 0, &\quad i,j \notin c_k; u,v= x, y, z,\end{aligned}$$ where ${\rm adj}(\Phi^{ij})$ denotes the adjoint of matrix $\Phi^{ij}$ such that $\Phi^{ij} {\rm adj}(\Phi^{ij})=| \Phi^{ij}|I$, here $I$ is identity matrix.
Another representation for the flexibility Hessian matrix ${\bf F}^{2}(c_k)$ can be defined as follows $$\begin{aligned}
\label{eq:Anisoflexibility5}
{\bf F}^{2}_{ij}(c_k) =& - \frac{1}{w_{j}} |\Phi^{ij}|(J_{3} - \Phi^{ij}), &\quad i,j \in c_k; i\neq j; u,v= x, y, z \\ {\bf F}^{2}_{ii}(c_k)=& \sum_{j=1}^N \frac{1}{w_{j}} |\Phi^{ij}|(J_3 - \Phi^{ij}), &\quad i \in c_k; u,v= x, y, z \\ {\bf F}^{2}_{ij}(c_k)=& 0, &\quad i,j \notin c_k; u,v= x, y, z,
\end{aligned}$$ where $J_3$ is a $3\times3$ matrix with every element being one.
One can achieve $3N_k$ eigenvectors for $N_k$ atoms in cluster $c_k$ by diagonalizing ${\bf F}^{\alpha}(c_k)$, $\alpha=1,2$. Note that, the diagonal part ${\bf F}^{\alpha}_{ii}(c_k)$, $\alpha=1,2$, has inherent information of all atoms in the system. As a result, the B-factors can be predicted by the following form: $$\begin{aligned}
\label{eq:Anisoflexibility2}
f_i^{\rm AF_\alpha} &=&{\rm Tr} \left({\bf F}^{\alpha}(c_k)\right)^{ii}, \\
&=& \left({\bf F}^{\alpha}(c_k)\right)^{ii}_{xx}+ \left({\bf F}^{\alpha}(c_k)\right)^{ii}_{yy}+ \left({\bf F}^{\alpha}(c_k)\right)^{ii}_{zz}, \quad \alpha=1,2.\end{aligned}$$ It was found that aFRI is much more accurate than ANM in protein B-factor prediction [@Opron:2014]. The anisotropic cluster analysis was found to play a significant role in study the local motion of RNA polymerase II translocation [@Opron:2016a].
### Spectral graph theory {#sec:spectral}
Spectral graph theory [@ChungOverview; @Mohar:1991laplacian; @Mohar:1997some; @von:2007tutorial] concerns the study and exploration of graphs through the eigenvalues and eigenvectors of matrices naturally associated with those graphs [@shi:2000; @meila:2001; @ng:2002; @azran:2006; @zelnik:2005]. Therefore, widely used GNM and ANM methods make use of spectral graph theory.
For a given graph $G(V,E)$ with $N$ nodes, one is interested in its matrix representation. Matrices $A$ and $D$ correspond to weighted adjacent matrix and weighted degree matrix, respectively. With this notation, one has the unnormalized graph Laplacian $$L=D-A.$$ It is often called admittance matrix, Kirchhoff matrix or discrete Laplacian. The graph matrix has several interesting properties. Firstly, it is symmetric and positive semi-definite. Secondly, it has $N$ non-negative, real-valued eigenvalues. Thirdly, the smallest eigenvalue is 0 and the corresponding eigenvector is constant vector $\mathds{1}$. Fourthly, for every vector ${\bf c} \in \mathbb{R}^N$, one has ${\bf c}^T L {\bf c}=\frac{1}{2}\sum_{i,j}w_{ij}(c_i-c_j)^2$, which can be derived from $$\begin{aligned}
\label{eq:couple_matrix22}
&& {\bf c}^T L {\bf c}= {\bf c}^TD{\bf c} -{\bf c}^T A{\bf c}= \sum_{i}c_i^2 d_i- \sum_{i,j}c_i c_i w_{ij}\\\nonumber
&& =\frac{1}{2}\left( \sum_{i}c_i^2 d_i-2\sum_{i,j}c_i c_j w_{ij}+ \sum_{j}c_j^2 d_j \right)=\frac{1}{2}\sum_{i,j}w_{ij}(c_i-c_j)^2,\end{aligned}$$ where $d_i=\sum_{j=1}^N w_{ij}$ is the degree of the $i$th vertex.
In spectral graph theory, two other kinds of Laplacian matrices [@ChungOverview] are also widely used. They are the normalized Laplacian matrix $$L_{\rm sym}=I-D^{-\frac{1}{2}}AD^{-\frac{1}{2}},$$ and random-walk normalized Laplacian matrix [@meila:2001; @Lovasz:1993; @Aldous:2002] $$L_{\rm rw}=I-D^{-1}A,$$ where $I$ is an identity matrix. Three different Laplacian matrices are tightly related to different ways of graph decompositions.
#### Graph decomposition and graph cut
A protein may have different domains. Identifying protein domains and analyzing their relative motions are important for studying protein functions. A protein complex involves different proteins. The study of protein complex can often be formulated as a graph decomposition problem as well.
In many situations, for a given graph $G(V,E)$, one wants to partition it into subgraphs, so that nodes within a subgroup are of similar properties and nodes in different subgroups are of different properties. Mathematically, if the weight is a measurement of similarity, i.e., large weight means a great similarity, an optimized partition means that edges within the same subgroup should have large weights and edges across subgroups should have small weights. State differently, one wants to find a way to cut the graph so that it will minimize the weights of edges connecting vertices in different subgroups.
Let us first consider a simple situation, divide $G$ into two subgroup $G_1$ and $\bar{G_1}$. The notation $\bar{G_1}$ denotes the complementary of ${G_1}$. One can define a partition ${\bf c}=\{c_i; i=1,2,...,N\}$ as, $$\begin{aligned}
\label{eq:couple_matrix2}
c_i=\begin{cases} \begin{array}{ll}
1 & {\rm if} ~i \in G_1,\\
-1 & {\rm if} ~i \in \bar{G_1}.\\
\end{array}
\end{cases}\end{aligned}$$ Therefore, if two nodes $v_i$ and $v_j$ are in the same subgroup, one will have $(c_i-c_j)^2=0$, otherwise $(c_i-c_j)^2=4$. In this way, a ${\rm Cut}(G_1,\bar{G_1})$, which is the total weights of edges connecting two subgroups, can be defined as following, $$\begin{aligned}
\label{eq:couple_matrix222}
&& {\rm Cut}(G_1,\bar{G_1})={\bf c}^T L {\bf c}= {\bf c}^TD{\bf c} -{\bf c}^T A{\bf c}= \sum_{i}c_i^2 d_i- \sum_{i,j}c_i c_i w_{ij}\\\nonumber
&& =\frac{1}{2}\left( \sum_{i}c_i^2 d_i-2\sum_{i,j}c_i c_j w_{ij}+ \sum_{j}c_j^2 d_j \right)=\frac{1}{4}\sum_{i,j}w_{ij}(c_i-c_j)^2.\end{aligned}$$ It can be also noticed that if one defines $W(G_1,\bar{G_1})=\sum_{i\in G_1,j \in \bar{G_1}} w_{ij}$, then one has $ {\rm Cut}(G_1,\bar{G_1})=W(G_1,\bar{G_1})$.
To obtain an optimized partition means to minimize the value of $ {\rm Cut}(G_1,\bar{G_1})$. However, the way of cut stated above does not consider the size of the subgraphs. In this way, the cut or graph composition can be very uneven in terms of the number of nodes in subgroup. For example, one extreme situation is that one of the subgraph may only have a few nodes (i.e., one or two nodes), while the other subgroup may have all the rest of nodes. To avoid this problem, three commonly defined cuts, namely, ratio cut, normalized cut and min-max cut, are proposed in the literature [@Hagen:1992new; @shi:2000; @Ding:2001min]. To facilitate the description, one can define $|G|$ as the total number of nodes in graph $G$ and ${\rm vol}(G)$ is the summation of all weights in $G$, i.e., ${\rm vol}(G)=\sum_{ij}w_{ij}$.
- Ratio cut is defined as [@Hagen:1992new] $$\begin{aligned}
{\rm Rcut}(G_1,\bar{G_1})=\frac{W(G_1,\bar{G_1})}{|G_1|} + \frac{W(G_1,\bar{G_1})}{|\bar{G_1}|}.\end{aligned}$$
- Normalized cut is defined as [@shi:2000] $$\begin{aligned}
{\rm Ncut}(G_1,\bar{G_1})=\frac{W(G_1,\bar{G_1})}{W(G_1,G_1)+W(G_1,\bar{G_1})} + \frac{W(G_1,\bar{G_1})}{W(\bar{G_1},\bar{G_1})+W(\bar{G_1},G_1)}.\end{aligned}$$
- Min-Max cut is given by [@Ding:2001min] $$\begin{aligned}
{\rm Mcut}(G_1,\bar{G_1})=\frac{W(G_1,\bar{G_1})}{W(G_1,G_1)} + \frac{W(G_1,\bar{G_1})}{W(\bar{G_1},\bar{G_1})}.\end{aligned}$$
These decompositions have found many applications in image segmentation [@shi:2000]. However, the impact of these cuts to protein domain partition is yet to be examined. An important issue is how to cut a given biomolecule to elucidate its biological function and predict its chemical and biological behavior.
#### Ratio cut and Laplaician matrix
To solve the optimization problem $$\begin{aligned}
\min \limits_{G_1 \subset G} {\rm Rcut}(G_1,\bar{G_1}),\end{aligned}$$ one can define the vector $\bf {c}$ as $$\begin{aligned}
\label{eq:indicator}
c_i=\begin{cases} \begin{array}{ll}
\sqrt{\frac{\bar{|G_1|}}{|G_1|}} & {\rm if} ~i \in G_1,\\
-\sqrt{\frac{|G_1|}{\bar{|G_1|}}} & {\rm if} ~i \in \bar{G_1}.\\
\end{array}
\end{cases}\end{aligned}$$ One can have $$\begin{aligned}
{\bf c}^T L {\bf c}&=&\frac{1}{2}\sum_{i,j}w_{ij}(c_i-c_j)^2 \\\nonumber
&=& \frac{1}{2}\sum_{i\in G_1,j \in \bar{G_1}} \left( \sqrt{\frac{\bar{|G_1|}}{|G_1|}} + \sqrt{\frac{|G_1|}{\bar{|G_1|}}} \right)^2 + \frac{1}{2}\sum_{i\in \bar{G_1},j \in G_1} \left( -\sqrt{\frac{\bar{|G_1|}}{|G_1|}} - \sqrt{\frac{|G_1|}{\bar{|G_1|}}} \right)^2 \\\nonumber
&=& W(G_1,\bar{G_1}) \left( \frac{\bar{|G_1|}}{|G_1|} + \frac{|G_1|}{\bar{|G_1|}} +2 \right) \\\nonumber
&=& W(G_1,\bar{G_1}) \left( \frac{\bar{|G_1|}+|G_1|}{|G_1|} + \frac{|G_1|+\bar{|G_1|}}{\bar{|G_1|}} \right) \\\nonumber
&=& |G| {\rm Rcut}(G_1,\bar{G_1})\end{aligned}$$ One also has $$\begin{aligned}
\sum_{i} c_i= \sum_{i\in G_1} \sqrt{\frac{\bar{|G_1|}}{|G_1|}} - \sum_{i \in \bar{G_1}}\sqrt{\frac{|G_1|}{\bar{|G_1|}}}=0\end{aligned}$$ Therefore, the vector ${\bf c}$ is orthogonal to the vector with common components (leading eigenvector of the Laplacian matrix). It is noted that $$\begin{aligned}
\|f\|^2= \sum_{i} c_i^2= \sum_{i\in G_1} \frac{\bar{|G_1|}}{|G_1|} - \sum_{i \in \bar{G_1}}\frac{|G_1|}{\bar{|G_1|}}=N.\end{aligned}$$ In this way, the minimization is equivalent to $$\begin{aligned}
\label{Eq:min}
\{ \min {\bf c}^T L {\bf c}~|~ {\bf c}~ {\rm satisfies ~ Eq. ~(\ref{eq:indicator})};~ {\bf c} \bot \mathds{1};~ \|f\|^2=N \}.\end{aligned}$$ As the entries of the solution vector are only allowed to take two particular values, Eq. (\[Eq:min\]) is a discrete optimization problem. One can discharge the discreteness condition and allow the vector to take any arbitrary values. This results in a relaxed problem $$\begin{aligned}
\{ \min \limits_{{\bf c}\in \mathbb{R}^N}{\bf c}^T L {\bf c}~|~ {\bf c} \bot \mathds{1};~ \|f\|^2=N \}.\end{aligned}$$ From the Rayleigh-Ritz theorem, it can be seen that the solution of this problem is the second smallest eigenvector ${\bf q}_2$ of the Laplacian matrix. In order to obtain a partition of the graph, one can choose, $$\begin{aligned}
\label{eq:cluster_eigenv2}
\begin{cases} \begin{array}{ll}
i \in G_1 (c_i= 1) &{\rm if} ~ ({\bf q}_2)_i\geq 0\\
i \in \bar{G_1} ( c_i=-1) &{\rm if} ~({\bf q}_2)_i<0
\end{array}.
\end{cases}\end{aligned}$$ Mathematically, the second smallest eigenvalue $\lambda_2$ is known as the algebraic connectivity (or Fiedler value) of a graph. The corresponding eigenvector of the second eigenvalue offers a near optimized partition. It becomes an interesting issue to design certain weighted Laplacian matrix so that a protein domain partition is optimal with respect to protein functions. Transferring the discrete Laplacian matrix to a continuous Laplacian operator and casting the domain separation problem into an optimization one are promising approaches. Certainly, these issues are also biologically significant in the exploration of protein structure-function relationship.
#### Normalized cut and normalized Laplacian matrix
One can define the vector ${\bf c}$ as $$\begin{aligned}
\label{eq:normalized_indicator}
c_i=\begin{cases} \begin{array}{ll}
\sqrt{\frac{{\rm vol}(\bar{G_1})}{{\rm vol}(G_1)}} & {\rm if} ~i \in G_1\\
-\sqrt{\frac{{\rm vol}(G_1)}{{\rm vol}(\bar{G_1})}} & {\rm if} ~i \in \bar{G_1}
\end{array}.
\end{cases}\end{aligned}$$ One can have $$\begin{aligned}
{\bf c}^T L {\bf c}&=&\frac{1}{2}\sum_{i,j}w_{ij}(c_i-c_j)^2 \\\nonumber
&=& \frac{1}{2}\sum_{i\in G_1,j \in \bar{G_1}} \left( \sqrt{\frac{{\rm vol}(\bar{G_1})}{{\rm vol}(G_1)}} + \sqrt{\frac{{\rm vol}(G_1)}{{\rm vol}(\bar{G_1})}} \right)^2 + \frac{1}{2}\sum_{i\in \bar{G_1},j \in G_1} \left( - \sqrt{\frac{{\rm vol}(\bar{G_1})}{{\rm vol}(G_1)}} - \sqrt{\frac{{\rm vol}(G_1)}{{\rm vol}(\bar{G_1})}} \right)^2 \\\nonumber
&=& |G| {\rm Ncut}(G_1,\bar{G_1}).\end{aligned}$$ Additionally, it is easy to see that $$\begin{aligned}
(D {\bf c})^T\mathds{1}=\sum_{i} d_i c_i= \sum_{i\in G_1} d_i \sqrt{\frac{{\rm vol}(\bar{G_1})}{{\rm vol}(G_1)}} - \sum_{i \in \bar{G_1}} d_i \sqrt{\frac{{\rm vol}(G_1)}{{\rm vol}(\bar{G_1})}}=0\end{aligned}$$ This means that vector ${\bf c}$ is orthogonal to constant one vector. Moreover, one can evaluate $$\begin{aligned}
{\bf c}^T D {\bf c} = \sum_{i} d_i c_i^2= \sum_{i\in G_1} d_i \frac{{\rm vol}(\bar{G_1})}{{\rm vol}(G_1)} + \sum_{i \in \bar{G_1}} d_i \frac{{\rm vol}(G_1)}{{\rm vol}(\bar{G_1})}={\rm vol}(G)\end{aligned}$$ In this way, the minimization process is equivalent to $$\begin{aligned}
\{\min {\bf c}^T L {\bf c}~|~ {\bf c}~ {\rm satisfies ~ Eq.~ (\ref{eq:normalized_indicator}) };~ D{\bf c} \bot \mathds{1};~ {\bf c}^T D {\bf c}={\rm vol}(G) \}\end{aligned}$$ This is also a discrete optimization problem, because the entries of the solution vector are only allowed to take two particular values. By discarding the discreteness condition and allowing the vector to be any arbitrary values, one results in a relaxed problem $$\begin{aligned}
\{ \min \limits_{{\bf c} \in \mathbb{R}^N}{\bf c}^T L {\bf c}~|~ D{\bf c} \bot \mathds{1};~ {\bf c}^T D {\bf c}={\rm vol}(G) \}\end{aligned}$$ Now one can substitute ${\bf c}'=D^{\frac{1}{2}}{\bf c}$. After substitution, the problem is $$\begin{aligned}
\{ \min \limits_{{\bf c}' \in \mathbb{R}^N}({\bf c'})^T D^{-\frac{1}{2}}L D^{\frac{1}{2}} {\bf c'}~|~ {\bf c}' \bot D^{\frac{1}{2}} \mathds{1};~ \|{\bf c'}\|={\rm vol}(G) \}.\end{aligned}$$ It can been seen that $D^{-\frac{1}{2}}L D^{\frac{1}{2}}=L_{\rm sym}$ and $D^{\frac{1}{2}} \mathds{1}$ is the first eigenvector of $L_{\rm sym}$. From the Rayleigh-Ritz theorem, it can be seen that the solution of this problem is the second smallest eigenvalue of $L_{\rm sym}$.
#### Graph Laplacian and continuous Laplace operator
It has been found that there is a connection between graph Laplacian and the continuous Laplace operator [@Belkin:2003; @Lafon:2004; @Belkin:2005; @Hein:2005graphs; @Gine:2006empirical]. Roughly speaking, if one chooses $w_{ij}=\frac{1}{r_{ij}^2}$ with $r_{ij}$ as the distance between node $v_i$ and node $v_j$, one can have $$\begin{aligned}
w_{ij}(c_i-c_j)^2=\left(\frac{c_i-c_j}{r_{ij}}\right)^2.\end{aligned}$$ The term $\frac{c_i-c_j}{r_{ij}}$ can be roughly viewed as a discretization of $\nabla c({\bf r})$. In this way, there is a connection between graph Laplacian and the continuous Laplace operator through this functional formulation, $$\begin{aligned}
{\bf c}^T L {\bf c}=<{\bf c}, L {\bf c}> \approx \int |\nabla c({\bf r})|^2 d{\bf r}.\end{aligned}$$
More specifically, one can set $w_{ij}= \Phi(r_{ij},\eta_{ij})$ as defined in Eqs. (\[eq:couple\_matrix1-1\]) and (\[eq:couple\_matrix1-12\]) $$\begin{aligned}
L_N { c({\bf r}_i)}=c({\bf r}_i) \sum_j \Phi(r_{ij},\eta_{ij})- \sum_j c({\bf r}_i) \Phi(r_{ij},\eta_{ij}),\end{aligned}$$ where $c({\bf r}_i)=c_i$ and ${\bf r}_i$ is the coordinate of $i$th node. This operator can be naturally extended to an integral operator $$\begin{aligned}
L_N {c({\bf r})}=c({\bf r}) \sum_j \Phi(|{\bf r}-{\bf r}_j|,\eta_{ij}) - \sum_j c({\bf r})\Phi(|{\bf r}-{\bf r}_j|,\eta_{ij}).\end{aligned}$$
If data points are sampled from a uniform distribution on a $k$-dimensional manifold $\mathcal{M}$, let set $w_{ij}= \Phi(r_{ij},\eta_{ij})
=e^{-\frac{r_{ij}^2}{4t}}$, with $t=t_N=N^{-\frac{1}{k+2+\alpha}}$, where $\alpha >0$, and assume $c(\bf {r}) \in C^{\infty}(\mathcal{M})$. Belkin [@Belkin:2005] found that there is a constant $C$, such that in probability [@Belkin:2003; @Belkin:2005] $$\begin{aligned}
\lim \limits_{N \rightarrow \infty}C\frac{(4\pi t_N)^{-\frac{k+2}{2}}}{N} L_N^{t_N}c({\bf r})=\Delta_{\mathcal{M}} c({\bf r}).\end{aligned}$$
The continuous Laplace operator has been widely utilized in many biophysical models, such as Poisson-Boltzmann theory for electrostatics [@Holst:1994], Laplace-Beltrami equation for molecular surface modeling [@Bates:2008; @Wei:2009], Poisson-Nernst-Planck equation for ion channel modeling [@Hyon:2010; @Wei:2012], and elasticity equation for macromolecular conformational change induced by electrostatic forces [@Zhou:2008d]. Obviously, these issues are associated with a graph problem. The modeling of biomolecular structure, function, dynamics and transport by combining graph theory and partial differential equation (PDE) is an open problem.
#### Modularity
Modularity is total summation of the weights within the group minus the expected one in an equivalent network with weight randomly placed.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![An illustration of modularity matrix, second-eigenvector formed matrix and domain indication matrix. The protein network for 2ABH C$_{\alpha}$ is constructed by using Gaussian network model with cut off distance 23 Å. (a) The illustration of modularity matrix in Eq. (\[eq:modularity\_matrix\]); (b) The illustration of the matrix formed by the second eigenvector, i.e., ${\bf q}_2 {\bf q}_2^T$; (c) The illustration of the index matrix ${\bf c} {\bf c}^T$. The vector ${\bf c}$ is generated from ${\bf q}_2$ by $\{c_i=1;~if~({\bf q}_2)_i \geq0\}$ and $\{c_i=-1;~if~({\bf q}_2)_i<0\}$. The parameter $\gamma=1$ is used in the modularity model.[]{data-label="fig:2abh_domain"}](2abh_domain.png "fig:"){width="99.00000%"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Mathematically, the modularity is defined as [@Newman:2004; @Newman:2006; @Fortunato:2010; @Jain:2010] $$\begin{aligned}
Q=\frac{1}{2{\rm vol}(G)} \sum_{ij} \left(A_{ij}-\gamma \frac{d_i d_j}{{\rm vol}(G)}\right) (\delta(c_i,c_j)+1)=\frac{1}{2{\rm vol}(G)} \sum_{ij} \left(A_{ij}- \gamma \frac{d_i d_j}{{\rm vol}(G)}\right) \delta(c_i,c_j),\end{aligned}$$ where $\gamma$ is a resolution parameter, which is designed to change the scale at which a network is clustered [@Fortunato:2010]. Here $\delta(c_i,c_j)=1$ if $c_i$ and $c_j$ are in the same subgroup ($c_i=c_j$), otherwise it equals to 0 ($c_i \neq c_j$). The term $A_{ij}$ is the weighted adjacent matrix component.
The modularity matrix is defined as: $$\begin{aligned}
\label{eq:modularity_matrix}
B_{ij}=A_{ij}-\frac{d_i d_j}{{\rm vol}(G)},\end{aligned}$$ then the above equation can be further simplified as $$\begin{aligned}
Q=\frac{1}{2{\rm vol}(G)}{\bf c}^T B {\bf c}.\end{aligned}$$
Again one assume that graph $G$ can be divided into two parts $G_1$ and $\bar{G_1}$ $$\begin{aligned}
&& Q =\frac{1}{{\rm vol}(G)}\left[ \left({\rm vol}(G)-\sum_{c_i \neq c_j}w_{ij} \right) -\frac{\gamma}{{\rm vol}(G)} \left(\sum_{c_i=c_j} d_i d_j \right) \right] \\ \nonumber
&& \quad =1-\frac{1}{{\rm vol}(G)} \left({\rm Cut}(G_1,\bar{G_1}) + \frac{\gamma}{{\rm vol}(G)} {\rm vol}(G_1)^2\right) \\ \nonumber
&& \quad =1-\gamma-\frac{1}{{\rm vol}(G)} \left({\rm Cut}(G_1,\bar{G_1}) -\frac{\gamma}{{\rm vol}(G)}{\rm vol}(G_1){\rm vol}(\bar{G_1})\right).\end{aligned}$$ One can define the total variation (TV): $|c|_{\rm TV}=\frac{1}{2} \sum_{i,j}w_{ij}|c_i-c_j|$, weighted $\ell_2$-norm $\|c\|^2_{\ell_2}= \sum_{i}d_{i}|c_i|^2$, and mean ${\rm mean}(c)=\frac{1}{{\rm vol}(G)} \sum_{i}d_{i}|c_i|$.
One can also define $c$ to be a function $\chi_{G_1}: G \rightarrow \{0,1\}$. This is the indicator function of a subsect $G_1 \subset G$. In this manner, one has $$\begin{aligned}
&& \quad |c|_{\rm TV}-\gamma\|c-{\rm mean}(c)\|_{\ell_2}^2 \\ \nonumber
&&= |\chi_{G_1}|_{\rm TV} -\gamma \|\chi_{G_1}-{\rm mean}(\chi_{G_1})\| \\ \nonumber
&&={\rm Cut}(G_1,\bar{G_1})-\gamma \left( \sum_i d_i \left| \chi_{G_1}-\frac{{\rm vol}(G_1)}{{\rm vol}(G)} \right|^2 \right) \\ \nonumber
&&={\rm Cut}(G_1,\bar{G_1})-\gamma \left( {\rm vol}(G_1)\left(1-\frac{{\rm vol}(G_1)}{{\rm vol}(G)}\right)^2+{\rm vol}(\bar{G_1}) \left(\frac{{\rm vol}(G_1)}{{\rm vol}(G)}\right)^2 \right) \\ \nonumber
&&={\rm Cut}(G_1,\bar{G_1})- \frac{\gamma}{{\rm vol}(G)} {\rm vol}(G_1) {\rm vol}(\bar{G_1}).\end{aligned}$$ Figure \[fig:2abh\_domain\] illustrates the modularity matrix, second-eigenvector formed matrix and domain indication matrix for protein 2ABH. Figure \[fig:2abh\_modularity2\] shows the protein domain decomposition using a modularity matrix.
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![An illustration of graph decomposition with modularity based eigenvectors. The protein network of 2ABH C$_{\alpha}$ atoms is constructed by using the Gaussian network model with cut off distance 23 Å. Based on this network, modularity matrix is constructed with parameter $\gamma=1$. The modularity eigenvector corresponding to the second lowest eigenvalue is used for protein domain decomposition. []{data-label="fig:2abh_modularity2"}](2abh_modularity.png "fig:"){width="40.00000%"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
One can have the theorem: maximizing the modularity functional $Q$ over all the partitions is equivalent to minimize$|c|_{\rm TV}-\gamma\|c-{\rm mean}(c)\|_{\ell_2}^2$ [@Hu:2013method] . Essentially, this is equal to a balance cut problem $$\begin{aligned}
\min \frac{{\rm Cut}(G_1,\bar{G_1})}{{\rm vol}(G_1){\rm vol}(\bar{G_1})}\end{aligned}$$
One can let $c=(c^1,c^2)$ by $c: G\rightarrow V^2$ with $V^2=\{(1,0),(0,1)\}$ [@Hu:2013method] . In this way, for each $c_i$, it has only a single entry equals 1. The minimizing problem can be solved through the following equation [@Merkurjev:2013; @Hu:2013method] $$\begin{aligned}
\frac{\partial c}{\partial t}=-(L c^1,L c^2)-\frac{1}{\epsilon^2}\nabla W_{\rm multi}(c)+\frac{\delta}{\delta c}\left( \gamma \| c-{\rm mean}(c) \| \right).\end{aligned}$$ Here $\nabla W_{\rm multi}(c)$ is the composition of function $W_{\rm multi}$ and $c$. Normally, the function $W_{\rm multi}$ is a multi-well potential [@Merkurjev:2013].
The field of spectral graph, modularity and related variation formation for biomolecular systems is completely open. There is much to be done on this interesting field. For example, one can formulate molecular design, such as drug design, protein design, and the design of protein-DNA and/or protein-RNA complexes, as graph cut problems. In drug design, one would like to optimize the protein-drug binding affinity for a given drug candidate and its target protein. If the selection of graph notes is also a part of optimization, one then would like to optimize protein-drug binding affinity, drug target selectivity, drug pharmacokinetics, drug toxicity, etc. To be more specific, one needs to consider a minimization process, $$\begin{aligned}
\label{eq:cutfreeenergy}
\min \limits_{{\bf r} \in {\mathbb R}^{3N}} {\rm Cut}(G_1({\bf r}),\bar{G_1}({\bf r}))+\gamma \Delta F ({\bf r}).\end{aligned}$$ Here $N$ is the total number of atoms in the complex, the parameter $\gamma$ is a scale parameter and $ \Delta F ({\bf r})$ is the free energy change. The function $G({\bf r})$ is denoted as the graph representation of protein-protein or protein-ligand complexes. The functions $G_1({\bf r})$ and $\bar{G_1}({\bf r})$ are graph models for the corresponding protein or ligand. They are all position-dependent and the minimization process is to find the best fitting configuration so that one can achieve a cut that minimizes the free energy change $\Delta F ({\bf r})$. In fact, biomolecular free energy minimization often leads to PDE based models for solvation, ion channel, membrane protein interaction, molecular machine assembly, to name only a few [@Wei:2009]. It should be noticed that in Eq. (\[eq:cutfreeenergy\]), all the three types of graph cuts might be used. This approach can be combined with techniques in other mathematical disciplines, such as those in dynamical systems, stochastic analysis, and differential equation, to address complex design problems, namely, drug design, protein design, RNA design, molecular machine design etc, in biomolecular systems.
### Molecular nonlinear dynamics {#sec:MND}
To introduce molecular nonlinear dynamics, one can consider a folding protein that constitutes $N$ particles and has the spatiotemporal complexity of ${\mathbb{R}}^{3N}\times \mathbb{R}^{+} $. Assume that the molecular mechanics of the protein is described by molecular nonlinear dynamics having a set of $N$ nonlinear oscillators of dimension ${\mathbb{R}}^{nN}\times \mathbb{R}^{+}$, where $n$ is the dimensionality of a single nonlinear oscillator. Let us consider an $n\times N$-dimensional nonlinear system for $N$ interacting chaotic oscillators [@KLXia:2014b] $$\begin{aligned}
\label{eq:couple_matrix}
\frac{d{\bf u}}{dt} &=& {\bf F}({\bf u})+ {G}{\bf u}, ~~~\end{aligned}$$ where ${\bf u}=({\bf u}_1,{\bf u}_2,\cdots, {\bf u}_N )^T$ is an array of state functions for $N$ nonlinear oscillators, ${\bf u}_j=(u_{j1},u_{j2}, \cdots, u_{jn})^T$ is an $n$-dimensional nonlinear function for the $j$th oscillator, $ {\bf F}({\bf u})=(f({\bf u}_1), f({\bf u}_2), \cdots, f({\bf u}_N))^T$ is an array of nonlinear functions of $N$ oscillators, and ${G}=\varepsilon L \otimes \Gamma$. Here, $\varepsilon$ is the overall interaction strength, $L$ is an $N\times N$ weighted Laplacian matrix and $\Gamma$ is an $n\times n$ linking matrix. Essentially, for each node in the biomolecular graph, one has an $n$-dimensional nonlinear oscillator. These oscillators are connected by the Laplacian matrix and a fixed $n\times n$ linking matrix.
For example, one can choose a set of $N$ Lorenz attractors [@Lorenz:1963] and a simple $3*3$ link matrix as following: $\mathbf{u}_i=(u_{i1},u_{i2},u_{i3})^T$, $$\begin{aligned}
\begin{cases} \nonumber
\frac{du_{i1}}{dt}=\alpha(u_{i2}-u_{i1}) \\ \label{oscillator}
\frac{du_{i2}}{dt}=\gamma u_{i1}-u_{i2}-u_{i1}u_{i3} \\ \nonumber
\frac{du_{i3}}{dt}=u_{i1}u_{i2}-\beta u_{i3}, i= 1, 2, \cdots, N
\end{cases},~\Gamma= \left( \begin{array}{ccc}
0 & 0 & 0 \\
1 & 0 & 0 \\
0 & 0 & 0
\end{array}
\right).\end{aligned}$$ If the Laplacian matrix shown in Fig. \[fig:GRAPH\] is used as the connectivity matrix, the nonlinear dynamic system for the first node is $$\begin{aligned}
\begin{cases} \nonumber
\frac{du_{11}}{dt}=\alpha(u_{12}-u_{11}) \\ \label{oscillator2}
\frac{du_{12}}{dt}=\gamma u_{11}-u_{12}-u_{11}u_{13}+\varepsilon (2u_{11}-u_{21}-u_{81})\\ \nonumber
\frac{du_{13}}{dt}=u_{11}u_{12}-\beta u_{13}
\end{cases}.\end{aligned}$$
#### Stability analysis {#Sec:Stability}
One can use the FRI kernel weighted Laplacian matrix to define the driving and response relation of nonlinear chaotic oscillators. Due to the synchronization of chaotic oscillators, an $N$-time reduction in the spatiotemporal complexity can be achieved, leading to an intrinsically low dimensional manifold (ILDM) of dimension ${\mathbb{R}}^{n}\times \mathbb{R}^{+}$. Formally, the $n$-dimensional ILDM is defined as $$\begin{aligned}
{\bf u}_1(t)={\bf u}_2(t)=\cdots = {\bf u}_N(t)={\bf s} (t),\end{aligned}$$ where ${{\bf s} (t)}$ is a synchronous state or reference state.
To understand the stability of the ILDM of protein chaotic dynamics, one can define a transverse state function as ${\bf w}(t)={\bf u}(t)-{\bf S}(t)$, where ${\bf S}(t)$ is a vector of $N$ identical components $({\bf s}(t),{\bf s}(t),\cdots, {\bf s}(t) )^T$. Obviously, the invariant ILDM is given by ${\bf w}(t)={\bf u}(t)-{\bf S}(t)={\bf 0}$. Therefore, the stability of the ILDM can be analyzed by $\frac{d {\bf w}(t)}{dt} =\frac {d {\bf u}(t)}{dt} -\frac {d {\bf S}(t)}{dt}$, which can be studied by the following linearized equation [@Pecora:1997; @GHu:1998] $$\begin{aligned}
\label{eqn:trans}
\frac{d{\bf w}}{dt} = ({\bf DF} ({\bf s}) + {G}){\bf w}, ~~~\end{aligned}$$ where ${\bf DF}({\bf s})$ is the Jacobian of ${\bf F}$.
To further analyze the stability of Eq. (\[eqn:trans\]), [one can diagonalize connectivity matrix ${L}$ ]{} $$\begin{aligned}
\label{eqn:trans2}
{L}{\bf \varphi}_j(t) =\lambda_j {\bf \varphi}_j(t), \quad j=1,2,\cdots, N,\end{aligned}$$ [where $\{{\bf \varphi}_j\}_{j=1}^N$ are eigenvectors and ${\lambda_j}_{j=1}^N$ are the associated eigenvalues. These eigenvectors span a vector space in which a transverse state vector has the expansion [@Pecora:1997; @GHu:1998]]{} $$\begin{aligned}
{\bf w}(t)=\sum_{j}{\bf q}_j(t)\phi_j(t).\end{aligned}$$ [Therefore, the stability problem of the ILDM is equivalent to the following stability problem]{} $$\begin{aligned}
\label{eqn:trans3}
\frac{d{\bf q}_j(t)}{dt} &=& ( Df ({\bf s}) + \varepsilon \lambda_j\Gamma){\bf q}_j(t), \quad j=1,2,\cdots,N,\end{aligned}$$ where $Df ({\bf s})$ is the diagonal component of ${\bf DF} ({\bf s})$. The stability of Eq. (\[eqn:trans3\]) is determined by the largest Lyapunov exponent $L_{\rm max}$, namely, $L_{\rm max} < 0$, which can be decomposed into two contributions $$L_{\rm max}=L_{\rm f}+L_{\rm c},$$ where $L_{\rm f}$ is the largest Lyapunov exponent of the original $n$ dimensional chaotic system $\frac{d{\bf s}}{dt} = { f}({\bf s})$, which can be easily computed for most chaotic systems. Here, $L_{\rm c}$ depends on $\lambda_j$ and $\Gamma$. The largest eigenvalue $\lambda_1$ equals 0, and its corresponding eigenvector represents the homogeneous motion of the ILDM, and all of other eigenvalues $\lambda_j, j=2,3,\cdots, N$ govern the transverse stability of the ILDM. Let us consider a simple case in which the linking matrix is the unit matrix ($\Gamma={\bf I}$). Then stability of the ILDM is determined by the second largest eigenvalue $\lambda_{2}$ (algebraic connectivity, or Fiedler value), which enables us to estimate the critical interaction strength $\varepsilon_{c}$ in terms of $\lambda_{2}$ and $L_{f}$ [@KLXia:2014b], $$\label{eq:expect}
\varepsilon_{c}=\frac{L_f}{- \lambda_{2}}.$$ The dynamical system reaches the ILDM when $\varepsilon > \varepsilon_{c}$ and is unstable when $\varepsilon \leq \varepsilon_{c}$. The eigenvalues of protein connectivity matrices are obtained with a standard matrix diagonalization algorithm. Molecular nonlinear dynamics has been recently developed as an efficient means for protein B-factor prediction [@KLXia:2014b]. It can be potentially used for protein domain separation without resorting to the matrix diagonalization.
The availability of more than a hundred thousands of interacting protein networks in the PDB provides living example problems for analyzing dynamical systems. Indeed, the connection between dynamical systems and spectral graph theory in mathematics, and the structure and function of macromolecules gives rise to exciting opportunities to further study dynamical systems and graph theory, and better analyze biomolecules.
Persistent homology {#sec:PHA}
--------------------
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![An illustration of topological change over different dense thresholds for a benzene molecule. From left to right, the density threshold is decreased.[]{data-label="fig:density"}](density.png "fig:"){width="60.00000%"}
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
As a branch of algebraic topology, persistent homology is a topological approach that utilizes algebraic algorithms to compute topological invariants in data[@Carlsson:2009; @Niyogi:2011; @BeiWang:2011; @Rieck:2012; @XuLiu:2012]. It is a working horse in the popular topological data analysis and has found its success in qualitative characterization or classification. Specifically, persistent homology describes geometric features of biomolecular date with topological invariants that persist over the systematic change of a scale parameter relevant to topological events. Such a change is called filtration. Figure \[fig:density\] illustrates the change of the topology of benzene molecule over different density thresholds. The idea is to capture topological structures continuously over a range of spatial scales during the filtration process. Unlike computational homology that is based on truly metric free or coordinate free representations, persistent homology can embed geometric information from protein data into topological invariants so that “birth" and “death" of geometric features, such as circles, rings, loops, pockets, voids and cavities can be monitored by topological measurements during the filtration process[@Edelsbrunner:2002; @Zomorodian:2005]
### Simplicial homology and persistent homology {#sec:SimplicialHomology}
An essential ingredient of persistent homology is simplicial homology which is built on simplicial complex. Simplicial complex is a finite set that consists of discrete vertices (nodes or atoms in a protein), edges (line segments or bonds in a biomolecule), triangles, and their high dimensional counterparts. Simplicial homology can be defined on simplicial complex to analyze and extract topological invariants. Then a filtration process is used to establish topological persistence from simplicial homology analysis[@Edelsbrunner:2002; @Zomorodian:2005].
#### Simplicial complex
A key component of simplicial complex $K$ is a $k$-simplex, $\sigma^k$, defined as the convex hall of $k+1$ affine independent nodes in $\mathbb{R}^N$ ($N>k$). Let $v_0,v_1,v_2,\cdots,v_k$ be $k+1$ affine independent points (or atoms in a biomolecule) and express a $k$-simplex $\sigma^k=\{v_0,v_1,v_2,\cdots,v_k\}$ as $$\begin{aligned}
\label{eq:couple_matrix12}
\sigma^k=\left\{\tau_0 v_0+\tau_1 v_1+ \cdots +\tau_k v_k \mid \sum^{k}_{i=0}\tau_i=1;0\leq \tau_i \leq 1,i=0,1, \cdots,k \right\}.\end{aligned}$$ Moreover, let us define an $i$-dimensional face of $\sigma^k$ as the convex hall formed by the nonempty subset of $i+1$ vertices from $\sigma^k$ ($k>i$). Clearly, a 0-simplex is a vertex, a 1-simplex is an edge, a 2-simplex is a triangle, and a 3-simplex represents a tetrahedron. One can also define the empty set as a (-1)-simplex.
A simplicial complex is constructed to combine these geometric components, including vertices, edges, triangles, and tetrahedrons together under certain rules. More specifically, a simplicial complex $K$ is a finite set of simplicies that satisfy two conditions. One is that any face of a simplex from $K$ is also in $K$ and the other is that the intersection of any two simplices in $K$ is either empty or shared faces. The dimension of a simplicial complex is defined as the maximal dimension of its simplicies. The underlying topological space $|K|$ is a union of all the simplices of $K$, i.e., $|K|=\cup_{\sigma^k\in K} \sigma^k$. Further, the concept of chain is introduced to associate this topological space with algebra groups.
#### Homology
One can denote a linear combination $\sum^{k}_{i}\alpha_i\sigma^k_i$ of $k$-simplex $\sigma^k_i$ as a $k$-chain $[\sigma^k]$. The coefficients $\alpha_i$ can be chosen from different fields, such as rational field $\mathbb{Q}$, real number field and complex number field, and from integers $\mathbb{Z}$ and prime integers $\mathbb{Z}_p$ with prime number $p$. For simplicity, one can consider the coefficients $\alpha_i$ are chosen from $\mathbb{Z}_2$, for which the addition operation between two chains is the modulo 2 addition for the coefficients of their corresponding simplices. The set of all $k$-chains of simplicial complex $K$ and the addition operation form an Abelian group $C_k(K, \mathbb{Z}_2)$. Therefore, the homology of a topological space is represented by a series of Abelian groups.
A boundary operation $\partial_k$ is defined as $\partial_k: C_k \rightarrow C_{k-1}$. Without orientation, the boundary of a $k$-simplex $\sigma^k=\{v_0,v_1,v_2,\cdots,v_k\}$ is $$\begin{aligned}
\partial_k \sigma^k = \sum^{k}_{i=0} \{ v_0, v_1, v_2, \cdots, \hat{v_i}, \cdots, v_k \},\end{aligned}$$ where the notation $\{v_0, v_1, v_2, \cdots ,\hat{v_i}, \cdots, v_k\}$ means that the $(k-1)$-simplex is generated by eliminating vertex $v_i$ from the sequence. When a boundary operator is applied twice, any $k$-chain will be mapped to a zero element, i.e., $\partial_{k-1}\partial_k= \emptyset$. As a special case, one has $\partial_0= \emptyset$. The $k$th cycle group $Z_k$ and the $k$th boundary group $B_k$ are the subgroups of $C_k$ and can be defined by means of the boundary operator, $$\begin{aligned}
&& Z_k={\rm Ker}~ \partial_k=\{c\in C_k \mid \partial_k c=\emptyset\}, \\
&&{ B_k={\rm Im} ~\partial_{k+1}= \{ c\in C_k \mid \exists d \in C_{k+1}: c=\partial_{k+1} d\}.}\end{aligned}$$ An element in the $k$th cycle group $Z_k$ or the $k$th boundary group $B_k$ is called the $k$th cycle or the $k$th boundary. One has $B_k\subseteq Z_k \subseteq C_k$ since the boundary of a boundary is always empty $\partial_{k-1}\partial_k= \emptyset$. Geometrically, the $k$th cycle is a $k$ dimensional loop or hole.
With all the above definitions, one can define the homology group. Specifically, the $k$th homology group $H_k$ is defined as the quotient group of the $k$th cycle group $Z_k$ and $k$th boundary group $B_k$: $H_k=Z_k/B_k$. Two $k$th cycle elements are called homologous if they are different by a $k$th boundary element. The $k$th Betti number represents the rank of the $k$th homology group, $$\begin{aligned}
\beta_k = {\rm rank} ~H_k= {\rm rank }~ Z_k - {\rm rank}~ B_k.\end{aligned}$$ From the fundamental theorem of finitely generated Abelian groups, the $k$th homology group $H_k$ can be given as a direct sum, $$\begin{aligned}
H_k= {Z}\oplus \cdots \oplus {Z} \oplus {Z}_{p_1}\oplus \cdots \oplus {Z}_{p_n}= {Z}^{\beta_k} \oplus {Z}_{p_1}\oplus \cdots \oplus {Z}_{p_n},\end{aligned}$$ where $\beta_k$ is the rank of the subgroup and is $k$th Betti number. Here $ {Z}_{p_i}$ is torsion subgroup with torsion coefficients $\{p_i| i=1,2,...,n\}$, the power of prime number.
Topologically, cycle element in $H_k$ forms a $k$-dimensional loop or ring that is not from the boundary of a higher dimensional chain element. The geometric meanings of Betti numbers in $\mathbb{R}^3$ are the follows: $\beta_0$ represents the number of isolated components (i.e., protein atoms), $\beta_1$ is the number of one-dimensional loop or ring, and $\beta_2$ describes the number of two-dimensional voids or cavities. Together, the Betti number sequence [ $\{\beta_0,\beta_1,\beta_2,\cdots \}$]{} gives the intrinsic topological property of biomolecular data.
#### $\check{\rm C}$ech complex, Rips complex and alpha complex
A key concept for the construction of simplicial complex from a point set of a given topological space is nerve. Basically, given an index set $I$ and open set ${\bf U}=\{U_i\}_{i\in I}$ that is a cover of a point set $X \in \mathbb{R}^N$, i.e., $X \subseteq \{U_i\}_{i\in I}$, the nerve [**N**]{} of [**U**]{} should satisfy two basic conditions. One is that $\emptyset \in {\bf N}$. Additionally, if $\cap_{j \in J} U_j \neq \emptyset $ for $J \subseteq I $, then $J \in {\bf N}$. Usually, for a given biomolecular dataset, the simplest way to construct a cover is to assign a ball of certain radius around each atom. If the biomolecular dataset is dense enough and the radius is large enough, then the union of all the balls has the capability to recover the underlying space for the biomolecule.
The nerve of a cover of the biomolecule constructed from the union of atomic balls is a $\check{\rm C}$ech complex for the biomolecule. More specifically, for a biomolecular dataset $X \in \mathbb{ R}^N$, one defines a cover of closed atomic balls ${\bf B}=\{B (x, r)\mid x \in X \}$ with radius $r$ and centered at $x$. The $\check{\rm C}$ech complex of $X$ with radius $r$ is denoted as $\mathcal{C}(X,r)$, which is the nerve of the closed ball set [**B**]{}, $$\begin{aligned}
\mathcal{C}(X,r) = \left\{ \sigma \mid \cap_{x \in \sigma} B (x,r) \neq \emptyset \right\}.\end{aligned}$$ One can relax $\check{\rm C}$ech complex conditions to generate a Vietoris-Rips complex, in which, a simplex $\sigma$ is constructed if the largest distance between any two atoms is at most $2r$. One can denote $\mathcal{R}(X,r)$ the Vietoris-Rips complex, or Rips complex [@Edelsbrunner:1994]. There is a sandwich relation for these abstract complexes, $$\begin{aligned}
\label{eq:SandwichRelation}
\mathcal{C}(X,r)\subset \mathcal{R}(X,r) \subset \mathcal{C}(X,\sqrt{2}r).\end{aligned}$$ In practical applications, Rips complex is preferred due to its computational convenience.
Another important geometric concept in computational geometry is alpha complex. Let $X$ be a biomolecular dataset in Euclidean space $\mathbb{R}^d$ and define the Voronoi cell of a point $x \in X$ as $$\begin{aligned}
V_x = \{ u\in R^d \mid |u-x|\leq |u-x'|, \forall x'\in X \}.\end{aligned}$$ Then the collection of all Voronoi cells for the biomolecule forms a Voronoi diagram. Further, the nerve of the biomolcular Voronoi diagram generates a Delaunay complex.
One can define $R(x,r)$ as the intersection of Voronoi cell $V_x$ with ball $B(x,\epsilon)$, i.e., $R(x,r)= V_x \cap B(x,r)$. The alpha complex $\mathcal{A}(X,r)$ of the dataset $X$ is defined as the nerve of cover $\cup_{x\in X} R(x,r)$, $$\begin{aligned}
\mathcal{A}(X,r) = \left\{ \sigma \mid \cap_{x \in \sigma} R (x,r) \neq \emptyset \right\}.\end{aligned}$$ Therefore, an alpha complex is a subset of the Delaunay complex.
-------------------------------------------------------------------------------------------------------------------------------------------------------
![Barcode representation of persistent homology analysis for fullerene molecule $C_{60}$.[]{data-label="fig:c60"}](c60.png "fig:"){width="60.00000%"}
-------------------------------------------------------------------------------------------------------------------------------------------------------
#### General filtration processes
In order to construct a simplicial homology from a dataset, a special parameter (like the radius $r$ mentioned above) is commonly used. However, to find a “suitable" value for this parameter, so that it can reveal the underlying manifold, is not straightforward. An elegant alternative approach is to carry out a filtration process [@Bubenik:2007; @edelsbrunner:2010; @Dey:2008; @Dey:2013; @Mischaikow:2013]. A suitable filtration is vital to the resulting persistent homology, which will be described in the following section. In practice, two commonly used filtration algorithms are the Euclidean-distance or correlation matrix based and density based ones. These basic filtration algorithms can be modified to achieve different goals in data analysis [@KLXia:2014c; @KLXia:2015c; @KLXia:2015d].
In the Euclidean-distance based filtration, one associates each atom with an ever-increasing radius to form an ever-growing ball for each atom. Various aforementioned complex construction algorithms can be utilized to identify the corresponding complexes. This filtration process can be formalized by the use of a distance matrix $\{d_{ij}\}$. Here the matrix element $d_{ij}$ represents the distance between atom $i$ and atom $j$. For diagonal terms, one assumes $d_{ij}=0$. With a filtration threshold $\varepsilon$, a 1-simplex is generated between atoms $i$ and $j$ if $d_{ij}\leq\varepsilon$. Higher dimensional complexes can also be created similarly.
Another important filtration process is the density based filtration process. In this process, the filtration goes along the increase or decrease of the density value. In this way, a series of isosurfaces are generated. Morse complex is used for the characterization of their topological invariants.
#### Persistent homology {#persistent-homology}
Persistent homology is an elegant mathematical theory to describe topological invariants from a series of topological spaces in various scales, that are generated by the filtration process. Persistent homology concerns a family of homologies, in which the connectivity of the given dataset is systematically reset according to a (scale) parameter. For a simplicial complex $K$, the filtration is defined as a nested sub-sequence of subcomplexes, $$\begin{aligned}
\varnothing = K^0 \subseteq K^1 \subseteq \cdots \subseteq K^m=K.\end{aligned}$$ The introduction of filtration leads to the creation of persistent homology. When the filtration parameter is a scale parameter, simplicial complexes generated from a filtration give a multiscale representation of the corresponding topological space, from which related homology groups can be evaluated to reveal topological features of the given dataset. Furthermore, the concept of persistence is introduced to measure the persistent length of topological features. The $p$-persistent $k$th homology group $K^i$ is $$\begin{aligned}
H^{i,p}_k=Z^i_k/(B_k^{i+p}\bigcap Z^i_k).\end{aligned}$$ Through the study of the persistent pattern of these topological features, the so called persistent homology is capable of capturing the intrinsic properties of the underlying protein topological space solely from the protein atomic coordinates.
To visualize the persistent homology results, many elegant representation methods have been proposed, including persistent diagram[@Edelsbrunner:2008persistent], persistent barcode[@Ghrist:2008], persistent landscape [@Bubenik:2015statistical], etc. In this paper, a barcode representation is used. The persistent barcode of an Euclidean-distance based filtration process of a fullerene molecular $c_{60}$ is shown in Fig. \[fig:c60\].
The combination of optimization and persistent homology was discussed in a recent work for biomolecular data analysis [@BaoWang:2016a]. The essentially idea to create an object functional for extracting certain geometric features in data. Then the use of variational principle to result in a differential equation, which is subsequently utilized to filtrate the biomolecular data. In this work, the minimization of the surface energy of biomolecules was the objective, which leads to the Laplace-Beltrami flow for filtration. In this manner, one can have connected persistent homology to other important mathematical subjects, such as partial differential equations, optimizations, and differential geometry [@BaoWang:2016a].
Object-oriented persistent homology is expected to play an important role in massive data analysis. In particular, this approach can be combined with a deep learning strategy to automatically extract desirable information in a semi-supervised or unsupervised learning framework.
### Multiscale persistent homology {#sec:mPHA}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Illustration of multiscale properties in an icosahedral viral particle capsid, which consists many hexagons and pentagons as shown on the left chart. A pentagon shown on the second left consists of five proteins. For a protein shown on second right, there are many residues indicated by different colors. Finally, for residue shown on the right, it has many atoms, including hexagonal and pentagonal rings. []{data-label="fig:Multiscale_illustration"}](Multiscale_illustration.png "fig:"){width="90.00000%"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The emergence of complexity in self-organizing biological systems frequently requires more comprehensive topological descriptions. Therefore, multiscale persistent homology, multiresolution persistent homology and multidimensional persistence become valuable for biological systems as well as many other complex systems.
It is noted that there is no need for a simplicial complex to be built exclusively on a set of atoms. It can be constructed for a biomolecule from its coarse-grained representation, namely, a set of amino acid residues. For protein complexes, such as a viral particle as shown in Fig. \[fig:Multiscale\_illustration\], a basic vortex in a simplicial complex can be either a protein, a residue (i.e., C$_{\alpha}$ representation) or an atom. Each of these selections gives rise to a different persistent homology at a given scale. To demonstrate the utility of persistent homology for cryo-EM data analysis, a microtubule intermediate structure, protein complex 1129 from electron microscopy data (EMD) is considered [@KLXia:2015b]. In this study, each point in the simplicial complex is a protein (tubulin).
An interesting problem in persistent homology studies is how to select appropriate scales and appropriate types of molecular information in filtration to better analyzing the structure, function and dynamics of subcellular organelles, molecular motors and multiprotein complexes.
### Topology based quantitative modeling
Traditional topological analysis analyzes data in terms of topological invariants, such as Euler characteristic, winding number, Betti numbers and so on, thus leads to so much reduction that the resulting information is hardly useful for complex real world problems. Geometric tools often become computationally intractable for macromolecules and their interactions due to the involved high degrees of freedom. As one can see, persistent homology embeds geometric information in topological invariants and bridges the gap between geometry and topology. However, persistent homology has been mainly employed for qualitative analysis, namely, characterization and classification. Only recently, persistent homology has been devised for quantitative analysis, mathematical modeling, and physical prediction [@KLXia:2014c; @KLXia:2015c; @KLXia:2015d; @BaoWang:2016a]. It has been shown that the length of intrinsic Betti 2 bar provides an excellent model for fullerene thermal energies [@BaoWang:2016a]. It has also been shown that accumulated Betti 0 and Betti 1 bar lengths offer highly accurate predictions of unfolding protein bond and total energies, respectively [@KLXia:2015c]. It is expected that persistent homology will continue to play a significant role in quantitative modeling and prediction.
Additionally, earlier work regard short-lived barcodes or topological invariants as [*noise*]{}. It has also been pointed out that, for biomolecular datsets, these short-lived topological invariants or non-persistent topological features, are part of topological fingerprints and have meaningful biophysical interpretations [@KLXia:2014c].
In general, topology based quantitative modeling and analysis will be a new trend in molecular bioscience and biophysics. Particularly, topological features are ideally suitable for machine learning based quantitative predictions of biomolecular functions.
Continuous apparatuses for biomolecules {#sec:continuous}
=======================================
As stated above, the topological study of scalar fields, particularly electron density data, has advanced the understanding of molecules tremendously. The theory of atoms in molecules (AIM) has provided an elegant and feasible partition of electron density field, so that one can mathematically rigorously define the atoms in molecule [@Bader:1985; @Bader:1990]. AIM can be generalized into a more general theory called quantum chemical topology (QCT), which employs the topological analysis in AIM for the study of other physically meaningful scalar fields [@Popelier:2005]. Additionally, geometric modeling and analysis, particularly the surface curvature analysis, has contributed a lot to the molecular visualization and structure characterization [@Whitley:2012]. Thanks to the geometric analysis, the establishment and further deeper understanding of structure-function relationship have been achieved. In this section, a brief discuss of geometric and topological analysis of scalar fields is presented.
Continuum representation of biomolecules plays an important role in their modeling, analysis and simulation. Volumetric biomolecular data are typically obtained from cryo-EM maps, quantum mechanical simulations and mathematical models that transform discrete datasets originally generated by X-ray crystallography or other means into continuous ones. Therefore, continuous mathematical approaches for analysis and modeling of biomolecules in the volumetric data form are as important as their discrete counterparts.
Geometric representation {#Sec:GeometricRep}
------------------------
Geometric modeling is a crucial ingredient of biophysics. Due to increasingly powerful high performance computers, geometric modeling has become an essential apparatus for biomolecular surface representation, visualization, surface and volumetric meshing, area and volume estimation, curvature analysis and filling the gap between macromolecular structural information and their theoretical models [@ZYu:2008; @XFeng:2012a; @XFeng:2013b; @KLXia:2014a; @JLi:2013; @Quine:2006intensity]. The visualization of macromolecules sheds light on biomolecular structure, function and interaction, including ligand-receptor binding sites, protein specification, drug binding, macromolecular assembly, protein-nucleic acid interactions, protein-protein binding hot spots, and enzymatic mechanism [@GRASP2; @Rocchia:2002; @NKWH07; @Decherchi:2013].
#### Non-smooth biomolecular surface representations
A number of molecular surface models has been proposed. Among them, the van der Waals surface (vdWS) is defined as the union of the atomic surfaces under a given atomic radius for each type of atoms. Solvent accessible surface (SAS) is defined as the trajectory of the center of a probe sphere moving around the van der Waals surface [@Lee:1971]. Because vdWs and SASs are non-smooth at intersection areas where two or more atoms join together, solvent excluded surface (SES) was introduced to generate relatively more smooth surfaces [@Richards:1977; @Connolly85]. The SES can be obtained by tracing the inward moving surface of a probe sphere rolling around the vdW surface. Connolly divided SES into two major parts, the contact areas formed by the subsets of the vdWs surface and the re-entrant surfaces, which contain toroidal patches and concave spherical triangles.
#### Smooth biomolecular surface representations
The SES of proteins admits geometric singularities, such as tips, sharp edges and self-intersecting surfaces [@Sanner:1996; @Yu:2007a] The construction of smooth biomolecular surfaces has been of considerable interest [@Blinn:1982; @Duncan:1993; @QZheng:2012; @ZYu:2008; @MXChen:2011]. The rigidity index in Eq. (\[rigidity1\]) has been extended into a continuous rigidity density [@KLXia:2013d; @Opron:2014] $$\begin{aligned}
\label{eq:rigidity3}
\mu^1(\mathbf{r})=\sum_{\substack{j=1}}^{N} w_{j} \Phi\left(\|\mathbf{r}-\mathbf{r}_j\|;\eta_{j}\right).\end{aligned}$$ Rigidity density (\[eq:rigidity3\]) serves as an excellent representation of molecular surfaces [@KLXia:2015e]. Gaussian surface was proposed with the Gaussian kernel [@Zap; @Grant:2007; @LLi:2013; @LinWang:2015]. Recently, Gaussian surface has been extended to a new class of surface densities equipped with a wide variety of FRI correlation kernels ($\Phi\left(\|\mathbf{r} -\mathbf{r}_j\|;\eta_{ j}\right)$) [@DDNguyen:2016b] $$\begin{aligned}
\label{rigidity2}
\mu^2(\mathbf{r}) =1-\prod_{\substack{j=1}}\left[1-w_{ j}\Phi\left(\|\mathbf{r} -\mathbf{r}_j\|;\eta_{ j}\right)\right].\end{aligned}$$ Two rigidity densities $\mu^\alpha(\mathbf{r}), ~\alpha=1,2$ may behave very differently. Therefore, one can normalize these densities by their maximal values $$\begin{aligned}
\label{normalization}
\bar{\mu}^{\alpha}(\mathbf{r})=\frac{\mu^{\alpha}(\mathbf{r})}{\max\limits_{\mathbf{r}\in \mathbb{R}^3} \mu^\alpha(\mathbf{r})}, \quad \alpha =1,2.\end{aligned}$$ As a result, the behaviors of two rigidity surfaces can be compared.
#### Discrete to continuum mapping
Many geometric and topological apparatuses are invented for continuous volumetric data. A typically examples include differential geometry and differential topology that deal with differentiable functions on differentiable manifolds. Cryo-EM data and electron quantum densities can be directly treated by mathematical tools devised from differentiable manifolds. However, a large variety of discrete macromolecular data originate from X-ray crystallography, NMR etc are not directly differentiable. Therefore, it desirable to transform discrete biomolecular datasets into continuous ones.
The rigidity densities defined in Eq. (\[eq:rigidity3\]) is differentiable for $\eta_{ j}>0$. Therefore, rigidity densities also serve as a discrete to continuum mapping. The resolution parameter can be exploited for generating multidimensional persistence as illustrated in Section \[sec:resultionPH\]. As a result, many mathematical techniques developed for continuous datasets can be employed to analyze discrete biomolecular datasets, such as X-ray crystallography data.
Additionally, the normalized rigidity density $\bar{\mu}^{1}(\mathbf{r})$ given in Eq. (\[normalization\]) can be used as solute domain indicators for implicit solvent models [@Holst:1994; @Baker:2001; @Geng:2007a; @Geng:2011], such as those used in the differential geometry based Poisson-Boltzmann theory [@Wei:2009; @ZhanChen:2010a; @ZhanChen:2010b].
Mathematically, the aforementioned discrete to continuum mapping is an interpolation using kernels. Many other techniques, such as splines, polynomials, wavelets, and Padé approximation can be used as well. Currently, there is little numerical analysis of the mapping in biomolecular context and further mathematical study is needed to improve the stability and efficiency of the mapping for large data sets.
Multiresolution and multidimensional persistent homology {#sec:resultionPH}
--------------------------------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Multiresolution representations of protein 1DYL. Different values of resolution parameter $\eta$ are used to generate rigidity density functions in different scales.[]{data-label="fig:1DYL"}](1DYL.png "fig:"){width="80.00000%"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Although persistent homology was originally built over simplicial complex of discrete data sets, homology and persistent homology in the cubical complex setting have been developed [@Kaczynski:2004; @Strombom:2007]. Therefore, one can apply these techniques to volumetric datasets directly, particularly, with an available software package, [Perseus](http://www.sas.upenn.edu/~vnanda/perseus/index.html) [@Mischaikow:2013]. The reader is referred to the literature for more comprehensive discussion and treatment [@Kaczynski:2004; @Strombom:2007].
Recently, persistent homology analysis of macromolecular volumetric datasets have been demonstrated [@KLXia:2014c; @KLXia:2015a; @KLXia:2015b]. In this paper, the multiresolution persistent homology [@KLXia:2015d; @KLXia:2015e] and multidimensional persistent homology [@KLXia:2015c] developed for volumetric macromolecular datasets are discussed.
#### Multiresolution persistent homology
As stated earlier, the basic idea of persistent homology is to exploit the topological changes of a given dataset at different scales of representations [@KLXia:2014c; @KLXia:2015a; @KLXia:2015b]. For the geometric representation given by Eq. (\[eq:rigidity3\]), rigidity density $\mu^1(\mathbf{r})$ depends on the resolution parameter $\eta$. The resolution parameter can be turned to emphasize the molecular features of scale $\eta$, see Fig. \[fig:1DYL\]. Resolution based continuous coarse-grained representations can be constructed for excessively large datasets in the spirit of wavelet multiresolution analysis. This approach is particularly valuable for representing viruses, protein complexes and subcellular organelles.
Since the geometric representation is controlled by the resolution parameter $\eta$ in Eq. (\[eq:rigidity3\]), one can develop multiresolution persistent homology (MPH) for macromolecular analysis. The essential idea is to match the scale of interest with appropriate resolution in the topological analysis. In contrast to the original persistent homology that is based on a uniform resolution of the point cloud data over the filtration domain, the MPH provides a mathematical microscopy of the topology at a given scale through a corresponding resolution. MPH can be utilized to reveal the topology of a given geometric scale and employed as a topological focus of lens. It becomes powerful when it is applied in conjugation with the data that has a multiscale nature, such as a multiprotein complex as shown in Fig. \[fig:Multiscale\_illustration\]. In this case, MPH can be used to extract the topological fingerprints either at atomic scale, residue scale, alpha helix and beta sheet scale, domain scale or at the protein scale.
Another very interesting multiresolution model is Mapper [@singh:2007; @Carlsson:2009]. This method is proposed for qualitative analysis, simplification and visualization of high dimensional data sets. It manages to reduce the complexity by using fewer points which can capture topological and geometric information at a specified resolution. Interestingly, Mapper also uses kernel functions. However, it should be noticed that it does not utilize the resolution parameter in the filtration process for persistent homology. It is expected that related subjects, such co-homology, Floer homology, Sheaf and K-theory, will find interesting applications in biomolecular systems.
#### Multidimensional persistent homology
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![An illustration of multidimensional persistent homology analysis of protein 2YGD. Left chart: protein 2YGD; Right chart: two-dimensional persistence of 2YGD. The horizontal axis denotes density and the vertical axis represents the logarithmic values of persistent Betti numbers.[]{data-label="fig:2ygd"}](2ygd.png "fig:"){width="80.00000%"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
There have been considerable interest in developing multidimensional persistent homology or multidimensional persistence. Here, two classes of multidimensional persistence algorithms for biomolecular data are discussed. One class of multidimensional persistence is generated by repeated applications of 1D persistent homology to high-dimensional data, such as those from protein folding, molecular dynamics, geometric partial differential equations, etc. The resulting high-dimensional persistent homology is a pseudo-multidimensional persistence and has been applied to protein folding analysis to identify topological transitions [@KLXia:2015c].
Another class of multidimensional persistence is created by the geometric representation given in Eq. (\[eq:rigidity3\]). As $\eta$ is an independent variable that can modify the geometry and topology of the underlying data set, one can carry out filtration with respect to both the original density and the resolution $\eta$ to obtain genuine 2D persistent homology [@KLXia:2015c]. Indeed, each $\eta$ value leads to a family of new simplicial complexes. Similarly, for an $N$-dimensional persistent homology, $N$-independent variables should be introduced for the filtration. Higher dimensional persistence has been demonstrated for macromolecular data [@KLXia:2015c]. Resolution induced persistent homology and multidimensional persistence have been applied to many biomolecular systems, including protein flexibility analysis, protein folding characterization, topological denoising, noise removal from cryo-EM data, and analysis of fullerene molecules. An example of multidimensional persistent homology is depicted in Fig. \[fig:2ygd\]. Clearly, the maps of multidimensional persistent homology given in Fig. \[fig:2ygd\] can be employed for deep learning, which is an open field.
Basically each independent parameter that regulates the filtration process contributes a genuine persistent dimension. When the dimension is higher than two, the result representation is no longer straightforward. Additionally, how to make use of multidimensional persistence in realistic applications is also an interesting problem.
Differential geometry theory of surfaces {#sec:DGA}
----------------------------------------
Differential geometry has fruitful applications in physics, particularly, general relativity and has found its success in biomolecular systems as well [@Bates:2008; @Wei:2009; @ZhanChen:2010a]. As stated earlier, in biophysical modeling, surface representation is a crucial subject and commonly used surface definitions lead to geometric singularities. Gaussian surface and general FRI rigidity surface are based on simple geometric ideas. In contrast, differential geometry theory of surfaces gives rise to natural description macromolecular surfaces [@Bates:2008; @Wei:2009; @ZhanChen:2010a]. This approach becomes powerful when it is combined with variation calculus for biomolecular modeling as shown in Section \[sec:scalar\_field\_geometry\]. In this section, a brief introduction is given about the differential geometry theory of surfaces using the notations and definitions from Ref. [@Kuhnel:2015].
#### Surface elements and immersion
For an open set $U \subset \mathds{R}^2$, a parametrized surface element is an immersion $f: U \rightarrow \mathds{R}^3$. Here $f$ is also known as a parametrization. One can call the elements of $U$ as parameters and their images under $f$ as points. If the rank of map $f$ is maximal, $f$ is an immersion. The point where the rank is not maximal, it is called a singular point or singularity [@Kuhnel:2015].
One can use the following notations for a parametrized surface element $f:U\rightarrow \mathds{R}^3, u \in U, p=f(u)$. $T_uU$ is the tangent space of $U$ at $u$, $T_uU=\{u\}\times \mathds{R}^2 $; $T_p\mathds{R}^3$ is the tangent space of $\mathds{R}^3$ at $p$, $T_p\mathds{R}^3=\{p\}\times \mathds{R}^3 $; $T_uf$ is the tangent space of $f$ at $p$, $T_uf:=Df|_u(T_uU) \subset T_{f(u)} \mathds{R}^3 $; and $\perp_uf$ is the normal space of $f$ at $p$ $T_uf \oplus \perp_uf =T_{f(u)} \mathds{R}^3 $. The element of $T_uf$ is called tangent vector and the element of $\perp_uf$ is the normal vector [@Kuhnel:2015].
#### First fundamental form
The first fundamental form $I$ is the inner product between two tangent vectors $X,Y$ in tangent planes $T_uf$, i.e., $I(X,Y):=<X,Y>$. For coordinate systems $f(u,v)=\left( x(u,v),y(u,v),z(u,v) \right)$, the first fundamental form can be described by the following tensor matrix [@Kuhnel:2015] $$(g_{ij})=\left( \begin{array}{cc}
E(u,v) & F(u,v) \\
F(u,v) & G(u,v)
\end{array}
\right)= \left(
\begin{array}{cc}
I(\frac{\partial f}{\partial u},\frac{\partial f}{\partial u}) & I(\frac{\partial f}{\partial u},\frac{\partial f}{\partial v}) \\
I(\frac{\partial f}{\partial u},\frac{\partial f}{\partial v}) & I(\frac{\partial f}{\partial v},\frac{\partial f}{\partial v}) \end{array}
\right)= \left(
\begin{array}{cc}
<\frac{\partial f}{\partial u},\frac{\partial f}{\partial u}> & <\frac{\partial f}{\partial u},\frac{\partial f}{\partial v}> \\
<\frac{\partial f}{\partial u},\frac{\partial f}{\partial v}> & <\frac{\partial f}{\partial v},\frac{\partial f}{\partial v}>
\end{array}
\right)$$ Also one can have the line element $$ds^2=E(u,v)du^2 +2F(u,v)dudv+G(u,v)dv^2$$ and the surface area $$dA=\sqrt{g}dudv,$$ where $g= {\rm Det}(g_{ij})$ is the determinant.
#### Gauss map
Since each plane is essentially determined by its normal vector, the curvature of the surface can be studied by the variation of the normal vector, i.e., Gauss map. For a surface element $f: U \rightarrow \mathds{R}^3$, the Gauss map is $v:U \rightarrow S^2$ and is defined by the formula [@Kuhnel:2015] $$v(u_1,u_2):=\frac{\frac{\partial f}{\partial u_1} \times \frac{\partial f}{\partial u_2}}{|\frac{\partial f}{\partial u_1} \times \frac{\partial f}{\partial u_2}|},$$ where $S^2$ denotes the unite sphere $S^2=\{(x,y,z)\in \mathds{R}^3 | x^2+y^2+z^2=1\}$.
#### Weingarten map
Let $f: U \rightarrow \mathds{R}^3$ be a surface element with Gauss map $v:U \rightarrow S^2 \in \mathds{R}^3$, and for every $u \in U$ the image plane of the linear map $Dv|_u:T_uU \rightarrow T_{v(u)}\mathds{R}^3$ is parallel to the tangent plane $T_uf$. By canonically identifying $T_{v(u)}\mathds{R}^3 \cong \mathds{R}^3 \cong T_{f(u)}\mathds{R}^3 $, one can have $Dv$ at every point as the map $Dv|_u:T_uU \rightarrow T_uf$. Moreover, by restricting to the image, one may view the map $Df|_u$ as a linear isomorphism $Df|_u:T_uU \rightarrow T_uf$. In this sense the inverse mapping $(Df|_u)^{-1}$ is well-defined and is also an isomorphism. The map $L:=-Dv\circ (Df)^{-1}$ defined point-wisely by $$L_u:=-(Dv|_u)\circ (Df|_u)^{-1}: T_uf \rightarrow T_uf$$ is called a Weingarten map or the shape operator of $f$. This map is independent of the parametrization of $f$, and it is self-adjoint with respect to the first fundamental form I.
#### Second and third fundamental form
Let $f: U \rightarrow \mathds{R}^3$ be a surface element with Gauss map $v:U \rightarrow S^2 \in \mathds{R}^3$. With the shape operator $L$ and tangent vectors $X$ and $Y$, the second fundamental $II$ is given by $$II(X,Y):=I(LX,Y)$$ and the third fundamental form is $$III(X,Y):=I(L^2X,Y)=I(LX,LY).$$ $II$ and $III$ are symmetric bilinear forms on $T_uf$ for every $u \in U$. The three fundamental forms $I$,$II$ and $III$ have a relation as, $$III-{\rm Tr}(L)II+{\rm Det}(L)I=0.$$ In coordinates $f(u,v)=\left( x(u,v),y(u,v),z(u,v) \right)$, three fundamental forms can be expressed as $$\begin{aligned}
&&I: g_{ij}=<\frac{\partial f}{\partial u_i},\frac{\partial f}{\partial u_j}>; \\
&&II: h_{ij}=<v,\frac{\partial^2 f}{\partial u_i \partial u_j}>=-<\frac{\partial v}{\partial u_i}, \frac{\partial f}{\partial u_j}>; \\
&&III: e_{ij}=<\frac{\partial v}{\partial u_i},\frac{\partial v}{\partial u_j}>.\end{aligned}$$
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Illustration of Gaussian curvature (left side) and mean curvature (right side) for an HIV-1 gp 120 trimer structure (EMD-5020).[]{data-label="fig:curvature_g_k"}](curvature_g_k.png "fig:"){width="50.00000%"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
#### Principal curvature
For a unite tangent vector $X \in T_uf$ and $I(X,X)=1$, it is a principal curvature direction for $f$ if $II(X,X)$ has a stationary value for $X$, or $X$ is an eigenvalue of the Weingarten map $L$. Further, the corresponding eigenvalue $\kappa$ is the principal curvature.
#### Gaussian and mean curvature
Gaussian curvature is the determinate $K={\rm Det}(L)=\kappa_1\cdot \kappa_2$ and mean curvature is the average value $H=\frac{1}{2}{\rm Tr}(L)=\frac{1}{2}(\kappa_1+\kappa_2)$. Gaussian and mean curvatures can be expressed in local coordinates as $$\begin{aligned}
&&K=\frac{ {\rm Det}(h_{ij})}{{\rm Det}(g_{ij})}=\frac{h_{11}h_{22}-h_{12}^2}{g_{11}g_{22}-g_{12}^2},\\
&&H=\frac{1}{2}\Sigma_i h^i_i=\frac{1}{2}\Sigma_{i,j} h_{ij} g^{ij}= \frac{1}{2 {\rm Det}(g_{ij})}(h_{11}g_{22}-2h_{12}g_{12}+h_{22}g_{11}).\end{aligned}$$ Gaussian and mean curvatures of an HIV virus fragment are illustrated in Fig. \[fig:curvature\_g\_k\]. In biophysics, the region with negative Gaussian is often associated membrane-protein interaction sites, while the region with negative mean curvatures on a protein surface is commonly regarded as a potential-ligand or protein-drug bonding site [@KLXia:2014a].
Differential geometry modeling and computation {#sec:scalar_field_geometry}
----------------------------------------------
With the advance of experimental technology, more than a hundred thousand of 3D macromolecular structural data has been accumulated. Differential geometry based surface modeling and curvature measurement are of essential importance to the geometric description and feature recognition of these 3D structural data [@Wei:2005; @Xu:DSM:2006]. In this section, a brief review is given to the differential geometry based modeling of macromolecular surfaces. Two algorithms for curvature calculations are also discussed.
### Minimal molecular surface {#Sec:MMSgeneration}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Comparison of the solvent excluded surface (Left chart) and the minimal molecular surface (Right chart) of protein 1PPL. The minimal molecular surface is free from geometric singularities.[]{data-label="fig:MMS"}](MMS.png "fig:"){width="50.00000%"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Minimal molecular surface (MMS) was introduced to construct surfaces free of geometric singularities via the variational principle [@Bates:2006; @Bates:2008]. In this approach, a hypersurface $S$ is defined to represent the biomolecular surface. Basically, one assigns each point with coordinate $(x,y,z)$ a value $S(x,y,z)$, which represents the domain information. It can be viewed as a characteristic function of the macromolecular domain. By using geometric measure theory, the surface energy functional can be expressed as [@Wei:2009] $$\begin{aligned}
\label{eq2surface}
G_{\rm{surface}}= \gamma {\rm{Area}}= \int_{{\mathbb R}^3} \gamma |\nabla S | d{\bf{r}},\end{aligned}$$ where $\gamma$ is the surface tension. As it is convenient for us to set up the total free functional as a 3D integral in ${\mathbb R}^3$, one can make use of the concept of a mean surface area [@Wei:2009; @ZhanChen:2011a] and the coarea formula [@coarea] on a smooth surface $$\begin{aligned}
\label{eqarea}
{\rm{Area}}=\int_0^1 \int_{S^{-1}(c)\bigcap\Omega} d\sigma dc = \int_{\Omega} | \nabla S({\bf{r}}) | d{\bf{r}}, \quad \Omega \subset {\mathbb R}^3.\end{aligned}$$ [Here the value of hypersurface function $S$ is distributed between 0 and 1, $S^{-1}(c)$ represents the inverse function of $S$ and $\Omega$ is defined as the whole domain.]{} The variation of Eq. (\[eq2surface\]) with respect to $S$ leads to the vanishing of surface-tension weighted mean curvature, $\nabla\cdot\left(\gamma \frac{\nabla S}{|\nabla S|}\right)=0$. The energy minimization of Eq. (\[eq2surface\]) can be realized by the introduction of an artificial time to obtain a generalized Laplace-Beltrami equation $$\begin{aligned}
\label{MeanCF}
\frac{\partial S}{\partial t}&=&|\nabla S|\nabla\cdot\left(\frac{\gamma\nabla S}{|\nabla S|}\right),\end{aligned}$$ The final MMS, which is free of geometric singularity, is obtained by extracting an iso-surface from the steady state hypersurface function. During each iteration, one can keep the value of $S$ in the van der Waals surface enclosed domain unchanged. Figure \[fig:MMS\] illustrates the difference between the solvent excluded surface and the minimal molecular surface of protein 1PPL.
In the earlier works, sophisticated computational algorithms have been developed to accelerate the construction of MMSs for large biomolecules [@Bates:2009; @ZhanChen:2010a; @ZhanChen:2010b; @XFeng:2012a]. Differential geometry based molecular surface modeling was extended to solvation analysis [@Wei:2009; @ZhanChen:2010a], including level set approaches [@Cheng:2007e; @Cheng2:2009], and ion channel transport [@Wei:2009; @Wei:2012; @Wei:2013; @DuanChen:2013]. Numerical aspects were examined in the literature [@SZhao:2011a; @SZhao:2014a]. Since these approaches work very well for biomolecular structure, function and dynamics, they will attract much attention in the future. However, a more detailed discussion of these issues is beyond the scope of the present review.
### Scalar field curvature evaluation {#sec:algorithm}
For a given set of volumetric biomolecular data, the efficient and accurate computation of curvatures is needed. The evaluation of curvature properties from iso-surface embedded volumetric data has been studied in geometric modeling, although the related techniques have not received much attention in computational biophysics. In this section, two popular algorithms are reviewed. Many other elegant methods, including isophote surface based curvature evaluation [@Verbeek:1993], Sander-Zucker approach [@Sander:1990], Direct surface mapping based approach [@Stokely:1992], piece-wise linear manifold techniques [@Stokely:1992], etc, are often used in the computer science community.
#### Algorithm I
Essentially, the first and second fundamental forms in the differential geometry are involved in the definition and evaluation of the curvatures. a brief discussion of the mathematical background is given [@Soldea:2006; @Bates:2008].
The surface of interest can be extracted from a level set with iso-value $S_0$, i.e., $S(x,y,z)=S_0$. One can assume $S$ to be non-degenerate, i.e., the norm of the gradient is non-zero at $S(x,y,z)=S_0$. Without loss of generality, one can assume that the projection onto $z$ is non-zero as well. Then the implicit function theorem states that locally, there exists a function $z=f(x,y)$, which parametrizes the surface as ${\bf S}(x,y)=(x,y,f(x,y))$. One can express the iso-value relation as $S(x,y,f(x,y))=S_0$. The differentiation with respect to $x$ and $y$ variables leads to two more equations $$\begin{aligned}
\nonumber
S_x(x,y,f(x,y))+S_z(x,y,f(x,y))f_x(x,y)=0,\\\nonumber
S_y(x,y,f(x,y))+S_z(x,y,f(x,y))f_y(x,y)=0,\end{aligned}$$ where $f_x(x,y)$ and $f_y(x,y)$ can be given by $$\begin{aligned}
\nonumber
f_x(x,y)=-\frac{S_x(x,y,z)}{S_z(x,y,z)}; \quad {\rm and} \quad f_y(x,y)=-\frac{S_y(x,y,z)}{S_z(x,y,z)}.\end{aligned}$$ One can define $E(x,y,z), F(x,y,z), G(x,y,z), L(x,y,z), M(x,y,z)$ and $ N(x,y,z)$ below to be the coefficients in the first and second fundamental forms. For simplicity, one can hide parameter labels. Their values for surface function ${\bf S}=(x,y,f)$ can be given as $$\begin{aligned}
\nonumber
E &=& \langle {\bf S}_x , {\bf S}_x\rangle=1+f_x^2=1+\frac{S_x^2}{S_z^2};\\\nonumber
F &=& \langle {\bf S}_x , {\bf S}_y\rangle=f_xf_y=\frac{S_x S_y}{S_z^2};\\\nonumber
G &=& \langle {\bf S}_y , {\bf S}_y\rangle=1+f_y^2=1+\frac{S_y^2}{S_z^2};\\\nonumber
L &=& \langle {\bf S}_{xx} , {\bf n}\rangle=\frac{2S_xS_zS_{xz}-S_x^2S_{zz}-S_z^2S_{xx}}{g^{\frac{1}{2}}S_z^2};\\\nonumber
M &=& \langle {\bf S}_{xy} , {\bf n}\rangle=\frac{S_xS_zS_{yz}+S_yS_zS_{xz}-S_xS_yS_{zz}-S_z^2S_{xy}}{g^{\frac{1}{2}}S_z^2};\\\nonumber
N &=& \langle {\bf S}_{yy} , {\bf n}\rangle=\frac{2S_yS_zS_{yz}-S_y^2S_{zz}-S_z^2S_{yy}}{g^{\frac{1}{2}}S_z^2}.\end{aligned}$$ The Gaussian curvature can be expressed as the ratio of the determinants of the second and first fundamental forms, $$\begin{aligned}
\nonumber
&&K=\frac{2S_xS_yS_{xz}S_{yz}+2S_xS_zS_{xy}S_{yz}+2S_yS_zS_{xy}S_{xz}}{g^2}-\frac{2S_xS_zS_{xz}S_{yy}+2S_yS_zS_{xx}S_{yz}+2S_xS_yS_{xy}S_{zz}}{g^2} \\\nonumber
&& \qquad \left. +\frac{S_z^2S_{xx}S_{yy}+S_x^2S_{yy}S_{zz}+S_yS_{xx}S_{zz}}{g^2} -\frac{S_x^2S_{yz}^2+S_y^2S_{xz}^2+S_z^2S_{xy}^2}{g^2} \right..\end{aligned}$$
Similarly, the mean curvature is given as the average second derivative with respect to the normal direction, $$\begin{aligned}
\label{eq:meanC}
H=\frac{2S_xS_yS_{xz}+2S_xS_zS_{xz}+2S_yS_zS_{yz}-(S_y^2+S_z^2)S_{xx}-(S_x^2+S_z^2)S_{yy}-(S_x^2+S_y^2)S_{zz}}{2g^{\frac{3}{2}}}.\end{aligned}$$ Note that curvature expressions become analytical when the implicit surface is given by the FRI rigidity density, Eq. (\[eq:rigidity3\]).
#### Algorithm II
An alternative algorithm for the curvature extraction from volumetric data is the Hessian matrix method [@Kindlmann:2003]. For volumetric data $S(x,y,z)$, one defines the surface gradient ${\bf g}$ and surface norm ${\bf n}$. $$\begin{aligned}
&& {\bf g} = \nabla S= (S_x,S_y,S_z)^T;\\
&& {\bf n = -\frac{g}{|g|}}.\end{aligned}$$ Here ${T}$ denoting the transpose. One further calculates the matrix $\nabla {\bf n}^T$, which can be expressed as $$\begin{aligned}
&& \nabla {\bf n}^T=- \nabla ( { \frac{{\bf g}}{|{\bf g}|}})=-( { \frac{\nabla {\bf g}^T}{|{\bf g}|}-\frac{{\bf g} \nabla^T |{\bf g}|}{|{\bf g}|^2}}) \\\nonumber
&& \qquad =-\frac{1}{|{\bf g}|}({H}-\frac{{\bf g} \nabla^T ({\bf g}^T{\bf g})^{\frac{1}{2}}}{|{\bf g}|})=-\frac{1}{|{\bf g}|}({H}-\frac{{\bf g} \nabla^T ({\bf g}^T{\bf g})}{2|{\bf g}|({\bf g}^T{\bf g})^{\frac{1}{2}}}) \\\nonumber
&& \qquad =-\frac{1}{|{\bf g}|}({H}-\frac{{\bf g}(2{\bf g}^TH)}{2|{\bf g}|^2})=-\frac{1}{|{\bf g}|}({ I}-\frac{{\bf g}{\bf g}^T}{|{\bf g}|^2}){H} \\\nonumber
&& \qquad =-\frac{1}{|{\bf g}|}({ I}-{\bf n}{\bf n}^T){H}=-\frac{1}{|{\bf g}|} {PH},\end{aligned}$$ where ${I}$ is the identity matrix, matrix $P=I-{\bf nn}^T$ and Hessian matrix ${H}$ is given by $$\begin{aligned}
{H}=\left[
\begin{array}{ccc}
\frac{\partial^2 S}{\partial^2 x} &\frac{\partial^2 S}{\partial x\partial y} &\frac{\partial^2 S}{\partial x\partial y} \\
\frac{\partial^2 S}{\partial x\partial y} &\frac{\partial^2 S}{\partial^2 y} &\frac{\partial^2 S}{\partial y\partial z}\\
\frac{\partial^2 S}{\partial x\partial z} &\frac{\partial^2 S}{\partial y\partial z} &\frac{\partial^2 S}{\partial^2 z}
\end{array}
\right].\end{aligned}$$ Geometrically, ${\bf nn}^T$ project ${\bf n}$ onto a one-dimensional span of ${\bf n}$. Here $I-{\bf nn}^T$ further projects onto the orthogonal space complement to the span of ${\bf n}$, which is the tangent plane.
In general, both ${ P}$ and ${H}$ are symmetric but ${\bf \nabla n}^T$ is not. If ${\bf q_1}$ lies in the tangent plane, $P{\bf q_1}={\bf q_1}$ and ${\bf q_1}^T P={\bf q_1}^T$. Therefore, for ${\bf q_2}$ and ${\bf q_1}$ in the tangent plane, one has $$\begin{aligned}
{\bf q_1}^T PH {\bf q_2=q_1}^T H {\bf q_2=q_2}^T H{\bf q_1=q_2}^T PH{\bf q_1}\end{aligned}$$
The restriction of $\nabla {\bf n}^T=-\frac{1}{|{\bf g}|} { PH}$ to the tangent plane is symmetric and thus there exists an orthonormal basis ${\bf p_1,p_2}$ for the tangent plane in which ${\bf n}^T$ is a 2\*2 diagonal matrix. This basis can be easily extended to an orthonormal basis for all $\{{\bf p_1}, {\bf p_2}, {\bf n} \}$. In this basis, the derivative of the surface normal is given by $$\begin{aligned}
{\bf \nabla n}^T=\left[
\begin{array}{ccc}
\kappa_1 &0 &\sigma_1 \\
0 &\kappa_2 &\sigma_2\\
0 &0 &0
\end{array}
\right].\end{aligned}$$ The diagonal term in the bottom row is zero because no change in normal ${\bf n}$ can lead to a change in length. Motion along ${\bf p_1}$ and ${\bf p_2}$ results in the change of ${\bf n}$ along the same directions, with a ratio of ${\kappa_1}$ and ${\kappa_2}$ respectively. Here ${\bf p_1}$ and ${\bf p_2}$ are the principal curvature directions, while ${\kappa_1}$ and ${\kappa_2}$ are the principal curvatures. When there is a change in normal ${\bf n}$, it tilts according to ${\sigma_1}$ and ${\sigma_2}$. One can further multiply ${\bf \nabla n}^T$ by ${ P}$ to diagonalize the matrix $$\begin{aligned}
G={\bf \nabla n}^T{ P}= {\bf \nabla n}^T \left[
\begin{array}{ccc}
1 &0 &0 \\
0 &2 &0\\
0 &0 &0
\end{array}
\right]=\left[
\begin{array}{ccc}
\kappa_1 &0 &0 \\
0 &\kappa_2 &0\\
0 &0 &0
\end{array}
\right].\end{aligned}$$
The surface curvature measurements are based on geometry tensor ${ G}$. In a volumetric data set or a scalar field, ${G}$ is known in terms of the Cartesian basis $(x, y, z)$. Matrix invariants provide the leverage to extract the desired curvature values $\kappa_1$ and $\kappa_2$ from ${ G}$, regardless of the coordinate frame of the principal curvature direction. The trace of ${ G}$ is $\kappa_1+ \kappa_2$. The Frobenius norm of ${ G}$, notated $|{ G}|_F$ and defined as $\sqrt{{\rm Tr}({ GG}^T)}$, is $\sqrt{\kappa_1^2+\kappa_2^2}$. Then $\kappa_1$ and $\kappa_2$ are found with the quadratic formula.
The two principal curvatures can be evaluated by the following procedure.
1. Calculate matrix $P=I-{\bf nn}^T$
2. Evaluate matrix $G=I-\frac{PHP}{|{\bf g}|}$, $$\begin{aligned}
{ G}=(g_{ij})_{(i,j=1,3)}\end{aligned}$$
3. Compute the trace $t$ and Frobenius norm $f$ of matrix ${ G}$; $$\begin{aligned}
&& t=g_{11}+g_{22}+g_{33};\\
&& f=\|{ G} \|=\sqrt{\sum_i \sum_j g_{ij}^2};\\
&& \kappa_1=\frac{t+\sqrt{2f^2-t^2}}{2};\\
&& \kappa_2=\frac{t-\sqrt{2f^2-t^2}}{2}.\end{aligned}$$
When the two principal curvatures are available, the Gaussian curvature $K$ and mean curvature $H$ can be obtained as $$\begin{aligned}
&&K=\kappa_1 \kappa_2;\\
&&H=\frac{\kappa_1+\kappa_2}{2}.\end{aligned}$$ Essentially, the Hessian matrix method generates the same results as Algorithm I derived from the first and second fundamental form. However, for FRI rigidity density given in Eq. (\[eq:rigidity3\]), Algorithm I is preferred as it is analytical and without matrix diagonalization.
### Analytical minimal molecular surface {#sec:density2}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![An illustration of analytical minimal molecular surfaces generated at different $\sigma$ values. From [ (a)]{} to [ (d)]{}, the $\sigma$ is chosen as 0.5, 0.6, 0.7 and 0.8, respectively. For all figures, isosurfaces are extracted at mean curvature isovalue of 0.001. []{data-label="fig:minimal_surface"}](minimal_surface.png "fig:"){width="50.00000%"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![An illustration of analytical minimal molecular surfaces generated at different $\sigma$ values. From [ (a)]{} to [ (d)]{}, the $\sigma$ is chosen as 0.5, 0.6, 0.7 and 0.8, respectively. For all figures, isosurfaces are extracted at mean curvature isovalue of 0.001. []{data-label="fig:minimal_surface"}](ms_scale.png "fig:"){width="60.00000%"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Minimal surfaces are ubiquitous in nature and are a fascinating topic for centuries. The Euler-Lagrange minimization of the hypersurface was proposed to generate minimal molecular surface (MMS) for biomolecules [@Bates:2008]. In general, MMS is a result of vanishing mean curvature and is free of geometric singularity. It is a powerful concept in biophysical modeling [@Bates:2008; @Wei:2009]. Nevertheless, it is still much more computationally expensive to construct MMS than to generate SES. In this work, a new minimal molecular surface, called analytical minimal molecular surface (AMMS) is proposed. In this approach, the rigidity density representation of biomolecular data by Eq. (\[eq:rigidity3\]) is employed. Then the mean curvature can be [*analytically*]{} computed by using Eq. (\[eq:meanC\]). Finally, AMMS can be constructed by setting the isosurface of the mean curvature to zero (or practically, a number very close zero).
One can consider the hexagonal ring and a fullerene C$_{20}$ molecule in the study of AMMS. The coordinates used for particles in the hexagonal ring are (0.000, 1.403, 0.000; -1.215, 0.701, 0.000; -1.215, -0.701, 0.000; 0.000, -1.403, 0.000; 1.215, -0.701, 0.000; and 1.215, 0.701, 0.000). The fullerene C$_{20}$ molecule has a highly symmetric cage structure made of 12 pentagons. Its atomic coordinates are (1.569, -0.657, -0.936; 1.767, 0.643, -0.472; 0.470, -0.665, -1.793; 0.012, 0.648, -1.826; 0.793, 1.467, -1.028; -0.487, -1.482, -1.216; -1.564, -0.657, -0.895; -1.269, 0.649, -1.277; -0.002, -1.962, -0.007; -0.770, -1.453, 1.036; -1.758, -0.638, 0.474; 1.288, -1.450, 0.163; 1.290, -0.660, 1.305; 0.012, -0.646, 1.853; 1.583, 0.645, 0.898; 0.485, 1.438, 1.194; -0.503, 0.647, 1.775; -1.606, 0.672, 0.923; -1.296, 1.489, -0.166; -0.010, 1.973, -0.006). The fullerene rigidity density is given by Eq. (\[eq:rigidity3\]). In both cases, the parameters $w_i$ and $\sigma_i$ are set to $1$ and $0.7$, respectively.
When the mean curvature isovalue equals to zero, the resulting surfaces are usually composed by several non-intersecting surfaces perpendicular to atomic bonds near the BCPs. When loosing the condition a little bit by setting the isovalue to 0.001 (or some other very small positive value), a better AMMS can be generated. Figure \[fig:minimal\_surface\] depicts the AMMS for the hexagonal ring and C$_{20}$ molecule. It is found that for the hexagonal ring, each atom is enclosed by a surface segment. These segments are tightly close to each other and only connect near BCPs. For the C$_{20}$ molecules, the surface appears to be better connected. However, a careful examination reveals the separation as well. Obviously, these surfaces are not minimal surfaces.
To generate better understanding of the minimal molecular surface generation, one can explore several $\sigma$ values for the hexagonal ring. Figure \[fig:minimal\_surface\] illustrates the results. From [ (a)]{} to [ (d)]{}, $\sigma$ is chosen as 0.5, 0.6, 0.7 and 0.8, respectively. All the isosurfaces are extracted at mean curvature equals to 0.001. It is seen that a relatively large $\sigma$ value produces a better minimal surface. The gaps between atomic segments are gradually narrow as the increase of $\sigma$ value and finally disappear, which gives rise to a good quality AMMS.
Scalar and vector field topology {#sec:scalar_field_topology}
--------------------------------
Topological approaches have become an integral part in data analysis, visualization, and mathematical modeling for volumetric as well as for vector valued data sets. In fact, the results of scalar and vector field topology coincide each other for gradient vector fields, although the respective mathematical approaches are originated from different fields. Vector field topology is usually developed for analyzing the streamlines of fluid flow generated by the velocity vector during the flow evolution. These techniques can be applied to macromolecular analysis for dealing with cryo-EM data and improving the structural construction.
Three dimensional scalar field data, particular the molecular structural data are widely available from many data sources. To decipher useful information from these data requires highly efficient methods and algorithms. Mathematical approaches, including topological tools, such as fundamental groups, homology theory, contour tress, Reeb graphs [@Dey:2013], Morse theory [@harker:mischaikow:mrozek:nanda; @Mischaikow:2013], etc, have a great potential for revealing the intrinsic connectivity or structure-function relationship. Topological data analysis (TDA) has gained much attention in the past decade. The traditional atoms in molecules (AIM) analysis [@Bader:1985; @Bader:1990] can be viewed as an application of TDA to electron density analysis. Historically, the topological study of scalar fields, particularly electron densities, has advanced the understanding of molecules and their chemical and biological functions tremendously. The theory of AIM developed by Bader and his coworkers has provided an elegant and feasible approach to define atoms in molecule. AIM can be generalized into a more general theory called quantum chemical topology (QCT), which applies the topological analysis in AIM to the study of other physically meaningful scalar fields [@Popelier:2005]. Mathematically, this topological analysis used in AIM is known as Morse theory. In general, all scalar fields used in QCT are some kind of Morse functions and more can be proposed as long as they satisfy the Mores function constraints. On the other hand, geometric modelings and analysis contribute a lot to the molecular visualization and structure-function relationship. Various surface definitions from geometric modelings facilitate the visualization and characterization of molecules. Geometric analysis has contributed enormously to the understanding of structure-function relationship. Among the geometric analysis, surface curvature analysis has provide great insights into solvation analysis, protein-protein interaction, drug design, etc. [@KLXia:2014a; @DDNguyen:2016c]
In the QCT theory, researchers explore atomic or molecular properties through the analysis of various properties at critical points. In geometric analysis, curvature estimation is normally done on a special surface, such as SES, SAS, Gaussian surface, FRI rigidity surface, etc. Even though these two approaches are very efficient and powerful in capturing and characterizing atomic and molecular information, enormous information embedded in the electron density scalar field has not been fully utilized. Question is how to improve geometric and topological analysis for biomolecules. Note that in persistent homology analysis, physical properties are analyzed by a systematic filtration process. Therefore, for a given scalar field, a better understanding can be achieved if the topological analysis includes not only some critical points or a special iso-surface, but also a series of iso-surfaces derived from a systematic evolution of iso-values. Further, more information can be extracted if one considers not only simply scalar fields, such as electron density, electron density Laplacian, etc, but also geometric properties like individual eigenvalues and various curvatures. These approaches are developed in the present work.
In the following, a brief review of some very basic concepts in AIM is given. Their connection with mathematical theories is discussed. Then, some new approaches are presented. Examples of applications are provided.
### Critical points and their classification {#sec:CP}
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![An illustration of four types of critical points, i.e., nucleic critical point (NCP), bond critical point (BCP), ring critical point (RCP) and cage critical point (CCP). [**(a)**]{} The demonstration of flowlines and CPs for a benzene molecule (Hydrogen atoms are not considered for simplicity). [**(a)**]{} The CPs for a Cubane (Hydrogen atoms are not considered for simplicity). []{data-label="fig:cp"}](cp.png "fig:"){width="60.00000%"}
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Three dimensional molecular electron density data can be collected through experimental tools, like electron microscopy, cryo-electron microscope, etc and also from theoretical models in quantum mechanics. One can denote electron density function $\rho$ and its domain $\Omega$. A quantitative method to analyze the topology of $\rho$ is to consider its first derivative, i.e., gradient $\nabla \rho$. At certain points, called critical points (CPs), this gradient vanishes. The characteristics of these points is determined by the second derivatives, which form the Hessian of electron density $\rho$. By diagonalization of the Hessian matrix, one can obtain three eigenvalues $\gamma_1 \leq \gamma_2 \leq \gamma_3$. Their sum equals to the Laplacian of electron density $\rho$. That is, locally [@Bader:1990] $$\begin{aligned}
\nabla^2 \rho= \gamma_1+\gamma_2 + \gamma_3=\frac{\partial^2 \rho}{\partial^2 x} +\frac{\partial^2 \rho}{\partial^2 y} +\frac{\partial^2 \rho}{\partial^2 z}.\end{aligned}$$ It can be noticed that the Hessian matrix is symmetric, therefore all the eigenvalues are real. Based on the positive and negative sign of these three eigenvalues, one defines rank and signature to characterize critical points. The rank of a CP is the number of non-zero eigenvalues, and a signature is the algebraic sum of the signs (+1 or -1) of the eigenvalues. In general, a CP is non-degenerate, meaning its three eigenvalues are non-zero (Rank$=3$). A degenerate critical point is unstable in the sense that even a small change in the function will cause it either to vanish or bifurcate into a non-degenerate CP.
Based on the rank and signature value, non-degenerate CPs in three dimensional scalar field can be classified into four basic types, namely, nucleic critical point (NCP), bond critical point (BCP), ring critical point (RCP) and cage critical point (CCP). Table \[tb:cp\] lists these values. An illustration of these four types of CPs can be found in Figure \[fig:cp\]. An NCP is a nucleic center of an atom. It is represented by a cyan-color point. A BCP is a bond center between two atoms and is marked with a red-color point. A RCP is usually found at the center of a ring structure and is colored in blue. A CCP is known as a cage critical point, and can only be found in the center of a cage structure. In general, an NCP is a local maximal point. BCP and RCP both are saddle points. A CCP is a local minimal point. In Sections \[sec:eigenvalue\] and \[sec:T\_C\], it will be demonstrated that these properties can be used to systematically characterize and analyze the eigenvalue and curvature isosuface information.
Topologically, the general structure connectivity can be characterized by the number of CPs. This is stated in the famous Poincaré-Hopf theorem as following: $$\begin{aligned}
\label{Eq:ph}
N_n-N_b+N_r-N_c=\chi (\rho),\end{aligned}$$ where $N_n$, $N_b$, $N_r$ and $N_c$ are numbers of NCPs, BCPs, RCPs and NCPs, respectively. Here $\chi(\rho)$ is the Euler characteristic. The Poincaré index is defined on a vector field. In the AIM theory, one studies the gradient vector field of electron density. Poincaré indices for NCP, BCP, RCP and NCP are $ 1, -1, 1$ and $-1$, respectively. More properties of this gradient vector field are discussed in the following section.
Equation (\[Eq:ph\]) can also be derived from simplicial complex analysis [@KLXia:2014c]. Essentially, NCP can be viewed as a point (0-simplex); BCP corresponds to a straight line (1-simplex); RCP is for a polygon (2-simplex); CCP is then for a polyhedron (3-simplex). In this manner, Euler characteristic can be directly employed and the above equation can be obtained as well.
Rank Signature Poincaré index Simplex Property
----- ------ ----------- ------------------ ----------- --------------
NCP 3 $-3$ 1 0-simplex local maxima
BCP 3 $-1$ $-1$ 1-simplex saddle
RCP 3 1 1 2-simplex saddle
CCP 3 3 $-1$ 3-simplex local minima
: The critical point can be classified into four basic types including: nucleic critical point (NCP), bond critical point (BCP), ring critical point (RCP) and cage critical point (CCP), as demonstrated in Fig. \[fig:cp\].
\[tb:cp\]
### Vector field topology
The gradient $\nabla \rho$ on the entire domain $\Omega$ forms a vector field. The topological property of this vector field can be explored by tools borrowed from dynamic systems or the mathematical analysis of fluid flows. These tools include integral line, separatrix, and Poincaré index, etc.
An integral line is a curve $l(t)$ of a function $f({\bf r})$ that satisfies $\frac{\partial l}{\partial t}=\nabla f(l(t))$. Its origin and destination can be defined as $${\rm org}(l)=\underset{t \rightarrow - \infty}{\lim} l(t)$$ and $${\rm dest}(l)=\underset{t \rightarrow \infty}{\lim} l(t).$$ Integral lines satisfy the following conditions:
1. Two integral lines are either disjoint or the same.
2. Integral lines cover all the manifold.
3. The limits, org$(l)$ and dest$(l)$, are critical points.
In general, integral lines represent the gradient flow between critical points. All the integral lines that share the same $org$ form a region called atomic basin. The whole electron density domain is subdivided into many atomic basins. The interface between these attraction basins is called inter-atomic surface (IAS). Mathematical, IAS is the separatrix of a gradient vector field. Each attraction basin includes one and only one atom. This is known as the quantum topological definition of an atom in a molecule.
IAS also satisfies the zero-flux condition $\nabla f({\bf r}) \cdot {\bf n( r)}=0$. Here $\nabla f({\bf r}) $ is a gradient vector and ${\bf n(r)}$ is the normal vector to the IAS.
#### Morse theory
It should be noticed that basic concepts like critical point, degeneracy, critical point classification, basin, etc. are the essential part of a mathematical tool called Morse theory [@harker:mischaikow:mrozek:nanda; @Mischaikow:2013]. In general, the atom in molecular method can be viewed as an application of Morse theory in molecular density field analysis. More specifically, various critical points, including NCP, BCP, RCP, and CCP, are the counterparts of peak point, saddle-1 point, saddle-2 point and valley point, respectively. The separatrix of a gradient vector field just provides a Morse decomposition of the underlying molecular scale field manifold [@harker:mischaikow:mrozek:nanda; @Mischaikow:2013].
### Topological characterization of chemical bonds {#sec:chemical_bond}
Chemical properties of molecular systems are profoundly determined or influenced by their atomic covalent bonds and noncovalent interactions. Usually, covalent bonds are much stronger and determine the structural integrity of a molecule. Noncovalent interactions are comparably weak but play important roles in macromolecular assembly, protein folding, macromolecular function, etc. Traditionally, the Laplacian of electron density can be used to interpret noncovalent interactions of a molecular system. Recently, signed electron density and reduced gradient, two scalar fields derived from electron density, have drawn much attention in quantum chemistry since they enable a qualitative visualization of noncovalent interactions even in complex molecular systems [@Johnson:2010; @contreras2011analysis; @contreras2011nciplot; @Gillet:2012; @Gunther:2014]. These approaches are reviewed below.
#### The Laplacian of electron density
The Laplacian of electron density $\rho ({\bf r})$ can be used to indicate the electron density concentration and depletion. Essentially, the density is locally concentrated where $\nabla^2 \rho({\bf r})<0$, and locally depleted where $\nabla^2 \rho({\bf r})>0$. In this manner, if one defines the function $L({\bf r})=-\nabla^2 \rho({\bf r})$, the maximum in $L({\bf r})$ denotes the maximum in the concentration of the density. The minimum in $L({\bf r})$ implies a depletion of density. The Laplacian of the electronic charge distribution, $L({\bf r})$, demonstrates the presence of local concentrations of charge in the valence shell of an atom in a molecule. These local maxima faithfully duplicate in number, location, and size of the spatially localized electron pairs of the valence shell electron pair repulsion (VSEPR) model. Thus the Laplacian of the charge density provides a physical basis for the Lewis and VSEPR models [@Bader:1988].
#### Identifying noncovalent interactions
The study of the Hessian matrix and its three eigenvalues has yielded many intriguing results [@Bader:1990]. It has been found that all eigenvalues ($\gamma_1({\bf r})$, $\gamma_2({\bf r})$ and $\gamma_3({\bf r})$) are negative in the vicinity of the nuclei centers. Away from these centers, the largest eigenvalue $\gamma_3({\bf r})$ becomes positive, and varies along the internuclear axis representing covalent bonds. It is also found that $\gamma_1({\bf r})$ and $\gamma_2({\bf r})$ describe the density variation orthogonal to this internuclear axis. More specially, $\gamma_1({\bf r})$ is always negative even it is away from the nuclei. While $\gamma_2({\bf r})$ can be either positive, meaning attractive interactions concentrating electron charge perpendicular to the bond, or negative, meaning repulsive interactions causing density depletion. Using this localized information, the signed electron density $\widetilde{\rho}({\bf r})$ is defined as $$\begin{aligned}
\label{Eq:sed}
\widetilde{\rho} ({\bf r})= {\rm sign}(\gamma_2({\bf r}))\rho({\bf r}).\end{aligned}$$ The signed electron density additionally enables the differentiation of attracting and repulsive interactions.
To further reveal weak noncovalent interactions, the reduced gradient is introduced as following [@Gillet:2012; @Gunther:2014] $$\begin{aligned}
\label{Eq:reduced_gradient}
s({\bf r})=\frac{1}{2(3\pi^2)^{\frac{1}{3}}}\frac{|\nabla \rho({\bf r})|}{\rho({\bf r})^\frac{4}{3}}.\end{aligned}$$ The reduced density gradient describes the deviation in atomic densities due to interactions and has found interesting applications in in analyzing biomolecular structure and function [@Johnson:2010; @Gillet:2012; @Gunther:2014].
Geometric-topological (Geo-Topo) fingerprints of scalar fields {#sec:GTF}
--------------------------------------------------------------
### Geo-Topo fingerprints of Hessian matrix eigenvalue maps {#sec:eigenvalue}
The QCT and AIM analyses of molecules are limited to special locations and given iso-surfaces. In this work, the Hessian eigenvalue analysis of a molecular density is considered, not only for a few critical points, but also for the whole domain. Additionally, the topology over a series of isosurfaces derived from a systematic evolution of isovalues are analyzed. More specifically, in the topological persistence of Hessian matrix eigenvalues, the topological properties of a series of isosurfaces, generated by varying the isovalue from the smallest to the largest, are systematically studied. Therefore, the isosurface value behaves like a filtration parameter. The topological transitions in this series of isosurfaces are emphasized. Another very important aspect of this analysis is to analyze the behavior of isosurfaces through its relation with the four types of CPs. It should also be noticed that these methods are greatly different from other interatomic surface methods [@Pendas:2003; @Popelier:1996], as these models focus on a special surface.
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Hessian matrix eigenvalue maps of the hexagonal ring model at cross section $Z=0$. The behaviors of three eigenvalues, i.e., $\gamma_1$, $\gamma_2$ and $\gamma_3$, are illustrated in [ (a)]{}, [ (b)]{} and [ (c)]{}, respectively. []{data-label="fig:C6_eigen_2d"}](C6_eigen_2d.png "fig:"){width="80.00000%"}
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Eigenvalue maps obtained from different isovalues (or level-set values) for a cubic structure. [ (a)]{} The isosurfaces for the first eigenvalue. The isovalues from [ ($a_1$)]{} to [ ($a_4$)]{} are -3.0, -1.5, 0.1 and 0.9. [ (b)]{} The isosurfaces for the second eigenvalue. The isovalues from [ ($b_1$)]{} to [ ($b_4$)]{} are -1.0, 0.5, 1.0 and 1.5. [ (c)]{} The isosurfaces for the third eigenvalue. The isovalues from [ ($c_1$)]{} to [ ($c_4$)]{} are -1.0, 1.5, 2.0 and 2.5. []{data-label="fig:Cubane_eigen_3d"}](C6_eigen_3d.png "fig:"){width="60.00000%"}
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Eigenvalue maps obtained from different isovalues (or level-set values) for a cubic structure. [ (a)]{} The isosurfaces for the first eigenvalue. The isovalues from [ ($a_1$)]{} to [ ($a_4$)]{} are -3.0, -1.5, 0.1 and 0.9. [ (b)]{} The isosurfaces for the second eigenvalue. The isovalues from [ ($b_1$)]{} to [ ($b_4$)]{} are -1.0, 0.5, 1.0 and 1.5. [ (c)]{} The isosurfaces for the third eigenvalue. The isovalues from [ ($c_1$)]{} to [ ($c_4$)]{} are -1.0, 1.5, 2.0 and 2.5. []{data-label="fig:Cubane_eigen_3d"}](Cubane_eigen_3d.png "fig:"){width="60.00000%"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
To illustrate the idea, two toy models, i.e., the hexagonal ring discussed earlier and a cubic structure are considered. The coordinates for the cubic structure are set to (1.245, 0.537, -0.073; 0.924, -0.995, 0.024; -0.123, -0.704, 1.155; 0.199, 0.828, 1.058; 0.123, 0.704, -1.155; -0.924, 0.995, -0.024; -1.245, -0.537, 0.073; and -0.199, -0.828, -1.058). The discrete to continuum mapping, Eq. (\[eq:rigidity3\]), is used to generate rigidity density. The parameters $\sigma$ and $w_i$ are chosen as $0.7$ Å and $1$ for all particles. In this approach, Hessian matrix is evaluated at each point of the computational domain and its eigenvalue is obtained everywhere as well, which forms an eigenvalue scalar field.
To have a general idea of the basic distribution of eigenvalues, one can study the Hessian matrix eigenvalue behavior of the hexagonal ring in a two-dimensional plane $Z=0$, as all its particles are located within this special plane. Results are illustrated in Fig. \[fig:C6\_eigen\_2d\]. Three eigenvalues, i.e., $\gamma_1$, $\gamma_2$ and $\gamma_3$, are demonstrated in subfigure [**(a)**]{}, [**(b)**]{} and [**(c)**]{}, respectively. It can be seen that for regions near NCPs, all three eigenvalues are negative. For regions near BCPs, $\gamma_1$ is always negative. While $\gamma_2$ is negative in the very closed neighborhood and gradual increases to be positive further away. Finally, $\gamma_3$ is always positive. For regions near RCPs, all three eigenvalues are positive. These results are consistent with findings in AIM [@Bader:1990].
To obtain more geometric insights, particularly eigenvalue behaviors on all four types of CPs, the cubic structure is considered. In this case, the CCP can be identified as having the positive $\gamma_1$ and $\gamma_2$.
As stated above, this approach is to extract a series of isosufaces of Hessian matrix through a filtration process. One can carefully observe these eigenvalue isosurface patterns to detect topological transitions. The geometric information is further analyzed and its relation with the four types of CPs is summarized into several unique characteristics, called Geo-Topo fingerprints. Using hexagonal ring and cubic structure models, the Geo-Topo fingerprints for eigenvalues are revealed.
Figures \[fig:C6\_eigen\_3d\] and \[fig:Cubane\_eigen\_3d\] demonstrate four unique patters for each eigenvalue. The subscripts $1$ to $4$ indicate four representative eigenvalue isovalues, from the small to the large. The notations $({\bf a})$, $({\bf b})$ and $({\bf c})$ represent eigenvalue $\gamma_1$, $\gamma_2$ and $\gamma_3$, respectively. As stated above, the filtration process delivers a series of isosurfaces. To avoid confusion, the filtration process always goes from the smallest value to the largest one. Only four representative isosurfaces that capture some unique topological features within the eigenvalue isosurface series, are selected. Through the comparison of these patterns, some common features can be extracted.
1. For all the three eigenvalues, regions around NCPs are always concentrated with negative values.
2. For $\gamma_1$, as the filtration goes, negative isosurfaces first appears near NCPs. Then they enlarge to incorporate regions near BCPs, being still negative. The regions near RCPs can be analyzed from two different perspectives: i.e., along the ring plane and perpendicular to the ring plane. Along the ring plane, $\gamma_1$ values gradually increase to about 0 as the eigenvalue isosurface propagate to RCPs. There is a sudden topological transition for the isosuface when its value passes through 0. When becoming positive, eigenvalue isosurfaces form ellipsoids surfaces perpendicular to the ring plane near all RCPs. These ellipsoids usually appear in pairs and symmetric to each other along the ring plane as indicated in Fig. \[fig:C6\_eigen\_3d\] $({\bf a})$. Finally, positive isosurface appears near the regions of CCPs.
3. For $\gamma_2$, as the filtration process goes, negative isosurfaces first appear near NCPs and then enlarge to incorporate regions near BCPs just like $\gamma_1$. However, positive isosurfaces appear much earlier. They occupy the regions around RCPs and CCPs. More interestingly, they form a loop around each bond near its BCP. To be more precise, these loops are perpendicular to atom bonds at BCPs and atom bonds pass through them at their centers. The isosurfaces gradually shrink and reduce to regions around RCPs as their values grows.
4. For $\gamma_3$, its negative isosurfaces concentrate only in regions around NCPs. Positive isosurfaces form a sphere slice around each atom. Isosurface with relative small positive value also appears around the RCPs and CCPs. As the filtration goes further, positive isosurfaces concentrate around regions near BCPs.
### Geo-Topo fingerprints of curvature maps {#sec:T_C}
In this work, the topological analysis of curvature maps are developed. One can still consider the hexagonal ring and the cubic structure discussed in the last section. The discrete to continuum mapping is carried out to generate the FRI rigidity density. Then curvatures are evaluated at all of the molecular FRI rigidity isosurfaces (or every point in the computational domain) to form a curvature map. At each curvature isovalue, topological analysis is applied. Here, topological analysis is twofold. One type of topological analysis is to identify topological critical points (i.e., 0-simplex, 1-simplex, 2-simplex, etc.). The other type is to carry out persistent homology analysis to track topological invariants during the density filtration of the curvature map.
#### Gaussian and mean curvature maps {#sec:K_H}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Isosurfaces of Gaussian and Mean curvature maps for a cubic structure. [ (a)]{} The isosurfaces built from same Gaussian curvature. The isovalues from [ ($a_{1}$)]{} to [ ($a_{4}$)]{} are -20.0, -2.0, 10.0 and 20.0. [ (b)]{} The isosurfaces built from same Mean curvature. The isovalues from [ ($b_{1}$)]{} to [ ($b_{4}$)]{} are -2.0, -1.0, 3.0 and 4.0. []{data-label="fig:Cubane_gaussian_mean_3d"}](C6_gaussian_mean_3d.png "fig:"){width="60.00000%"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Isosurfaces of Gaussian and Mean curvature maps for a cubic structure. [ (a)]{} The isosurfaces built from same Gaussian curvature. The isovalues from [ ($a_{1}$)]{} to [ ($a_{4}$)]{} are -20.0, -2.0, 10.0 and 20.0. [ (b)]{} The isosurfaces built from same Mean curvature. The isovalues from [ ($b_{1}$)]{} to [ ($b_{4}$)]{} are -2.0, -1.0, 3.0 and 4.0. []{data-label="fig:Cubane_gaussian_mean_3d"}](Cubane_gaussian_mean_3d.png "fig:"){width="60.00000%"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
It is seen from the above analysis that with four types of CPs, Geo-Topo fingerprints are extracted and the behaviors of the Hessian matrix eigenvalue isosurfaces can been characterized very well. However, before the employment of this technique in Gaussian and mean curvature isosurface analysis, it is helpful to review a little bit more the four types of CPs. As stated in Section \[sec:CP\], NCPs and CCPs are locally maximal and locally minimal, respectively. Actually, for NCPs, all three eigenvalues have negative signs, while for CCPs, all three eigenvalues have positive signs. BCPs and RCPs are saddle points, which means that their eigenvalues have both positive and negative values. Geometrically, the curvature is isotropic near NCPs and CCPs, but anisotropic near BCPs and RCPs. So in this analysis, the isosurfaces near BCPs and RCPs can be divided into two types, namely, A-type or V-type. A-type isosurface means it is along atomic bonds or ring planes. V-type isosurface means it is vertical or perpendicular to the atomic bonds or ring planes. For instance, the isosurface near the RCP in Fig. \[fig:C6\_gaussian\_mean\_3d\] [**($a_1$)**]{} is an A-type, but it becomes a V-type in Fig. \[fig:C6\_gaussian\_mean\_3d\] [**($a_3$)**]{}. Another important property is that A-type and V-type isosurfaces are usually with different signs. That is, if A-type of isosurface is obtained from a positive isovlaue, the associated V-type isosurface can only be obtained from a negative isosurface. With this notation, one can extract some Geo-Topo fingerprints for Gaussian and mean curvatures. The basic results are summarized as below.
1. For a Gaussian curvature field, it has negative A-type isosurfaces near RCPs and negative V-type isosurfaces near BCPs. In contrast, positive isosurfaces enclose regions near BCPs and CCPs. Positive V-type isosurfaces can be found in RCPs and positive A-type isosurfaces are found near BCPs. Finally, positive isosurfaces are found in NCPs.
2. For a mean curvature field, it has negative isosurfaces near RCPs and CCPs. Negative V-type isosurfaces can be found near BCPs. Positive isosurfaces are found in NCPs. Finally, A-type positive isosurfaces can be found near BCPs.
#### Maximal and minimal curvature maps {#sec:k1k2}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Isosurfaces for maximal and minimal curvature maps for a cubic structure. [**(a)**]{} The isosurfaces built from same maximal curvature. The isovalues from [**($b_{1}$)**]{} to [**($b_{4}$)**]{} are -2.0, 1.0, 3.0 and 4.0. [**(b)**]{} The isosurfaces built from same minimal curvature. The isovalues from [**($a_{1}$)**]{} to [**($a_{4}$)**]{} are -10.0, -4.0, 3.0 and 4.0. []{data-label="fig:Cubane_k1k2_3d"}](C6_k1k2_3d.png "fig:"){width="60.00000%"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Isosurfaces for maximal and minimal curvature maps for a cubic structure. [**(a)**]{} The isosurfaces built from same maximal curvature. The isovalues from [**($b_{1}$)**]{} to [**($b_{4}$)**]{} are -2.0, 1.0, 3.0 and 4.0. [**(b)**]{} The isosurfaces built from same minimal curvature. The isovalues from [**($a_{1}$)**]{} to [**($a_{4}$)**]{} are -10.0, -4.0, 3.0 and 4.0. []{data-label="fig:Cubane_k1k2_3d"}](Cubane_k1k2_3d.png "fig:"){width="60.00000%"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
One can also carry out a throughout investigation of maximal and minimal curvature maps. Isosurfaces near NCPs, BCPs, RCPs and CCPs are analyzed in the same manner as they were done in the above two sections. Figures \[fig:C6\_k1k2\_3d\] and \[fig:Cubane\_k1k2\_3d\] illustrate results for the hexagonal ring and the cubic structure, respectively. Main results about their Geo-Topo fingerprints are summarized below:
1. For maximal curvature map, it has negative isosurfaces near CCPs and negative V-type isosurfaces near RCPs. Positive isosurfaces enclose regions near BCPs and NCPs. Positive A-type isosurfaces can be found in RCPs.
2. For minimal curvature map, it has negative isosurfaces near RCPs and negative V-type isosurfaces near BCPs. Negative isosurfaces also enclose region near CCPs. Positive isosurfaces are found in NCPs. Finally, A-type positive isosurfaces can be found near BCPs.
It can be noticed that all curvature representations, i.e., Gaussian curvature, mean curvature and two principal curvatures, are quite different from Hessian matrix eigenvalue distributions. In general, negative curvatures usually do not occur near NCPs, this is particularly true for maximal and mean curvatures. They are more likely to appear near BCPs, RCPs and CCPs. In contrast, positive isosurfaces are concentrated in the atomic basin (i.e., NCPs).
NCP BCP RCP CCP
------------ ----- ------------------------ ------------------------ -----
$\gamma_1$ N N P P
$\gamma_2$ N N; P-Loop P P
$\gamma_3$ N P P P
$K$ P A-type (P); V-type (N) V-type (P); A-type (N) P
$H$ P A-type (P); V-type(N) N N
$\kappa_1$ P P A-type (P); V-type (N) N
$\kappa_2$ P A-type (P); V-type (N) N N
: Geo-Topo fingerprints for eigenvalue and curvature maps. For simplicity, notations “P” and “N” mean positive and negative, respectively. Loop means the ring structure around the bond (see Section \[sec:eigenvalue\] for detail). A-type and V-type means the isosurface is [*along with*]{} or [*vertical to*]{} atomic bonds or ring planes. Four types of critical points (CPs) are nucleic critical point (NCP), bond critical point (BCP), ring critical point (RCP) and cage critical point (CCP). Here $\gamma_1$, $\gamma_2$ and $\gamma_3$ are three Hessian matrix eigenvalues. Gaussian and mean curvatures are represented by $K$ and $H$, respectively. The two principal curvatures are maximal curvature $\kappa_1$ and minimal curvature $\kappa_2$.
\[tb:geometric\_fingerprint\]
To have a more general understanding of the properties of eigenvalue and curvature maps, all of the above-mentioned results are summarized in Table \[tb:geometric\_fingerprint\]. To show the consistency of results, one can take some simply tests. For instance, one can compare the results of $\kappa_1$ and $\kappa_2$ with $K$. Since $K=\kappa_1 \kappa_2$, their signs of isosurface at four types of CPs should match with each other. For NCP, $\kappa_1$ and $\kappa_2$ are both positive, and multiplying together yields a positive $K$, exactly as it is found in the table. For all other three types of CPs, they all match very well.
Eigenvector field analysis {#sec:Tensor field}
--------------------------
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![The 4th to 11th eigenmodes of protein 2XHF. The eigenmodes are evaluated based on the rigidity density of 2XHF. A threshold value as 60% of the Maximum density is chosen. []{data-label="fig:Density_2XHF"}](2XHF_mode.png "fig:"){width="80.00000%"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Tensor fields widely exist in natural world. For biomolecules, researchers are particular interested in their motions. As stated in the previous sections, various methods, including molecular dynamics, anisotropic network model (ANM), anisotropic FRI (aFRI), etc, have been developed to explore the dynamics of the biomolecular systems. However, all the above mentioned approaches rely on the discrete representation and can not be used in the study of motions of continuous biomolecular density function or data, particularly electron density data, such as cryo-electron microscopy (cryo-EM) maps. To analyze the motion of density profiles, a virtual particle model (VPM) is proposed to systematically explore the anisotropic motions of continuous density.
### Virtual particle model
Both ANM and aFRI depend on their graph or discrete representation of biomolecules. It is not obvious how to construct a continuum model for a continuous density function to characterize its anisotropic motions. Previously, vector quantization (VQ) algorithm [@Gray:1984vector] is employed to decompose the electron density map of a biological molecule into a set of finite Voronoi cells. It is then combined with ANM to explore the dynamic of the cryo-EM data [@Tama:2002exploring; @Ming:2002describe].
In this section, virtual particles are introduced and defined for each small volume or element of a density data. To be more specific, the domain of the density function can be discretized into many elements. In general, the discretization can be non-uniformed and the elements may have different sizes. One can associate each element with a virtual particle, which is centered at the element center but having a continuous density profile. One can assume that all virtual particles are correlated with each other, but the correlations between them decrease with the distance or follow prescribed relations. The anisotropic motions of virtual particles are obtained. Similar to ANM and aFRI approaches, such anisotropic motions are evaluated from the eigenmodes of the anisotropic Kirchhoff matrix.
One can assume that the density function of interest is given by $\mu({\bf r})$, which can be either a cryo-EM density map or a rigidity density computed from atomic coordinates by using the discrete to continuum mapping as shown in Section \[Sec:GeometricRep\]. One can consider two virtual particles centered at ${\bf r}_I$ and ${\bf r}_J$ and enclosed by the volume elements of $\Omega_I$ and $\Omega_J$, respectively. A special scaling parameter $\gamma({\bf r}_I,{\bf r}_J,\Omega_1,\Omega_2,\eta_{IJ})$ is proposed as following: $$\begin{aligned}
\label{rigidity_potential8}
\gamma({\bf r}_I,{\bf r}_J,\Omega_I,\Omega_J,\eta_{IJ})=\left(1+a\int_{\Omega_I}\mu({\bf r}) d{\bf r}\right)\left(1+a\int_{\Omega_J}\mu({\bf r}) d{\bf r}\right)
\Phi(|{\bf r}_I-{\bf r}_J|, \eta_{IJ}),\end{aligned}$$ where the coefficient $a$ is a normalization factor, $\Phi(|{\bf r}_I-{\bf r}_J|,\eta_{IJ})$ is a FRI correlation function and $ \eta_{IJ}$ is the characteristic length of elements. In contrast, the $\eta_j$ in the discrete to continuum mapping is the characteristic length of atomic distances.
Three realizations of VPM by constructing three anisotropic Kirchhoff matrices are proposed. First, one can modify ANM to construct a VPM anisotropic Kirchhoff matrix. For each matrix element $H_{IJ}$, a local $3\times3$ Hessian matrix for ANM in Eq. (\[eq:multi-kirchoff1\]) is formed as $$\begin{aligned}
\label{eq:multi-kirchoff128}
H_{IJ} = -\frac{\gamma({\bf r}_I,{\bf r}_J,\Omega_I,\Omega_J,\eta_{IJ})}{r^2_{IJ}}\left[ \begin{array}{ccc}
(x_J-x_I)(x_J-x_I) &(x_J-x_I)(y_J-y_I) &(x_J-x_I)(z_J-z_I)\\
(y_J-y_I)(x_J-x_I) &(y_J-y_I)(y_J-y_I) &(y_J-y_I)(z_J-z_I)\\
(z_J-z_I)(x_J-x_I) &(z_J-z_I)(y_J-y_I) &(z_J-z_I)(z_J-z_I)
\end{array}\right] ~ \forall ~ I \neq J.
\end{aligned}$$ The diagonal elements are constructed following Eq. (\[eq:multi-kirchoff1\_diagonal\]). The test indicates this ANM based VPM works very well. However, the demonstration of this test is skipped.
Additionally, one can modify the aFRI to generate two other realizations of VPM. It is very natural for one to derive the continuous aFRI by making use of the local anisotropic matrix $\Phi^{IJ}$ defined in Eqs. (\[eq:Anisorigidity1\]) and (\[eq:afri\_local\_Hessian\]). However, in general applications, the correlation $\Phi(\|{\bf r}_I-{\bf r}_J \|; \eta_{IJ})$ can be more specifically defined to describe the interaction between each pair of virtual particles. To construct the final matrix, one can multiply the scaling parameter to the local flexibility Hessian matrix ${\bf F}^{1}$ and ${\bf F}^{2}$, the corresponding two generalized Hessian matrices can be expressed as following: $$\begin{aligned}
\label{eq:Anisoflexibility8}
{\bf F}^{1}_{IJ} =& - \frac{1}{w_{J}} {\rm adj}(\Phi^{IJ}) \gamma({\bf r}_I,{\bf r}_J,\Omega_I,\Omega_J,\eta_{IJ}), &\quad I\neq J; \\ \label{eq:Anisoflexibilityy4}
{\bf F}^{1}_{II}=& \sum_{J=1}^N \frac{1}{w_{J}} {\rm adj}(\Phi^{IJ}) \gamma({\bf r}_I,{\bf r}_J,\Omega_I,\Omega_J,\eta_{IJ}), &\quad I=J\end{aligned}$$ and $$\begin{aligned}
\label{eq:Anisoflexibility58}
{\bf F}^{2}_{IJ} =& - \frac{1}{w_{J}} |\Phi^{IJ}|(J_{3} - \Phi^{IJ}) \gamma({\bf r}_I,{\bf r}_J,\Omega_I,\Omega_J,\eta_{IJ}), &\quad I\neq J; \\ {\bf F}^{2}_{II}=& \sum_{J=1}^N \frac{1}{w_{J}} |\Phi^{IJ}|(J_3 - \Phi^{IJ}) \gamma({\bf r}_I,{\bf r}_J,\Omega_I,\Omega_J,\eta_{IJ}), &\quad I=J,\end{aligned}$$ where ${\rm adj}(\Phi^{IJ})$ denotes the adjoint of matrix. Here $J_3$ is a $3\times3$ matrix with every element being one.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![The first three nontrivial eigenmodes of Cyo-EM data EMD 8295. A threshold value of 0.08 is used in the model to map out the biomolecule. Modes 4, 5 and 6 are demonstrated in (a), (b) and (c), respectively. []{data-label="fig:EM8295"}](EM8295.png "fig:"){width="80.00000%"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![The first three nontrivial eigenmodes of Cyo-EM data EMD 1590. A threshold value of 0.05 is used in the model. Modes 4, 5 and 6 are demonstrated in (a), (b) and (c), respectively. []{data-label="fig:EM1590"}](EM1590.png "fig:"){width="80.00000%"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
It should be noticed that clustering idea in the discrete aFRI matches with the idea of discretization of the continuous density function very well. However, only the simple global form is highlighted in Eqs. (\[eq:Anisoflexibility8\]) and (\[eq:Anisoflexibility58\]).
The proposed model is tested to evaluate the anisotropic motion of density volumetric data. One only needs to consider those density elements in the anisotropic motion analysis that their density values are larger than a threshold, which is the suggested value for biomolecular visualization. This truncation dramatically reduces the number of entries in final Kirchhoff matrix.
It has been verified that the VPM is able to recover ANM and aFRI when the elements are very small and each element contains at most one real particle (say a C$_\alpha$). The demonstration of this result is omitted.
The proposed method is illustrated with three examples, i.e., protein 2XHF, cryo-EM maps EMD8295 and EMD3266. For protein 2XHF, the original data contain discrete atomic coordinates. The discrete to continuum mapping is used to generate rigidity density $\mu({\bf r})$. In this transformation, the Lorentz kernel as in Eq. (\[eq:couple\_matrix24\]) is chosen with $\epsilon=2$ and $\eta_j=1.0$ Å. To construct VPM Hessian matrix, a threshold is chosen to be $40\%$ of the maximal density value. One can discretize the computational domain by using the element size of 3.0 Å. In the mode analysis, the second aFRI form ${\bf F}^{2}$ is used. The Lorentz kernel in Eq. (\[eq:couple\_matrix24\]) is chosen with $\epsilon=2$ and $\eta_{IJ}=12$Å. Mode 4 to mode 11 are illustrated in Fig. \[fig:Density\_2XHF\]. The results are similar to those obtained with discrete aFRI [@DDNguyen:2016b].
For cryo-EM density map EMD8295, the data have a dimension of $326.4*326.4*326.4$Å$^3$. A mesh of 40\*40\*40 is used to discretize domain. The threshold of 0.08 is used to result in a total number of 639 nonzero elements in the matrix. Again the second aFRI form ${\bf F}^{2}$ with the Lorentz kernel is used. The parameter used are $\epsilon=2$ and $\eta_{IJ}=40$ Å. Modes 4, 5 and 6 are illustrated in Fig. \[fig:EM8295\].
For EMD1590, the region is of the size $436*436*436$Å$^3$. A mesh of 25\*25\*25 is employed for the discretization. The biomolecular domain is chosen by using the threshold of 0.05 and there are 394 nonzero elements in the final matrix. The second aFRI form ${\bf F}^{2}$ with the Lorentz kernel is used with $\epsilon=2$ and $\eta_{IJ}=60$ Å in matrix construction. Modes 4, 5 and 6 are illustrated in Fig. \[fig:EM1590\].
It can be noticed that VPM can be applied to other systems, such as stability analysis of cells, tissues, and some elastic systems with appropriate definitions of correlations functions. However, this aspect is beyond the scope of the present review.
### Eigenvector analysis
Mathematically, vector field can be analyze by Poincaré index [@Poincare:1890], winding number, Morse index [@harker:mischaikow:mrozek:nanda; @Mischaikow:2013] and more interestingly, the Conley index [@Conley:1978; @Mischaikow:2002; @Chen:2012morse; @Manolescu:2013conley]. The essential idea of all these methods is to explore the behavior of the vectors around critical points.
To give a brief introduction of Conley index, one can consider a manifold $M$ and its associated vector field. If one expresses the vector field in terms of a differential equation $\dot{x}=V(x)$. The solution can be expressed as a function $\phi: {\bf R}\times M \rightarrow M$. This solution is a flow that satisfies $\phi(0,x)=x$ for all $x \in M$. One can also define the trajectory as $\phi ({\bf R},x):=\bigcup_{t \in {\bf R}} \phi(t,x)$. With this setting, an invariant set $S \subset M$ is $\phi (R,S)=S$. Two basic types of invariant sets are fixed points and periodic points, see Table \[tb:conley\]. One can define an isolating neighborhood $N \subset M$ as for every $x \in \partial N$, there exists $\epsilon > 0$, such that one has either $\phi((-\epsilon,0),x)\bigcap N =\emptyset$ or $\phi((0,\epsilon),x)\bigcap N =\emptyset$. The exit set of an isolating block $N$ is $L=\{{x \in \partial N |\phi((0, \epsilon),x)\bigcap N =\emptyset}\}$. The pair $(N,L)$ is called an index pair. The Conley index of an invariant set $S$ is the relative homology of the index pair $(N,L)$, i.e., $CH_*(S):=H_*(N,L)$.
Cases Conely index
--------------------------- --------------
Attracting fixed point (1 0 0)
Saddle fixed point (0 1 0)
Repelling fixed point (0 0 1)
attracting periodic orbit (1 1 0)
Repelling periodic orbit (0 1 1)
: Conley index for various types of invariant sets
\[tb:conley\]
It should be noticed that the VPM eigenvector field is different from the traditional vector field. For an individual eigenvector, one has local vectors associated all virtual particles. All these local vectors join together to form a unique global vector field, which is able to capture the collective motions of the biomolecule of interest. However, locally, various fixed points can also be identified as demonstrated in Figs. \[fig:Density\_2XHF\], \[fig:EM8295\] and \[fig:EM1590\]. Rigorous analysis of these eigenvectors is still an open problem.
Demonstrations {#sec:theory}
--------------
In this section, the utility of Geo-Topo algorithms, namely, topological analysis of Hessian matrix eigenvalue and curvature maps is illustrated by a few case studies. Additionally, the persistent homology analysis of molecular Hessian matrix eigenvalue maps is also demonstrated.
### Case studies {#sec:density}
Three molecules, a fullerene, an alpha helix and beta sheet, to illustrate Geo-Topo methods are considered.
#### Fullerene C$_{20}$
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Hessian matrix eigenvalue surfaces obtained from different isovalues (or level-set values). [ (a)]{} The isosurfaces for the first eigenvalue. The isovalues from [ ($a_{1}$)]{} to [ ($a_{4}$)]{} are -3.0, -2.0, -1.0 and 0.1. [ (b)]{} The isosurfaces for the second eigenvalue. The isovalues from [ ($b_{1}$)]{} to [ ($b_{4}$)]{} are -1.0, -0.01, 0.5 and 1.0. [ (c)]{} The isosurfaces for the third eigenvalue. The isovalues from [ ($c_{1}$)]{} to [ ($c_{4}$)]{} are -1.0, 1.0, 1.8 and 2.0. []{data-label="fig:C20_eigen_3d"}](C20_eigen_3d.png "fig:"){width="60.00000%"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Using the proposed Geo-Topo fingerprint, one can study more complicated structures. The first one is the fullerene C$_{20}$ considered in last section. Again, it rigidity density is described by Eq. (\[eq:rigidity3\]) and parameter $w_j$ and $\eta_j$ are set to $1$ and $0.7$, respectively.
First the Hessian matrix eigenvalue isosurfaces of C$_{20}$ is studied. Figure \[fig:C20\_eigen\_3d\] has illustrated four representative isosurfaces for each eigenvalue. Subscripts $1$ to $4$ indicate four isovalues from small to large. The notations $({\bf a})$ to $({\bf c})$ represent eigenvalues $\gamma_1$ to $\gamma_3$. One can see that their isosurface behaviors are consistent with descriptions of the Geo-Topo fingerprints summarized in Table \[tb:geometric\_fingerprint\]. State differently, the Geo-Topo fingerprints are able to capture the essential Geo-Topo properties of C$_{20}$ isosurfaces.
More specifically, for $\gamma_1$, negative isosurfaces enclose all NCPs and BCPs. Anisotropic isosurfaces are found near RCPs with negative A-type and positive V-type behaviors. CCPs are enclosed with positive isosurfaces. For $\gamma_2$, negative isosurfaces enclose all NCPs and bond regions of BCPs. Positive loops can be found around BCP bonds. RCPs and CCPs are enclosed by positive isosurfaces. For $\gamma_3$, negative isosurfaces [*only*]{} enclose atomic basin of NCPs, whereas BCPs, RCPs and CCPs are all enclosed by positive isosurfaces.
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Isosurfaces for Gaussian and mean curvature maps of C$_{20}$. [ (a)]{} The isosurfaces built from the Gaussian curvature of C$_{20}$. The isovalues from [ ($a_{1}$)]{} to [ ($a_{4}$)]{} are -2.0, 2.0, 3.0 and 5.0. [ (b)]{} The isosurfaces built from the Mean curvature of C$_{20}$. The isovalues from [ ($b_{1}$)]{} to [ ($b_{4}$)]{} are -1.0, 0.001, 1.0 and 2.0. []{data-label="fig:C20_gaussian_mean_3d"}](C20_k1k2_3d.png "fig:"){width="60.00000%"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Isosurfaces for Gaussian and mean curvature maps of C$_{20}$. [ (a)]{} The isosurfaces built from the Gaussian curvature of C$_{20}$. The isovalues from [ ($a_{1}$)]{} to [ ($a_{4}$)]{} are -2.0, 2.0, 3.0 and 5.0. [ (b)]{} The isosurfaces built from the Mean curvature of C$_{20}$. The isovalues from [ ($b_{1}$)]{} to [ ($b_{4}$)]{} are -1.0, 0.001, 1.0 and 2.0. []{data-label="fig:C20_gaussian_mean_3d"}](C20_gaussian_mean_3d.png "fig:"){width="60.00000%"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Figures \[fig:C20\_k1k2\_3d\] and \[fig:C20\_gaussian\_mean\_3d\] illustrates four representative isosurfaces for Gaussian, mean and two principal curvatures. Here the detailed analysis is omitted. However, just as eigenvalue isosurfaces, the curvature isosurfaces can be well-described by the Geo-Topo fingerprints summarized in Table \[tb:geometric\_fingerprint\].
#### An $\alpha$-helix structure
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![An illustration of second eigenvalue, third eigenvalue and mean curvature of a coarse-grained representation (C$_{\alpha}$) of an alpha helix. [ (a)]{} The isosurfaces for the second eigenvalue. The isovalues from [ ($a_{1}$)]{} to [ ($a_{4}$)]{} are -2.0, -0.05, 0.05 and 0.1. [ (b)]{} The isosurfaces for the third eigenvalue. The isovalues from [ ($b_{1}$)]{} to [ ($b_{4}$)]{} are -0.05, 0.2, 0.3 and 0.35. [ (c)]{} The isosurfaces for the mean curvature. The isovalues from [ ($c_{1}$)]{} to [ ($c_{4}$)]{} are -1.0, -0.1, 0.1 and 3.0. []{data-label="fig:1c26_helix"}](1c26_helix.png "fig:"){width="60.00000%"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
So far, all cases examined are about highly symmetric molecular structures. In this part, the Geo-Topo analysis is applied to irregular biomolecular structures, particularly protein structure. It is known that there are four distinct levels of protein structure, namely primary structure, which is a sequence of amino acids in the polypeptide chain; secondary structure, which is an $\alpha$-helix or a $\beta$-strand; tertiary structure, which refers to the three-dimensional structure of a monomeric and multimeric protein molecule; and quaternary structure, which is the three-dimensional structure of a multi-subunit protein complex. The Geo-Topo analysis of proteins for protein secondary structures is demonstrated.
First, one can consider one of protein secondary structures, i.e., an $\alpha$-helix segment. This segment is extracted from protein with ID 1C26. The coarse-grained (CG) representation is employed and 19 C$_{\alpha}$ atoms from the $335$th residue to $353$th residue in chain A are used. The distance between two adjacent C$_{\alpha}$ atoms are about $3.8$ Å and the $\eta$ used in the CG rigidity density model is $2.0$ Å.
In stead of listing all results of eigenvalues and curvatures, only three geometric parameters of interest are examined, including eigenvalue $\gamma_2$, eigenvalue $\gamma_3$ and mean curvature. For each of them, four representative isosurfaces are extracted. The results are illustrated in Fig. \[fig:1c26\_helix\].
It can be found that results are very consistent with the Geo-Topo fingerprints. For $\gamma_2$, positive loops around atomic bonds are found. For $\gamma_3$, large eigenvalues are still concentrated in regions near BCPs. For mean curvature, one can still find positive A-type and negative V-type isosurfaces near BCPs. Also, large positive values are concentrates around NCPs and indicate the topological connectivity at Figs. \[fig:1c26\_helix\](a$_1$), (c$_3$) and (c$_4$).
It also should be noticed that unlike the regular symmetric structures studied in previous sections, RCPs and CCPs are more complicated in $\alpha$-helix structure. Moreover, since the characteristic distance $\eta$ is chosen as $2.0$ Å, bond effect (or topological connectivity) is observed not only between adjacent two atoms but also between atoms in close distance, as indicated in Fig. \[fig:1c26\_helix\]${ (a_3)}$. At meantime, strongest topological connectivity (largest $\gamma_2$ isovalues) is as usual found between adjacent two atoms as indicated in Fig. \[fig:1c26\_helix\]${(a_4)}$. Interestingly, in the case of $\gamma_3$, positive isosurface forms a strip that is parallel to the backbone of the $\alpha$-helix as demonstrated in Fig. \[fig:1c26\_helix\]${ (b_3)}$. The largest isovalues are concentrated around BCPs not between adjacent two atoms, but neighboring two atoms in adjacent two circles as illustrated in Fig. \[fig:1c26\_helix\]${ (b_4)}$. However, if characteristic distance $\eta$ was chosen as $1.5$ Å, the largest isovalues of $\gamma_3$ would move to BCP regions between adjacent two atoms. This is due to the multiscale nature of the model. Further detailed discussion of this multiscale property is beyond the scope of this paper.
#### A $\beta$-sheet structure
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![An illustration of second eigenvalue, third eigenvalue and mean curvature of a coarse-grained representation (C$_{\alpha}$) of a $\beta$-sheet. [ (a)]{} The isosurfaces for the second eigenvalue. The isovalues from [ (a$_{1}$)]{} to [ (a$_{4}$)]{} are -0.3, -0.05, 0.05 and 0.1. [ (b)]{} The isosurfaces for the third eigenvalue. The isovalues from [ (b$_{1}$)]{} to [ (b$_{4}$)]{} are -0.1, 0.2, 0.3 and 0.35. [ (c)]{} The isosurfaces for the mean curvature. The isovalues from [ (c$_{1}$)]{} to [ (c$_{4}$)]{} are -1.0, -0.1, 0.5 and 3.0. []{data-label="fig:4uw4_sheet"}](4uw4_sheet.png "fig:"){width="60.00000%"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Another important protein secondary structure is $\beta$-sheet. In this part, three adjacent $\beta$-strands from protein with ID 4UW4 are considered. Again, the CG representation is used and three stands include residues from $575$ to $586$, $589$ to $600$ and $603$ to $615$ in chain A. Just as the analysis in the $\alpha$-helix structure, the characteristic distance $\eta$ is chosen as $2.0$ Å. One can examine the second eigenvalue $\gamma_2$, the third eigenvalue $\gamma_3$ and mean curvature in the Geo-Topo analysis. For each of them, four representative isosurfaces are extracted. The results are illustrated in Fig. \[fig:4uw4\_sheet\].
Just as the $\alpha$-helix case, results for $\beta$-sheet are also very consistent with the Geo-Topo fingerprints. Positive loops around BCPs for $\gamma_2$ can be observed. For $\gamma_3$, small negative isosurfaces indicate NCPs. Large $\gamma_3$ eigenvalues are concentrated in regions near BCPs. Positive A-type and negative V-type isosurfaces near BCPs are found in mean curvature. Large positive values are concentrates around NCPs. Further, as the characteristic distance is chosen as $2.0 $ Å, the bond connection between the stands or sheets are amplified. It can be clearly observed in Fig. \[fig:4uw4\_sheet\] ${ (b_4)}$.
### Persistent homology for scalar field analysis {#sec:PHA_scalar_field}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Barcodes for three eigenvalue maps of benzene molecule. From [ (a)]{} to [ (c)]{}, the barcodes are for $\lambda_1$, $\lambda_2$ and $\lambda_3$, respectively. In each subfigure, from top to bottom, the results are for $\beta_0$, $\beta_1$ and $\beta_2$, respectively. []{data-label="fig:C6_eigen_ph"}](C6_eigen_ph.png "fig:"){width="90.00000%"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Finally, the topological persistence in the scalar fields generated by Hessian matrix eigenvalues and curvatures is studied. The hexagonal ring is used for persistent homology analysis. The barcodes for eigenvalues $\lambda_1$, $\lambda_2$ and $\lambda_3$ are demonstrated in Figs. \[fig:C6\_eigen\_ph\] (a)-(c). In each subfigure, results for $\beta_0$, $\beta_1$ and $\beta_2$ are presented. The filtration goes from the smallest value to the largest in the persistent homology analysis.
For $\lambda_1$, at very small values, it has $6$ $\beta_0$ bars, i.e., $6$ independent components. Topologically, it represents small $\lambda_1$ negative values concentrating around $6$ NCPs. As the filtration progresses, a loop is formed, which leads to a bar in $\beta_1$. Further down the filtration process, $\beta_0$ isosurface begins to shrink to two balls perpendicular to the RCP. This contributes two bars in $\beta_2$.
For $\lambda_2$, again $6$ $\beta_0$ bars are found in the earliest stage of the filtration. Here $6$ independent components quickly combine to form a loop as filtration progresses. More interesting, individual loops around BCPs form when filtration value is around the range from 0 to 1. Finally, isosurfaces shrink into the RCP.
For $\lambda_3$, smallest negative values are concentrated around NCPs and contribute $6$ $\beta_0$ bars at the earliest stage of the filtration. Once the isosurface value becomes positive, a new component emerges due to the generation of a new isosurface at the boundary region. This new isosurface also contributes to a long persisting $\beta_2$ bar. Further down the filtration, $6$ more $\beta_0$ bars occur, each representing a very narrow ring region around atom bonds. These regions are relatively small and quickly disappear. After that, $6$ “hat" regions attaching to the original NCP isosurfaces emerge, contributing to $6$ small loops. They quickly become detached and shrink away. At same time, $6$ isosurfaces form near the BCP and gradually disappear. Together they contribute 12 independent $\beta_2$ bars.
It can be seen that barcodes in Figs. \[fig:C6\_eigen\_ph\] (a)-(c) give a detailed account of the full spectrum of the isosurface evolution process in Figs. \[fig:C6\_eigen\_3d\] (a$_1)$-(a$_4)$.
Concluding remarks
==================
Every field in natural science, engineering, medicine, finance and social sciences becomes quantitative when it is getting mature. Mathematics is essential for all quantitative fields. Being regarded the last scientific forefront, biological sciences, particularly molecular biology and structural biology, have accumulated gigantic among of data in terms of biomolecular structures, activity relations and genetic sequences in the past few decades and are transforming from qualitative and phenomenological to quantitative and predictive. Such a transformation offers unprecedented opportunities for mathematically driven advances in biological sciences [@Wei:2016].
Geometry, topology, and graph theory are some of the core mathematics and have been naturally playing a unique role in molecular biology and molecular biophysics. In this paper, we present a brief review of geometric, topological, and graph theory apparatuses that are important to the contemporary molecular biology and biophysics. We first discuss the discrete methods and models, including graph theory, Gaussian network models, anisotropic network model, normal mode analysis, flexibility-rigidity index, molecular nonlinear dynamics, spectral graph theory and persistent homology for biomolecular modeling and analysis. Additionally, we describe continuous algorithms and theories, including, discrete to continuum mapping, multidimensional persistent homology for volumetric data sets, geometric modeling of biomolecules, differential geometry theory of surfaces, curvature analysis, atoms in molecule theory, and quantum chemical topology theory. Attention is paid to the connections between existing biophysical approaches and standard mathematical subjects, such as, Morse theory, Poincaré Hopf index, differential topology, etc. Open problems and potential new directions are point out in discussions.
Two new models, namely the analytical minimal molecular surface and virtual particle model, and two new methods, i.e., Hessian eigenvalue map and curvature map, are introduced for biomolecular modeling and analysis. These new approaches were inspirited by the subject under review during our preparation of this review. For simplicity, only the proof-of-principle applications are demonstrated for all methods, models, theories and algorithms covered in this review.
Selected mathematical topics in geometry, topology, and graph theory are based on our limited knowledge and understanding of mathematics and molecular biophysics. Many subjects in geometry, topology, and graph theory that have found much success in molecular biology and biophysics have not been covered in this review. One of these subjects is knot theory, particularly the DNA knot theory, which is also a very important ingredient of topological modeling of biomolecules [@vologodskii:1992; @pohl:1980; @fuller:1971; @ABates:2005; @sumners:1992; @darcy:2001; @arsuaga:2002; @buck:2007; @IKDarcy:2013; @RBrasher:2013; @Schlick:1992trefoil]. Knot theory is an area of geometric topology that deals with knots and links. Mathematically, a knot is an embedding of circles or its homeomorphisms in the three-dimensional (3D) Euclidean space, ${\mathbb R}^3$. Physically, DNA, as a genetic material, exists usually in two very long strands that intertwine to form chromatins, bind with histones to build nucleosomes, tie into knots, and are subjected to successive coiling before package into chromosomes. The loss of knots in chromosomes can cause Angelman and Prader-Willi syndromes. DNA knot theory has been a very important topic in applied topology. However, in this review, focus is given to geometric and topological methods and models for atomistic biomolecular data. Therefore, DNA knot theory is not covered.
Another relevant subject that have not been considered in this review is biomolecular interaction network models, including protein interaction networks, metabolic networks and transcriptional regulatory networks. Obviously, part of the mathematical foundation of these models is graph theory. Protein interaction networks are designed to study protein-protein interactions [@rain:2001; @giot:2003]. Normally, in these models, proteins are represented as nodes and physical interaction between them are represented by edges. To make the network more reliable, data from different sources are combined together and different rules are applied for the identification of protein interactions. A metabolic network considers all metabolic and physical processes happened within a cell [@jeong:2000; @overbeek:2000]. This network comprises the chemical reactions of metabolism, the metabolic pathways, as well as the regulatory interactions that guide these reactions. Transcriptional regulatory networks describe the regulatory interactions between genes [@lee:2002; @salgado:2006]. In this network, each gene is represented by a node and the regulation relations are represented by edges. The exclusion of this subject is also due to our focus on atomistic biomolecular data.
Topological graph theory concerns immersions of graphs as well as the embedding of graphs in surfaces, spatial embeddings of graphs, and graphs as topological spaces. It has had much success in the mathematical modeling of DNA recombination, DNA-RNA interactions, protein folding and protein-protein interactions [@Angeleska:2009]. The exclusion of this subject is due to our insufficient knowledge and understanding. For the same reason, fascinating applications of combinatorics, algebra and tiling theory in the modeling of viral capsid self assembly [@Jonoska:2009; @Twarock:2008; @Angeleska:2009; @SHarvey:2013; @CHeitsch:2014; @Sadre-Marandi:2014] have not been covered in our review.
A continuous, differentiable curve can be embedded in a three-dimensional Euclidean space and its kinematic properties, such as the derivatives of tangent, normal, and binormal unit vectors, can be described by Frenet-Serret formulas in differential geometry [@crenshaw1993orientation]. Discrete Frenet-Serret frame offers an efficient description of amino-acid and/or nucleic acid chains [@quine2004mathematical; @hu2011discrete]. We believe that discrete Frenet-Serret frame can be easily applied to RNA chains, microtubules, nucleosomes, chromatins, active chromosomes and metaphase chromosomes.
An emergent approach is to combine machine learning with geometry, topology and/or graph theory for analyzing biomolecular data [@ZXCang:2015; @Kovacev-Nikolic:2016]. Machine learning is a cutting edge computer science and statistical tool originally developed for pattern recognition and artificial intelligence. Its combination with mathematical apparatuses leads to extremely powerful approaches to massive biomolecular data challenges, such as the blind predictions of solvation free energies [@BaoWang:2016FFTS; @BaoWang:2016HPK] and protein-ligand binding affinity prediction [@BaoWang:2016FFTB]. For example, topological learning algorithm that utilizes exclusively persistent homology and machine learning for protein-ligand binding affinity predictions outperforms all the existing eminent methods in computational biophysics over massive binding data sets [@ZXCang:2016b]. However, this subject is at its early stage and is evolving too fast to have a conclusive review at present.
It is worth mentioning that the geometric, topological and graph theory apparatuses discussed in this review can be employed in conjugation with partial differential equation (PDE), which is widely used in computational biophysics, to model biomolecular systems. Certainly geometric modeling is often a prerequisite in the PDE models of electrostatics, solvation, ion channels, membrane-protein interactions, and biomolecular elasticity [@Wei:2009]. As mentioned in Section \[sec:spectral\], the graph cut problem can be formulated as a free energy minimization. Such a formulation makes it possible to combine spectral graph theory with PDE approaches for a wide range of biophysical modeling for biomolecular systems, including solvation, ion channel, biomembrane, protein-ligand binding, protein-protein interaction and protein-nucleic acid interaction. Finally, connection between algebraic topology and differential geometry, including Laplace-Beltrami operator, has been made [@BaoWang:2016a]. Essentially, one defines an object function to optimize certain biophysical properties, which leads to a Laplace-Beltrami operator that generates a multiscale representation of the initial data and offers an object-oriented filtration process for persistent homology. The resulting differential geometry based object-oriented persistent homology is able to preserve desirable geometric features in the evolutionary filtration and enhances the corresponding topological persistence. However, how to design object-oriented persistent homology to automatically extract desirable features in the original biomolecular data during the filtration process is still an open problem.
Indeed, the application aspects of geometry, topology and graph theory have become a driven force for the development of abstract geometry, topology, homology, graph theory in recent years. It is expected that a versatile variety of pure mathematics concepts, methods and techniques will find their cutting edges in the transcend description of biomolecular structure, function, dynamics and transport. This health interaction between mathematics and molecular bioscience will benefit both fields and attract young researchers for generations to come.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported in part by a Start-Up Grant from The Nanyang Technological University (KLX), NSF IIS- 1302285 (GWW), NSF DMS-1160352 (GWW), NIH R01GM-090208 (GWW) and MSU Center for Mathematical Molecular Biosciences Initiative (GWW).
[100]{}
P. K. Agarwal, H. Edelsbrunner, J. Harer, and Y. Wang. Extreme elevation on a 2-manifold. , 36(4):553–572, 2006.
D. Aldous and J. Fill. Reversible [Markov]{} chains and random walks on graphs, 2002.
N. Alexandrov and I. Shindyalov. : protein domain parser. , 19(3):429–430, 2003.
M. P. Allen and D. J. Tildesley. . Oxford: Clarendon Press, 1987.
A. Amadei, A. Linssen, and H. JC Berendsen. Essential dynamics of proteins. , 17(4):412–425, 1993.
C. B. Anfinsen. Einfluss der configuration auf die wirkung den. , 181:223 – 230, 1973.
A. Angeleska, N. Jonoska, and M. Saito. Dna rearrangement through assembly graphs. , 157:3020–3037, 2009.
J. Arsuaga, M. V[á]{}zquez, S. Trigueros, D. W. Sumners, and J. Roca. Knotting probability of [DNA]{} molecules confined in restricted volumes: [DNA]{} knotting in phage capsids. , 99(8):5373–5377, 2002.
A. R. Atilgan, S. R. Durrell, R. L. Jernigan, M. C. Demirel, O. Keskin, and I. Bahar. Anisotropy of fluctuation dynamics of proteins with an elastic network model. , 80:505 – 515, 2001.
A. Azran and Z. Ghahramani. A new approach to data driven clustering. In [*Proceedings of the 23rd international conference on Machine learning*]{}, pages 57–64. ACM, 2006.
R. F. Bader. Atoms in molecules. , 18(1):9–15, 1985.
R. F. Bader. . Wiley Online Library, 1990.
R. F. Bader, R. J. Gillespie, and P. J. MacDougall. A physical basis for the [VSEPR]{} model of molecular geometry. , 110(22):7329–7336, 1988.
I. Bahar, A. R. Atilgan, M. C. Demirel, and B. Erman. Vibrational dynamics of proteins: Significance of slow and fast modes in relation to function and stability. , 80:2733 – 2736, 1998.
I. Bahar, A. R. Atilgan, and B. Erman. Direct evaluation of thermal fluctuations in proteins using a single-parameter harmonic potential. , 2:173 – 181, 1997.
N. A. Baker. Improving implicit solvent simulations: a [Poisson]{}-centric view. , 15(2):137–43, 2005.
N. A. Baker, D. Sept, S. Joseph, M. J. Holst, and J. A. McCammon. Electrostatics of nanosystems: Application to microtubules and the ribosome. , 98(18):10037–10041, 2001.
A. D. Bates and A. Maxwell. . Oxford University Press, USA, 2005.
P. W. Bates, Z. Chen, Y. H. Sun, G. W. Wei, and S. Zhao. Geometric and potential driving formation and evolution of biomolecular surfaces. , 59:193–231, 2009.
P. W. Bates, G. W. Wei, and S. Zhao. The minimal molecular surface. , \[q-bio.BM\], 2006.
P. W. Bates, G. W. Wei, and Shan Zhao. Minimal molecular surfaces and their applications. , 29(3):380–91, 2008.
U. Bauer, M. Kerber, and J. Reininghaus. Distributed computation of persistent homology. , 2014.
K. Beketayev, G. H. Weber, M. Haranczyk, P.T. Bremer, M. Hlawitschka, and B. Hamann. Topology-based visualization of transformation pathways in complex chemical systems. In [*Computer Graphics Forum*]{}, volume 30, pages 663–672. Wiley Online Library, 2011.
M. Belkin. . PhD thesis, The University of Chicago, 2003.
M. Belkin and P. Niyogi. Towards a theoretical foundation for [Laplacian]{}-based manifold methods. In [*International Conference on Computational Learning Theory*]{}, pages 486–500. Springer, 2005.
Paul Bendich, Herbert Edelsbrunner, and Michael Kerber. Computing robustness and persistence for images. , 16:1251–1260, 2010.
Paul Bendich and John Harer. Persistent intersection homology. , 11(3):305–336, 2011.
Janine Bennett, Fabien Vivodtzev, and Valerio Pascucci, editors. . Mathematics and Visualization. Springer-Verlag Berlin Heidelberg, 2015.
C.A.S. Bergstrom, M. Strafford, L. Lazorova, A. Avdeef, K. Luthman, and P. Artursson. Absorption classification of oral drugs based on molecular surface properties. , 46:558–570, 2003.
S. Biasotti, L. De Floriani, B. Falcidieno, P. Frosini, D. Giorgi, C. Landi, L. Papaleo, and M. Spagnuolo. Describing shapes by geometrical-topological properties of real functions. , 40(4):12, 2008.
F. Biegler-K[ö]{}nig and J. Sch[ö]{}nbohm. Update of the [AIM]{}2000-program for atoms in molecules. , 23(15):1489–1494, 2002.
J. Blinn. A generalization of algebraic surface drawing. , 1(3):235–256, 1982.
R. Brasher, R. G. Scharein, and M. Vazquez. New biologically motivated knot table. , 41:606–611, 2013.
P. T. Bremer, V. Pascucci I. Hotz, and R. Peikert, editors. . Mathematics and Visualization. Springer International Publishing, 2014.
B. Brooks and M. Karplus. Harmonic dynamics of proteins: normal modes and fluctuations in bovine pancreatic trypsin inhibitor. , 80(21):6571–6575, 1983.
B. R. Brooks, R. E. Bruccoleri, B. D. Olafson, D.J. States, S. Swaminathan, and M. Karplus. Charmm: A program for macromolecular energy, minimization, and dynamics calculations. , 4:187–217, 1983.
B. R. Brooks, D. Jane[ž]{}i[č]{}, and M. Karplus. Harmonic analysis of large systems. [I. Methodology]{}. , 16(12):1522–1542, 1995.
Peter Bubenik. Statistical topological data analysis using persistence landscapes. , 16(1):77–102, 2015.
Peter Bubenik and Peter T. Kim. A statistical approach to persistent homology. , 19:337–362, 2007.
D. Buck and E. Flapan. Predicting knot or catenane type of site-specific recombination products. , 374(5):1186–1199, 2007.
Z. X. Cang and G. W. Wei. . , 2016.
Z. X. Cang and G. W. Wei. . , 2016.
Zixuan Cang, Lin Mu, Kedi Wu, Kris Opron, Kelin Xia, and Guo-Wei Wei. A topological approach to protein classificationy. , 3:140–162, 2015.
G. Carlsson. Topology and data. , 46(2):255–308, 2009.
G. Carlsson, T. Ishkhanov, V. Silva, and A. Zomorodian. On the local behavior of spaces of natural images. , 76(1):1–12, 2008.
G. Carlsson and A. Zomorodian. The theory of multidimensional persistence. , 42(1):71–93, 2009.
G. Carlsson, A. Zomorodian, A. Collins, and L. J. Guibas. Persistence barcodes for shapes. , 11(2):149–187, 2005.
Gunnar Carlsson and Vin De Silva. Zigzag persistence. , 10(4):367–405, 2010.
Gunnar Carlsson, Vin de Silva, and Dmitriy Morozov. Zigzag persistent homology and real-valued functions. In [*Proc. 25th Annu. ACM Sympos. Comput. Geom.*]{}, pages 247–256, 2009.
S. L. Chan and E. O. Purisima. Molecular surface generation using marching tetrahedra. , 11:1268–1277, 1998.
H. W. Chang, S. Bacallado, V. S. Pande, and G. E. Carlsson. Persistent topology and metastable state in conformational dynamics. , 8(4):e58699, 2013.
Fr[é]{}d[é]{}ric Chazal, David Cohen-Steiner, Marc Glisse, Leonidas J. Guibas, and Steve Oudot. Proximity of persistence modules and their diagrams. In [*Proc. 25th ACM Sympos. on Comput. Geom.*]{}, pages 237–246, 2009.
Fr[é]{}d[é]{}ric Chazal, Leonidas J. Guibas, Steve Y. Oudot, and Primoz Skraba. Persistence-based clustering in riemannian manifolds. In [*Proceedings of the 27th annual ACM symposium on Computational geometry*]{}, SoCG ’11, pages 97–106, 2011.
Duan Chen, Zhan Chen, Changjun Chen, W. H. Geng, and G. W. Wei. : A software package for electrostatic analysis. , 32:657 – 670, 2011.
Duan Chen, Zhan Chen, and G. W. Wei. Quantum dynamics in continuum for proton transport [II: Variational]{} solvent-solute interface. , 28:25 – 51, 2012.
Duan Chen and G. W. Wei. Quantum dynamics in continuum for proton transport—[Generalized]{} correlation. , 136:134109, 2012.
Duan Chen and G. W. Wei. Quantum dynamics in continuum for proton transport [I: Basic]{} formulation. , 13:285–324, 2013.
Guoning Chen, Qingqing Deng, Andrzej Szymczak, Robert S Laramee, and Eugene Zhang. Morse set classification and hierarchical refinement using conley index. , 18(5):767–782, 2012.
Minxin Chen and Benzhuo Lu. Tmsmesh: A robust method for molecular surface mesh generation using a trace technique. , 7:203–212, 2011.
Minxin Chen, Bin Tu, and Benzhuo Lu. Triangulated manifold meshing method preserving molecular surface topology. , 38:411–418, 2012.
Wenyu Chen, Jianmin Zheng, and Yiyu Cai. Kernel modeling for molecular surfaces using a uniform solution. , 42:267–278, 2010.
Z. Chen, N. A. Baker, and G. W. Wei. Differential geometry based solvation models [I]{}: Eulerian formulation. , 229:8231–8258, 2010.
Z. Chen, N. A. Baker, and G. W. Wei. Differential geometry based solvation models [II]{}: Lagrangian formulation. , 63:1139– 1200, 2011.
Z. Chen and G. W. Wei. Differential geometry based solvation models [III]{}: Quantum formulation. , 135:194108, 2011.
L. T. Cheng, Joachim Dzubiella, Andrew J. McCammon, and B. Li. Application of the level-set method to the implicit solvation of nonpolar molecules. , 127(8), 2007.
Li-Tien Cheng, Yang Xie, Joachim Dzubiella, J. Andrew McCammon, Jianwei Che, and Bo Li. . , 5:257–266, 2009.
F. Chiti and C. M. Dobson. Protein misfolding, functional amyloid, and human disease. , 75:333 – 366, 2006.
F. Chung. . American Mathematical Society, 1997.
J. Cioslowski and G. H. Liu. Topology of electron-electron interactions in atoms and molecules. [II]{}. the correlation cage. , 110(4):1882–1887, 1999.
David Cohen-Steiner, Herbert Edelsbrunner, and John Harer. Stability of persistence diagrams. , 37(1):103–120, 2007.
David Cohen-Steiner, Herbert Edelsbrunner, and John Harer. Extending persistence using poincar[é]{} and lefschetz duality. , 9(1):79–103, 2009.
David Cohen-Steiner, Herbert Edelsbrunner, John Harer, and Dmitriy Morozov. Persistent homology for kernels, images, and cokernels. In [*Proceedings of the Twentieth Annual ACM-SIAM Symposium on Discrete Algorithms*]{}, SODA 09, pages 1011–1020, 2009.
Charles Conley. . American Mathematical Society,Providence, R.I, 1978.
M. L. Connolly. Depth buffer algorithms for molecular modeling. , 3:19–24, 1985.
Julia Contreras-Garc[í]{}a, Erin R Johnson, Shahar Keinan, Robin Chaudret, Jean-Philip Piquemal, David N Beratan, and Weitao Yang. Nciplot: a program for plotting noncovalent interaction regions. , 7(3):625–632, 2011.
Julia Contreras-Garc[í]{}a, Weitao Yang, and Erin R Johnson. Analysis of hydrogen-bond interaction potentials from the electron density: integration of noncovalent interaction regions. , 115(45):12983–12990, 2011.
Hugh C Crenshaw and Leah Edelstein-Keshet. Orientation by helical motion—ii. changing the direction of the axis of motion. , 55(1):213–230, 1993.
P.B. Crowley and A. Golovin. Cation[-]{}pi interactions in protein-protein interfaces. , 59:231–239, 2005.
Q. Cui. Combining implicit solvation models with hybrid quantum mechanical/molecular mechanical methods: A critical test with glycine. , 117(10):4720, 2002.
Q. Cui and I. Bahar. . Chapman and Hall/CRC, 2010.
Q. Cui, G. J. Li, J. Ma, and M. Karplus. A normal mode analysis of structural plasticity in the biomolecular motor f(1)-atpase. , 340(2):345 – 372, 2004.
Y. Dabaghian, F. Memoli, L. Frank, and G. Carlsson. A topological paradigm for hippocampal spatial map formation using persistent homology. , 8(8):e1002581, 08 2012.
I. K. Darcy. Biological distances on [DNA]{} knots and links: applications to [XER]{} recombination. , 10(02):269–294, 2001.
I. K. Darcy and M. Vazquez. Determining the topology of stable [ protein-DNA]{} complexes. , 41:601–605, 2013.
S. J. Darnell, L. LeGault, and J. C. Mitchell. server: interactive forecasting of protein interaction hot spots. , 36:W265–W269, 2008.
Bhaskar DasGupta and Jie Liang. . John Wiley & Sons, 2016.
Vin de Silva, Dmitriy Morozov, and Mikael Vejdemo-Johansson. Persistent cohomology and circular coordinates. , 45:737–759, 2011.
S. Decherchi and W. Rocchia. . , 8:e59744, 2013.
O. N. A. Demerdash, M. D. Daily, and J. C. Mitchell. Structure-based predictive models for allosteric hot spots. , 5:e1000531, 2009.
Omar N. A. Demerdash and Julie C. Mitchell. . , [80]{}([7]{}):[1766–1779]{}, [JUL]{} [2012]{}.
T. K. Dey, K. Y. Li, J. Sun, and C. S. David. Computing geometry aware handle and tunnel loops in 3d models. , 27, 2008.
Tamal K Dey, Fengtao Fan, and Yusu Wang. Computing topological persistence for simplicial maps. In [*Proc. 30th Annu. Sympos. Comput. Geom. (SoCG)*]{}, pages 345–354, 2014.
Tamal K. Dey and Y. S. Wang. Reeb graphs: Approximation and persistence. , 49(1):46–73, 2013.
Barbara Di Fabio and Claudia Landi. A mayer-vietoris formula for persistent homology with an application to shape recognition in the presence of occlusions. , 11:499–527, 2011.
C. H. Q. Ding, X. F. He, H. Y. Zha, M. Gu, and H. D. Simon. A min-max cut algorithm for graph partitioning and data clustering. In [*Data Mining, 2001. ICDM 2001, Proceedings IEEE International Conference on*]{}, pages 107–114. IEEE, 2001.
A.I. Dragan, C.M. Read, E.N. Makeyeva, E.I. Milgotina, M.E.A. Churchill, C. Crane[-]{}Robinson, and P.L. Privalov. Dna binding and bending by hmg boxes: Energetic determinants of specificity. , 343:371–393, 2004.
B. S. Duncan and A. J. Olson. Shape analysis of molecular surfaces. , 33:231–238, 1993.
H. Edelsbrunner and J. Harer. Persistent homology-a survey. , 453:257–282, 2008.
H. Edelsbrunner, D. Letscher, and A. Zomorodian. Topological persistence and simplification. , 28:511–533, 2002.
H. Edelsbrunner and E. P. Mucke. Three-dimensional alpha shapes. , 13:43–72, 1994.
Herbert Edelsbrunner and John Harer. . American Mathematical Soc., 2010.
H. Federer. . , 93:418–491, 1959.
X. Feng, K. L. Xia, Y. Y. Tong, and G. W. Wei. Multiscale geometric modeling of macromolecules [II:]{} lagrangian representation. , 34:2100–2120, 2013.
Xin Feng, Kelin Xia, Yiying Tong, and Guo-Wei Wei. Geometric modeling of subcellular structures, organelles and large multiprotein complexes. , 28:1198–1223, 2012.
D. Fera, N. Kim, N. Shiffeldrim, J. Zorn, U. Laserson, H. H. Gan, and T. Schlick. : [RNA]{}-as-graphs web resource. , 5(1):1, 2004.
P. J. Flory. Statistical thermodynamics of random networks. , 351:351 – 378, 1976.
S. Fortunato. Community detection in graphs. , 486(3):75–174, 2010.
Patrizio Frosini. A distance for similarity classes of submanifolds of a [Euclidean]{} space. , 42(3):407–416, 1990.
Patrizio Frosini and Claudia Landi. Size theory as a topological tool for computer vision. , 9(4):596–603, 1999.
Patrizio Frosini and Claudia Landi. Persistent betti numbers for a noise tolerant shape-based approach to image retrieval. , 34:863–872, 2013.
Issei Fujishiro, Yuriko Takeshima, Taeko Azuma, and Shigeo Takahashi. Volume data mining using 3d field topology analysis. , 20(5):46–51, 2000.
F. B. Fuller. The writhing number of a space curve. , 68(4):815–819, 1971.
M. Gameiro, Y. Hiraoka, S. Izumi, M. Kramar, K. Mischaikow, and V. Nanda. Topological measurement of protein compressibility via persistence diagrams. , 32:1–17, 2014.
H. H. Gan, D. Fera, J. Zorn, N. Shiffeldrim, M. Tang, U. Laserson, N. Kim, and T. Schlick. : [RNA]{}-as-graphs database—concepts, analysis, and features. , 20(8):1285–1291, 2004.
A. E. Garc[í]{}a. Large-amplitude nonlinear motions in proteins. , 68(17):2696, 1992.
W. Geng and G. W. Wei. Multiscale molecular dynamics using the matched interface and boundary method. , 230(2):435–457, 2011.
Weihua Geng, Sining Yu, and G. W. Wei. Treatment of charge singularities in implicit solvent models. , 127:114106, 2007.
Z. Nevin Gerek and S. Banu Ozkan. A flexible docking scheme to explore the binding selectivity of pdz domains. , 19:914–928, 2010.
R. Ghrist. Barcodes: [The]{} persistent topology of data. , 45:61–75, 2008.
N. Gillet, R. Chaudret, J. Contreras-Garc[i]{}́a, W. T. Yang, B. Silvi, and J. P. Piquemal. Coupling quantum interpretative techniques: another look at chemical mechanisms in organic reactions. , 8(11):3993–3997, 2012.
E. Gin[é]{} and V. Koltchinskii. Empirical graph [Laplacian]{} approximation of [Laplace Beltrami]{} operators: Large sample results. In [*High dimensional probability*]{}, pages 238–259. Institute of Mathematical Statistics, 2006.
L. Giot, J. S. Bader, C. Brouwer, A. Chaudhuri, B. Kuang, Y. Li, Y.L. Hao, C.E. Ooi, B. Godwin, E. Vitols, et al. A protein interaction map of drosophila melanogaster. , 302(5651):1727–1736, 2003.
N. Go, T. Noguti, and T. Nishikawa. Dynamics of a small globular protein in terms of low-frequency vibrational modes. , 80:3696 – 3700, 1983.
J. A. Grant and B. T. Pickup. A gaussian description of molecular shape. , 99:3503–3510, 1995.
J. A. Grant, B. T. Pickup, M. T. Sykes, C. A. Kitchen, and A. Nicholls. The [Gaussian Generalized Born]{} model: application to small molecules. , 9:4913–22, 2007.
J. Andrew Grant, Barry T. Pickup, and Anthony Nicholls. A smooth permittivity function for [[Poisson-Boltzmann]{}]{} solvation methods. , 22(6):608–640, 2001.
R. Gray. Vector quantization. , 1(2):4–29, 1984.
D. G[ü]{}nther, A. Jacobson, J. Reininghaus, H. P. Seidel, O. Sorkine-Hornung, and T. Weinkauf. Fast and memory-efficient topological denoising of [2D and 3D]{} scalar fields. , 20:12, 2014.
J. T. Guo, D. Xu, D. Kim, and Y. Xu. Improving the performance of [DomainParser]{} for structural domain partition using neural network. , 31(3):944–952, 2003.
L. Hagen and A. B. Kahng. New spectral methods for ratio cut partitioning and clustering. , 11(9):1074–1085, 1992.
B. Halle. Flexibility and packing in proteins. , 99:1274–1279, 2002.
Shaun Harker, Konstantin Mischaikow, Marian Mrozek, and Vidit Nanda. Discrete morse theoretic algorithms for computing homology of complexes and maps. , pages 1–34, 2013.
S. C. Harvey, Y. Zeng, and C. E. Heitsch. The icosahedral rna virus as a grotto: organizing the genome into stalagmites and stalactite. , Chapter 7:163–172, 2013.
S. Hayward and B. L. De Groot. Normal modes and essential dynamics. , pages 89–106, 2008.
M. Hein, J. Y. Audibert, and U. Von Luxburg. From graphs to manifolds–weak and strong pointwise consistency of graph [Laplacian]{}. In [*International Conference on Computational Learning Theory*]{}, pages 470–485. Springer, 2005.
C. Heitsch and S. Poznanovic. Combinatorial insights into rna secondary structure, in [ N. Jonoska and M. Saito]{}, editors. , Chapter 7:145–166, 2014.
G. Henkelman, A. Arnaldsson, and H. J[ó]{}nsson. A fast and robust algorithm for [Bader]{} decomposition of charge density. , 36(3):354–360, 2006.
K. Hinsen. Analysis of domain motions by approximate normal mode calculations. , 33:417 – 429, 1998.
K. Hinsen. Structural flexibility in proteins: impact of the crystal environment. , 24:521 – 528, 2008.
L. Holm and C. Sander. Mapping the protein universe. , 273(5275):595, 1996.
Michael Holst. PhD thesis, California Institute of Technology, 1994.
D. Horak, S Maletic, and M. Rajkovic. Persistent homology of complex networks. , 2009(03):P03034, 2009.
G. Hu, J. H. Yang, and W. J. Liu. Instability and controllability of linearly coupled oscillators: Eigenvalue analysis. , 58:4440– 4453, 1998.
H. Y. Hu, T. Laurent, M. A. Porter, and A. L. Bertozzi. A method based on total variation for network modularity optimization using the [MBO]{} scheme. , 73(6):2224–2246, 2013.
Shuangwei Hu, Martin Lundgren, and Antti J Niemi. Discrete frenet frame, inflection point solitons, and curve visualization with applications to folded proteins. , 83(6):061908, 2011.
YunKyong Hyon, Bob Eisenberg, and Chun Liu. . , 9:459–475, 2010.
R.M. Jackson and M.J. Sternberg. Dna binding and bending by hmg boxes: Energetic determinants of specificity. , 250:258–275, 1995.
A. K. Jain. Data clustering: 50 years beyond k-means. , 31(8):651–666, 2010.
H. Jeong, B. Tombor, R. Albert, Z. N. Oltvai, and A. L. Barab[á]{}si. The large-scale organization of metabolic networks. , 407(6804):651–654, 2000.
E. R. Johnson, S. Keinan, P. Mori-Sanchez, J. Contreras-Garcia, A. J. Cohen, and W.T. Yang. Revealing noncovalent interactions. , 132(18):6498–6506, 2010.
N. Jonoska and G. McColm. Complexity classes for self-assembling flexible tiles. , 410:332–346, 2009.
T. Kaczynski, K. Mischaikow, and M. Mrozek. . Springer-Verlag, 2004.
Tomasz Kaczynski, Konstantin Mischaikow, and Marian Mrozek. , volume 157 of [*Applied Mathematical Sciences*]{}. Springer-Verlag, New York, 2004.
P. M. Kasson, A. Zomorodian, S. Park, N. Singhal, L. J. Guibas, and V. S. Pande. Persistent voids a new structural metric for membrane fusion. , 23:1753–1759, 2007.
T. A. Keith and R. F. Bader. Topological analysis of magnetically induced molecular current distributions. , 99(5):3669–3682, 1993.
T. A. Keith, R. F. Bader, and Y. Aray. Structural homeomorphism between the electron density and the virial field. , 57(2):183–198, 1996.
O. Keskin, I. Bahar, D. Flatow, D. G. Covell, and R. L. Jernigan. Molecular mechanisms of chaperonin groel-groes function. , 41:491 – 501, 2002.
N. Kim, N. Shiffeldrim, H. H. Gan, and T. Schlick. Candidates for novel [RNA]{} topologies. , 341(5):1129–1144, 2004.
G. Kindlmann, R. Whitaker, T. Tasdizen, and T. M[ö]{}ller. Curvature-based transfer functions for direct volume rendering: methods and applications. , 2003.
A. Kitao, F. Hirata, and N. G[ō]{}. The effects of solvent on the conformation and the collective motions of protein: normal mode analysis and molecular dynamics simulations of melittin in water and in vacuum. , 158(2-3):447–472, 1991.
M. Kohout, K. Pernal, F. R. Wagner, and Y. Grin. Electron localizability indicator for correlated wavefunctions. [I. Parallel-spin]{} pairs. , 112(5-6):453–459, 2004.
W. L. Koltun. Precision space-filling atomic models. , 3:667–679, 1965.
D. A. Kondrashov, A. W. Van Wynsberghe, R. M. Bannen, Q. Cui, and Jr. G. N. Phillips. Protein structural variation in computational models and crystallographic data. , 15:169 – 177, 2007.
Violeta Kovacev-Nikolic, Peter Bubenik, Dragan Nikolić, and Giseon Heo. Using persistent homology and dynamical distances to analyze protein binding. , 15(1):19–38, 2016.
Bala Krishnamoorthy, Scott Provan, and Alexander Tropsha. A topological characterization of protein structure. In [*Data Mining in Biomedicine, Springer Optimization and Its Applications*]{}, pages 431–455, 2007.
L.A. Kuhn, M. A. Siani, M. E. Pique, C. L. Fisher, E. D. Getzoff, and J. A. Tainer. The interdependence of protein surface topography and bound water molecules revealed by surface accessibility and fractal density measures. , 228:13–22, 1992.
W. K[ü]{}hnel. . American mathematical Society, 2015.
S. Kundu, J. S. Melton, D. C. Sorensen, and Jr. G. N. Phillips. Dynamics of proteins in crystals: comparison of experiment with simple models. , 83:723 – 732, 2002.
S. Kundu, D. C. Sorensen, and G. N. Phillips. Automatic domain decomposition of proteins by a [Gaussian]{} network model. , 57(4):725–733, 2004.
Z. Zhang L. Li, C. Li and Emil Alexov. On the dielectric "constant” of proteins: Smooth dielectric function for macromolecular modeling and its implementation in [DelPhi]{}. , 9:2126–2136, 2013.
S. S. Lafon. . PhD thesis, Yale University, 2004.
M. Leboeuf, A. M. K[ö]{}ster, K. Jug, and D. R. Salahub. Topological analysis of the molecular electrostatic potential. , 111(11):4893–4905, 1999.
B. Lee and F. M. Richards. The interpretation of protein structures: estimation of static accessibility. , 55(3):379–400, 1971.
H Lee, H. Kang, M. K. Chung, B. Kim, and D. S. Lee. Persistent brain network homology from the perspective of dendrogram. , 31(12):2267–2277, Dec 2012.
T. I. Lee, N. J. Rinaldi, F. Robert, D. T. Odom, Z. Bar-Joseph, G. K. Gerber, N. M. Hannett, C. T. Harbison, C. M. Thompson, I. Simon, et al. Transcriptional regulatory networks in saccharomyces cerevisiae. , 298(5594):799–804, 2002.
M. Levitt, C. Sander, and P. S. Stern. The normal modes of a protein: Native bovine pancreatic trypsin inhibitor. , 24(S10):181–199, 1983.
M. Levitt, C. Sander, and P. S. Stern. Protein normal-mode dynamics: Trypsin inhibitor, crambin, ribonuclease and lysozyme. , 181(3):423 – 447, 1985.
D. W. Li and R. Br[ü]{}schweiler. All-atom contact model for understanding protein dynamics from crystallographic b-factors. , 96(8):3074–3081, 2009.
G. H. Li and Q. Cui. A coarse-grained normal mode approach for macromolecules: an efficient implementation and application to [Ca(2+)-ATPase]{}. , 83:2457 – 2474, 2002.
J. Li, P. Mach, and P. Koehl. Measuring the shapes of macromolecules - and why it matters. , 8:e201309001, 2013.
Lin Li, Chuan Li, and Emil Alexov. On the modeling of polar component of solvation energy using smooth gaussian-based dielectric function. , 13:10.1142/S0219633614400021, 2014.
V.J. LiCata and N.M. Allewell. Functionally linked hydration changes in escherichia coli aspartate transcarbamylase and its catalytic subunit. , 36:10161–10167, 1997.
C. P. Lin, S. W. Huang, Y. L. Lai, S. C. Yen, C. H. Shih, C. H. Lu, C. C. Huang, and J. K. Hwang. Deriving protein dynamical properties from weighted protein contact number. , 72(3):929–935, 2008.
Beibei Liu, Bao Wang, Rundong Zhao, Yiying Tong, and Guo Wei Wei. . , 2015.
Xu Liu, Zheng Xie, and Dongyun Yi. A fast algorithm for constructing topological structure in large data. , 14:221–238, 2012.
D. R. Livesay, S. Dallakyan, G. G. Wood, and D. J. Jacobs. . , [576]{}:[468–476]{}, [2004]{}.
J. R. L[ó]{}pez-Blanco, O. Miyashita, F. Tama, and P. Chac[ó]{}n. Normal mode analysis techniques in structural biology. , 2014.
E. N. Lorenz. Deterministic nonperiodic flow. , 20:130– 141, 1963.
L. Lov[á]{}sz. Random walks on graphs. , 2:1–46, 1993.
J. P. Ma. Usefulness and limitations of normal mode analysis in modeling dynamics of biomolecular complexes. , 13:373 – 180, 2005.
P. Mach and P. Koehl. Geometric measures of large biomolecules: Surface, volume, and pockets. , 32:3023–3038, 2011.
Ciprian Manolescu. The conley index, gauge theory, and triangulations. , 13(2):431–457, 2013.
J. A. McCammon, B. R. Gelin, and M. Karplus. Dynamics of folded proteins. , 267:585–590, 1977.
M. Meila and J. B. Shi. A random walks view of spectral segmentation. In [*In Tenth International Workshop on Artificial Intelligence and Statistics [AISTATS]{}*]{}, 2001.
E. Merkurjev, T. Kostic, and A. L. Bertozzi. An [MBO]{} scheme on graphs for classification and image processing. , 6(4):1903–1930, 2013.
P. G. Mezey. Catchment region partitioning of energy hypersurfaces,[I]{}. , 58(4):309–330, 1981.
D. M. Ming, Y. F. Kong, M. A. Lambert, Z. Huang, and J. P. Ma. How to describe protein motion without amino acid sequence and atomic coordinates. , 99(13):8620–8625, 2002.
K. Mischaikow, M Mrozek, J. Reiss, and A. Szymczak. Construction of symbolic dynamics from experimental time series. , 82:1144–1147, 1999.
K. Mischaikow and V. Nanda. Morse theory for filtrations and efficient computation of persistent homology. , 50(2):330–353, 2013.
Konstantin Mischaikow and Marian Mrozek. . Elsevier, 2002.
B. Mohar. Some applications of [Laplace]{} eigenvalues of graphs. In [*Graph symmetry*]{}, pages 225–275. Springer, 1997.
B. Mohar, Y. Alavi, G. Chartrand, and O. R. Oellermann. The laplacian spectrum of graphs. , 2(871-898):12, 1991.
A. G. Murzin, S. E. Brenner, T. Hubbard, and C. Chothia. : a structural classification of proteins database for the investigation of sequences and structures. , 247(4):536–540, 1995.
Vidit Nanda. Perseus: the persistent homology software. Software available at <http://www.sas.upenn.edu/~vnanda/perseus>.
V. Natarajan, P. Koehl, Y. Wang, and B. Hamann. Visual analysis of biomolecular surfaces. In L. Linsen, H. Hagen, and B. Hamann, editors, [*Mathematical Methods for Visualization in Medicine and Life Science*]{}, pages 237–256. Springer Verlag, 2008.
M. E. J. Newman. Modularity and community structure in networks. , 103(23):8577–8582, 2006.
M. E. J. Newman and M. Girvan. Finding and evaluating community structure in networks. , 69(2):026113, 2004.
A. Y. Ng, M. I. Jordan, and Y. Weiss. On spectral clustering: [Analysis]{} and an algorithm. , 2:849–856, 2002.
Duc D Nguyen and G. W. Wei. The impact of surface area, volume, curvature and lennard-jones potential to solvation modeling. , submitted 2016.
Duc D Nguyen, K. L. Xia, and G. W. Wei. Generalized flexibility-rigidity index. , 144:234106, 2016.
P. Niyogi, S. Smale, and S. Weinberger. A topological view of unsupervised learning from noisy data. , 40:646–663, 2011.
J. N. Onuchic, Z. Luthey-Schulten, and P. G. Wolynes. Theory of protein folding: The energy landscape perspective. , 48:545–600, 1997.
K. Opron, K. L. Xia, and G. W. Wei. Fast and anisotropic flexibility-rigidity index for protein flexibility and fluctuation analysis. , 140:234105, 2014.
Kristopher Opron, K. L. Xia, Z. Burton, and G. W. Wei. Flexibility-rigidity index for protein-nucleic acid flexibility and fluctuation analysis. , 37:1283–1295, 2016.
Kristopher Opron, K. L. Xia, and G. W. Wei. Communication: Capturing protein multiscale thermal fluctuations. , 142(211101), 2015.
C. A. Orengo, A. D. Michie, S. Jones, D. T. Jones, M. B. Swindells, and J. M. Thornton. –a hierarchic classification of protein domain structures. , 5(8):1093–1109, 1997.
Steve Y. Oudot and Donald R. Sheehy. igzag [Z]{}oology: [R]{}ips [Z]{}igzags for [H]{}omology [I]{}nference. In [*Proc. 29th Annual Symposium on Computational Geometry*]{}, pages 387–396, June 2013.
R. Overbeek, N. Larsen, G. D. Pusch, M. D’Souza, E. Selkov Jr, N. Kyrpides, M. Fonstein, N. Maltsev, and E. Selkov. : integrated system for high-throughput genome sequence analysis and metabolic reconstruction. , 28(1):123–125, 2000.
D. Pachauri, C. Hinrichs, M.K. Chung, S.C. Johnson, and V. Singh. Topology-based kernels with application to inference problems in alzheimer’s disease. , 30(10):1760–1770, Oct 2011.
J. K. Park, Robert Jernigan, and Zhijun Wu. Coarse grained normal mode analysis vs. refined gaussian network model for protein residue-level structural fluctuations. , 75:124 –160, 2013.
L. M. Pecora, T. L. Carroll, G. A. Johnson, and D. J. Mar. Fundamentals of synchronization in chaotic systems, concepts and applications. , 7:520– 543, 1997.
A. M. Pend[á]{}s and V. Lua[ñ]{}a. Curvature of interatomic surfaces. i. fundamentals. , 119(15):7633–7642, 2003.
J. A. Perea, A. Deckard, S. B. Haase, and J. Harer. Sw1pers: Sliding windows and 1-persistence scoring; discovering periodicity in gene expression time series data. , 16:257, 2015.
J. A. Perea and J. Harer. Sliding windows and persistence: An application of topological methods to signal analysis. , 15:799–838, 2015.
D. Petrey and B. Honig. : Visualization, surface properties, and electrostatics of macromolecular structures and sequences. , 374:492–509, 2003.
W. F. Pohl. and differential geometry. , 3(1):20–27, 1980.
J. H. Poincaré. Sur le probleme des trois corps et les équations de la dynamique. divergence des s�ries de m. lindstedt. , 13:A3–A270, 1890.
P. L. Popelier. On the differential geometry of interatomic surfaces. , 74(6):829–838, 1996.
P. L. Popelier. Quantum chemical topology: on bonds and potentials. In [*Intermolecular forces and clusters I*]{}, pages 1–56. Springer, 2005.
P. L. Popelier, F. M. Aicken, and S. E. O’Brien. Atoms in molecules. , 1:143–198, 2000.
J. R. Quine, S. Achuthan, T. Asbury, R. Bertram, M.S. Chapman, J. Hu, and T.A. Cross. Intensity and mosaic spread analysis from [PISEMA]{} tensors in solid-state [NMR]{}. , 179(2):190–198, 2006.
JR Quine, Timothy A Cross, Michael S Chapman, and Richard Bertram. Mathematical aspects of protein structure determination with nmr orientational restraints. , 66(6):1705–1730, 2004.
A. J. Rader, D. H. Vlad, and I. Bahar. Maturation dynamics of bacteriophage hk97 capsid. , 13:413 – 421, 2005.
J. C. Rain, L. Selig, H. De Reuse, V. Battaglia, C. Reverdy, S. Simon, G. Lenzen, F. Petel, J. Wojcik, V. Sch[ä]{}chter, Y. Chemama, A. Labigne, and P. Legrain. The protein-protein interaction map of [Helicobacter]{} pylori. , 409(6817):211–215, 2001.
T.M. Raschke, J. Tsai, and M. Levitt. Quantification of the hydrophobic interaction by simulations of the aggregation of small hydrophobic solutes in water. , 98:5965–5969, 2001.
F. M. Richards. Areas, volumes, packing, and protein structure. , 6(1):151–176, 1977.
Bastian Rieck, Hubert Mara, and Heike Leitte. Multivariate data analysis using persistence-based filtering and topological signatures. , 18:2382–2391, 2012.
Vanessa Robins. Towards computing homology from finite approximations. In [*Topology Proceedings*]{}, volume 24, pages 503–532, 1999.
W. Rocchia, S. Sridharan, A. Nicholls, E Alexov, A Chiabrera, and B. Honig. . , 23:128 – 137, 2002.
F. Sadre-Marandi, J. Liu, S. Tavener, and C. Chen. Generating vectors for the lattice structures of tubular and conical viral capsids. , 2:128–140, 2014.
H. Salgado, A. Santos-Zavaleta, S. Gama-Castro, M. Peralta-Gil, M. I. Pe[ñ]{}aloza-Sp[í]{}nola, A. Mart[í]{}nez-Antonio, P. D. Karp, and J. Collado-Vides. The comprehensive updated regulatory network of escherichia coli k-12. , 7(1):1, 2006.
P. T. Sander and S. W. Zucker. Inferring surface trace and differential structure from [3D]{} images. , 12(9):833–854, 1990.
M. F. Sanner, A. J. Olson, and J. C. Spehner. Reduced surface: An efficient way to compute molecular surfaces. , 38:305–320, 1996.
T. Schlick and W. K. Olson. Trefoil knotting revealed by molecular dynamics simulations of supercoiled [DNA]{}. , 257(5073):1110–1115, 1992.
H. L. Schmider and A. D. Becke. Chemical content of the kinetic energy density. , 527(1):51–61, 2000.
M Schroder and R. J. Kaufman. The mammalian unfolded protein response. , 74:739 – 789, 2005.
J. B. Shi and J. Malik. Normalized cuts and image segmentation. , 22(8):888–905, 2000.
X. Shi and P. Koehl. Geometry and topology for modeling biomolecular surfaces. , 50:1–34, 2011.
V. D. Silva and R Ghrist. Blind swarms for coverage in 2-d. In [*In Proceedings of Robotics: Science and Systems*]{}, page 01, 2005.
B. Silvi and A. Savin. Classification of chemical bonds based on topological analysis of electron localization functions. , 371(6499):683–686, 1994.
G. Singh, F. M[é]{}moli, and G. E. Carlsson. Topological methods for the analysis of high dimensional data sets and [ 3D]{} object recognition. In [*SPBG*]{}, pages 91–100, 2007.
G. Singh, F. Memoli, T. Ishkhanov, G. Sapiro, G. Carlsson, and D. L. Ringach. Topological analysis of population activity in visual cortex. , 8(8), 2008.
L. Skjaerven, S. M. Hollup, and N. Reuter. Normal mode analysis for proteins. , 898:42 – 48, 2009.
O. Soldea, G. Elber, and E. Rivlin. Global segmentation and curvature analysis of volumetric data sets using trivariate b-spline functions. , 28(2):265 – 278, 2006.
G. Song and R. L. Jernigan. vgnm: a better model for understanding the dynamics of proteins in crystals. , 369(3):880 – 893, 2007.
R. S. Spolar and M. T. Record Jr. Coupling of local folding to site[-]{}specific binding of proteins to dna. , 263:777–784, 1994.
E. M. Stokely and S. Y. Wu. Surface parametrization and curvature measurement of arbitrary [3-D]{} objects: five practical methods. , 14(8):833–840, 1992.
D. Strombom. . , 2007.
D. W. Sumners. Knot theory and [DNA]{}. In [*Proceedings of Symposia in Applied Mathematics*]{}, volume 45, pages 39–72, 1992.
F. Tama and C. K. Brooks III. Diversity and identity of mechanical properties of icosahedral viral capsids studied with elastic network normal mode analysis. , 345:299 – 314, 2005.
F. Tama and Y. H. Sanejouand. Conformational change of proteins arising from normal mode calculations. , 14:1 – 6, 2001.
F. Tama, M. Valle, J. Frank, and C. K. Brooks III. Dynamic reorganization of the functionally active ribosome explored by normal mode analysis and cryo-electron microscopy. , 100(16):9319 – 9323, 2003.
F. Tama, W. Wriggers, and C. L. Brooks. Exploring global distortions of biological macromolecules and assemblies from low-resolution structural information and elastic network theory. , 321(2):297–305, 2002.
M. Tasumi, H. Takenchi, S. Ataka, A. M. Dwidedi, and S. Krimm. Normal vibrations of proteins: Glucagon. , 21:711 – 714, 1982.
Andrew Tausz, Mikael Vejdemo-Johansson, and Henry Adams. Javaplex: A research software package for persistent (co)homology. Software available at <http://code.google.com/p/javaplex>, 2011.
M. M. Tirion. Large amplitude elastic motions in proteins from a single-parameter, atomic analysis. , 77:1905 – 1908, 1996.
R. Twarock and N. Jonoska. Blueprints for dodecahedral dna cages. , 41:304043 –304057, 2008.
V. Uversky and A. K. Dunker. Controlled chaos. , 322:1340 – 1341, 2008.
P. W. Verbeek and L. J. Van Vliet. Curvature and bending energy in digitized [2D and 3D]{} images. In [*8th Scandinavian Conference on Image Analysis, Tromso, Norway*]{}, 1993.
S. Veretnik and I. Shindyalov. Computational methods for domain partitioning of protein structures. In [*Computational Methods for Protein Structure Prediction and Modeling*]{}, pages 125–145. Springer New York, 2007.
N. Volkmann. Methods for segmentation and interpretation of electron tomographic reconstructions. In [*Methods Enzymol*]{}, volume 483, pages 31–46, 2010.
A. Vologodskii. . CRC Press, 1992.
U. Von Luxburg. A tutorial on spectral clustering. , 17(4):395–416, 2007.
B. Wang and G. W. Wei. Parameter optimization in differential geometry based solvation models. , 143:134119, 2015.
B. Wang and G. W. Wei. Object-oriented persistent homology. , 305:276–299, 2016.
Bao Wang, Chengzhang Wang, and G. W. Wei. Feature functional theory - solvation predictor (fft-sp) for the blind prediction of solvation free energy. , submitted 2016.
Bao Wang, Zhixiong Zhao, and G. W. Wei. Automatic parametrization of non-polar implicit solvent models for the blind prediction of solvation free energies. , 145:124110, 2016.
Bao Wang, Zhixiong Zhao, and G. W. Wei. Feature functional theory - binding predictor (fft-bp) for the blind prediction of binding free energy. , submitted 2016.
Bei Wang, Brian Summa, Valerio Pascucci, and M. Vejdemo-Johansson. Branching and circular features in high dimensional data. , 17:1902–1911, 2011.
Lin Wang, Lin Li, and Emil Alexov. predictions for proteins, [RNAs and DNAs with the Gaussian]{} dielectric function using [DelPhiPKa]{}. , 83:2186–2197, 2015.
Y. Wang, A. J. Rader, I. Bahar, and R. L. Jernigan. Global ribosome motions revealed with elastic network model. , 147:302 – 314, 2004.
A. Warshel and M. Levitt. Theoretical studies of enzymic reactions: Dielectric, electrostatic and steric stabilization of the carbonium ion in the reaction of lysozyme. , 103:227–249, 1976.
G. W. Wei. Wavelets generated by using discrete singular convolution kernels. , 33:8577 – 8596, 2000.
G. W. Wei. Differential geometry based multiscale models. , 72:1562 – 1622, 2010.
G. W. Wei, Y. H. Sun, Y. C. Zhou, and M. Feig. Molecular multiresolution surfaces. , pages 1 – 11, 2005.
Guo Wei Wei. Multiscale, multiphysics and multidomain models [I: Basic]{} theory. , 12(8):1341006, 2013.
Guo Wei Wei. Mathematical molecular bioscience and biophysics. , 49(7), September 2016.
Guo-Wei Wei, Qiong Zheng, Zhan Chen, and Kelin Xia. Variational multiscale models for charge transport. , 54(4):699 – 754, 2012.
S. H. White and W. C. Wimley. Membrane protein folding and stability: Physical principles. , 28:319–365, 1999.
David Whitley. . Royal Society of Chemistry, 2012.
K. L. Xia, X. Feng, Y. Y. Tong, and G. W. Wei. Multiscale geometric modeling of macromolecules i: Cartesian representation. , 275:912–936, 2014.
K. L. Xia, X. Feng, Y. Y. Tong, and G. W. Wei. Persistent homology for the quantitative prediction of fullerene stability. , 36:408–422, 2015.
K. L. Xia, K. Opron, and G. W. Wei. Multiscale multiphysics and multidomain models — [ Flexibility]{} and rigidity. , 139:194109, 2013.
K. L. Xia, K. Opron, and G. W. Wei. Multiscale [Gaussian network model (mGNM)]{} and multiscale anisotropic network model [(mANM)]{}. , 143:204106, 2015.
K. L. Xia and G. W. Wei. Molecular nonlinear dynamics and protein thermal uncertainty quantification. , 24:013103, 2014.
K. L. Xia and G. W. Wei. Persistent homology analysis of protein structure, flexibility and folding. , 30:814–844, 2014.
K. L. Xia and G. W. Wei. Multidimensional persistence in biomolecular data. , 36:1502–1520, 2015.
K. L. Xia and G. W. Wei. Persistent topology for [cryo-EM]{} data analysis. , 31:e02719, 2015.
K. L. Xia, Z. X. Zhao, and G. W. Wei. Multiresolution persistent homology for excessively large biomolecular datasets. , 143:134103, 2015.
K. L. Xia, Z. X. Zhao, and G. W. Wei. Multiresolution topological simplification. , 22:1–5, 2015.
C. Xu, D. Tobi, and I. Bahar. Allosteric changes in protein structure computed by a simple mechanical model: hemoglobin t <–> r2 transition. , 333:153 – 168, 2003.
Guoliang Xu, Qing Pan, and Chandrajit L. Bajaj. Discrete surface modeling using partial differential equations. , 23(2):125–145, 2006.
L. W. Yang and C. P. Chng. Coarse-grained models reveal functional dynamics–[I]{}. elastic network models–theories, comparisons and perspectives. , 2:25 – 45, 2008.
Yang, A Rader, Xiong Liu, Cristopher Jursa, Shann Chen, Hassan Karimi, and Ivet Bahar. online computation of structural dynamics using the gaussian network model. , 34(Web Server issue):W24–W31, 2006.
Y. Yao, J. Sun, X. H. Huang, G. R. Bowman, G. Singh, M. Lesnick, L. J. Guibas, V. S. Pande, and G. Carlsson. Topological methods for exploring low-density states in biomolecular folding pathways. , 130:144115, 2009.
S. N. Yu and G. W. Wei. Three-dimensional matched interface and boundary [(MIB)]{} method for treating geometric singularities. , 227:602–632, 2007.
Z. Yu, M. Holst, T. Hayashi, C. L. Bajaj, M. H. Ellisman, J. A. McCammon, and M. Hoshijima. Three-dimensional geometric modeling of membrane-bound organelles in ventricular myocytes: Bridging the gap between microscopic imaging and mathematical simulation. , 164:304–313, 2008.
Z. Y. Yu, M. Holst, Y. Cheng, and J. A. McCammon. Feature-preserving adaptive mesh generation for molecular shape modeling and simulation. , 26:1370–1380, 2008.
Lihi Zelnik-manor and Pietro Perona. Self-tuning spectral clustering. In [*Advances in Neural Information Processing Systems 17*]{}, pages 1601–1608. MIT Press, 2004.
Y. Zhang, H. Yu, J. H. Qin, and B. C. Lin. A microfluidic dna computing processor for gene expression analysis and gene drug synthesisn. , 3(044105), 2009.
Shan Zhao. Pseudo-time-coupled nonlinear models for biomolecular surface representation and solvation analysis. , 27:1964–1981, 2011.
Shan Zhao. Operator splitting [ADI]{} schemes for pseudo-time coupled nonlinear solvation simulations. , 257:1000 – 1021, 2014.
Q. Zheng, S. Y. Yang, and G. W. Wei. . , 28:291–316, 2012.
Qiong Zheng, Duan Chen, and G. W. Wei. Second-order [Poisson-Nernst-Planck]{} solver for ion transport. , 230:5239 – 5262, 2011.
Qiong Zheng and G. W. Wei. . , 134:194101, 2011.
W. Zheng, B. R. Brooks, and D. Thirumalai. Allosteric transitions in the chaperonin groel are captured by a dominant normal mode that is most robust to sequence variations. , 93:2289 – 2299, 2007.
W. J. Zheng and S. Doniach. A comparative study of motor-protein motions by using a simple elastic-network model. , 100(23):13253 – 13258, 2003.
Y. C. Zhou, M. J. Holst, and J. A. McCammon. A nonlinear elasticity model of macromolecular conformational change induced by electrostatic forces. , 340:135–164, 2008.
A. Zomorodian and G. Carlsson. Computing persistent homology. , 33:249–274, 2005.
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Biological systems are known to communicate by diffusing chemical signals in the surrounding medium. However, most of the recent literature has neglected the *electron transfer* mechanism occurring amongst living cells, and its role in cell-cell communication. Each cell relies on a continuous flow of electrons from its electron donor to its electron acceptor through the electron transport chain to produce energy in the form of the molecule adenosine triphosphate, and to sustain the cell’s vital operations and functions. While the importance of biological electron transfer is well-known for individual cells, the past decade has also brought about remarkable discoveries of multi-cellular microbial communities that transfer electrons between cells and across centimeter length scales, *e.g.*, biofilms and multi-cellular bacterial cables. These experimental observations open up new frontiers in the design of electron-based communications networks in microbial communities, which may coexist with the more well-known communication strategies based on molecular diffusion, while benefiting from a much shorter communication delay. This paper develops a stochastic model that links the electron transfer mechanism to the energetic state of the cell. The model is also extensible to larger communities, by allowing for electron exchange between neighboring cells. Moreover, the parameters of the stochastic model are fit to experimental data available in the literature, and are shown to provide a good fit.'
author:
- 'Nicolò Michelusi, Sahand Pirbadian, Mohamed Y. El-Naggar and Urbashi Mitra [^1] [^2] [^3]'
bibliography:
- 'IEEEabrv.bib'
- 'Refs.bib'
title: |
A Stochastic Model for Electron Transfer\
in Bacterial Cables
---
Introduction
============
Biological systems are known to communicate by diffusing chemical signals in the surrounding medium. One example is *quorum sensing* [@Bassler; @Visick15082005; @Nealson], where the concentration of certain signature chemical compounds emitted by the bacteria is used to estimate the bacterial population size, so as to simultaneously activate a certain collective behavior. [More recently, molecular communication has been proposed as a viable communication scheme for nanodevices and nanonetworks, and is under IEEE standards consideration [@P1906].]{} The performance evaluation, optimization and design of molecular communications systems opens up new challenges in the information theory [@bush2010nanoscale; @Nakano; @Akyildiz1; @Rose]. The achievable capacity of the chemical channel using molecular communication is investigated in [@Eckford07; @Kadloor], under Brownian motion, and in [@Einolghozati], under a diffusion channel. In [@Fekri], a new architecture for networks of bacteria to form a data collecting network is described, and aspects such as reliability and speed of convergence of consensus are investigated. In [@Arjmandi; @Mosayebi], a new molecular modulation scheme for nanonetworks is proposed and analyzed, based on the idea of time-sharing between different types of molecules in order to effectively suppress the interference. In [@Oiwa], an in-vitro molecular communication system is designed and, in [@Kuran201086], an energy model is proposed, based on molecular diffusion.
While communication via chemical signals has been the focus of most prior investigations, experimental evidence on the microbial emission and response to three physical signals, *i.e.*, sound waves, electromagnetic radiation and electric currents, suggests that physical modes of microbial communication could be widespread in nature [@Reguera2]. In particular, communication exploiting electron transfer in a bacterial network has previously been observed in nature [@Pfeffer] and in bacterial colonies in lab [@Kato]. This multi-cellular communication is usually triggered by extreme environmental conditions, *e.g.*, lack of electron donor (ED) or electron acceptor (EA), in turn resulting in various gene expression levels and functions in different cells within the community, and enables the entire community to survive under harsh conditions. Electron transfer is fundamental to cellular respiration: each cell relies on a continuous flow of electrons from an ED to an EA through the cell’s electron transport chain (ETC) to produce energy in the form of the molecule adenosine triphosphate (ATP), and to sustain its vital operations and functions. This strategy, known as *oxidative phosphorylation*, is employed by all respiratory microorganisms. In this regard, we can view the flow of one electron from the ED to the EA as an energy unit which is harvested from the surrounding medium to power the operations of the cell, and stored in an internal “rechargeable battery” (energy queue, *e.g.*, see the literature on *energy harvesting* for wireless communications and references therein [@Gunduz; @MichelusiEH; @Liu]). While the importance of biological electron transfer and oxidative phosphorylation is well-known for individual cells, the past decade has also brought about remarkable discoveries of multi-cellular microbial communities that transfer electrons between cells and across much larger length scales than previously thought [@Naggar]. Within the span of only a few years, observations of microbial electron transfer have jumped from nanometer to centimeter length scales, and the structural basis of this remarkably long-range transfer has evolved from recently discovered molecular assemblies known as *bacterial nanowires* [@Naggar; @Naggar2; @Pirbadian], to entire macroscopic architectures, including biofilms and multi-cellular bacterial cables, consisting of thousands of cells lined up end-to-end in marine sediments [@Reguera; @Pfeffer] (see Fig. \[figb\]). Therein, the cells in the deeper regions of the sediment where the ED is located extract more electrons, while the cells in the upper layers, where Oxygen (an EA) is more abundant, have a heightened transfer of electrons to the EA.
The survival of the whole system relies on this division of labor, with the intermediate cells operating as “relays” of electrons to coordinate this collective response to the spatial separation of ED and EA. It is worth noticing that other biological cable-like mechanisms exist in nature, enabling cell-cell communication: *tunneling nanotubes* connect two animal cells for transport of organelles and membrane vesicles and create complex networks of interconnected cells [@Rustom]; in the bacterial world, *Myxococcus xanthus* cells form membrane tubes that connect cells to one another in order to transfer outer membrane content [@Remis].
These experimental observations raise the possibility of an electron-based communications network in microbial communities, which may coexist with the more well-known communication strategies based on molecular diffusion [@Naggar; @Reguera2; @Arjmandi]. For microbes, the advantage of electron-based communications is clear: in contrast to the relatively slow diffusion of whole molecules via Brownian motion, electron transfer is a rapid process that enables cells to quickly sense and respond to their environment. As an example of a communications architecture based on electron transfer, consider a system composed of an ED terminal (transmitter, or electron source) which operates as the signal encoder, an EA terminal (receiver, or electron sink) and the network of bacteria; the electron signal, encoded by the ED terminal and input into the network, is then relayed in a multi-hop fashion, following the natural laws of electron transfer within each cell and across neighboring cells, which this paper aims at modeling; the flow of electrons is finally collected at the EA terminal. Such an electron signal, coupled with the energetic state of each cell, can be “decoded” by the individual cells to activate a certain desired gene expression. For instance, in a biofilm formed on a surface, bacteria interact with each other and with a solid phase terminal EA via electron transfer, which serves *both* as a respiratory advantage and a communications scheme for bacteria to adapt to their environment. Additionally, electron transfer can be employed in place of molecular diffusion for quickly transporting information in nanonetworks. In particular, information can be encoded in the concentration of electrons released by the encoder into the bacterial cable, using a technique termed *concentration shift keying* [@Kuran5962989; @Arjmandi]. The additional challenge with respect to molecular diffusion is that the electron is *both* an energy carrier involved in the energy production for the cell to sustain its functionalities, and an information carrier, which enables the transport of information between nanodevices, thus introducing additional constraints in the encoded signal. Electron-based communication presents significant advantages, as discussed above, but this phenomenon also raises new intriguing questions. While a single cell can extract enough free energy to power life’s reactions by exploiting the redox potential difference between ED oxidation and EA reduction, how can the same potential difference be used to power an entire multicellular assembly such as the *Desulfobulbaceae* bacterial cables [@Pfeffer]? Specifically, can intermediate cells survive without access to chemical ED or EA, by exploiting the potential difference between cells in the deeper sediment (sulfide oxidizers) and cells in the oxic zone (oxygen reducers)? For a cable consisting of thousands of cells this appears unlikely, since the free energy available for an intermediate cell is inversely proportional to the total number of cells. Are additional, yet unknown, electron sources and sinks necessary to maintain the whole community? These questions necessitate flexible models that analyze emerging experimental data in order to elucidate the energetics of individual cells, as presented here, and, eventually, whole bacterial cables or biofilms.
In order to enable the modeling and control of such microbial communications network and guide future experiments, in this paper we set out to develop a stochastic queuing theoretic model that links electron transfer to the energetic state of the cell (*e.g.*, ATP concentration or energy charge potential). We show how the proposed model can be extended to larger communities (*e.g.*, cables, biofilms), by allowing for electron transfer between neighboring cells. In particular, we analyze the stochastic model for an isolated cell, which is the building block of multi-cellular networks, and provide an example of the application of the proposed framework to the computation of the cell’s lifetime. Finally, we design a parameter estimation framework and fit the parameters of the model to experimental data available in the literature. The prediction curves are compared to experimental ones, showing a good fit. This paper represents a preliminary essential modeling step towards the design and analysis of bacterial communications networks, and provides the ground to model and control bacterial interactions (*e.g.*, gene expressions) induced by the electron transfer signal, and to analyze information theoretic aspects, such as the interplay between information capacity and lifetime of the cells, as well as communication reliability and delay. This paper is organized as follows. In Sec. \[stochmodel\], we present a stochastic model for the cell, and for the interconnection of cells via electron transfer. In Sec. \[isolatedcell\], we specialize the model to the case of an isolated cell. In Sec. \[celllifetime\], we present an application of the proposed framework to compute the lifetime of an isolated cell. In Sec. \[estimation\], we present a parameter estimation framework and fit the parameters of the model to experimental data. Finally, Sec. \[future\] presents some future work and Sec. \[conclu\] concludes the paper.
Stochastic cell model {#stochmodel}
=====================
In this section, we describe a continuous time stochastic model for the dynamics of electron flow and ATP production and consumption within a single cell, represented in Fig. \[fig1\]. This is the building block of more complex multi-cellular systems, *e.g.*, a bacterial cable, also represented in Fig. \[fig1\]. The cell is modeled as a system with an input electron flow coming either from the ED via molecular diffusion, or from a neighboring cell via electron transfer, and an output flow of electrons leaving either toward the EA via molecular diffusion, or toward the next cell in the cable via electron transfer. We first review these well known biological and physical mechanisms and then provide our new stochastic model. Inside the cell, the conventional pathway of electron flow, enabled by the presence of the ED and the EA, is as follows (see the numbers in Fig. \[fig1\]):
1. ED molecules permeate inside the cell via molecular diffusion;
2. The presence of these ED molecules inside the cell results in reactions that produce electron-containing carriers (*e.g.*, NADH). These are collected in the *internal electron carrier pool* (IECP, Fig. \[fig1\]). The electron carriers diffusively transfer electrons to the ETC, which is partially localized in the cell inner membrane;
3. The electrons originating from the electron carriers flow through the ETC and are discarded by either a soluble and internalized EA (*e.g.*, molecular Oxygen) or are transferred through the periplasm to the outer membrane and deposited on an extracellular EA;
4. The electron flow through the ETC results in the production of a proton concentration gradient (proton motive force [@Lane]) across the inner membrane of the cell;
5. The proton motive force is utilized by an inner membrane protein called ATP synthase to produce ATP as an energy reserve that will later be used for various functions in the cell. The ATP produced in this way is collected in the *ATP pool* in Fig. \[fig1\], and used by the cell to sustain its vital operations and functions.
Alternatively, when the cells are organized in multi-cellular structures, *e.g.*, bacterial cables, an additional pathway of electron flow may exist, termed *intercellular electron transfer* (IET), which involves only a transfer of electrons between neighboring cells, as opposed to molecules (ED and EA) diffusing through the cell membrane. In this regime, one or both of the ED and the EA are replaced by neighboring cells in a network of interconnected cells. In other words, IET can be substituted for the ED or the EA, enabling cells to survive even in the absence of the ED or the EA. In this case, the pathway for the electrons is as follows:
6. High-energy[^4] electrons localized in the outer-membrane of a neighboring cell are transported to the host cell, and utilized in its ETC to produce ATP. Therefore, the electrons creating the proton motive force are not originating from the chemical carriers such as NADH, but instead are entering directly from the neighboring cell;
7. The electrons subsequently leave the ETC and move to the outer-membrane of the host cell, and are transferred to another neighboring cell that, in turn, uses these electrons to produce ATP.
As a result, this cooperative strategy creates a multi-cellular ETC that utilizes IET to distribute electrons throughout an entire bacterial network. These electrons originate from the ED localized on one end of the network to the locally available EA on the other end. The collective electron transport through this network provides energy for all cells involved to maintain their vital operations. The conventional ED-EA and IET processes may coexist, depending on the availability of both ED and EA in the medium where the cell is operating and on the connectivity of the cell to neighboring ones. For instance, if the concentration of ED and EA is sufficiently large, only the conventional pathway is used by the cell for ATP production. In contrast, if such concentration is too small to support ATP production, only IET from/to neighboring cells may be active. In accordance with the steps outlined above, we propose the following stochastic model for the cell, as depicted in Fig. \[fig1\]. This model incorporates four pools:
1. The *IECP*, containing the electron carrier molecules (*e.g.*, NADH) produced as a result of ED diffusion across the cell membrane and chemical processes occurring inside the cell;
2. The *ATP pool*, containing all the ATP molecules produced as a result of electron flow from the electron carriers through the ETC to the EA;
3. The *external membrane pool*, which involves the extracellular respiratory pathway of the cell in the outer membrane. This part of the ETC typically includes heme-containing c-type cytochromes that facilitate electron transfer outside of the inner membrane and into the terminal EA. In fact, the accumulation of these c-type cytochromes in the outer membrane forms the external membrane pool. In order to incorporate the case of IET into this model, we assume that the external membrane pool is further divided into two parts:
1. *High energy external membrane* (HEEM), which contains high energy electrons coming from previous cells in the cable;
2. *Low energy external membrane* (LEEM), which collects low energy electrons that have been used to synthesize ATP, before they are transferred to a neighboring cell.
Each pool in this model has a corresponding inflow and outflow of electrons that connect that pool to the others, and one cell to the next in the cable:
1. The IECP gains electrons from ED molecules diffusing into the cell and transforming into electron carriers through a series of reactions; we model this as a flow with rate $\lambda_{CH}$ joining the IECP in Fig. \[fig1\]. The electrons leave this pool to the ETC (cell inner membrane) to produce ATP, modeled as another flow with rate $\mu_{CH}$ leaving the IECP in Fig. \[fig1\];
2. Alternatively, electrons are transferred from neighboring cells into the HEEM, corresponding to the flow with rate $\lambda_{EXT}^{(H)}$ in Fig. \[fig1\]. These electrons leave this pool to the ETC (cell inner membrane) to produce ATP, modeled as another flow with rate $\mu_{EXT}^{(H)}$ leaving the IECP in Fig. \[fig1\];
3. The electron flow out of the first pool (either the IECP or the HEEM) directly causes the synthesis of ATP, so that the overall flow into the ATP pool is $\mu_{CH}+\mu_{EXT}^{(H)}$. On the other hand, ATP consumption via ATP hydrolysis within the cell through various functions is responsible for the ATP molecules leaving the ATP pool, with rate $\mu_{ATP}$;
4. As a simplification, we assume that there are two major pathways for the electron output of the ETC: internalized molecular Oxygen in aerobic conditions and transport to the external membrane in anaerobic conditions. The former case, modeled as a flow with rate $\mu_{OUT}$ leaving the cell to the EA in Fig. \[fig1\], does not involve the external membrane pool but only the EA. In contrast, the latter involves the extracellular respiration pathway, which includes the external membrane. The electrons in this case are collected in the LEEM, *i.e.*, the flow with rate $\lambda_{EXT}^{(L)}$ in Fig. \[fig1\]. The electrons in this pool can, in turn, be transferred to neighboring cells, modeled as a flow with rate $\mu_{EXT}^{(L)}$ leaving the LEEM of cell 1 to the HEEM of cell 2 in Fig. \[fig1\], or to solid phase terminal EAs, not represented in Fig. \[fig1\].
In addition, because typical values for transfer rates between electron carriers (*e.g.*, outer-membrane cytochromes) on the cell exterior are relatively high [@Pirbadian], one can assume that the external membranes of neighboring cells have high transfer rates between one another, *i.e.*, when IET is active, we have $\mu_{EXT}^{(L)}=\lambda_{EXT}^{(H)}=\infty$, so that any electron collected in the LEEM is instantaneously transferred to the HEEM of the neighboring cell in the cable. Under these assumptions, we can simplify the model by combining the LEEM and HEEM pools of Fig. \[fig1\] together, so that any pair of neighboring cells share a single pool for IET. On the other hand, if the cell is isolated, no IET occurs, hence $\mu_{EXT}^{(L)}=0$ and/or $\lambda_{EXT}^{(H)}=0$. This latter case will be studied in more detail in Sec. \[isolatedcell\].
We model the cell as a finite state machine, and characterize the state of the cell and its stochastic evolution. The *internal state* of a given cell at time $t$ is defined as $$\begin{aligned}
\label{Si}
\mathbf s_I(t)=\left(m_{CH}(t),n_{ATP}(t),q_{EXT}^{(L)}(t),q_{EXT}^{(H)}(t)\right),\end{aligned}$$ where:
- $m_{CH}(t)$ is the number of electrons in the IECP that will participate in the synthesis of ATP; these electrons are carried by ED units which diffuse through the membrane into the cell (*e.g.*, lactate), and are bonded to electron carriers within the cell (*e.g.*, NADH); $m_{CH}(t)$ takes value in the set $\mathcal M_{CH}\equiv\{0,1,\dots,M_{CH}\}$, where $M_{CH}$ is the *electro-chemical storage capacity* of the cell;
- $n_{ATP}(t)$ is the number of ATP molecules within the cell, taking value in the set $\mathcal N_{AXP}\equiv\{0,1,\dots,N_{AXP}\}$, where $N_{AXP}$ is the overall number of ATP plus ADP molecules in the cell, which is assumed to be constant over time; $N_{AXP}$ also represents the maximum number of ATP molecules which can be present within the cell at any time (when no ADP is present);
- $q_{EXT}^{(H)}(t)$ is the number of electrons in the HEEM, taking value in the set $\mathcal Q_{EXT}^{(H)}\equiv \{0,1,\dots,Q_{EXT}^{(H)}\}$, where $Q_{EXT}^{(H)}$ is the electron “storage capacity” of the HEEM;
- $q_{EXT}^{(L)}(t)$ is the number of electrons in the LEEM, taking value in the set $\mathcal Q_{EXT}^{(L)}\equiv \{0,1,\dots,Q_{EXT}^{(L)}\}$, where $Q_{EXT}^{(L)}$ is the electron “storage capacity” of the LEEM.
For simplicity, we assume that all the quantities related to the state of the cell and to the flows of electrons/molecules are in terms of equivalent number of electrons involved, rather than molecular units. Hence, for instance, the ATP level in the ATP pool, $n_{ATP}(t)$, actually represents the equivalent number of electrons involved in the synthesis of the corresponding quantity of ATP available in the cell. Similarly, the level of NADH in the IECP, $m_{CH}(t)$, is expressed in terms of the equivalent number of electrons carried by the electron carriers, which actively synthesize ATP. A similar interpretation holds for the flows (of electrons, rather than molecules or ${\,\,\,\mathrm{mM}}$, where ${\,\,\,\mathrm{M}}$ stands for “1 molar”). Transition from one representation (electrons) to the other (molecules or ${\,\,\,\mathrm{mM}}$) is possible by appropriate scaling.
Moreover, while in the following analysis we assume that one “unit” corresponds to one electron, this can be generalized to the case where one “unit” corresponds to $N_E$ electrons, so that, *e.g.*, $n_{ATP}$ units in the ATP pool correspond to $N_E n_{ATP}$ electrons.
Note that, if the cell is connected to other cells in a larger community, the low (respectively, high) energy external membrane is shared with the high (low) energy external membrane of the neighboring cell, owing to the *high transfer rate approximation*, as explained above. Additionally, we denote the state of death of the cell as $\text{DEAD}$ (to be specified later). The state space of the cell is denoted as $$\begin{aligned}
\mathcal S_I\equiv \left(\mathcal M_{CH}\times\mathcal N_{AXP}\times\mathcal Q_{EXT}^{(L)}\times\mathcal Q_{EXT}^{(H)}\right)\cup\{\text{DEAD}\}.\end{aligned}$$ Note that the behavior of the cell is influenced by the concentration of the ED and the EA in the surrounding medium. Therefore, we also define the *external state* of the cell as $\mathbf s_{E}(t)=(\sigma_{D}(t),\sigma_{A}(t))$, where $\sigma_D(t)$ and $\sigma_A(t)$ are, respectively, the external concentration of the ED and the EA. For simplicity, we assume that $\mathbf s_{E}(t)$ is an exogenous process, not influenced by the cell dynamics, *i.e.*, the consumption of the ED and the EA by the cell does not influence their concentration in the surrounding medium. This requires that the medium in which cells are suspended is continuously being replaced by fresh medium containing a constant amount of the ED and the EA. Otherwise a high cell concentration would use up all the resources in the time-scales relevant to this model. This aspect will be considered in future work, and is beyond the scope of the current paper.
-- -- -- --
-- -- -- --
The internal state process of cell $i$, $\mathbf s_I^{(i)}(t)\in\mathcal S_I$ (see Eq. \[Si\]), is time-varying and stochastic; $\mathbf s_I^{(i)}(t)$ evolves as a consequence of electro/chemical reactions occurring within the cell, chemical diffusion through the cell membrane, and IET from the neighboring cell $i-1$ to the neighboring cell $i+1$. The evolution of $\mathbf s_I^{(i)}(t)$ is also influenced by the external state $\mathbf s_E^{(i)}(t)$ experienced by the cell. We define the following processes affecting the evolution of $\mathbf s_I^{(i)}(t)$, all of which, for analytical tractability, are modeled as Poisson processes with state-dependent rates; these processes are represented in Fig. \[fig1\] and the corresponding state transitions are depicted in Fig. \[fig3\]:
- through the membrane: ED molecules carry electrons to synthesize ATP, which are stored in the IECP; this process occurs with rate $\lambda_{CH}(\mathbf s_I^{(i)}(t);\mathbf s_{E}^{(i)}(t))$ \[electrons/s\]. Whenever an ED diffuses through the membrane within the cell (say, at time $t$), the state $m_{CH}^{(i)}(t)$ increases by one unit (Fig. \[fig3\].a), so that the internal state moves from $\mathbf s_I^{(i)}(t)=(m_{CH},n_{ATP},q_{EXT}^{(L)},q_{EXT}^{(H)})$ at time $t$ to $\mathbf s_I^{(i)}(t^+)=(m_{CH}+1,n_{ATP},q_{EXT},q_{EXT})$ at time instant $t^+$;
- from the neighboring cell $i-1$: the electron is collected in the HEEM, so that the corresponding state increases by one unit and $\mathbf s_I^{(i)}(t^+)=(m_{CH},n_{ATP},q_{EXT}^{(L)},q_{EXT}^{(H)}+1)$ (Fig. \[fig3\].b); note that this process is coupled with the *anaerobic ATP synthesis* (see definition below) process of the neighboring cell $i-1$ from which the electron is transferred; in fact, owing to the high transfer rate approximation, the LEEM of cell $i-1$ is shared with the HEEM of cell $i$, so that the rate of electron flow into the HEEM of cell $i$ is $\lambda_{EXT}^{(L)}(\mathbf s_I^{(i-1)}(t);\mathbf s_E^{(i-1)}(t))$;
- : this process involves the transfer of one electron from the IECP to the internal membrane to synthesize ATP, with rate $\mu_{CH}(\mathbf s_I^{(i)}(t),\mathbf s_E^{(i)}(t))$ \[electrons/s\]. Correspondingly, one molecule of ATP is generated; the electron then leaves the internal membrane and follows either the aerobic pathway (*i.e.*, it is captured by an internalized EA, such as Oxygen, see Fig. \[fig3\].c), with overall rate $\mu_{OUT}(\mathbf s_I^{(i)}(t);\mathbf s_E^{(i)}(t))$, or the anaerobic one (Fig. \[fig3\].d) and is collected in the LEEM, with overall rate $\lambda_{EXT}(\mathbf s_I^{(i)}(t);\mathbf s_E^{(i)}(t))$ (note that this is also the HEEM of cell $i+1$). If the aerobic pathway is followed, the new state becomes $\mathbf s_I^{(i)}(t^+)=(m_{CH}-1,n_{ATP}+1,q_{EXT}^{(L)},q_{EXT}^{(H)})$ (Fig. \[fig3\].c). Otherwise (anaerobic pathway), the new state becomes $\mathbf s_I^{(i)}(t^+)=(m_{CH}-1,n_{ATP}+1,q_{EXT}^{(L)}+1,q_{EXT}^{(H)})$ (Fig. \[fig3\].d);
- : this process involves the transfer of one electron from the HEEM to the internal membrane to synthesize ATP, with rate $\mu_{EXT}^{(H)}(\mathbf s_I^{(i)}(t),\mathbf s_E^{(i)}(t))$ \[electrons/s\]. Afterwards, the electron follows a similar path as in the conventional ATP synthesis, *i.e.*, either it is captured by an internalized EA (aerobic pathway), with overall rate $\mu_{OUT}(\mathbf s_I^{(i)}(t);\mathbf s_E^{(i)}(t))$, or it is collected in the LEEM of the cell (anaerobic pathway), with overall rate $\lambda_{EXT}^{(L)}(\mathbf s_I^{(i)}(t);\mathbf s_E^{(i)}(t))$. In the former case, the new state becomes $\mathbf s_I^{(i)}(t^+){=}(m_{CH},n_{ATP}{+}1,q_{EXT}^{(L)},q_{EXT}^{(H)}{-}1)$ (Fig. \[fig3\].e); in the latter, $\mathbf s_I^{(i)}(t^+){=}(m_{CH},n_{ATP}{+}1,q_{EXT}^{(L)}{+}1,q_{EXT}^{(H)}-1)$ (Fig. \[fig3\].f);
- : this process provides energy for cellular functions, and occurs with rate $\mu_{ATP}(\mathbf s_I^{(i)}(t);\mathbf s_E^{(i)}(t))$ \[electrons/s\]; when one molecule of ATP is consumed, the state $n_{ATP}^{(i)}(t)$ decreases by one unit, so that $\mathbf s_I^{(i)}(t^+)=(m_{CH},n_{ATP}-1,q_{EXT}^{(L)},q_{EXT}^{(H)})$ (Fig. \[fig3\].g);
- process, with rate $\delta(\mathbf s_I^{(i)}(t);\mathbf s_E^{(i)}(t))$: if death occurs, the new state becomes $\mathbf s_I^{(i)}(t^+)=\text{DEAD}$, from which the cell cannot recover any longer, *i.e.*, $\mathbf s_I^{(i)}(\tau)=~\text{DEAD},\ \forall\tau>~t$.
Flow Constraints {#constr}
----------------
Note that the rates of the different flows involved need to satisfy some constraints, induced by the queuing model employed. In particular, if some queue is empty (respectively, saturated), the rate of the corresponding outbound (respectively, inbound) flow must be zero, so that, for instance, for the flows out of and into the ATP pool, the following condition must hold: $$\begin{aligned}
&\mu_{ATP}(m_{CH},0,q_{EXT}^{(L)},q_{EXT}^{(H)};\mathbf s_E)=0, \ \text{(outbound flow)},\\
&\mu_{CH}(m_{CH},N_{AXP},q_{EXT}^{(L)},q_{EXT}^{(H)};\mathbf s_E)\\&
+\mu_{EXT}^{(H)}(m_{CH},N_{AXP},q_{EXT}^{(L)},q_{EXT}^{(H)};\mathbf s_E){=}0, \ \text{(inbound flow)}.\end{aligned}$$ A similar consideration holds for the other queues and the corresponding flows. Moreover, $$\begin{aligned}
&\mu_{EXT}^{(H)}(\mathbf s_I;\mathbf s_E){+}\mu_{CH}(\mathbf s_I;\mathbf s_E)
{=}
\lambda_{EXT}^{(L)}(\mathbf s_I;\mathbf s_E){+}\mu_{OUT}(\mathbf s_I;\mathbf s_E),\end{aligned}$$ since each electron leaving either the IECP or the HEEM to synthesize ATP either follows the aerobic pathway to the EA or the anaerobic one to the LEEM.
We further assume that $$\begin{aligned}
&\lambda_{CH}(\mathbf s_I;\sigma_D,\sigma_A)=\sigma_D\lambda_{CH}(\mathbf s_I;1,\sigma_A),
\\
&\mu_{OUT}(\mathbf s_I;\sigma_D,\sigma_A)=\sigma_A\mu_{OUT}(\mathbf s_I;\sigma_D,1),\end{aligned}$$ thus capturing the fact that the molecular diffusion rate is proportional to the ED (respectively, EA) concentration. This assumption is supported by Fick’s law of diffusion [@smith2003foundations], which states that the diffusion rate is linearly dependent on the concentration differential between inside and outside. It follows that, if no ED is present ($\sigma_D=0$), then $\lambda_{CH}(\mathbf s_I;0,\sigma_A)=0$ and no ED diffusion may occur. Similarly, if no EA is present ($\sigma_A=0$), then $\mu_{OUT}(\mathbf s_I;\sigma_D,0)=0$ and no EA diffusion may occur. In Sec. \[parammodel\], a parametric model for these flows is presented, based on which the model is fit to experimental data.
Isolated cell model {#isolatedcell}
===================
In the most general case, electron transport in a series of interconnected single-cell organisms is represented by the proposed stochastic model. However, this model can also explain the electron transport behavior of a single cell, which is the building block of the general multi-cell system. The experimental investigation of a multi-cellular network of bacteria is very challenging, in fact:
1. In order to build a chain of interconnected cells, single-cell organisms have to be placed in each other’s proximity. Placing multiple cells next or close to each other in a controlled way that maintains the intercellular contact is very difficult in practice and requires cellular manipulation techniques such as optical tweezers [@Liu2], as well as nanofabricated micron-scaled chambers designed specifically to hold these communities in place;
2. *In vivo* characterization of the energetic and electron transfer properties of an individual cell within this chain independently from the other cells requires complex chemical and optical assays that have never been used in such complicated systems.
Therefore, instead of the most general case of the model (multi-cell system), we start by investigating the properties of single, isolated cells. Using a few simplifying assumptions, the general model can be reduced to a single cell model which can be more easily matched against experimental results. In addition, the single-cell experiments are not hindered by the practical issues mentioned above, which makes them easier to perform. In this way, we can characterize the properties of the individual components, which will help us better understand the electron transport in multi-cellular systems.
In the case of an isolated cell, the IET process is not active, and $\lambda_{EXT}^{(H)}(t)=\mu_{EXT}^{(L)}(t)=0$. As a result, the HEEM gets depleted, and the LEEM gets filled. Therefore, after a transient phase, the cell reaches the configuration depicted in Fig. \[fig5\], where the HEEM is empty, and the LEEM is fully charged. In the following treatment, we assume that the transient phase is concluded, hence $q_{EXT}^{(L)}(t)=Q_{EXT}^{(L)}$ and $q_{EXT}^{(H)}(t)=0,\ \forall t$, so that the state $(q_{EXT}^{(L)}(t),q_{EXT}^{(H)}(t))=(Q_{EXT}^{(L)},0)$ of the external membrane can be neglected. Assuming that the cell operates in this configuration, we thus redefine its internal state as $\mathbf s_I(t)=(m_{CH}(t),n_{ATP}(t))$.
[MM]{} &
Ê The corresponding Markov chain and state transitions are depicted in Fig. \[fig5\]. From the continuous-time process described in Sec. \[stochmodel\], we now generate a discrete-time process, as detailed below. Initially, we assume that the external state $\mathbf s_E(t)$ is fixed, *i.e.*, $\sigma_D(t)=\sigma_D,\ \forall t$ and $\sigma_A(t)=\sigma_A,\ \forall t$. The case where $\mathbf s_E(t)$ is piecewise constant will be considered in Sec. \[senonconstant\]. The discretization is obtained by sampling the state process $\mathbf s_I(t)$ at specific times, corresponding to one of the events described in Sec. \[stochmodel\], specialized to the case of an isolated cell: molecular diffusion; conventional aerobic ATP synthesis; ATP consumption; death. Starting from time $t=0$ in state $\mathbf s_I(0)\in\mathcal S_I$, we define $T_k$ as the time instant corresponding to the occurrence of the $k$th event since time $0$, and $\mathbf S_k$ as the corresponding state at time instant $T_k^+$ (*i.e.*, right after the corresponding transition occurs). In particular, $T_0=0$ and $\mathbf S_0=\mathbf s_I(0)$. Note that, by sampling, we have transformed the continuous-time stochastic process into a discrete-time Markov chain, with finite state space $\mathcal S_I$. However, the duration of the $k$th time-slot, $T_{k+1}-T_{k}$, is not fixed but is a random variable which depends on the inter-arrival time of the events described in Sec. \[senonconstant\]. In the subsequent sections, we first derive the transition probabilities of the underlying discrete time Markov chain and the inter-arrival times of the events, thus leading to a full-characterization of the stochastic dynamics of $\mathbf s_I(t)$. We then provide an example of applicability of this framework to the computation of the lifetime of the cell. Finally, in Sec. \[estimation\], we present a parameter estimation framework and match the model to experimental data available in [@Ozalp].
Transition Probabilities and inter-arrival times {#TXprob}
------------------------------------------------
In this section, we compute the transition probability of the underlying discrete-time Markov chain, and the distribution of the inter-arrival times in the corresponding continuous time system. To this end, let $\mathbf S_k=i\footnote{In this section, $i$ is an index corresponding to a specific state in $\mathcal S_I$.}\in\mathcal S_I\setminus\{\text{DEAD}\}$ be the state of the cell at time $T_k^+$. We compute the transition probability $$\begin{aligned}
\label{prob}
\mathbb P(\mathbf S_{k+1}=j,T_{k+1}>\tau |\mathbf S_k=i,T_k=t),\end{aligned}$$ for some $j\in\mathcal S_I$, $\tau\geq t$ (note that, due to the memoryless property of Poisson processes, the event $\mathbf S_{k+1}=j,T_{k+1}>\tau $ conditioned on $\mathbf S_k=i,T_k=t$ is independent of the realization of $\{(\mathbf S_j,T_j),0\leq j<k\}$). Let $\lambda_{i,j}$ be the transition rate from state $i$ to state $j$, which depends on the specific event which triggers the transition. For instance, if $i$ corresponds to $(m_{CH},n_{ATP})$ and $j$ to $(m_{CH},n_{ATP}-1)$, then a transition from state $i$ to state $j$ occurs if the ATP consumption event occurs, with rate $\lambda_{i,j}=\mu_{ATP}(\mathbf s_I;\mathbf s_E)$. The transition from state $i$ to state $j$ can be interpreted as follows. Let $E_{i,s}$ be the event which triggers the transition from $i$ to $s$, and $t+W_{i,s}$ be the time when such event occurs (with respect to the reference time-position $t$). From the properties of Poisson processes, we have that $W_{i,s}$ is an exponential random variable, with pdf $f_{W_{i,s}}(w)=\lambda_{i,s}e^{-\lambda_{i,s}w}$, and that $\{W_{i,s},\forall s\}$ are mutually independent. Then, the system moves to state $j$ if $t+W_{i,j}<t+W_{i,s},\ \forall s\neq j$, *i.e.*, the event $E_{i,j}$ is the first one to occur, which thus triggers the transition. Therefore, the probability (\[prob\]) is equivalent to $$\begin{aligned}
\label{p1}
&\mathbb P(\mathbf S_{k+1}=j,T_{k+1}>\tau |\mathbf S_k=i,T_k=t)
\\&
=
\mathbb P(t+W_{i,j}>\tau,W_{i,j}<W_{i,s},\ \forall s\neq j |\mathbf S_k=i,T_k=t)
\nonumber\\&
=
\int_{\tau-t}^\infty
\lambda_{i,j}e^{-\lambda_{i,j}w}
\prod_{s\neq j}\mathbb P(W_{i,s}>w)\mathrm dw
=
\frac{\lambda_{i,j}}{R_i}e^{-R_i(\tau-t)},{\nonumber}\end{aligned}$$ where we have defined the total flow from state $i$, $R_i=\sum_{s}\lambda_{i,s}$, we have marginalized with respect to $W_{i,j}$, we have used the independence among $\{W_{i,s},\forall s\}$ and $\mathbb P(W_{i,s}>w)=e^{-\lambda_{i,s}w}$. From (\[p1\]), we thus obtain the transition probability $\mathbb P(\mathbf S_{k+1}=j|\mathbf S_k=i)$ by letting $\tau=t$ in (\[p1\]) and by noticing that the resulting expression is independent of $t$, *i.e.*, $$\begin{aligned}
&\mathbb P(\mathbf S_{k+1}=j|\mathbf S_k=i,T_k=t)=
\frac{\lambda_{i,j}}{R_i}
=\mathbb P(\mathbf S_{k+1}=j|\mathbf S_k=i).\end{aligned}$$ We now compute the distribution of the inter-arrival time $T_{k+1}-T_{k}$ as $$\begin{aligned}
&\mathbb P(T_{k+1}-T_k>\tau-t |\mathbf S_k=i,\mathbf S_{k+1}=j,T_k=t)
{\nonumber}\\&=
\frac{\mathbb P(\mathbf S_{k+1}=j,T_{k+1}>\tau |\mathbf S_k=i,T_k=t)}
{\mathbb P(\mathbf S_{k+1}=j|\mathbf S_k=i,T_k=t)}
=e^{-R_i(\tau-t)}.\end{aligned}$$ Note that the resulting expression is independent of $\mathbf S_{k+1}$ and of time $t$, since the process is stationary. We can thus write $$\begin{aligned}
&\mathbb P(T_{k+1}-T_k>\tau-t |\mathbf S_k=i)=e^{-R_i(\tau-t)}.\end{aligned}$$
We define the $(|\mathcal S_I|-1)\times (|\mathcal S_I|-1)$ transition probability matrix $\mathbf T$ of the underlying discrete-time Markov chain within $\mathcal S_I\setminus\{\text{DEAD}\}$, with entries $\mathbf T(i,j)=\mathbb P(\mathbf S_{k+1}=j|\mathbf S_k=i),\ i,j\in\mathcal S_I\setminus\{\text{DEAD}\}$ (we do not consider transitions from $\text{DEAD}$, since this is absorbing). The transition probability from $i\in\mathcal S_I\setminus\{\text{DEAD}\}$ to $\text{DEAD}$ is then given by $1-\mathbf e_i^T\mathbf T\mathbf 1$, where $\mathbf 1$ is the column vector of all ones, and $\mathbf e_i$ equals $1$ in the position corresponding to state $i$, and zero otherwise.
State distribution of the system at time $t>0$
----------------------------------------------
Given the analysis of the underlying discrete-time Markov chain and of the inter-arrival times in the previous section, we are now able to compute the state distribution of the system at a generic time $t$, given that $\mathbf S_I(0)=i$. We define $$\begin{aligned}
\mathbb P_t(j|i)=\mathbb P(\mathbf S_I(t)=j|\mathbf S_I(0)=i),\ j\in\mathcal S_I\setminus\{\text{DEAD}\}.
\end{aligned}$$ In order to compute it, let $0<h<t$. By the memoryless property of Poisson processes, $$\begin{aligned}
\mathbb P_t(j|i)&=\!\!\!\!\!\sum_{s\in\mathcal S_I\setminus\{\text{DEAD}\}}\mathbb P(\mathbf S_I(t)=j,\mathbf S_I(t-h)=s|\mathbf S_I(0)=i)
{\nonumber}\\&
=\!\!\!\!\!
\sum_{s\in\mathcal S_I\setminus\{\text{DEAD}\}}\mathbb P_h(j|s)\mathbb P_{t-h}(s|i).
\end{aligned}$$ It follows that, $\forall i,j\in\mathcal S_I\setminus\{\text{DEAD}\}$, $$\begin{aligned}
\mathbb P_t(j|i)-\mathbb P_{t-h}(j|i)
=\!\!\!\!\!
\sum_{s\in\mathcal S_I\setminus\{\text{DEAD}\}}(\mathbb P_h(j|s)-\delta_{j,s})\mathbb P_{t-h}(s|i).
\end{aligned}$$ Then, dividing by $h$ and taking the limit for $h\to 0$, we obtain $$\begin{aligned}
\label{p2}
\frac{\mathrm d\mathbb P_{t}(j|i)}{\mathrm d t}
=
\sum_{s\in\mathcal S_I\setminus\{\text{DEAD}\}}\lim_{h\to 0}\frac{\mathbb P_h(j|s)-\delta_{j,s}}{h}
\mathbb P_{t}(s|i).
\end{aligned}$$ Note that $\underset{h\to 0}\lim\frac{\mathbb P_h(j|s)-\delta_{j,s}}{h}\!=\!\lambda_{s,j}$, and $\underset{h\to 0}\lim\frac{\mathbb P_h(s|s)-\delta_{s,s}}{h}\!=\!-R_s$. Substituting in (\[p2\]), we obtain the system of differential equations $$\begin{aligned}
\label{p3}
\frac{\mathrm d\mathbb P_{t}(j|i)}{\mathrm d t}
=
\!\!\!\!\!\sum_{s\in\mathcal S_I\setminus\{\text{DEAD},j\}}\!\!\!\!\!\lambda_{s,j}\mathbb P_{t}(s|i)
-R_j\mathbb P_{t}(j|i),\quad\forall i,j.
\end{aligned}$$ Letting $\mathbf P_{t}$ be the $(|\mathcal S_I|-1)\times(|\mathcal S_I|-1)$ matrix with components $\mathbf P_{t}(i,j)=\mathbb P_{t}(j|i),\ i,j\in\mathcal S_I\setminus\{\text{DEAD}\}$, we can rewrite the system of differential equations (\[p3\]) as $$\begin{aligned}
\label{p4}
\mathbf P_{t}^\prime
=
\mathbf P_{t}\mathbf A,
\end{aligned}$$ where we have defined the *flow matrix* $\mathbf A$ with components $\mathbf A(s,j)=\lambda_{s,j}$ for $j\neq s$ and $\mathbf A(j,j)=-R_j$, and $\mathbf P_{t}^\prime$ represents the first-order derivative of $\mathbf P_{t}$ with respect to time. Note that $\mathbf A=\mathbf R(\mathbf T-\mathbf I)$, where $\mathbf T$ is the transition matrix of the underlying discrete-time Markov chain within $\mathcal S_I\setminus\{\text{DEAD}\}$, derived in the previous section, $\mathbf R$ is the *rate matrix*, a diagonal matrix with entries $\mathbf R(i,i)=R_i$, and $\mathbf I$ is the unit matrix. Moreover, by Gershgorin’s circle Theorem [@Golub], all eigenvalues of $\mathbf A$ are non-positive. The general solution to (\[p4\]) subject to $\mathbf P_0=\mathbf I$ is $$\begin{aligned}
\label{gensol}
\mathbf P_t=\exp\{\mathbf At\},\end{aligned}$$ where we have defined the matrix exponential $\exp\{\mathbf At\}=\sum_{k=0}^{\infty}\frac{t^k}{k!}\mathbf A^k$. Note that such solution guarantees a feasible transition probability matrix, *i.e.*, $[\mathbf P_t]_{i,j}\geq 0$ and $\sum_{j}[\mathbf P_t]_{i,j}\leq~1$.
Numerical evaluation of $\mathbf P_t$
-------------------------------------
Unfortunately, from our numerical evaluations, we have verified that $\mathbf A$ can seldom be diagonalized. Therefore, we employ an alternative solution to efficiently compute $\mathbf P_t$. Let $\Delta\ll 1$ and $n=\lceil t/\Delta\rceil$. Then, the general solution can be approximated as $$\begin{aligned}
\label{approx}
\!\!\!\!\mathbf P_t\!=\!\left[\exp\{\mathbf A\Delta\}\right]^{n}\!\exp\{\mathbf A(t\!-\!\Delta n)\}
\!\simeq\!\left[\exp\{\mathbf A\Delta\}\right]^{n}\!=\!\mathbf P_\Delta^{n},\!\!\end{aligned}$$ where we have used the approximation $\exp\{\mathbf A(t-\Delta n)\}\simeq\mathbf I$, which holds for $\Delta\ll 1$. Moreover, since we assume $\Delta\ll 1$, we approximate the matrix exponential $\mathbf P_\Delta=\exp\{\mathbf A\Delta\}$ with the first order Taylor approximation $$\begin{aligned}
\label{pdelta}
\mathbf P_\Delta\simeq\mathbf I+\Delta\mathbf A=
\mathbf I-\Delta\mathbf R(\mathbf I-\mathbf T)
\triangleq \tilde{\mathbf P}_\Delta.\end{aligned}$$ Note that the approximation $\tilde{\mathbf P}_\Delta$ of $\mathbf P_\Delta$ is a feasible transition matrix with non-negative entries, if $\Delta<\min_i\{1/R_i\}$.
Extension to $\mathbf s_E(t)$ piecewise constant {#senonconstant}
------------------------------------------------
In this section, we extend the previous analysis to the case where the external ambient state is piecewise constant, *i.e.*, $\mathbf s_E(t)=\mathbf s_{E,n},\ \forall n\in [\tau_n,\tau_{n+1}),\ \forall n\geq 0$, where $0=\tau_0$ and $\tau_n<\tau_{n+1},\ \forall n\geq 0$. This analysis is of interest for the following experimental evaluation: the ED concentration is varied in order to measure the response in terms of fluctuations in the ATP level within the cell.
For this case, it is straightforward to derive the probability of the cell being in state $\mathbf s_I(t)=j\in\mathcal S_I\setminus\{\text{DEAD}\}$ at time $t\in [\tau_n,\tau_{n+1})$, for some $n\geq 0$, given $\mathbf s_I(0)=i\in\mathcal S_I\setminus\{\text{DEAD}\}$. To this end, let $\mathbf T_n$ be the transition probability matrix within $\mathcal S_I\setminus\{\text{DEAD}\}$, $\mathbf R_n$ be the rate matrix, and $\mathbf A_n=\mathbf R_n(\mathbf T_n-\mathbf I)$ be the flow matrix when $\mathbf s_E(t)=\mathbf s_{E,n}$. Then, $\forall t\in [\tau_n,\tau_{n+1})$ we have $$\begin{aligned}
\mathbf P_t=\left[\prod_{m=0}^{n-1}\exp\{\mathbf A_m(\tau_{m+1}-\tau_m)\}\right]
\times \exp\{\mathbf A_n(t-\tau_n)\}.\end{aligned}$$ where we have defined $\prod_{m=0}^{n-1}\mathbf C_m=\mathbf C_0\times \mathbf C_1\times\cdots\times\mathbf C_{n-1}$, and we have used the fact that, from the Markov property, $$\begin{aligned}
&\mathbb P(\mathbf s_I(t)=j|\mathbf s_I(0)=i)
=\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
\sum_{s_0,s_1,\dots,s_{n}\in\mathcal S_I\setminus\{\text{DEAD}\}}\!\!\!\!\!\!\!\!\!\!
\mathbb P(\mathbf s_I(t)=j|\mathbf s_I(\tau_n)=s_n)
\\&\times
\prod_{m=0}^{n-1}
\mathbb P(\mathbf s_I(\tau_{m+1})=s_{m+1}|\mathbf s_I(\tau_{m})=s_{m}),\end{aligned}$$ and, since $\mathbf s_E(\tau)$ is constant in the time interval $[\tau_m,\tau_{m+1})$, the probability $\mathbb P(\mathbf s_I(\tau_{m+1})=s_{m+1}|\mathbf s_I(\tau_{m})=s_{m})$ can be computed as in Sec. \[TXprob\].
Application to cell-lifetime computation, Isolated cell {#celllifetime}
=======================================================
For every cell in the bacterial chain, it is possible that, at some point in time, due to variations in the energetic state of the cell and changes to the supply of the ED and the EA, the cell reaches a state where its ATP consumption rate reaches a minimum value (*e.g.*, zero). Once a cell enters this state, it is considered dead and its ATP consumption rate may not restore to normal values, thus jeopardizing the overall functionality of the cable. Accordingly, the time it takes for a cell to reach this irreversible state is defined as the *lifetime of the cell*. This quantity can be measured experimentally by using indicators of cellular respiratory activity. In an experimental setup where cells in a bacterial chain can be characterized on an individual basis, cellular lifetime is one of the easiest measurable quantities that contains a significant amount of information regarding the specific properties of the target cell. In this section, we apply the stochastic model presented in Sec. \[isolatedcell\] to the computation of the lifetime of an isolated cell, defined as follows.
The lifetime of the cell, $L$, is defined as $$\begin{aligned}
L=\min\{t>0:\mathbf S_I(t)=\text{\em{DEAD}}\}.
\end{aligned}$$
Equivalently, letting $k^*=~\min\{k>0:\mathbf S_k=\text{DEAD}\}$, we have $L=T_{k^*}$. In this section, we compute the probability density function (pdf) of the lifetime, $f_L(t;\boldsymbol{\pi}_0)$, as well as the expected lifetime $\mathbb E[L|\boldsymbol{\pi}_0]$, given some initial state distribution $\boldsymbol{\pi}_0(i),i\in\mathcal S_I\setminus\{\text{DEAD}\}$. $f_L(t;\boldsymbol{\pi}_0)$ is given by (we use $\mathbb P$ to denote also a pdf) $$\begin{aligned}
&f_L(t;\boldsymbol{\pi}_0)
=\mathbb P(L=t|\boldsymbol{\pi}_0)
\\&
=\sum_{k=0}^{\infty}\mathbb P(L=t,\text{Death occurs at the $(k+1)$th event}|\boldsymbol{\pi}_0).
{\nonumber}\end{aligned}$$ Note that the event $(L=t,$Death occurs at the $(k+1)$th event$)$ is equivalent to $$\begin{aligned}
\mathbf S_k\in\mathcal S_I\setminus\{\text{DEAD}\},\mathbf S_{k+1}=\text{DEAD},T_{k+1}=t,
\end{aligned}$$ *i.e.*, the cell is alive upon occurrence of the $k$th event, and dies upon occurrence of the $(k+1)$th event. Therefore, we obtain $$\begin{aligned}
\label{lifetime}
f_L(t;\boldsymbol{\pi}_0)=
\sum_{k=0}^{\infty}\sum_{i\in\mathcal S_I\setminus\{\text{DEAD}\}}g_k(i,t),
\end{aligned}$$ where we have defined $g_k(i,t)\triangleq\mathbb P(\mathbf S_k=i,\mathbf S_{k+1}=\text{DEAD},T_{k+1}=t|\boldsymbol{\pi}_0)$. In order to compute $g_k(i,t)$, we first determine, for $k\geq 0$ and $s\in\mathcal S_I\setminus\{\text{DEAD}\}$, $$\begin{aligned}
&h_k(s,t)\triangleq \mathbb P(\mathbf S_k=s,T_k=t|\boldsymbol{\pi}_0).
\end{aligned}$$ For $k=0$, this is given by $h_0(s,t)=\boldsymbol{\pi}_0(s)\delta(t)$, where $\delta(t)$ is the Kronecker delta function. For $k>0$, we have $$\begin{aligned}
&h_k(s,t)\!=\!
\sum_{j}\!\int_0^t \mathbb P(\mathbf S_k=s,T_k=t,\mathbf S_{k-1}\!=\!j,T_{k-1}\!=\tau|\boldsymbol{\pi}_0)\mathrm d\tau
\nonumber\\&
=
\sum_{j} \int_0^t \mathbb P(T_k-T_{k-1}=t-\tau|\mathbf S_{k-1}=j)
\nonumber\\&
\qquad\times
\mathbb P(\mathbf S_k=s|\mathbf S_{k-1}=j,T_{k-1}=\tau)
h_{k-1}(j,\tau)\mathrm d\tau
\nonumber\\&
=
\sum_{j} \int_0^t R_j e^{-R_j(t-\tau)}
\mathbf T(j,s)
h_{k-1}(j,\tau)\mathrm d\tau.
\end{aligned}$$ It follows that $$\begin{aligned}
&g_k(s,t)=\mathbb P(\mathbf S_k=s,\mathbf S_{k+1}=\text{DEAD},T_{k+1}=t)
{\nonumber}\\&
=
\int_0^t \mathbb P(\mathbf S_k=s,\mathbf S_{k+1}=\text{DEAD},T_{k+1}=t,T_k=\tau)\mathrm d\tau
\nonumber
\\&
=
\int_0^t R_s e^{-R_s(t-\tau)}\left(1-\sum_j\mathbf T(s,j)\right)h_k(s,\tau)
\mathrm d\tau,
\end{aligned}$$ where $1-\sum_j\mathbf T(s,j)$ is the transition probability to state $\text{DEAD}$, from state $s$. Substituting in (\[lifetime\]), we obtain $$\begin{aligned}
& f_L(t;\boldsymbol{\pi}_0)
\!=\!\sum_{s}\!\!\int_0^t R_s e^{-R_s(t-\tau)}\!\!\left(1-\sum_j\mathbf T(s,j)\right)H(s,\tau)
\mathrm d\tau,
\end{aligned}$$ where we have defined $$\begin{aligned}
&H(s,t)\triangleq\sum_{k=0}^{\infty}f_k(s,t)
\\&
=h_0(s,t)+
\sum_{j} \int_0^t R_j e^{-R_j(t-\tau)}\mathbf T(j,s)H(j,\tau)\mathrm d\tau
\nonumber\\&
=
\boldsymbol{\pi}_0(s)\delta(t)+
\sum_{j} \int_0^t R_j e^{-R_j(t-\tau)}\mathbf T(j,s)H(j,\tau)\mathrm d\tau.{\nonumber}\end{aligned}$$ Then, we obtain $$\begin{aligned}
& \mathbb E[L]
\!=\!\int_0^{\infty}\!\!\!\!t\sum_{s}\int_0^t\!\!R_s e^{-R_s(t-\tau)}\!\!\left(\!\!1\!-\!\sum_j\mathbf T(s,j)\!\!\right)H(s,\tau)
\mathrm d\tau\mathrm dt
\nonumber\\&
\!=\!\sum_{s}\!\int_0^{\infty}\!\!\!
R_s e^{R_s\tau}\!\!\left(1-\sum_j\mathbf T(s,j)\!\!\right)\!\!H(s,\tau)
\int_\tau^\infty te^{-R_st}\mathrm d t
\mathrm d\tau.
\end{aligned}$$ Using the fact that $ \int_\tau^\infty te^{-R_st}\mathrm d t
{=}
\frac{e^{-R_s\tau}}{R_s}\left(\tau{+}\frac{1}{R_s}\right)
$, we obtain $$\begin{aligned}
& \mathbb E[L]
=
\sum_{s}\left(1-\sum_j\mathbf T(s,j)\right)\int_0^{\infty}\tau H(s,\tau)\mathrm d\tau
{\nonumber}\\&
\quad+\frac{1}{R_s}\sum_{s}\left(1-\sum_j\mathbf T(s,j)\right)\int_0^{\infty}H(s,\tau)\mathrm d\tau
\nonumber\\&
=
\sum_{s}\left(1-\sum_j\mathbf T(s,j)\right)Q(s)
{\nonumber}\\&
\quad+\sum_{s}\frac{1-\sum_j\mathbf T(s,j)}{R_s}\sum_{k=0}^\infty \mathbb P(S_k=s|\boldsymbol{\pi}_0),
\end{aligned}$$ where we have defined $Q(s)\triangleq\int_0^{\infty}\tau H(s,\tau)\mathrm d\tau$. This term can be computed as $$\begin{aligned}
&Q(s)=\int_0^{\infty}t H(s,t)\mathrm dt
{\nonumber}\\&
=
\sum_{j} R_j\mathbf T(j,s)
\int_0^\infty e^{R_j\tau} H(j,\tau)
\int_\tau^\infty t e^{-R_jt}\mathrm dt
\mathrm d\tau{\nonumber}\\&=\nonumber
\sum_{j}\mathbf T(j,s)
\int_0^\infty H(j,\tau)\left(\tau+\frac{1}{R_j}\right)
\mathrm d\tau
{\nonumber}\\&
=
\sum_{j}\mathbf T(j,s)Q(j)
+
\sum_{j}\frac{\mathbf T(j,s)}{R_j}\sum_{k=0}^\infty \mathbb P(S_k=j|\boldsymbol{\pi}_0).\end{aligned}$$ Let $\mathbf Q$ be a row vector with elements $Q(j)$, and $\mathbf x{=}(\mathbf I{-}\mathbf T)\mathbf 1$ be the column vector associated to transitions from the transient states to the DEAD state. We obtain $$\begin{aligned}
Q(s)=\mathbf Q\mathbf e_s
=
\mathbf Q\mathbf T\mathbf e_s
+
\boldsymbol{\pi}_0^T(\mathbf I-\mathbf T)^{-1}\mathbf R^{-1}\mathbf T\mathbf e_s,\end{aligned}$$ where we have used the fact that $
\sum_{k=0}^\infty \mathbb P(S_k=s|\boldsymbol{\pi}_0)
=
\boldsymbol{\pi}_0^T(\mathbf I-\mathbf T)^{-1}\mathbf e_s
$. Therefore, we obtain $$\begin{aligned}
\mathbf Q
=
\boldsymbol{\pi}_0^T(\mathbf I-\mathbf T)^{-1}\mathbf R^{-1}\mathbf T(\mathbf I-\mathbf T)^{-1}.\end{aligned}$$ Substituting in the expression of the expected lifetime, we obtain $$\begin{aligned}
& \mathbb E[ L]
=
\boldsymbol{\pi}_0^T(\mathbf I-\mathbf T)^{-1}\mathbf R^{-1}(\mathbf I-\mathbf T)^{-1}\mathbf x.
\end{aligned}$$ Finally, we use the fact that $\mathbf x=(\mathbf I-\mathbf T)\mathbf 1$, yielding the expression of the expected lifetime $$\begin{aligned}
& \mathbb E[ L]
=
\boldsymbol{\pi}_0^T(\mathbf I-\mathbf T)^{-1}\mathbf R^{-1}\mathbf 1.
\end{aligned}$$
Parameter estimation and Experimental validation {#estimation}
================================================
As an example of experiments related to our stochastic model, Ozalp et al. [@Ozalp] have measured *in vivo* levels of ATP and NADH in the yeast *Saccharomyces cerevisiae*, as they abruptly add ED to a suspension of starved yeast cells. Since we have theoretically investigated the single cell system, this work, which is performed on a culture of mutually-independent yeast cells, can be used as a test for our stochastic model. Although the ETC in yeast is not exactly identical to the bacterial counterpart, the principles on which the model is based upon are conserved between yeast and bacteria. These include the involvement of an ED, electron carriers such as NADH, ETC and an EA, which in the case of yeast is molecular Oxygen.
We have extracted the measured quantities from [@Ozalp] in the form of ATP and NADH concentrations as a function of time. In accordance with [@Ozalp], we have assumed that yeast cells are initially starved and, at some point in time, the ED is added to the cell suspension. This triggers an increase in ATP and NADH production as well as ATP consumption. In extracting the data, we have averaged out the small oscillations in NADH and ATP concentrations in time, since these are mainly caused by an enzyme involved in the metabolic pathway that is specific to yeast and does not exist in the bacterial strain that we are interested in, *Shewanella oneidensis* MR-1. Therefore, in matching the experimental data from yeast to our model, we have only taken into account the large scale variations of the levels of ATP and NADH over time.
Let $\{(\mathbf s_{I,k}^{i},t_k),\ k=0,1,\dots,N\}$ be the time-series of the state of cell $i$ at times $t_k$, where $0=t_0<t_1<\dots<t_N$. Let $\mathbf s_E(t)$ be the known profile of the concentration of the external ED and EA, which we assume to be piecewise constant, as in Sec. \[senonconstant\], and the same for all cells. In particular, we assume that $\mathbf s_E(t)=\mathbf s_{E,k},\ \forall t\in [t_k,t_{k+1}),\ \forall k=0,1,\dots, N-1$, so that the external state is constant in the time interval between two consecutive measurements. The measurement collected in [@Ozalp] at time $t_k$ is $$\begin{aligned}
\label{ts}
& \text{NADH}_k=\alpha_{NADH}\frac{1}{M}\sum_{i=1}^M m_{CH}^{i}(t_k)+\tilde{w}_k^{(NADH)},
\nonumber\\
& \text{ATP}_k=\alpha_{ATP}\frac{1}{M}\sum_{i=1}^M n_{ATP}^{i}(t_k)+\tilde{w}_k^{(ATP)},\end{aligned}$$ where $\text{NADH}_k$ is the measurement of NADH (typically, fluorescence level [@Ozalp]), whereas $\text{ATP}_k$ is the measurement of ATP (typically, in ${\,\,\,\mathrm{mM}}$ [@Ozalp]); the constants $\alpha_{NADH}$ and $\alpha_{ATP}$ account for the conversion in the unit of measurements of NADH and ATP, respectively, from the stochastic model presented in this paper (electron units) to the experimental setup (fluorescence level and ${\,\,\,\mathrm{mM}}$, respectively); and $\tilde{w}_k^{(NADH)}$ and $\tilde{w}_k^{(ATP)}$ are zero mean Gaussian noise samples, each i.i.d. over time, with variance $\sigma_{NADH}^2$ and $\sigma_{ATP}^2$, respectively. A practical assumption is that $M\gg~1$, so that $\frac{1}{M}\sum_{i=1}^M m_{CH}^{i}(t_k)\simeq\mathbb E[m_{CH}(t_k)]$ and $\frac{1}{M}\sum_{i=1}^M n_{ATP}^{i}(t_k)\simeq\mathbb E[n_{ATP}(t_k)]$, where the expectation is computed with respect to the state distribution at time $t_k$, given by $\boldsymbol{\pi}_0^T\mathbf P_{t_k}$. Letting $\mathbf y_k=[\alpha_{NADH}^{-1}\text{NADH}_k,\alpha_{ATP}^{-1}\text{ATP}_k]$, $\mathbf w_k=[\alpha_{NADH}^{-1}\tilde{w}_k^{(NADH)},\alpha_{ATP}^{-1}\tilde{w}_k^{(ATP)}]$ and $\mathbf Z\in\mathbb R^{(|\mathcal S_I|-1)\times 2}$ with $j$th row $[\mathbf Z]_{j,:}=[m_{CH}(j),n_{ATP}(j)]$, where $m_{CH}(j)$ and $n_{ATP}(j)$ are the NADH and ATP levels in the state corresponding to index $j$, we thus obtain $$\begin{aligned}
\label{obsmodel}
\mathbf y_k=\boldsymbol{\pi}_0^T\mathbf P_{t_k}\mathbf Z+\mathbf w_k\simeq
\boldsymbol{\pi}_0^T\prod_{j=1}^{k}\mathbf P_{\Delta,j-1}^{n_j-n_{j-1}}\mathbf Z+\mathbf w_k,\end{aligned}$$ where $\prod_{j=1}^{k}\mathbf P_{\Delta,j-1}^{n_j-n_{j-1}}
=\mathbf P_{\Delta,0}^{n_1}\times
\mathbf P_{\Delta,1}^{n_2-n_1}
\times\dots\times
\mathbf P_{\Delta}^{n_k-n_{k-1}}$, $n_j=\lceil t_j/\Delta\rceil$, with $n_0=0$, and $\mathbf P_{\Delta,j-1}$ is the transition matrix with time-step size $\Delta$, when the external state takes value $\mathbf s_{E,j-1}$. In the last step, we have used the approximation (\[approx\]). Herein, we assume that $\mathbf w_k\sim\mathcal N(\mathbf 0,\sigma_w^2\mathbf I_2)$, *i.e.*, $\alpha_{NADH}^{-2}\sigma_{NADH}^2=\alpha_{ATP}^{-2}\sigma_{ATP}^2=\sigma_w^2$.
Parametric model {#parammodel}
----------------
The statistics of the system, defined by the transition probability matrix $\mathbf P_{t}$, is determined by the rates $\lambda_{CH}$, $\mu_{CH}$ and $\mu_{ATP}$. In this section, we present a parametric model for these rates, based on biological constraints. Specifically, we let $$\begin{aligned}
\label{flows}
\!\!\!\!\!\!\begin{array}{l}
\lambda_{CH}(\mathbf s_I(t);\mathbf s_E(t))=\gamma\sigma_{D}(t)+\rho\left(1-\frac{m_{CH}(t)}{M_{CH}}\right)\sigma_{D}(t),\\
\mu_{CH}(\mathbf s_I(t);\mathbf s_E(t))=\zeta\left(1-\frac{n_{ATP}(t)}{N_{AXP}}\right),\\
\mu_{ATP}(\mathbf s_I(t);\mathbf s_E(t))=\beta\sigma_{D}(t),
\end{array}\!\!\end{aligned}$$ where $\gamma,\rho,\zeta,\beta\in\mathbb R_+$ are parameters, that we want to estimate, and $\mathbb R_+$ is the set of non-negative reals. The NADH generation rate $\lambda_{CH}$ primarily depends on the concentration of available ED, as explained in Sec. \[constr\]. Additionally, it depends on the number of available NAD molecules in the cell, since the ED reacts with NAD to form NADH. The more NAD molecules are available, the higher the rate of the NADH-producing reaction. Moreover, the larger the ATP level, the smaller the ATP generation rate $\mu_{CH}(\mathbf s_I(t);\mathbf s_E(t))$. This is true because the ATP synthase, the protein responsible for ATP production, transforms ADP into ATP. Since the sum of ATP and ADP molecules in the cell is conserved, a higher ATP level corresponds to a lower ADP level. Therefore, as there is more ATP available in the cell, there are less ADP molecules available for ATP synthase to produce additional ATP molecules, which, in turn, results in a smaller ATP production rate. Finally, the larger the ED concentration, the larger the ATP consumption rate $\mu_{ATP}(\mathbf s_I(t);\mathbf s_E(t))$. This is shown to be true experimentally, for instance in [@Ozalp]. The reason behind this correlation is that cellular operations that consume ATP (*e.g.*, ATP-ases) are directly regulated by the ED concentration. Note that $\lambda_{CH}$, $\mu_{CH}$ and $\mu_{ATP}$ further need to satisfy the constraints listed in Sec. \[constr\]. We assume that all the cells are alive throughout the experiment, and set the death rate $\delta(\mathbf s_I(t);\mathbf s_E(t))=0$.
We define the parameter vector $\mathbf x=[\gamma,\rho,\zeta,\beta]$, which is estimated via maximum likelihood (ML) in the next section. Therefore, the flow matrix $\mathbf A$, defined in (\[p4\]), is a linear function of the entries of $\mathbf x$. We write such a dependence as $\mathbf A(\mathbf x,\mathbf s_{E})$. Similarly, from (\[pdelta\]), we write $\mathbf P_\Delta(\mathbf x,\mathbf s_E)=\mathbf I+\Delta\mathbf A(\mathbf x,\mathbf s_E)$.
Maximum Likelihood estimate of $\mathbf x$
------------------------------------------
For a given time series $\{(\mathbf y_k,t_k),\ k=0,1,\dots,N\}$, and the piecewise constant profile of the external state $\mathbf s_E(t)$, in this section we design a ML estimator of $\mathbf x$. Since the initial distribution $\boldsymbol{\pi}_0$ is unknown, we also estimate it jointly with $\mathbf x$. Note that, since the death rate is zero, the entries of $\boldsymbol{\pi}_0$ need to sum to one, *i.e.*, $\mathbf 1^T\boldsymbol{\pi}_0=1$. Moreover, we further enforce the constraints $\boldsymbol{\pi}_0^T\mathbf Z=\mathbf y_0$, *i.e.*, the expected values of the NADH and ATP pools at time $t_0$ equal the measurement $\mathbf y_0$. Therefore, we have the linear equality constraint $\boldsymbol{\pi}_0^T[\mathbf Z,\mathbf 1]=[\mathbf y_0,1]$, and the inequality constraint $\boldsymbol{\pi}_0\geq 0$ (component-wise). We denote the constraint set as $\mathcal P$, so that $\boldsymbol{\pi}_0\in\mathcal P$. Due to the Gaussian observation model (\[obsmodel\]), the ML estimate of $(\mathbf x,\boldsymbol{\pi}_0)$ is given by $$\begin{aligned}
\label{ML}
(\hat{\mathbf x},\hat{\boldsymbol{\pi}}_0)=\underset{\mathbf x\geq 0,\boldsymbol{\pi}_0\in\mathcal P}{\arg\min} f(\mathbf x,\boldsymbol{\pi}_0),
\end{aligned}$$ where we have defined the negative log-likelihood cost function $$\begin{aligned}
f(\mathbf x,\boldsymbol{\pi}_0)\triangleq
\frac{1}{2}\sum_{k=0}^N \left\|\mathbf y_k-
\boldsymbol{\pi}_0^T
\prod_{j=1}^{k}\mathbf P_{\Delta}(\mathbf x,\mathbf s_{E,j-1})^{n_j-n_{j-1}}
\mathbf Z\right\|_F^2\!\!.\end{aligned}$$ For a fixed $\mathbf x$, the optimization over $\boldsymbol{\pi}_0$ is a quadratic programming problem, which can be solved efficiently using, *e.g.*, interior-point methods [@Boyd; @Boggs]. On the other hand, for fixed $\boldsymbol{\pi}_0$, the optimization over $\mathbf x$ is a non-convex optimization problem. Therefore, we resort to a gradient descent (GD) algorithm to optimize over $\mathbf x$, which only guarantees convergence to a *local* optimum. Finally, we employ an iterative method to solve (\[ML\]), *i.e.*, we optimize over $\boldsymbol{\pi}_0$ for the current estimate of $\mathbf x$, then we optimize over $\mathbf x$ for the current estimate of $\boldsymbol{\pi}_0$, and so on. The derivative of $f(\mathbf x,\boldsymbol{\pi}_0)$ with respect to $\mathbf x_j$ is given by $$\begin{aligned}
&
\left[\nabla_{\mathbf x}f(\mathbf x,\boldsymbol{\pi}_0)\right]_j=
\frac{\mathrm d}{\mathrm d\mathbf x_j}f(\mathbf x,\boldsymbol{\pi}_0)
\nonumber\\&
=
-\sum_{k=0}^N\boldsymbol{\pi}_0^T\frac{\mathrm d
\left[\prod_{j=1}^{k}\mathbf P_{\Delta}(\mathbf x,\mathbf s_{E,j-1})^{n_j-n_{j-1}}\right]
}{\mathrm d\mathbf x_j}\mathbf Z
{\nonumber}\\&
\times
\left(\mathbf y_k-\boldsymbol{\pi}_0^T\prod_{j=1}^{k}\mathbf P_{\Delta}(\mathbf x,\mathbf s_{E,j-1})^{n_j-n_{j-1}}
\mathbf Z\right)^T.\end{aligned}$$ We further assume that the intervals satisfy $t_{k+1}-t_k=T,\ \forall k$, so that $n_k=kn,\ \forall k$, and we enforce $n=2^b$, for some integer $b>0$. This can be accomplished by appropriately choosing $\Delta\ll 1$. Then, the derivative $\frac{\mathrm d
\left[\prod_{j=1}^{k}\mathbf P_{\Delta}(\mathbf x,\mathbf s_{E,j-1})^{n}\right]
}{\mathrm d\mathbf x_j}$ can be efficiently computed recursively as $$\begin{aligned}
&\frac{\!\mathrm d\!\!\left[\overset{k}{\underset{j=1}{\prod}}\!\mathbf P_{\!\Delta\!}(\mathbf x,\mathbf s_{E,j-1})^{n}\right]\!}{\mathrm d\mathbf x_j}
{=}
\frac{\!\mathrm d\!\!\left[\overset{k-1}{\underset{j=1}{\prod}}\mathbf P_{\!\Delta\!}(\mathbf x,\mathbf s_{E,j-1})^{n}\right]\!}{\mathrm d\mathbf x_j}\mathbf P_{\!\Delta\!}(\mathbf x,\mathbf s_{E,k-1})^{n}
\nonumber\\&
+
\prod_{j=1}^{k-1}\mathbf P_{\Delta}(\mathbf x,\mathbf s_{E,j-1})^{n}\frac{\mathrm d\mathbf P_{\Delta}(\mathbf x,\mathbf s_{E,k-1})^{n}}{\mathrm d\mathbf x_j},\end{aligned}$$ where the derivative $\frac{\mathrm d\mathbf P_{\Delta}(\mathbf x,\mathbf s_{E})^{n}}{\mathrm d\mathbf x_j}$ can be efficiently computed recursively as $$\begin{aligned}
&\frac{\mathrm d\mathbf P_{\Delta}(\mathbf x,\mathbf s_{E})^{2^{b}}}{\mathrm d\mathbf x_j}
=
\frac{\mathrm d\mathbf P_{\Delta}(\mathbf x,\mathbf s_{E})^{2^{b-1}}}{\mathrm d\mathbf x_j}\mathbf P_{\Delta}(\mathbf x,\mathbf s_{E})^{2^{b-1}}
{\nonumber}\\&\quad
+
\mathbf P_{\Delta}(\mathbf x,\mathbf s_{E})^{2^{b-1}}
\frac{\mathrm d\mathbf P_{\Delta}(\mathbf x,\mathbf s_{E})^{2^{b-1}}}{\mathrm d\mathbf x_j}.\end{aligned}$$ Finally, let $\hat{\mathbf x}_p$ be the estimate of $\mathbf x$ at the $p$th iteration of the GD algorithm. Then, the GD algorithm updates the ML estimate of $\mathbf x$ as $$\begin{aligned}
&\hat{\mathbf x}_{p+1}=(\hat{\mathbf x}_p-\mu_p\nabla_{\mathbf x} f(\hat{\mathbf x}_p,\hat{\mathbf e}_p))^+,\end{aligned}$$ where $0<\mu_k\ll 1$ is the (possibly, time-varying) step size and we have defined $(v)^+=\max\{v,0\}$, applied to each entry, so that a non-negativity constraint is enforced (in fact, the entries of $\mathbf x$ need to be non-negative).
Results
-------
We use the algorithm outlined above to fit the parameter vector $\mathbf x$ to the experimental data. While, in principle, the capacities of both the IECP ($M_{CH}$) and the ATP pool ($N_{AXP}$) need to be estimated, we found that $M_{CH}=N_{AXP}=20$ provides a good fit, and good trade-off between convergence of the estimation algorithm and fitting. Note that the ATP capacity of the cell culture is $3.6{\,\,\,\mathrm{mM}}$ and the concentration of cells is $10^{12} {\,\,\,\mathrm{cells/liter}}$ (see [@Ozalp]). It follows that the capacity of the ATP pool of each cell is $2.16 \times10^{9}{\,\,\,\mathrm{molecules/cell}}$. Since, approximately, $2.5$ ATP molecules are created by the flow of $2$ electrons in the ETC (see [@Biochemistry Sec. 18.6]), the ATP pool may carry $1.728 \times10^{9}{\,\,\,\mathrm{electrons/cell}}$. Therefore, one “unit” in the stochastic model corresponds to $N_E=0.864\times10^{8}{\,\,\,\mathrm{electrons}}$, or $1.08\times10^{8}{\,\,\,\mathrm{ATP\ molecules}}$. Similarly, since each NADH molecule carries $2$ electrons which actively participate in the ETC, we have that one “unit” corresponds to $0.432\times10^{8}{\,\,\,\mathrm{NADH\ molecules}}$. The time-series $\{(\text{ATP}_k,\text{NADH}_k,t_k)\}$ is first extracted from [@Ozalp], where $\text{ATP}_k$ is in ${\,\,\,\mathrm{mM}}$ (stars in Fig. \[figmodel\].a), $\text{NADH}_k$ is a fluorescence level ($\times 10^{-6}$, stars in Fig. \[figmodel\].a), which we assume to be linearly proportional to the NADH level. The time-series is then converted to feasible values in the stochastic model. In particular, since the ATP capacity of the cell culture is $3.6{\,\,\,\mathrm{mM}}$, and assuming that the IECP capacity of the cell culture is $\max_k \text{NADH}_k=12.985$ \[fluorescence $\times 10^{-6}$\], *i.e.*, the maximum level reached in the NADH measurements, the time-series is $$\begin{aligned}
\label{conversion}
&y_k^{(NADH)}=\frac{\text{NADH}_k\ {\,\,\,\mathrm{[fluorescence\ x\ 10^{-6}]}}}{12.985\ {\,\,\,\mathrm{[fluorescence\ x\ 10^{-6}]}}}M_{CH},
{\nonumber}\\&
y_k^{(ATP)}=\frac{\text{ATP}_k\ {\,\,\,\mathrm{[mM]}}}{3.6\ {\,\,\,\mathrm{[mM]}}}N_{AXP},$$ so that, from (\[ts\]), $\alpha_{NADH}\simeq 0.650{\,\,\,\mathrm{[fluorescence\ x\ 10^{-6}]}}$ and $\alpha_{ATP}=0.18{\,\,\,\mathrm{[mM]}}$. Note that both $y_k^{(NADH)}$ and $y_k^{(ATP)}$ are dimensionless quantities. The time-series $\{(\mathbf y_k,t_k)\}$, where $\mathbf y_k=[y_k^{(NADH)},y_k^{(ATP)}]$, is then fed into the estimation algorithm. The EA concentration (molecular Oxygen) is assumed to be constant throughout the experiment, and sufficient to sustain reduction. On the other hand, the ED concentration profile, extrapolated from [@Ozalp], is zero at time $t<80{\,\,\,\mathrm{s}}$, when cells are starved, $30{\,\,\,\mathrm{mM}}$ at $t=80{\,\,\,\mathrm{s}}$, when glucose is added to the starved cells, and constantly decreases until it becomes zero at time $t\simeq 1300{\,\,\,\mathrm{s}}$, when cells become starved again.
\[rem1\] In the parameter estimation, the samples after $t\simeq 1300{\,\,\,\mathrm{s}}$ are discarded, since cells are starved after that time, which, in turn, results in increased cell lysis occurring in the cell suspension. The cellular material released by lysis can be used by other intact cells as ED or EA. For this reason, and since the extent of cell lysis in unpredictable in the cell culture, the concentrations of the ED and the EA cannot be accurately determined after $t\simeq 1300{\,\,\,\mathrm{s}}$, rendering the corresponding experimental data useless.
-- --
-- --
With this approach, the estimated parameters, used in (\[flows\]) to compute the corresponding flow rates, are given by $$\begin{aligned}
\left\{
\begin{array}{ll}
\hat\gamma=0&{\,\,\,\mathrm{units/mM/s}},\\
\hat\rho=2.31\times 10^{-3}& {\,\,\,\mathrm{units/mM/s}},\\
\hat\zeta=4.866\times 10^{-3}&{\,\,\,\mathrm{units/s}},\\
\hat\beta=0.850\times 10^{-3}&{\,\,\,\mathrm{units/mM/s}},
\end{array}
\right.\end{aligned}$$ where “units”, equivalent to $0.864\times10^{8}{\,\,\,\mathrm{electrons}}$, refers to the number of slots being occupied/emptied in the respective queue of the stochastic model, and ${\,\,\,\mathrm{mM}}$ refers to the glucose ED concentration. In particular, the “units” can be converted back into the original scale (\[conversion\]), which is related to the overall cell culture. The corresponding quantity related to a single cell is then obtained by converting one “unit” to the corresponding molecular quantity (1 unit=$1.08\times10^{8}{\,\,\,\mathrm{ATP\ molecules}}$=$0.432\times10^{8}{\,\,\,\mathrm{NADH\ molecules}}$).
Figs. \[figmodel\].a and \[figmodel\].b plot, respectively, the ATP and NADH time-series, related to the cell culture, and the predicted values based on our proposed stochastic model. We observe a good fit in the time-interval $t\in [0{\,\,\,\mathrm{s}},1300{\,\,\,\mathrm{s}}]$. [The corresponding standard deviation of the error between the prediction and experimental curves is 0.1325\[mM\] and 1.7736\[Fluorescence$\times 10^{-6}$\], respectively. The prediction error observed in the two figures can be explained by both cell lysis occurring in the bacterial population and the resulting distortion in the ATP and NADH levels, as explained in Remark \[rem1\], and by the *bias* introduced by our specific choice of the parametric model (\[flows\]), which may not be sufficiently accurate to capture higher order fluctuations. The investigation of other parametric models to improve the prediction accuracy is left for future work.]{} Figs. \[figmodel\].c and \[figmodel\].d plot, respectively, the expected ATP and NADH generation and consumption rates over time, related to a single cell. These biophysical parameters were not directly measured in the experiments by [@Ozalp], but can be predicted by our proposed model. We notice that the ATP generation/consumption rate is of the order of $5\times 10^5$/$3\times 10^6$, whereas the NADH generation/consumption rate is of the order of $2\times 10^6$/$2\times 10^5$ (molecules per cell per second). These values are indeed physical, and consistent with known metabolic rates in yeast [@Teusink], which further motivates the development of this stochastic model as a predictive tool for microbial energetics.
Future work {#future}
===========
Our current experimental work involves measuring ATP and NADH levels in isolated single cells of bacterium *S. oneidensis* MR-1. This organism is capable of extracellular electron transfer by utilizing its array of outer-membrane multi-heme cytochromes and, due to its unique properties, presents a great model organism for this study. Similarly to the previous experiment done on the yeast *Saccharomyces cerevisiae* [@Ozalp], whose data have been used in our experimental validation in Sec. \[estimation\], the ED will be abruptly added to a culture of starved bacterial cells, and, subsequently, the cellular ATP and NADH levels will be measured over time. As opposed to [@Ozalp], we are developing an experimental setup in which the amount of the ED and the EA can be maintained at any desired level, therefore producing additional unprecedented data to be compared with our theoretical model.
[The next step in our experiments will be to assemble bacterial cables, *e.g.*, using similar techniques as in [@Wang], and perform ATP and NADH measurements in these cells as the availability and the type of the ED and the EA is varied. In order to control and stimulate the growth of bacterial cables, a population of cells will be initially grown in an ED/EA rich medium, and subsequently moved to an environment with limited amounts of ED/EA, causing cell growth to stop. Then, the cells will be placed in a microfluidic medium where they can be moved, *e.g.*, via optical tweezers [@Liu2], to form a one-dimensional bacterial cable. Adjusting the environmental parameters will then induce the production of electron transfer components in the bacteria, thus enabling long-range electron transfer in the cable. By keeping the concentration of ED/EA along the cable small, the bacteria are forced to use the externally provided solid-state electrodes as the electron source/sink and maintain the collective electron transfer through the cable. The role of each cell in this collective behavior is in the form of establishing direct cell-cell contact and facilitating electron transfer to and from the adjacent cells, and cooperation of every single cell in the system is necessary to provide enough ED/EA to sustain the entire network. Solid-phase electrodes poised to a desired electric potential can be used as the EA for such a cable. The rate of electron transfer to such an electrode can be controlled by adjusting its potential, and this electron transfer rate to the electrode can be accurately measured. Similar manipulations of the ED and the EA and their availability and the subsequent measurements on the state of the cells and the transfer rates within the chain will result in a vast amount of quantitative data to validate our stochastic model for IET.]{}
Conclusions {#conclu}
===========
In this paper, we have presented a stochastic model for electron transfer in bacterial cables. In particular, we have specialized the stochastic model to the case of an isolated cell, which is the building block of more complex bacterial cables, and we have provided an example of the application to the computation of the cell’s lifetime. Moreover, we have designed a parameter estimation framework, based on a parametric description of the model, guided by biological constraints. The parameters were fit to experimental data available in the literature, demonstrating the capability of the proposed stochastic model to predict salient features related to the energetic state of the cells, such as ATP generation and consumption rates. This study is a first step towards addressing questions of more communications theoretic relevance, such as the interplay between information capacity of a microbial community and lifetime of the cells, reliability and delay in electron-based nanonetworks.
[^1]: N. Michelusi and U. Mitra are with the Ming Hsieh Department of Electrical Engineering, University of Southern California, Los Angeles, USA; M. Y. El-Naggar and S. Pirbadian are with the Department of Physics and Astronomy, University of Southern California, Los Angeles, USA; emails:
[^2]: N. Michelusi and U. Mitra acknowledge support from one or all of these grants: ONR N00014-09-1-0700, CCF-0917343, CCF-1117896, CNS-1213128, AFOSR FA9550-12-1-0215, and DOT CA-26-7084-00. S. Pirbadian and M. Y. El-Naggar acknowledge support from NASA Cooperative Agreement NNA13AA92A and grant DE-FG02-13ER16415 from the Division of Chemical Sciences, Geosciences, and Biosciences, Office of Basic Energy Sciences of the US Department of Energy. N. Michelusi is in part supported by AEIT (Italian association of electrical engineering) through the research scholarship “Isabella Sassi Bonadonna 2013”.
[^3]: Parts of this work have appeared in [@CISS].
[^4]: Note that the terms *high* and *low* referred to the energy of electrons are used here only in relative terms, *i.e.*, relative to the redox potential at the cell surface. In bacterial cables, the redox potential slowly decreases along the cable, thus inducing a net flow of electrons from one end to the opposite one.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper, we consider sums of class numbers of the type\
$\sum_{m\equiv a\pmod{p}} H\left(4n-m^2\right)$, where $p$ is an odd prime, $n\in {\mathbb{N}},$ and $a\in {\mathbb{Z}}$. By showing that these are coefficients of mixed mock modular forms, we obtain explicit formulas. Using these formulas for $p=5$ and $7$, we then prove a conjecture of Brown et al. in the case that $n=\ell$ is prime.
address:
- |
Mathematical Institute\
University of Cologne\
Weyertal 86-90\
50931 Cologne\
Germany
- 'Department of Mathematics, University of Hong Kong, Pokfulam, Hong Kong'
author:
- Kathrin Bringmann
- Ben Kane
title: Sums of class numbers and mixed mock modular forms
---
Introduction and statements of results
======================================
Let $H(n)$ denote the $n$th
Hurwitz class number
, i.e., the number of equivalence classes of positive definite binary quadratic forms of discriminant $-n$ with the class containing $x^2+y^2$ weighted by $\frac{1}{2}$ and the class containing $x^2+xy+y^2$ weighted by $\frac{1}{3}$. Moreover, by convention $H(0)=-\frac{1}{12}$. Certain congruence classes appear as coefficients of weight $\frac{3}{2}$ Eisenstein series. However, the generating function for all Hurwitz class numbers $$\label{Hgen}
\mathcal{H}(q):=\sum_{n\geq 0} H(n)q^n$$ is not itself modular, but rather mock modular [@HZ]. Roughly speaking, this means that $\mathcal{H}$ may be naturally “completed” to a non-holomorphic modular form (a further description is given in Section \[sec:Hurwitz\]). Mock modular forms have since shown up in a variety of applications. To name a few examples, Ramanujan’s mock theta functions have been shown to be mock modular forms [@ZwegersThesis], they have led to asymptotic and exact formulas in partition theory [@BringmannOnoInvent; @BringmannOnoAnnals], they are related to Lie superalgebras [@BringmannFolsom; @BringmannOnoKac; @Liesuper], and they are connected to the quantum theory of black holes [@BM; @DabMurZag].
In this paper, we prove conjectures reminiscent of the famous identity (cf. p. 154 of [@Eichler]) $$\label{eqn:1case}
\sum_{|m|<2\sqrt{\ell}} H\left(4\ell-m^2\right) = 2\ell,$$ where $\ell$ is an odd prime. More specifically, for a prime $p$, $a\in {\mathbb{Z}}$, and $n\in {\mathbb{N}}$, this paper is focused on sums of the type $$H_{a,p}(n):=\sum_{\substack{|m|\leq 2\sqrt{n}\\ m\equiv a\pmod{p}}} H\left(4n-m^2\right).$$
A number of identities for $H_{a,p}(n)$ were obtained in [@Hurwitz] for the special cases that $n=\ell$ is prime and $p=2$, $3$, $5$, or $7$. To give an example indicative of the results in [@Hurwitz], they proved in the special case that $p=5$ and $n=\ell$ is prime that $$H_{a,5}(\ell)= \begin{cases}
\frac{\ell-3}{2}&\text{if }a\equiv 0\pmod{5},\text{ and } \ell \equiv 4\pmod{5},\\
\frac{\ell-1}{2}&\text{if }a\equiv \pm 1\pmod{5},\text{ and } \ell \equiv 3\pmod{5},\\
\frac{\ell-1}{2}&\text{if }a\equiv \pm 2\pmod{5},\text{ and } \ell \equiv 2\pmod{5}.
\end{cases}$$ In the cases $p=5$ and $p=7$, they were unable to completely classify $H_{a,p}(\ell)$, but conjecture a number of pleasant identities similar to based on computer data.
\[conj:5case\] For every $a,L\in {\mathbb{Z}}$, there exist constants $c_1,c_2\in {\mathbb{Q}}$ (given explicity in ) such that for every prime $\ell\equiv L\pmod{5}$, we have $$H_{a,5}(\ell) = c_{1} \ell + c_{2}.$$
When restricted to certain congruence classes for $a$ and $\ell\pmod{7}$, they conjecture a similar formula for $H_{a,7}(\ell)$.
\[conj:7case\] If $L=3,5,6$ and $a\in {\mathbb{Z}}$ or $(a,L)\equiv (\pm 1,1)\pmod{7}$, then there exist constants $c_1,c_2\in {\mathbb{Q}}$ (given explicitly in ) such that for every prime $\ell\equiv L\pmod{7}$, we have $$H_{a,7}(\ell) = c_{1} \ell + c_{2}.$$
We settle these conjectures in this paper.
\[thm:conj\] Conjectures \[conj:5case\] and \[conj:7case\] are true.
Theorem \[thm:conj\] is implied by a more general theorem, which we next describe. The key ingredient is to use the mock modularity of the Hurwitz class number generating function, whereas the authors of [@Hurwitz] only took advantage of the modularity of certain congruence classes. In particular, $H_{a,p}(n)$ are the coefficients of what are known as
mixed mock modular forms
, which are products of mock modular forms with modular forms.
We construct explicit mixed mock modular forms with these coefficients from the product of $\mathcal{H}(q)$ with unary theta functions $\vartheta_{a,N}(\tau)$ defined in (where $N\in {\mathbb{N}}$ and $q:=e^{2\pi i \tau}$ throughout). To completely describe these, for a series $f(\tau)=\sum_{n\in {\mathbb{Z}}} a(n) q^n$, we require the $d$th $U$-operator $f|U(d)(\tau):=\sum_{n\in {\mathbb{Z}}} a(nd) q^n$ and the twist of a series $f$ by a character $\chi$, i.e., $f\otimes \chi(\tau):=\sum_{n\in {\mathbb{Z}}} a(n)\chi(n) q^n$. It is straightforward to see that for every $a\in{\mathbb{Z}}$ and prime $p$, one has $$\label{eqn:classsums}
\sum_{\substack{n\geq 0\\ p\nmid n}} H_{a,p}(n) q^n= \left(\mathcal{H}(q)\vartheta_{a,p}\left(\tau\right)\right)\Big|U(4)\otimes \chi_{p}^2.$$ Our main theorem expresses the right-hand side of in terms of generating functions for explicit divisor sums. For this, define $$\mathcal{G}_{N,r}(q):=\sum_{n\geq 1}\sum_{\substack{dd'=n\\ d\equiv \pm r\pmod{N}\\ d'> d}} dq^n + \sum_{n\geq 1} \left(Nn-r\right)q^{\left(Nn-r\right)^2}.$$ Moreover, we let $S_{N,r}$ denote the operator $f | S_{N,r}(\tau) = \sum_{n\in {\mathbb{Z}}} a(Nn+r) q^{Nn+r}$ which sieves coefficients congruent to $r$ modulo $N$.
\[thm:gencomplete\] For every $a\in {\mathbb{Z}}$ and odd prime $p$, we have that $$\label{eqn:gencomplete2}
\left(\mathcal{H}(q)\vartheta_{a,p}\left(\tau\right)\right)\Big|U(4)\otimes \chi_{p}^2 + \sum_{\substack{b\pmod{p}\\ b\not\equiv \pm a\pmod{p}}} \mathcal{G}_{p,a+b}(q)\Big|S_{p,a^2-b^2}$$ is a weight $2$ holomorphic modular form on $\Gamma_0\left(p^2\right)\cap \Gamma_1(p)$. Moreover, if $a=0$, then is a weight 2 modular form on $\Gamma_0\left(p^2\right)$.
This paper is organized as follows. In Section \[sec:Hurwitz\], we recall Hirzebruch and Zagier’s completion of the class number generating function and introduce important series known as Appell-Lerch sums. In Section \[sec:mixed\], we show how to complete the functions on either side of to obtain non-holomorphic modular forms. We conclude the paper by proving Theorem \[thm:gencomplete\] and Theorem \[thm:conj\] in Section \[sec:main\].
Hurwitz class numbers, Lerch sums, and known facts {#sec:Hurwitz}
==================================================
As mentioned in the introduction, the generating function for the Hurwitz class numbers is a mock modular form. More precisely, Hirzebruch and Zagier [@HZ] proved that one can complete $\mathcal{H}$ by adding a certain simple sum involving the incomplete Gamma function $$\Gamma\left(s;y\right):=\int_{y}^{\infty} t^{s-1}e^{-t}dt.$$ The completed function is non-holomorphic, but belongs to a special class of functions known as harmonic weak Maass forms. To define these, we require the
weight $k$ hyperbolic Laplacian
(denoting $\tau=x+iy$ with $x,y\in {\mathbb{R}}$ throughout) $$\Delta_{k}:=-y^2\left(\frac{\partial^2}{\partial x^2} +\frac{\partial^2}{\partial y^2}\right) +iky\left(\frac{\partial}{\partial x} + i \frac{\partial}{\partial y}\right).$$ Weight $k$
harmonic weak Maass forms
for $\Gamma\subset{{\text {\rm SL}}}_2({\mathbb{Z}})$ are real analytic functions $\mathcal{F}:{\mathbb{H}}\to{\mathbb{C}}$ satisfying the following properties:
1. $\mathcal{F}|_{k} \gamma\left(\tau\right) = \mathcal{F}\left(\tau\right)$ for every $\gamma\in \Gamma$,
2. $\Delta_{k}\left(\mathcal{F}\right)=0$,
3. $\mathcal{F}$ has at most linear exponential growth at each cusp of $\Gamma$.
Here $|_k$ is the usual weight $k$ slash operator. The class number generating function also belongs to a distinguished subspace, consisting of those forms whose $n$th coefficient (in the Fourier expansion in $x$) vanishes unless $(-1)^{k-\frac{1}{2}}n\equiv 0,1\pmod{4}$, known as
Kohnen’s plus space
[@Kohnen].
We collect the modularity properties of $\mathcal{H}$ in the following theorem which can be easily concluded from Theorem 2 of [@HZ].
\[thm:Hcomplete\] The function $$\label{Hshadow}
\widehat{\mathcal{H}}(\tau):=\mathcal{H}(q) + \frac{1}{4\sqrt{\pi}}\sum_{n>0}n\Gamma\left(-\frac12; 4\pi n^2 y\right)q^{-n^2}+\frac{1}{8\pi \sqrt{y}}$$ is a weight $\frac{3}{2}$ harmonic weak Maass form on $\Gamma_0(4)$ in Kohnen’s plus space.
A number of other completions help us to prove Theorem \[thm:gencomplete\]. To this end, for $u\in {\mathbb{C}}\setminus \left({\mathbb{Z}}\tau+{\mathbb{Z}}\right)$ we define the
multivariable Appell function
[@Liesuper; @ZwegersAppell] $$\label{eqn:A2def}
A_{\ell}\left(u,v;\tau\right):=e^{\pi i\ell u}\sum_{n\in {\mathbb{Z}}} \frac{(-1)^{\ell n}q^{\frac{\ell}{2} n\left(n+1\right)}e^{2\pi i n v}}{1-e^{2\pi i u} q^n}.$$ Directly from the definition, we have $$\begin{aligned}
\nonumber
(-1)^{\ell}A_{\ell}(u+1,v;\tau) &= A_{\ell}(u,v+1;\tau) = A_{\ell}(u,v;\tau),\\
\label{eqn:Ashift}
A_{\ell}\left(u+\tau,v+\ell\tau;\tau\right)&=(-1)^{\ell}q^{-\frac{\ell}{2}}e^{2\pi i v}A_{\ell}(u,v;\tau).\end{aligned}$$ In order to add a non-holomorphic function which “completes” $A_{\ell}$ to satisfy modularity, we define $$\begin{aligned}
\vartheta\left(z;\tau\right)&:=\sum_{n\in \frac{1}{2}+{\mathbb{Z}}}e^{\pi i n^2\tau+2\pi i n\left(z+\frac{1}{2}\right)},\\
R\left(u;\tau\right)&:=\sum_{n\in \frac{1}{2}+{\mathbb{Z}}} \left({\operatorname{sgn}}\left(n\right) - E\left(\left(n+\frac{\operatorname{Im}(u)}{y}\right)\sqrt{2y}\right)\right)\left(-1\right)^{n-\frac{1}{2}}e^{-\pi i n^2\tau-2\pi i n u },\\
E\left(z\right)&:=2\int_0^z e^{-\pi t^2}dt.\end{aligned}$$ We note that for $u\neq 0$, we also have the useful formula $$\label{eqn:Egamma}
E(u)={\operatorname{sgn}}(u)\left(1-\frac{e^{-\pi u^2}}{\pi |u|}+\frac{1}{2\sqrt{\pi}}\Gamma\left(-\frac12; \pi u^2\right)\right).$$
Theorem 2.2 of [@ZwegersAppell] yields the transformation properties of the completion of $A_{\ell}$: $$\begin{gathered}
\label{eqn:A2complete}
\widehat{A}_{\ell}\left(u,v;\tau\right):=A_{\ell}(u, v; \tau)\\
+\frac{i}{2}\sum_{k=0}^{\ell-1} e^{2\pi iku}\vartheta\left(v+k\tau+\frac{\ell-1}{2}; \ell\tau\right)R\left(\ell u-v-k\tau-\frac{\ell-1}{2}; \ell\tau\right).\end{gathered}$$
The function $\widehat{A}_{\ell}$ satisfies $$\label{eqn:A2modularity}
\widehat{A}_{\ell}\left(\frac{u}{c\tau+d},\frac{v}{c\tau+d};\frac{a\tau+b}{c\tau+d}\right)=\left(c\tau+d\right) e^{\frac{\pi i c}{c\tau+d}\left(-\ell u^2+2uv\right)}\widehat{A}_\ell\left(u,v; \tau\right).$$ Moreover, for every $m_1,m_2,n_1,n_2\in {\mathbb{Z}}$, we have $$\begin{gathered}
\label{eqn:A2elliptic}
\widehat{A}_{\ell}\left(u+n_1\tau+m_1,v+n_2\tau+m_2;\tau\right)\\
= (-1)^{\ell\left(n_1+m_1\right)}e^{2\pi i u\left(\ell n_1 -n_2\right)}e^{-2\pi i n_1v}q^{\frac{\ell}{2}n_1^2-n_1n_2}\widehat{A}_{\ell}\left(u,v;\tau\right).\end{gathered}$$
Mixed mock modular forms and non-holomorphic completions {#sec:mixed}
========================================================
In this section, we show how to add a non-holomorphic function to “complete” each of the functions defined in the introduction to obtain a non-holomorphic modular form. In order to write down these completions, for $N\in {\mathbb{N}}$ it is helpful to define the function $$R_{a,N}(\tau):=\sum_{\substack{n\in {\mathbb{Z}}\\ n \equiv a\pmod{N}}} |n|\Gamma\left(-\frac{1}{2};4\pi n^2 y\right) q^{-n^2}.$$ Note that $R_{a,N}$ only depends on $a\pmod{N}$ and $$R_{-a,N}=R_{a,N}.$$ We further define the
unary theta functions
$$\label{eqn:unth}
\vartheta_{a,N}\left(\tau\right):=\sum_{\substack{m\in {\mathbb{Z}}\\ m\equiv a\pmod{N}}}q^{m^2}.$$
In this subsection, we complete $\left(\mathcal{H}(q)\vartheta_{2a,p}(\tau)\right)\big|U(4)\otimes\chi_p^2$.
\[lem:Htwist\] For every $a,b\in {\mathbb{Z}}$ and odd prime $p$, the function $$\begin{gathered}
\label{eqn:Htwist}
\left(\mathcal{H}(q)\vartheta_{2a,p}(\tau)\right)\Big|U(4)\otimes\chi_p^2+\frac{1}{8\sqrt{\pi}}\sum_{\substack{b\pmod{p}\\ b\not\equiv \pm a \pmod{p}}}\sum_{k=0}^1 R_{2b+kp,2p}\left(\frac{\tau}{4}\right)\vartheta_{2a+kp,2p}\left(\frac{\tau}{4}\right)\\
+ \frac{\vartheta_{2a,2p}\left(\frac{\tau}{4}\right)}{4\pi\sqrt{y}}\end{gathered}$$ satisfies weight $2$ modularity for $\Gamma:=\Gamma_0\left(p^2\right)\cap \Gamma_1(p)$.
By Proposition 2.1 of [@Shimura], $\vartheta_{2a,p}$ is modular of weight $\frac{1}{2}$ on $\Gamma_0\left(4p^2\right)\cap \Gamma_1(p)$.
We first write $$\label{eqn:HThrewrite}
\left(\mathcal{H}(q)\vartheta_{2a,p}(\tau)\right)\big|U(4)\otimes\chi_p^2= \sum_{1\leq r\leq p-1} \left(\mathcal{H}(q)\vartheta_{2a,p}(\tau)\right)\big|U(4)\big|S_{p,r}.$$ By Lemma 1 of [@Li], $(\widehat{\mathcal{H}}(q)\vartheta_{2a,p}(\tau))|U(2)$ fulfills weight 2 modularity on $\Gamma_0(2p^2)\cap \Gamma_1(2p)$. Since the $n$th coefficient of $(\widehat{\mathcal{H}}(q)\vartheta_{2a,p}(\tau))|U(2)$ is zero unless $n$ is even, Lemma 4 of [@Li] implies that $(\widehat{\mathcal{H}}(q)\vartheta_{2a,p}(\tau))|U(4)$ satisfies weight 2 modularity on $\Gamma_0(p^2)\cap \Gamma_1(p)$. Moreover, since modular forms on $\Gamma_1(p)$ split into modular forms on $\Gamma_0(p)$ with Nebentypus, rewriting $\chi_p^2 = 1-U(p)V(p)$ (where as usual $f|V(d)(\tau):=f(d\tau)$), Lemma 1 of [@Li] implies that the level goes down with $U(p)$ and back up with $V(p)$, so that overall the group becomes $\Gamma$.
By Theorem \[thm:Hcomplete\], to complete each summand $\left(\mathcal{H}(q)\vartheta_{2a,p}(\tau)\right)\big|U(4)\big|S_{p,r}$ on the right-hand side of , one must add $$\label{eqn:Sievecomplete}
\left(\frac{1}{4\sqrt{\pi}}\sum\limits_{\substack{n>0\\ m\equiv 2a\pmod{p}}} n\Gamma \left(-\frac12; 4\pi n^2 y\right) q^{m^2-n^2}+\frac{1}{8\pi\sqrt{y}}\sum_{m\equiv 2a\pmod{p}}q^{m^2}\right)\Bigg|U(4)\Bigg|S_{p,r}.$$ Due to $m\equiv 2a\pmod{p}$ and the congruences implied by $U(4)$ and $S_{p,r}$, the congruence conditions on $n$ and $m$ are equivalent to $m\equiv n\pmod{2}$, $m\equiv 2a\pmod{p}$, and $n\equiv \pm 2b \pmod{p}$, where $b$ satisfies $a^2-b^2\equiv r\pmod{p}$. If no such $b$ exists, then equals zero. We may thus assume that such a $b$ exists. Whenever $b\not\equiv 0\pmod{p}$, the fact that $p$ is odd implies that equals $$\frac{1}{8\sqrt{\pi}}\left(\sum_{k=0}^1\sum_{\pm} R_{\pm 2b+kp,2p}\left(\frac{\tau}{4}\right)\vartheta_{2a+kp,2p}\left(\frac{\tau}{4}\right)\right).$$
Moreover, in the case that $b\equiv 0\pmod{p}$, equals $$\frac{1}{8\sqrt{\pi}}\left(\sum_{k=0}^1 R_{kp,2p}\left(\frac{\tau}{4}\right)\vartheta_{2a+kp,2p}\left(\frac{\tau}{4}\right)\right)+\frac{\vartheta_{2a,2p}\left(\frac{\tau}{4}\right)}{4\pi\sqrt{y}}.$$ To finish the proof, one then sums over all choices of $b$ to obtain .
In this section, we complete $\sum_{\pm}\mathcal{G}_{p,a\pm b}(q)|S_{p,a^2-b^2}$.
We begin with a lemma, which gives a more useful form ($\ell\in {\mathbb{Z}}$) for $$\mathcal{R}_{\ell,p}(v;\tau):=\left[\frac{d}{dv} \left(e^{\frac{\pi i\ell v}{p}}q^{-\frac{\ell^2}{4}}R\left(p\ell\tau-v-\frac12; 2p^2\tau\right)\right)\right]_{v=0}.$$
\[lem:diffR\] If $p$ is an odd prime and $\ell\in{\mathbb{Z}}$ satisfies $-p< \ell\leq p$, then $$\mathcal{R}_{\ell,p}(v;\tau)=\sqrt{\pi}\sum_{n\in\frac12+{\mathbb{Z}}}\left\lvert n+\frac{\ell}{2p}\right\rvert \Gamma\left(-\frac12; 4\pi p^2\left(n+\frac{\ell}{2p}\right)^2 y\right) q^{-p^2\left(n+\frac{\ell}{2p}\right)^2}+\frac{\delta_{\ell=p}}{p\sqrt{y}}.$$
We have $$\begin{gathered}
e^{\frac{\pi i\ell v}{p}}q^{-\frac{\ell^2}{4}}R\left(p\ell\tau-v-\frac12; 2p^2\tau\right)\\
=i\sum_{n\in\frac12+{\mathbb{Z}}}\left({\operatorname{sgn}}(n)-E\left(2p\left(n+\frac{\ell}{2p}-\frac{\text{Im}(v)}{2p^2y}\right)\sqrt{y}\right)\right)
q^{-p^2\left(n+\frac{\ell}{2p}\right)^2}e^{2\pi iv\left(n+\frac{\ell}{2p}\right)}.\end{gathered}$$ We now note that whenever $n+\frac{\ell}{2p}\neq 0$, we have $${\operatorname{sgn}}(n)={\operatorname{sgn}}\left(n+\frac{\ell}{2p}\right).$$ Differentiating and plugging in yields the claim, after a straightforward calculation.
In order to complete the right-hand side of , we pair the terms $b$ and $-b$ and determine the associated completion.
\[lem:Gcompletegen\] Suppose that $a,b\in {\mathbb{Z}}$ and $p$ is an odd prime with $a\not\equiv \pm b\pmod{p}$.
1. If $b\not\equiv 0\pmod{p}$, then the function $$\widehat{\mathcal{G}}_{p,a,b}(q):=\sum_{\pm}\mathcal{G}_{p,a\pm b}(q)\Big|S_{p,a^2-b^2}-\frac{1}{4\sqrt{\pi}}\sum_{k=0}^1 R_{kp+2b,2p}\left(\frac{\tau}{4}\right)\vartheta_{kp+2a,2p}\left(\frac{\tau}{4}\right)$$ satisfies weight $2$ modularity for $\Gamma$.
2. If $b\equiv 0\pmod{p}$, then the function $$\widehat{\mathcal{G}}_{p,a,0}(q):=\mathcal{G}_{p,a}(q)\Big|S_{p,a^2}-\frac{1}{8\sqrt{\pi}}\sum_{k=0}^1 R_{kp,2p}\left(\frac{\tau}{4}\right)\vartheta_{kp+2a,2p}\left(\frac{\tau}{4}\right) - \frac{\vartheta_{2a,2p}\left(\frac{\tau}{4}\right)}{4\pi\sqrt{y}}$$ satisfies weight $2$ modularity for $\Gamma$.
3. In the case that $a\equiv 0\pmod{p}$, the function $$\sum_{0<b<\frac{p}{2}}\widehat{\mathcal{G}}_{p,0,b}\Big|S_{p,-b^2}$$ furthermore satisfies weight 2 modularity for $\Gamma_0\left(p^2\right)$.
We first assume that $b\not\equiv 0\pmod{p}$ and may assume without loss of generality that $a+b<p$, since the definition of $\mathcal{G}_{p,a+b}$ only depends on $a$ and $b$ modulo $p$. Recalling the definition of $A_2$, we obtain $$\begin{aligned}
\mathcal{G}_{p, a+b}(q)\Big| S_{p,a^2-b^2}=&\sum_{n\equiv a^2-b^2 \pmod{p}}\sum\limits_{{dd'=n\atop{d\equiv \pm (a+b)\pmod{p}}}\atop{d'> d}}d q^n\\
=&\sum\limits_{n\geq 0\atop{m\geq 0}}\left(pn+(a+b)\right)q^{\left(pn+(a+b)\right)\left(pn+p+(a-b)+pm\right)}\\
&+\sum_{\substack{n\geq 1\\ m\geq 0}}\left(pn-(a+b)\right)q^{\left(pn-(a+b)\right)\left(pn+(b-a)+pm\right)}\\
=&\sum_{n\in{\mathbb{Z}}}\frac{\left(pn+(a+b)\right)q^{\left(pn+(a+b)\right)\left(pn+p+(a-b)\right)}}{1-q^{p\left(pn+(a+b)\right)}}\\
=&\frac{p}{2\pi i}\left[\frac{d}{dv}\left(A_2\left(\left(a+b\right)p\tau, v+2ap\tau; p^2\tau\right) e^{\frac{2\pi i\left(a+b\right)v}{p}}q^{a^2-b^2}\right)\right]_{v=0}.\end{aligned}$$ The analogous calculation for $-b$ follows by . A similar calculation yields a uniform equation for $b\equiv 0\pmod{p}$ and $a\not\equiv 0\pmod{p}$ as well. Overall, we obtain $$\mathcal{G}_{p, a\pm b}(q)\Big| S_{p,a^2-b^2}=\frac{p}{2\pi i}\left[\frac{d}{dv}\left(A_2\left(\left(b\pm a\right)p\tau, v\pm 2ap\tau; p^2\tau\right) e^{\frac{2\pi i\left(b\pm a\right)v}{p}}q^{a^2-b^2}\right)\right]_{v=0}.$$
We next prove the modularity of $$\label{eqn:hatG}
\widehat{G}_b(\tau):=\frac{p}{2\pi i}\left[\frac{d}{dv}\left(\widehat{A}_2\left(\left(a+b\right)p\tau, v+2ap\tau; p^2\tau\right) e^{\frac{2\pi i\left(a+b\right)v}{p}}q^{a^2-b^2}\right)\right]_{v=0}.$$ For $\left(\begin{smallmatrix} \alpha &\beta\\ \gamma&\delta\end{smallmatrix}\right)\in \Gamma$, we use equation followed by with $m_1=(a+b)\beta p$, $m_2=a\beta p$, $n_1=\left(b+a\right)\left(\frac{\alpha-1}{p}\right)\in {\mathbb{Z}}$, and $n_2=2a\left(\frac{\alpha-1}{p}\right)\in{\mathbb{Z}}$ to yield $$\widehat{G}_b\left(\frac{\alpha\tau+\beta}{\gamma\tau+\delta}\right) = \left(\gamma\tau+\delta\right)^2\widehat{G}_b(\tau).$$ Hence $\widehat{G}_b$ satisfies weight 2 modularity.
It remains to compute the non-holomorphic part of $\sum_{\pm}\widehat{G}_{\pm b}$. By and the definition of $\widehat{G}_b$, we have $$\begin{gathered}
\label{eqn:diffpm}
\sum_{\pm} \widehat{G}_{\pm b}(\tau) = \sum_{\pm} \mathcal{G}_{p, a\pm b}(q)\Big|S_{p,a^2-b^2} + \frac{p}{4\pi}\Bigg[\frac{d}{dv}\sum_{\pm}\sum_{k=0}^{1} \vartheta\left(v\pm 2ap\tau+kp^2\tau +\frac{1}{2};2p^2\tau\right)\\
\times R\left(2bp\tau -v-kp^2\tau -\frac{1}{2};2p^2\tau\right)e^{2\pi i\left(a^2-b^2+ k\left(b\pm a \right)p\right)\tau} e^{\frac{2\pi i (b\pm a)v}{p}} \Bigg]_{v=0}.\end{gathered}$$ Here we double count $b\equiv 0\pmod{p}$ to get a uniform formula. We now rewrite $$\begin{gathered}
q^{\frac{1}{4}\left(kp\pm 2a\right)^2}e^{\frac{2\pi i v}{p}\left(\frac{kp}{2} \pm 2a\right)} \vartheta\left(v\pm 2ap\tau+kp^2\tau +\frac{1}{2};2p^2\tau\right)\\
=-\sum_{m\in \frac{1}{2}+{\mathbb{Z}}} q^{\left(mp+\frac{kp}{2}\pm a\right)^2} e^{\frac{2\pi iv}{p}\left(mp+\frac{kp}{2}\pm a\right)}.\end{gathered}$$ However, we have that $$\begin{gathered}
\sum_{\pm}\left[\frac{d}{dv}\sum_{m\in \frac{1}{2}+{\mathbb{Z}}} q^{\left(mp+\frac{kp}{2}\pm a\right)^2} e^{\frac{2\pi iv}{p}\left(mp+\frac{kp}{2}\pm a\right)}\right]_{v=0}\\
=\frac{2\pi i}{p}\sum_{\pm}\sum_{m\in \frac{1}{2}+{\mathbb{Z}}}\left(mp+\frac{kp}{2}\pm a\right) q^{\left(mp+\frac{kp}{2}\pm a \right)^2}=0,\end{gathered}$$ which can be seen by making the change of variables $m\to -m-k$. Moreover, by taking $m\to -m-k$, we also see that $$\sum_{\pm}\sum_{m\in \frac{1}{2}+{\mathbb{Z}}} q^{\left(mp+\frac{kp}{2}\pm a \right)^2}=2\sum_{m\in \frac{1}{2}+{\mathbb{Z}}} q^{\frac{1}{4}\left(2mp+kp+ 2a \right)^2}=2\vartheta_{(1-k)p+2a,2p}\left(\frac{\tau}{4}\right).$$ It follows that equals $$-\frac{p}{2\pi}\sum_{k=0}^1 \mathcal{R}_{2b-kp,p}\left(v;\tau\right)\vartheta_{(1-k)p+2a,2p}\left(\frac{\tau}{4}\right).$$ By Lemma \[lem:diffR\], we may rewrite this as $$\begin{gathered}
-\delta_{b=p}\frac{\vartheta_{2a,2p}\left(\frac{\tau}{4}\right)}{2\pi\sqrt{y}}-\frac{1}{4\sqrt{\pi}}\sum_{k=0}^1 \vartheta_{(1-k)p+2a,2p}\left(\frac{\tau}{4}\right)\\
\times \sum_{n\in \frac{1}{2}+{\mathbb{Z}}} \left|2np-kp+2b\right|\Gamma\left(-\frac{1}{2};\pi\left(2np-kp+2b\right)^2y\right)q^{-\frac{1}{4}\left(2np-kp+2b\right)^2}\\
=-\frac{1}{4\sqrt{\pi}}\sum_{k=0}^1 R_{(1-k)p+2b,2p}\left(\frac{\tau}{4}\right)\vartheta_{(1-k)p+2a,2p}\left(\frac{\tau}{4}\right)-\delta_{b=p}\frac{\vartheta_{2a,2p}\left(\frac{\tau}{4}\right)}{2\pi\sqrt{y}}.\end{gathered}$$ Statements (1) and (2) now follow from Lemma \[lem:diffR\] with $\ell=2b-kp$.
We next prove part (3). The claim is equivalent to showing that $$g(\tau):=\frac{4\pi i}{p} \sum_{0<b<\frac{p}{2}}\widehat{\mathcal{G}}_{p,0,b}\Big|S_{p,-b^2}=\left[\frac{d}{dv}\sum_{b\pmod{p}^\ast}e^{\frac{2\pi ib v}{p}}q^{-b^2}\widehat{A}_2\left(bp\tau, v; p^2\tau\right)\right]_{v=0}$$ satisfies weight 2 modularity for $\Gamma_0\left(p^2\right)$. Here the sum runs over those $b\pmod{p}$ with $(b,p)=1$. Note that the sum only depends on $b\pmod{p}$ because, for $b'\equiv b\pmod{p}$, we may use with $n_1=\frac{b'-b}{p}$. However, a simple calculation yields $$\begin{gathered}
g\left(\frac{\alpha\tau+\beta}{\gamma\tau+\delta}\right)=(\gamma\tau+\delta)^2\left[\frac{d}{dv}\sum_{b\pmod{p}^\ast}q^{-\alpha^2b^2}e^{\frac{2\pi i b\alpha v}{p}}\widehat{A}_2\left(b\alpha p\tau, v; p^2\tau\right)\right]_{v=0}\\
=\left(\gamma\tau+\delta\right)^2 g(\tau),\end{gathered}$$ where we used that $b\alpha$ runs $\pmod{p}^\ast$ if $b$ does.
Proof of Theorem \[thm:gencomplete\] and Theorem \[thm:conj\] {#sec:main}
=============================================================
In this section, we prove our main theorem and then give explicit identities for $$\sum_{
\substack{n\geq 0\\ p\nmid n}} H_{a,p}(n)q^n$$ for certain fixed choices of $a$ and $p$. In particular, since the divisor sums occurring in Theorem \[thm:gencomplete\] are particularly simple for primes, we obtain the desired conjectures leading to Theorem \[thm:conj\]. We begin with the proof of our main theorem.
By fixing $a$ and summing over all congruence classes for $b\not\equiv \pm a \pmod{p}$ in Lemma \[lem:Gcompletegen\], we see that the non-holomorphic parts cancel the non-holomorphic part from Lemma \[lem:Htwist\] and hence the sum is a (weakly) holomorphic modular form. Every coefficient of our new overall function may be written as a linear combination of class numbers and divisor sums and hence grows polynomially. Thus we have a holomorphic modular form of weight $2$, yielding Theorem \[thm:gencomplete\].
To compute explicit identities, we use the following lemma, which follows from the valence formula.
\[lem:modcomplete\] If $p$ is an odd prime, $a\in {\mathbb{Z}}$, and $f$ is a holomorphic modular form of weight $2$ on $\Gamma$, then $$\left(\mathcal{H}(q)\vartheta_{a,p}\left(\tau\right)\right)\Big|U(4)\otimes \chi_{p}^2 + \sum_{\substack{b\pmod{p}\\ b\not\equiv \pm a\pmod{p}}} \mathcal{G}_{p,a+b}(q)\Big|S_{p,a^2-b^2}=f(\tau)$$ if and only if the first $\frac{p}{6}\left(p^2-1\right)$ Fourier coefficients agree.
Moreover, if $a\equiv 0\pmod{p}$ and $f$ satisfies weight $2$ modularity for $\Gamma_0(p^2)$, then the above identity holds if and only if it holds for the first $\frac{p}{6} \left(p+1\right)$ coefficients.
Denoting $\sigma(n):=\sum_{d\mid n} d$, we write the Eisenstein series part of the modular forms from Lemma \[lem:modcomplete\] (in the special cases $p=3,5,7$) in terms of $$\mathcal{D}(q):=\sum_{n=1}^{\infty} \sigma\left(n\right)q^n.$$ Since $\mathcal{D}$ is essentially a constant multiple of the weight $2$ Eisenstein series $E_2$, it is well-known that $$\label{Dshadow}
\widehat{\mathcal{D}}(\tau):=\mathcal{D}(q)-\frac{1}{24}+\frac{1}{8\pi y}$$ transforms like a modular form of weight $2$ on ${{\text {\rm SL}}}_2({\mathbb{Z}})$. In particular, since every non-trivial character $\chi$ satisfies $\chi(0)=0$, the function $\mathcal{D}\otimes \chi$ is a weight 2 holomorphic modular form. More precisely, if the modulus of $\chi$ is $m\in {\mathbb{N}}$, then $\mathcal{D}\otimes \chi$ is a weight 2 holomorphic modular form on $\Gamma_0(m^2)$ (cf. Proposition 2.8 of [@OnoBook]). Furthermore, if $r\not\equiv 0\pmod{p}$, then a straightforward calculation shows that $\mathcal{D}\big|S_{p,r}$ is a holomorphic modular form of weight $2$ on $\Gamma$.
It is well-known (cf. Section 7.2, Example 2 in [@DabMurZag]) that $$\left(\mathcal{H}\vartheta_{0,1}\right)\Big|U(4)=2\mathcal{D}-\mathcal{G}_{1,0}-\frac{1}{12},$$ while one sees directly that $\vartheta_{0,1}=\sum_{a\pmod{p}} \vartheta_{a,p}$ and $\vartheta_{a,p}=\vartheta_{-a,p}$. Hence we only need to determine formulas for the modular forms from Lemma \[lem:modcomplete\] whenever $0\leq a\leq \frac{p-3}{2}$ to obtain them for all $a\in {\mathbb{Z}}$. After constructing such modular forms, we use to conclude Theorem \[thm:conj\]. For simplicity, we only work out the exact identities for $a=0$.
$p=3$
-----
We include the case $p=3$ since the formulas are particularly simple in this case and because it indicates the general method well. Using the fact that $\mathcal{D}\otimes \chi_3$ is a holomorphic modular form of weight $2$ for $\Gamma_0(9)$, Lemma \[lem:modcomplete\] implies that $$\label{eqn:3}
\left(\mathcal{H}\vartheta_{0,3}\right)\Big|U(4)\otimes\chi_3^2=-2\mathcal{G}_{3, 1}\Big|S_{3, 2}+\mathcal{D}\otimes \chi_3^2-\frac14\mathcal{D}\otimes \chi_3\left(1+\chi_3\right).$$ The $n$th coefficient of $-2\mathcal{G}_{3, 1}\Big|S_{3, 2}$ is $0$ unless $n\equiv 2\pmod{3}$, in which case it equals $$-2\sum\limits_{\substack{ d\equiv \pm 1\pmod{3}\\ d\mid n, d<\frac{n}{d}}}d.$$ The $n$th coefficient of $\mathcal{D}\otimes \chi_3^2$ is $0$ if $3|n$ and otherwise $\sigma(n)$. Finally, the $n$th coefficient of $-\frac14\mathcal{D}\otimes \chi_3(1+\chi_3)$ is $0$ unless $n\equiv 1\pmod{3}$, in which case it equals $-\frac12 \sigma(n)$. Thus the overall $n$th coefficient on the right-hand side of is $$\begin{cases}
\frac12 \sigma(n)&\quad\text{ if } n\equiv 1\pmod{3},\\
\sigma(n)-2\sum\limits_{d\equiv \pm 1\pmod{3}\atop{d\mid n, d<\frac{n}{d}}}d &\quad\text{ if } n\equiv 2\pmod{3}.
\end{cases}$$ Comparing with , we get in particular for a prime $\ell>3$ $$H_{0,3}(\ell)=
\begin{cases}
\frac{\ell+1}{2}&\quad\text{ if } \ell\equiv 1\pmod{3},\\
\ell-1 &\quad\text{ if } \ell\equiv 2\pmod{3}.
\end{cases}$$
$p=5$
-----
The precise version of Conjecture \[conj:5case\] in [@Hurwitz] is given by the following.
\[conj:5\] For a prime $\ell$ and $a\in {\mathbb{Z}}$ one has that $$\label{eqn:5conjecture}
H_{a,5}(\ell)=
\begin{cases}
\frac{\ell+1}{2}&\text{if }a\equiv 0\pmod{5},\text{ and } \ell \equiv 1\pmod{5},\\
\frac{\ell+1}{3}&\text{if }a\equiv 0\pmod{5},\text{ and } \ell \equiv 2,3\pmod{5},\\
\frac{\ell+1}{3}&\text{if }a\equiv \pm 1\pmod{5},\text{ and } \ell \equiv 1,2\pmod{5},\\
\frac{5\ell+5}{12}&\text{if }a\equiv \pm 1\pmod{5},\text{ and } \ell \equiv 4\pmod{5},\\
\frac{5\ell-7}{12}&\text{if }a\equiv \pm 2\pmod{5},\text{ and } \ell \equiv 1\pmod{5},\\
\frac{\ell+1}{3}&\text{if }a\equiv \pm 2\pmod{5},\text{ and } \ell \equiv 3,4\pmod{5}.
\end{cases}$$
We prove Conjecture \[conj:5\] by showing the following more precise version.
\[cor:5case\] One has that $$\begin{aligned}
\left(\mathcal{H}\vartheta_{0,5}\right)\Big|U(4)\otimes\chi_5^2=&\frac{1}{2}\mathcal{D}\otimes\chi_5^2-\frac{1}{12}\mathcal{D}\otimes \chi_5\left(\chi_5-1\right) - 2 \mathcal{G}_{5,1}\Big|S_{5,4} - 2 \mathcal{G}_{5,2}\Big|S_{5,1},\\
\left(\mathcal{H}\vartheta_{1,5}\right)\Big|U(4)\otimes\chi_5^2=&\frac{1}{3}\mathcal{D}\otimes\chi_5^2+\left(\frac{1}{6}\mathcal{D}-\mathcal{G}_{5,1}-\mathcal{G}_{5,2}\right)\Big|S_{5,3}\\
&\qquad\qquad\qquad\qquad\quad+\left( \frac{1}{12}\mathcal{D}-\frac{1}{2}\mathcal{G}_{5,2}-\frac{1}{2}\mathcal{G}_{5,3}\right)\Big|S_{5,4}.\end{aligned}$$
In particular, the conjectured formula is true.
The holomorphic modular forms occurring by twisting $\mathcal{D}$ with a character of modulus 5 have level $\Gamma_0(25)$, while $\mathcal{D}|S_{5,a}$ ($a\not\equiv 0\pmod{5}$) is a holomorphic modular form for $\Gamma$. After checking $20$ coefficients, Lemma \[lem:modcomplete\] yields the equalities claimed in the corollary. To obtain , we simply apply an analysis similar to that used in the case for $p=3$ above to obtain the explicit coefficients.
$p=7$
-----
Conjecture \[conj:7case\] follows from the following more precise version.
\[conj:7\] For a prime $\ell$ and $a\in{\mathbb{Z}}$ one has that $$\label{eqn:7conjecture}
H_{a,7}(\ell)=
\begin{cases}
\frac{\ell+1}{3}&\text{if }a\equiv \pm 1\pmod{7},\text{ and } \ell \equiv 1\pmod{7},\\
\frac{\ell+1}{4}&\text{if }a\equiv \pm 1\pmod{7},\text{ and } \ell \equiv 3,6\pmod{7},\\
\frac{\ell+1}{4}&\text{if }a\equiv \pm 2\pmod{7},\text{ and } \ell \equiv 3,5\pmod{7},\\
\frac{\ell+1}{4}&\text{if }a\equiv \pm 3\pmod{7},\text{ and } \ell \equiv 5,6\pmod{7},\\
\end{cases}$$
The argument is analogous to the above cases for $p=3$ and $p=5$, except that there are cusp forms in the spaces of interest. For example, for $a=0$, we require the weight $2$ newform (of level $49$) $g_7$ associated to the elliptic curve $y^2+xy=x^3-x^2-2x-1$ (which is denoted $49A1$ and has CM by ${\mathbb{Q}}\left(\sqrt{-7}\right)$). The coefficients of $g_7$ are multiplicative and Parry has explicitly written $g_7$ in terms of the two variable Ramanujan theta function.
After comparing the first $56$ coefficients in the identity, Lemma \[lem:modcomplete\] immediately yields the following corollary.
\[cor:7casea=0\] One has that $$\begin{gathered}
\left(\mathcal{H}\vartheta_{0,7}\right)\Big|U(4)\otimes\chi_7^2=\frac{1}{4}\mathcal{D}\otimes\chi_7^2+\frac{1}{24}\mathcal{D}\otimes \chi_7\left(\chi_7-1\right) \\
-2 \mathcal{G}_{7,2}\Big|S_{7,3} -2 \mathcal{G}_{7,4}\Big|S_{7,-2} -2 \mathcal{G}_{7,1}\Big|S_{7,-1} + \frac{1}{4}g_7.\end{gathered}$$
We do not work out the details for the cusp forms for $a\not\equiv 0 \pmod{7}$, but rather list the resulting identity when restricting to coefficients in certain congruence classes where the coefficients of the cusp forms are all zero. This suffices to prove Conjecture \[conj:7\].
\[cor:7case\] The following identities hold.
1. For $r\equiv 1,3,6\pmod{7}$, one has $$\left(\mathcal{H}\vartheta_{1,7}\right)\Big|U(4)\Big|S_{7,r} = \frac{1}{4}\mathcal{D} +
\begin{cases}
\frac{1}{12}\mathcal{D}-\mathcal{G}_{7,2}-\mathcal{G}_{7,3}&\text{if }r\equiv 1\pmod 7,\\
0 & \text{if }r\equiv 3,6\pmod{7}.
\end{cases}$$
2. For $r\equiv 3,5\pmod{7}$, one has $$\left(\mathcal{H}\vartheta_{2,7}\right)\Big|U(4)\Big|S_{7,r} = \frac{1}{4}\mathcal{D}.$$
3. For $r\equiv 5,6\pmod{7}$, one has $$\left(\mathcal{H}\vartheta_{3,7}\right)\Big|U(4)\Big|S_{7,r} = \frac{1}{4}\mathcal{D}.$$
In particular, Conjecture \[conj:7\] is true.
[99]{} K. Bringmann and A. Folsom,
Almost harmonic Maass forms and Kac Wakimoto characters,
J. reine und angew. Math., accepted for publication. K. Bringmann and S. Murthy,
On the positivity of black hole degeneracies in string theory
, Commun. Number Theory Phys., recommended for publication. K. Bringmann and K. Ono,
The $f(q)$ mock theta function conjecture and partition ranks
, Invent. Math.
165
(2006), 243–266. K. Bringmann and K. Ono,
Some characters of Kac and Wakimoto and nonholomorphic modular functions
, Math. Ann.
345
(2009), 547–558. K. Bringmann and K. Ono,
Dyson’s rank and Maass forms
, Ann. of Math.
171
(2010), 419–449. B. Brown, N. Calkin, T. Flowers, K. James, E. Smith, and A. Stout,
Elliptic curves, modular forms, and sums of Hurwitz class numbers
, J. Number Theory
128
(2008), 1847–1863. A. Dabholkar, S. Murthy, and D. Zagier,
Quantum black holes, wall crossing and mock modular forms,
submitted for publication. M. Eichler,
On the class of imaginary quadratic fields and sums of divisors of natural numbers,
J. Indian Math. Soc. **19** (1956), 153–180. F. Hirzebruch and D. Zagier,
Intersection numbers of curves on Hilbert modular surfaces and modular forms of Nebentypus,
Invent. Math. **36** (1976), 57–113. W. Kohnen,
Modular forms of half-integral weight on $\Gamma_0(4)$,
Math. Ann. **248** (1980), 249–266. W. Li,
Newforms and functional equations
, Math. Ann. **212** (1975), 285–315. K. Mahlburg,
Partition congruences and the Andrews-Garval-Dyson crank
, Proc. Natl. Acad. Sci. USA **102** (2005), 15373–15376. K. Ono,
The web of modularity: arithmetic of the coefficients of modular forms and $q$-series
, CBMS Regional Conference Series in Mathematics, 102. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 2004. W. Parry,
A negative result on the representation of modular forms by theta series
, J. Reine Angew. Math.
310
(1979), 151–170. A. Semikhatov, A. Taormina, and I. Tipunin,
Higher-level Appell functions, modular transformations, and characters
, Comm. Math. Phys. **255** (2005), 469–512. G. Shimura,
On modular forms of half integral weight,
Ann. of Math.
97
(1973), 440–481. D. Zagier,
Introduction to modular forms
in “From Number Theory to Physics”, Springer-Verlag, Heidelberg (1992), 238–291. S. Zwegers,
Multivariable Appell functions
, preprint. S. Zwegers,
Mock theta functions
, Ph.D. thesis, Utrecht University (2002).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The collision problem is to decide whether a function $X:\left\{
1,\ldots,n\right\} \rightarrow\left\{ 1,\ldots,n\right\} $ is one-to-one or two-to-one, given that one of these is the case. We show a lower bound of $\Omega\left( n^{1/5}\right) $ on the number of queries needed by a quantum computer to solve this problem with bounded error probability. The best known upper bound is $O\left( n^{1/3}\right) $, but obtaining any lower bound better than $\Omega\left( 1\right) $ was an open problem since 1997. Our proof uses the polynomial method augmented by some new ideas. We also give a lower bound of $\Omega\left( n^{1/7}\right) $ for the problem of deciding whether two sets are equal or disjoint on a constant fraction of elements. Finally we give implications of these results for quantum complexity theory.
author:
- 'Scott Aaronson[^1]'
title: Quantum Lower Bound for the Collision Problem
---
Introduction
============
The power of quantum computing has been intensively studied for a decade [@bv; @bbbv; @grover; @shor; @bbcmw; @ambainis; @simon]. Apart from possible applications—such as speeding up combinatorial search [@grover] and breaking public-key cryptography [@shor]—a major motivation for this work has been to better understand quantum theory itself. Thus, researchers have tried to discover not just the capabilities of quantum computing but also the limitations. This task is difficult, though; proving (for example) that quantum computers cannot solve $NP$-complete problems in polynomial time would imply $P\neq NP$.
A popular alternative is to study restricted models of computation, and particularly the *query model*, in which one counts only the number of queries to the input, not the number of computational steps. An early result of Bennett, Bernstein, Brassard, and Vazirani [@bbbv] showed that a quantum computer needs $\Omega\left( \sqrt{n}\right) $ queries to search a list of $n$ items for one marked item. (This bound is tight, as evidenced by Grover’s algorithm [@grover].) Subsequently, Beals et al. [@bbcmw], Ambainis [@ambainis], and others obtained lower bounds for many other problems.
But one problem, the collision problem, resisted attempts to prove a lower bound [@bdhhmsw; @ambainis]. Because of its simplicity, the problem was widely considered a benchmark for our understanding of quantum query complexity. The collision problem of size $n$, or $\operatorname*{Col}_{n}$, is defined as follows. Let $X=x_{1}\ldots x_{n}$ be a sequence of $n$ integers drawn from $\left\{ 1,\ldots,n\right\} $, with $n$ even. We are guaranteed that either
1. $X$ is one-to-one (that is, a permutation of $\left\{
1,\ldots,n\right\} $), or
2. $X$ is two-to-one (that is, each element of $\left\{
1,\ldots,n\right\} $ appears in $X$ twice or not at all).
The problem is to decide whether (1) or (2) holds.
We show that $Q_{2}\left( \operatorname*{Col}_{n}\right) =\Omega\left(
n^{1/5}\right) $, where $Q_{2}$ is bounded-error quantum query complexity as defined by Beals et al. [@bbcmw]. Details of the oracle model are given in Section \[prelim\]. The best known upper bound, due to Brassard, Høyer, and Tapp [@bht], is $O\left( n^{1/3}\right) $; thus, our bound is probably not tight. Previously, though, no lower bound better than the trivial $\Omega\left( 1\right) $ bound was known. How great a speedup quantum computers yield for the problem was apparently first asked by Rains [@rains].
Previous lower bound techniques failed for the problem because they depended on a function’s being sensitive to many disjoint changes to the input. For example, Beals et al. [@bbcmw] showed that for all total Boolean functions $f$, $Q_{2}\left( f\right) =\Omega\left( \sqrt{\operatorname*{bs}\left( f\right) }\right) $, where $\operatorname*{bs}\left( f\right)
$ is the block sensitivity, defined by Nisan [@nisan] to be, informally, the maximum number of disjoint changes (to any particular input $X$) to which $f$ is sensitive. In the case of the collision problem, though, every one-to-one input differs from every two-to-one input in at least $n/2$ places, so the block sensitivity is $O\left( 1\right) $. Ambainis’ adversary method [@ambainis], as currently formulated, faces a related obstacle. In that method we consider the algorithm and input as a bipartite quantum state, and upper-bound how much the *entanglement* of the state can increase via a single query. Yet under the simplest measures of entanglement, the algorithm and input can become highly entangled after $O\left( 1\right) $ queries, again because every one-to-one input is far from every two-to-one input.
Our proof is an adaptation of the polynomial method, introduced to quantum computing by Beals et al. [@bbcmw]. Their idea was to reduce questions about quantum algorithms to easier questions about multivariate polynomials. In particular, if a quantum algorithm makes $T$ queries, then its acceptance probability is a polynomial over the input bits of degree at most $2T$. So by showing that any polynomial approximating the desired output has high degree, one obtains a lower bound on $T$.
To lower-bound the degree of a multivariate polynomial, a key technical trick is to construct a related *univariate* polynomial. Beals et al. [@bbcmw], using a lemma due to Minsky and Papert [@mp], replace a polynomial $p\left( X\right) $ (where $X$ is a bit string) by $q\left(
\left| X\right| \right) $ (where $\left| X\right| $ denotes the Hamming weight of $X$), satisfying $$q\left( k\right) =\operatorname*{EX}\limits_{\left| X\right| =k}p\left(
X\right) $$ and $\deg\left( q\right) \leq\deg\left( p\right) $.
We construct the univariate polynomial in a different way. We consider a uniform distribution over $k$-to-one inputs, where $k$ might be greater than $2$. Even though the problem is to distinguish $k=1$ from $k=2$, the acceptance probability must lie between $0$ and $1$ for all $k$, and that is a surprisingly strong constraint. We show that the acceptance probability is *close* to a univariate polynomial in $k$ of degree at most $2T$. We then obtain a lower bound by generalizing a classical approximation theory result of Ehlich and Zeller [@ez] and Rivlin and Cheney [@rc]. Much of the proof deals with the complication that $k$ does not divide $n$ in general.
Shi [@shi] has recently improved our method to obtain a lower bound of $\Omega\left( n^{1/4}\right) $ for the collision problem.
The paper is organized as follows. Section \[motivation\] motivates the collision lower bound within quantum computing, pointing out connections to collision-resistant hash functions, the nonabelian hidden subgroup problem, and information erasure. Section \[prelim\] gives technical preliminaries, Section \[bivar\] proves the crucial fact that the acceptance probability is “almost” a univariate polynomial, and Section \[lb\] completes the lower bound argument. In Appendix \[setcomp\] we show a lower bound of $\Omega\left( n^{1/7}\right) $ for the *set comparison problem*, a variant of the collision problem that is needed for the application to information erasure.
Motivation\[motivation\]
========================
The most immediate implication of the collision lower bound is that certain problems, notably breaking cryptographic hash functions, are not in $BQP$ relative to an oracle. A second implication is that a nonstandard quantum oracle model proposed by Kashefi et al. [@kashefi] is exponentially more powerful than the usual oracle model. A third implication, in our view the most interesting one, concerns the computational power of so-called *dynamical quantum theories*. That implication will be discussed in detail in another paper.
Oracle Hardness Results
-----------------------
The original motivation for the collision problem was to model *(strongly) collision-resistant hash functions* in cryptography. There is a large literature on collision-resistant hashing; see [@damgard; @bsp] for example. When building secure digital signature schemes, it is useful to have a family of hash functions $\left\{
H_{i}\right\} $, such that finding a distinct $\left( x,y\right) $ pair with $H_{i}\left( x\right) =H_{i}\left( y\right) $ is computationally intractable. A quantum algorithm for finding collisions using $O\left(
\operatorname*{polylog}\left( n\right) \right) $ queries would render *all* hash functions insecure against quantum attack in this sense. (Shor’s algorithm [@shor] already renders hash functions based on modular arithmetic insecure.) Our result indicates that collision-resistant hashing might still be possible in a quantum setting.
The collision problem also models the *nonabelian hidden subgroup problem*, of which graph isomorphism is a special case. Given a group $G$ and subgroup $H\leq G$, suppose we have oracle access to a function $f:G\rightarrow\mathbb{N}$ such that for all $g_{1},g_{2}\in G$, $f\left(
g_{1}\right) =f\left( g_{2}\right) $ if and only if $g_{1}$ and $g_{2}$ belong to the same coset of $H$. Is there then an efficient quantum algorithm to determine $H$? If $G$ is abelian, the work of Simon [@simon] and Shor [@shor] implies an affirmative answer. If $G$ is nonabelian, though, efficient quantum algorithms are known only for special cases [@eh; @gsvv]. An $O\left( \operatorname*{polylog}\left( n\right)
\right) $-query algorithm for the collision problem would yield a polynomial-time algorithm to distinguish $\left| H\right| =1$ from $\left|
H\right| =2$, which does not exploit the group structure at all. Our result implies that no such algorithm exists.
Information Erasure\[erasure\]
------------------------------
Let $f:\left\{ 0,1\right\} ^{n}\rightarrow\left\{ 0,1\right\} ^{m}$ with $m\geq n$ be a one-to-one function. Then we can consider two kinds of quantum oracle for $f$:
1. a *standard oracle*, one that maps $\left| x\right\rangle
\left| z\right\rangle $ to $\left| x\right\rangle \left| z\oplus f\left(
x\right) \right\rangle $, or
2. an *erasing oracle* (as recently proposed by Kashefi et al. [@kashefi]), which maps $\left| x\right\rangle $ to $\left| f\left(
x\right) \right\rangle $, in effect “erasing” $\left| x\right\rangle $.
Intuitively erasing oracles seem at least as strong as standard ones, though it is not clear how to simulate the latter with the former without also having access to an oracle that maps $\left| y\right\rangle $ to $\left|
f^{-1}\left( y\right) \right\rangle $. The question that concerns us here is whether erasing oracles are *more* useful than standard ones for some problems. One-way functions provide a clue: if $f$ is one-way, then (by assumption) $\left| x\right\rangle \left| f\left( x\right) \right\rangle
$ can be computed efficiently, but if $\left| f\left( x\right)
\right\rangle $ could be computed efficiently given $\left| x\right\rangle
$ then so could $\left| x\right\rangle $ given $\left| f\left( x\right)
\right\rangle $, and hence $f$ could be inverted. But can we find, for some problem, an exponential gap between query complexity given a standard oracle and query complexity given an erasing oracle?
In Appendix \[setcomp\] we extend the collision lower bound to show an affirmative answer. Define the *set comparison problem* of size $n$, or $\operatorname*{SetComp}_{n}$, as follows. We are given as input two sequences, $X=x_{1}\ldots x_{n}$ and $Y=y_{1}\ldots y_{n}$, such that for each $i$, $x_{i},y_{i}\in\left\{ 1,\ldots,2n\right\} $. A query has the form $\left( b,i\right) $, where $b\in\left\{ 0,1\right\} \ $and $i\in\left\{
1,\ldots,n\right\} $, and produces as output $\left( 0,x_{i}\right) $ if $b=0$ and $\left( 1,y_{i}\right) $ if $b=1$. $\ $Sequences $X$ and $Y$ are both one-to-one; that is, $x_{i}\neq x_{j}$ and $y_{i}\neq y_{j}$ for all $i\neq j$. We are furthermore guaranteed that either
1. $X$ and $Y$ are equal as sets (that is, $\left\{ x_{1},\ldots,x_{n}\right\} =\left\{ y_{1},\ldots,y_{n}\right\} $) or
2. $X$ and $Y$ are far as sets (that is, $\left| \left\{
x_{1},\ldots,x_{n}\right\} \cup\left\{ y_{1},\ldots,y_{n}\right\} \right|
\geq1.1n$).
As before the problem is to decide whether (1) or (2) holds.
This problem can be solved with high probability in a constant number of queries using an erasing oracle, by using a trick similar to that of Watrous [@watrous] for verifying group non-membership. First, using the oracle, we prepare the uniform superposition$$\frac{1}{\sqrt{2n}}\sum_{i\in\left\{ 1,\ldots,n\right\} }\left( \left|
0\right\rangle \left| x_{i}\right\rangle +\left| 1\right\rangle \left|
y_{i}\right\rangle \right) \text{.}$$ We then apply a Hadamard gate to the first register, and finally we measure the first register. If $X$ and $Y$ are equal as sets, then interference occurs between every $\left( \left| 0\right\rangle \left| z\right\rangle
,\left| 1\right\rangle \left| z\right\rangle \right) $ pair and we observe $\left| 0\right\rangle $ with certainty. But if $X$ and $Y$ are far as sets, then basis states $\left| b\right\rangle \left| z\right\rangle $ with no matching $\left| 1-b\right\rangle \left| z\right\rangle $ have probability weight at least $1/10$, and hence we observe $\left|
1\right\rangle $ with probability at least $1/20$.
In Appendix \[setcomp\] we show that $Q_{2}\left( \operatorname*{SetComp}_{n}\right) =\Omega\left( n^{1/7}\right) $; that is, no efficient quantum algorithm using a standard oracle exists for this problem.
Preliminaries\[prelim\]
=======================
Let $A$ be a quantum query algorithm. A basis state of $A$ is written $\left| \Psi,i,z\right\rangle $. Then a query replaces each $\left|
\Psi,i,z\right\rangle $ by $\left| \Psi\oplus x_{i},i,z\right\rangle $, where $x_{i}$ is exclusive-OR’ed into some specified location of $\Psi$ (which we cannot assume to be all $0$’s). We assume without loss of generality that every basis state queries at every step. Between queries, the algorithm can perform any unitary operation that does not depend on the input. At the end $z$ is measured in the standard basis; if $z=1$ the algorithm returns ‘one-to-one’ and if $z=2$ it returns ‘two-to-one.’ The total number of queries is denoted $T$. Also, we assume for simplicity that all amplitudes are real; this restriction is without loss of generality [@bv].
Let $\alpha_{X,\Psi,i,z}^{\left( t\right) }$ be the amplitude of basis state $\left| \Psi,i,z\right\rangle $ after $t$ queries when the input is $X$. Also, let $\Delta\left( x_{i},h\right) =1$ if $x_{i}=h$, and $\Delta\left( x_{i},h\right) =0$ if $x_{i}\neq h$. Let $P\left( X\right)
$ be the probability that $A$ returns $z=2$ when the input is $X$. Then we obtain a simple variant of the main lemma of Beals et al. [@bbcmw].
\[poly\]$P\left( X\right) $ is a multilinear polynomial of degree at most $2T$ over the $\Delta\left( x_{i},h\right) $.
We show, by induction on $t$, that for all basis states $\left|
\Psi,i,z\right\rangle $, $\alpha_{X,\Psi,i,z}^{\left( t\right) }$ is a multilinear polynomial of degree at most $t$ over the $\Delta\left(
x_{i},h\right) $. Since $P\left( X\right) $ is a sum of squares of $\alpha_{X,\Psi,i,z}^{\left( t\right) }$, the lemma follows.
The base case ($t=0$) holds since, before making any queries, each $\alpha_{X,\Psi,i,z}^{\left( 0\right) }$ is a degree-$0$ polynomial over the $\Delta\left( x_{i},h\right) $. A unitary transformation on the algorithm part replaces each $\alpha_{X,\Psi,i,z}^{\left( t\right) }$ by a linear combination of $\alpha_{X,\Psi,i,z}^{\left( t\right) }$, and hence cannot increase the degree. Suppose the lemma holds prior to the $t^{th}$ query. Then$$\alpha_{X,\Psi,i,z}^{\left( t+1\right) }=\sum_{1\leq h\leq n}\alpha
_{X,\Psi\oplus h,i,z}^{\left( t\right) }\Delta\left( x_{i},h\right) ,$$ and we are done.
A remark on notation: we sometimes use brackets ($a_{b\left[ c\right] }$) rather than nested subscripts ($a_{b_{c}}$).
Reduction to Bivariate Polynomial\[bivar\]
==========================================
Call the point $\left( g,N\right) \in\Re^{2}$ an $\left( n,T\right)
$-*quasilattice point* if and only if
1. $g$ and $N$ are integers, with $g$ dividing $N$,
2. $1\leq g\leq\sqrt{n}$,
3. $n\leq N\leq n+n/\left( 10T\right) $, and
4. if $g=1$ then $N=n$.
For quasilattice point $\left( g,N\right) $, define $\mathcal{D}_{n}\left(
g,N\right) $ to be the uniform distribution over all size-$n$ subfunctions of $g$-1 functions having domain $\left\{ 1,\ldots,N\right\} $ and range a subset of $\left\{ 1,\ldots,n\right\} $. More precisely: to draw an $X$ from $\mathcal{D}_{n}\left( g,N\right) $, we first choose a set $S\subseteq\left\{ 1,\ldots,n\right\} $ with $\left| S\right| =N/g\leq
n$ uniformly at random. We then choose a $g$-1 function $\widehat
{X}=\widehat{x}_{1}\ldots\widehat{x}_{N}$ from $\left\{ 1,\ldots,N\right\}
$ to $S$ uniformly at random. Finally we let $x_{i}=\widehat{x}_{i}$ for each $1\leq i\leq n$.
Let $P\left( g,N\right) $ be the probability that algorithm $A$ returns $z=2$ when the input is chosen from $\mathcal{D}_{n}\left( g,N\right) $:$$P\left( g,N\right) =\operatorname*{EX}\limits_{X\in\mathcal{D}\left[
n\right] \left( g,N\right) }P\left( X\right) .$$ We then have the following surprising characterization:
\[univ\]For all sufficiently large $n$ and if $T\leq\sqrt{n}/3$, there exists a bivariate polynomial $q\left( g,N\right) $ of degree at most $2T$ such that if $\left( g,N\right) $ is a quasilattice point, then$$\left| P\left( g,N\right) -q\left( g,N\right) \right| <0.182$$ (where the constant $0.182$ can be made arbitrarily small by adjusting parameters).
Let $I$ be a product of $\Delta\left( x_{i},h\right) \ $variables, with degree $r\left( I\right) $, and let $I\left( X\right) \in\left\{
0,1\right\} $ be $I$ evaluated on input $X$. Then define$$\gamma\left( I,g,N\right) =\operatorname*{EX}\limits_{X\in\mathcal{D}\left[
n\right] \left( g,N\right) }I\left( X\right)$$ to be the probability that monomial $I$ evaluates to $1$ when the input is drawn from $\mathcal{D}_{n}\left( g,N\right) $. Then by Lemma \[poly\], $P\left( X\right) $ is a polynomial of degree at most $2T$ over $X$, so$$P\left( g,N\right) =\operatorname*{EX}\limits_{X\in\mathcal{D}\left[
n\right] \left( g,N\right) }P\left( X\right) =\operatorname*{EX}\limits_{X\in\mathcal{D}\left[ n\right] \left( g,N\right) }\sum
_{I:r\left( I\right) \leq2t}\beta_{I}I\left( X\right) =\sum_{I:r\left(
I\right) \leq2T}\beta_{I}\gamma\left( I,g,N\right)$$ for some coefficients $\beta_{I}$.
We now calculate $\gamma\left( I,g,N\right) $. Assume without loss of generality that for all $\Delta\left( x_{i},h_{1}\right) ,\Delta\left(
x_{j},h_{2}\right) \in I$, either $i\neq j$ or $h_{1}=h_{2}$, since otherwise $\gamma\left( I,g,N\right) =0$.
Define the “range” $Z\left( I\right) $ of $I$ to be the set of all $h$ such that $\Delta\left( x_{i},h\right) \in I$. Let $w\left( I\right)
=\left| Z\left( I\right) \right| $; then we write $Z\left( I\right)
=\left\{ z_{1},\ldots,z_{w\left( I\right) }\right\} $. $\ $Clearly $\gamma\left( I,g,N\right) =0$ unless $Z\left( I\right) \in S$, where $S$ is the range of $\widehat{X}$. By assumption,$$\frac{N}{g}\geq\frac{n}{\sqrt{n}}\geq2T\geq r\left( I\right)$$ so the number of possible $S$ is $\dbinom{n}{N/g}$ and, of these, the number that contain $Z$ is $\dbinom{n-w\left( I\right) }{N/g-w\left(
I\right) }$.
Then, conditioned on $Z\in S$, what is the probability that $\gamma\left(
I,g,N\right) =1$? The total number of $g$-1 functions with domain size $N$ is $N!/\left( g!\right) ^{N/g},$ since we can permute the $N$ function values arbitrarily, but must not count permutations that act only within the $N/g$ constant-value blocks of size $g$.
Among these functions, how many satisfy $\gamma\left( I,g,N\right) =1$? Suppose that, for each $1\leq j\leq w\left( I\right) $, there are $r_{j}\left( I\right) $ distinct $i$ such that $\Delta\left( x_{i},z_{j}\right) \in I$. Clearly$$r_{1}\left( I\right) +\cdots+r_{w\left( I\right) }\left( I\right)
=r\left( I\right) .$$ Then we can permute the $\left( N-r\left( I\right) \right) !$ function values outside of $I$ arbitrarily, but must not count permutations that act only within the $N/g$ constant-value blocks, which have size either $g$ or $g-r_{i}\left( I\right) $ for some $i$. So the number of functions for which $\gamma\left( I,g,N\right) =1$ is$$\frac{\left( N-r\left( I\right) \right) !}{\left( g!\right)
^{N/g-w\left( I\right) }{\displaystyle\prod\nolimits_{i=1}^{w\left( I\right) }}
\left( g-r_{i}\left( I\right) \right) !}.$$
Putting it all together, $$\begin{aligned}
\gamma\left( I,g,N\right) & =\frac{\dbinom{n-w\left( I\right)
}{N/g-w\left( I\right) }}{\dbinom{n}{N/g}}\cdot\frac{\left( N-r\left(
I\right) \right) !\left( g!\right) ^{N/g}}{\left( g!\right)
^{N/g-w\left( I\right) }N!{\displaystyle\prod\nolimits_{i=1}^{w\left( I\right) }}
\left( g-r_{i}\left( I\right) \right) !}\\
& =\frac{\left( N-r\left( I\right) \right) !\left( n-w\left( I\right)
\right) !\left( N/g\right) !}{N!n!\left( N/g-w\left( I\right) \right)
!}\cdot\frac{\left( g!\right) ^{w\left( I\right) }}{{\displaystyle\prod\nolimits_{i=1}^{w\left( I\right) }}
\left( g-r_{i}\left( I\right) \right) !}\\
& =\frac{\left( N-r\left( I\right) \right) !}{N!}\frac{\left( n-w\left(
I\right) \right) !}{n!}\cdot{\displaystyle\prod\limits_{i=0}^{w\left( I\right) -1}}
\left( \frac{N}{g}-i\right)
{\displaystyle\prod\limits_{i=1}^{w\left( I\right) }}
\left[ g{\displaystyle\prod\limits_{j=1}^{r\left[ i\right] \left( I\right) -1}}
\left( g-j\right) \right] \\
& =\frac{\left( N-2T\right) !}{N!}\frac{\left( n-w\left( I\right)
\right) !}{n!}\cdot{\displaystyle\prod\limits_{i=r\left( I\right) }^{2T-1}}
\left( N-i\right)
{\displaystyle\prod\limits_{i=0}^{w\left( I\right) -1}}
\left( N-gi\right)
{\displaystyle\prod\limits_{i=1}^{w\left( I\right) }}
{\displaystyle\prod\limits_{j=1}^{r\left[ i\right] \left( I\right) -1}}
\left( g-j\right) \\
& =\frac{\left( N-2T\right) !n!}{N!\left( n-2T\right) !}\widetilde
{q}_{n,T,I}\left( g,N\right)\end{aligned}$$ where$$\widetilde{q}_{n,T,I}\left( g,N\right) =\frac{\left( n-w\left( I\right)
\right) !\left( n-2T\right) !}{\left( n!\right) ^{2}}\cdot{\displaystyle\prod\limits_{i=r\left( I\right) }^{2T-1}}
\left( N-i\right)
{\displaystyle\prod\limits_{i=0}^{w\left( I\right) -1}}
\left( N-gi\right)
{\displaystyle\prod\limits_{i=1}^{w\left( I\right) }}
{\displaystyle\prod\limits_{j=1}^{r\left[ i\right] \left( I\right) -1}}
\left( g-j\right)$$ is a bivariate polynomial of total degree at most$$\left( 2T-r\left( I\right) \right) +w\left( I\right) +\left( r\left(
I\right) -w\left( I\right) \right) =2T.$$ (Note that in the case $r_{i}\left( I\right) >g$ for some $i$, this polynomial evaluates to $0$, which is what it ought to do.) Hence$$P\left( g,N\right) =\sum_{I:r\left( I\right) \leq2T}\beta_{I}\gamma\left(
I,g,N\right) =\frac{\left( N-2T\right) !n!}{N!\left( n-2T\right)
!}q\left( g,N\right)$$ where$$q\left( g,N\right) =\sum_{I:r\left( I\right) \leq2T}\beta_{I}\widetilde
{q}_{n,T,I}\left( g,N\right) .$$
Clearly$$\frac{\left( N-2T\right) !n!}{N!\left( n-2T\right) !}\leq1.$$ Since $N\leq n+n/\left( 10T\right) $ and $T\leq\sqrt{n}/3$, we also have$$\begin{aligned}
\frac{\left( N-2T\right) !n!}{N!\left( n-2T\right) !} & \geq\left(
\frac{n-2T+1}{N-2T+1}\right) ^{2T}\\
& \geq\left( 1-\frac{n/\left( 10T\right) }{N-2T+1}\right) ^{2T}\\
& \geq\left( 1-\frac{n}{10\left[ n-\left( 2T+1\right) /n\right]
}\frac{1}{T}\right) ^{2T}\\
& \geq\exp\left\{ -\frac{1}{5}\frac{n}{n-\left( 2T+1\right) /n}\right\} \\
& \geq0.818\end{aligned}$$ for all sufficiently large $n$.
Thus, since $0\leq P\left( g,N\right) \leq1$,$$\left| P\left( g,N\right) -q\left( g,N\right) \right| <0.182$$ and we are done.
Lower Bound\[lb\]
=================
We are now ready to prove a lower bound for the collision problem. To do so, we generalize an approximation theory result due to Rivlin and Cheney [@rc] and (independently) Ehlich and Zeller [@ez]. That result was applied to query complexity by Nisan and Szegedy [@ns] and later by Beals et al. [@bbcmw].
\[thetheorem\]$Q_{2}\left( \operatorname*{Col}_{n}\right) =\Omega\left(
n^{1/5}\right) .$
Let $g$ have range $1\leq g\leq G$. Then the quasilattice points $\left(
g,N\right) $ all lie in the rectangular region $R=\left[ 1,G\right]
\times\left[ n,n+n/\left( 10T\right) \right] $. Recalling the polynomial $q\left( g,N\right) $ from Lemma \[univ\], define$$d\left( q\right) =\max_{\left( g,N\right) \in R}\left( \max\left\{
\left| \frac{\partial q}{\partial g}\right| ,\frac{n}{10T\left( G-1\right)
}\cdot\left| \frac{\partial q}{\partial N}\right| \right\} \right) .$$ Suppose without loss of generality that we require $P\left( 1,n\right)
\leq1/10$ and $P\left( 2,n\right) \geq9/10$ (that is, algorithm $A$ distinguishes 1-1 from 2-1 functions with error probability at most $1/10$). Then, since$$\left| P\left( g,N\right) -q\left( g,N\right) \right| <0.182$$ by the Intermediate Value Theorem we have$$d\left( q\right) \geq\max_{1\leq g\leq2}\frac{\partial q}{\partial
g}>0.8-2\left( 0.182\right) =\allowbreak0.436.$$ $\,$
An inequality due to Markov (see [@cheney; @ns]) states that, for a univariate polynomial $p$, if $b_{1}\leq p\left( x\right) \leq b_{2}$ for all $a_{1}\leq x\leq a_{2}$, then$$\max_{a\left[ 1\right] \leq x\leq a\left[ 2\right] }\left|
\frac{dp\left( x\right) }{dx}\right| \leq\frac{b_{2}-b_{1}}{a_{2}-a_{1}}\deg\left( p\right) ^{2}.$$ Clearly for every point $\left( \widehat{g},\widehat{N}\right) \in R$, there exists a quasilattice point $\left( g,N\right) $ for which $\left|
g-\widehat{g}\right| \leq1$ and $\left| N-\widehat{N}\right| \leq G$. For take $g=\left\lceil \widehat{g}\right\rceil $—or, in the special case $\widehat{g}=1$, take $g=2$, since there is only one quasilattice point with $g=1$.
Furthermore, since $P\left( g,N\right) $ represents an acceptance probability at such a point, we have$$-0.182<q\left( g,N\right) <1.182.$$ Observe that for all $\left( \widehat{g},\widehat{N}\right) \in R$, $$-0.182-\left( \frac{10TG\left( G-1\right) }{n}+1\right) d\left( q\right)
<q\left( \widehat{g},\widehat{N}\right) <1.182+\left( \frac{10TG\left(
G-1\right) }{n}+1\right) d\left( q\right) .$$ For consider a quasilattice point close to $\left( \widehat{g},\widehat
{N}\right) $, and note that the maximum-magnitude derivative is at most $d\left( q\right) $ in the $g$ direction and $10T\left( G-1\right)
d\left( q\right) /n$ in the $N$ direction.
Let $\left( g^{\ast},N^{\ast}\right) $ be a point in $R$ at which the weighted maximum-magnitude derivative $d\left( q\right) $ is attained. Suppose first that the maximum is attained in the $g$ direction. Then $q\left( g,N^{\ast}\right) $ (with $N^{\ast}$ constant) is a univariate polynomial with$$\left| \frac{dq\left( g,N^{\ast}\right) }{dg}\right| >0.436$$ for some $1\leq g\leq G$. So$$\begin{aligned}
2T & \geq\deg\left( q\left( g,N\right) \right) \\
& \geq\deg\left( q\left( g,N^{\ast}\right) \right) \\
& \geq\sqrt{\frac{d\left( q\right) \left( G-1\right) }{1.364+2d\left(
q\right) \left( 1+10TG\left( G-1\right) /n\right) }}\\
& \geq\sqrt{\frac{0.436\left( G-1\right) n}{2.236n+8.720TG\left(
G-1\right) }}\\
& =\Omega\left( \min\left\{ \sqrt{G},\sqrt{\frac{n}{TG}}\right\} \right)
.\end{aligned}$$
Similarly, suppose the maximum $d\left( q\right) $ is attained in the $N$ direction. Then $q\left( g^{\ast},N\right) $ (with $g^{\ast}$ constant) is a univariate polynomial with$$\left| \frac{dq\left( g^{\ast},N\right) }{dN}\right| >\frac{0.436T\left(
G-1\right) }{n}$$ for some $n\leq N\leq n+n/\left( 10T\right) $. So$$2T\geq\sqrt{\frac{\left( 10T\left( G-1\right) /n\right) d\left( q\right)
n/\left( 10T\right) }{1.364+2d\left( q\right) \left( 1+10TG\left(
G-1\right) /n\right) }}\geq\Omega\left( \min\left\{ \sqrt{G},\sqrt{\frac{n}{TG}}\right\} \right) .$$
One can show that the lower bound on $T$ is optimized when we take $G=n^{2/5}\leq\sqrt{n}$. Then$$\begin{aligned}
T & =\Omega\left( \min\left\{ n^{1/5},\frac{\sqrt{n}}{\sqrt{T}n^{1/5}}\right\} \right) ,\\
T & =\Omega\left( n^{1/5}\right)\end{aligned}$$ and we are done.
Acknowledgments
===============
I am grateful to Yaoyun Shi, Ronald de Wolf, Umesh Vazirani, Ashwin Nayak, and Andris Ambainis for helpful comments; to Leonard Schulman, Lawrence Ip, Jordan Kerenidis, and John Preskill for discussions during earlier stages of this work; and to James Lee, Alex Halderman, and Elham Kashefi for discussions and references regarding Section \[motivation\].
[9]{} A. Ambainis. Quantum lower bounds by quantum arguments. *Proceedings of STOC’2000*, pages 636–643, 2000. Journal version to appear in *Journal of Computer and System Sciences*. quant-ph/0002066[^2].
R. Beals, H. Buhrman, R. Cleve, M. Mosca, and R. de Wolf. Quantum lower bounds by polynomials. *Proceedings of FOCS’98*, pages 352–361, 1998. quant-ph/9802049.
C. Bennett, E. Bernstein, G. Brassard, and U. Vazirani. Strengths and weaknesses of quantum computing. *SIAM Journal on Computing*, 26(5):1510–1523, 1997. quant-ph/9701001.
E. Bernstein and U. Vazirani. Quantum complexity theory. *SIAM Journal on Computing*, 26(5):1411–1473, 1997.
G. Brassard, P. Høyer, and A. Tapp. Quantum algorithm for the collision problem. *ACM SIGACT News (Cryptology Column)*, 28:14–19, 1997. quant-ph/9705002.
H. Buhrman, C. Dürr, M. Heiligman, P. Høyer, F. Magniez, M. Santha, and R. de Wolf. Quantum algorithms for element distinctness. *Proceedings of IEEE Conference on Computational Complexity (CCC’2001)*, pages 131–137, 2001. quant-ph/0007016.
S. Bakhtiari, R. Safavi-Naini, and J. Pieprzyk. Cryptographic hash functions: a survey. Technical Report 95-09, Department of Computer Science, University of Wollongong, July 1995. Available at ftp://ftp.cs.uow.edu.au/pub/papers/1995/tr-95-09.ps.Z.
E. W. Cheney. *Introduction to approximation theory*, McGraw-Hill, 1966.
I. B. Damgård. Collision free hash functions and public key signature schemes. *Proceedings of Eurocrypt’87*, Volume 304 of *Lecture Notes in Computer Science* (Springer-Verlag), 1988.
M. Ettinger and P. Høyer. On quantum algorithms for noncommutative hidden subgroups. *Advances in Applied Mathematics*, 25(3):239–251, 2000.
H. Ehlich and K. Zeller. Schwankung von Polynomen zwischen Gitterpunkten. *Mathematische Zeitschrift*, 86:41–44, 1964.
M. Grigni, L. Schulman, M. Vazirani, and U. Vazirani. Quantum mechanical algorithms for the nonabelian hidden subgroup problem. *Proceedings of STOC’2001*, pages 68–74, 2001.
L. K. Grover. A fast quantum mechanical algorithm for database search. *Proceedings of STOC’96*, pages 212–219, 1996. quant-ph/9605043.
E. Kashefi, A. Kent, V. Vedral, and K. Banaszek. On the power of quantum oracles, 2001. quant-ph/0109104.
M. Minsky and S. Papert. *Perceptrons*, MIT Press, 1988. First appeared in 1968.
N. Nisan. CREW PRAMs and decision trees. *SIAM Journal on Computing*, 20(6):999-1007, 1991.
N. Nisan and M. Szegedy. On the degree of Boolean functions as real polynomials. *Computational Complexity*, 4:301–313, 1994.
E. Rains. Talk given at AT&T, Murray Hill, New Jersey, on March 12, 1997.
T. J. Rivlin and E. W. Cheney. A comparison of Uniform Approximations on an interval and a finite subset thereof. *SIAM Journal on Numerical Analysis*, 3(2):311–320, 1966.
Y. Shi. Improving the lower bound on the collision problem to $n^{1/4}$. Manuscript, 2001.
P. Shor. Polynomial-time algorithms for prime factorization and discrete logarithms on a quantum computer. *SIAM Journal on Computing*, 26(5):1484–1509, 1997. quant-ph/9508027.
D. Simon. On the power of quantum computation. *Proceedings of FOCS’94*, pages 116–123, 1994.
J. Watrous. Succinct quantum proofs for properties of finite groups. *Proceedings of FOCS’2000*, pages 537–546, 2000. cs.CC/0009002.
Appendix: Set Comparison\[setcomp\]
===================================
Here we show that $Q_{2}\left( \operatorname*{SetComp}_{n}\right)
=\Omega\left( n^{1/7}\right) $, where $\operatorname*{SetComp}_{n}$ is the set comparison problem of size $n$ as defined in Section \[erasure\]. We need only redo the proof of Lemma \[univ\]; then Theorem \[thetheorem\] goes through largely unchanged.
The idea is the following. We need a distribution of inputs with a parameter $g$, such that the inputs are one-to-one when $g=1$ or $g=2$—since otherwise the problem of distinguishing $g=1$ from $g=2$ would be ill-defined for erasing oracles. On the other hand, the inputs must *not* be one-to-one for all $g>2$—since otherwise the lower bound for standard oracles would apply also to erasing oracles, and we could not obtain a separation between the two. Finally, the algorithm’s acceptance probability must be close to a polynomial in $g$.
Our solution is to consider $\kappa\left( g\right) $-to-one inputs, where $$\kappa\left( g\right) =4g^{2}-12g+9.$$ is a quadratic with $\kappa\left( 1\right) =\kappa\left( 2\right) =1$. The total range of the inputs (on sequences $X$ and $Y$ combined) has size roughly $n/g$; thus, we can tell the $g=1$ inputs apart from the $g=2$ inputs using an erasing oracle, even though $\kappa\left( g\right)
$ is the same for both. The disadvantage is that, because $\kappa\left(
g\right) $ increases quadratically rather than linearly in $g$, the quasilattice points become sparse more quickly. That is what weakens the lower bound from $\Omega\left( n^{1/5}\right) $ to $\Omega\left(
n^{1/7}\right) $. We note that, using the ideas of Shi [@shi], one can improve our lower bound on $Q_{2}\left( \operatorname*{SetComp}_{n}\right)
$ to $\Omega\left( n^{1/6}\right) $.
Call $\left( g,N,M\right) \in\Re^{3}$ an $\left( n,T\right)
$*-super-quasilattice point* if and only if
1. $g$ is an integer in $\left[ 1,n^{1/3}\right] $,
2. $N$ and $M$ are integers in $\left[ n,n\left( 1+1/\left(
100T\right) \right) \right] $,
3. $g$ divides $N$,
4. if $g=1$ then $N=n$,
5. $\kappa\left( g\right) $ divides $M$, and
6. if $g=2$ then $M=n$.
For super-quasilattice point $\left( g,N,M\right) $, we draw input $\left(
X,Y\right) =\left( x_{1}\ldots x_{n},y_{1}\ldots y_{n}\right) $ from distribution $\mathcal{L}_{n}\left( g,N,M\right) $ as follows. We first choose a set $S\subseteq\left\{ 1,\ldots,2n\right\} $ with $\left|
S\right| =2N/g\leq2n$ uniformly at random. We then choose two sets $S_{X},S_{Y}\subseteq S$ with $\left| S_{X}\right| =\left| S_{X}\right|
=M/\kappa\left( g\right) \leq\left| S\right| $, uniformly at random and independently. Next we choose $\kappa\left( g\right) $-1 functions $\widehat{X}=\widehat{x}_{1}\ldots\widehat{x}_{N}$ $:\left\{ 1,\ldots
,M\right\} \rightarrow S_{X}$ and $\widehat{Y}=\widehat{y}_{1}\ldots
\widehat{y}_{N}$ $:\left\{ 1,\ldots,M\right\} \rightarrow S_{Y}$ uniformly at random and independently. Finally we let $x_{i}=\widehat{x}_{i}$ and $y_{i}=\widehat{y}_{i}$ for each $1\leq i\leq n$.
Define sets $X_{S}=\left\{ x_{1},\ldots,x_{n}\right\} $ and $Y_{S}=\left\{
y_{1},\ldots,y_{n}\right\} $. Suppose $g=1$ and $N=M=n$; then by Chernoff bounds,$$\Pr_{\left( X,Y\right) \in\mathcal{L}\left[ n\right] \left( 1,n,n\right)
}\left[ \left| X_{S}\cup Y_{S}\right| <1.1n\right] \leq2e^{-n/10}.$$ Thus, if algorithm $A$ can distinguish $\left| X_{S}\cup Y_{S}\right|
=n$ from $\left| X_{S}\cup Y_{S}\right| \geq1.1n$ with probability at least $9/10$, then it can distinguish $\left( X,Y\right) \in\mathcal{L}_{n}\left( 1,n,n\right) $ from $\left( X,Y\right) \in\mathcal{L}_{n}\left( 2,n,n\right) $ with probability at least $9/10-2e^{-n/10}$. So a lower bound for the latter problem implies an equivalent lower bound for the former.
Define $P\left( X,Y\right) $ to be the probability that the algorithm returns that $X$ and $Y$ are far on input $\left( X,Y\right) $, and let$$P\left( g,N,M\right) =\operatorname*{EX}\limits_{\left( X,Y\right)
\in\mathcal{L}\left[ n\right] \left( g,N,M\right) }P\left( X,Y\right) .$$ We then have
\[univ2\]For all sufficiently large $n$ and if $T\leq n^{1/3}/8$, there exists a trivariate polynomial $q\left( g,N,M\right) $ of degree at most $8T$ such that if $\left( g,N,M\right) $ is a super-quasilattice point, then$$\left| P\left( g,N,M\right) -q\left( g,N,M\right) \right| <\varepsilon$$ for some constant $0<\varepsilon<1/2$.
By analogy to Lemma \[poly\], $P\left( X,Y\right) $ is a multilinear polynomial of degree at most $2T$ over variables of the form $\Delta\left(
x_{i},h\right) $ and $\Delta\left( y_{i},h\right) $. Let $I\left(
X,Y\right) =I_{X}\left( X\right) I_{Y}\left( Y\right) $ where $I_{X}$ is a product of $r_{X}\left( I\right) $ distinct $\Delta\left(
x_{i},h\right) \ $variables and $I_{Y}$ is a product of $r_{Y}\left(
I\right) \ $distinct$\ \Delta\left( y_{i},h\right) $ variables. Let $r\left( I\right) =r_{X}\left( I\right) +r_{Y}\left( I\right) $. Define$$\gamma\left( I,g,N,M\right) =\operatorname*{EX}\limits_{\left( X,Y\right)
\in\mathcal{L}\left[ n\right] \left( g,N,M\right) }I\left( X,Y\right) ;$$ then$$P\left( g,N,M\right) =\sum_{I:r\left( I\right) \leq2T}\beta_{I}\gamma\left( I,g,N,M\right)$$ for some coefficients $\beta_{I}$.
We now calculate $\gamma\left( I,g,N,M\right) $. As before we assume there are no pairs of variables $\Delta\left( x_{i},h_{1}\right) ,\Delta\left(
x_{i},h_{2}\right) \in I$ with $h_{1}\neq h_{2}$.
Let $Z_{X}\left( I\right) $ be the range of $I_{X}$ and let $Z_{Y}\left(
I\right) $ be the range of $I_{Y}$. Then let $Z\left( I\right)
=Z_{X}\left( I\right) \cup Z_{Y}\left( I\right) $. Let $w_{X}\left(
I\right) =\left| Z_{X}\left( I\right) \right| $, $w_{Y}\left( I\right)
=\left| Z_{Y}\left( I\right) \right| $, and $w\left( I\right) =\left|
Z\left( I\right) \right| $. By assumption$$\frac{N}{g}\geq\frac{M}{\kappa\left( g\right) }\geq\frac{1}{4}n^{1/3}\geq2T$$ so$$\Pr\left[ Z\left( I\right) \subseteq S\right] =\frac{\dbinom{2n-w\left(
I\right) }{2N/g-w\left( I\right) }}{\dbinom{2n}{2N/g}}.$$ Then the probability that $Z_{X}\left( I\right) \subseteq S_{X}$ given $Z\left( I\right) \subseteq S$ is$$\frac{\dbinom{2N/g-w_{X}\left( I\right) }{M/\kappa\left( g\right)
-w_{X}\left( I\right) }}{\dbinom{2N/g}{M/\kappa\left( g\right) }}$$ and similarly for the probability that $Z_{Y}\left( I\right) \subseteq
S_{Y}$ given $Z\left( I\right) \subseteq S$.
Let $r_{X,1}\left( I\right) ,\ldots,r_{X,w\left[ X\right] \left(
I\right) }\left( I\right) $ be the multiplicities of the range elements in $Z_{X}\left( I\right) $, so that $r_{X,1}\left( I\right) +\cdots
+r_{X,w\left[ X\right] \left( I\right) }\left( I\right) =r_{X}\left(
I\right) $. Then$$\Pr\left[ I_{X}\left( X\right) \,\,|\,\,Z_{X}\left( I\right) \subseteq
S_{X}\right] =\frac{\left( M-r_{X}\left( I\right) \right) !}{M!}{\displaystyle\prod\limits_{i=1}^{w\left[ X\right] \left( I\right) }}
{\displaystyle\prod\limits_{j=0}^{r\left[ X,i\right] \left( I\right) -1}}
\left( \kappa\left( g\right) -j\right)$$ and similarly for $\Pr\left[ I_{Y}\left( Y\right) \,\,|\,\,Z_{Y}\left(
I\right) \subseteq S_{Y}\right] $.
Putting it all together,$$\begin{aligned}
\gamma\left( I,g,N,M\right) = & \frac{\dbinom{2n-w\left( I\right)
}{2N/g-w\left( I\right) }}{\dbinom{2n}{2N/g}}\frac{\left( M-r_{X}\left(
I\right) \right) !}{M!}\frac{\left( M-r_{Y}\left( I\right) \right)
!}{M!}\frac{\dbinom{2N/g-w_{X}\left( I\right) }{M/\kappa\left( g\right)
-w_{X}\left( I\right) }}{\dbinom{2N/g}{M/\kappa\left( g\right) }}\times\\
&
{\displaystyle\prod\limits_{i=1}^{w\left[ X\right] \left( I\right) }}
{\displaystyle\prod\limits_{j=0}^{r\left[ X,i\right] \left( I\right) -1}}
\left( \kappa\left( g\right) -j\right) \frac{\dbinom{2N/g-w_{Y}\left(
I\right) }{M/\kappa\left( g\right) -w_{Y}\left( I\right) }}{\dbinom
{2N/g}{M/\kappa\left( g\right) }}{\displaystyle\prod\limits_{i=1}^{w\left[ Y\right] \left( I\right) }}
{\displaystyle\prod\limits_{j=0}^{r\left[ Y,i\right] \left( I\right) -1}}
\left( \kappa\left( g\right) -j\right) \\
= & \frac{\left( 2n-w\left( I\right) \right) !}{\left( 2n\right)
!}\frac{\left( M-r_{X}\left( I\right) \right) !}{M!}\frac{\left(
M-r_{Y}\left( I\right) \right) !}{M!}\frac{\left( 2N/g-w_{X}\left(
I\right) \right) !}{\left( 2N/g-w\left( I\right) \right) !}\frac{\left(
2N/g-w_{Y}\left( I\right) \right) !}{\left( 2N/g\right) !}\theta
_{I}\left( g,M\right)\end{aligned}$$ where$$\begin{aligned}
& \theta_{I}\left( g,M\right) =\\
&
{\displaystyle\prod\limits_{i=0}^{w\left[ X\right] \left( I\right) -1}}
\left( M-i\kappa\left( g\right) \right)
{\displaystyle\prod\limits_{i=1}^{w\left[ X\right] \left( I\right) }}
{\displaystyle\prod\limits_{j=1}^{r\left[ X,i\right] \left( I\right) -1}}
\left( \kappa\left( g\right) -j\right)
{\displaystyle\prod\limits_{i=0}^{w\left[ Y\right] \left( I\right) -1}}
\left( M-i\kappa\left( g\right) \right)
{\displaystyle\prod\limits_{i=1}^{w\left[ Y\right] \left( I\right) }}
{\displaystyle\prod\limits_{j=1}^{r\left[ Y,i\right] \left( I\right) -1}}
\left( \kappa\left( g\right) -j\right)\end{aligned}$$ is a bivariate polynomial in $\left( g,M\right) $ of total degree at most $2r\left( I\right) $.
Thus$$\begin{aligned}
\gamma\left( I,g,N,M\right) = & \frac{\left( 2n-w\left( I\right)
\right) !}{\left( 2n\right) !}\left[ \frac{\left( M-2T\right)
!n!}{M!\left( n-2T\right) !}\right] ^{2}\left( \frac{\left( n-2T\right)
!}{n!}\right) ^{2}\prod_{i=r\left[ X\right] \left( I\right) }^{2T-1}\left( M-i\right) \prod_{i=r\left[ Y\right] \left( I\right)
}^{2T-1}\left( M-i\right) \times\\
& \frac{\left( 2N/g-w_{X}\left( I\right) \right) \cdots\left(
2N/g-\left( w\left( I\right) -1\right) \right) }{\left( 2N/g\right)
\left( 2N/g-1\right) \cdots\left( 2N/g-\left( w_{Y}\left( I\right)
-1\right) \right) }\theta_{I}\left( g,M\right) \\
= & \frac{\left( 2n\right) ^{2T}}{\left( 2N\right) \left( 2N-g\right)
\cdots\left( 2N-\left( 2T-1\right) g\right) }\left[ \frac{\left(
M-2T\right) !n!}{M!\left( n-2T\right) !}\right] ^{2}\widetilde{q}_{n,T,I}\left( g,N,M\right)\end{aligned}$$ where$$\begin{aligned}
\widetilde{q}_{n,T,I}\left( g,N,M\right) = & \frac{\left( 2n-w\left(
I\right) \right) !}{\left( 2n\right) !\left( 2n\right) ^{2T}}\left(
\frac{\left( n-2T\right) !}{n!}\right) ^{2}g^{w\left[ X\right] \left(
I\right) +w\left[ Y\right] \left( I\right) -w\left( I\right) }\theta_{I}\left( g,M\right) \prod_{i=r\left[ X\right] \left( I\right)
}^{2T-1}\left( M-i\right) \times\\
& \prod_{i=r\left[ Y\right] \left( I\right) }^{2T-1}\left( M-i\right)
\prod_{i=w\left[ X\right] \left( I\right) }^{w\left( I\right) -1}\left(
2N-ig\right) \prod_{i=w\left[ Y\right] \left( I\right) }^{2T-1}\left(
2N-ig\right)\end{aligned}$$ is a trivariate polynomial in $\left( g,N,M\right) $ of total degree at most$$\left( 4T-r\left( I\right) \right) +2r\left( I\right) +\left(
w_{X}\left( I\right) +w_{Y}\left( I\right) -w\left( I\right) \right)
+\left( w\left( I\right) -w_{X}\left( I\right) \right) +\left(
2T-w_{Y}\left( I\right) \right) \leq8T.$$
Thus$$P\left( g,N,M\right) =\frac{\left( 2n\right) ^{2T}}{{\displaystyle\prod\limits_{i=0}^{2T-1}}
\left( 2N-gi\right) }\left[ \frac{\left( M-2T\right) !n!}{M!\left(
n-2T\right) !}\right] ^{2}q\left( g,N,M\right)$$ where $q\left( g,N,M\right) $ is a polynomial of total degree at most $8T$. The argument that $q$ approximates $P$ to within a constant is analogous to that of Lemma \[univ\]; note that$$\left| \frac{\left( 2n\right) ^{2T}}{{\displaystyle\prod\limits_{i=0}^{2T-1}}
\left( 2N-gi\right) }-1\right| =O\left[ \left( 1+\frac{1}{T}+\frac{gT}{n}\right) ^{2T}\right] =O\left( 1\right)$$ since $g\leq n^{1/3}\ $and $T\leq n^{1/3}/8$.
\[thetheorem2\]$Q_{2}\left( \operatorname*{SetComp}_{n}\right)
=\Omega\left( n^{1/7}\right) $.
\[Proof sketch\]The proof is analogous to that of Theorem \[thetheorem\]. Let $g\in\left[ 1,G\right] $ for some $G\leq n^{1/3}$. Then the super-quasilattice points $\left( g,N,M\right) $ all lie in $R=\left[
1,G\right] \times\left[ n,n+n/\left( 100T\right) \right] ^{2}$. Define $d(q)$ to be$$\max_{\left( g,N,M\right) \in R}\left( \max\left\{ \left| \frac{\partial
q}{\partial g}\right| ,\frac{n/100T}{\left( G-1\right) }\left|
\frac{\partial q}{\partial N}\right| ,\frac{n/100T}{\left( G-1\right)
}\left| \frac{\partial q}{\partial M}\right| \right\} \right) .$$ Then $d\left( q\right) \geq\delta$ for some constant $\delta>0$, by Lemma \[univ2\].
For every point $\left( \widehat{g},\widehat{N},\widehat{M}\right) \in R$, there exists a super-quasilattice point $\left( g,N,M\right) $ such that $\left| g-\widehat{g}\right| \leq1$, $\left| N-\widehat{N}\right| \leq G$, and $\left| M-\widehat{M}\right| \leq\kappa\left( G\right) $. Hence, $q\left( \widehat{g},\widehat{N},\widehat{M}\right) $ can deviate from $\left[ 0,1\right] $ by at most$$O\left( \left( \frac{TG^{3}}{n}+1\right) d\left( q\right) \right) .$$
Let $\left( g^{\ast},N^{\ast},M^{\ast}\right) $ be a point in $R$ at which $d\left( q\right) $ is attained. Suppose $d\left( q\right) $ is attained in the $g$ direction; the cases of the $N$ and $M$ directions are analogous. Then $q\left( g,N^{\ast},M^{\ast}\right) $ is a univariate polynomial in $g$, and$$\begin{aligned}
8T & \geq\deg\left( q\left( g,N^{\ast},M^{\ast}\right) \right) \\
& =\Omega\left( \sqrt{\frac{d\left( q\right) G}{1+d\left( q\right)
+d\left( q\right) TG^{3}/n}}\right) \\
& =\Omega\left( \min\left\{ \sqrt{G},\sqrt{\frac{n}{TG^{2}}}\right\}
\right) .\end{aligned}$$
One can show that the bound is optimized when we take $G=n^{2/7}\leq n^{1/3}$. Then$$\begin{aligned}
T & =\Omega\left( \min\left\{ n^{1/7},\frac{\sqrt{n}}{\sqrt{T}n^{2/7}}\right\} \right) ,\\
T & =\Omega\left( n^{1/7}\right) .\end{aligned}$$
[^1]: Computer Science Department, University of California, Berkeley, CA 94720-1776. Email: `[email protected]`. Supported in part by a National Science Foundation Graduate Fellowship and by the Institute for Quantum Information at the California Institute of Technology.
[^2]: Available at `www.arxiv.org`.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The hyperfine structure of the interstitial muonium (Mu) in rutile (TiO$_2$, weakly $n$-type) has been identified by means of a muon spin rotation technique. The angle-resolved hyperfine parameters exhibit a tetragonal anisotropy within the $ab$ plane and axial anisotropy with respect to the $\langle 001\rangle$ ($\hat{c}$) axis. This strongly suggests that the Mu is bound to O (forming an OH bond) at an off-center site within a channel along the $\hat{c}$ axis, while the unpaired Mu electron is localized around the neighboring Ti site. The hyperfine parameters are quantitatively explained by a model that considers spin polarization of the unpaired electron at both the Ti and O sites, providing evidence for the formation of Mu as a Ti-O-Mu complex paramagnetic state. The disappearance of the Mu signal above $\sim$10 K suggests that the energy necessary for the promotion of the unpaired electron to the conduction band by thermal activation is of the order of $10^1$ meV. These observations suggest that, while the electronic structure of Mu (and hence H) differs considerably from that of the conventional shallow level donor described by the effective mass model, Mu supplies a loosely bound electron, and thus, serves as a donor in rutile.'
author:
- 'K. Shimomura'
- 'R. Kadono'
- 'A. Koda'
- 'K. Nishiyama'
- 'M. Mihara'
title: 'Electronic structure of Mu-complex donor state in rutile TiO$_2$'
---
[^1]
Introduction
============
Hydrogen (H) is a ubiquitous impurity in a wide variety of semiconductors, including elemental (e.g., Si) and binary compound (e.g., GaAs) materials, which comprise the primary basis for current industrial applications. In these covalent systems, H is known to be amphoteric, forming an acceptor or donor level in $n$- or $p$-type materials, respectively [@Patterson:88; @Pankove:90; @Myers:92; @Chow:98]. Meanwhile, a recent theoretical prediction that H could behave as an independent electron donor to induce $n$-type conductivity in ZnO [@Walle:00], along with subsequent support from various experiments involving muon spin rotation ([$\mu$SR]{}) [@Cox:01; @Shimomura:02], electron paramagnetic resonance (EPR), electron-nuclear double resonance (ENDOR) measurements [@Hofmann:02], and additional conventional techniques [@Gil:99; @Shi:04; @Vilao:05], have prompted the development of a generalized hypothesis of H as a shallow donor dopant in wide gap semiconductors [@Kilic:02; @Walle:03; @Peacock:03].
As in the case of ZnO, [$\mu$SR]{} studies have made a significant contribution to this subject by providing spectroscopic information on muonium (Mu; an analog of a neutral H atom where the proton is replaced by a positive muon) observed in a wide variety of semiconductors. While the dynamical aspects (e.g., diffusion) of Mu and H may differ considerably, because of the relatively low mass of Mu ($m_\mu \simeq m_p/9$, with $m_p$ being the proton mass), the local electronic structure of Mu is virtually equivalent to that of H if a small correction to account for the difference in the reduced mass ($\sim$0.5%) is made. The recent discovery of novel Mu states having extremely small hyperfine parameters ($\sim$10$^{-4}$ times smaller than that of Mu in vacuum) and low ionization energy ($\sim$10$^1$ meV) in several compound semiconductors, including CdS [@Gil:99], ZnSe [@Vilao:05], InN [@Davis:03], and GaN [@Shimomura:04], also supports the hypothesis that Mu (and hence H) can act as a donor in these compounds.
The generalized hypothesis also predicts shallow donor H in various metal oxides, most notably in rutile [TiO$_2$]{}, where the origin of the unintentional $n$-type conductivity exhibited by the as-grown crystal is the focus of intense research. While the electrical and optical properties of [TiO$_2$]{} are subject to both intrinsic and extrinsic defects, this material has a rich variety of potential applications in both electronic and opto-electronic devises, and detailed understanding of the H behavior is crucially important to the material functionality in this context.
As shown in Fig. \[fig1\](a), rutile has a tetragonal structure with Ti$^{4+}$ ions occupying the body-center position, where each Ti ion is coordinated by six ligand oxygen atoms (O$^{2-}$) comprising a slightly distorted octahedron. The TiO$_6$ octahedra are linked by corner-sharing to form O chains parallel to the $\hat{c}$ axis. Regions of low electron density exist between the chains, forming the so-called “$c$ channel". Earlier studies have shown that interstitial H resides in the $c$ channel, leading to the formation of an OH bond (with a length of 0.109 nm) perpendicular to the $\hat{c}$ axis [@Anderson:73]. The open channels cause anisotropy in the diffusion process, and may allow fast diffusion of smaller ions parallel to the tetragonal axis. It has recently been reported that the H diffusion can be strongly enhanced via irradiation of the crystal by infrared light that matches the OH bond stretch mode [@Spahr:10].
![(Color online) (a) Schematic illustration of rutile [TiO$_2$]{} crystal structure, showing a unit cell with $a=b=0.4593$ nm and $c=0.2959$ nm. (b) Atomic position of interstitial hydrogen (H), where H forms OH bonds (0.109 nm) at four inequivalent H sites. (c)–(e) Signs of hyperfine parameters ($A_i$) expected for the magnetic dipolar field ($B_{\rm dip}$) generated by an electron spin localized on Ti atom under an external field ($\vec{B}$). (see text).[]{data-label="fig1"}](tio2-fig1.eps){width="0.95\linewidth"}
Here, we report on the electronic structure of Mu in single-crystalline rutile. Although the existence of a Mu state with an extremely small hyperfine (HF) parameter has been reported elsewhere [@Shimomura:05; @Cox:06], little is known about the local electronic structure. In this study, we show that the HF parameters exhibit strong anisotropy within the $ab$ plane as well as along the $\hat{c}$ axis. The angular dependence of the HF parameters is perfectly in line with that expected for Mu occupying the sites that are common to the interstitial H and forming OMu bonds. More interestingly, the relative signs of these parameters indicate that the HF interaction is predominantly determined by the magnetic dipolar fields generated by the unpaired Mu electron located away from the muon site. This is qualitatively consistent with a recent report on the H-related paramagnetic center, in which the HF interaction is attributed to the Ti$^{3+}$\[OH\]$^-$ complex state with the $d$ electron being located on the Ti ion nearest to the OH base [@Brant:11]. Meanwhile, a detailed comparison of the Mu and H HF structures reveals a distinct difference between the complex centers, where a contribution from the residual spin polarization at the nearest neighboring O site (antiparallel to that of the Ti ion) is suggested in the Mu case. Through consideration of the temperature ($T$) dependence of the Mu yield, we also show that the unpaired electron bound to the Mu complex state requires small energy for promotion to the conduction band. Based on these observations, we argue that interstitial Mu/H is one of the primary origins of unintentional $n$-type conductivity in rutile [TiO$_2$]{}. This is despite its complex electronic structure, which differs significantly from that for the conventional effective-mass-like donor.
Experiment
==========
A conventional [$\mu$SR]{} experiment was conducted at the former Booster Meson Facility in Tsukuba (at KEK) and at J-PARC MUSE. A 4-MeV muon beam that was almost 100% spin-polarized parallel to the beam direction was implanted into single-crystalline rutile wafers \[30 mm $\times$ 30 mm, 1.0-mm thick, grown using the Verneuil process\], which were “as received" from a local vendor (Furuuchi Chemical Co.). The definitions of the azimuthal and polar angles ($\phi$, $\theta$, respectively) for the external magnetic field ($\vec{B}$) are shown in Fig. \[fig1\](a). For the $\phi$-dependence measurements, a wafer having a normal axis orientation of $\langle001\rangle$ ($\parallel \hat{c}$) was used, where the initial muon spin polarization ($\vec{P}_\mu$) was parallel to $\hat{c}$, and the $\hat{a}$ axis was rotated contrary to the $\vec{B}$ direction within the $ab$ plane (i.e., $\theta=90^\circ$). Meanwhile, the $\theta$-dependence was measured by tilting the $\vec{B}$ direction towards the $\hat{c}$ axis (except for $\theta=0$, for which another $\langle100\rangle$ crystal was used), where the angle was tuned by an additional field parallel to the $\hat{c}$ axis. In both cases, the $\vec{B}$-parallel component of the HF parameter was observed in terms of the frequency shift of the satellite signals.
Results
=======
Within the $T$ range of the present measurements (2 K $\le T\le$ 290 K), a single diamagnetic state ($\mu^+=$ Mu$^+$) was observed above $\sim$10 K as a signal precessing with a frequency $\nu_0=\gamma_\mu B/2\pi$ (comprising the central line), where $\gamma_\mu$ ($= 2\pi\times135.53$ MHz/T) is the muon gyromagnetic ratio and $B=|\vec{B}|$. The depolarization rate for this signal was almost independent of $T$ with a Gaussian damping ($\sim$0.02 $\mu$s$^{-1}$); this was unambiguously attributed to random local fields exerted by the nuclear magnetic moments of the $^{47}$Ti (natural abundance: 7.4%) and $^{49}$Ti (5.4%) nuclei. Meanwhile, the [$\mu$SR]{} time spectra exhibited a remarkable change below $\sim$10 K. A typical example (observed at 5 K, $B\simeq32.5$ mT) is shown in Fig. \[fig2\](a), where a beat pattern due to multiple frequency components is clearly observed in the signal amplitude.
![(Color online) (a) Example of [$\mu$SR]{} time spectrum observed in [TiO$_2$]{} at 5 K under an applied magnetic field of 32.5 mT, where the field ($\vec{B}$) is rotated from $\langle100\rangle$ axis ($\hat{a}$) by $\phi=30^\circ$ around the $\hat{c}$ axis ($\theta=90^\circ$). Solid curve shows a result of curve fit. (b) Fast Fourier-transform (FFT) of the data shown in (a). (c) FFT of the time spectra with $\vec{B}\parallel \hat{c}$ at 5 K. []{data-label="fig2"}](tio2-fig2.eps){width="0.95\linewidth"}
As is evident in Fig. \[fig2\](b), one can trace two pairs of satellite lines in the fast Fourier-transformed spectra ($\nu_{i\pm}$, $i=1$, 2), which are situated almost symmetrically around the central line ($\nu_0\simeq4.41$ MHz). A preliminary analysis showed that the least-square curve fits of the [$\mu$SR]{} time spectra, assuming one pair of satellites, did not reproduce the data for those obtained with $\vec{B}$ oriented away from the $\langle100\rangle$ direction, yielding a poor $\chi^2$. However, fits with two sets of satellites were found to yield a satisfactory result with significantly improved $\chi^2$ in the majority of cases. It was also found that the splitting ($\nu_{i+}-\nu_{i-}$) of these satellites remained unchanged when the applied field was reduced to 10 mT, indicating that $B$ was in the high field-limit range, where the splitting was independent of the $B$ value. These observations clearly indicate that a significant fraction of implanted muons form a paramagnetic Mu state characterized by an extremely small HF parameter (almost 10$^{-4}$ times smaller than the vacuum value of $4.463\times10^3$ MHz) with significant anisotropy in the electronic structure.
The results of curve fits for the time spectra assuming multiple precession frequency components ($\nu_{i\pm}$) are summarized in Fig. \[fig3\], for both (a) $\phi$ and (b) $\theta$ dependence. (We assumed that the signal amplitudes of $\nu_{i+}$ and $\nu_{i-}$ were common for each pair in the curve fits.) As can be clearly seen in Fig. \[fig3\](a), $\nu_{i\pm}$ exhibits a strong anisotropy with a characteristic angle dependence that is symmetric over $\phi\simeq45^\circ$. Considering the fact that there are four inequivalent H (Mu) sites with different principal axis orientations for anisotropic HF parameters \[see Fig. \[fig1\](b)\], four sets of satellite pairs discerned by different $\phi$-dependence can be expected. More specifically, it is predicted that the $\phi$ dependence of the HF parameters can be expressed in the form $$\begin{aligned}
\nu_{i\pm}(\phi)&\simeq&\nu_0\pm\frac{1}{2}A_\perp(\phi),\label{mufrq}\\
A_{\perp}(\phi)&=&A_1\cos^2\Phi+A_2\sin^2\Phi,\label{Aperp}\end{aligned}$$ where $\Phi=\phi\pm\phi_0$ or $\phi\pm\phi_0\pm\frac{\pi}{2}$, and $\phi_0$ is the offset angle from the $\hat{a}$ axis. It must be noted that the double sign correspondence between $\nu_{i\pm}$ and the right hand side of Eq. (\[mufrq\]) is not unique, because the sign of $A_{\perp}(\phi)$ depends on that of $A_i$ and, thus, can change over the $\phi$ range. While our measurements are insensitive to the absolute sign of $A_i$, they can determine the relative sign between $A_1$ and $A_2$. The solid curves shown in Fig. \[fig3\](a) are the result of a simultaneous fit for those curves using common values of $A_i$ and $\phi_0$, where $A_1A_2<0$ is inferred. This strongly suggests that the HF interaction is predominantly determined by the magnetic dipolar field generated by the off-site electron spin(s). (This scenario closely resembles the case of Mu-substituted free radicals often observed in unsaturated organic compounds.) Moreover, considering that Eqs. (\[mufrq\]) and (\[Aperp\]) predict eight lines when $|A_1|\neq |A_2|$, Fig. \[fig3\](a) indicates that $|A_1|$ is almost equal to $|A_2|$ within the frequency resolution, which is limited by the observation time range ($\sim$0.1 MHz). The obtained values for $A_i$ and $\phi$ are shown in Table \[tab1\]. Hereafter, we assume that $A_1<0$ (see below).
The $\theta$ dependence is shown in Fig. \[fig3\](b), and is characterized by a narrow splitting of $\nu_{1\pm}$ at $\theta=0$ \[see Fig. \[fig2\](c)\], along with a monotonous increase with increasing $\theta$. Only one pair of satellites has been identified in this data set, most likely because of the narrow splitting relative to the frequency resolution. It is also notable that $|\nu_{1-}-\nu_0|$ tends to be smaller than $|\nu_{1+}-\nu_0|$ for some unknown reason. We tentatively attribute this to systematic uncertainty in the present measurement, and adopt $\nu_{1+}$ in order to deduce the HF parameters using a form similar to Eq. (\[Aperp\]), such that $$A_\parallel(\theta) = A_3\cos^2 \theta+A_\perp(0)\sin^2 \theta,$$ where the sign of $A_3$ is left undetermined. The value of $A_3$ deduced from the curve fit is shown in Table I, where $|A_\perp(0)|$ has been determined to be 1.314(15) MHz (which is in perfect agreement with the value deduced from the $\phi$ dependence).
![(Color online) (a) Angle dependence of muon spin precession frequency observed at 5 K for (a) azimuthal angle ($\phi$) and (b) polar angle ($\theta$). Filled symbols: $\nu_{i\pm}$ satellite lines, open symbols: $\nu_0$ central line. The solid curves are obtained via curve fits using general forms of the anisotropic hyperfine parameters, while the dashed curves are calculated for a given set of parameters (see text).[]{data-label="fig3"}](tio2-fig3.eps){width="0.8\linewidth"}
Mu $A_1$ (MHz) $A_2$ (MHz) $A_3$ (MHz) $\phi_0$ (deg)
------------------------- --------------- --------------- --------------- ----------------
$A_{\perp}(\phi)$ $-1.29(6)$ $+1.29(6)$ – $25.5(1.4)$
$A_{\parallel}(\theta)$ – – $-0.17(2)$ –
H -1.276(3)$^*$ +1.961(3)$^*$ -1.076(3)$^*$ 22.1
: Mu hyperfine parameters deduced from curve fits of data shown in Fig. \[fig3\] (where $|A_1|=|A_2|$ was assumed). The $A_i$’s for the H-related center [@Brant:11] are scaled against the gyromagnetic ratio of $\mu^+$ (indicated by an asterisk, see text). The $A_i$ signs for Mu are assumed to be common to those for H. \[tab1\]
Discussion
==========
Electronic Structure
--------------------
It would be intriguing to compare the observed electronic structure with that of a H-related paramagnetic center observed via EPR/ENDOR in a “lightly reduced" rutile sample [@Brant:11]. According to the proposed model, the center consists of a substitutional Ti$^{3+}$ ion adjacent to a substitutional \[OH\]$^{-}$ molecular ion (which is identical to the atomic configuration shown in Fig. \[fig1\], apart from the Ti valency). The HF parameters for the electron-$^1$H nucleus (proton) interaction determined by the ENDOR spectra are reported to be $A_1=-0.401(1)$ MHz, $A_2=+0.616(1)$ MHz, and $A_3=-0.338(1)$ MHz with their principal axes being displaced from the $\langle110\rangle$ or $\langle1\overline{1}0\rangle$ axes by 22.9$^\circ$ (or by 22.1$^\circ$ from the $\hat{a}$ axis). As illustrated in Figs. \[fig1\](c)–(e), the $A_i$ signs are consistent with the magnetic dipolar field generated by a $d$ electron centered around the Ti site, and their relative magnitudes are in reasonable agreement with those expected for a point-like dipole at the Ti site, i.e., $A_1$:$A_2$:$A_3$ = $-1$:$+2$:$-1$. These values are estimated based on the diagonal terms of the dipole tensor, $A_1\propto-|\overline{\mu}_e|/r^3$, $A_2\propto+2|\overline{\mu}_e|/r^3$, and $A_3\propto-|\overline{\mu}_e|/r^3$ for an effective magnetic moment $|\overline{\mu}_e|$, with $r$ being the distance between the H and Ti atoms.
Provided that Mu forms a complex state identical to that of H, the corresponding HF parameters can be predicted by simply scaling these values for H using the factor $\gamma_{\mu}/\gamma_{\rm p}=3.1832$ (as shown in Table I, disregarding a minor difference of $\phi_0$), where $\gamma_{\rm p}=2\pi\times42.5774$ MHz/T is the proton gyromagnetic ratio. Adopting the reasonable assumption that the $A_i$ signs are common between the Mu and H complexes, comparison of the $A_i$ values given in Table I immediately leads us to the conclusion that the Mu complex state differs from that of H.
In the classical limit, the unpaired electron is regarded as a point-like magnetic dipole situated at the nearest neighboring Ti site, and the HF parameters (per unit Bohr magneton, $\mu_B$) are given by the equation $$A_i=\frac{\gamma_\mu}{2\pi}g_e\mu_B\frac{3\cos^2\tau_i-1}{2r^3},$$ where $g_e$ is the electron $g$ factor, ${\bf r}$ ($r=|{\bf r}|$) is the coordinate of the Ti$^{3+}$ atom with a muon at the origin, and $\tau_i$ is the angle between ${\bf r}$ and the symmetry axis ($\tau_1=\tau_3=\pi/2$, $\tau_2=0$). Assuming that the distance to the nearest neighboring Ti$^{3+}$ atom $r_{\rm nn}=0.226$ nm (i.e., an unrelaxed crystal lattice) and that a full $1\mu_B$ moment (with $g_e=2$) is present, we have $A_1^{\rm nn}=A_3^{\rm nn}=-10.88$ MHz and $A_2^{\rm nn}=+21.77$ MHz, where nn indicates the nearest neighbor. These values are far greater than the experimental values, which strongly suggests that the spatial distribution must be considered for the spin polarization, e.g., using $$A_i=\frac{\gamma_\mu}{2\pi}g_e\mu_B\int d^3{\bf r}\rho_{\rm s}({\bf r})\frac{3\cos^2\tau_i-1}{2r^3},\label{SD}$$ where ${\bf r}$ is now the electron coordinate, and $\rho_{\rm s}$ is the spin density corresponding to the difference between the charge densities of the spin-up and spin-down electron(s) [@deWalle:93]. Comparison of the $A_i^{\rm nn}$ \[obtained using Eq. (\[SD\]) for $\rho_{\rm s}({\bf r})=\frac{1}{4\pi}\delta(r-r_{\rm nn})$\] with the experimental values implies that $\rho_{\rm s}({\bf r})$ may have a considerable spread to reduce the effective moment size $|\mu_e|\simeq\mu_B\int d^3{\bf r}\rho_{\rm s}({\bf r})/r^3$.
![(Color online) Hyperfine interactions expected for the magnetic dipolar field ($B_{\rm dip}$) generated by the spin polarization of both the Ti and O atoms under an external field ($\vec{B}$) parallel to the (a) $A_1$, (b) $A_2$, and (c) $A_3$ axes, respectively (see text).[]{data-label="fig4"}](tio2-fig4.eps){width="0.95\linewidth"}
Here, we show that both magnitudes and signs of the HF parameters can be understood by considering additional spin polarization of the O atom next to the muon. As illustrated in Fig. \[fig4\], assuming that spin polarization at the Ti site induces small polarization of the ligand O antiparallel to the Ti polarization (usually expected for the Heisenberg-type exchange interaction), the $A_i$ incorporating the contribution of the magnetic dipole at the O site are given approximately by $$\begin{aligned}
A_1&\simeq&\frac{\gamma_\mu}{2\pi}|\overline{\mu}_e|\left[-\frac{1}{r^3_{\rm nn}}-\frac{2\epsilon}{r^3_{\rm O}}\right],\nonumber\\
A_2&\simeq&\frac{\gamma_\mu}{2\pi}|\overline{\mu}_e|\left[+\frac{2}{r^3_{\rm nn}}+\frac{\epsilon}{r^3_{\rm O}}\right],\label{dipo}\\
A_3&\simeq&\frac{\gamma_\mu}{2\pi}|\overline{\mu}_e|\left[-\frac{1}{r^3_{\rm nn}}+\frac{\epsilon}{r^3_{\rm O}}\right],\nonumber\end{aligned}$$ where $\epsilon$ is the relative magnitude of the O spin polarization against that of Ti and $r_{\rm O}$ is the OMu bond length (= 0.109 nm). As summarized in Table \[tab2\], simulated $A_i$ for $|\overline{\mu}_e|=0.050\mu_B$ and $\epsilon=0.077$ (i.e., $\epsilon|\overline{\mu}_e|=0.0039\mu_B$) exhibit almost perfect agreement with the experimental values. The model also provides a natural explanation for the small magnitude of $A_3$ that the dipole fields from Ti and O atoms nearly cancel at the muon site \[see Fig. \[fig4\](c)\], suggesting the possibility that the current assumption on the $A_3$ sign may be irrelevant. We note that such residual spin polarization of ligand O atoms is usually expected in the case of strong $d$-$p$ hybridization, as has been observed for a variety of transition metal oxides [@Chung:03].
$A_1$ (MHz) $A_2$ (MHz) $A_3$ (MHz) $|\overline{\mu}_e|$ $\epsilon$
------ ------------- ------------- ------------- ---------------------- ------------
Exp. $-1.29(6)$ $+1.29(6)$ $-0.17(2)$ – –
Sim. $-1.29$ $+1.09$ $-0.17$ $0.050\mu_B$ 0.077
: Comparison of experimental and simulated Mu hyperfine parameters assuming antiferromagnetic spin polarization at nn O site, where the experimental values of $A_1$ and $A_3$ were used to evaluate $A_2$, $|\overline{\mu}_e|$, and $\epsilon$ in the simulation. \[tab2\]
We currently speculate that the difference in the electronic structure between Mu and H is primarily due to the local environment of the specimen. While the present Mu complex state was observed in an “as received" crystal, the H-related paramagnetic state was observed after a reduction process (annealing at 600$^\circ$ C for 10 min in a N$_2$ atmosphere [@Brant:11]). Considering the fact that the reduction process also yields O vacancies, it may be of interest to recall the recent investigation based on density-functional theory that suggested that H is more stable at the O vacancy site, where it adopts a negatively charged state to form a Ti$^{3+}$H$_{\rm VO}^-$ complex state [@Filippone:09]. Thus, it is natural to consider the possibility that the paramagnetic state observed via EPR/ENDOR might correspond to such a state, although the consistency with experimental results must be carefully examined. The fact that the relative magnitudes of the $A_i$’s for the H-related center correspond to the case of $\epsilon\simeq0$ in Eq. (\[dipo\]) (i.e., $A_1$:$A_2$:$A_3$ = $-1$:$+2$:$-1$) seems to constitute supporting evidence, as it suggests that the relevant H atom has no neighboring O atom.
The fact that the hyperfine interaction is dominated by a magnetic dipolar interaction also places a strong constraint on the diffusion of $\mu^+$ in the $c$ channel, because the $A_i$ are strongly dependent on the distance to the electron ($\propto1/r^3$). The diffusive motion of $\mu^+$ against a stationary $d$ electron at the Ti site would immediately lead to the strong damping of satellite signals due to the fluctuation of $r$ and hence of $A_i$.
Thermal Property
----------------
The $T$ dependence of the fractional yield ($f$) for the Mu complex state is shown in Fig. \[fig5\]. The sum of the signal amplitude between the Mu complex state and the diamagnetic state (not shown) is almost independent of $T$, suggesting that the Mu complex is converted to a diamagnetic state above $\sim$10 K. Here, it must be noted that $f\simeq 0.4$ ($<1$) does not necessarily indicate two different muon sites with different charge states. In non-metallic compounds, it is usually expected that the initial yield of the paramagnetic state upon muon implantation is predominantly determined by the density of the short-lived free electrons that are produced via radiolysis near the end of the muon radiation track. While this density varies between compounds, it is independent of $T$ because the radiolysis is an athermal process. We assume that the muon site in [TiO$_2$]{} is unique (as shown in Fig. \[fig1\]b) over the entire observed $T$ range, where the initial yield of the Mu complex is controlled by the athermal electron density, which is independent of $T$.
In general, the diamagnetic state can be either positively or negatively charged. If the electronic energy levels ($E_\mu^{0/+}$, $E_\mu^{-/0}$) associated with the Mu complex are situated near the top of the valence band, a Mu$^-$ state can be expected at lower $T$ because the Fermi level ($E_F$) is likely to be situated far above the mid-gap in the present $n$-type specimen (i.e., $E_\mu^{0/+}<E_\mu^{-/0}<E_F$). In that case, an [*increase*]{} in the yield of the Mu complex would be observed with elevating temperature due to the hole capture process (${\rm Mu}^- +h^+\rightarrow{\rm Mu}$); however, this is contrary to the actual observed behavior. Thus, the behavior of the Mu complex strongly suggests that the process relevant to the promotion of the diamagnetic state is electron release, ${\rm Mu}\rightarrow{\rm Mu}^+ + e^-$ (i.e., $E_\mu^{0/+}<E_F<E_\mu^{-/0}$), and that the Mu complex state can serve as an electron donor.
Provided that the ionization of the Mu complex state is driven by an Arrhenius-type activation process, the disappearance of the Mu signal above $\sim$10 K suggests that the activation energy ($E_a$) is of the order of $10^1$ meV. This is in line with certain earlier reports suggesting unidentified shallow level states (e.g, via optical absorption [@Pascual:78] or infrared absorption spectroscopy on deuterated rutile[@Herklotz:11]). In any case, it must be remembered that the interpretation of $E_a$ depends on the detailed neutral charge state formation process of the Mu complex, and also on the kind of equilibrium state realized for the Mu complex formation. At one end of the range, $E_a$ represents a direct transition from the defect level to the bottom of the conduction band, while at the other extreme, in equilibrium, it indicates a transition from the defect level to the Fermi level (i.e., $E_a\simeq E_F-E_\mu^{0/+}$). Since the origin of the $n$-type conductivity in the present specimen is unknown, the present value of $E_a$ should be interpreted as a lower bound for the actual defect level.
![(Color online) Temperature dependence of fractional yield for muonium complex state deduced from signal amplitude.[]{data-label="fig5"}](tio2-fig5.eps){height="0.6\linewidth"}
Summary and Conclusion
======================
We have demonstrated that the electronic structure of the interstitial Mu center in rutile is characterized by extremely small and highly anisotropic hyperfine parameters. These parameters are predominantly determined by magnetic dipolar interaction with the unpaired Mu electron, which is primarily situated at the Ti site. The hyperfine parameters are quantitatively explained by a model that considers a small residual spin polarization of the O atom (which is antiparallel to that of the Ti atom), suggesting that the electronic structure should be interpreted as being a Ti-O-Mu complex state. The extremely small effective moment size of the unpaired electron ($\sim$0.05$\mu_B$ at the Ti site, $\sim$0.0039$\mu_B$ at the O site) as well as the small activation energy required for its promotion to the conduction band, implies that the Mu complex (and hence the corresponding H state) can serve as an electron donor. This strongly suggests that H is one of the primary origins of unintentional $n$-type conductivity in rutile [TiO$_2$]{}.
We would like to thank the KEK-MSL staff for their technical support during the muon experiment. We would also like to express our gratitude to K. Yoshizawa, Y. Iwazaki, and S. Tsuneyuki for helpful discussion and for providing us with the result of their first-principle calculation for Mu/H in [TiO$_2$]{} prior to publication. This work was partially supported by the KEK-MSL Inter-University Research Program (2012A0110, 2012B0032).
B. D. Patterson, Rev. Mod. Phys. [**60**]{}, 69 (1988).
, ed. by J. Pankove and N. M. Johnson (Academic, New York 1990); S. J. Pearton, J. W. Corbett and M. Stavola, [*Hydrogen in Crystalline Semiconductors*]{} (Springer, 1992).
S. M. Myers, M. I. Baskes, H. K. Birnbaum, J. W. Corbett, G. G. DeLeo, S. K. Estreicher, E. E. Haller, P. Jena, N. M. Johnson, R. Kirchheim, S. J. Pearton, and M. J. Stavola, Rev. Mod. Phys. [**64**]{}, 559 (1992).
K. H. Chow, B. Hitti, and R. F. Kiefl, in *$\mu$SR on Muonium in Semiconductor and Its Relation to Hydrogen*, ed. by M. J. Stavola, [*Semiconductors and Semimetals*]{}, Vol. 51 (Academic, New York, 1998), p. 137.
C. G. Van de Walle, Phys. Rev. Lett. [**85**]{}, 1012 (2000).
S. F. J. Cox, E. A. Davis, S. P. Cottrell, P. J. C. King, J. S. Lord, J. M. Gil, H. V. Alberto, R. C. Viläo, J. Piroto Duarte, N. Ayres de Campos, A. Weidinger, R. L. Lichti and S. J. C. Irvine, Phys. Rev. Lett. [**86**]{}, 2601 (2001).
K. Shimomura, K. Nishiyama and R. Kadono, Phys. Rev. Lett. [**89**]{}, 255505 (2002).
D. M. Hofmann, A. Hofstaetter, F. Leiter, H. Zhou, F. Henecker, B. K. Meyer, S. B. Orlinskii, J. Schmidt, and P. G. Baranov, Phys. Rev. Lett. [**88**]{}, 045504 (2002).
J. M. Gil, H. V. Alberto, R. C. Viläo, J. Piroto Duarte, P. J. Mendes, L. P. Foreira, N. Ayres de Campos, A. Weidinger, J. Krause, Ch. Niedermayer, and S. F. J. Cox, Phys. Rev. Lett. [**83**]{}, 5294 (1999).
G. Alvin Shi, M. Saboktakin, M. Stavola, and S. J. Pearton, Appl. Phys. Lett. [**85**]{}, 5601 (2004).
R. C. Viläo, H. V. Alberto, J. Piroto Duarte, J. M. Gil, A. Weidinger, N. Ayres de Campos, R. L. Lichti, K. H. Chow, and S. F. Cox, Phys. Rev. B 72, 235203 (2005).
Ç. Kiliç and A. Zunger, Appl. Phys. Lett. [**81**]{}, 73 (2002).
C. G. Van de Walle and J. Neugebauer, Nature [**423**]{}, 626 (2003).
P. W. Peacock and J. Robertson, Appl. Phys. Lett. [**83**]{}, 2025 (2003)
E. A. Davis, S. F. J. Cox, R. L. Lichti, and C. G. Van de Walle, Appl. Phys. Lett. [**82**]{}, 592 (2003).
K. Shimomura, R. Kadono, K. Ohishi, M. Mizuta, M. Saito, K. H. Chow, B. Hitti, and R. L. Lichti, Phys. Rev. Lett. [**92**]{}, 135505 (2004).
P. O. Anderson, E. L. Kollberg, and A. Jelenski, Phys. Rev. B [**8**]{}, 4956 (1973).
E. J. Spahr, L. Wen, M. Stavola, L. A. Boatner, L. C. Feldman, N. H. Tolk, and G. Lüpke, Phys. Rev. Lett. [**104**]{}, 205901 (2010).
K. Shimomura, R. Kadono, A. Koda, K. Ohishi, and K. Nishiyama, in KEK Progress Report 2005-2 (KEK-MSL Report 2004), pp. 17 (2005).
S. F. Cox, J. L. Gavartin, J. S. Lord, S. P. Cottrell, J. M. Gil, H. V. Alberto, J. Piroto Duarte, R. C. Viläo, N. Ayres de Campos, D. J. Keeble, E. A. Davis, M. Charlton, and D. P. van der Werf, J. Phys.: Condens. Matter, [**18**]{}, 1079 (2006).
A. T. Brant, S. Yang, N. C. Giles, and L. E. Halliburton, J. Appl. Phys. [**110**]{}, 053714 (2011).
C. G. Van de Walle and L. Pavesi, Phys. Rev. B [**47**]{}, 4256 (1993).
See, for example, E. M. L. Chung, G. J. McIntyre, D. McK. Paul, G. Balakrishnan, and M. R. Lees, Phys. Rev. B [**68**]{}, 144410 (2003).
F. Filippone, G. Mattioli, P. Alippi, and A. A. Bonapasta, Phys. Rev. B [**80**]{}, 245203 (2009).
J. Pascual, J. Camassel, and H. Mathieu, Phys. Rev. B [**18**]{}, 5606 (1978).
F. Herklotz, E. V. Lavrov, and J. Weber, Phys. Rev. B [**83**]{}, 235202 (2011).
[^1]: Corresponding author: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $(M^{n},g)$ denote a closed Riemannian manifold ($n\geq3$) which admits a metric of negative curvature (not necessarily equal to $g$). Let $\omega_{1}:=\omega_{0}+\pi^{*}\sigma$ denote a twisted symplectic form on $TM$, where $\sigma\in\Omega^{2}\left(M\right)$ is a closed $2$-form and $\omega_{0}$ is the symplectic structure on $TM$ obtained by pulling back the canonical symplectic form $dx\wedge dp$ on $T^{*}M$ via the Riemannian metric. Let $\Sigma_{k}$ be the hypersurface $|v|=\sqrt{2k}$. We prove that if $n$ is odd and the Hamiltonian structure $(\Sigma_{k},\omega_{1})$ is Anosov with $C^{1}$ weak bundles then $(\Sigma_{k},\omega_{1})$ is stable if and only if it is contact. If $n$ is even and in addition the Hamiltonian structure is $1/2$-pinched, then the same conclusion holds. As a corollary we deduce that if $g$ is negatively curved, strictly $1/4$-pinched and $\sigma$ is not exact then the Hamiltonian structure $(\Sigma_{k},\omega_{1})$ is never stable for all sufficiently large $k$.'
address: 'Department of Pure Mathematics and Mathematical Statistics, University of Cambridge, Cambridge CB3 0WB, England'
author:
- 'Will J. Merry and Gabriel P. Paternain'
bibliography:
- 'C:/Users/Will/Documents/Lyx/willbibtex.bib'
title: Stability of Anosov Hamiltonian Structures
---
Introduction
============
Let $\Sigma$ be a closed oriented manifold of dimension $2n-1$. A *Hamiltonian structure* on $\Sigma$ is a closed $2$-form $\omega$ such that $\omega^{n-1}\ne0$. Its kernel $\ker\,\omega$ defines an orientable $1$-dimensional foliation.
A natural condition to impose on a Hamiltonian structure is *stability*; this asserts the existence of a $1$-form $\lambda$ such that $\ker\,\omega\subseteq\ker\, d\lambda$ and such that $\lambda\wedge\omega^{n-1}>0$. The $1$-form $\lambda$ is known as a *stabilizing $1$-form*.
A stronger condition one might like to impose is the following: $\left(\Sigma,\omega\right)$ is called of *contact type* if we can find a $1$-form $\lambda$ such that $d\lambda=\omega$ and $\lambda\wedge\omega^{n-1}>0$. In particular $\lambda$ is a stabilizing $1$-form and $\lambda$ is a contact form on $\Sigma$. Note that $\left(\Sigma,\omega\right)$ can be of contact type only if $\omega$ is exact. The stability condition first appeared in [@HoferZehnder1994] as a condition for which the Weinstein conjecture could be proved. More recently, stability has been recognized as a key condition to produce compactness results in Symplectic Field Theory [@BourgeoisEliashbergHoferWysockiZehnder2003; @CieliebakMohnke2005; @EliashbergKimPolterovich2006] and Rabinowitz Floer homology [@CieliebakFrauenfelderPaternain2009]. This paper is motivated by the desire to generalize a result in the latter reference, as we explain below.
Let $F$ be any vector field spanning $\ker\,\omega$. We say that $\left(\Sigma,\omega\right)$ is an *Anosov* Hamiltonian structure if the flow $\phi_{t}$ of $F$ is Anosov. Recall that this asserts the existence of a $d\phi_{t}$-invariant splitting $$T\Sigma=\mathbb{R}F\oplus E^{s}\oplus E^{u},$$ where $\mathbb{R}F$ is the $1$-dimensional distribution spanned by $F$, and such that there exist constants $C,\mu>0$ such that for all $x\in\Sigma$ and $t\geq0$,$$\left|d_{x}\phi_{t}(\xi)\right|\leq C\left|\xi\right|e^{-\mu t}\ \mbox{for }\xi\in E^{s}(x);$$ $$\left|d_{x}\phi_{-t}\left(\xi\right)\right|\leq C\left|\xi\right|e^{-\mu t}\ \mbox{for }\xi\in E^{u}(x).$$ The Anosov condition is invariant under time changes, and so is independent of the choice of vector field $F$. In other words, it is intrinsic to the Hamiltonian structure $(\Sigma,\omega)$. The *weak bundles* $E^{+}:=\mathbb{R}F\oplus E^{s}$ and $E^{-}:=\mathbb{R}F\oplus E^{u}$ are also invariant under time changes.\
We say that $\phi_{t}:\Sigma\rightarrow\Sigma$ is $1/2$-*pinched* (or $1$*-bunched* [@Hasselblatt1994a]) if there exist positive constants $C,A,a$ with $A<2a$ such that
$$\frac{1}{C}|\xi|e^{-At}\leq|d_{x}\phi_{t}(\xi)|\leq C|\xi|e^{-at}\;\;\mbox{{\rm for}}\;\xi\in E^{s}\;\;\mbox{{\rm and}}\; t\geq0,$$ $$\frac{1}{C}|\xi|e^{-At}\leq|d_{x}\phi_{-t}(\xi)|\leq C|\xi|e^{-at}\;\;\mbox{{\rm for}}\;\xi\in E^{u}\;\;\mbox{{\rm and}}\; t\geq0.$$
We say that an Anosov Hamiltonian structure satisfies the $1/2$-pinching condition if there exists *some* vector field $F$ spanning $\ker\,\omega$ whose flow satisfies the $1/2$-pinching condition.\
Here is the situation we are actually interested in. Let $(M,g)$ be a closed Riemannian manifold that admits a background metric of negative curvature (possibly different from $g$) and $\pi:TM\rightarrow M$ the tangent bundle. Throughout the paper we let $\omega_{0}$ denote the symplectic form on $TM$ obtained by pulling back the canonical symplectic form $dx\wedge dp$ on $T^{*}M$ via the Riemannian metric. The form $\omega_{0}$ is exact; if $\alpha\in\Omega^{1}(TM)$ denotes the $1$-form defined by $$\alpha_{v}(\xi)=\left\langle d_{v}\pi(\xi),v\right\rangle ,\label{eq:alpha}$$ then it is well known that $\omega_{0}=-d\alpha$. Suppose $\sigma\in\Omega^{2}\left(M\right)$ is a closed $2$-form on $M$. Given $\varepsilon\in\mathbb{R}$ we define $$\omega_{\varepsilon}:=\omega+\varepsilon\pi^{*}\sigma.$$ Let $F_{\varepsilon}$ denote the symplectic gradient of the Hamiltonian $$H(x,v)=\frac{1}{2}\left|v\right|^{2}$$ with respect to $\omega_{\varepsilon}$ and let $\phi_{t}^{\varepsilon}$ denote the flow of $F_{\varepsilon}$ with respect to $\omega_{\varepsilon}$. Note that $\phi_{t}^{0}$ is simply the geodesic flow. This flow models the motion of a particle of unit mass and charge $\varepsilon$ under the effect of a magnetic field, whose *Lorentz force* $Y:TM\rightarrow TM$ is the bundle map uniquely determined by $$\sigma_{x}(u,v)=\left\langle Y_{x}(u),v\right\rangle \label{eq:lorentz}$$ for all $u,v\in T_{x}M$ and all $x\in M$.
The family $\left\{ \omega_{\varepsilon}\right\} $ for $\varepsilon\in\left[0,1\right]$ interpolates between the standard symplectic form $\omega_{0}$ and the form $\omega_{1}$. The form $\omega_{1}$ is called a *twisted symplectic structure* [@ArnoldGivental1990] and the flow $\phi_{t}^{1}$ is called a *twisted geodesic flow* or a *magnetic flow*.
Let $\Sigma_{k}=H^{-1}(k)$. We are interested in Anosov Hamiltonian structures of the form $(\Sigma_{k},\omega_{1})$.\
Here is the main result we present. Let $n=\dim\, M$.\
**Theorem A.** *Suppose $(\Sigma_{k},\omega_{1})$ is an Anosov Hamiltonian structure and $n\geq3$. Assume in addition:*
- *If $n$ is odd, $(\Sigma_{k},\omega_{1})$ has weak bundles of class $C^{1}$;*
- *If $n$ is even, $(\Sigma_{k},\omega_{1})$ is $1/2$-pinched.*
*Then $(\Sigma_{k},\omega_{1})$ is stable if and only if it is of contact type. In particular, if $(\Sigma_{k},\omega_{1})$ is stable, then $\sigma$ must be exact.*\
The last statement in the theorem can be seen as follows. Since $n=\dim\, M\geq3$, the Gysin sequence of the sphere bundle $\pi|\Sigma_{k}:\Sigma_{k}\rightarrow M$ shows that $(\pi|\Sigma_{k})^{*}:H^{2}(M,\mathbb{R})\rightarrow H^{2}(\Sigma_{k},\mathbb{R})$ is an isomorphism for $n\geq4$ and injective for $n=3$. Since $\omega_{1}=-d\alpha+\pi^{*}\sigma$, the assertion that $\omega_{1}$ is exact on $\Sigma_{k}$ implies that $\pi^{*}\sigma|\Sigma_{k}$ is exact. Putting this together we conclude that $\sigma$ is exact.
The bunching condition is a necessary one in the even dimensional case. Indeed, consider the twisted geodesic flow $\phi_{t}^{1}$ on compact quotients of complex hyperbolic space with $\sigma$ given by the Kähler form. Then for $k$ sufficiently large, $\phi_{t}^{1}|\Sigma_{k}$ is Anosov, and $\Sigma_{k}$ is stable but not contact ($\sigma$ is not exact). The flow $\phi_{t}^{1}$ is algebraic and thus the stable and unstable bundles are real analytic. A stabilizing 1-form $\lambda$ can be defined by setting $\lambda(F_{1})=1$ and $\ker\,\lambda=E_{1}^{s}\oplus E_{1}^{u}$. The flow $\phi_{t}^{1}$ is not $1/2$-pinched since it has 2:1 resonances. It seems a reasonable conjecture that these are in fact the only cases where $C^{1}$ weak bundles is not sufficient to ensure that the conclusion of Theorem A holds. A well known theorem [@HirschPughShub1977] states that the $1/2$-pinching condition implies that the weak bundles are of class $C^{1}$. However the pinching condition is strictly stronger than requiring the weak bundles to be of class $C^{1}$ as the example of complex hyperbolic space described above shows.\
We will prove in Proposition \[pro:ANOSOV 12PINCHED\] that if $g$ is negatively curved and strictly $1/4$-pinched, for $k$ sufficiently large, the flow $\phi_{t}^{1}:\Sigma_{k}\rightarrow\Sigma_{k}$ is Anosov and $1/2$-pinched. Thus, as a corollary of Theorem A we obtain the following.\
**Corollary B.** *Suppose $n\geq3$ and $g$ is negatively curved and strictly $1/4$-pinched. Then for any $k$ sufficiently large, the hypersurface $\Sigma_{k}\subset TM$ is* ***not*** *stable if $\sigma$ is not exact.*\
The corollary was first proved in [@CieliebakFrauenfelderPaternain2009 Theorem 1.4] for the case of $n$ even (but not Theorem A) and this previous result was the motivation for the present paper. It shows that the stability condition may fail for whole intervals of energy levels.
A caveat about the word “stable” is in order. One of the most remarkable features about Anosov systems is that they are structurally stable. This means that nearby systems are orbitally equivalent via a *homeomorphism* which in general is *not* $C^{1}$. The stability condition in Symplectic Geometry is equivalent to the existence of a thickening of the hypersurface with smoothly conjugate characteristic foliations [@CieliebakMohnke2005 Lemma 2.3]. In some sense it is this additional smoothness that is being exploited in Theorem A. One can put this into different words as follows: The existence of the form 1-form $\lambda$ means that one can find a parametrization of the characteristic foliation such that the hyperplane bundle $E^{s}\oplus E^{u}$ is smooth. In general, $E^{s}\oplus E^{u}$ is only Hölder continuous.
The assumption that $M$ admits a background metric of negative curvature is most likely superfluous. We use it to construct a conjugacy between our flow and the geodesic flow of a metric of negative curvature on $M$. We use this conjugacy to show that the fundamental group $\pi_{1}(M)$ acts as a ‘North-South dynamics’ (see Section \[sec:North-South-Dynamics\] and in particular Theorem \[thm:north south dynamic\]) on the space of stable leaves on the universal covering, and that the space of leaves admit a ‘flip map’ (see the discussion before the proof of Theorem \[strong\]). In order to remove the assumption of negative curvature one would need to prove Theorem \[thm:north south dynamic\] and Theorem \[strong\] directly instead of constructing such a conjugacy.
*Acknowlegment*: We thank the referee for Remark \[Referee remark\].
\[sec:Preliminaries-on-Hamiltonia\]Preliminaries on Hamiltonian structures, holonomy and the Kanai connection
=============================================================================================================
A *Hamiltonian structure* is a pair $\left(\Sigma,\omega\right)$ where $\Sigma^{2n-1}$ is a closed oriented manifold and $\omega$ is a closed $2$-form on $\Sigma$ such that $\omega^{n-1}\ne0$ everywhere.
If $\left(\Sigma,\omega\right)$ is a Hamiltonian structure then $\ker\,\omega$ defines an orientable $1$-dimensional foliation of $\Sigma$, which we call the *characteristic foliation*.
Let $\phi_{t}:N\rightarrow N$ be a smooth flow on a closed Riemannian manifold $N$, and let $F$ denote its infinitesimal generator. We say that $\phi_{t}$ is *Anosov* if there exists a $d\phi_{t}$-invariant splitting $$T\Sigma=\mathbb{R}F\oplus E^{s}\oplus E^{u},$$ where $\mathbb{R}F$ is the $1$-dimensional distribution spanned by $F$, and such that there exist constants $C,\mu>0$ such that for all $x\in\Sigma$ and $t\geq0$,$$\left|d_{x}\phi_{t}(\xi)\right|\leq C\left|\xi\right|e^{-\mu t}\ \mbox{for }\xi\in E^{s}(x);$$ $$\left|d_{x}\phi_{-t}\left(\xi\right)\right|\leq C\left|\xi\right|e^{-\mu t}\ \mbox{for }\xi\in E^{u}(x).$$
The Anosov property is invariant under time changes (see [@deLaLlaveMarcoMoryion1986 Lemma 1.2] or [@KatokHasselblatt1995 Proposition 17.4.5]), and hence we can define a Hamiltonian structure $\left(\Sigma,\omega\right)$ to be *Anosov* if the flow $\phi_{t}$ of *any* vector field $F$ spanning $\ker\,\omega$ is Anosov.
We say that $\phi_{t}:\Sigma\rightarrow\Sigma$ is $1/2$-*pinched* (or $1$*-bunched* [@Hasselblatt1994a]) if there exist positive constants $C,A,a$ with $A<2a$ such that
$$\frac{1}{C}|\xi|e^{-At}\leq|d_{x}\phi_{t}(\xi)|\leq C|\xi|e^{-at}\;\;\mbox{{\rm for}}\;\xi\in E^{s}\;\;\mbox{{\rm and}}\; t\geq0,$$ $$\frac{1}{C}|\xi|e^{-At}\leq|d_{x}\phi_{-t}(\xi)|\leq C|\xi|e^{-at}\;\;\mbox{{\rm for}}\;\xi\in E^{u}\;\;\mbox{{\rm and}}\; t\geq0.$$
We say that an Anosov Hamiltonian structure satisfies the $1/2$-pinching condition if there exists *some* vector field $F$ spanning $\mbox{{\rm ker}}\,\omega$ whose flow satisfies the $1/2$-pinching condition.\
The $1/2$-pinching condition is a natural one to study, and should be thought of as a statement about being ‘strongly hyperbolic’. Unsurprisingly, an Anosov system possessing this enhanced degree of hyperbolicity enjoys greater regularity. More specifically, write $E^{+}$ and $E^{-}$ for the *weak stable* and *weak unstable* bundles $E^{s}\oplus\mathbb{R}F$ and $E^{u}\oplus\mathbb{R}F$ respectively. If an Anosov Hamiltonian structure is $1/2$-pinched, then $E^{+}$ and $E^{-}$ are of class $C^{1}$ [@HirschPughShub1977]. The next example will be crucial for the proof of Corollary B.
\[exa:example sec curv\]Let $(M,g)$ be a closed manifold with negative sectional curvature $K$. Then the geodesic flow $\phi_{t}:SM\rightarrow SM$ is Anosov (see [@KatokHasselblatt1995 Section 17.6]). In fact, we can say more. By compactness we can find constants $k_{1}\geq k_{0}>0$ such that $$-k_{1}^{2}\leq K\leq-k_{0}^{2}.$$ Then, comparison theorems show that [@Klingenberg1995 Theorem 3.2.17] (see also [@Knieper2002 Proposition 3.2]) there is a constant $C>0$ such that $$\frac{1}{C}|\xi|e^{-k_{1}t}\leq|d_{v}\phi_{t}(\xi)|\leq C|\xi|e^{-k_{0}t}\;\;\mbox{{\rm for}}\;\xi\in E^{s}(v)\;\;\mbox{{\rm and}}\; t\geq0,\label{eq:a1}$$ $$\frac{1}{C}|\xi|e^{-k_{1}t}\leq|d_{v}\phi_{-t}(\xi)|\leq C|\xi|e^{-k_{0}t}\;\;\mbox{{\rm for}}\;\xi\in E^{u}(v)\;\;\mbox{{\rm and}}\; t\geq0.\label{eq:a2}$$ We see that $\phi_{t}$ is $1/2$-pinched as long as $k_{1}<2k_{0}$. Therefore the geodesic flow of a metric whose sectional curvature satisfies $-4\leq K<-1$ is $1/2$-pinched.
We return to two more definitions regarding Hamiltonian structures.
Let $\left(\Sigma,\omega\right)$ denote a Hamiltonian structure. We say that $\left(\Sigma,\omega\right)$ is *stable* if there exists a $1$-form $\lambda$ such that $\ker\,\omega\subseteq\ker\, d\lambda$ and $\lambda\wedge\omega^{n-1}>0$. $\lambda$ is called a *stabilizing $1$-form*. Note that if $\lambda$ is a stabilizing $1$-form and $F$ is any vector field tangent to $\ker\,\omega$ we have $i_{F}d\lambda=0$. If $F$ is normalized so that $\lambda(F)=1$ then we say $F$ is the *Reeb vector field* of **$\lambda$**; note that the Reeb vector field is unique**.**
A stronger condition is the following.
Let $\left(\Sigma,\omega\right)$ denote a Hamiltonian structure. We say that $\left(\Sigma,\omega\right)$ is of *contact type* if there exists a $1$-form $\lambda$ such that $d\lambda=\omega$ and $\lambda\wedge\omega^{n-1}>0$. Note if $\left(\Sigma,\omega\right)$ is of contact type then it is certainly stable, and that if $\left(\Sigma,\omega\right)$ is of contact type then $\omega$ is exact.
Now let $\left(\Sigma,\omega\right)$ denote a stable Anosov Hamiltonian structure with weak bundles of class $C^{1}$. Let $\lambda$ be a stabilizing $1$-form and let $F$ be the Reeb vector field of $\lambda$. Let $\phi_{t}$ denote the flow of $F$; then $\phi_{t}$ is Anosov and $T\Sigma=\mathbb{R}F\oplus E^{s}\oplus E^{u}$, with $\ker\,\lambda=E^{s}\oplus E^{u}$. Then the weak bundles are $C^{1}$; since $\lambda$ is $C^{\infty}$ (and so the bundle $\ker\,\lambda$ is of class $C^{\infty}$), it follows that the strong bundles $E^{s}$ and $E^{u}$ are also of class $C^{1}$. The importance of this is that, as we shall see in Theorem \[thm:(the-Kanai-connection)\] below, under these conditions there exists a unique connection $\nabla$ on $\Sigma$ called the *Kanai connection* which satisfies certain desirable properties. This was originally introduced by Kanai in [@Kanai1988]; see also [@Kanai1993; @FeresKatok1990].
\[sec:holonomy\]Holonomy {#secholonomyholonomy .unnumbered}
------------------------
We briefly recall the concept of holonomy transport along the weak (un)stable foliations defined by an Anosov flow.
Throughout the remainder of this section, let $N$ denote a closed manifold of dimension $2n-1$ and $\phi_{t}:N\rightarrow N$ an Anosov flow on $N$ with infinitesimal generator $F$. We also assume that $E:=E^{s}\oplus E^{u}$ is smooth and admits a smooth symplectic form $\omega$ which is $\phi_{t}$-invariant. We extend this form $\omega$ to a form defined on all of $TN$ by requiring that $i_{F}\omega=0$. As above $\lambda$ is the 1-form defined by $\ker\,\lambda=E$ and $\lambda(F)=1$.
It is well known that the subbundles $E^{s}$ and $E^{u}$, together with the weak bundles $E^{+}$ and $E^{-}$, are all integrable. Namely, given any $x\in N$, we define $$W^{s}(x):=\left\{ y\in N\,:\,\mbox{dist}(\phi_{t}x,\phi_{t}y)\rightarrow0\mbox{ as }t\rightarrow\infty\right\}$$ and $$W^{u}(x):=\left\{ y\in N\,:\,\mbox{dist}(\phi_{t}x,\phi_{t}y)\rightarrow0\mbox{ as }t\rightarrow-\infty\right\} .$$ The sets $W^{s}(x)$ and $W^{u}(x)$ are injectively immersed manifolds called the *strong (un)stable manifolds* at $x$ and satisfy$$T_{x}W^{s}(x)=E^{s}(x),\ \ \ T_{x}W^{u}(x)=E^{u}(x).$$ These define foliations $\mathcal{W}^{s}$ and $\mathcal{W}^{u}$ of $N$, called the *strong (un)stable foliations*. We assume throughout this section that these foliations $\mathcal{W}^{s},\mathcal{W}^{u}$ are of class $C^{1}$.
Similarly, given $x\in N$, we define $$W^{+}(x):=\bigcup_{t\in\mathbb{R}}\phi_{t}[W^{s}(x)]=\bigcup_{t\in\mathbb{R}}W^{s}(\phi_{t}x)$$ and $$W^{-}(x):=\bigcup_{t\in\mathbb{R}}\phi_{t}[W^{u}(x)]=\bigcup_{t\in\mathbb{R}}W^{u}(\phi_{t}x),$$ which are then the *weak (un)stable manifolds* at $x$. They satisfy$$T_{x}W^{+}(x)=E^{+}(x),\ \ \ T_{x}W^{-}(x)=E^{-}(x),$$ and define foliations $\mathcal{W}^{+}$ and $\mathcal{W}^{-}$ of $M$.\
Consider the foliations $\mathcal{W}^{u}$ and $\mathcal{W}^{+}$. It is well known that these are transverse to each other (i.e. $E^{u}(x)\cap E^{+}(x)=\{0\}$) and have complementary dimensions $n-1$ and $n$ respectively. The same is of course true of the foliations $\mathcal{W}^{s}$ and $\mathcal{W}^{-}$, and all of what we say below can be repeated for them.
By a *foliation chart* we mean a diffeomorphism $\varphi:U\rightarrow(-1,1)^{n-1}\times(-1,1)^{n}$ of the form $x\mapsto(\varphi_{1}(x),\varphi_{2}(x))$ where $U\subseteq N$ is open, such that the connected components of $\mathcal{W}^{u}|U$ are given by $\varphi_{2}=\mbox{const}$ and the connected components of $\mathcal{W}^{+}|U$ are given by $\varphi_{1}=\mbox{const}$. We then call $U$ a *foliated neighborhood*.
Let $ $$(\varphi,U)$ denote a foliated chart defined on $N$. Given $x\in U$, let $W_{U}^{u}(x):=W^{u}(x)\cap U$. Suppose $y\in W_{U}^{+}(x)$ lies in the same connected component of $W_{U}^{+}(x)$ as $x$. We want to define the *holonomy map $\mathcal{H}_{x,y}^{U}:W_{U}^{u}(x)\rightarrow W_{U}^{u}(y)$ along the leaves of $\mathcal{W}^{+}$* . This is defined as follows. Suppose $p\in W_{U}^{u}(x)$. Then there exists a unique point $q\in W^{+}(p)\cap W_{U}^{u}(y)$, and we define $\mathcal{H}_{x,y}^{U}(p)=q$. The map $\mathcal{H}_{x,y}^{U}$ is of class $C^{r}$ for $r\geq1$ if the foliations $\mathcal{W}^{u}$ and $\mathcal{W}^{+}$ are also of class $C^{r}$.
More generally, suppose $\gamma:[0,T]\rightarrow N$ is a smooth curve such that $\gamma(0)=x$ and $\gamma(t)\in W^{+}(x)$ for all $t$. Let $y:=\gamma(T)\in W^{+}(x)$. Then we can define $\mathcal{H}_{x,y}^{\gamma}$ by covering the image of $\gamma$ with foliated charts $(\varphi_{i},U_{i})$ for $i=1,\dots,l$ with $x\in U_{1}$ and $y\in U_{l}$, and choosing points $0=t_{0},\dots,t_{l}=T$ such that $\gamma(t_{i-1}),\gamma(t_{i})\in U_{i}$ for each $1\leq i\leq l$ and then setting $$\mathcal{H}_{x,y}^{\gamma}:=\mathcal{H}_{\gamma(t_{l-1}),y}^{U_{l}}\circ\mathcal{H}_{\gamma(t_{l-2}),\gamma(t_{l-1})}^{U_{l-1}}\circ\dots\circ\mathcal{H}_{x,\gamma(t_{1})}^{U_{1}}.$$ It can be shown that $\mathcal{H}_{x,y}^{\gamma}$ only depends on the homotopy class of $\gamma\in W^{+}(x)$. One can check that if $x,y,z$ are in the image of $\gamma$, then after suitably restricting the domains of definition, it holds that$$\mathcal{H}_{y,z}^{\gamma}\circ\mathcal{H}_{x,y}^{\gamma}=\mathcal{H}_{x,z}^{\gamma}.$$ Moreover, since the foliations are $\phi_{t}$-invariant, for any curve $\gamma$ in the weak unstable foliation $\mathcal{W}^{+}$ we have$$\phi_{t}\circ\mathcal{H}_{x,y}^{\gamma}\circ\phi_{-t}=\mathcal{H}_{\phi_{t}x,\phi_{t}y}^{\phi_{t}\circ\gamma}.\label{eq:hol eq}$$
Next, we consider the differential of $\mathcal{H}_{x,y}^{\gamma}$, known as the *holonomy transport along $\gamma$*, $$H_{x,y}^{\gamma}(p):=d_{p}\mathcal{H}_{x,y}^{\gamma}:T_{p}W^{u}(x)\rightarrow T_{\mathcal{H}_{x,y}^{\gamma}(p)}W^{u}(y).$$ In particular we write $$H_{x,y}^{\gamma}:E^{u}(x)\rightarrow E^{u}(y)$$ for the map $H_{x,y}^{\gamma}(x):T_{x}W^{u}(x)\rightarrow T_{y}W^{u}(y)$. Note that differentiating gives
$$d_{y}\phi_{t}\circ H_{x,y}^{\gamma}\circ d_{\phi_{-t}x}\phi_{-t}=H_{\phi_{t}x,\phi_{t}y}^{\phi_{t}\circ\gamma}\ \ \ \mbox{as maps }E^{u}(\phi_{t}x)\rightarrow E^{u}(\phi_{t}y).\label{eq:diff hol eq}$$
We say that a vector field $X\in\Gamma(E^{u})$ is *invariant under the holonomy transport along the leaves of* **$\mathcal{W}^{+}$** if for any curve $\gamma$ from $x$ to $y$ contained in $W^{+}(x)$ it holds that $$H_{x,y}^{\gamma}(p)(X(p))=X(\mathcal{H}_{x,y}^{\gamma}(p))\ \ \ \mbox{for all\ }p\in W^{u}(x).$$
The Kanai connection {#the-kanai-connection .unnumbered}
--------------------
We now recall the definition and main features of the Kanai connection.
Let $I$ be the $(1,1)$-tensor on $N$ given by $I(v)=-v$ for $v\in E^{s}$, $I(v)=v$ for $v\in E^{u}$ and $I(F)=0$. Consider the symmetric non-degenerate bilinear form given by $$h(X,Y):=\omega(X,IY)+\lambda\otimes\lambda(X,Y).$$ The pseudo-Riemannian metric $h$ is of class $C^{1}$ and thus there exists a unique $C^{0}$ affine connection $\nabla$ such that:
1. $h$ is parallel with respect to $\nabla$;
2. $\nabla$ has torsion $\omega\otimes F$.
\[thm:(the-Kanai-connection)\] The Kanai connection $\nabla$ has the following properties:
1. $\nabla$ is $\phi_{t}$-invariant, $\nabla\omega=0$, $\nabla_{F}=L_{F}$ and $\nabla F=0$.
2. The Anosov splitting is invariant under $\nabla$, that is, if $X_{s}\in\Gamma(E^{s}),X_{u}\in\Gamma(E^{u})$ and $Y$ is any vector field on $N$ then $$\nabla_{Y}X_{s}\in\Gamma(E^{s}),\ \ \ \nabla_{Y}X_{u}\in\Gamma(E^{u}).$$
3. The restriction of $\nabla$ to each leaf of the foliations $\mathcal{W}^{s}$ and $\mathcal{W}^{u}$ of $N$ is flat (note that restriction of the connection to the leaves of the stable and unstable foliations is smooth so it makes sense to talk about its curvature).
4. Parallel transport along curves on the weak stable and unstable manifolds coincide with the holonomy transport determined by the stable and unstable foliations.
\[-indeo of gamma\]Let us observe that since we know that the restriction of $\nabla$ to each leaf of $\mathcal{W}^{s}$ and $\mathcal{W}^{u}$ is flat, it follows that $H_{x,y}^{\gamma}$ is independent of $\gamma$. Thus we will omit $\gamma$ from the notation and simply write $H_{x,y}$.
\[The-holonomy-transport=00003D00003D00003Ddif\]The holonomy transport $H_{x,\phi_{t}x}$ is given by $d_{x}\phi_{t}|E^{u}(x)$, that is,$$H_{x,\phi_{t}x}=d_{x}\phi_{t}|E^{u}(x)\ \ \ \mbox{as maps }E^{u}(x)\rightarrow E^{u}(\phi_{t}x).$$
Fix $x\in N$, and let $\Gamma(t):=\phi_{t}x$. Fix $\xi\in E^{u}(x)$, and let $ $$V(t)$ denote the vector field along $\Gamma(t)$ defined by $V(t)=d_{x}\phi_{t}(\xi)$. It suffices to show that $V$ is parallel: $\nabla_{\dot{\Gamma}}V\equiv0$.
Note that $\dot{\Gamma}(t)=F(\phi_{t}x)$. Since $\nabla_{F}=L_{F}$ we have $$\nabla_{F(\phi_{t}x)}V(t)=\frac{d}{dt}\Bigl|_{t=0}d_{\phi_{t}x}\phi_{-t}(V(\phi_{t}x))=\frac{d}{dt}\Bigl|_{t=0}\xi=0.$$ and the lemma follows.
\[sec:Constructing-the-Invariant\]Constructing the Invariant Subbundles
=======================================================================
Throughout this section let $\left(\Sigma,\omega\right)$ denote a stable Anosov Hamiltonian structure of dimension $2n-1$ where $n\geq2$, $\lambda$ a stabilizing $1$-form for $(\Sigma,\omega)$, $F$ the Reeb vector field of $\lambda$, and $\phi_{t}$ the flow of $F$. If $n$ is odd, suppose that the weak (un)stable bundles are of class $C^{1}$. If $n$ is even, assume that the Anosov Hamiltonian structure is $1/2$-pinched. In either case, the Kanai connection $\nabla$ is defined. The goal in this section is to construct a subbundle of $E^{u}$ that is invariant under both $\phi_{t}$ and the holonomy transport along the leaves of the weak stable foliation $\mathcal{W}^{+}$. It is the existence of this subbundle that we will then exploit in Section \[sec:Proof-of-Theorem A\] in order to prove Theorem A from the introduction. The main ideas for these constructions come from [@Hamenstaedt1995; @Feres1991; @FeresKatok1990; @Kanai1993].\
In the even dimensional case we will need to know that $d\lambda$ is parallel with respect to the Kanai connection. This is the only place in the paper where we will actually use the $1/2$-pinching condition (as opposed to just $C^{1}$ weak (un)stable bundles). The following lemma is due to Kanai ([@Kanai1993 Lemma 3.2]).
\[lem:.dlambda is parallel\]Suppose $\phi_{t}$ is a time change of a $1/2$-pinched Anosov flow. Then $\nabla(d\lambda)=0$.
Suppose $\tau$ is any invariant $(0,3)$-tensor annihilated by $F$, i.e. $i_{F}\tau=0$. We claim that $\tau$ must vanish. Note that if $\psi_{t}$ is any time change of $\phi_{t}$, then $\psi_{t}$ also leaves $\tau$ invariant since $F$ annihilates $\tau$, so in the proof below without loss of generality, we may assume that $\phi_{t}$ itself is $1/2$-pinched.
To see that $\tau$ vanishes consider for example a triple of vectors $(\xi_{1},\xi_{2},\xi_{3})$ where $\xi_{1},\xi_{2}\in E^{s}(x)$ but $\xi_{3}\in E^{u}(x)$. Then there is a constant $C>0$ such that $$\begin{aligned}
|\tau_{x}(\xi_{1},\xi_{2},\xi_{3})| & =|\tau_{\phi_{t}x}(d_{x}\phi_{t}(\xi_{1}),d_{x}\phi_{t}(\xi_{2}),d_{x}\phi_{t}(\xi_{3}))|\\
& \leq Ce^{(A-2a)t}|\xi_{1}||\xi_{2}||\xi_{3}|,\end{aligned}$$ By the $1/2$-pinching condition the last expression tends to zero as $t\to\infty$ and therefore $\tau_{x}(\xi_{1},\xi_{2},\xi_{3})=0$. The same will happen for other possible triples $(\xi_{1},\xi_{2},\xi_{3})$ when we let $t\to\pm\infty$.
Since $d\lambda$ and $\nabla$ are $\phi_{t}$-invariant, so is $\nabla(d\lambda)$. Since $i_{F}d\lambda=0$, $\nabla(d\lambda)$ is also annihilated by $F$ (to see that $\nabla_{F}(d\lambda)=0$ use that $d\lambda$ is $\phi_{t}$-invariant and that $\nabla_{F}=L_{F}$). Hence by the previous argument applied to $\tau=\nabla(d\lambda)$ we conclude that $\nabla(d\lambda)=0$ as desired.
There exists a smooth $\phi_{t}$-invariant bundle map $L:E\rightarrow E$ such that for $X,Y\in\Gamma(E)$, $$d\lambda(X,Y)=\omega(LX,Y)=\omega(X,LY).$$ Moreover $L$ preserves the decomposition of $E=E^{s}\oplus E^{u}$, that is, $L=L^{s}+L^{u}$ where $L^{s}:E^{s}\rightarrow E^{s}$ and $L^{u}:E^{u}\rightarrow E^{u}$.
Since $d\lambda$ is $\phi_{t}$-invariant and annihilated by $F$, there exists a $\phi_{t}$-invariant smooth section $L$ of $E^{*}\otimes E$ such that the stated equation holds. It remains to check that $L$ preserves the decomposition, that is, $L$ commutes with $I$. But this is clear: if $X_{s},Y_{s}\in\Gamma(E^{s})$ then $$d\lambda(X_{s},Y_{s})=X_{s}\lambda(Y_{s})-Y_{s}\lambda(X_{s})-\lambda\bigl(\left[X_{s},Y_{s}\right]\bigr);$$ using integrability of $E^{s}$ and the fact that $\ker\,\lambda=E$, we see that $$0=d\lambda(X_{s},Y_{s})=\omega(LX_{s},Y_{s}),$$ and hence $LX_{s}\in\Gamma(E^{s})$. The same argument applies with sections of $E^{u}$.
The construction of the invariant subbundle will depend on the parity of $n$. We will begin with the easier case, when $n$ is even.
The case of $n$ even {#the-case-of-n-even .unnumbered}
--------------------
Since $\dim\,\Sigma=2n-1$ and $n$ is even, we have $\dim\, E^{u}=n-1$ an odd number. Thus for any $x\in\Sigma$, the map $L_{x}^{u}:E^{u}(x)\rightarrow E^{u}(x)$ admits a real eigenvalue $\rho_{x}$. In fact, we have the following result.
\[lem:eigenvalue lemma\]There exists $\rho\in\mathbb{R}$ such that $\rho_{x}=\rho$ for all $x\in\Sigma$.
\[standard anosov rem\]In the proof of the lemma we will make use of the fact that $\phi_{t}$ is a *transitive* flow. To see this, we first note that since $\phi_{t}$ preserves a probability measure, the *non-wandering set* $\Omega(\phi)$ of $\phi_{t}$ is necessarily equal to all of $\Sigma$ (see for instance [@Anosov1967] or [@KatokHasselblatt1995 Chapter 18]). It is then a standard result that an Anosov flow whose non-wandering set is the whole space is transitive (see for instance [@KatokHasselblatt1995 p.576]). We also remark that since $\phi_{t}$ is Anosov the set of periodic points is dense in the non-wandering set $\Omega(\phi)$, and hence dense in $\Sigma$; we will use this observation in the next subsection.
For $k\geq0$, let $a_{k}(x)$ denote the coefficient of $t^{k}$ in the characteristic polynomial $p_{x}(t)$ of $L_{x}^{u}$. Then $a_{k}:\Sigma\rightarrow\mathbb{R}$ is continuous and $\phi_{t}$-invariant. Since $\phi_{t}$ is transitive, $a_{k}$ is constant. Thus the characteristic polynomial $p_{x}(t)$ of $L_{x}^{u}$ is independent of $x$, and so each $L_{x}^{u}$ admits the same eigenvalues.
Let therefore $\rho_{0}\in\mathbb{R}$ be a common eigenvalue of the maps $\left\{ L_{x}^{u}\right\} $. Let $$P_{\rho_{0}}(x):=\left\{ \xi\in E^{u}(x)\,:\, L_{x}^{u}\xi=\rho_{0}\xi\right\} \ne\{0\}.\label{eq:P ro 0}$$
The map $x\mapsto P_{\rho_{0}}(x)$ defines a $C^{1}$ subbundle of $E^{u}$. Moreover for $x\in\Sigma$, the restriction of $P_{\rho_{0}}$ to $W^{u}(x)$ is integrable.
Since $d\lambda$ is $\nabla$-parallel by Lemma \[lem:.dlambda is parallel\], $P_{\rho_{0}}(x)$ is invariant under the parallel transport of $\nabla$. More precisely, given a curve $\gamma$ from $x$ to $y$, if $\mathcal{P}_{x,y}^{\gamma}:T_{x}\Sigma\rightarrow T_{y}\Sigma$ denotes the parallel transport along $\gamma$ with respect to $\nabla$, then $$\mathcal{P}_{x,y}^{\gamma}P_{\rho_{0}}(x)\subseteq P_{\rho_{0}}(y),$$ since for $\xi\in E^{u}(x)$, if $L_{x}^{u}\xi=\rho_{0}\xi$ then $$L_{y}^{u}\mathcal{P}_{x,y}^{\gamma}(\xi)=\rho_{0}\mathcal{P}_{x,y}^{u}(\xi).$$
Indeed, we have for any vector fields $X,Y,Z$ that:$$\begin{aligned}
0 & = & \bigl(\nabla_{X}d\lambda\bigr)\left(Y,Z\right)\\
& = & \nabla_{X}(d\lambda(Y,Z))-d\lambda\bigl(\nabla_{X}Y,Z\bigr)-d\lambda\bigl(Y,\nabla_{X}Z\bigr)\\
& = & \nabla_{X}\bigl(\omega(LY,Z)\bigr)-\omega\bigl(L\bigl(\nabla_{X}Y\bigr),Z\bigr)-\omega\bigl(LY,\nabla_{X}Z\bigr)\\
& = & \nabla_{X}\bigl(\omega\left(LY,Z\right)\bigr)-\omega\bigl(\nabla_{X}(LY),Z\bigr)-\omega\bigl(LY,\nabla_{X}Z\bigr)+\omega\bigl(\bigl(\nabla_{X}L\bigr)Y,Z\bigr)\\
& = & \nabla_{X}\omega\bigl(LY,Z\bigr)+\omega\bigl(\bigl(\nabla_{X}L\bigr)Y,Z\bigr)\\
& = & \omega\bigl(\bigl(\nabla_{X}L\bigr)Y,Z\bigr),\end{aligned}$$ where the last equality used the fact that $\nabla\omega=0$. Thus $L$ is parallel, and hence $L^{u}$ and $L^{s}$ are also parallel. Thus: $$\begin{aligned}
L_{y}^{u}\mathcal{P}_{x,y}^{\gamma}(\xi) & = & \mathcal{P}_{x,y}^{\gamma}\bigl(L_{x}^{u}\xi\bigr)\\
& = & \mathcal{P}_{x,y}^{\gamma}(\rho_{0}\xi)\\
& = & \rho_{0}\mathcal{P}_{x,y}^{\gamma}(\xi).\end{aligned}$$ It remains to check that the restriction of $P_{\rho_{0}}$ to $W^{u}(x)$ is integrable. Note that if $X$ and $Y$ are parallel sections of $E^{u}$ over $W^{u}(x)$ then since $\nabla$ has no torsion over $E^{u}$, we have $$0=\nabla_{X}Y-\nabla_{Y}X=\left[X,Y\right].$$
The case of $n$ odd {#the-case-of-n-odd .unnumbered}
-------------------
We now want to construct a $\phi_{t}$-invariant subbundle $P_{\rho_{0}}\subseteq E^{u}$ that is invariant under holonomy transport along the leaves of $\mathcal{W}^{+}$ for the case when $n$ is an odd integer. As before we construct the maps $L_{x}^{u}:E^{u}(x)\rightarrow E^{u}(x)$; however in this case since $\dim\, E^{u}(x)=n-1$ is even, it is no longer necessarily the case that $L_{x}^{u}$ admits a real eigenvalue, and so our previous construction will not work.
\[Referee remark\]Recall we are only assuming that $\phi_{t}$ is a time change of a $1/2$-pinched Anosov flow in the even dimensional case. Thus in the odd dimensional case Lemma \[lem:.dlambda is parallel\] is not available to us. If however we did assume that $\nabla(d\lambda)=0$ then we could dramatically simplify the treatment of the odd-dimensional case both in this section and in Section \[sec:Proof-of-Theorem A\]. Indeed, whilst in the odd-dimensional case the characteristic polynomial $p(t)$ of $L^{u}$ no longer necessarily admits a real eigenvalue, it is reducible over $\mathbb{R}$ to a product of quadratic factors, say $p(t)=q_{1}(t)\dots q_{k}(t)$. We may then define sub-bundles $P_{i}:=\ker\, q_{i}(L^{u})$ of $E^{u}$, which have constant non-zero dimension. Since $\nabla(d\lambda)=0$ these are parallel and hence integrable. We thank the referee for this observation.
Let $E_{\mathbb{C}}^{u}:=E^{u}\otimes\mathbb{C}$ denote the complexification of $E^{u}$, and let $\mathbb{L}:E_{\mathbb{C}}^{u}\rightarrow E_{\mathbb{C}}^{u}$ denote the complex linear extension of $L^{u}$. Given $\rho\in\mathbb{C}$, we set $$Q_{\rho}(x):=\left\{ \xi\in E^{u}(x)\,:\,\xi=\mbox{Re}\,\zeta\mbox{ for some }\zeta\in E_{\mathbb{C}}^{u}\mbox{ with }\mathbb{L}_{x}\zeta=\rho\zeta\right\} .$$ By the same arguments as those used in the proof of Lemma \[lem:eigenvalue lemma\], there exists an open $\phi_{t}$-invariant set $O\subseteq\Sigma$ with the property that for any fixed $\rho\in\mathbb{C}$, the dimension of the subspaces $Q_{\rho}(x)$ is constant for $x\in O$. Thus $Q_{\rho}|O$ defines a continuous $\phi_{t}$-invariant subbundle of $E^{u}|O$ for all $\rho\in\mathbb{C}$.
Now define a new bundle $P_{\rho}$ over all of $\Sigma$ by setting $$P_{\rho}(x):=\bigcap\left\{ H_{y,x}Q_{\rho}(y)\,:\, y\in O\cap W^{s}(x)\right\} .\label{eq:P ro}$$ We call $P_{\rho}$ the *holonomy intersection* of $Q_{\rho}$ (see [@Hamenstaedt1995 p685], as well as [@Feres1991 Section 8]). Note that $P_{\rho}$ is clearly holonomy invariant. The next result is Lemma $2.7$ in [@Hamenstaedt1995].
\[thm:hamen2-1\]For all $\rho\in\mathbb{C}$, $P_{\rho}$ is a continuous $\phi_{t}$-invariant subbundle of $E^{u}$ over $\Sigma$.
Let $$K:=\min\left\{ \dim\, P_{\rho}(x)\,:\, x\in\Sigma\right\} .$$ Since the dimension of $K$ is constant along the leaves of $\mathcal{W}^{s}$, there exists $x_{0}\in O$ with $\dim P_{\rho}(x_{0})=K$. We can choose points $x_{1},\dots,x_{\ell}\in W^{s}(x_{0})\cap O$ such that $$P_{\rho}(x_{0})=\bigcap_{i=1}^{\ell}H_{x_{i},x_{0}}Q_{\rho}(x_{i}).$$ Since $Q_{\rho}$ and the holonomy maps $H_{x_{i},x_{0}}$ are continuous there exists an open neighborhood $O_{0}\subseteq O\cap W^{-}(x_{0})$ of $x_{0}$ such that the following holds. Given $y_{0}\in O_{0}$ there exist open neighborhoods $U_{i}$ of $x_{i}$ such that the connected component of $x_{i}$ in the set $U_{i}\cap W^{s}(y_{0})\cap W^{u}(x_{i})$ consists of a single point $y_{i}$, and moreover that $$\dim\,\bigcap_{i=1}^{\ell}H_{y_{i},y_{0}}Q_{\rho}(y_{0})\leq K.$$ By minimality of $K$ this forces$$P_{\rho}(y_{0})=\bigcap_{i=1}^{\ell}H_{y_{i},y_{0}}Q_{\rho}(y_{0});$$ in particular this proves we can find a neighborhood $V$ of $x_{0}$ in $\Sigma$ such that the restriction of $P_{\rho}$ to $V$ is a continuous subbundle of $E^{u}|V$. But then since $$\Sigma=\bigcup_{y\in V}W^{s}(y),$$ it follows that $P_{\rho}$ is a continuous subbundle of $E^{u}$ over all of $\Sigma$.
Unfortunately, it is not necessarily the case that $P_{\rho}$ is of positive dimension. To ensure this, we will need to choose $\rho\in\mathbb{C}$ carefully. Here is the general idea; the precise construction (due to Hamenstädt [@Hamenstaedt1995 Section 2]) is somewhat technical. We choose a periodic point $q$ of $\phi_{t}$, with period $T>0$, say. Then $d_{q}\phi_{T}:T_{q}\Sigma\rightarrow T_{q}\Sigma$ induces a map $A_{q}^{u}:E^{u}(q)\rightarrow E^{u}(q)$. Let $\mathbb{A}_{q}$ denote the complex linear extension of $A_{q}^{u}$ to a map $E_{\mathbb{C}}^{u}(q)\rightarrow E_{\mathbb{C}}^{u}(q)$, and let $\sigma\in\mathbb{R},\sigma>1$ denote the minimal absolute value of an eigenvalue of $\mathbb{A}_{q}$. We are interested in the subspace $S(q)\subseteq E^{u}(q)$ consisting of the subspace spanned by the union of the eigenspaces of $\mathbb{A}_{q}$ corresponding to eigenvalues of absolute value $\sigma$.
We then use holonomy transport to carry $S(q)$ to subspaces $S(x)\subseteq E^{u}(x)$ for $x\in W^{+}(q)$, thus creating a distribution $S$ over $E^{u}|W^{+}(q)$.\
What is the point of this construction? Suppose now $q\in O$ (where $O$ is the open set defined earlier on which $Q_{\rho}$ is a continuous subbundle of $E^{u}|O$ for all $\rho\in\mathbb{C}$ - such $q$ always exist since the set of periodic points of $\phi_{t}$ is dense in $\Sigma$, see Remark \[standard anosov rem\]). Then one shows that $L_{q}^{u}:E^{u}(q)\rightarrow E^{u}(q)$ preserves the subbundle $S(q)$, and thus $\mathbb{L}_{q}:E_{\mathbb{C}}^{u}(q)\rightarrow E_{\mathbb{C}}^{u}(q)$ preserves the complexification $S_{\mathbb{C}}(q)\subseteq E_{\mathbb{C}}^{u}(q)$. Moreover if $\rho_{0}$ is an eigenvalue of $\mathbb{L}_{q}|S_{\mathbb{C}}(q)$, we will show that the subspace $$Q_{\rho_{0}}(q)\cap S(q)\subseteq E^{u}(q)$$ (which is necessarily of positive dimension) is contained in $P_{\rho_{0}}(q)$; in other words, for this choice of $\rho_{0}$, $P_{\rho_{0}}$ is of positive dimension, and this gives us our desired subbundle of $E^{u}$ for the case where $n$ is odd.\
We will now begin with the details of the construction.\
Since $q$ is periodic of period $T$, $d_{q}\phi_{T}:T_{q}\Sigma\rightarrow T_{q}\Sigma$ defines a hyperbolic linear map $A_{q}:E(q)\rightarrow E(q)$ which preserves $E^{s}(q)$ and $E^{u}(q)$, and so defines maps $A_{q}^{s}:E^{s}(q)\rightarrow E^{s}(q)$ and $A_{q}^{u}:E^{u}(q)\rightarrow E^{u}(q)$. Let $\mathbb{A}_{q}:E_{\mathbb{C}}^{u}(q)\rightarrow E_{\mathbb{C}}^{u}(q)$ denote the linear map induced by $A_{q}^{u}$. Let $\sigma\in\mathbb{R}$ denote the smallest absolute value of an eigenvalue of $\mathbb{A}_{q}$ (note $\sigma>1$ as $d\phi_{t}|E^{u}$ is expanding).
Now let us set $$S(q):=\mbox{span}\left\{ \xi\in E^{u}(q)\,:\,\exists\,\zeta\in E_{\mathbb{C}}^{u}(q),\rho\in\mathbb{C},\left|\rho\right|=\sigma,\mbox{ with }\xi=\mbox{Re}\,\zeta,\ \mathbb{A}_{q}\zeta=\rho\zeta\right\} .$$
Then for $x\in W^{+}(q)$ we define $$S(x):=H_{q,x}[S(q)].$$ Then $S$ is a $C^{1}$-subbundle of $E^{u}|W^{+}(q)$ and moreover using and Lemma \[The-holonomy-transport=00003D00003D00003Ddif\] we see $d_{x}\phi_{t}[S(x)]\subseteq S(\phi_{t}x)$ for all $x\in W^{+}(q)$; in particular $A_{q}^{u}$ maps $S(x)$ to itself.\
\
We now make a little digression into some elementary linear algebra. Given an invertible complex linear endomorphism $A:\mathbb{C}^{m}\rightarrow\mathbb{C}^{m}$, decompose $\mathbb{C}^{m}=\bigoplus_{j=1}^{k}V^{\rho_{j}}$ into the root spaces of $A$, and let $\left\{ \xi_{1}^{\rho_{1}},\dots,\xi_{d(\rho_{1})}^{\rho_{1}},\xi_{1}^{\rho_{2}},\dots,\xi_{d(\rho_{2})}^{\rho_{2}},\dots,\xi_{1}^{\rho_{k}},\dots,\xi_{d(\rho_{k})}^{\rho_{k}}\right\} $ denote a basis of $\mathbb{C}^{m}$ such that $A$ is in Jordan normal form with respect to this basis. That is, we have $d(\rho_{j})=\dim\, V^{\rho_{j}}$, $A\xi_{1}^{\rho_{j}}=\rho_{j}\xi_{1}^{\rho_{j}}$ for $j=1,\dots,k$ and $A\xi_{i}^{\rho_{j}}=\xi_{1}^{\rho_{j}}+\xi_{2}^{\rho_{j}}+\dots+\xi_{i-1}^{\rho_{j}}+\rho_{j}\xi_{i}^{\rho_{j}}$ for $i>1$ and $j=1,\dots,k$. Now set $$\eta_{i}^{\rho_{j}}:=\frac{\rho_{j}}{\left|\rho_{j}\right|}\xi_{i}^{\rho_{j}},$$ and define a Hermitian inner product $\left\langle \cdot,\cdot\right\rangle $ on $\mathbb{C}^{m}$ by declaring $\{\eta_{i}^{\rho_{j}}\}$ to be a unitary basis.
Let $\sigma:=\min\{\left|\rho_{j}\right|\,:j=1\dots,k\}$, and set $$S:=\mbox{span}_{\mathbb{C}}\left\{ \xi_{1}^{\rho_{j}}\,:\,\left|\rho^{j}\right|=\sigma\right\} .$$ Then for any $\xi,\eta\in S$ we have $$\left\langle A\xi,A\eta\right\rangle =\sigma^{2}\left\langle \xi,\eta\right\rangle ,$$ The following result is essentially due to Hamenstädt.
Let $\xi,\eta\in\mathbb{C}^{n}$, with $\eta\in S$, $\eta\ne0$ and $\xi\notin S$. Then $$\lim_{k\rightarrow\infty}\frac{\left|A^{k}\xi\right|}{\left|A^{k}\eta\right|}=\infty.\label{eq:required limit}$$
The crux of the proof is the following formula, whose proof can be found in [@Hamenstaedt1995 Corollary 2.3]. Suppose $\zeta\in\mathbb{C}^{n}$ is a root vector for $A$ with eigenvalue $\rho\ne0$, that is, there exists $j\in\mathbb{N}$ such that$$(A-\rho\mbox{\mbox{Id}})^{j}\zeta=0\ \ \ \mbox{but}\ \ \ (A-\rho\mbox{\mbox{Id}})^{j-1}\zeta\ne0.$$ Then $$\lim_{k\rightarrow\infty}\frac{A^{k}\zeta}{\rho^{k}k^{j-1}}=\frac{1}{\rho^{j-1}(j-1)!}(A-\sigma\mbox{Id})^{j-1}\zeta.$$ In particular, if either:
1. $\rho,\rho'$ are eigenvalues of $A$ with $\left|\rho\right|<\left|\rho'\right|$ and $\zeta,\zeta'\in\mathbb{C}^{N}$ are root vectors for $A$ with eigenvalues $\rho,\rho'$ or
2. $\zeta$ and $\zeta'$ are both root vectors for $A$ with eigenvalue $\rho$, such that there exists $j\in\mathbb{N}$ with $$(A-\rho\mbox{Id})^{j}\zeta=0\ \ \ \mbox{but}\ \ \ (A-\rho\mbox{Id})^{j}\zeta'\ne0,$$
then $$\lim_{k\rightarrow\infty}\frac{\left|A^{k}\zeta'\right|}{\left|A^{k}\zeta\right|}=\infty.$$ This implies the lemma.
We return now to the problem at hand and prove the following result ([@Hamenstaedt1995 Lemma 2.5]).
\[lu preserves\]The map $L^{u}$ preserves the bundle $S$ over $E^{u}|W^{+}(q)$.
Before getting started on the proof, we introduce an auxilliary inner product that will be helpful for this result and its sequel. Let us fix an inner product $\left\langle \cdot,\cdot\right\rangle _{q}$ on $E^{u}(q)$ with the property that if $\xi,\eta\in S(q)$ then $\left\langle A_{q}^{u}\xi,A_{q}^{u}\eta\right\rangle _{q}=\sigma^{2}\left\langle \xi,\eta\right\rangle _{q}$ (such an inner product exists by the discussion above). Then extend $\left\langle \cdot,\cdot\right\rangle _{q}$ to an inner product $\left\langle \cdot,\cdot\right\rangle _{x}$ for all $x\in W^{s}(q)$ by$$\left\langle \xi,\eta\right\rangle _{x}:=\left\langle H_{x,q}(\xi),H_{x,q}(\eta)\right\rangle _{q}\ \ \ \mbox{for }x\in W^{s}(q),\;\xi,\eta\in E^{u}(x).$$ Let $\left|\cdot\right|_{x}$ denote the norm induced by $\left\langle \cdot,\cdot\right\rangle _{x}$ for $x\in W^{s}(q)$.
*(of Proposition \[lu preserves\])*
First note that by for any $x\in W^{s}(q)$ and $k\in\mathbb{Z}$, $$H_{\phi_{kT}x,q}\circ d_{x}\phi_{kT}=\left(A_{q}^{s}\right)^{k}\circ H_{x,q}\mbox{ as maps }E^{u}(x)\rightarrow E^{u}(q).\label{eq:important H+}$$ Thus if $\xi\in E^{u}(x)$, $\xi\notin S(x)$ and $\eta\in S(x),\eta\ne0$ we have $$\begin{aligned}
\frac{\left|d_{x}\phi_{kT}(\xi)\right|_{\phi_{kT}x}}{\left|d_{x}\phi_{kT}(\eta)\right|_{\phi_{kT}x}} & = & \frac{\left|H_{\phi_{kT}x,q}\bigl(d_{x}\phi_{kT}(\xi)\bigr)\right|_{q}}{\left|H_{\phi_{kT}x,q}\bigl(d_{x}\phi_{kT}(\eta)\bigr)\right|_{q}}\\
& = & \frac{\left|\left(A_{q}^{u}\right)^{k}H_{x,q}(\xi)\right|_{q}}{\left|\left(A_{q}^{u}\right)^{k}H_{x,q}(\eta)\right|_{q}},\end{aligned}$$ and then since $H_{x,q}(\eta)\in S(q)$ but $H_{x,q}(\xi)\notin S(q)$, the previous lemma tells us that$$\lim_{k\rightarrow\infty}\frac{\left|d_{x}\phi_{kT}(\xi)\right|_{\phi_{kT}x}}{\left|d_{x}\phi_{kT}(\eta)\right|_{\phi_{kT}x}}=\infty.$$ Now let $B$ denote the ball $$B=\left\{ x\in W^{s}(q)\,:\,\mbox{dist}(q,x)\leq1\right\} .$$ Using continuity and compactness of $B$, we see that the operator norm of $L^{u}$ with respect to the norm $\left|\cdot\right|$ on $E^{u}|B$ is uniformly bounded on $B$.
Suppose now for contradiction that there exists $x\in W^{s}(q)$ and $\xi\in S(x)$ with $L_{x}^{u}\xi\notin S(x)$. Then since $L_{x}^{u}$ is $\phi_{t}$-invariant we have $$L_{\phi_{kT}x}^{u}\left(d_{x}\phi_{kT}\left(\xi\right)\right)=d_{x}\phi_{kT}\left(L_{x}^{u}\xi\right)$$ for all $k\geq0$, and thus $$\lim_{k\rightarrow\infty}\frac{\left|L_{\phi_{kT}x}^{u}\bigl(d_{x}\phi_{kT}\left(\xi\right)\bigr)\Bigr)\right|_{\phi_{kT}x}}{\left|d_{x}\phi_{kT}\left(\xi\right)\right|_{\phi_{kT}x}}=\infty.$$ In other words, the operator norms with respect to $\left|\cdot\right|$ of the endomorphisms $L_{\phi_{kT}x}^{u}$ of $E^{u}(\phi_{kT}x)$ tends to infinity as $k\rightarrow\infty$, contradicting the fact that $\phi_{kT}x\in B$ for $k$ large enough.
Thus $L^{u}$ preserves $S$ over $E^{u}|W^{s}(q)$. To complete the proof we need to show $L^{u}$ preserves $S$ over all of $E^{u}|W^{+}(q)$. This however is clear, since $L^{u}$ is $\phi_{t}$-invariant.
Given $x\in W^{+}(q)$ we have $$L_{x}^{u}=H_{q,x}\circ L_{q}^{u}\circ H_{x,q}\mbox{ as maps }S(x)\rightarrow S(x).$$ In other words, when restricted to the subbundle $S$ over $W^{+}(q)$, the map $L^{u}$ ‘commutes with holonomy’. \[prop:comm\]
As before, since $L^{u}$ commutes with $d\phi_{t}$ it suffices to verify the assertion for $x\in W^{s}(q)$. Thus given $x\in W^{s}(q)$, define $C_{x}\in\mbox{End}\, S$ by $$C_{x}=L_{x}^{u}-H_{q,x}\circ L_{q}^{u}\circ H_{x,q}.$$ To complete the proof we show that $C_{x}=0$ for all $x\in W^{s}(q)$. To do this we shall show that the function $\beta$ defined on $W^{s}(q)$ by $$\beta(x)=\left\Vert C_{x}\right\Vert _{x}$$ (where here $\left\Vert \cdot\right\Vert _{x}$ denotes the operator norm on $E^{u}(x)$ with respect to $\left\langle \cdot,\cdot\right\rangle _{x}$) is invariant under $\phi_{T}$. Since $\beta$ is continuous and $\beta(q)=0$ it then follows $\beta$ is identically zero, and hence $C_{x}=0$ for all $x\in W^{s}(q)$.
Since $L^{u}$ is $\phi_{t}$-invariant we have $$C_{\phi_{T}x}\bigl(d_{x}\phi_{T}(\xi)\bigr)=d_{x}\phi_{T}\bigl(C_{x}(\xi)\bigr)$$ for all $x\in W^{s}(q)$ and $\xi\in S(x)$. Since (using )
$$\begin{aligned}
\left|d_{x}\phi_{T}(\xi)\right|_{x} & = & \left|H_{x,q}\bigl(d_{x}\phi_{T}(\xi)\bigr)\right|_{q}\\
& = & \left|A_{q}^{u}\bigl(H_{x,q}^{+}(\xi)\bigr)\right|_{q}\\
& = & \sigma\left|H_{x,q}^{+}(\xi)\right|_{q}\\
& = & \sigma\left|\xi\right|_{x},\end{aligned}$$
for all $\xi\in S(x)$ and $x\in W^{s}(q)$, for $k\geq0$ we have$$\begin{aligned}
\frac{\left|C_{\phi_{kT}x}\circ d_{x}\phi_{kT}(\xi)\right|_{\phi_{kT}x}}{\left|d_{x}\phi_{kT}(\xi)\right|_{\phi_{kT}x}} & = & \frac{\left|d_{x}\phi_{kT}\circ C_{x}(\xi)\right|_{\phi_{kT}x}}{\sigma^{k}\left|\xi\right|_{x}}\\
& = & \frac{\left|C_{x}(\xi)\right|_{x}}{\left|\xi\right|_{x}};\end{aligned}$$ which proves the assertions about the map $\beta$ stated above.
The point of all this work is the following result.
\[thm:hamen2\]Let $\rho_{0}\in\mathbb{C}$ be an eigenvalue of $\mathbb{L}_{q}|S_{\mathbb{C}}(q)$. Then the dimension of $P_{\rho_{0}}$ is strictly positive.
From the previous proposition it follows that the non-trivial subspace $Q_{\rho_{0}}(q)\cap S(q)$ is contained in $P_{\rho_{0}}(q)$, and hence in the notation of Theorem \[thm:hamen2-1\], the integer $K=\dim\, P_{\rho_{0}}$ is strictly positive.
We conclude this section with the following construction, again due to Hamenstädt [@Hamenstaedt1995 p686]. If $V$ is a real vector space and $V_{i}\subseteq V$ are even dimensional subspaces admitting almost complex structures $\mathbb{J}_{i}$, let $\overline{\bigcap}_{i}(V_{i},\mathbb{J}_{i})$ denote the largest subspace $W\subseteq\bigcap_{i}V_{i}$ which is invariant under the $\mathbb{J}_{i}$ and such that $\mathbb{J}_{i}|W=\mathbb{J}_{j}|W$ for all $i,j$. We call $\overline{\bigcap}_{i}(V_{i},\mathbb{J}_{i})$ the *complex intersection* of the $(V_{i},\mathbb{J}_{i})$.
Now suppose the $\rho_{0}$ we found above happens to lie in $\mathbb{C}\backslash\mathbb{R}$. Then if we consider the operator $$\mathbb{J}:=\frac{1}{\mbox{Im}\,\rho_{0}}\left(L^{u}-(\mbox{Re}\,\rho_{0})\mbox{Id}\right),$$ it is easy to see that $\mathbb{J}$ defines an almost complex structure on $Q_{\rho_{0}}$.
If follows that if we define a new bundle $P_{\rho_{0}}^{\mathbb{C}}$ over all of $\Sigma$ by setting $$P_{\rho_{0}}^{\mathbb{C}}(x):=\overline{\bigcap}\left\{ (H_{y,x}Q_{\rho_{0}}(y),\mathbb{J}_{y})\,:\, y\in W^{s}(x)\right\} ,\label{P ro complex}$$ then the same proof as above shows that $P_{\rho_{0}}^{\mathbb{C}}$ is a continuous $\phi_{t}$-invariant subundle of $E^{u}$ over $\Sigma$ of positive dimension. We shall use this observation later.
\[sec:North-South-Dynamics\]North-South Dynamics
================================================
In this section we return to the situation described in the introduction. The main goal of this section is to prove Theorem \[thm:north south dynamic\] below; we will use heavily the assumption that $M$ admits a metric of negative curvature. As stated in the introduction however it should be possible to prove Theorem \[thm:north south dynamic\] without using this assumption.
Throughout this section $(M,g)$ denotes a closed $n$-dimensional Riemannian manifold with tangent bundle $\pi:TM\rightarrow M$. We first begin with a quick summary of the geometry of the tangent bundle that we will need throughout what follows.
The geometry of $TM$ {#the-geometry-of-tm .unnumbered}
--------------------
The *vertical bundle* $V\subseteq TTM$ is given by $$V(v)=\ker\left\{ d_{v}\pi:T_{v}TM\rightarrow T_{x}M\right\} ,$$ where for convenience throughout this paragraph an arbitrary vector $v\in TM$ is assumed to lie in $T_{x}M$. The Riemannian metric $g$ on $M$ determines a direct summand $H$ of the vertical bundle $V$, called the *horizontal bundle* together with isomorphisms$$T_{v}TM\cong H(v)\oplus V(v)\cong T_{x}M\oplus T_{x}M.$$ Let $\nabla$ denote the Levi-Civita connection on $(M,g)$. We make this isomorphism explicit as follows: given $\xi\in T_{v}TM$ associate a point $(\xi_{H},\xi_{V})\in T_{x}M\oplus T_{x}M$, where $$\xi_{H}=d_{v}\pi(\xi)$$ and $$\xi_{V}=K_{v}(\xi).$$ Here $K:TTM\rightarrow TM$ is the connection map of $\nabla$, defined as follows: given $\xi\in T_{v}TM$, choose a curve $Z:(-\varepsilon,\varepsilon)\rightarrow TM$ such that $Z(0)=v$ and $\dot{Z}(0)=\xi$. Then $$K_{v}(\xi)=\nabla_{t}Z(0),$$ where $\nabla_{t}$ denotes the covariant derivative along the curve $\pi\circ Z$.
$H(v)$ is thus defined to be the set of $\xi\in T_{v}TM$ such that $\xi_{V}=0$, and similarly $V(v)$ is simply the set of $\xi\in T_{v}TM$ such that $\xi_{H}=0$. Clearly the map $\xi\mapsto\xi_{H}$ defines an isomorphism $H(v)\rightarrow T_{x}M$ and similarly $\xi\mapsto\xi_{V}$ defines an isomorphism $V(v)\rightarrow T_{x}M$. In general we shall slightly abuse notation and write $\xi=(\xi_{H},\xi_{V})$ to indicate this identification.\
It is easy to see that given $\xi,\eta\in TTM$ we have$$\omega_{0}(\xi,\eta)=\left\langle \xi_{H},\eta_{V}\right\rangle -\left\langle \xi_{V},\eta_{H}\right\rangle ,$$ where as before $\omega_{0}$ denotes the canonical symplectic form on $TM$. We define the *Sasaki metric* $g_{TM}$ on $TM$ by setting $$\left\langle \xi,\eta\right\rangle _{TM}:=\left\langle \xi_{H},\eta_{H}\right\rangle +\left\langle \xi_{V},\eta_{V}\right\rangle ,$$ so that $\omega_{0}$ and $g_{TM}$ are compatible.
North-South dynamics {#north-south-dynamics .unnumbered}
--------------------
As before, let $H:TM\rightarrow M$ denote the energy Hamiltonian $(x,v)\mapsto\frac{1}{2}\left|v\right|^{2}$, and let $F_{1}$ denote the symplectic gradient of $H$ with respect to $\omega_{1}$. Let $\phi_{t}^{1}:TM\rightarrow TM$ denote the flow of $F_{1}$. Let $\Sigma_{k}=H^{-1}(k)$ denote a closed energy level set, where $k$ is a regular value of $H$, and assume that $\phi_{t}^{1}|\Sigma_{k}$ is Anosov (we do not need to assume $C^{1}$ weak (un)stable bundles at this point). Write the Anosov splitting of $T\Sigma_{k}$ as $$T\Sigma_{k}=\mathbb{R}F_{1}\oplus E_{1}^{s}\oplus E_{1}^{u}.$$ We first quote the following theorem from [@PaternainPaternain1994], which will very useful in what follows.
\[thm:pat&pat theorem\]If $\Sigma_{k}$ is an Anosov energy level then the weak (un)stable bundles $E_{1}^{+}$ and $E_{1}^{-}$ are transverse to the vertical subbundle $V$.
Now let $\widetilde{M}$ denote the universal covering of $M$, and let $\widetilde{\Sigma}_{k}$ denote the pullback of $\Sigma_{k}$ to $T\widetilde{M}$. Then $\widetilde{\Sigma}_{k}$ is a smooth connected hypersurface of $T\widetilde{M}$ that intersects each tangent space $T_{x}\widetilde{M}$ in a sphere containing the origin in its interior (since the same is true of $\Sigma_{k}$). Let $\widetilde{\sigma}$ denote the pullback of $\sigma$ to $\widetilde{M}$ and let $\widetilde{\omega}_{0}$ denote the natural symplectic form on $T\widetilde{M}$. Let $\widetilde{\omega}_{1}:=\widetilde{\omega}_{0}+\widetilde{\pi}^{*}\widetilde{\sigma}$, where $\widetilde{\pi}:T\widetilde{M}\rightarrow\widetilde{M}$ is the footpoint map. We will let $\widetilde{F}_{1}$ denote the symplectic gradient of the lifted Hamiltonian $(x,v)\mapsto\frac{1}{2}\left|v\right|^{2}$ with respect to $\widetilde{\omega}_{1}$. We will write $\widetilde{\phi}_{t}^{1}:T\widetilde{M}\rightarrow T\widetilde{M}$ for the flow of $\widetilde{F}_{1}$. By assumption $\widetilde{\phi}_{t}^{1}|\widetilde{\Sigma}_{k}$ is Anosov, and we will write the Anosov splitting as $$T\widetilde{\Sigma}_{k}=\mathbb{R}\widetilde{F}_{1}\oplus\widetilde{E}_{1}^{s}\oplus\widetilde{E}_{1}^{u}.$$ Similarly let us denote by $\widetilde{V}$ and $\widetilde{H}$ denote the vertical and horizontal subbundles of $TT\widetilde{M}$.\
As before let $\mathcal{W}^{s},\mathcal{W}^{u},\mathcal{W}^{+}$ and $\mathcal{W}^{-}$ denote the four foliations of $\Sigma_{k}$ defined by the subbundles $E^{s},E^{u},E^{+}$ and $E^{-}$ respectively. We can lift these to foliations $\widetilde{\mathcal{W}}^{s},\widetilde{\mathcal{W}}^{u},\widetilde{\mathcal{W}}^{+}$ and $\widetilde{\mathcal{W}}^{-}$ of $\widetilde{\Sigma}_{k}$. Let $\mathcal{L}^{+}=\widetilde{\Sigma}_{k}/\widetilde{\mathcal{W}}^{+}$ and $\mathcal{L}^{-}=\widetilde{\Sigma}_{k}/\mathcal{\widetilde{\mathcal{W}}}^{-}$ denote the spaces of weak stable and unstable leaves respectively. The fundamental group $\pi_{1}(M)$ (regarded as covering transformations of $\widetilde{M}$) acts on $\widetilde{\Sigma}_{k}$ freely and properly discontinuously by permuting the orbits of $\widetilde{\phi}_{t}^{1}$. Since elements of $\pi_{1}(M)$ act by isometries, the action on $\widetilde{\Sigma}_{k}$ must send weak (un)stable leaves to weak (un)stable leaves, and thus $\pi_{1}(M)$ induces an action on $\mathcal{L}^{+}$ and $\mathcal{L}^{-}$.\
The aim of this section is to prove the following result.
\[thm:north south dynamic\]Suppose $M$ admits a metric of negative curvature. The fundamental group $\pi_{1}(M)$ acts on both $\mathcal{L}^{+}$ and $\mathcal{L}^{-}$ as a ‘North-South dynamic’. By this we mean the following: for all $\varphi\in\pi_{1}(M)$, there exists two fixed leaves $\widetilde{W}_{1}^{+},\widetilde{W}_{2}^{+}\in\mathcal{L}^{+}$ and two fixed leaves $\widetilde{W}_{1}^{-},\widetilde{W}_{2}^{-}\in\mathcal{L}^{-}$ such that for all $\widetilde{W}^{\pm}\in\mathcal{L}^{\pm}$ it holds that $$\lim_{n\rightarrow\infty}\varphi^{n}[\widetilde{W}^{\pm}]=\widetilde{W}_{1}^{\pm},\ \ \ \lim_{n\rightarrow\infty}\varphi^{-n}[\widetilde{W}^{\pm}]=\widetilde{W}_{2}^{\pm}.$$
Consider the fibration $$\pi|\Sigma_{k}:\Sigma_{k}\rightarrow M$$ of $\left(n-1\right)$-spheres. As a direct consequence of Theorem \[thm:pat&pat theorem\], we see that the foliations $\mathcal{W}^{+}$ and $\mathcal{W}^{-}$ are transverse to the fibres of the fibration $\Sigma_{k}\rightarrow M$. This implies that $\widetilde{\mathcal{W}}^{+}$ and $\widetilde{\mathcal{W}}^{-}$ are transverse to the fibration $\widetilde{\pi}|\widetilde{\Sigma}_{k}:\widetilde{\Sigma}_{k}\rightarrow\widetilde{M}$.
Given $x\in\widetilde{M}$, the fibre $(\widetilde{\pi}|\widetilde{\Sigma}_{k})^{-1}(x)$$ $ is compact. Thus we can apply the following theorem of Ehresmann.
Let $F\rightarrow E\overset{p}{\rightarrow}B$ be a fibre bundle and $\mathcal{F}$ a foliation of $E$ transverse to the fibres. Suppose $F$ is compact. Then for every leaf $L$ of $\mathcal{F}$, $p|L:L\rightarrow B$ is a covering map.
For a proof, see [@CamachoNeto1985 p91]. Now fix $v\in\widetilde{\Sigma}_{k}$. Since $\widetilde{M}$ is simply connected, $\widetilde{\pi}|\widetilde{W}^{+}(v)$$ $ is a diffeomorphism, and thus $\widetilde{W}^{+}(v)$ simply connected. Thus $\widetilde{W}^{+}(v)$ intersects each fibre of the fibration $\widetilde{\Sigma}_{k}\rightarrow\widetilde{M}$ in precisely one point, and thus each leaf $\widetilde{W}^{+}(v)$ is diffeomorphic to $\widetilde{M}$. Thus $\widetilde{M}$ is diffeomorphic to $\mathbb{R}^{n}$, and the space of stable leaves $\mathcal{L}^{+}$ can be identified topologically with the $\left(n-1\right)$-sphere. Of course the same applies to $\mathcal{L}^{-}$.\
Now we recall the concept of the ideal boundary of $\widetilde{M}$. For this, let $g_{0}'$ denote a metric of negative curvature on $M$ (whose existence we assume in this section), and lift $g_{0}'$ to a metric $g_{0}$ on $\widetilde{M}$ of negative curvature.
The *ideal boundary* ***$\widetilde{M}_{g_{0}}(\infty)$*** of $(\widetilde{M},g_{0})$ is given by $\widetilde{M}_{g_{0}}(\infty):=\Lambda_{g_{0}}(\widetilde{M})/\sim$, where $\Lambda_{g_{0}}(\widetilde{M})$ denotes the set of $g_{0}$-geodesics[^1] $c:\mathbb{R}\rightarrow\widetilde{M}$ of $\widetilde{M}$, and $c_{1}\sim c_{2}$ if and only if $$\mbox{dist}_{\textrm{HD}}(c_{1}[\mathbb{R}^{+}],c_{2}[\mathbb{R}^{+}])<\infty;$$ here $\mbox{dist}_{\textrm{HD}}$ denotes the *Hausdorff distance* defined by $$\mbox{dist}_{\textrm{HD}}(U,V):=\inf\left\{ r\in\mathbb{R}\,:\, U\subseteq B(V,r),V\subseteq B(U,r)\right\} ,$$ where $U,V\subseteq\widetilde{M}$ and $B(U,r):=\left\{ x\in\widetilde{M}\,:\,\mbox{dist}_{g_{0}}(x,U)\leq r\right\} $.
Given $x\in\widetilde{M}$ and $v\in T_{x}\widetilde{M}$, let $c_{v}:\mathbb{R}\rightarrow\widetilde{M}$ denotes the unique $g_{0}$-geodesic adapted to $v$, and let $c_{v}(\infty)\in\widetilde{M}_{g_{0}}(\infty)$ denote the corresponding element of $\widetilde{M}_{g_{0}}(\infty)$. If $c_{v}^{-1}$ is the geodesic obtained by going along $c_{v}$ backwards, let $c_{v}(-\infty)$ denote the element of $\widetilde{M}_{g_{0}}(\infty)$ corresponding to $c_{v}^{-1}$.
Let $S^{g_{0}}\widetilde{M}$ denote the unit sphere bundle of $(\widetilde{M},g_{0})$. Fix a point $x\in\widetilde{M}$, and consider the map $s_{x}:S_{x}^{g_{0}}\widetilde{M}\rightarrow\widetilde{M}_{g_{0}}^{g_{0}}(\infty)$ sending $v\mapsto c_{v}(\infty)$. Then $s_{x}$ is a bijection, and we define a topology on $\widetilde{M}_{g_{0}}(\infty)$ so that that $s_{x}$ becomes a homeomorphism; thus $\widetilde{M}_{g_{0}}(\infty)\cong S^{n-1}$. This topology is independent of the choice of $x$, since $s_{y}\circ s_{x}^{-1}:S_{x}^{g_{0}}\widetilde{M}\rightarrow S_{y}^{g_{0}}\widetilde{M}$ is a homeomorphism.\
The next thing we require is the concept of a quasi-geodesic.
A curve $\gamma:[a,b]\rightarrow\widetilde{M}$ is an *quasi-geodesic* of $(\widetilde{M},g_{0})$ if there exist $P,Q\in\mathbb{R}^{+}$ such that $$\frac{1}{P}\left|s-t\right|-Q\leq\mbox{dist}_{g_{0}}(\gamma(s),\gamma(t))\leq P\left|s-t\right|+Q$$ for all $s,t\in[a,b]$. If we need to be explicit about the constants $P,Q$, we call such a quasi-geodesic a *$(P,Q)$-quasi-geodesic*.
We now quote two theorems which explain why this is relevant to the situation in hand. The first is due to Peyerimhoff and Siburg ([@PeyerimhoffSiburg2003 Theorem 2.9]).
\[thm:Pey and Sib\]Suppose $\phi_{t}^{1}|\Sigma_{k}$ is Anosov. Then there exists a constant $P_{k}\in\mathbb{R}^{+}$ such that the projection to $\widetilde{M}$ of any orbit of $\widetilde{\phi}_{t}^{1}$ is a $(P_{k},0)$-quasi-geodesic.
In fact, Theorem \[thm:Pey and Sib\] is stated in a somewhat different form in [@PeyerimhoffSiburg2003]: there they assert that when $k$ is greater than a certain critical value $c(g,\sigma)$ known as *Mañé’s critical value* then the projection to $\widetilde{M}$ of any minimizing orbit of $\widetilde{\phi}_{t}^{1}$ is a quasi-geodesic. See for instance [@ContrerasIturriagaPaternainPaternain1998; @BurnsPaternain2002] for the definition of $c(g,\sigma)$, where it is proved that when $\phi_{t}^{1}$ is Anosov, it necessarily holds that $k>c(g,\sigma)$, and that in this case, every orbit of $\widetilde{\phi}_{t}^{1}$ is minimizing.
We can build the *quasi-ideal boundary* $\widetilde{M}_{g_{0}}^{*}(\infty)$ in much the same way using quasi-geodesics. If $\gamma:\mathbb{R}\rightarrow\widetilde{M}$ is an $g_{0}$-quasi-geodesic, we write $\gamma^{*}(\infty)$ to denote the corresponding element of $\widetilde{M}_{g_{0}}^{*}(\infty)$. Note that any geodesic is automatically a quasi-geodesic, and thus we have a natural map $\widetilde{M}_{g_{0}}(\infty)\hookrightarrow\widetilde{M}_{g_{0}}^{*}(\infty)$ carrying an equivalence class of geodesics to the corresponding equivalence class of quasi-geodesics. This is the second theorem we quote here; a proof may be found in [@Knieper2002 Theorem 2.2].
\[thm:Knei\]The inclusion $\widetilde{M}_{g_{0}}(\infty)\hookrightarrow\widetilde{M}_{g_{0}}^{*}(\infty)$ is a bijection.
We will use this to show the following key result, whose proof is essentially that of Theorem $2.12$ in [@Knieper2002]. The result however is originally due to Gromov (see [@Gromov2000]), and also independently due to Ghys ([@Ghys1984 Theorem 4.5]). Let $\psi_{t}$ denote the geodesic flow of $(\widetilde{M},g_{0})$ and let $\psi_{t}'$ denote the geodesic flow of $(M,g_{0}')$. Let $S^{g_{0}'}M$ denote the unit sphere bundle of $(M,g_{0}')$.
\[thm:gromov\]$\phi_{t}^{1}|\Sigma_{k}$ and $\psi'_{t}|S^{g_{0}'}M$ are topologically conjugate.
Given $v\in\widetilde{\Sigma}_{k}$, let $\gamma_{v}:=\widetilde{\pi}\circ\widetilde{\phi}_{t}^{1}v$. Then $\gamma_{v}$ determines an element $\gamma_{v}^{*}(\infty)\in\widetilde{M}_{g_{0}}^{*}(\infty)$ by Theorem \[thm:Pey and Sib\], and thus by Theorem \[thm:Knei\] a unique element $\xi_{v}\in\widetilde{M}_{g_{0}}(\infty)$. Let $\xi_{v}^{-1}\in\widetilde{M}_{g_{0}}(\infty)$ denote the element corresponding to $\gamma_{v}^{*}(-\infty)\in\widetilde{M}_{g_{0}}^{*}(\infty)$.
Suppose $\zeta,\xi\in\widetilde{M}_{g_{0}}(\infty)$. Then there exists a unique $g_{0}$-geodesic $c$ such that $c(\infty)=\zeta$ and $c(-\infty)=\xi$. Let $\mathbb{P}(c):\widetilde{M}\rightarrow\widetilde{M}$ denote orthogonal projection onto $c$, and use this to define a map $\mathbb{P}(\zeta,\xi):\widetilde{M}\rightarrow S^{g_{0}}\widetilde{M}$ by $$\mathbb{P}(\zeta,\xi)(x)=\dot{c}(t)\ \ \ \mbox{where\ \ \ }\mathbb{P}(c)(x)=c(t).$$ Now define $G_{0}:\widetilde{\Sigma}_{k}\rightarrow S^{g_{0}}\widetilde{M}$ by setting $$G_{0}(v):=\mathbb{P}(\xi_{v},\xi_{v}^{-1})(\widetilde{\pi}v),$$ Then $G_{0}$ is continuous and surjective but in general not injective: there may exist two points $v,v'$ on the same orbit of $\widetilde{\phi}_{t}^{1}$ that have the same orthogonal projection onto the $g_{0}$-geodesic $c$ determined by $\xi_{v}=\xi_{v'}$ and $\xi_{v}^{-1}=\xi_{v'}^{-1}$. In order to achieve injectivity we ‘average’ $G_{0}$. For this look at the map $\rho:\mathbb{R}\times\widetilde{\Sigma}_{k}\rightarrow S^{g_{0}}\widetilde{M}$ defined by $$G_{0}(\widetilde{\phi}_{t}^{1}v)=\psi_{\rho(t,v)}(G_{0}(v)).$$ Then $\rho$ satisfies the *cocycle property*, that is, $$\rho(t+t',v)=\rho(t,\widetilde{\phi}_{t'}^{1}v)+\rho(t',v),$$ as is easily checked. Now choose $\tau\in\mathbb{R}^{+}$ such that $\rho(\tau,v)>0$ for all $v\in\widetilde{\Sigma}_{k}$, and then let $r(v)$ denote the average$$r(v):=\frac{1}{\tau}\int_{0}^{\tau}\rho(t,v)dt.$$ Next define $G_{\tau}:\widetilde{\Sigma}_{k}\rightarrow S^{g_{0}}\widetilde{M}$ by $$G_{\tau}(v)=\psi_{r(v)}(G_{0}(v)).$$ We claim that $G_{\tau}$ is injective. For this observe that if $$f(t):=r(\widetilde{\phi}_{t}^{1}v)+\rho(t,v)$$ then $f$ is monotone increasing. Indeed, $$\begin{aligned}
f'(t) & = & \frac{1}{\tau}\int_{0}^{\tau}\rho'(u+t,v)du\\
& = & \frac{1}{\tau}(\rho(\tau+t,v)-\rho(t,v))\\
& = & \frac{1}{\tau}\rho(\tau,\widetilde{\phi}_{t}^{1}v)>0.\end{aligned}$$ The claim then follows from the computation$$\begin{aligned}
G_{\tau}(\widetilde{\phi}_{t}^{1}v) & = & \psi_{r(\psi_{t}v)}(G_{0}(\widetilde{\phi}_{t}^{1}v))\\
& = & \psi_{r(\psi_{t}v)+\rho(t,v)}(G_{0}(v))\\
& = & \psi_{f(t)}(G_{0}(v)).\end{aligned}$$ Finally, in order to deduce the stronger statement that $\phi_{t}^{1}|\Sigma_{k}$ and $\psi'_{t}|S^{g_{0}'}M$ are also topologically conjugate, one simply notes that $G_{0}$ is obviously equivariant under the action of $\pi_{1}(M)$, and hence so is $G_{\tau}$; thus $G_{\tau}$ descends to $M$ to define an orbit equivalence $G'_{\tau}$ from $\phi_{t}^{1}|\Sigma_{k}$ and $\psi'_{t}|S^{g_{0}'}M$.
It is now easy to prove Theorem \[thm:north south dynamic\]. Indeed, it is well known (see for instance [@Klingenberg1995 Theorem 3.8.13]) that Theorem \[thm:north south dynamic\] holds in the case of a geodesic flow of a negatively curved manifold. Thus if $M$ admits a metric $g_{0}'$ of negative curvature, Theorem \[thm:gromov\] gives us an orbit equivalence between $\phi_{t}^{1}|\Sigma_{k}$ and the geodesic flow $\psi_{t}'$ of $(M,g_{0}')$, and via this orbit equivalence we see Theorem \[thm:north south dynamic\] holds in our case too.
\[sec:Proof-of-Theorem A\]Proof of Theorem A
============================================
We will now prove Theorem A. Our proof of the theorem will depend on the parity of $n$; moreover, a separate argument will be required to deal with the cases $n=3$ and $n=7$. We will start with the case where $n$ is even.
The case when $n$ is an even integer. {#the-case-when-n-is-an-even-integer. .unnumbered}
-------------------------------------
In the even dimensional case, recall that we assume $\phi_{t}^{1}|\Sigma_{k}$ is $1/2$-pinched. Since $(\Sigma_{k},\omega_{1})$ is stable, there exists an invariant subbundle $P_{\rho_{0}}$ of $E^{u}$ as constructed in Section \[sec:Constructing-the-Invariant\]; see . The maximal integral submanifolds of $P_{\rho_{0}}$ define a foliation $\mathcal{P}(v)$ of class $C^{1}$ on $W^{u}(v)$. Since $P$ is invariant under parallel transport, it is also invariant under holonomy transport and thus the foliations $\mathcal{P}(v)$ glue together to give a foliation $\mathcal{P}$ of class $C^{1}$ on $\Sigma_{k}$ that is invariant under the parallel transport of $\nabla$ and thus also the holonomy maps. Thus $\mathcal{P}$ can be lifted to $\widetilde{\Sigma}_{k}$ and then projected to a foliation $\mathcal{P}'$ of $\mathcal{L}^{+}$ of positive dimension.
Since $\mathcal{P}$ is invariant under $\phi_{t}$, $\mathcal{P}'$ is invariant under the action of $\pi_{1}(M)$. Here Theorem \[thm:north south dynamic\] of the previous section comes into play: $\pi_{1}(M)$ acts on $\mathcal{L}^{+}$ as a North-South dynamics. A theorem of Foulon [@Foulon1994] states that there are no non-trivial $C^{0}$ foliations of the sphere $S^{n-1}$ which are invariant under North-South dynamics. Since we know that $\mathcal{P}'$ is of positive dimension, we must have $\mathcal{P}'=\mathcal{L}^{+}$, and hence the subbundle $P_{\rho_{0}}$ is equal to $E^{u}$. From this it is easy to deduce that $d\lambda=\rho_{0}\omega_{1}$, where $\rho_{0}\in\mathbb{R}$ defines $P_{\rho_{0}}$; see .
To complete the proof in this case it remains to rule out the possibility that $\rho_{0}=0$. Suppose for contradiction that this is the case. Then the $1$-form $\lambda$ is closed. Let $\mu$ denote a Borel probability measure on $\Sigma_{k}$. Recall that $\mu$ determines a *$1$-current* $l_{\mu}$ by $$l_{\mu}\left(\beta\right)=\int_{\Sigma_{k}}\beta\left(F_{1}\right)d\mu\,:\,\beta\in\Omega^{1}(\Sigma_{k}).$$ We say that $\mu$ is *exact as a current* if $l_{\mu}\left(\beta\right)=0$ whenever $\beta$ is closed. Now let $\mu_{L}$ denote the Liouville measure of $\Sigma_{k}$ (defined precisely below). Then by Lemma \[lem:exact as a current\] below, $\mu_{L}$ is exact as a current. But $$l_{\mu_{L}}\left(\lambda\right)=\int_{\Sigma_{k}}\lambda\left(F_{1}\right)d\mu_{L}\ne0;$$ contradiction. Thus $\rho_{0}\ne0$; this completes the proof in the even dimensional case.\
For completeness let us give a precise definition of the Liouville measure $\mu_{L}$ and prove that it is indeed exact as a current. Whilst in general there may exist many invariant volume forms on an energy level $\Sigma_{k}$, and thus many invariant probability measures, in the special case where the energy level $\Sigma_{k}$ is Anosov, the Liouville measure is the unique smooth invariant probability measure. It can be defined as follows. Let $X\in TTM|\Sigma_{k}$ denote a vector field such that $\omega_{1}(X,F_{1})=1$ (such a vector field always exists since $\Sigma_{k}$ is a regular energy level). Observe that $$i_{F_{1}}i_{X}\omega_{1}^{n}|\Sigma_{k}=\omega_{1}^{n-1},$$ and hence if $\Theta:=i_{X}\omega_{1}^{n}$ then $\Theta$ is a volume form on $\Sigma_{k}$. Now observe that $\omega_{1}^{n-1}$ is exact for $n\geq3$. Indeed $$\omega_{1}^{n-1}=(\omega_{0}+\pi^{*}\sigma)^{n-1}=(\omega_{0})^{n-1}+(n-1)\pi^{*}\sigma\wedge(\omega_{0})^{n-2}.$$ On the right-hand side the first term is exact. For $n\geq3$ the second term is exact as well and the claim follows (when $n=2$, $\omega_{1}$ is exact if $M$ is not the 2-torus, but we do not need this here). Then $i_{F_{1}}\Theta=d\tau$ and $\phi_{t}^{1}$ preserves the volume form $\Theta$. Let $\mu_{L}$ denote the smooth invariant probability measure induced by $\Theta$; $\mu_{L}$ is called the Liouville measure.
\[lem:exact as a current\]The Liouville measure $\mu_{L}$ of $\Sigma_{k}$ is exact as a current.
Let $\beta\in\Omega^{1}(\Sigma_{k},\mathbb{R})$ denote any closed $1$-form and let $A$ be the integral of $\Theta$. Then$$\begin{aligned}
A\, l_{\mu_{L}}(\beta) & = & A\,\int_{\Sigma_{k}}\beta(F_{1})d\mu_{L}\\
& = & \int_{\Sigma_{k}}\beta(F_{1})\Theta\\
& = & \int_{\Sigma_{k}}i_{F_{1}}\Theta\wedge\beta\\
& = & \int_{\Sigma_{k}}d\tau\wedge\beta\\
& = & \int_{\Sigma_{k}}d(\tau\wedge\beta),\end{aligned}$$ and this last integral is zero by Stokes’ theorem.
The case when $n$ is an odd integer. {#the-case-when-n-is-an-odd-integer. .unnumbered}
------------------------------------
We now proceed to the second case, where $n$ is odd. We no longer need to assume that $\phi_{t}^{1}|\Sigma_{k}$ is $1/2$-pinched, only that the weak (un)stable bundles are of class $C^{1}$. The next result is from [@Feres1991 Lemma 2].
\[lem:useful lemma from Feres\]Let $K\subseteq\widetilde{E}_{1}^{u}$ be a continuous distribution of $k$-dimensional planes defined everywhere on $\widetilde{\Sigma}_{k}$. Then $K$ projects onto a continuous field $Q$ of $k$-planes on $S^{n-1}$.
Let $C\subseteq\widetilde{M}$ denote a closed submanifold of $\widetilde{M}$ diffeomorphic to $S^{n-1}$. For each $x\in C$, let $\nu(x)\in T_{x}\widetilde{M}$ denote the inward pointing normal to $C$ at $x$. Thus $K(\nu(x))\subseteq\widetilde{E}_{1}^{u}(\nu(x))$. Then note that given $x\in C$ and $v\in\widetilde{E}_{1}^{s}(x)$, the map $d_{v}\widetilde{\pi}|\widetilde{E}_{1}^{u}:\widetilde{E}_{1}^{u}(v)\rightarrow v^{\perp}\subseteq T_{x}\widetilde{M}$ is a linear isomorphism. The desired continuous field $Q$ of $k$-planes is then given by$$Q(x):=d_{\nu(x)}\widetilde{\pi}\left[K(\nu(x))\right]\subseteq T_{x}C.$$
For the case where $M$ is odd dimensional we will use the bundle $P_{\rho_{0}}$ defined by , where this time $\rho_{0}\in\mathbb{C}$ is given by Corollary \[thm:hamen2\]. In this case however the argument is initially simpler. Indeed, by the following topological result, after lifting $P_{\rho_{0}}$ to a continuous distribution on $\widetilde{E}_{1}^{u}$, we see immediately that $P_{\rho_{0}}=E_{1}^{u}$.
For $n$ odd, $S^{n-1}$ admits no $k$-plane distribution for $1\leq k\leq n-2$.
See for instance [@Steenrod1951 Theorem 27.18] for a proof.
We could alternatively deduce the same result by observing as before that the space of leaves $\mathcal{L}^{+}:=\widetilde{\Sigma}_{k}/\widetilde{\mathcal{W}}^{+}$ is topologically an $\left(n-1\right)$-sphere, and since $P_{\rho_{0}}$ is invariant under the holonomy transport, $P_{\rho_{0}}$ determines a continuous distribution of $k$-planes on $S^{n-1}$. The proof given above does *not* use the fact that $P_{\rho_{0}}$ is invariant under holonomy transport.
Unfortunately this does not quite nail the result as in the even dimensional case. If $\rho_{0}\in\mathbb{R}$, we deduce $L^{u}=\rho_{0}\mbox{Id}$, and the desired contradiction follows just as in the even-dimensional case.
If however $\rho_{0}\notin\mathbb{R}$ then more work is required. Consider the $\phi_{t}^{1}$-invariant almost complex structure $\mathbb{J}$ on $E^{u}$ of class $C^{1}$ defined by$$\mathbb{J}:=\frac{1}{\mbox{Im}\,\rho_{0}}\left(L^{u}-(\mbox{Re}\,\rho_{0})\mbox{Id}\right).\label{eq:acs}$$ Lifting $\mathbb{J}$ to an almost complex structure on $\widetilde{E}_{1}^{u}$, and then using the construction from Lemma \[lem:useful lemma from Feres\], we see that $\mathbb{J}$ induces an almost complex structure on $S^{n-1}$. This immediately implies that $n=3$ or $n=7$, since the only spheres that admit almost complex structures are $S^{2}$ and $S^{6}$.
It thus remains to eliminate the cases $n=3$ and $n=7$, and we will tackle these separately.
The case where $n=3$ or $n=7$. {#the-case-where-n3-or-n7. .unnumbered}
------------------------------
The first step in the proof of these two special cases is to show that the existence of a $\phi_{t}^{1}$-invariant almost complex structure $\mathbb{J}$ on $E^{u}$ forces both $E^{s}$ and $E^{u}$ to be of class $C^{\infty}$. The next two results are due to Hamenstädt; see Corollary $2.11$ and Corollary $2.12$ of [@Hamenstaedt1995].
The almost complex structure $\mathbb{J}$ on $E^{u}$ is parallel with respect to the Kanai connection.
Since $\mathbb{J}$ is $\phi_{t}^{1}$-invariant and $\nabla_{F_{1}}=L_{F_{1}}$, we certainly have $\nabla_{F_{1}}\mathbb{J}=0$ and thus it suffices to show that $\nabla_{X_{s}}\mathbb{J}=0$ for all $X_{s}\in\Gamma(E^{s})$ and $\nabla_{X_{u}}\mathbb{J}=0$ for all $X_{u}\in\Gamma(E^{u})$. We know that $\mathbb{J}$ is invariant under the holonomy maps $H_{x,y}$ (since otherwise the subbundle $P_{\rho_{0}}^{\mathbb{C}}$ from would be of positive dimension - this contradict the fact that an even dimensional sphere does not have non-trivial subbundles) and thus as holonomy transport is the same as parallel transport for $\nabla$, we see that $\nabla_{X_{s}}\mathbb{J}=0$ for all $X_{s}\in\Gamma(E^{s})$. We can define a new almost complex structure $\mathbb{J}'$ on $E^{s}$ by the equation $$\omega(X_{u},\mathbb{J}'X_{s})=\omega(\mathbb{J}X_{u},X_{s}).$$ Then if $Y_{s}\in\Gamma(E^{s})$, applying $Y_{s}$ to the previous equation and using that $\nabla\omega=0$ shows that $$\omega(X_{u},(\nabla_{Y_{s}}\mathbb{J}')X_{s})=\omega((\nabla_{Y_{s}}\mathbb{J})X_{u},X_{s}),\label{eq:j}$$ and thus $\nabla_{Y_{s}}\mathbb{J}'=0$.
But now the point is the following: we could repeat all of what we have done above but working with $L^{s}$ not $L^{u}$, and thus obtain an almost complex structure $\widetilde{\mathbb{J}}$ on $E^{s}$. The same argument would then show that this almost complex structure $\widetilde{\mathbb{J}}$ is parallel on $E^{u}$, that is, $\nabla_{X_{u}}\widetilde{\mathbb{J}}=0$ for $X_{u}\in\Gamma(E^{u})$. But straight from the definition, it is clear that $\widetilde{\mathbb{J}}=\mathbb{J}'$, and thus $\mathbb{J}'$ is parallel on $E^{u}$ as well. But then taking $Y_{u}\in\Gamma(E^{u})$ and plugging it into , we see that $\nabla_{Y_{u}}\mathbb{J}=0$, and this completes the proof.
\[strong\]Suppose the strong unstable bundle $E_{1}^{u}$ admits an almost complex structure $\mathbb{J}$ defined as in . Then $\Sigma_{k}$ admits a real analytic structure, for which the strong (un)stable bundles $E_{1}^{s}$ and $E_{1}^{u}$ are both real analytic. Moreover, this real analytic structure is $C^{1}$ diffeomorphic to the underlying smooth structure of $\Sigma_{k}$.
Before starting the proof, let us recall the following facts about the spaces of leaves. Recall that the space $\mathcal{L}^{+}$ and $\mathcal{L}^{-}$ of leaves are defined to be the quotient spaces $\mathcal{L}^{+}:=\widetilde{\Sigma}_{k}/\sim_{+}$ and $\mathcal{L}^{-}:=\widetilde{\Sigma}_{k}/\sim_{-}$, where $\sim_{\pm}$ are the equivalence relations on $\widetilde{\Sigma}_{k}$ defined by $v\sim_{\pm}w$ if and only if $w\in\widetilde{W}^{\pm}(v)$ respectively. We let $p_{\pm}:\widetilde{\Sigma}_{k}\rightarrow\mathcal{L}^{\pm}$ denote the quotient maps, and give $\mathcal{L}^{\pm}$ the quotient topology induced by $p_{\pm}$. For each $x\in\widetilde{M}$ the restriction $p_{+}|\widetilde{\Sigma}_{k}\cap T_{x}\widetilde{M}$ and $p_{-}|\widetilde{\Sigma}_{k}\cap T_{x}\widetilde{M}$ are homeomorphisms onto $\mathcal{L}^{+}$ and $\mathcal{L}^{-}$ respectively; thus $\mathcal{L}^{+}$ and $\mathcal{L}^{-}$ both admit the structure of topological manifolds, and are homeomorphic to $S^{n-1}$.
The key fact that we will need is that there exists a homeomorphism $F:\mathcal{L}^{+}\rightarrow\mathcal{L}^{-}$ with the following property: given any leaves $\widetilde{W}^{+}\in\mathcal{L}^{+}$ and $\widetilde{W}^{-}\in\mathcal{L}^{-}$, the intersection $\widetilde{W}^{+}\cap\widetilde{W}^{-}$ contains a unique flow line of $\widetilde{\phi}_{t}^{1}$ unless $\widetilde{W}^{-}=F(\widetilde{W}^{+})$, in which case $\widetilde{W}^{+}\cap\widetilde{W}^{-}=\emptyset$. We will call $F$ a *flip map*.
To prove this we take advantage of the fact that $\widetilde{M}$ admits a metric $g_{0}$ of negative curvature again. As before let $S^{g_{0}}\widetilde{M}$ denote the unit sphere bundle of $(\widetilde{M},g_{0})$ and $\psi_{t}:S^{g_{0}}\widetilde{M}\rightarrow S^{g_{0}}\widetilde{M}$ the geodesic flow of $(\widetilde{M},g_{0})$. Let $\mathcal{L}_{g_{0}}^{+}$ and $\mathcal{L}_{g_{0}}^{-}$ denote the space of stable and unstable leaves determined by $\psi_{t}$. Defining similar equivalence relations $\sim_{+}^{g_{0}}$ and $\sim_{-}^{g_{0}}$ on $S^{g_{0}}\widetilde{M}$, we can realise $\mathcal{L}_{g_{0}}^{+}$ and $\mathcal{L}_{g_{0}}^{-}$ as quotients of $S^{g_{0}}\widetilde{M}$; let $p_{+}^{g_{0}}:S^{g_{0}}\widetilde{M}\rightarrow\mathcal{L}_{g_{0}}^{+}$ and $p_{-}^{g_{0}}:S^{g_{0}}\widetilde{M}\rightarrow\mathcal{L}_{g_{0}}^{-}$ denote the projections. Theorem \[thm:north south dynamic\] ensures we have a homeomorphism $G=G_{\tau}:\widetilde{\Sigma}_{k}\rightarrow S^{g_{0}}\widetilde{M}$ that conjugates the orbits of $\widetilde{\phi}_{t}^{1}$ and $\psi_{t}$, and this conjugacy then induces maps $G_{+}:\mathcal{L}^{+}\rightarrow\mathcal{L}_{g_{0}}^{+}$ and $G_{-}:\mathcal{L}^{-}\rightarrow\mathcal{L}_{g_{0}}^{-}$. $ $It is well known that the map $S^{g_{0}}\widetilde{M}\rightarrow S^{g_{0}}\widetilde{M}$ sending $v\mapsto-v$ induces a flip map $F_{0}:\mathcal{L}_{g_{0}}^{+}\rightarrow\mathcal{L}_{g_{0}}^{-}$ for the flow $\psi_{t}$, and thus if $F:=G_{-}^{-1}\circ F_{0}\circ G_{+}$ then $F:\mathcal{L}^{+}\rightarrow\mathcal{L}^{-}$ is a flip map for $\widetilde{\phi}_{t}^{1}$.
Among other things, the existence of a flip map gives us another way to view the holonomy transport. Given $v\in\widetilde{\Sigma}_{k}$, the restriction $p_{+}|\widetilde{W}^{u}(v)$ is a homeomorphism onto the set $\mathcal{L}^{+}\backslash F(\widetilde{W}^{+}(v))$. Thus there is a unique map $p_{+}^{v}:\widetilde{\Sigma}_{k}\backslash F(\widetilde{W}^{+}(v))\rightarrow\widetilde{W}^{u}(v)$ such that $p_{+}\circ p_{+}^{v}=p_{+}$. Given $w\in\widetilde{W}^{+}(v)$, if $$V:=\widetilde{W}^{u}(w)\backslash F(\widetilde{W}^{+}(w))\label{eq:def of v}$$ then it is easy to see that $p_{+}^{v}|V$ is precisely the restriction to $V$ of the holonomy map $\widetilde{\mathcal{H}}_{w,v}^{\gamma}:\widetilde{W}^{u}(w)\rightarrow\widetilde{W}^{u}(v)$. We will use this fact in the proof below.
*(of Theorem \[strong\])*
Fix $v\in\Sigma_{k}$, and suppose $X$ and $Y$ are parallel sections of $E^{u}$ over $W^{u}(v)$. Then $[X,Y]=0$ as $\nabla$ is torsion free when restricted to $W^{u}(v)$. Since $\mathbb{J}$ is parallel by the previous lemma, so are $\mathbb{J}X$ and $\mathbb{J}Y$; hence $[\mathbb{J}X,\mathbb{J}Y]=[\mathbb{J}X,Y]=[X,\mathbb{J}Y]=0$; in other words the Nijenhuis tensor $N_{\mathbb{J}}$ of $\mathbb{J}$ vanishes, and consequently $\mathbb{J}$ is integrable (see for instance, [@KobayashiNomizu1963a Chapter IX, Theorem 2.4]). Thus we obtain a complex structure on $W^{u}(v)$. Using the almost complex structure $\mathbb{J}'$ defined in the previous lemma (which we shall now also refer to simply as $\mathbb{J}$) and repeating the same argument gives us complex structures on the stable manifolds $W^{s}(v)$.
Let us now pass to the universal cover. We have shown that each of the strong stable and unstable leaves $\widetilde{W}^{s}$ and $\widetilde{W}^{u}$ admit $\pi_{1}(M)$-equivariant complex structures. We will use this to define the structure of a complex manifold on the spaces $\mathcal{L}^{+}$ and $\mathcal{L}^{-}$ of leaves. We have shown above that given $v\in\widetilde{\Sigma}_{k}$, the map $$p_{+}|\widetilde{W}^{u}(v):\widetilde{W}^{u}(v)\rightarrow U(v):=\mathcal{L}^{+}\backslash F(\widetilde{W}^{+}(v))$$ is a homeomorphism. Let $\varphi_{v}:U(v)\rightarrow\widetilde{W}^{u}(v)\cong\mathbb{C}^{m}$ (here $n=2m+1$) denote the inverse map. We want to define the structure of a complex manifold on $\mathcal{L}^{+}$ by taking $\{(\varphi_{v},U(v)\,:\, v\in\widetilde{\Sigma}_{k}\}$ to be an atlas. The fact that the overlap maps $\varphi_{v}\circ\varphi_{w}^{-1}$ are holomorphic where defined follows immediately from the fact that if $v\in\widetilde{\Sigma}_{k}$, $w\in\widetilde{W}^{+}(v)$, and $V\subseteq\widetilde{W}^{u}(v)$ is as defined above in then $$d_{w}(p_{+}^{v}|V)=\widetilde{H}_{w,v}|V,$$ and thus $p_{+}^{v}|V$ is holomorphic, since the lifted complex structure $\widetilde{\mathbb{J}}$ is invariant under the holonomy transport maps.
Similarly $\mathcal{L}^{-}$ admits the structure of a complex manifold. Note that by construction the complex structure on $\mathcal{L}^{+}$ and $\mathcal{L}^{-}$ is $\pi_{1}(M)$-equivariant. Next, we claim that the orbit space $\widetilde{\Sigma}_{k}/\widetilde{\phi}_{t}^{1}$ is homeomorphic to $\mathcal{L}^{+}\times\mathcal{L}^{-}\backslash K$ where $K$ is a closed set. Set $$\Delta:=\left\{ (p_{+}^{g_{0}}(v),p_{-}^{g_{0}}(v))\in\mathcal{L}_{g_{0}}^{+}\times\mathcal{L}_{g_{0}}^{-}\,:\, v\in S^{g_{0}}\widetilde{M}\right\} .$$ The following is well known: given a flow line $C_{v}(t)=\psi_{t}v$, the map $$C_{v}\mapsto(p_{+}^{g_{0}}(v),F_{0}(p_{-}^{g_{0}}(v)))\in\mathcal{L}_{g_{0}}^{+}\times\mathcal{L}_{g_{0}}^{-}$$ defines a homeomorphism between the orbit space $S^{g_{0}}\widetilde{M}/\psi_{t}$ of $\psi_{t}$ and $\mathcal{L}_{g_{0}}^{+}\times\mathcal{L}_{g_{0}}^{-}\backslash\Delta$. It follows that if $$K:=\left\{ (p_{+}(v),p_{-}(v))\in\mathcal{L}^{+}\times\mathcal{L}^{-}\,:\, v\in\widetilde{\Sigma}_{k}\right\} ,$$ then given a flow line $\Gamma_{v}(t)=\widetilde{\phi}_{t}^{1}v$, the map $$\Gamma_{v}\mapsto(p_{+}(v),F(p_{-}(v)))\in\mathcal{L}^{+}\times\mathcal{L}^{-}$$ defines a homeomorphism from $\widetilde{\Sigma}_{k}/\widetilde{\phi}_{t}^{1}$ to $\mathcal{L}^{+}\times\mathcal{L}^{-}\backslash K$.
This essentially completes the proof. Indeed, we have shown that $\mathcal{L}^{+}$ and $\mathcal{L}^{-}$ carry $\pi_{1}(M)$-equivariant complex structures, that the orbit space $\widetilde{\Sigma}_{k}/\widetilde{\phi}_{t}^{1}$ is $C^{1}$ equivalent to the complex manifold $\mathcal{L}^{+}\times\mathcal{L}^{-}\backslash K$ provided with the product complex structure, and that this correspondence is equivariant under the action of $\pi_{1}(M)$. Thus $\Sigma_{k}$ carries a real analytic structure for which the bundles $E^{+}$ and $E^{-}$ are real analytic, which by construction is $C^{1}$ diffeomorphic to the underlying smooth structure on $\Sigma_{k}$.
Write $\rho_{0}=\sigma+it$, so that $\mathbb{J}=\sigma^{-1}(L^{u}-t\mbox{Id})$. Thus $L^{u}=\sigma J+t\mbox{Id}$, and thus if $K:=\left|\rho_{0}\right|^{-2}\cdot(-\sigma\mathbb{J}+t\mbox{Id})$ then $L^{u}\circ K=K\circ L^{u}=\mbox{Id}$. Then since $d\lambda(\cdot,\cdot)=\omega_{1}(L^{s}(\cdot),\cdot)$ and $\omega_{1}^{n-1}\ne0$, it follows that $(d\lambda)^{n-1}\ne0$ and thus $\lambda\wedge(d\lambda)^{n-1}$ is a volume form.
Now we quote the following theorem of Benoist, Foulon and Labourie ([@BenoistFoulonLabourie1992], Theorem $1$).
Let $N$ be a closed manifold of odd dimension $2n-1$ with $n\geq3$ an odd integer. Let $\phi_{t}:N\rightarrow N$ be an Anosov flow with infinitesimal generator $F$. Suppose that the strong bundles $E^{s}$ and $E^{u}$ are of class $C^{\infty}$. Define a $1$-form $\lambda$ by $\ker\,\lambda=E^{s}\oplus E^{u}$ and $\lambda(F)=1$ and suppose $\lambda\wedge(d\lambda)^{n-1}$ is a volume form. Then there exists a unique cohomology class $\eta\in H^{1}(N,\mathbb{R})$ and a closed $1$-form $\beta$ representing $\eta$ such that $1+\beta(F)>0$ and such that if $Y:=(1+\beta(F))^{-1}F$ then the flow of $Y$ is $C^{\infty}$ conjugate to the geodesic flow of a Riemannian manifold of constant negative curvature.
Thus we conclude that there exists a Riemannian manifold $(N,h_{0})$ of constant negative curvature, a $C^{1}$-diffeomorphism $G:SN\rightarrow\Sigma_{k}$ and a constant $c>0$ such that $$\rho_{t}\circ G=G\circ\tau_{ct},\label{eq:conjugacy}$$ where $\rho_{t}$ is the flow of the vector field $Y=(1+\beta(F_{1}))^{-1}F_{1}$, where $\beta\in\Omega^{1}(\Sigma_{k})$ is the closed $1$-form given by the theorem above, and $\tau_{t}:SN\rightarrow SN$ is the geodesic flow of $N$. In fact, by the main result in [@Hamenstaedt1993], $G$ is of class $C^{2}$.
Actually for the case $n=3$ we could bypass the above and conclude immediately that $\phi_{t}^{1}$ is $C^{\infty}$-orbit equivalent to the geodesic flow of a closed three-dimensional hyperbolic manifold using recent work of Fang [@Fang2007]. Namely, for $n=3$ (so $\dim\, E^{u}=2$), an almost complex structure is the ‘same’ as having a conformal structure, and thus the fact that $E^{u}$ admits a $\phi_{t}^{1}$-invariant almost complex structure $\mathbb{J}$ implies that $\phi_{t}^{1}$ is *quasiconformal* (see [@Fang2007]). Then [@Fang2007 Theorem 3] tells us that up to finite covers, $\phi_{t}^{1}$ is $C^{\infty}$-orbit equivalent to the suspension of a symplectic hyperbolic automorphism or to the geodesic flow of a closed three-dimensional hyperbolic manifold. The former is impossible by Theorem \[thm:north south dynamic\]. This method does not appear to work for $n=7$ however.
Let $\Omega_{\textrm{inv}}^{2}(\Sigma_{k})$ denote the set of $2$-forms on $\Sigma_{k}$ that are invariant under $\phi_{t}^{1}$. It is easy to see that $2$-forms invariant under $\phi_{t}^{1}$ are precisely the same as the $2$-forms invariant under $\rho_{t}$.
Similarly let $\Omega_{\textrm{inv}}^{2}(SN)$ denote the set of $2$-forms on $SN$ that are invariant under $\tau_{t}$. Note that $\omega_{1}$ and $d\lambda$ both lie in $\Omega_{\textrm{inv}}^{2}(\Sigma_{k})$, and hence by the $2$-forms $G^{*}\omega_{1}$ and $G^{*}d\lambda$ on $SN$ both lie in $\Omega_{\textrm{inv}}^{2}(SN)$.\
The following result is due to Kanai ([@Kanai1993 Claim 3.3]).
\[pro:kanai prop\]If $n\geq3$ then $\Omega_{\textrm{\emph{inv}}}^{2}(SN)$ is $1$-dimensional (where $\dim\, N=2n-1$), spanned by the canonical symplectic form $\omega_{N}$ on $TN$. If $n=2$ then $\Omega_{\textrm{\emph{inv}}}^{2}(SN)$ is $2$-dimensional, spanned by $\omega_{N}$ and another $2$-form $\psi_{N}$. This form can be uniquely characterized as follows. Let $\Pi:S\mathbb{H}^{3}\rightarrow SN$ denote the universal covering of $SN$, where $\mathbb{H}^{3}$ denotes hyperbolic $3$-space. Then there exists a $2$-form $\psi\in\Omega^{2}(S\mathbb{H}^{3})$ which we will define below, and $\psi_{N}$ is then the unique $2$-form on $SN$ such that $p^{*}\psi_{N}=\psi$.
The theorem immediately implies the result for the case $n=7$. Indeed, we deduce that $G^{*}\omega_{1}=c_{1}\omega_{N}$ and $G^{*}d\lambda=c_{2}\omega_{N}$ for two constants $c_{1},c_{2}\in\mathbb{R}$ with $c_{1},c_{2}\ne0$, and from this it is clear that $d\lambda=\rho\omega_{1}$ for some constant non-zero $\rho\in\mathbb{R}$. The case $n=3$ still requires a little work though.\
Now we define the $2$-form $\psi$ on $S\mathbb{H}^{3}$ mentioned in the statement of Theorem \[pro:kanai prop\]. The Jacobi equation for $\mathbb{H}^{3}$ is given by $$\ddot{J}-J=0,$$ (where the dot denotes the covariant derivative along $\gamma$) and thus it is easy to see that given a geodesic $\gamma$, the normal Jacobi fields along $\gamma$ are linear combinations of the fields $$J(t)=e^{\pm t}U(t),$$ where $U(t)$ is any parallel normal vector field along $\gamma$. Let $\phi_{t}^{\mathbb{H}}$ denote the geodesic flow on $\mathbb{H}^{3}$. Then $\phi_{t}^{\mathbb{H}}$ is Anosov (see Example \[exa:example sec curv\]). Given $v\in S\mathbb{H}^{3}$, let $\gamma_{v}$ denote the unique geodesic adapted to $v$. Then, given $\xi\in T_{v}S\mathbb{H}^{3}$, let $J_{\xi}$ denote the unique Jacobi field along $\gamma_{v}$ with $J_{\xi}(0)=d_{v}\pi_{\mathbb{H}}(\xi)$ and $\dot{J}_{\xi}(0)=K_{\mathbb{H},v}(\xi)$ (here $\pi_{\mathbb{H}}:S\mathbb{H}^{3}\rightarrow\mathbb{H}^{3}$ and $K_{\mathbb{H}}:TT\mathbb{H}^{3}\rightarrow T\mathbb{H}^{3}$ denote the footpoint and connection maps of $\mathbb{H}^{3}$ respectively.
Using the fact that $d\phi_{t}^{\mathbb{H}}(\xi)=(J_{\xi}(t),\dot{J}_{\xi}(t))$ (see for instance [@Paternain1999 Section 1.5]), it easily follows that the Anosov splitting $S\mathbb{H}^{3}=\mathbb{R}F_{\mathbb{H}}\oplus E_{\mathbb{H}}^{s}\oplus E_{\mathbb{H}}^{u}$ is given by $$E_{\mathbb{H}}^{s}(v)=\left\{ \xi\in T_{v}S\mathbb{H}^{3}\,:\,\xi_{H}=-\xi_{V}\right\} ,$$ and similarly $$E_{\mathbb{H}}^{u}(v)=\left\{ \xi\in T_{v}S\mathbb{H}^{3}\,:\,\xi_{H}=\xi_{V}\right\} ,$$ $$\mathbb{R}F_{\mathbb{H}}(v)=\left\{ \xi\in T_{v}S\mathbb{H}^{3}\,:\,\xi_{V}=0,\xi_{H}=av,a\in\mathbb{R}\right\} .$$ Using this decomposition we can define an almost complex structure on the subbundle $E_{\mathbb{H}}=E_{\mathbb{H}}^{s}\oplus E_{\mathbb{H}}^{u}$. Indeed, fix $v\in S\mathbb{H}^{3}$ and note that the isomorphism $$T_{v}T\mathbb{H}^{3}\cong T_{x}\mathbb{H}^{3}\oplus T_{x}\mathbb{H}^{3}$$ described at the beginning of Section \[sec:North-South-Dynamics\] restricts to define an isomorphism$$E_{\mathbb{H}}(v)\cong v^{\perp}\oplus v^{\perp},$$ where $$v^{\perp}:=\left\{ w\in T_{x}\mathbb{H}^{3}\,:\,\left\langle w,v\right\rangle =0\right\} .$$ Now $v^{\perp}$ is $2$-dimensional, and let $\{e_{1}(v),e_{2}(v)\}\subset v^{\perp}$ be an orthonormal basis such that $\{e_{1}(v),e_{2}(v),v\}$ is a positively oriented basis of $T_{x}\mathbb{H}^{3}$. This allows us to define a map $i_{v}:v^{\perp}\rightarrow v^{\perp}$ by $i_{v}e_{1}(v)=e_{2}(v)$ and $i_{v}e_{2}(v)=-e_{1}(v)$.
Now define $\mathbb{J}_{v}:E_{\mathbb{H}}^{s}(v)\rightarrow E_{\mathbb{H}}^{s}(v)$ by $$\mathbb{J}_{v}(\xi_{H},-\xi_{H})=(i_{v}\xi_{H},-i_{v}\xi_{H}),$$ and define $\mathbb{J}_{v}:E_{\mathbb{H}}^{u}(v)\rightarrow E_{\mathbb{H}}^{u}(v)$ by $$\mathbb{J}_{v}(\xi_{H},\xi_{H})=(-i_{v}\xi_{H},-i_{v}\xi_{H}).$$ This defines an almost complex structure on $E_{\mathbb{H}}$. For convenience, extend $\mathbb{J}$ to all of $TS\mathbb{H}^{3}$ by letting $\mathbb{J}|\mathbb{R}F_{\mathbb{H}}\equiv0$.
Finally, we define $\psi\in\Omega^{2}(S\mathbb{H}^{3})$ by $$\psi_{v}(\xi,\eta)=\omega_{\mathbb{H}}|_{v}(\mathbb{J}_{v}\xi,\eta)\ \ \ \mbox{for\ \ \ }\xi,\eta\in T_{v}S\mathbb{H}^{3}.$$ This form $\psi$ is the $2$-form referred in the statement of Theorem \[pro:kanai prop\]. Note that $\psi$ is closed because $d\psi$ is an invariant 3-form which must vanish by the proof of Lemma \[lem:.dlambda is parallel\].\
The sphere bundle $S\mathbb{H}^{3}$ is trivial: $S\mathbb{H}^{3}=\mathbb{H}^{3}\times S^{2}$. Given $x\in\mathbb{H}^{3}$, let $S_{x}$ be the 2-sphere of unit vectors at $x$. Let us make the following observation which shows that $[\psi]\in H^{2}(S\mathbb{H}^{3},\mathbb{R})=\mathbb{R}$ is not zero.
\[lem:The-class- not trivial\]For all $x\in\mathbb{H}^{3}$ we have $$\int_{S_{x}}\psi\ne0.$$
Fix $x\in\mathbb{H}^{3}$ and $v\in S_{x}\mathbb{H}^{3}$. Let us take two vectors $\xi,\eta\in V(v)$, with say, $$\xi=(0,\xi_{V})\in T_{x}\mathbb{H}^{3}\oplus T_{x}\mathbb{H}^{3},\ \ \ \eta=(0,\eta_{V})\in T_{x}\mathbb{H}^{3}\oplus T_{x}\mathbb{H}^{3}.$$ It suffices to observe that $\psi_{v}(\xi,\eta)\ne0$ if $\xi$ and $\eta$ are not colinear. We need to express $\xi$ as a sum of elements of $E_{\mathbb{H}}^{s}(v)$ and $E_{\mathbb{H}}^{u}(v)$: this is easily done by noting$$\xi=\underset{\in E_{\mathbb{H}}^{s}(v)}{\underbrace{\frac{1}{2}(-\xi_{V},\xi_{V})}}+\underset{\in E_{\mathbb{H}}^{u}(v)}{\underbrace{\frac{1}{2}(\xi_{V},\xi_{V})}}.$$ Then we have$$\mathbb{J}_{v}(\xi)=\frac{1}{2}(-i_{v}\xi_{V},+i_{v}\xi_{V})+\frac{1}{2}(-i_{v}\xi_{V},-i_{v}\xi_{V})=-(i_{v}\xi_{V},0).$$ Thus:$$\begin{aligned}
\psi_{v}(\xi,\eta) & = & \omega_{\mathbb{H}}|_{v}(\mathbb{J}_{v}\xi,\eta)\\
& = & -\left\langle i_{v}\xi_{V},\eta_{V}\right\rangle ,\end{aligned}$$ and this is non-zero if $\xi$ and $\eta$ are not colinear.
From this it is now easy to complete the proof in the case $n=3$. Since both $\omega_{0}$ and $d\lambda$ are exact (recall $\omega_{0}=-d\alpha$; see ) it follows that both $\Pi^{*}G^{*}\omega_{1}$ and $\Pi^{*}G^{*}d\lambda$ vanish in $H^{2}(S\mathbb{H}^{3},\mathbb{R})$ (since $\omega_{1}$ is equal to $\omega_{0}$ plus the pullback of a form on the base), and thus they must both be multiples of $\omega_{\mathbb{H}}$; that is, $$\Pi^{*}G^{*}\omega_{1}=c_{1}\omega_{\mathbb{H}},\ \ \ \Pi^{*}G^{*}d\lambda=c_{2}\omega_{\mathbb{H}}\,:\, c_{1},c_{2}\in\mathbb{R}\backslash\{0\}.$$ Hence again $d\lambda=\rho\omega_{1}$ for some non-zero $\rho\in\mathbb{R}$, which completes the proof for the case $n=3$, and thus finally completes the proof of Theorem A.
\[sec:Twisted-Geodesic-Flows\]Proof of the Corollary B {#sectwisted-geodesic-flowsproof-of-the-corollary-b .unnumbered}
------------------------------------------------------
Now we prove Corollary B from the introduction.
\[pro:ANOSOV 12PINCHED\]Suppose $g$ is negatively curved and strictly $1/4$-pinched. For sufficiently large $k$, the Hamiltonian structure $\left(\Sigma_{k},\omega_{1}\right)$ is Anosov and satisfies the $1/2$-pinching condition.
First of all we will show that $\left(\Sigma_{k},\omega_{1}\right)$ is an Anosov Hamiltonian structure for $k$ sufficiently large. It is well known that $\phi_{t}^{0}:SM\rightarrow SM$ is Anosov. By structural stability of Anosov flows, there exists $\delta_{0}>0$ such that $\phi_{t}^{\varepsilon}:SM\rightarrow SM$ is Anosov for all $\varepsilon\in(-\delta_{0},\delta_{0})$ (see for instance [@KatokHasselblatt1995 Corollary 18.2.2]).
Consider now the map $$h_{\varepsilon}:TM\rightarrow TM\,,\, v\mapsto\varepsilon v.$$ Then the observation is that $$h_{\varepsilon}^{*}\omega_{\varepsilon}=\varepsilon\omega_{1},\ \ \ h_{\varepsilon}^{*}H=\varepsilon^{2}H.$$ Now $\phi_{t}^{\varepsilon}:SM\rightarrow SM$ is Anosov if and only if the flow (temporarily called) $\psi_{t}$ of $h_{\varepsilon}^{*}H$ with respect to $h_{\varepsilon}^{*}\omega_{\varepsilon}$ is Anosov on $(h_{\varepsilon}^{*}H)^{-1}(1/2)$. But $\psi_{t}$ is the Hamiltonian flow of $\left(x,v\right)\mapsto\frac{\varepsilon^{2}}{2}\left|v\right|^{2}$ with respect to the symplectic form $\varepsilon\omega_{1}$. But it is easy to see that $\psi_{t}=\phi_{\varepsilon t}^{1}$; hence we have shown that $$\phi_{t}^{\varepsilon}:SM\rightarrow SM\mbox{ is Anosov }\Leftrightarrow\phi_{t}^{1}:\Sigma_{1/2\varepsilon^{2}}\rightarrow\Sigma_{1/2\varepsilon^{2}}\mbox{ is Anosov.}$$ Thus in particular for $$k>\frac{1}{2}\delta_{0}^{-2},$$ $\left(\Sigma_{k},\omega_{1}\right)$ is an Anosov Hamiltonian structure.
In order to prove that $\left(\Sigma_{k},\omega_{1}\right)$ is $1/2$-pinched for $k$ large we note that an equivalent claim is that $(\Sigma_{1/2},\omega_{\varepsilon\sigma})$ is $1/2$-pinched for small $\varepsilon$. An inspection of the proof of (\[eq:a1\]) and (\[eq:a2\]) in [@Klingenberg1995 Theorem 3.2.17] (see Example \[exa:example sec curv\]) shows that if we do the same analysis for the magnetic Jacobi (or Riccati) equation we obtain numbers $k_{1}(\varepsilon)$ and $k_{0}(\varepsilon)$ for which (\[eq:a1\]) and (\[eq:a2\]) hold. These numbers will be as close as we wish to $k_{1}(0)=2$ and $k_{0}(0)=\sqrt{-\max\, K}>1$ if $\varepsilon$ is small enough and the $1/2$-pinching condition follows (see Example \[exa:example sec curv\] again).
From this it is easy to prove Corollary B. Suppose $k$ is chosen large enough such that $(\Sigma_{k},\omega_{1})$ is an Anosov Hamiltonian structure satisfying the $1/2$-pinching condition, and suppose for contradiction that $(\Sigma_{k},\omega_{1})$ is stable, and let $\lambda$ be a stabilizing $1$-form. By Theorem A, $(\Sigma_{k},\omega_{1})$ is contact, that is, $d\lambda=\rho\omega_{1}$ for some non-zero $\rho\in\mathbb{R}$. In particular, $\omega_{1}$ is exact which in turn implies that $\sigma$ is exact.
[^1]: For clarity we will use the letter $c$ to stand for $g_{0}$-geodesics and $\gamma$ for the projection to $\widetilde{M}$ of flow lines of $\widetilde{\phi}_{t}^{1}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper, we consider the local bases of primitive nonpowerful sign pattern matrices, show that there are “[*gaps*]{}" in the local base set and characterize some sign pattern matrices with given local bases.
[**AMS Classification:**]{} 05C50
[**Keywords:**]{} Gap; Primitive and nonpowerful; Signed digraph; Local base
author:
- |
Guanglong Yu$^{a, b}$[^1] Zhengke Miao$^b$ [^2]\
\
[$^a$Department of Mathematics, Yancheng Teachers University, Yancheng, 224002, China]{}\
[$^b$Department of Mathematics, Xuzhou Normal University, Xuzhou, 221116, China]{}
title: '[**On the local base set of primitive and nonpowerful signed digraphs[^3]**]{}'
---
Introduction
============
In this paper, we permit loops but no multiple arcs in a digraph. We denote by $V(S)$ the vertex set and denoted by $E(S)$ the arc set for a digraph $S$. A digraph is called a $signed$ digraph if its each edge is assigned one of the signs $-1$ and $1$. In a signed digraph, the sign of a directed walk $W=v_{0}e_{1}v_{1}e_{2}\cdots
e_{k}v_{k}\ (e_{i}=(v_{i-1}$, $v_{i})$, $1\le i\le k)$, denoted by $\mathrm{sgn}$$(W)$, is $\prod\limits_{ i =1}^k $sgn$(e_{i})$. The underlying graph of a signed digraph $S$, denote by $|S|$, is obtained by replacing the sign of each negative edge (with sign $-1$) with sign $1$.
\[de1.02\]A strongly connected digraph $S$ is primitive if there exists a positive integer $k$ such that for any two vertices $v_{i}, v_{j}$ (not necessarily distinct), there exists a directed walk of length $k$ from $v_{i}$ to $v_{j}$. The least such $k$ is called the primitive index of $S$, and is denoted by $\mathrm{exp}(S)$.
As a result, we know that, in a primitive digraph, there exist the least positive integer $k$ such that there is a directed walk of length $t$ from $v_{i}$ to $v_{j}$ for any integer $t\geq l$ is called the local primitive index from $v_{i}$ to $v_{j}$. The least $k$ is called the the local primitive index from $v_{i}$ to $v_{j}$, denoted by $\mathrm{exp}_{S}(v_{i}, v_{j})$. $\mathrm{exp}_{S}(v_{i})=\max\limits_{v_{j}\in
V(S)}\{\mathrm{exp}_{S}(v_{i}, v_{j})\}$ is called the local primitive index at $v_{i}$. Therefore, $\mathrm{exp}(S)=\max\limits_{v_{i}\in
V(S)}\{\mathrm{exp}_{S}(v_{i})\}$.
\[de1.4\]Assume that $W_{1}$, $W_{2}$ are two directed walks in signed digraph $S$. They are called a pair of $SSSD$ walks if they have the same initial vertex, the same terminal vertex and the same length, but they have different sign.
\[de1.6\]A signed digraph $S$ is primitive and nonpowerful if there exists a positive integer $l$ such that for any integer $t\geq l$, there are a pair of $SSSD$ walks of length $t$ from any vertex $v_{i}$ to any vertex $v_{j}(v_{i}, v_{j}\in V(S))$. The least such $l$ is called the base of $S$, denoted by $l(S)$.
As a result, in a primitive and nonpowerful signed digraph $S$, for $u, v \in V(S)$, there exists an integer $k$ such that their is a pair of $SSSD$ walks of length $t$ from $u$ to $v$ for any integer $ t\geq k$. The least such $k$ is called the local base from $u$ to $v$, denoted by $l_{S}(u, v)$. $l_{S}(u)=\max\limits_{v\in V(S)}\{l_{S}(u, v)\}$ is called the local base at vertex $u$. Therefore, $$l(S) = \max\limits_{u\in V(S)}l_{S} (u)= \max\limits_{u,v\in
V(S)}l_{S} (u, v).$$
The primitivity of a digraph have been studied extensively which is closely related to many other problems in various areas of pure and applied mathematics (for example, see [@BLiu]-[@55]). For a primitive and nonpowerful signed digraph, the base always seems being not equal to its primitive index, and studying the base needs more treatment (see [@Li], [@ShaoYou], [@YSZ]). Simultaneously, for a primitive and nonpowerful signed digraph, we find that the local base always seems different from its local primitive index (see [@W.M]). In [@YSZ], we find that studying the base or local base of a signed digraph is of great significance for communication science and for studying the properties of sign matrices.
In this paper, we consider the local bases of primitive nonpowerful sign pattern matrices. The paper is organized as follows: Section 1 introduces the basic ideas of patterns and their supports; Section 2 introduces series of working lemmas; Section 3 shows that there are some gaps in the local base set and characterizes some digraphs with given local bases.
Preliminaries
=============
We first introduce some notations. We denoted by $L(W)$ the length of a directed walk, and denote by $d(v_{i}, v_{j})$ or $d_{S}(v_{i}, v_{j})$ the distance from $v_{i}$ to $v_{j}$ in signed digraph $S$. We denote by $C_{k}$ or [*$k$-cycle*]{} a directed cycle with length $k$, and denote by $P_{k}$ a directed path of order $k$. A cycle with even (odd) length is called an [*even cycle*]{} ([*odd cycle*]{}). The length of the shortest directed cycle in a digraph is called the [*girth*]{} of this digraph. When there is no ambiguity, a directed walk, a directed path or a directed cycle will be called a walk, a path or a cycle. A walk is called a [*positive (negative) walk*]{} if its sign is positive (negative). The union of digraphs $H$ and $G$ is the digraph $G\bigcup H$ with vertex set $V(G)\bigcup V(H)$ and arc set $E(G)\bigcup E(H)$. The intersection $G\bigcap H$ of digraphs $H$ and $G$ is defined analogously. If $p$ is a positive integer and if $C$ is a cycle, then $pC$ denotes the walk obtained by traversing through $C$ $p$ times. If a cycle $C$ passes through the end vertex of $W$, $W\bigcup pC$ denotes the the walk obtained by going along $W$ and then going around the cycle $C$ $p$ times; $pC\bigcup W$ is similarly defined. We use the notation $v\stackrel{k}{\longrightarrow}u$ ($v\stackrel{k}{\not\longrightarrow}u$) to denote that there exists a (exists no) directed walk with length $k$ from vertex $v$ to $u$. For a digraph $S$, let $R_{k}(v)=\{u|$ $v\stackrel{k}{\longrightarrow}u$, $u\in V(S)\}$. For a vertex subset $T$ in a digraph $S$, let $T\stackrel{k}{\longrightarrow}u$ mean that there exists a $s\in T$ such that $s\stackrel{k}{\longrightarrow}u$.
For a strongly connected digraph $S$ with order $n$, let $C(S)$ denote the cycle length set.
\[de2.1\]Let $\{s_{1}$, $s_{2}$, $\cdots$, $s_{\lambda} \}$ be a set of distinct positive integers with gcd$(s_{1}$, $s_{2}$, $\cdots$, $s_{\lambda})$ = 1. The Frobenius number of $
s_{1}$, $s_{2}$, $\cdots$, $s_{\lambda} $, denoted by $\phi(s_{1},
s_{2}, \cdots, s_{\lambda})$, is the smallest positive integer $m$ such that for all positive integers $k\geq m$, there are nonnegative integers $a_{i} \ (i = 1, 2, \cdots, \lambda)$ such that $k=\sum
\limits_{i = 1}^ \lambda$ $ a_{i}s_{i}$.
It is well known that
[**([@LB])**]{} \[le2.2\]If gcd$(s_{1}, s_{2}) = 1$, then $\phi(s_{1}, s_{2}) = (s_{1} -
1)(s_{2} - 1)$.
From Definition \[de2.1\], it is easy to see that $\phi(s_{1},
s_{2},\cdots,s_{\lambda})\leq \phi(s_{i}, s_{j})$ if there exist $s_{i}, s_{j}\in\{ s_{1}$, $s_{2}$, $\cdots$, $s_{\lambda} \}$ such that gcd$(s_{i}, s_{j}) = 1$. So if $\min \{s_{i}:1\leq i\leq
\lambda\} = 1$, then $\phi(s_{1}, s_{2},\cdots,s_{\lambda})=0$.
[**([@Kim])**]{} \[le2.5\]A digraph $S$ with $C(S)=\{p_{1}, p_{2}, \cdots, p_{t}\}$ is primitive if and only if $S$ is strongly connected $\gcd(p_{1}, p_{2}, \cdots, p_{t})=1$.
For a primitive digraph $S$, suppose $C(S)=\{p_{1}$, $p_{2}$, $\ldots$, $p_{u}\}$. Let $d_{C(S)}(v_{i}$, $v_{j})$ denote the length of the shortest walk from $v_{i}$ to $v_{j}$ which meets at least one $p_{i}$-cycle for each $i$, $i=1$, $2, \cdots, u$. Such a shortest directed walk is called a $C(S)$-walk from $v_{i}$ to $v_{j}$. And further, $d_{C(S)}(v_{i})$, $d_{i}(C(S))$ and $d(C(S))$ are defined as follows: $d_{C(S)}(v_{i})=\max\{d_{C(S)}(v_{i}$, $v_{j})$: $v_{j}\in V(S)\}$, $d(C(S))=\max\{d_{C(S)}(v_{i}$, $v_{j})$: $v_{i}$, $v_{j}\in V(S)\}$, $d_{i}(C(S))$ $(1\leq i \leq
n)$ is the $ith$ smallest one in $\{d_{C(S)}(v_{i})| 1\leq i \leq
n\}$, $d_{n}(C(S))=d(C(S))$. In particular, if $C(S)=\{p, q\}$, $d(C(S))$ can be simply denoted by $d\{p, q\}.$
[**([@BLiu])**]{} \[le2.6\]Let $S$ be a primitive digraph of order $n$ and $C(S)=\{p_{1}$, $p_{2}$, $\ldots$, $p_{u}\}$. Then $\mbox{exp}(v_{i},v_{j}) \leq d_{C(S)}(v_{i},v_{j})+\phi(p_{1},p_{2},\ldots,p_{u})$ for $v_{i}$, $v_{j}\in V(S)$. And furthermore, we have $\mathrm{exp}(S)\leq d(C(S))+\phi(p_{1},p_{2},\ldots,p_{u})$.
[**([@ShaoYou])**]{} \[le2.101\]Let $S$ be a primitive nonpowerful signed digraph. Then $S$ must contain a $p_{1}$-cycle $C_{1}$ and a $p_{2}$-cycle $C_{2}$ satisfying one of the following two conditions:
\(1) $p_{i}$ is odd, $p_{j}$ is even and sgn$C_{j}=-1$ ($i,j = 1,2$; $i\neq j)$.
\(2) $p_{1}$ and $p_{2}$ are both odd and sgn$C_{1}=-$sgn$C_{2}$.
$C_{1}$, $C_{2}$ satisfying condition (1) or (2) are always called [*a distinguished cycle pair*]{}. It is easy to prove that $W_{1}=p_{2}C_{1}$ and $W_{2}=p_{1}C_{2}$ have the same length $p_1p_2$ but different sign if $p_{1}$-cycle $C_{1}$ and $p_{2}$-cycle $C_{2}$ are a distinguished cycle pair, namely $(\mbox{sgn}C_{1})^{p_{2}} =
-((\mbox{sgn}C_{2})^{p_{1}}).$
[**([@W.M])**]{}\[le2.12\]Let $S$ be a primitive nonpowerful signed digraph of order $n$ and $u \in
V(S)$. If there exists a pair of $SSSD$ walks with length $r$ from $u$ to $u$, then $l_{S}(u) \leq \mbox{exp}_{S}(u) + r.$
[**([@W.M])**]{}\[le2.13\]Let $S$ be a primitive nonpowerful signed digraph of order $n$, then we have $l_{S}(k) \leq l_{S}(k-1) +
1$ for $ 2\leq k \leq n$.
Let $D_1$ consists of cycle $(v_{n}$, $v_{n-1}$, $\cdots$ , $v_{2}$, $v_{1}$, $v_{n})$ and arc $(v_{1}$, $v_{n-1})$ and $D_{2}$ = $D_{1}\bigcup \{(v_{2}$, $v_{n})\}$. Then we have the next lemma.
[**([@W.M])**]{}\[le2.14.01\]Let $S$ be a primitive nonpowerful signed digraph of order $n$ with $D_{1}$ as its underlying digraph. Then we have $l_{S}(k) = 2n^{2}
-4n + k + 2$ for $1\leq k \leq n$.
[**([@W.M])**]{}\[le2.14.02\]Let $S$ be a primitive nonpowerful signed digraph of order $n \geq
3$ with $D_{2}$ as its underlying digraph. Then we have:
\(1) If the (only) two cycles of length $n-1$ of $S$ have different signs, then $$l_{S}(k)\leq\left \{\begin{array}{ll}
2n^{2}
-2n + k + 1,\ & \ 1\leq k\leq n-1;
\\ n^{2}-n,\ & \ k= n. \end{array}\right.$$
\(2) If the (only) two cycles of length $n-1$ of $S$ have the same sign, then $l_{S}(k) = 2n^{2}-4n + k + 1$ for $1\leq k \leq n$.
[**([@W.M])**]{}\[le2.14\]Let $S$ be a primitive nonpowerful signed digraph with order $n\geq6$ whose underlying digraph is neither isomorphic to $D_1$ nor to $D_2$, then $l_{S}(k)\leq 2n^{2}- 6n +k +
4$ for $1\leq k \leq n$.
[**([@YMS])**]{}\[le2.14.0\] (i) Let $A$ be a primitive nonpowerful square sign pattern with order $n\geq6$. If $C(S(A))=\{p, q\} \
(p<q\leq n, p+q>n)$ and the cycles with the same length have the same sign in $S(A)$, then $p(2q-1)
\leq l(A) \leq 2p(q-1)+n.$
\(i) Let $n\geq 6$, and let $p, q$ be integers satisfying $p< q\leq
n, p+q\geq n$ and $\gcd(p, q)=1$. Then there exists a primitive nonpowerful square sign pattern matrix $A$ with order $n$ such that $C(S(A))=\{p, q\}$ and $l(A)=k$ for each $k\in [p(2q-1) ,
2p(q-1)+n]$, namely, $$[p(2q-1) , 2p(q-1)+n] \subseteq E_{n}^{l}$$ where $E_{n}^{l}=\{l(A)|A$ is a $n\times n$ primitive nonpowerful sign pattern matrix $\}$.
[**([@YZ])**]{}\[th3.5\] Let $S$ be a primitive nonpowerful signed digraph of order $n\geq6$. If there exists some $k\ (1\leq k\leq n)$ such that $l_{S}(k)\geq \displaystyle
\frac{3}{2}n^{2}-3n+k+4$, then we have the results as follows:
(i)$|C(S)|=2$. Suppose $C(S)=\{p_{1},p_{2}\}(p_{1}<p_{2})$, then $\gcd(p_{1},p_{2})=1, p_{1}+p_{2}>n;$
\(ii) In $S$, all $p_{1}-$cycles have the same sign, all $p_{2}-$cycles have the same sign, and every pair of $p_{1}-$cycle and $p_{2}-$cycle form a distinguished cycle pair.
[**([@YZ])**]{}\[th4.3\] Let $S$ be a primitive nonpowerful signed digraph of order $n\geq6$. If there exists some $k\ (1\leq k\leq n)$ such that $l_{S}(k)\geq \displaystyle
\frac{3}{2}n^{2}-3n+k+4$, then $$l_{S}(k)\leq \left \{\begin{array}{ll}
(2n-1)p_{1},\ &
p_{2}=n, 1\leq k\leq p_{1};\\
\\ (2n-2)p_{1}+k,\ &
p_{2}=n, p_{1}+1\leq k\leq n;\\
\\ n+2p_{1}(p_{2}-1),\ & p_{2}\leq n-1, 1\leq k\leq n.\end{array}\right.$$ where $C(S)=\{p_{1},p_{2}\}, p_{1}<p_{2}$.
\[le2.16\]Let $D_{k, i}$ consists of cycle $C_{n}=(v_{1}$, $v_{n}$, $v_{n-1}$, $v_{n-2}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n-k})$, $(v_{2}$, $v_{n-k+1})$, $\ldots$, $(v_{i}$, $v_{n-k+i-1})\ (1\leq i\leq \min\{k+1, n-k-1\})$ (see Fig. 3.1) where $\gcd(n, n-k)=1$. Then $\mbox{exp}_{D_{k,
i}}(m)=\mbox{exp}_{D_{k, i}}(v_{m})=(n-2)(n-k)+1-i+m$ for $1\leq
m\leq n$.
(94.33,37.33) (41.33,20.00)(39.33,6.67)(65.00,7.33) (41.33,20.00)(39.33,36.67)(65.00,35.00) (65.33,35.00)(93.33,36.33)(91.33,20.33) (91.33,20.33)(92.33,5.67)(65.33,7.33) (61.00,7.33)[(5,3)[30.00]{}]{} (84.33,9.33)[(-2,3)[17.11]{}]{} (49.33,8.67) (91.00,15.33) (61.33,7.33) (84.33,9.33) (91.33,25.33) (67.67,35.00) (41.67,15.33) (41.33,25.67) (48.00,9.33)[(-2,1)[3.00]{}]{} (41.33,15.67)[(0,1)[5.67]{}]{} (42.33,28.00)[(3,4)[1.75]{}]{} (52.67,34.67)[(1,0)[5.33]{}]{} (68.00,35.00)[(1,0)[6.33]{}]{} (91.33,25.33)[(0,-1)[6.67]{}]{} (86.00,32.00)[(4,-3)[3.00]{}]{} (90.33,14.33)[(-2,-3)[1.78]{}]{} (74.33,7.33)[(-1,0)[4.33]{}]{} (60.67,7.33)[(-1,0)[4.67]{}]{} (49.33,8.67)[(6,1)[41.67]{}]{} (49.33,9.00)[(1,0)[3.00]{}]{} (81.33,19.33)[(3,2)[4.67]{}]{} (77.67,19.67)[(-2,3)[2.00]{}]{} (47.67,6.33)[(0,0)\[cc\][$v_{1}$]{}]{} (39.00,15.00)[(0,0)\[cc\][$v_{n}$]{}]{} (37.00,25.67)[(0,0)\[cc\][$v_{n-1}$]{}]{} (61.33,5.00)[(0,0)\[cc\][$v_{2}$]{}]{} (85.33,7.00)[(0,0)\[cc\][$v_{i}$]{}]{} (96.00,15.00)[(0,0)\[cc\][$v_{n-k}$]{}]{} (98.33,26.00)[(0,0)\[cc\][$v_{n-k+1}$]{}]{} (67.33,37.33)[(0,0)\[cc\][$v_{n-k+i-1}$]{}]{} (65.67,0.00)[(0,0)\[cc\][Fig. 3.1. $D_{k,i}$]{}]{}
It is easy to see $D_{k, i}$ is primitive by lemma \[le2.5\]. Also, it is not difficult to check that $$R_{n-k-(i-2)}(v_{1})\supseteq\left \{\begin{array}{ll}
\{v_{n}, v_{n-k}, v_{k}\},\ & i=1;\\
\\ \{v_{i-1}, v_{n-k+(i-1)}, v_{k+i-1}\},\ & 2\leq i\leq \min\{k+1,n-k-1\}.
\end{array}\right.$$
If $|\bigcup_{t=1}^{j} R_{t(n-k)-(i-2)}(v_{1})|<n$, we assert that $$|\bigcup_{t=1}^{j} R_{t(n-k)-(i-2)}(v_{1})\setminus
\bigcup_{t=1}^{j-1} R_{t(n-k)-(i-2)}(v_{1})|\geq 1.$$ Otherwise, $|\bigcup_{t=1}^{+\infty} R_{t(n-k)-(i-2)}(v_{1})|<n$, which contradicts that $D_{k, i}$ is primitive.
By the assertion above, we get $|\bigcup_{t=1}^{n-2}
R_{t(n-k)-(i-2)}(v_{1})|=n$ because of $|R_{1}(v_{1})|\geq 3$. So $\mbox{exp}_{D_{k, i}}(v_{1})\leq (n-2)(n-k)+2-i.$
If $i-1<k$, then $n-k+i-1<n$ and $d(C(D_{k, i}))=d_{C(D_{k,
i})}(v_{n},v_{n-k+i})=n+k-i.$ By Lemma \[le2.6\], we get $$\mbox{exp}(D_{k, i})\leq
d(C(D_{k, i}))+\phi(n,n-k)$$$$=d_{C(D_{k,
i})}(v_{n},v_{n-k+i})+(n-1)(n-k-1)=n+k-i+(n-1)(n-k-1).$$
Now we prove that there is no directed walk of length $n+k-i+\phi(n,n-k)-1$ from $v_{n}$ to $v_{n-k+i}$. Otherwise, suppose $W$ is a directed walk of length $n+k-i+\phi(n,n-k)-1$ from $v_{n}$ to $v_{n-k+i}$. Let $P_{1}$ denote the path from $v_{n}$ to $v_{n-k+i}$ on cycle $C_{n}$, then $$\mid P_{1}\mid = d(v_{n},
v_{n-k+i})=k-i$$ and $P_{1}$ meet only $n$-cycle not any $(n-k)$-cycle. $W$ must contain $P_{1}\bigcup C_{n}$, some $(n-k)-$cycles and some $n$-cycles, namely $$n+k-i+\phi(n,n-k)-1=k-i+n+a_{1}n+a_{2}(n-k)\ \ (a_{j}\geq
0,\ j=1,2)$$ and $$\phi(n,n-k)-1=a_{1}n+a_{2}(n-k)\ \ (a_{j}\geq 0,\
j=1,2)$$ which contradicts the definition of $\phi(n,n-k)$. So there is no directed walk of length $n+k-i+\phi(n,n-k)-1$ from $v_{n}$ to $v_{n-k+i}$, and further, we have $$\mbox{exp}(D_{k, i})=\mbox{exp}_{D_{k, i}}(v_{n})=\mbox{exp}_{D_{k, i}}(v_{n},v_{n-k+i})=n+k-i+(n-1)(n-k-1)
.$$ Notice that $$\mbox{exp}_{D_{k, i}}(v_{m})\leq \mbox{exp}_{D_{k, i}}(v_{1})+m-1, 1\leq
m\leq n,$$ $$n+k-i+(n-1)(n-k-1)-((n-2)(n-k)+2-i)=n-1,$$ thus $$\mbox{exp}_{D_{k, i}}(v_{1})\geq
\mbox{exp}_{D_{k, i}}(v_{n})-(n-1)=(n-2)(n-k)+2-i,$$ so we have $$\mbox{exp}_{D_{k, i}}(v_{1})=(n-2)(n-k)+2-i$$ and $$\mbox{exp}_{D_{k, i}}(m)=\mbox{exp}_{D_{k, i}}(v_{m})=(n-2)(n-k)+1-i+m\ (1\leq m\leq n).$$
If $k=i-1$, then $n-k+(i-1)=n$, we have $d(C(D_{k, i}))=d_{C(D_{k,
i})}(v_{n},v_{1})=n-1.$ Analogous to the proof of the case $n-k+(i-1)<n$, we can prove $$\mbox{exp}(D_{k, i})=\mbox{exp}_{D_{k, i}}(v_{n})=\mbox{exp}_{D_{k, i}}(v_{n},v_{1})=
d_{C(D_{k, i})}(v_{n},v_{1})+\phi(n,n-k)=(n-1)(n-k),$$ $$\mbox{exp}_{D_{k, i}}(v_{1})=(n-2)(n-k)+2-i,$$ and $\mbox{exp}_{D_{k, i}}(m)=
\mbox{exp}_{D_{k, i}}(v_{m})=(n-2)(n-k)+1-i+m$ for $1\leq m\leq n.\
\ \ \ \Box$
If $n$ is odd, let $\mathscr{L}$ consist of cycle $C_{n}=(v_{1}$, $v_{n}$, $v_{n-1}$, $v_{n-2}$, $v_{n-3}$, $\ldots$, $v_{2}$, $v_{1})\ (n\geq 6)$ and arcs $(v_{1}$, $v_{n-2})$, $(v_{3}$, $v_{n})$. For any positive integer $n$, let $F$ consist of cycle $C_{n-1}=(v_{1}$, $v_{n}$ , $v_{n-1}$, $v_{n-3}$, $v_{n-4}$, $\ldots$, $v_{2}$, $v_{1})\ (n\geq 6)$ and arcs $(v_{1}$, $v_{n-2})$, $(v_{n-2}$, $v_{n-3})$; let $F_{1}$ consist of cycle $(v_{1}$, $v_{n-1}$, $v_{n-2}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n-2})$, $(v_{2}$, $v_{n})$, $(v_{n}$, $v_{n-1})$; let $F_{2}$ consist of cycle $(v_{1}$, $v_{n}$, $v_{n-2}$, $v_{n-3}$, $v_{n-4}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n-2})$, $(v_{n}$, $v_{n-1})$, $(v_{n-1}$, $v_{n-3})$; let $F_{3}$ consist of cycle $(v_{1} $, $v_{n-2}$, $v_{n-3}$, $v_{n-4}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n-1})$, $(v_{n-1}$, $v_{n-2})$, $(v_{1}$, $v_{n})$, $(v_{n}$, $v_{n-2})$; let $F_{i}^{'}$ $(2\leq i\leq n-3)$ consist of cycle $(v_{1}$, $v_{n-1}$, $v_{n-2}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n-2})$, $(v_{i+1}$, $v_{n})$, $(v_{n}$, $v_{i-1})$; let $F_{4}$ consist of cycle $(v_{1}$, $v_{n-1}$, $v_{n-2}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n-2})$, $(v_{1}$, $v_{n})$, $(v_{n}$, $v_{n-3})$; let $F_{5}$ consist of cycle $(v_{1}$, $v_{n-1}$, $v_{n-2}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n-2})$, $(v_{2}$, $v_{n})$, $(v_{n}$, $v_{n-2})$; let $F_{6}$ consist of cycle $(v_{1}$, $v_{n-1}$, $v_{n-2}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n})$, $(v_{n}$, $v_{n-3})$, $(v_{2}$, $v_{n-1})$; let $F_{7}$ consist of cycle $(v_{1}$, $v_{n-1}$, $v_{n-2}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n-2})$, $(v_{3}$, $v_{n})$, $(v_{n}$, $v_{n-1})$; let $\mathscr{B}_{1}$ consist of cycle $C_{n}=(v_{1}$, $v_{n}$, $v_{n-1}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n-3})$, $(v_{3}$, $v_{n-1})$; let $\mathscr{B}_{2}$ consist of cycle $C_{n}=(v_{1}$, $v_{n}$, $v_{n-1}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n-3})$, $(v_{4}$, $v_{n})$; let $\mathscr{B}_{3}$ consist of cycle $C_{n}=(v_{1}$, $v_{n}$, $v_{n-1}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n-3})$, $(v_{2}$, $v_{n-2})$, $(v_{4}$, $v_{n})$; let $\mathscr{B}_{3}$ consist of cycle $C_{n}=(v_{1}$, $v_{n}$, $v_{n-1}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n-3})$, $(v_{2}$, $v_{n-2})$, $(v_{4}$, $v_{n})$; let $\mathscr{B}_{4}$ consist of cycle $C_{n}=(v_{1}$, $v_{n}$, $v_{n-1}$, $\ldots$, $v_{2}$, $v_{1})$ and arcs $(v_{1}$, $v_{n-3})$, $(v_{3}$, $v_{n-1})$, $(v_{4}$, $v_{n})$.
\[le2.16.1\](1) Suppose that $n$ is odd. Then $\mathrm{exp}_{\mathscr{L}}(k)=\mathrm{exp}_{\mathscr{L}}(v_{k})=(n-1)(n-3)+k-1$ for $1\leq k\leq n$.
\(2) $$\mbox{exp}_{F}(m)=\mbox{exp}_{F}(v_{m})=\left
\{\begin{array}{ll}
n^{2}-5n+7+m,\ & {\mbox {if}}\ 1\leq m\leq n-2;
\\ n^{2}-5n+6+m,\ & {\mbox {if}}\ n-1\leq m\leq n. \end{array}\right.$$
\(3) $\mbox{exp}_{F_{1}}(k)=\mbox{exp}_{F_{1}}(v_{k})=n^{2}-5n+6+k$ for $1\leq k\leq n$.
\(4) $$\mbox{exp}_{F_{2}}(k)=\mbox{exp}_{F_{2}}(v_{k})=\left
\{\begin{array}{ll}
n^{2}-5n+7+k,\ & {\mbox {if}}\ 1\leq k\leq n-2;
\\ n^{2}-5n+6+k,\ & {\mbox {if}}\ n-1\leq k\leq n. \end{array}\right.$$
\(5) $$\mbox{exp}_{F_{3}}(k)=\mbox{exp}_{F_{3}}(v_{k})=\left
\{\begin{array}{ll}
n^{2}-5n+6+k,\ & {\mbox {if}}\ 1\leq k\leq n-1;
\\ n^{2}-5n+5+k,\ & {\mbox {if}}\ k=n. \end{array}\right.$$
\(6) $$\mbox{exp}_{F_{i}^{'}}(k)=\left \{\begin{array}{ll}
\mbox{exp}_{F_{i}^{'}}(v_{k})=n^{2}-5n+6+k,\ & {\mbox {if}}\ 1\leq k\leq i;
\\ \mbox{exp}_{F_{i}^{'}}(v_{n})=n^{2}-5n+5+k,\ & {\mbox {if}}\ k=i+1;
\\ \mbox{exp}_{F_{i}^{'}}(v_{k-1})=n^{2}-5n+5+k,\ & {\mbox {if}}\ i+2\leq k\leq n. \end{array}\right.$$
\(7) $$\mbox{exp}_{F_{4}}(k)=\left \{\begin{array}{ll}
\mbox{exp}_{F_{4}}(v_{k})=n^{2}-5n+6+k,\ & {\mbox {if}}\ 1\leq k\leq n-2;
\\ \mbox{exp}_{F_{4}}(v_{n})=n^{2}-4n+4,\ & {\mbox {if}}\ k=n-1;
\\ \mbox{exp}_{F_{4}}(v_{n-1})=n^{2}-4n+5,\ & {\mbox {if}}\ k=n. \end{array}\right.$$
\(8) $$\mbox{exp}_{F_{5}}(k)=\left \{\begin{array}{ll}
\mbox{exp}_{F_{5}}(v_{k})=n^{2}-5n+6+k,\ & {\mbox {if}}\ 1\leq k\leq n-1;
\\ \mbox{exp}_{F_{5}}(v_{n-1})=n^{2}-4n+5,\ & {\mbox {if}}\ k=n. \end{array}\right.$$
\(9) $$\mbox{exp}_{F_{6}}(k)=\left \{\begin{array}{ll}
\mbox{exp}_{F_{6}}(v_{k})=n^{2}-5n+6+k,\ & {\mbox {if}}\ 1\leq k\leq n-2;
\\ \mbox{exp}_{F_{6}}(v_{n})=n^{2}-4n+4,\ & {\mbox {if}}\ k=n-1;
\\ \mbox{exp}_{F_{6}}(v_{n-1})=n^{2}-4n+5,\ & {\mbox {if}}\ k=n. \end{array}\right.$$
\(10) $\mbox{exp}_{F_{7}}(k)=\mbox{exp}_{F_{7}}(v_{k})= n^{2}-5n+5+k$ for $1\leq k\leq n$.
\(11) $\mbox{exp}_{\mathscr{B}_{1}}(k)=\mbox{exp}_{\mathscr{B}_{1}}(v_{k})=(n-1)(n-4)+k$ for $1\leq k\leq n$.
\(12) $\mbox{exp}_{\mathscr{B}_{2}}(k)=\mbox{exp}_{\mathscr{B}_{2}}(v_{k})=(n-3)^{2}+n+k-6$ for $1\leq k\leq n$.
\(13) $\mbox{exp}_{\mathscr{B}_{3}}(k)=\mbox{exp}_{\mathscr{B}_{3}}(v_{k})=(n-3)^{2}+n+k-6$ for $1\leq k\leq n$.
\(14) $\mbox{exp}_{\mathscr{B}_{4}}(k)=\mbox{exp}_{\mathscr{B}_{4}}(v_{k})=(n-3)^{2}+n+k-6$ for $1\leq k\leq n$.
\(1) It is not difficult to check that $R_{n-3}(v_{1})= \{v_{2},
v_{4}, v_{n}\}.$ Similar to the proof of Lemma \[le2.16\], we can prove $|\bigcup_{t=0}^{n-3} R_{t(n-2)+n-3}(v_{1})|=n,$ so $\mathrm{exp}(v_{1})\leq(n-1)(n-3)$ and $\mathrm{exp}(v_{n})\leq\mathrm{exp}(v_{1})+d(v_{n}, v_{1})\leq
(n-1)(n-2),$ and further, we get $\mathrm{exp}(v_{n})=\mathrm{exp}(v_{n}, v_{1})=(n-1)(n-2).$ So $\mathrm{exp}(v_{1})=(n-1)(n-3)$ and $$\mathrm{exp}(k)=\mathrm{exp}(v_{k})=(n-1)(n-3)+k-1\ (1\leq k\leq
n).$$
\(2) It is not difficult to check that $$R_{t(n-2)+2}(v_{1})=\left \{\begin{array}{ll}
\{v_{n-1}, v_{n-3}\},\ & t=0;\\
\\ \{v_{n}, v_{n-2}, v_{n-3}\},\ & t=1.
\end{array}\right.$$
Similar to the proof of Lemma \[le2.16\], if $|\bigcup_{t=0}^{j}
R_{t(n-2)+2}(v_{1})|<n$, we can prove $$|\bigcup_{t=0}^{j} R_{t(n-2)+2}(v_{1}) \setminus
\bigcup_{t=0}^{j-1} R_{t(n-2)+2}(v_{1})|\geq 1\ (j\geq3).$$ So we have $|\bigcup_{t=0}^{n-3} R_{t(n-2)+2}(v_{1})|=n$ because of $|R_{2}(v_{1})\bigcup R_{(n-2)+2}(v_{1})|= 4$ and $$\mbox{exp}_{F}(v_{1})\leq (n-3)(n-2)+2.$$ It is easy to check that $d(C(F))=d_{C(F)}(v_{n}, v_{n-1})=n.$ By Lemma \[le2.6\], thus we have $$\mbox{exp}_{F}(n)=\mbox{exp}_{F}(v_{n})\leq d(C(F))+\phi (n-1,
n-2)=n^{2}-4n+6.$$ Because of $d_{C(F)}(v_{n-2})=d_{C(F)}(v_{n-2},v_{n-1})=n-1,$ just as the proof of Lemma \[le2.16\], we get $$\mbox{exp}_{F}(n)=\mbox{exp}_{F}(v_{n})=\mbox{exp}_{F}(v_{n},
v_{n-1})=n^{2}-4n+6,$$ $$\mbox{exp}_{F}(v_{n-2})=d_{C(F)}(v_{n-2},v_{n-1})+\phi (n-1,
n-2)=(n-3)(n-2)+n-1,$$ and get $\mbox{exp}_{F}(v_{1})=(n-3)(n-2)+2,$ $$\mbox{exp}_{F}(m)=\mbox{exp}_{F}(v_{m})=\left \{\begin{array}{ll}
n^{2}-5n+7+m,\ & {\mbox {if}}\ 1\leq m\leq n-2;
\\ n^{2}-5n+6+m,\ & {\mbox {if}}\ n-1\leq m\leq n. \end{array}\right.$$
In a same way, we can prove (3)-(14). $\ \ \ \ \Box$
Gaps and characterizations of some digraphs with given local bases
==================================================================
\[th6.1\] Let $\gcd(n$, $n-k)=1$ and $S_{k, i}$ be a primitive nonpowerful signed digraph with underlying digraph $D_{k, i}\ (1\leq i\leq \min\{k+1, n-k-1\})$. If all $(n-k)$-cycles have the same sign, then $l_{S_{k,
i}}(m)=l_{S_{i}}(v_{m})=(2n-2)(n-k)+1-i+m\ (1\leq m\leq n).$
Every pair of $(n-k)$-cycle and $n$-cycle form a distinguished cycle pair because $S_{k, i}$ is a primitive nonpowerful signed digraph. By Lemmas \[le2.12\], \[le2.16\], we get $$l_{S_{k, i}}(v_{1})\leq
\mbox{exp}_{S_{k, i}}(v_{1})+n(n-k)=(2n-2)(n-k)+2-i.$$ Because of $d(v_{m},\ v_{1})=m-1$, we have $l_{S_{k, i}}(v_{m})\leq l_{S_{k,
i}}(v_{1})+m-1$ for $1\leq m\leq n$.
[**Case 1**]{} $i-1<k$, then $n-k+i-1<n$.
Now we prove that there is no pair of $SSSD$ walks of length $(2n-2)(n-k)-i+n$ from $v_{n}$ to $v_{n-k+i}$.
Otherwise, suppose $W_{1}, W_{2}$ are a pair of $SSSD$ walks with length $(2n-2)(n-k)-i+n$ from $v_{n}$ to $v_{n-k+i}$. Let $P$ be the unique path from $v_{n}$ to $v_{n-k+i}$ on cycle $C_{n}$. Then each $W_{j}\ (j=1, 2)$ must consists of $P\bigcup C_{n}$, some $n$-cycles and some $(n-k)$-cycles, namely $$|W_{j}|=(2n-2)(n-k)-i+n=n+k-i+a_{i}n+b_{i}(n-k)\ (a_{j}, b_{j}\geq 0,\ j=1, 2).$$ Because of $\gcd(n$, $n-k)=1$, so $(a_{1}-a_{2}
)n=(b_{2}-b_{1})(n-k),\ n|(b_{2}-b_{1}),\ (n-k)|(a_{1}-a_{2})$, and then $b_{2}-b_{1}=nx,\ a_{1}-a_{2}=(n-k)x$ for some integer $x$.
We assert $x=0$.
If $x\geq 1$, then $b_{2}\geq n$, thus we have $$(2n-2)(n-k)-i+n=n+k-i+a_{2}n+(b_{2}-n)(n-k)+n(n-k)$$ and $\phi(n, n-k)-1=a_{2}n+(b_{2}-n)(n-k),$ which contradicts the definition of $\phi(n, n-k)$. In a same way, we can get analogous contradiction when $x\leq -1$. Thus the assertion $x=0$ is proved.
So $W_{1}, W_{2}$ have the same sign because $b_{2}=b_{1},\
a_{1}=a_{2}$ and all $(n-k)-$cycles have the same sign. This contradicts $W_{1}, W_{2}$ are a pair of $SSSD$ walks. Thus there are no pair of $SSSD$ walks of length $(2n-2)(n-k)-i+n$ from $v_{n}$ to $v_{n-k+i}$, and so $$l(S_{k, i})= l_{S_{k, i}}(v_{n})=(2n-2)(n-k)+1+n-i.$$
Because of $(2n-2)(n-k)+1+n-i-((2n-2)(n-k)+2-i)=n-1,$ $$l_{S_{i}}(v_{m})\leq l_{S_{i}}(v_{1})+m-1(1\leq m\leq n),$$ we get $l_{S_{i}}(v_{1})\geq l_{S_{i}}(v_{n})-(n-1),$ so $l_{S_{i}}(v_{1})=(n-2)(2n-k)+2-i,$ and thus we have $l_{S_{i}}(m)=l_{S_{i}}(v_{m})=(2n-2)(n-k)+1-i+m$ for $1\leq m\leq
n$ by Lemma \[le2.13\].
[**Case 2**]{} $k=i-1$, then $n-k+i-1=n$.
As the proof of case 1, we can prove there is no pair of $SSSD$ walks of length $(2n-2)(n-k)-i+n$ from $v_{n}$ to $v_{1}$, and $l_{S_{k, i}}(v_{n})= (2n-2)(n-k)+1+n-i$, $$l_{S_{k, i}}(m)=l_{S_{k,
i}}(v_{m})=(2n-2)(n-k)+1-i+m\ (1\leq m\leq n).$$ $\ \ \ \Box$
If $n$ is odd, let $\mathscr{T}$ be a primitive nonpowerful signed digraph with underlying digraph $\mathscr{L}$, in which all $(n-2)$-cycles have the same sign. For any positive integer $n$, let $\mathscr{S}_{0}$ be a primitive nonpowerful signed digraph with underlying digraph $F$, in which all $(n-1)$-cycles have the same sign, all $(n-2)$-cycles have the same sign; let $\mathscr{S}_{1}$ be a primitive nonpowerful signed digraph with underlying digraph $F_{1}$, in which all $(n-1)$-cycles have the same sign, all $(n-2)$-cycles have the same sign; let $\mathscr{S}_{2}$ be a primitive nonpowerful signed digraph with underlying digraph $F_{2}$, in which all $(n-1)$-cycles have the same sign, all $(n-2)$-cycles have the same sign; let $\mathscr{S}_{3}$ be a primitive nonpowerful signed digraph with underlying digraph $F_{3}$, in which all $(n-1)$-cycles have the same sign, all $(n-2)$-cycles have the same sign; let $\mathscr{S}_{4}$ be a primitive nonpowerful signed digraph with underlying digraph $F_{4}$, in which all $(n-1)$-cycles have the same sign, all $(n-2)$-cycles have the same sign; let $\mathscr{S}_{5}$ be a primitive nonpowerful signed digraph with underlying digraph $F_{5}$, in which all $(n-1)$-cycles have the same sign, all $(n-2)$-cycles have the same sign; let $\mathscr{S}_{6}$ be a primitive nonpowerful signed digraph with underlying digraph $F_{6}$, in which all $(n-1)$-cycles have the same sign, all $(n-2)$-cycles have the same sign; let $\mathscr{S}_{7}$ be a primitive nonpowerful signed digraph with underlying digraph $F_{7}$, in which all $(n-1)$-cycles have the same sign, all $(n-2)$-cycles have the same sign; let $\mathscr{S}_{i}$ be a primitive nonpowerful signed digraph with underlying digraph $F^{'}_{i}$, in which all $(n-1)$-cycles have the same sign, all $(n-2)$-cycles have the same sign; let $\mathscr{Q}_{1}$ be a primitive nonpowerful signed digraph with underlying digraph $\mathscr{B}_{1}$, in which all $(n-3)$-cycles have the same sign; let $\mathscr{Q}_{2}$ be a primitive nonpowerful signed digraph with underlying digraph $\mathscr{B}_{2}$, in which all $(n-3)$-cycles have the same sign; let $\mathscr{Q}_{3}$ be a primitive nonpowerful signed digraph with underlying digraph $\mathscr{B}_{3}$, in which all $(n-3)$-cycles have the same sign; let $\mathscr{Q}_{4}$ be a primitive nonpowerful signed digraph with underlying digraph $\mathscr{B}_{4}$, in which all $(n-3)$-cycles have the same sign.
\[th6.1.1\] (1) $l_{\mathscr{T}}(k)=l_{\mathscr{T}}(v_{k})=2n(n-3)+k+2\ (1\leq
k\leq n).$
\(2) $$l_{\mathscr{S}_{0}}(k)=l_{\mathscr{S}_{0}}(v_{k})=\left \{\begin{array}{ll}
2n^{2}-8n+9+k,\ & \ 1\leq k\leq n-2;
\\ 2n^{2}-8n+8+k,\ & \ n-1\leq k \leq n. \end{array}\right.$$
\(3) $l_{\mathscr{S}_{1}}(k)=l_{\mathscr{S}_{1}}(v_{k})=2n^{2}-8n+8+k\
(1\leq k\leq n).$
\(4) $$l_{\mathscr{S}_{2}}(k)=l_{\mathscr{S}_{2}}(v_{k})=\left \{\begin{array}{ll}
2n^{2}-8n+9+k,\ & \ 1\leq k \leq n-2;
\\ 2n^{2}-8n+8+k,\ & \ n-1\leq k \leq n. \end{array}\right.$$
\(5) $$l_{\mathscr{S}_{3}}(k)=l_{\mathscr{S}_{3}}(v_{k})=\left \{\begin{array}{ll}
2n^{2}-8n+8+k,\ & \ 1\leq k \leq n-1;
\\ 2n^{2}-7n+7,\ & \ k= n. \end{array}\right.$$
\(6) $$l_{\mathscr{S}_{4}}(k)=\left \{\begin{array}{ll}
l_{\mathscr{S}_{4}}(v_{k})=
2n^{2}-8n+8+k,\ & \ 1\leq k\leq n-2;
\\ l_{\mathscr{S}_{4}}(v_{n})=2n^{2}-7n+6,
\ & \ k=n-1;
\\l_{\mathscr{S}_{4}}(v_{n-1})=2n^{2}-7n+7,
\ & \ k=n. \end{array}\right.$$
\(7) $$l_{\mathscr{S}_{5}}(k)=l_{\mathscr{S}_{3}}(v_{k})=\left \{\begin{array}{ll}
2n^{2}-8n+8+k,\ & \ 1\leq k \leq n-1;
\\ 2n^{2}-7n+7,\ & \ k= n. \end{array}\right.$$
\(8) $$l_{\mathscr{S}_{6}}(k)=\left \{\begin{array}{ll}
l_{\mathscr{S}_{6}}(v_{k})=
2n^{2}-8n+8+k,\ & \ 1\leq k\leq n-2;
\\ l_{\mathscr{S}_{6}}(v_{n})=2n^{2}-7n+6,
\ & \ k=n-1;
\\l_{\mathscr{S}_{6}}(v_{n-1})=2n^{2}-7n+7,
\ & \ k=n. \end{array}\right.$$
\(9) $l_{\mathscr{S}_{7}}(k)=l_{\mathscr{S}_{7}}(v_{k})=2n^{2}-8n+7+k\
(1\leq k\leq n).$
\(10) $$l_{\mathscr{S}_{i}}(k)=\left \{\begin{array}{ll}
l_{\mathscr{S}_{i}}(v_{k})=2n^{2}-8n+8+k,\ & \ 1\leq k\leq i;
\\ l_{\mathscr{S}_{i}}(v_{n})=2n^{2}-8n+8+k,
\ & \ k=i+1;
\\ l_{\mathscr{S}_{i}}(v_{k-1})=2n^{2}-8n+7+k,
\ & \ i+2\leq k\leq n. \end{array}\right.$$
\(11) $l_{\mathscr{Q}_{1}}(k)=l_{\mathscr{Q}_{1}}(v_{k})=2n^{2}-8n+4+k(1\leq
k\leq n).$
\(12) $l_{\mathscr{Q}_{2}}(k)=l_{\mathscr{Q}_{2}}(v_{k})=2n^{2}-8n+3+k\
(1\leq k\leq n).$
\(13) $l_{\mathscr{Q}_{3}}(k)=l_{\mathscr{Q}_{3}}(v_{k})=2n^{2}-8n+3+k\
(1\leq k\leq n).$
\(14) $l_{\mathscr{Q}_{4}}(k)=l_{\mathscr{Q}_{4}}(v_{k})=2n^{2}-8n+3+k\
(1\leq k\leq n).$
\(1) By Lemma \[le2.16.1\], we get $l_{\mathscr{T}}(v_{1})\leq(n-1)(n-3)+n(n-2)$ and $l_{\mathscr{T}}(v_{n})\leq l_{\mathscr{T}}(v_{1}) +d(v_{n},
v_{1})\leq(2n-1)(n-2).$ As the proof of Case 1 in Theorem \[th6.1\], we can prove $$l_{\mathscr{T}}(v_{n})=l_{\mathscr{T}}(v_{n}, v_{1})=(2n-1)(n-2),\
l_{\mathscr{T}}(v_{1})=(n-1)(n-3)+n(n-2)$$ and $l_{\mathscr{T}}(k)=l_{\mathscr{T}}(v_{k})=2n(n-3)+k+2\ (1\leq k\leq
n).$ In a same way, we can prove the Theorems (2)– (14) $\ \ \
\Box$
\[th6.11\] Let $S$ be a primitive nonpowerful signed digraph with order $n(n\geq 14)$. Then we have:
\(1) There is no $S$ such that $l_{s}(k)\in [2n^{2}-8n+10+k,
2n^{2}-4n +k]$ for $1\leq k\leq n-2$ and no $S$ such that $l_{S}(k)\in [2n^{2}-8n+9+k, 2n^{2}-4n +k]$ for $n-1\leq k\leq n$ if $n$ is an positive even integer.
\(2) If $n$ is an positive odd integer, there is no $S$ such that $l_{S}(k)\in [2n^{2}-6n+5+k, 2n^{2}-4n +k]$ for $1\leq k\leq n$;
there is no $S$ such that $l_{S}(k)\in [2n^{2}-8n+10+k, 2n^{2}-6n
+k+1]$ for $1\leq k\leq n-2$;
there is no $S$ such that $l_{S}(k)\in [2n^{2}-8n+9+k, 2n^{2}-6n
+k+1]$ for $n-1\leq k\leq n$; and further, we have:
\(i) $l_{S}(k)=2n^{2}-6n+4+k(1\leq k\leq n)$ if and only if $|S|\cong
D_{2, 1}$;
\(ii) $l_{S}(k)=2n^{2}-6n+3+k(1\leq k\leq n)$ if and only if $|S|\cong D_{2, 2}$, the cycles with the same length have the same sign in $S$;
\(iii) $l_{S}(k)=2n^{2}-6n+2+k(1\leq k\leq n)$ if and only if $|S|\cong D_{2, 3}$ or $|S|\cong \mathscr{L}$, the cycles with the same length have the same sign in $S$.
\(3) (i) $l_{S}(k)=2n^{2}-8n+9+k(1\leq k\leq n-2)$ if and only if $|S|\cong F$ or $|S|\cong F_{2}$, the cycles with the same length have the same sign in $S$;
\(ii) $l_{S}(k)=2n^{2}-8n+8+k(1\leq k\leq n)$ if and only if $|S|\cong F_{1}$, the cycles with the same length have the same sign in $S$;
$l_{S}(k)=2n^{2}-8n+8+k(1\leq k\leq n-2)$ if and only if $|S|$ is isomorphic to one of $\{F_{1}, F_{3}, F_{4}, F_{5}, F_{6},
F^{'}_{n-3}\}$, the cycles with the same length have the same sign in $S$;
$l_{S}(k)=2n^{2}-8n+8+k(1\leq k\leq n-1)$ if and only if $|S|$ is isomorphic to one of $\{F_{1}, F_{3}, F_{5}\}$, the cycles with the same length have the same sign in $S$;
$l_{S}(k)=2n^{2}-8n+8+k(n-1\leq k\leq n)$ if and only if $|S|$ is isomorphic to one of $\{F, F_{1}, F_{2}\}$, the cycles with the same length have the same sign in $S$.
\(iii) $l_{S}(k)=2n^{2}-8n+7+k(1\leq k\leq n)$ if and only if $|S|\cong F_{7}$, the cycles with the same length have the same sign in $S$; $l_{S}(k)=2n^{2}-8n+7+k(n-1\leq k\leq n)$ if and only if $|S|$ is isomorphic to one of $\{F_{4}, F_{6},
F_{7}\}\bigcup\{F^{'}_{i}| 2\leq i\leq n-3\}$, the cycles with the same length have the same sign in $S$; $l_{S}(k)=2n^{2}-8n+7+k(k=
n)$ if and only if $|S|$ is isomorphic to one of $\{F_{3}$, $F_{4}$, $F_{5}$, $F_{6}$, $F_{7}\}\bigcup\{F^{'}_{i}| 2\leq i\leq n-3\}$, the cycles with the same length have the same sign in $S$.
\(4) $2n^{2}-8n+6+k$ if and only if $|S|\cong D_{3, 1}$; $2n^{2}-8n+5+k$ if and only if $|S|\cong D_{3, 2}$, the cycles with the same length have the same sign in $S$; $2n^{2}-8n+4+k$ if and only if $|S|\cong D_{3, 3}$ or $|S|\cong \mathscr{B}_{1}$, the cycles with the same length have the same sign in $S$; $2n^{2}-8n+3+k$ if and only if $|S|$ is isomorphic to one of $\{D_{3, 4}, \mathscr{B}_{2}, \mathscr{B}_{3}, \mathscr{B}_{4}\}$, the cycles with the same length have the same sign in $S$.
\(5) For any positive integer $n$, there is no $S$ such that $l_{s}(k)\in [2n^{2}-9n+13, 2n^{2}-8n +2+k]$ for $1\leq k\leq n$.
Note that $n\geq 14$, then $2n^{2}-9n+12\geq \displaystyle
\frac{3}{2}n^{2}-3n+k+4.$ By Lemma \[th3.5\], then $C(S)=\{p_{1},p_{2}\}, p_{1}<p_{2}, p_{1}+p_{2}>n$, all the $p_{1}-$cycles have the same sign, all the $p_{2}-$cycles have the same sign in $S$. By Lemma \[th4.3\], we know that for $1\leq
k\leq n$, $$l_{S}(k)\leq \left \{\begin{array}{ll}
(2n-2)p_{1}+n\leq 2n^{2}-9n+8,\ & \ p_{2}=n, p_{1}\leq n-4;
\\ n+2p_{1}(p_{2}-1)\leq 2n^{2}-9n+12,\ & \ p_{1}\leq n-3, p_{2}\leq n-1.\end{array}\right.$$ So, if $l_{S}(k)\geq 2n^{2}-9n+13$, there are just the following cases:
\(1) $p_{2}=n, p_{1}= n-1$;
\(2) $p_{2}=n, p_{1}= n-2$;
\(3) $p_{2}=n, p_{1}= n-3$;
\(4) $p_{2}=n-1, p_{1}= n-2$.
Then the theorem follows from the Lemmas \[le2.14.01\]–\[le2.14\], Theorems \[th6.1\], \[th6.1.1\]. $\ \ \ \Box$
[**Acknowledgment**]{}
Many thanks to the referees for their kind reviews and helpful suggestions.
[^1]: E-mail addresses: [email protected] (Yu).
[^2]: Corresponding author: [email protected].
[^3]: Supported by NSFC (Nos. 11271315, 11171290, 11171728) and Jiangsu Qing Lan Project (2014A).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Despite the success of single-agent reinforcement learning, multi-agent reinforcement learning (MARL) remains challenging due to complex interactions between agents. Motivated by decentralized applications such as sensor networks, swarm robotics, and power grids, we study policy evaluation in MARL, where agents with jointly observed state-action pairs and private local rewards collaborate to learn the value of a given policy. In this paper, we propose a double averaging scheme, where each agent iteratively performs averaging over both space and time to incorporate neighboring gradient information and local reward information, respectively. We prove that the proposed algorithm converges to the optimal solution at a global geometric rate. In particular, such an algorithm is built upon a primal-dual reformulation of the mean squared projected Bellman error minimization problem, which gives rise to a decentralized convex-concave saddle-point problem. To the best of our knowledge, the proposed double averaging primal-dual optimization algorithm is the first to achieve fast finite-time convergence on decentralized convex-concave saddle-point problems.'
author:
- 'Hoi-To Wai, Zhuoran Yang, Zhaoran Wang, Mingyi Hong[^1]'
bibliography:
- 'intro.bib'
- 'to.bib'
- 'ref.bib'
title: 'Multi-Agent Reinforcement Learning via Double Averaging Primal-Dual Optimization'
---
[^1]: H.-T. Wai is with Chinese University of Hong Kong, Hong Kong. Z. Yang is with Princeton University, NJ, USA. Z. Wang is with Northwestern University, IL, USA. M. Hong is with University of Minnesota, MN, USA. E-mails: `[email protected]`, `[email protected]`, `[email protected]`, `[email protected]`
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Cold Fermionic atoms with three different hyperfine states confined in optical lattices show pronounced Atomic Density Waves (ADWs). These ADWs are pinned due to the confining potential that traps the atoms in the optical lattice and can be considered a crystal of strongly bound trions. We show that the crystalline phase is incompressible and robust against SU(3) symmetry breaking interaction. We also show that it is generic in the presence of the trap due to its incompressible characteristics. A simulation of the time evolution of the fermionic gas after the trapping potential is switch off shows that the formation of the trionic crystal is marked by a freezing of the dynamics in the lattice.'
author:
- 'Rafael A. Molina$^{1,2}$'
- Jorge Dukelsky$^1$
- 'Peter Schmitteckert$^{3}$'
title: 'Crystallization of trions in SU(3) cold atom gases trapped in optical lattices'
---
Ultracold Bose or Fermi gases confined in artificial optical lattices have become an important instrument for investigating the physics of strong correlations. The parameters and dimensionality of these systems can be tuned with very high precision and unprecedent control [@review] by means of an atomic Feshbach resonance or in an optical lattice by changing the depth of the wells. One of the main achievements in this subject is the experimental observation of a superfluid to Mott insulator transition in a three-dimensional optical lattice with bosonic $^{87}Rb$ atoms [@Greiner02]. Very interesting experimental results have also been obtained for fermions [@Stoferle06].
High spin fermions can be directly studied with cold atoms in more than two hyperfine states. This kind of systems could give rise to new exotic phases in optically trapped cold atoms. The general structure of high-spin Cooper pairs has already been theoretically addressed [@Ho99]. More recently, the emergence of triplets and quartets (three or four fermion bound states) has been explored [@Wu05; @Lecheminant05; @Kam05; @Rapp07; @Roux08]. In the case of SU(3), fermions with three hyperfine (color) states, these investigations have suggested that it would be possible to observe the transition from baryonic matter at low densities to a color superconducting state at higher densities [@Hofstetter04; @Rapp07]. At least two alkali atoms $^6Li$ and $^{40}K$ seem to be possible candidates for the experimental realization of a SU(3) fermionic lattice with attractive interactions[@Rapp07].
The paradigmatic model used for exploring these exotic features is the one-dimensional N-component fermionic Hubbard model with a local attractive interaction [@Hofstetter04]. A low-energy effective theory can be derived through linearization of the dispersion relation like in the usual SU(2) Hubbard Hamiltonian [@Giamarchibook; @Assaraf99]. This theory predicts new scattering channels in the half-filled case, [*i.e.*]{} the Umklapp scattering terms that do not conserve momentum but conserve quasi-momentum. In a recent work by Capponi [*et al.*]{} [@Capponi] the bosonization parameters of SU(N) theories have been investigated by means of large scale DMRG calculations. Regarding SU(3) their conclusion was that, for low densities, the dominant instability is that of a trionic superfluid and that color superconductivity is strongly suppressed. There is a competition between the superfluid trionic behavior and ADW order with the former phase dominating for generic fillings, and the latter probably appearing at half-filling.
The main goal of this letter is to study the emergence of a trionic crystalline phase in three color atoms loaded in optical lattices. We will show how the harmonic confinement potential can pin down an ADW at high densities in optical lattices, that is the precursor of crystal phase of trions (CPT). As the system gains energy due to local Umklapp scattering, the confined fermionic cloud deforms itself in the lattice to be at half-filling in the center of the trap. Similar effects have also been reported in the repulsive and attractive SU(2) cases [@Rigol; @Molina07], however for three color systems the appearance of an ADW is enhanced dramatically. The critical value of the attractive interaction $U$ for the CPT to be fully formed is approximately $U=-4$. Two main properties of this phase appear: a) it is an incompressible state, and b) the dynamics is almost frozen due to the suppression of the hopping of the trions between lattice wells. Fermionic color pairs are only dominant at weak coupling (a small region around $U=-1$).
The low energy physics of cold fermionic atoms with three different hyperfine states trapped in an optical lattice can be described by an SU(3) generalization of the Hubbard Hamiltonian, $$\begin{aligned}
H&=&-t\sum_{\left\langle ij\right\rangle \alpha} \left(f_{i\alpha}^{\dagger}f_{j\alpha }+{\mathrm{h.c.}}\right)+ \nonumber
\sum_{i\alpha\neq\beta} \frac{U_{\alpha\beta}}{2} n_{i\alpha}n_{i\beta} \\
&+&V\sum_{i\alpha}(i-L/2)^2~n_{i\alpha}.
\label{eq:Hamiltonian}\end{aligned}$$ The sums over $\alpha$ and $\beta$ go over the three colors The operators $f_{i\alpha}^{\dagger}$ and $f_{i\alpha }$ are the creation and destruction operators of an atom in site $i$ with color $\alpha$. We consider different values of the on-site interaction between the different color pairs $U_{\alpha\beta}$ to be able to include SU(3) symmetry breaking terms as they would probably be present in future experiments. The site label $i$ goes from $0$ to $L-1$, with $L$ being the total number of lattice sites. All energies are expressed in units of $t$ ($t=1$).
We study the ground state properties of the Hamiltonian (\[eq:Hamiltonian\]) with the DMRG algorithm [@DMRG] that provides very accurate numerical results. The modifications of the DMRG procedure due to the trap potential that breaks the translational symmetry have been described in previous publications [@schmitteckert; @schmitteckert_werner; @Molina07]. In order to obtain enough accuracy we have kept up to 1200 states in each iteration.
In Figure \[fig:l60q36d0\] we show the atom density for equal color populations $N=N_r+N_g+N_b=36$, parameters $V=0.003$ , $L=60$, and different values of the attractive SU(3) interaction $U=U_{rg}=U_{rb}=U_{gb}$. The figure clearly shows the formation of the CPT as $|U|$ increases. For $U \leq -4$ the phase is fully developed. Increasing $|U|$ compresses the density as interaction energy gains over kinetic energy and the particles prefer to be in the center of the trap. However, the crystal phase is incompressible and, once it is formed, increasing the value of the attraction does not compress the particle density any further.
![(Color online) Density for different values of $U$ for $L=60$, $N=36$, and $V=0.001333$.[]{data-label="fig:l60q36d0"}](figure1.eps){width="8cm"}
The formation of these phases can be studied by means of local correlation functions that measure the number of pairs and the number of trions. For the number of pairs we define $$P_{i \alpha \beta}=\left< n_{i \alpha} n_{i \beta} \right>-
\left< n_{ir} n_{ig} n_{ib}\right>.
\label{eq:pairs}$$ In the non-polarized we will be interested in the total number of pairs at each site i, $$P_i=P_{irg}+P_{irb}+P_{igb}.
\label{eq:total_pairs}$$ The local correlation that counts the number of trions is $$T_i=\left< n_{ir} n_{ig} n_{ib}\right>.$$ For non-interacting SU(3) fermions $P_{i\alpha\beta}=\left< n_{i\alpha}\right> \left< n_{i\beta}\right>-\left< n_{ir} \right> \left< n_{ig} \right>\left < n_{ib}\right>$ and $T_i=\left< n_{ir} \right> \left< n_{ig} \right>\left < n_{ib}\right>$.
These local correlations can be summed over all the sites on the system to obtain the total number of fermions in pairs (2P) and the total number of fermions in trions (3T). The number of uncorrelated fermions (F) is then $F=Q-2P-3T$.
The proper finite size scaling of the systems, as described in [@Damle], is given by $N^2V=constant$ when $N
\rightarrow \infty$ and $V \rightarrow 0$. We checked the validity this scaling law by comparing results for different sizes. For a given value of $U$ the results for the fraction of fermions in trions $3T/N$ and the fraction of fermions in pairs $2P/N$ as a function of the inverse size $1/N$ can be easily fitted by a second order polynomial. In this way we can extrapolate the values of trion and pair fractions in the thermodynamic limit. Figure \[fig:extrapolated\_correlations\] shows these results. It can be seen that the number of trions increases rapidly with the interaction strength until almost all fermions are present as trions for $U<-4$. The number of pairs is only dominant in a small region around $U=-1$ suggesting a color superconducting phase. An example of the quality of the extrapolation procedure is shown in the inset of the figure.
![(color online) Extrapolated total number of fermions in pairs (2P) and in trions(3T) compared with the single fermions(F) for different sizes as a function of $|U|$. Inset: Fitting second order polynomial for $2P/N$ as a function of $1/N$ for $U=-2$.[]{data-label="fig:extrapolated_correlations"}](figure2.eps){width="8cm"}
To understand better the origin of the ADW state we will compare the DMRG ground state energy with several approximations. In the strong interaction limit it is possible to treat the hopping term as a perturbation. Ignoring the effect of the trap, the zeroth order energy of each trion is $3U$. The first non-zero correction appears in second order perturbation theory giving a lowering of $3t^2/U$ for two next to nearest-neighbor trions, due to quantum fluctuations of the fermions within the trions. If two trions sit in neighboring sites this term is completely suppressed. Therefore, it acts as an effective repulsion between nearest-neighbor trions, being at the origin of the CPT at half-filling [@Zhao07]. Unlike the SU(2) case, the first contribution to hopping appears in the third order of perturbations, thus a CPT will always dominate for strong enough attraction.
Adding the effect of the trap we have calculated the energy of the CPT state, defined as $\prod_{i'}c^{\dagger}_{i',r} c^{\dagger}_{i',g} c^{\dagger}_{i',b} \left|0\right>$, where the $i'$ labels the even sites closer to the trap center until all the fermions are accounted for. We have also calculated the energy of a Fermi liquid (FL) state, defined as a Slater determinant in the basis that diagonalizes the hopping term plus the trapping potential. The CPT state valid in the strong coupling limit and the FL state appropriate for weak coupling limit are both Slater determinants independent of the attractive interaction $U$. Therefore, it is expected that a Hartree-Fock (HF) approximation that looks for the best Slater determinant for a given value of $U$ will interpolate between the two limits, and it will provide a better approximation for intermediate interactions. In Fig. \[fig:HF\] the ground state energy $E_0$ as a function of $|U|$ for the three different approximations is compared with the DMRG results. As expected, the HF energies are very accurate for small and large values of $|U|$. They deviate from the DMRG results appreciably in a region around $U=-2$ where the intermediate color superconductivity phase appears to take place.
![(Color online) Absolute value of the ground state energy $|E_0|$ in logarithmic scale as a function of $|U|$ for the case with $L=40$, $V=0.003$, and $N=24$. DMRG results are compared with a HF approximation, a CPT, and a Fermi liquid state of trapped trions.[]{data-label="fig:HF"}](figure3.eps){width="8cm"}
When the effect of the trap is not taken into account the ADW order is non-generic occurring only at exactly half-filling. According to our simulations the CPT state [*is*]{} generic in the presence of a trap and occurs for a wide range of values of the number of particles in the trap, the confining potential, and the interaction strength. To address this issue we have calculated the extension of the fermionic cloud in the trap, defined as $\sigma_n=\sqrt{\sum_i n_i(i-L/2)^2}$, as we vary the confining potential for different values of the interaction strength. Results of these calculations for $N=36$ are shown in Fig. \[fig:compres\]. The horizontal dotted line denotes the value of $\sigma_n$ for the CPT while the continuous horizontal line at the bottom denotes the Band Insulator (BI) state with all atoms filling the bottom of the trap. For $U \leq -4$, when the value of the CPT is reached, there is a plateau of $\sigma_n$ as a function of $N^2V$ marking the appearance of an incompressible state. For $U \leq -6$, new plateaus for lower $\sigma_n$ and higher values of $V$ appear. These other plateaus correspond to the presence of a band insulator in the center and symmetric CPTs at the sides of the trap. The crossover from the CPT to the BI as we increase the confinement consists on a number of steps in which each of the trions in the CPTs in the outer side of the cloud is forced to move the BI in the center (an analogous phenomenom has been recently described for spin 3/2 atoms [@Roux08]).
![(Color online) Normalized size of the fermionic cloud $\sigma_n/N$ as a function of $N^2V$ for different values of $U$ and $N=36$. []{data-label="fig:compres"}](figure4.eps){width="8cm"}
In what follows we will explore the evolution of the CPT when the trap is opened. We expect that its dynamics should be frozen since the effective hopping of a bound trion is extremely small (third order in perturbation theory). We have prepared a state of five trions made out of 15 fermions in a lattice of size $L=40$ and $v=0.005$. The harmonic confinement potential is switched off at $t=0$, and the state evolves with the rest of the Hamiltonian (\[eq:Hamiltonian\]). The temporal evolution is computed with the time dependent DMRG algorithm [@Petert]. In Fig. \[fig:time1\] we show the evolution in natural time units of the density for two selected values of $U$: $U=-2$ in the color superconducting phase and $U=-8$ well in the CPT. As a general trend, we observe that the expansion of the cloud almost ceases for $U\leq -6$. In order to quantify this statement we have calculated the velocity of the most external trion as a function of $U$. The velocity is computed by means of a linear fit of the position of the external peak as a function of time. The results are depicted in Fig. \[fig:vel\] where can see how the velocity is strongly suppressed for $U<-4$.
![(Color online) Evolution of the ground state density after a sudden switch off of the confinement potential for the case with $L=40$, $V=0.005$, and $Q=15$. Left panel: $U=-2$. Right panel: $U=-8$. []{data-label="fig:time1"}](figure5a.eps "fig:"){width="4cm" height="4cm"} ![(Color online) Evolution of the ground state density after a sudden switch off of the confinement potential for the case with $L=40$, $V=0.005$, and $Q=15$. Left panel: $U=-2$. Right panel: $U=-8$. []{data-label="fig:time1"}](figure5b.eps "fig:"){width="4cm" height="4cm"}
As already mentioned, a possible candidate for realizing the trionic phase is $^6Li$ for which the magnetic field dependence of the three scattering lengths has been measured [@exp6li]. The attractive $U$ interaction in some appropriate regions of magnetic field can be estimated as $U_{rg}=U_0$, $U_{rb}=1.23U_0$, and $U_{gb}=1.06U_0$ [@Rapp07]. Furthermore, the effective value of $U_0$ can be controlled changing the depth of the optical lattice wells. We have numerically checked the effects of this experimental anisotropy in the interaction. The CPT persists without noticeable changes even for stronger SU(3) symmetry breaking interactions.
![(Color online) Velocity of expansion as a function of the interaction in a case with $N=15$.[]{data-label="fig:vel"}](figure6.eps){width="8cm"}
In summary, we have shown that an CPT of trions is a very robust and generic phase of atoms with three hyperfine states confined in optical lattices. Due to the incompressible properties of this state it should be easily observed for experiments with high-density of fermionic atoms. The appearance of this state is marked by a freezing of the size of the cloud when the trapping potential is switched off, as we have shown through a full time simulations of the evolution. In future experiments of ultracold fermionic atoms the existence of the CPT could be revealed by light-scattering diffraction experiments. Moreover, radio-frequency spectroscopy could be used to probe the dissociation dynamics of the trions.
The authors acknowledge discussions with G. Zarand, P. Lecheminant, and R. Thomale. This work is supported in part by Spanish Government grant No. FIS2006-12783-C03-01 and by grant CAM-CSIC No. CCG07-CSIC/ESP-1962. RAM contract is financed by CSIC and the European Comission through the I3p program and by a Jose Castillejo Grant of the Spanish Ministry of Research.
[xx]{}
M. Lewenstein [*et al.*]{} Adv. Phys. [**56**]{}, 246 (2007).
M. Greiner [*et al.*]{} Nature [**415**]{}, 39 (2002).
T. Stöferle, H. Moritz, K. Gunter, M. Kohl, T. Esslinger, Phys. Rev. Lett. [**96**]{}, 030401 (2006); J.K. Chin [*et al.*]{} Nature [**443**]{}, 961 (2006).
T.-L. Ho, and S. Yip, Phys. Rev. Lett. [**82**]{}, 247 (1999).
C.J. Wu, Phys. Rev. Lett. [**95**]{}, 266404 (2005).
P. Lecheminant, E. Boulat, and P. Azaria, Phys. Rev. Lett. [**95**]{}, 240402 (2005).
H. Kamei, and K. Miyake, J. Phys. Soc. Jpn. [**74**]{}, 1911 (2005).
A. Rapp, G. Zaránd, C. Honerkamp, and W. Hofstetter, Phys. Rev. Lett. [**98**]{}, 160405 (2007).
G. Roux, S. Capponi, P. Lecheminant, and P. Azaria, cond-mat/0807.0412.
C. Honerkamp, and W. Hofstetter, Phys. Rev. Lett. [**92**]{}, 170403 (2004); Phys. Rev. B [**70**]{}, 094521 (2004).
T. Giamarchi, [*Quantum Physics in One Dimension*]{} (Clarendon Press, Oxford, UK, 2004).
R. Assaraf, P. Azaria, M. Caffarel, and P. Lecheminant, Phys. Rev. B [**60**]{}, 2299 (1999).
S. Capponi, G. Roux, P. Lecheminant, P. Azaria, E. Boulat, and S. R. White, Phys. Rev. A [**77**]{}, 013624 (2008).
M. Rigol, A. Muramatsu, G.G. Batrouni, and R.T. Scalettar, Phys. Rev. Lett. [**91**]{}, 130403 (2003).
R.A. Molina, J. Dukelsky, and P. Schmitteckert, Phys. Rev. Lett. [**99**]{}, 080404 (2007).
S. R. White, Phys. Rev. Lett. [**69**]{}, 2863 (1992).
P. Schmitteckert, Ph.D Thesis. Universität Augsburg (1996); P. Schmitteckert in Lecture Notes in Physics Vol. 528, ed. by I. Peschel, X. Wang, M. Kaulke and K. Hallberg, ISBN 3-540-66129-8 (1999). P. Schmitteckert, and R. Werner, Phys. Rev. B [**69**]{}, 195115 (2004).
J. Zhao, K. Ueda, and X. Wang, Jap. Phys. Soc. J. [**76**]{}, 114711 (2007).
P. Schmitteckert, Phys. Rev. B [**70**]{}, 121302(R) (2004).
K. Damle, T. Senthil, S. N. Majundar, and S. Sachdev, [**36**]{}, 7 (1996).
M. Bartenstein [*et al.*]{} Phys. Rev. Lett. [**94**]{}, 103201 (2005).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In [Source Location]{} ([SL]{}) problems the goal is to select a minimum cost source set $S$ such that the connectivity from $S$ to any node $v$ is at least the demand $d_v$ of $v$. In [Network Augmentation]{} ([NA]{}) problems we are given a graph $G=(V,E)$ and an edge-set $F$ on $V$, edge-costs on $F$ or node-costs on $V$, and connectivity requirements $\{r_{sv}:sv \in D\}$ on a set $D$ of “demand edges”. The goal is to compute a minimum cost edge-set $I \subseteq F$, such that in the graph $G+I$, the connectivity of each $sv \in D$ is at least $r_{sv}$. In [Rooted NA]{} $D$ is a star with center $s$, and in [$a$-Based NA]{} $F$ is a star with center $a$. We suggest a notion of $q$-connectivity, where every node $u$ has capacity $q_u \geq 1$ that represents the resistance of $u$ to failures. We observe that a large class of [SL]{} problems, including the variants that appear in the literature, is a particular case of $q$-connectivity [$s$-Based Rooted NA]{} problems. We use this to derive some approximation algorithms for [SL]{} from those for [NA]{}, as well as to derive some new results for [SL]{} problems. Some of our results are as follows.
- We observe that directed [SL]{} with unit costs is $\Omega(\log n)$-hard to approximate even with $0,1$ demands, while some versions of [SL]{} with uniform demands can be solved in polynomial time.
- We show that the $H(|D|)$-approximation algorithm of [@KN-aug] for edge costs [$a$-Based NA]{} has ratio $H(\alpha)$ for the $q$-connectivity version with node-costs, where $p_{\max}$ is the maximum number of parallel edges in $F$, $\alpha=\min\{r(D),|D| \cdot p_{\max}\}$, and $H(j)=\sum_{i=1}^j 1/i$ is the $j$th harmonic number. This gives a simple unifying $H(d(V))$-approximation algorithm for several [SL]{} variants considered separately by Sakashita, Makino, and Fujishige [@SMF]. Moreover, we will show that this is so whenever the connectivity function is submodular.
- For undirected [$a$-Based NA]{} we give an approximation algorithm with ratio $\min\{k,p_{\max} \ln k\} \cdot O(\ln k)$ for node-costs and $O(\ln^2 k)$ for edge-costs, where $k=\max_{v \in V}d_v$ is the maximum demand; the latter improves the ratio $\tilde{O}(k)$ of Fukunaga [@Fuk].
- We consider a variant of [SL]{} where we also have edge-costs and flow-cost bounds $\{b_v: v \in V\}$, and require that for every node $v$, the minimum cost of a flow of value $d_v$ from $S$ to $v$ is at most $b_v$. We show that the problem admits approximation ratio $H(d(V))+H(c(E))$.
author:
- Guy Kortsarz
- Zeev Nutov
title: |
A unified approach to source location\
via rooted network augmentation
---
Introduction
============
Problems considered and relations between them
----------------------------------------------
In this paper we suggest a unifying approach to handle [Source Location]{} problems via [Rooted Network Augmentation]{} problems. In [Source Location]{} problems the goal is to select a minimum cost source set $S$ such that the connectivity from $S$ to any other node $v$ is at least the demand $d_v$ of $v$. In [Network Augmentation]{} problems the goal is to augment a given graph by a minimum-cost edge-set such that the new graph satisfies prescribed connectivity requirements. Formally, the generic versions of these problems are as follows. Given a function $w=\{w_u:u \in U\}$ on a groundset $U$ and $U' \subseteq U$, let $w(U')=\sum_{u \in U'}w_u$. If $w$ is a cost function on $U$ and $I$ is an edge-set on $U$, then the cost (or the node-costs) $w(I)$ of $I$ is defined to be the cost of the set of the endnodes of $I$. Also, let $w_{\max}=\max_{u \in U} w_u$ denote the maximum value of $w$, and assume that all the input numbers are integers.
In [NA]{} problems, typical connectivity functions are as follows.
- [*Edge-connectivity $\lambda_G(s,v)$*]{} is the maximum number of pairwise edge disjoint $sv$-paths in $G$.
- [*Node-connectivity $\kappa_G(s,v)$*]{} is the maximum number of pairwise internally disjoint $sv$-paths in $G$.
- [*$Q$-connectivity $\lambda_G^Q(s,v)$*]{} for given $Q \subseteq V$, is the maximum number of $sv$ paths no two of which have an edge or an internal node in $Q$ in common. Note that $Q$-connectivity reduces to edge-connectivity if $Q=\emptyset$, and to node-connectivity if $Q=V$; namely, $\lambda_G^\emptyset(s,v)=\lambda_G(s,v)$ and $\lambda_G^V(s,v)=\kappa_G(s,v)$.
The corresponding versions of [NA]{} are edge-connectivity [NA]{} when $f_{sv}(I)=\lambda_{G+I}(s,v)$, node-connectivity [NA]{} when $f_{sv}(I)=\kappa_{G+I}(s,v)$, and $Q$-connectivity [NA]{} when $f_{sv}(I)=\lambda^Q_{G+I}(s,v)$.
Most papers that considered [SL]{} problems defined $(S,v)$-edge-connectivity $\lambda_G(S,v)$ as the maximum number of pairwise edge-disjoint $(S,v)$-paths. On the other hand, several definitions were used for $(S,v)$-node-connectivity $\kappa(S,v)$; in most of these definitions, $\kappa(S,v)=\infty$ if $v \in S$. Here we suggest a definition that captures previous definitions as particular cases.
Note that $\lambda_G^Q(s,v)$ is the maximum $sv$-flow value, where edges and nodes in $Q$ have unit capacities, while the capacity of the nodes in $V \setminus Q$ is $\infty$. Hence the $Q$-connectivity function $\lambda_G^Q(s,v)$ is the max-flow/min-cut value function in $G$ with node-capacities in $\{1,\infty\}$ and unit edge capacities. In Definition \[d:pq\] below, this is generalized as follows. Every node $u$ has capacity $q_u$ (the resistance of $u$ to failures), and an amount $p_u$ of flow-supply that $u$ can deliver to any other node (including itself) if $u$ is chosen to be in the source set $S$.
\[d:pq\] Let $G=(V,E)$ be a graph with node-capacities $\{q_u:u \in V\}$. For $S \subseteq V$ and $v \in V$ the $(S,v)$-$q$-connectivity $\lambda_G^q(S,v)$ is the maximum flow value from $S \setminus \{v\}$ to $v$ (every edge has capacity $1$). Equivalently, $\lambda_G^q(S,v)$ is the minimum capacity of a cut $C \subseteq E \cup V \setminus \{v\}$ such that $G \setminus C$ has no $(S,v)$-path. Given [*flow-supplies*]{} $\{p_u:u \in S\}$, the $(S,v)$-$(p,q)$-connectivity $\lambda_G^{p,q}(S,v)$ is the maximum $(s,v)$-flow value $\lambda_{G'}^q(s,v)$ in the network $G'$ obtained by adding to $G$ a new node $s$ and connecting it to every $u \in S$ with $p_u$ edges; hence if $p_u \geq q_u$ for every $u \in S$ then: $\lambda_G^{p,q}(S,v)=p_v+\lambda_G^q(S,v)$ if $v \in S$, and $\lambda_G^{p,q}(S,v)=\lambda_G^q(S,v)$ otherwise.
Note that $\lambda_G(S,v)=\lambda_G^{p,q}(S,v)$ for $p_u= \infty$ and $q_u=\infty$ for every $u \in V$, and that $\lambda_G^Q(s,v)=\lambda_G^q(s,v)$ for $q_u=1$ if $u \in Q \setminus \{s,v\}$ and $q_u=\infty$ otherwise. Now we mention some node-connectivity functions $\psi$ that appear in the literature, and show that they are particular case of the $(p,q)$-connectivity function $\lambda_G^{p,q}$ with $p_u,q_u \in \{1,\infty\}$ and $q_u \leq p_u$.
1. $\kappa(S,v)$ is the maximum number of $(S,v)$-paths no two of which have a common node in $V \setminus (S \cup v)$ if $v \notin S$, and $\kappa(S,v)=\infty$ otherwise; equivalently, $\kappa(S,v)$ is the minimum size $|C|$ of a cut $C \subseteq E \cup V \setminus (S \cup \{v\})$ such that $G \setminus C$ has no $(S,v)$-path.\
For directed graphs, [$\kappa$-SL]{} is equivalent to [$\lambda$-SL]{} by the following (approximation ratio preserving) standard reduction: replace every $v \in V$ by two nodes $v^{in}, v^{out}$ connected by an edge $v^{in}v^{out}$, and replace every edge $uv \in E$ by an edge $u^{out}v^{in}$; each node $v^{out}$ inherits the cost and the demand of $v$, while $v^{in}$ has cost $\infty$ and demand $0$. It is not hard to verify that $S$ is a feasible solution to the original instance with connectivity function $\kappa$ if, and only if, $S^{out}=\{v^{out}:v \in S\}$ is a feasible solution to the obtained instance with the edge-connectivity function $\lambda$. For undirected graphs, we do not see that $\kappa$ is a particular case of $\lambda_G^{p,q}$, but we are also not aware on any work on [*undirected*]{} [$\kappa$-SL]{} (except that it is shown in [@III] that the problem can be solved in polynomial time for $k=\max_{v \in V} d_v \leq 2$ and is NP-hard for $k \geq 3$).
2. $\hat{\kappa}(S,v)$ is the maximum number of $(S,v)$-paths no two of which have a common node in $V \setminus \{v\}$ if $v \notin S$, and $\hat{\kappa}(S,v)= \infty$ otherwise; equivalently, $\hat{\kappa}(S,v)$ is the minimum size $|C|$ of a cut $C \subseteq E \cup V \setminus (\{v\})$ such that $G \setminus C$ has no $(S,v)$-path.\
Then $\hat{\kappa}(S,v)=\lambda_G^{p,q}(S,v)$ for $p_u=\infty$ and $q_u=1$ for every $u \in V$.
3. $\kappa'(S,v)=\hat{\kappa}(S,v)$ if $v \notin S$ and $\kappa'(S,v)=1+ \hat{\kappa}(S \setminus \{v\},v)$ if $v \in S$; equivalently, $\kappa'(S,v)$ is the minimum size $|C|$ of a cut $C \subseteq E \cup V \setminus (\{v\})$ such that $G \setminus C$ has no $(S \setminus \{v\},v)$-path.\
Then $\kappa'(S,v)=\lambda_G^{p,q}(S,v)$ for $p_u=1$ and $q_u=1$ for every $u \in V$.
Given an instance of [SL]{} or of [NA]{}, let $k$ denote the maximum demand $d_{\max}=\max_{v \in V} d_v$ or the maximum requirement $r_{\max}=\max_{uv \in D}r_{uv}$. Note that in [SL]{} problems with $\psi(S,v)=\lambda_G^{p,q}(S,v)$, we may always assume that $p_u \leq k$, and it is also reasonable to assume that $1 \leq q_u \leq p_u$ for every $u \in V$ (as in the above versions). We consider a relation between such $(p,q)$-connectivity versions of [SL]{} and $q$-connectivity versions of [NA]{}, that formally are defined as follows.
Now we define several particular important cases of [NA]{}.
The [$a$-Based Connectivity Augmentation]{} problem was defined in [@KN-aug], where it was also shown to admit an $H(r(D))$-approximation algorithm. The study of this problem in [@KN-aug] is motivated by the following observation.
\[[@KN-aug]\] \[o:aug\] If directed [$a$-Based Connectivity Augmentation]{} admits approximation ratio $\rho$, then directed [Connectivity Augmentation]{} admits the following approximation ratios: $2\rho$ if $Q \neq V$, and $2k \rho$ if $Q=V$.
For an edge-set/graph $J$ let $\delta_J(X)$ denote the set of edges in $J$ from $X$ to $V \setminus X$. This paper is motivated by a recent paper of Fukunaga [@Fuk], that defined the connectivity function $\kappa'$, and observed that [$\kappa'$-SSL]{} is equivalent to the particular case of [$s$-Based Rooted SNA]{} with node-connectivity requirements and [*edge-costs*]{}, and with $\delta_G(s)=\emptyset$. Here, by the same reduction as in Definition \[d:pq\] (for a proof see Section \[s:hard\]), we further observe the following.
\[o:reduction\] For both directed and undirected graphs, [SSL]{} is equivalent to [$s$-Based Rooted SNA]{} with node-costs and with $\delta_G(s)=\emptyset$ and $c(s)=0$.
We also consider the following generalization of [SSL]{}. Given an instance of [SSL]{} and edge-costs $\{c_e:e \in E\}$ let $\mu_G(S,v)$ denote the minimum cost of an edge set $F \subseteq E$ such that $\lambda^{p,q}_{(V,F)}(S,v) \geq d_v$, where $\mu_G(S,v)=\infty$ if no such edge set $F$ exists (namely, if $\lambda^{p,q}_G(S,v) < d_v$).
Previous work
-------------
The previously best known approximation ratios and hardness of approximation results for [SL]{} problems with connectivity functions $\lambda,\kappa,\hat{\kappa},\kappa'$, are summarized in Table 1; see also a survey in [@NI].
\[tbl:ratios\]
[ $c$ & $d$]{}
------------------ --------------------------------- --------------------------------- --------------------------------- ---------------------------------
[*Undirected*]{} [*Directed*]{} [*Undirected*]{} [*Directed*]{}
GC & GD $\Theta(\ln d(V))$ [@BKP; @SMF] $\Theta(\ln d(V))$ [@BKP; @SMF] $\Theta(\ln d(V))$ [@BKP; @SMF] $\Theta(\ln d(V))$ [@BKP; @SMF]
GC & UD in P [@AIMF] $O(\ln d(V))$ [@BKP] $O(\ln d(V))$ [@BKP] $O(\ln d(V))$ [@BKP]
UC & GD in P [@AIMF] $O(\ln d(V))$ [@BKP] $O(\ln d(V))$ [@BKP] $O(\ln d(V))$ [@BKP]
UC & UD in P [@TSSA] in P [@IMAH] $O(\ln d(V))$ [@BKP] $O(\ln d(V))$ [@BKP]
GC & GD $\Theta(\ln d(V))$ [@SMF] $\Theta(\ln d(V))$ [@SMF] $O(\ln d(V))$ [@Fuk] $O(\ln d(V))$ [@Fuk]
$O(k \ln k)$ [@Fuk] $O(k \ln k)$ [@Fuk]
GC & UD in P [@NII] in P [@NII]
UC & GD $O(\ln d(V))$ [@SMF] $O(\ln d(V))$ [@SMF] $O(\ln d(V))$ [@Fuk] $O(\ln d(V))$ [@Fuk]
$O(k)$ [@Ishii]
UC & UD in P [@NII] in P [@NII]
: Previous approximation ratios and hardness of approximation thresholds for [SL]{} problems. GC and UC stand for general and uniform costs, GD and UD stand for general and uniform demands, respectively.
Some additional results are as follows. Ishii, Fujita, and Nagamochi [@IFN2; @IFN] showed that undirected [SSL]{} with $\hat{\kappa}$ can be solved in polynomial time for $k \leq 3$, but is NP-hard if there exists a vertex $v \in V$ with $d(v) \geq 4$. Barasz, Becker, and Frank [@BBF] gave a strongly polynomial time algorithm for edge-connectivity directed [SSL]{} with uniform demands. Several generalizations of source location problems can be found in [@SMNF; @IM; @MUMF; @III].
Our results
-----------
Our first result is an easy consequence from Observation \[o:reduction\], and it illustrates the usefulness of the relation between the two problems.
\[t:hard\] Directed [SSL]{} for $k=1$ and unit costs is $\Omega(\log n)$-hard to approximate. Directed/undirected [$\kappa'$-SSL]{} with uniform demands can be solved in polynomial time.
Observations \[o:aug\] and \[o:reduction\] motivate the study of the [$a$-Based SNA]{} problem. Interestingly, the algorithms of [@SMF] for [SL]{} with connectivity functions $\lambda$,$\kappa,\hat{\kappa}$, and the algorithm of [@KN-aug] for [$a$-Based Connectivity Augmentation]{} both use the same method of reducing the problem to a [Submodular Cover]{} problem. Now we see that this is not a coincidence, since by Observations \[o:aug\] and \[o:reduction\], both problems are particular cases of the [$a$-Based SNA]{} problem. Furthermore, we will show by a simple proof that such a reduction is possible whenever the connectivity function is submodular and non-decreasing. A set function $f$ on $U$ is submodular if $f(X)+f(Y) \geq f(X \cap Y)+f(X \cup Y)$ for all $X,Y \subseteq U$, and $f$ is non-decreasing if $f(X) \leq f(Y)$ for all $X \subseteq Y \subseteq U$. Our next result is obtained by observing that the algorithm of [@KN-aug] for [$a$-Based Connectivity Augmentation]{} extends to [$a$-Based NA]{} with arbitrary submodular non-decreasing connectivity function, for both edge-costs and node-costs, as follows. Recall that for an [$a$-Based SNA]{} instance, $p_{\max}$ denotes the maximum number of parallel edges in $F$, and that $p_{\max} \leq k$.
\[t:aug\] If for a directed [$a$-Based SNA]{} instance each function $f_{uv}(I)$ is submodular and non-decreasing, then the problem admits an approximation algorithm with ratio $H(\alpha)$, where $\alpha = \max\limits_{e \in F} \sum_{uv \in D} [f_{uv}(\{e\})-f_{uv}(\emptyset)]$ in the case of edge costs, and $\alpha = \max\limits_{z \in V} \sum_{uv \in D}[f_{uv}(\delta_F(z))-f_{uv}(\emptyset)]$ in the case of node costs.
In Section \[s:a\] we observe that for [$a$-Based SNA]{} instances, the set-function on $F$ defined by $f_{uv}(I)=\lambda^q_{G+I}(u,v)$ is submodular and non-decreasing, and that $\alpha \leq |D|$ in the case of edge-costs and $\alpha \leq \min\{r(D),|D| \cdot p_{\max}\}$ in the case of node-costs, where $p_{\max}$ is the maximum number of parallel edges in $F$.
\[t:a\] Directed [$a$-Based SNA]{} admits approximation ratios $H(|D|)$ for edge-costs and $H(\min\{r(D),|D| \cdot p_{\max}\})$ for node-costs. Thus directed [SSL]{} admits approximation ratio $H(\min\{r(D),|D| \cdot p_{\max}\})$.
For both [SSL]{} and [SNA]{}, an approximation ratio $\rho$ for directed graphs implies approximation ratio $2\rho$ for undirected graphs. Usually, undirected connectivity problems are easier to approximate than the directed ones. Directed [SSL]{} is already [Set-Cover]{} hard even for $k=1$ and unit costs, but for undirected [SSL]{} with $k=1$ (and even with $k=2$) it is not hard to obtain a polynomial time algorithm. Hence a natural question is whether undirected [SSL]{} admits an approximation ratio that depends on $k$ only. As was mentioned, Fukunaga [@Fuk] observed that $Q$-connectivity [SSL]{} with $Q=V$ is equivalent to [$s$-Based Rooted SNA]{} with edge-costs. Since undirected $Q$-connectivity [Rooted SNA]{} with edge-costs admits ratio $O(k\ln k)$ [@N-focs], so is $Q$-connectivity [SSL]{} with $Q=V$. However, [Connectivity Augmentation]{} admits ratio $O(\ln^2k)$ for rooted requirements [@N-aug], and a natural question to ask is whether the algorithm of [@N-aug] extends to [$s$-Based Rooted SNA]{}.
\[t:main’\] Undirected [$a$-Based SNA]{} admits the following approximation ratios, where $H(k)=\sum_{i=1}^k 1/i$ denotes the $k$th Harmonic number.
- For edge-costs, $\sum_{\ell=1}^k \frac{H(\Delta_\ell)}{k-\ell+1} \leq H(k) \cdot H(\Delta_k)=O(\ln^2k)$, where $\Delta_\ell={(4\ell-3)}^2$; [$s$-Based Rooted SNA]{} admits ratio $\sum_{\ell=1}^k \frac{H(2\ell-1)}{k-\ell+1} \leq H(k) \cdot H(2k-1)$.
- For node-costs, $\sum_{\ell=1}^k H(\Delta_\ell) \min\left\{\frac{p_{\max}}{k-\ell+1},1\right\} \leq
\min\{k,p_{\max} H(k)\} \cdot H(\Delta_k)$; [$s$-Based Rooted SNA]{}, and thus also undirected [SSL]{}, admits ratio\
$\sum_{\ell=1}^k H(2\ell-1) \min\left\{\frac{p_{\max}}{k-\ell+1},1\right\} \leq
\min\{k,p_{\max} \cdot H(k)\} \cdot H(2k-1)$.
Part (ii) implies that undirected [SSL]{} admits approximation ratio that depends on $k$ only; this ratio is better than the one implied by Theorem \[t:a\] for $k=o\left(\frac{\ln n}{\ln \ln n}\right)$. Furthermore, it improves the ratio $O(k \ln k)$ of Fukunaga [@Fuk] for $\kappa'$-[SSL]{} to $O(\ln^2 k)$, since we have $p_{\max}=1$ in this case (for $\hat{\kappa}$-[SSL]{} we get the same ratio $O(k \ln k)$ as in [@Fuk], since $p_{\max}=k$ in this case).
Finally, we consider the [SSL with Flow-Cost Bounds]{} problem.
\[t:fc-bounds\] [SSL with Flow-Cost Bounds]{} admits approximation ratio $H(d(V))+H(c(E))$.
Theorems \[t:hard\], \[t:aug\], \[t:a\], \[t:main’\], and \[t:fc-bounds\], are proved in Sections \[s:hard\], \[s:aug\], \[s:a\], \[s:main’\], and \[s:fc-bounds\], respectively.
Proof of Theorem \[t:hard\] {#s:hard}
===========================
We start by proving Observation \[o:reduction\]. The reduction is essentially the one in Definition \[d:pq\]. Given an instance of [SSL]{} construct an instance of [$s$-Based Rooted SNA]{} as follows: add to $G$ a new node $s$ of cost $0$, and for every $v \in V$ set $r_{sv}=d_v$ and put $p_v$ edges from $s$ to $v$ into $F$. Conversely, given an instance of [$s$-Based Rooted SNA]{} with node-costs and with $\delta_G(s)=\emptyset$ and $c(s)=0$, construct an instance of [SSL]{} as follows. Remove $s$ from $G$, and for every $v \in V$ set $p_v$ to be the number of edges in $F$ from $s$ to $v$ and $d_v=r_{sv}$. In both directions, it is not hard to see that $S$ is a solution to the [SSL]{} instance, if, and only if, the edge set $I$ of all edges in $F$ from $s$ to $S$ is a solution to the [$s$-Based Rooted SNA]{} instance, and clearly $I$ and $S$ have the same node-cost.
Note that in the above reduction, we have the following.
- The case of uniform demands (namely, $d_v=k$ for all $v \in V$) in [SSL]{} corresponds to the case of rooted uniform requirements (namely $r_{sv}=k$ for all $v \in V \setminus \{s\}$) in [$s$-Based Rooted SNA]{}.
- The case of unit costs (namely, $c_v=1$ for all $v \in V$) in [SSL]{} corresponds to the version of [$s$-Based Connectivity Augmentation]{} with node-costs, when we can pick $p_v$ edges from $s$ to $v$ by a cost of $1$.
Directed [Rooted SNA]{} with edge-costs and uniform requirements $r_{sv}=k$ for all $v \in V$ can be solved in polynomial time [@FT]; this easily implies that also undirected [Rooted $s$-Based SNA]{} with edge-costs can be solved in polynomial time. Thus the same holds for [$\kappa'$-SSL]{}, since [$\kappa'$-SSL]{} is a particular case of [$s$-Based Rooted SNA]{} with edge-costs.
Frank [@Frank] showed that [*directed*]{} [$s$-Based Connectivity Augmentation]{} with $\delta_G(s)=\emptyset$ is NP-hard. Using a slight modification of his reduction we can show that the problem is in fact [Set-Cover]{} hard to approximate, and thus is $\Omega(\log n)$-hard to approximate. Given an instance of [Set-Cover]{}, where a family $A$ of sets needs to cover a set $B$ of elements, construct the corresponding directed bipartite graph $G'=(A \cup B,E')$, by putting an edge from every set to each element it contains. The graph $G=(V,E)$ is obtained from $G'$ by adding $M$ copies of $B$, connecting $A$ to each copy in the same way as to $B$, and adding a new node $s$. Let $F=\{sv:v \in V\}$, $c(e)=1$ for every $e \in F$, and $r_{sv}=0$ if $v \in A$ and $r_{sv}=1$ otherwise. It is easy to see that if $I \subseteq F$ is a feasible solution to the obtained [Rooted $s$-Based SNA]{} instance, then either $I$ corresponds to a feasible solution to the [Set-Cover]{} instance, or $|I| \geq M$. The $\Omega(\log n)$-hardness follows for $M$ large enough, say $|M|={(|A|+|B|)}^2$, and $|A|=|B|$. Since for $k=1$ all connectivity functions of [SSL]{} are equivalent, we get $\Omega(\log n)$ hardness for directed [SSL]{} with $k=1$ and unit costs.
Proof of Theorem \[t:aug\] {#s:aug}
==========================
We use a result due to Wolsey [@W] about a performance of a greedy algorithm for submodular covering problems. A generic covering problem is as follows.
The [Submodular Cover]{} problem is a special case when the function $g$ is submodular and non-decreasing, and $c(S)=\sum_{v \in S} c(v)$ for some $c:U\rightarrow \mathbb{R}^+$. Wolsey [@W] proved that then, the greedy algorithm, that as long as $g(A)<g(U)$ repeatedely adds to $A$ an element $u \in U \setminus A$ with maximum $\frac{g(A \cup \{u\})-g(A)}{c_u}$, has approximation ratio $H\left(\max_{u \in U} g(\{u\})-g(\emptyset)\right)$.
We start with the case of edge-costs. Then the function $g$ is defined in the same way as in [@KN-aug]: $U=F$ and for $I \subseteq F$ $$g(I)=\sum_{uv \in D}\min\{r(u,v),f_{uv}(I)\} .$$ It is not hard to verify that $g$ is non-decreasing, and that $I$ is a feasible solution to an [NA]{} instance if and only if $g(I)=g(F)=r(D)$. Also, $g(\{e\}) - g(\emptyset) \leq \sum_{uv \in D} [f_{uv}(\{e\}) - f_{uv}(\emptyset)]$ for any $e \in F$. We show that $g$ is submodular. It is known (c.f. [@Sch]) that if $h$ is submodular, then $\min\{r,h\}$ is submodular for any constant $r$. Thus the function $h_{uv}(I)=\min\{r(u,v),f_{uv}(I)\}$ is submodular. As a sum of submodular functions is also submodular, we obtain that $g$ is submodular.
Now let us consider node-costs. For $S \subseteq V$ let $F_S$ denote the set of edges in $F$ from $a$ to $S$, and let $f'_{uv}(S)=f_{uv}(F_S)$. We have $U=V$ and for $S \subseteq V$ let $$g'(S)= \sum_{uv \in D}\min\{r(u,v),f'_{uv}(S)\} \ .$$ As in edge-costs case, it is not hard to verify that $g'$ is non-decreasing and that $S$ is a feasible solution to an [NA]{} instance if and only if $g'(S)=g'(V)=r(D)$. Also, $g'(\{z\}) - g'(\emptyset) \leq \sum_{uv \in D} [f_{uv}(\delta_F(z)) - f_{uv}(\emptyset)]$ for any $z \in V$. We show that $g'$ is submodular. We claim that the submodularity of $f(I)$ implies that $f'(S)$ is submodular. This is not true in general, but holds if $F$ is a star, and hence for [$a$-Based NA]{} instances. More precisely, it is not hard to verify the following statement, that finishes the proof of Theorem \[t:aug\].
\[l:star\] Let $(V,F)$ be a graph and let $f$ be a submodular set function on $F$. If $F$ is a star, then the set function $f'(S)=f(F_S)$ defined on $V$ is also submodular.
Proof of Theorem \[t:a\] {#s:a}
========================
All graphs in this section are assumed to be directed. Theorem \[t:a\] will follow from Theorem \[t:aug\] and the following Lemma, whose parts were implicitly proved in [@KN-aug].
\[l:max\] For any directed [$a$-Based SNA]{} instance, for any $s,v \in V$, the set-function $f_{sv}(I)=\lambda^q_{G+I}(s,v)$ on $F$ is submodular and non-decreasing. Furthermore, $f_{sv}(\{e\}) \leq f_{sv}(\emptyset)+1$ for any $e \in F$, and $f'_{sv}(\{z\})-f'_{sv}(\emptyset) \leq |\delta_F(z)| \leq p_{\max}$ for any $z \in V$.
In the rest of this section we prove Lemma \[l:max\]. Let $s,v \in V$. It is easy to see that $f(I)=f_{sv}(I)=\lambda^q_{G+I}(s,v)$ is non-decreasing, and that the second statement in the lemma holds. Thus it remains to prove that $f(I)$ is submodular. For that, we will use the following known characterization of submodularity, c.f. [@Sch]:\
[*A set-function $f$ on $F$ is submodular if, and only if*]{} $$f(I_0 \cup \{e\})+f(I_0 \cup \{e'\}) \geq f(I_0)+f(I_0 \cup \{e,e'\}) \ \ \ \forall I_0 \subset F, e,e' \in F \setminus I_0 $$ Let us fix $I_0 \subseteq F$. Revising our notation to $G \gets G+I_0$, $F \gets F \setminus I_0$, and denoting $h(I)=f(I_0 \cup I)-f(I_0)$, we get that $f$ is submodular if, and only if $$h(\{e\})+h(\{e'\}) \geq h(\{e,e'\}) \ \ \ \forall e,e' \in F \ .$$ In our setting, $h(I)=\lambda^q_{G+I}(s,v)-\lambda^q_G(s,v)$ is the increase in the $(s,v)$-$q$-connectivity as a result of adding $I$ to $G$. Thus $0 \leq h(I) \leq |I|$ for any $I \subseteq F$, so $0 \leq h(\{e,e'\}) \leq 2$. If $h(\{e,e'\})=0$, then we are done; if $h(\{e,e'\})=1$, then we need to show that $h(\{e\})=1$ or $h(\{e'\})=1$; and if $h(\{e,e'\})=2$, then we need to show that $h(\{e\})=1$ and $h(\{e'\})=1$. We prove the following general statement, that implies the above.
\[l:J\] Let $G=(V,E)$ be a directed graph with node capacities $\{q_v:v \in V\}$, let $I$ be a set of edges on $V$ disjoint to $E$, let $s,t \in V$, and let $h=\lambda^q_{G+I}(s,t)-\lambda^q_G(s,t)$. Then there is $J \subseteq I$ of size $|J| \geq h$ such that $\lambda^q_{G+\{e\}}(s,t)=\lambda^q_G(s,t)+1$ for every $e \in J$.
Since we consider directed graphs, it is sufficient to prove the lemma for the case of edge-connectivity. For that, apply the following standard reduction that eliminates node capacities: replace every $v \in V \setminus \{s,t\}$ by two nodes $v^{in}, v^{out}$ connected by $q_v$ parallel edges from $v^{in}$ to $v^{out}$ and replace every $uv \in E \cup F$ by an edge from $u^{out}$ to $v^{in}$. Hence we will prove the lemma for the edge connectivity function $\lambda$. Let us say that $S \subseteq V$ is [*tight*]{} if $s \in S$, $v \notin S$, and $|\delta_G(S)|=\lambda_G(s,v)$. Let ${\cal F}$ be the family of tight sets. By Menger’s Theorem ${\cal F}$ is non-empty. It is known that ${\cal F}$ is a ring family, namely, the intersection of all the sets in ${\cal F}$ is nonempty, and if $X,Y \in {\cal F}$ then $X \cap Y, X \cup Y \in {\cal F}$. Then ${\cal F}$ has a unique inclusion-minimal set $S_{\min}$ and a unique inclusion-maximal set $S_{\max}$. Let $J=\{uv \in I:u \in S_{\min},v \in V \setminus S_{\max}\}$ be the set of edges in $I$ that go from $S_{\min}$ to $V \setminus S_{\max}$. By Menger’s Theorem, $|J| \geq h$, and $\lambda_{G+\{e\}}(s,t)=\lambda_G(s,t)+1$ for any $e \in J$. The statement follows.
Proof of Theorem \[t:main’\] {#s:main'}
============================
Here we prove Theorem \[t:main’\]. All graphs in this and the next section are assumed to be undirected, unless stated otherwise. We start by considering the edge-costs case, and then will show that it implies the node-costs case by reductions.
An ordered pair $\hat{X}=(X,X^+)$ of subsets of a groundset $V$ is called a [*biset*]{} if $X \subseteq X^+$; $X$ is the [*inner part*]{} and $X^+$ is the [*outer part*]{} of $\hat{X}$, and $\Gamma(\hat{X})=X^+ \setminus X$ is the [*boundary*]{} of $\hat{X}$. An edge $e$ covers a biset $\hat{X}$ if it has one endnode in $X$ and the other in $V \setminus X^+$. For a biset $\hat{X}$ and an edge-set/graph $J$ let $\delta_J(\hat{X})$ denote the set of edges in $J$ covering $\hat{X}$.
Given an instance of [SNA]{} and a biset $\hat{X}$ on $V$, let the requirement of $\hat{X}$ be $r(\hat{X})=\max\{r_{uv}: uv \in \delta_D(\hat{X})\}$ if $\delta_D(\hat{X}) \neq \emptyset$ and $r(\hat{X})=0$ otherwise. By the $q$-connectivity version of Menger’s Theorem (c.f. [@KN-sur]), $I \subseteq F$ is a feasible solution to an [SNA]{} instance if, and only if, $|\delta_I(\hat{X})| \geq h(\hat{X})$ for every bisets $\hat{X}$ on $V$, where $h$ is a biset-function defined by $$\label{e:p}
h(\hat{X})= \max\{r(\hat{X})-(q(\Gamma(\hat{X}))+|\delta_G(\hat{X})|),0\}$$ Let ${\cal P}_h$ denote the polytope of “fractional edge-covers” of $h$, namely, $${\cal P}_h=\left\{x \in \mathbb{R}^F:
x\left(\delta_F(\hat{Y})\right) \geq h(\hat{Y}) \ \forall \mbox{ biset } \hat{Y} \mbox{ on } V, \
0 \leq x_e \leq 1 \ \forall e \in F\right\} \ .$$ Let $\tau^*(h)$ denote the optimal value of a standard LP-relaxation for edge covering $h$ by a minimum cost edge set, namely, $\tau^*(h)=\min\left\{\sum_{e \in F} c_e x_e: x \in {\cal P}_h \right\}$. As an intermediate problem, we consider [SNA]{} instances when we seek to increase the connectivity by $1$ for every $uv \in D$, namely, when $r_{uv} = \lambda^q_G(u,v)+1$ for all $uv \in D$.
Given a [$D$-SNA]{} instance, we say that a biset $\hat{X}$ is [*tight*]{} if $h(\hat{X})=1$, where $h$ is defined by (\[e:p\]). [$D$-SNA]{} is equivalent to the problem of finding a minimum cost edge-cover of the biset family ${\cal F}=\{\hat{X}:h(\hat{X})=1\}$ of tight bisets. Thus the following generic problem includes the [$D$-SNA]{} problem.
For a biset-family ${\cal F}$ let $\tau^*({\cal F})$ denote the optimal value of a standard LP-relaxation for edge covering ${\cal F}$ by a minimum cost edge set, namely, $\tau^*({\cal F})=\tau^*(h)$ for $h(\hat{X})=1$ if $\hat{X} \in {\cal F}$ and $h(\hat{X})=0$ otherwise.
\[p:scale\] Suppose that [$a$-Based $D$-SNA]{} with edge-costs admits a polynomial time algorithm that computes a solution of cost at most $\rho(k) \tau^*({\cal F})$, where ${\cal F}$ is the family of tight bisets. Then [$a$-Based SNA]{} admits a polynomial time algorithm that computes a solution $I$ such that:
- For edge-costs, $c(I) \leq \tau^*(h) \cdot \sum_{\ell=1}^k \frac{\rho(\ell)}{k-\ell+1}$, where $h$ is defined by (\[e:p\]).
- For node-costs, $c(I) \leq {\sf opt} \cdot \sum_{\ell=1}^k \rho(\ell) \cdot \min\left\{\frac{p_{\max}}{k-\ell+1},1\right\}$.
We start with the edge-costs case. Consider the following sequential algorithm. Start with $I=\emptyset$. At iteration $\ell=1, \ldots,k$, add to $I$ and remove from $F$ an edge-set $I_\ell \subseteq F$ that increases by $1$ the $q$-connectivity of $G+I$ on the set of demand edges $D_\ell=\{sv:\lambda_{G+I}^q(s,v)=r(s,v)-k+\ell-1, sv \in D\}$, by covering the corresponding biset-family ${\cal F}_\ell$ using the $\rho$-approximation algorithm. After iteration $\ell$, we have $\lambda_{G+I}^q(s,v) \geq r(s,v)-k+\ell$ for all $sv \in D$. Consequently, after $k$ iterations $\lambda_{G+I}^q(s,v) \geq r(s,v)$ holds for all $sv \in D$, thus the computed solution is feasible. The approximation ratio follows from the following two observations.
- $c(I_\ell) \leq \rho(\ell) \cdot \tau^*({\cal F}_\ell)$. This is so since $\lambda(s,v) \leq \ell-1$ for every $sv \in D_\ell$, hence the maximum requirement at iteration $\ell$ is at most $\ell$.
- $\tau^*({\cal F}_\ell) \leq \frac{\tau^*(h)}{k-\ell+1}$. To see this, note that if $\hat{Y} \in {\cal F}_\ell$ and $x \in {\cal P}_h$ then $x(\delta(\hat{Y})) \geq k-\ell+1$, by Menger’s Theorem. Thus $x/(k-\ell+1)$ is a feasible solution for the LP-relaxation for edge-covering ${\cal F}_\ell$, of value $c \cdot x/(k-\ell+1)$.
Consequently, $c(I) =
\sum_{\ell=1}^k c(I_\ell) \leq \sum_{\ell=1}^k \rho(\ell) \cdot \frac{\tau^*(h)}{k-\ell+1}=
\tau^*(h) \cdot \sum_{\ell=1}^k \frac{\rho(\ell)}{k-\ell+1}$.
Now let us consider the case of node-costs. Then we convert node-costs into edge-costs by assigning to every edge $e=av$ the cost $c'(e)=c(v)$. Let ${\sf opt}'$ denote the optimal solution value of the edge-costs instance obtained. Clearly, ${\sf opt} \leq {\sf opt}' \leq p_{\max} \cdot {\sf opt}$. Note that any inclusion minimal solution to an [$a$-Based $D$-SNA]{} instance has no parallel edges. This implies that $c(I_\ell) \leq \rho(\ell) \cdot {\sf opt}$ and that $c(I_\ell)=c'(I_\ell)$. The latter implies $c(I_\ell)=c'(I_\ell) \leq \rho(\ell) \cdot \frac {{\sf opt}'}{k-\ell+1} \leq
\rho(\ell) \cdot {\sf opt} \cdot \frac {p_{\max}}{k-\ell+1}$, and the statement for the node-costs case follows.
In the next section we prove the following theorem, that together with Proposition \[p:scale\] finishes the proof of Theorem \[t:main’\].
\[t:D\] For edge-costs, undirected [$a$-Based $D$-SNA]{} admits a polynomial time algorithm that computes a feasible solution $I$ of cost $c(I) \leq H(\Delta_k) \cdot \tau^*({\cal F})$, and $c(I) \leq H(2k-1) \cdot \tau^*({\cal F})$ if $D$ is a star with center $s=a$, where ${\cal F}$ is the family of tight bisets.
Proof of Theorem \[t:D\]
========================
Recall that [$D$-SNA]{} reduces to [Biset-Family Edge-Cover]{} with ${\cal F}$ being the family of tight bisets; in the case of rooted requirements, it is sufficient to cover the biset-family ${\cal F}^s=\{\hat{X} \in {\cal F}: s \in V \setminus X^+\}$. Biset-families arising from [SNA]{} instances have some special properties, that are summarized in the following definitions.
The intersection and the union of two bisets $\hat{X},\hat{Y}$ is defined by $\hat{X} \cap \hat{Y} = (X \cap Y, X^+ \cap Y^+)$ and $\hat{X} \cup \hat{Y} = (X \cup Y,X^+ \cup Y^+)$. The biset $\hat{X} \setminus \hat{Y}$ is defined by $\hat{X} \setminus \hat{Y}=(X \setminus Y^+,X^+ \setminus Y)$. We write $\hat{X} \subseteq \hat{Y}$ and say that [*$\hat{Y}$ contains $\hat{X}$*]{} if $X \subseteq Y$ and $X^+ \subseteq Y^+$. Let ${\cal C}_{\cal F}$ denote the inclusion-minimal bisets in ${\cal F}$.
\[d:uncrossable\] Two bisets $\hat{X},\hat{Y}$ covered by an edge-set $D$ are [*$D$-independent*]{} if for any $xx',yy' \in D$ such that $xx'$ covers $\hat{X}$ and $yy'$ covers $\hat{Y}$, $\{x,x'\} \cap \Gamma(\hat{Y}) \neq \emptyset$ or $\{y,y'\} \cap \Gamma(\hat{X}) \neq \emptyset$; otherwise, $\hat{X},\hat{Y}$ are [*$D$-dependent*]{}. We say that a biset family ${\cal F}$ is [*$D$-uncrossable*]{} if $D$ covers ${\cal F}$ and if for any $D$-dependent $\hat{X},\hat{Y} \in {\cal F}$ the following holds: $$\label{e:uncross}
\hat{X} \cap \hat{Y},\hat{X} \cup \hat{Y} \in {\cal F} \mbox{ or }
\hat{X} \setminus \hat{Y},\hat{Y} \setminus \hat{X} \in {\cal F} \ .$$ Similarly, given a set $T \subseteq V$ of terminals, we say that $\hat{X}, \hat{Y}$ are [*$T$-independent*]{} if $X \cap T \subseteq \Gamma(\hat{Y})$ or if $Y \cap T \subseteq \Gamma(\hat{X})$, and $\hat{X},\hat{Y}$ are [*$T$-dependent*]{} otherwise. We say that ${\cal F}$ is $T$-uncrossable if $T$ covers the set-family of the inner parts of ${\cal F}$, and if (\[e:uncross\]) holds for any $T$-dependent $\hat{X},\hat{Y} \in {\cal F}$.
A biset-family ${\cal F}$ is symmetric if $\hat{X} \in {\cal F}$ implies $(V \setminus X^+,V \setminus X) \in {\cal F}$. Clearly, the family of tight bisets is symmetric. We will use the the following fundamental statement, that was implicitly proved in [@N-aug].
\[[@N-aug]\] The family ${\cal F}$ of tight bisets is $D$-uncrossable, and if $D$ is a star with center $s$ and leaf-set $T$ then $\{\hat{X} \in {\cal F}: s \notin X^+\}$ is $T$-uncrossable.
For a biset-family ${\cal C}$ let $\gamma_{\cal C}=\max\{|\Gamma(\hat{C})|: \hat{C} \in {\cal C}\}$. Note that if ${\cal F}$ is the family of tight bisets then $\gamma_{\cal F} \leq k-1$. Given an instance of [Biset-Family Edge-Cover]{}, we will assume that the family ${\cal C}$ of the inclusion members of ${\cal F}$ can be computed in polynomial time. We note that for ${\cal F}$ being the family of tight sets, this step can be implemented in polynomial time, c.f. [@N-aug]. Under this assumption, we prove the following generalization of Theorem \[t:D\].
\[t:main\] For edge/node-costs, [$a$-Based Biset-Family Edge-Cover]{} admits a polynomial time algorithm that computes a cover $I$ of ${\cal F}$ such that:
- $c(I) \leq H\left({(4\gamma_{\cal C}+1)}^2\right) \cdot \tau^*({\cal F})$ if ${\cal F}$ is symmetric and $D$-uncrossable.
- $c(I) \leq H(2\gamma_{\cal C}+1) \cdot \tau^*({\cal F})$ if ${\cal F}$ is $T$-uncrossable and $a \in V \setminus X^+$ for all $\hat{X} \in {\cal F}$.
In the rest of this section we prove Theorem \[t:main\].
A node set $U \subseteq V$ is a [${\cal C}$-transversal]{} of a hypergraph (set-family) ${\cal C}$ on $V$ if $U$ intersects every set in ${\cal C}$; if ${\cal C}$ is a biset-family then $U$ should intersect the inner part of every member of ${\cal C}$. Given costs $\{c_v:v \in V\}$, let $t^*({\cal C})$ denote the minimum value of a fractional ${\cal C}$-transversal, namely: $$t^*({\cal C})=\min\{\sum_{v \in V} c_v x_v:
x(C) \geq 1 \ \ \forall \hat{C} \in {\cal C}, \ x(v) \geq 0 \ \forall v \in V\} \ .$$
In [@N-aug], the following is proved.
\[[@N-aug]\] \[t:2\] Let ${\cal F}$ be a biset-family and let ${\cal C}$ be the family of the inclusion members of ${\cal F}$. Then the maximum degree in the hypergraph $\{C:\hat{C} \in {\cal C}\}$ is at most:
- ${(4\gamma_{\cal C}+1)}^2$ if ${\cal F}$ is $D$-uncrossable.
- $2\gamma_{\cal C}+1$ if ${\cal F}$ is $T$-uncrossable.
Given a hypergraph $(V,{\cal C})$ with node-costs, the greedy algorithm computes in polynomial time a ${\cal C}$-transversal $U \subseteq V$ of cost $c(U) \leq H(\Delta({\cal C})) t^*({\cal C})$, where $\Delta({\cal C})$ is the maximum degree of the hypergraph (c.f. [@Lov]).
\[l:ratio\] If an edge-set $I$ covers a biset-family ${\cal F}$ then the set of endnodes of $I$ is a transversal of ${\cal F}$.
\[l:feasible\] Let ${\cal F}$ be a biset family on $V$ and $I$ a star with center $a$ on a transversal $U \subseteq V$ of ${\cal F}$. Then $I$ covers ${\cal F}$ in each one of the following cases.
- ${\cal F}$ is symmetric and $a \notin \Gamma(\hat{X})$ for all $\hat{X} \in {\cal F}$.
- $a \in V \setminus X^+$ for all $\hat{X} \in {\cal F}$.
Let $\hat{X} \in {\cal F}$. Then $a \in X$ or $a \in V \setminus X^+$. If $a \in V \setminus X^+$, then since $U$ is a transversal of ${\cal C}$, there is $u \in U \cap X$. If $a \in X$, then if ${\cal F}$ is symmetric, then there $u \in U \cap (V \setminus X^+)$. In both cases, there is an edge $au \in I$, and this edge covers $\hat{X}$.
The algorithm as in Theorem \[t:main\], for both edge-costs and node-costs is as follows, where in the case of node-costs we may assume that the cost of $a$ is zero.
The solution computed is feasible by Lemma \[l:feasible\]. The approximation ratio follows from Theorem \[t:2\] and Lemma \[l:ratio\].\
Proof of Theorem \[t:fc-bounds\] {#s:fc-bounds}
================================
[SSL with Flow-Cost Bounds]{} is a special case of the following generalization of [Submodular Cover]{}, where we have two progress functions: $$f(S)= \sum_{v \in V}\min\{\lambda^{p,q}_G(S,v),d_v\} \ \mbox{ and} \
g(S)= \sum_{v \in V}\min\{-\mu^{p,q}_G(S,v),-b_v\} .$$ It is easy to see that $S$ is a feasible solution to [SSL with Flow-Cost Bounds]{} if and only if both $f(S)=f(V)=\sum_{v \in V} d_v$ and $g(S)=g(V)=-\sum_{v \in V} b_v$. Note that $\max_{u \in U} f(\{u\})-f(\emptyset) \leq d(V)$, and that $\max_{u \in U} g(\{u\})-g(\emptyset) \leq c(E)$ if $g(\emptyset) \neq -\infty$. The function $f$ is submodular since for any $v \in V$ the function $f_v(S)=\lambda^{p,q}_G(S,v)$ is submodular, as can be deduced from Lemmas \[l:star\] and \[l:max\]. The function $g$ is submodular since for any $v \in V$ the function $g_v(S)=\lambda^{p,q}_G(S,v)$ is submodular; this is proved in [@BKP] for the case of edge-connectivity, and the proof for $(p,q)$-connectivity is similar. Also, both functions are non-decreasing and admit a polynomial time evaluation oracle.
There are several natural approaches to solve the [Double Submodular Cover]{} problem using the greedy algorithm of Wolsey [@W]. One is to apply the greedy algorithm with the function $f+g$. Another possibility is to solve two instances of [Submodular Cover]{}, one with function $f$ and the other with function $g$, returning the union of the solutions $S_f$ and $S_g$ computed. However, in both case the ratio guarantee may be unbounded if $g(\emptyset)=-\infty$, which happens in the case of [SSL with Flow-Cost Bounds]{}.
Note that in [Double Submodular Cover]{} instances that arise from [SSL with Flow-Cost Bounds]{} have the following property: if $f(S_f)=f(U)$ then $g(S_f) \neq -\infty$ for any $S \supseteq S_f$. Therefore, the following approach works. We take the set $S_f$ into our solution, and consider the residual [Submodular Cover]{} problem with groundset $V \setminus S_f$ and set function $h(S)=g(S_f \cup S)$, $S \subseteq V \setminus S_f$. The function $h$ is submodular if $g$ is, and we get approximation ratio $$H\left(\max_{v \in V} f(\{v\})-f(\emptyset)\right)+H\left(\max_{v \in V} h(\{v\})-h(\emptyset)\right)
\leq H(d(V))+H(c(E)) \ .$$
Clearly, the approach described can be generalized to the case when we have many non-decreasing submodular functions, under the assumption that there exists an ordering $f_1,f_2, \ldots$ of the functions such that for any $i$, if $f_j(S)=f(U)$ for every $j \leq i$, then $f_{j+1}(S') \neq -\infty$ for any $S' \supseteq S$.
[**Acknowledgment**]{} The second author thank Takuro Fukunaga and an anonymous referee for many useful comments.
[10]{}
K. Arata, S. Iwata, K. Makino, and S. Fujishige. Locating sources to meet flow demands in undirected networks. , 42:54––68, 2002.
J. Bar-Ilan, G. Kortsarz, and D. Peleg. Generalized submodular cover problems and applications. , 250(1-2):179–200, 2001.
M. Barasz, J. Becker, and A. Frank. An algorithm for source location in directed graphs. , 33:221––230, 2005.
A. Frank. Augmenting graphs to meet edge-connectivity requirements. , 5(1):25–53, 1992.
A. Frank and E. Tardos. An application of submodular flows. , 114/115:329–348, 1989.
T. Fukunaga. Approximating minimum cost source location problems with local vertex-connectivity demands. In [*TAMC*]{}, pages 428–439, 2011.
T. Ishii. Greedy approximation for source location problem with vertex-connectivity requirements in undirected graphs. , 7:570–578, 2009.
T. Ishii, H. Fujita, and H. Nagamochi. Source location problem with local [$3$]{}-vertex-connectivity requirements. In [*Proc. 3rd Hungarian–Japanese Symposium on Discrete Mathematics and Its Applications*]{}, pages 368––377, 2003.
T. Ishii, H. Fujita, and H. Nagamochi. Minimum cost source location problem with local [$3$]{}-vertex-connectivity requirements. , 372:81––93, 2007.
T. Ishii and K. Makino. Posi-modular systems with modulotone requirements under permutation constraints. , 2(1):61–76, 2010.
H. Ito, M. Ito, Y. Itatsu, H. Uehara, and M. Yokoyama. Source location problems considering vertex-connectivity and edge-connectivity simultaneously. , 40:63––70, 2002.
H. Ito, K. Makino, K. Arata, S. Honami, Y. Itatsu, and S. Fujishige. Source location problem with flow requirements in directed networks. , 18(4):427––435, 2003.
G. Kortsarz and Z. Nutov. PWS, 2007.
G. Kortsarz and Z. Nutov. Tight approximation algorithm for connectivity augmentation problems. , 74(5):662–670, 2008.
L. Lovász. On the ratio of optimal integral and fractional covers. , 13:383–390, 1975.
S. Mamada, T. Uno, K. Makino, and S. Fujishige. An [$O(n\log^2 n)$]{} algorithm for the optimal sink location problem in dynamic tree networks. , 154(16):2387–2401, 2006.
H. Nagamochi and T. Ibaraki. . Cambridge University Press, 2008.
H. Nagamochi, T. Ishii, and H. Ito. Minimum cost source location problem with vertex-connectivity requirements in digraphs. , 80:287––294, 2001.
Z. Nutov. Approximating minimum cost connectivity problems via uncrossable bifamilies. To appear in [*Transactions on Algorithms*]{}. Preliminary version in [*FOCS*]{} 2009, 417–426.
Z. Nutov. Approximating node-connectivity augmentation problems. , 63(1-2):398–410, 2012.
M. Sakashita, K. Makino, and S. Fujishige. Minimum cost source location problems with flow requirements. , 50:555–583, 2008.
M. Sakashita, K. Makino, H. Nagamochi, and S. Fujishige. Minimum transversals in posi-modular systems. , 23:858–871, 2009.
A. Schrijver. . Springer, 2004.
H. Tamura, M. Sengoku, S. Shinoda, and T. Abe. Location problems on undirected flow networks. , E73:1989––1993, 1990.
L. A. Wolsey. An analysis of the greedy algorithm for the submodular set covering problem. , 2:385––393, 1982.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Phylogenetic networks have gained prominence over the years due to their ability to represent complex non-treelike evolutionary events such as recombination or hybridization. Popular combinatorial objects used to construct them are triplet systems and cluster systems, the motivation being that any network $N$ induces a triplet system $\mathcal R(N)$ and a softwired cluster system $\mathcal S(N)$. Since in real-world studies it cannot be guaranteed that all triplets/softwired clusters induced by a network are available, it is of particular interest to understand whether subsets of $\mathcal R(N)$ or $\mathcal S(N)$ allow one to uniquely reconstruct the underlying network $N$. Here we show that even within the highly restricted yet biologically interesting space of level-1 phylogenetic networks it is not always possible to uniquely reconstruct a level-1 network $N$ even when all triplets in $\mathcal R(N)$ or all clusters in $\mathcal S(N)$ are available. On the positive side, we introduce a reasonably large subclass of level-1 networks the members of which are uniquely determined by their induced triplet/softwired cluster systems. Along the way, we also establish various enumerative results, both positive and negative, including results which show that certain special subclasses of level-1 networks $N$ can be uniquely reconstructed from proper subsets of $\mathcal R(N)$ and $\mathcal S(N)$. We anticipate these results to be of use in the design of algorithms for phylogenetic network inference.'
author:
- 'P. Gambette, K.T. Huber, S. Kelk'
bibliography:
- 'GambetteHuberKelk\_revision\_4april2016.bib'
date: '4th April, 2016'
title: 'On the challenge of reconstructing level-1 phylogenetic networks from triplets and clusters'
---
[^1]
Introduction
============
cases where the taxa are suspected to have undergone reticulate evolutionary events such as hybridization or recombination have been found to not always be appropriate [@S75]. The need for structures capable of appropriately dealing with such data sets combined with the fact that different evolutionary processes have given rise to them has resulted in the introduction of a number of more general structures for representing evolutionary relationships. under the name these include [@BSS06], recombination graphs [@H90], galled trees [@gusfield2004optimal; @WZZ01], normal networks [@W10], regular networks [@BSS04], tree-sibling networks [@CLRV08], level-$k$ networks [@JNS06; @IKKSHB09], median networks [@B94] and NeighborNets [@BM03], to name just a few, which all generalize a phylogenetic tree in one way or another.
Apart from median networks and NeighborNets which are a special type of split-based phylogenetic network, the basic graph-theoretical structure underpinning a phylogenetic network is a rooted directed acyclic graph (DAG) that has a unique root and whose set of sinks is a given set of taxa. of the combinatorially simplest types of phylogenetic network but still complicated enough to be of interest to Evolutionary Biology is that of a binary level-$1$ network (see Fig. \[fig:blobby-shapely\] for an example).
![\[fig:saturated-shapely\] A phylogenetic network $N$ on $X=\{1,\ldots, 10\}$ []{data-label="fig:blobby-shapely"}](blobby-shapely.eps)
have attracted a considerable amount of interest in the literature (see e.g.[@JNS06; @gusfield2004optimal; @rossello2009all; @HIKS2011]) and can informally be thought of as rooted DAGs with . . However, this simplicity has proven to be deceptive as the combinatorial structure of such networks has turned out to be more complicated than originally thought (see e.g. [@GH12; @HM13]).
In this paper, we [*cluster systems*]{}, that is, collections of non-empty subsets of , [*triplet systems*]{}, that is, binary phylogenetic trees on just three leaves. for reconstructing phylogenetic networks (see e.g.[@M09] and [@HRS10] for ), we are particularly interested in finding bounds on the minimum size of a triplet system/cluster system to “uniquely determine” a level-1 network. For trees this question is well understood. Specifically, for a phylogenetic tree $T$ on $n \geq 3$ leaves it is well-known that $T$ is uniquely determined by its induced triplet system $\mathcal R(T)$ (leading to an upper bound of ${n\choose 3}$ for such a minimum-sized set) and that $n-2$ carefully chosen triplets from $\mathcal R(T)$ suffice to uniquely reconstruct $T$ when $T$ is binary (see Theorem 3 of [@Steel1992] and its Corollary). For this case, it is also well-known that $T$ is uniquely determined by its induced cluster system $\mathcal C(T)$ and that for a minimum-sized cluster system to uniquely determine $T$, it must have $|\mathcal C(N)|=2n-1$ elements.
As we shall see, the situation is more complicated for binary level-1 networks. Every level-1 network $N$ induces a triplet system $\mathcal R(N)$ and a certain cluster system $\mathcal S(N)$ called the [*softwired cluster system of $N$*]{} (see [@huson2011survey] for background) but their ability to fully capture the topological structure of $N$ is not as strong as one might hope. Let us say that a binary level-1 network $N$ is [*encoded*]{} by its induced triplet system if for every binary level-1 network $N'$ such that $\mathcal R(N')=\mathcal R(N)$, we have $N = N'$. Continuing, we say that a binary level-1 network is [*4-outwards*]{} if its underlying graph does not have a cycle of length four or less. It is precisely the 4-outwards binary level-1 networks $N$ that are encoded by $\mathcal R(N)$ as well as $\mathcal S(N)$ [@GH12] (where we define a binary level-1 network to be encoded by its induced softwired cluster system in an analogous way).
Intriguingly, if $\mathcal R(N')=\mathcal R(N)$ is replaced by $\mathcal R(N)\subseteq \mathcal R(N')$ (as is the case in our formalization of “uniquely determining”) then the assumption that $N$ is $4$-outwards is no longer strong enough to guarantee uniqueness. A similar observation holds for $\mathcal S(N)$ (see Sections \[sec:triplets-define\] and \[sec:define-clusters\] for examples for both cases). However, the situation changes for both if, in addition to being $4$-outwards we require that $N$ is [*saturated*]{}, that is, none of its vertices is incident with more than one cut arc (Theorem \[theo:shapely-saturated-triplet\] and Theorem \[theo:shapely-saturated-cluster\]). Simple networks on $n\geq 4$ leaves are 4-outwards, saturated networks that have precisely one cycle in their underlying graph. . As the network on four leaves depicted in Fig. \[fig:new\] indicates, this bound is however not tight because five triplets suffice in that case (which can be checked by a simple case analysis). Given that any binary level-1 network $N$ contains at least one triplet for any three of its leaves and so $|\mathcal R(N)|\geq {n\choose 3}$ holds, this suggests that at least for simple phylogenetic networks there is a considerable amount of redundancy in $\mathcal R(N)$ with regards to reconstructing $N$ from $\mathcal R(N)$. To establish a similar result for general binary level-1 networks $N$ might not be in view of Proposition \[prop:triplet-system-size\] which suggests that $|\mathcal R(N)|$ is not easily expressible in terms of a natural parameter associated a phylogenetic network $N$ namely its number of non-trivial cut arcs (see Section \[arcs-vertices-galls\]). This is somewhat surprising in view of the close relationship between the triplet system induced by a binary level-1 network $N$ and its associated softwired cluster system $\mathcal S(N)$ (see e.g.[@GH12 Proposition 2 and Theorem 1] for details concerning this relationship) because the size of $\mathcal S(N)$ *is* closely related to the number of cut arcs of $N$ (Theorem \[theo:countingClusters\]). As in the case of triplet systems, it is easy to find examples of binary level-1 networks $N$ that indicate that there is redundancy in the softwired cluster system induced by $N$ with regards to uniquely determining $N$. Again focusing on simple networks $N$, we show that at most $n$ carefully chosen (softwired) clusters induced by $N$ suffice to uniquely determine $N$ (Corollary \[cor:defineSimple\]). However, we do not know if this bound is sharp.
Given that in phylogenetic one is hardly ever guaranteed to have all triplets/clusters induced by a (as yet unknown) phylogenetic network , the above observations have profound consequences for phylogenetic network reconstruction. One of the most important that a phylogenetic network reconstructed a triplet or cluster system need not be the network that gave rise to .
The paper is organized as follows. In the next section, we present basic terminology of relevance to this paper including the definition of a level-$k$ network and that of a gall in a level-1 network. In Section \[arcs-vertices-galls\], we define cut arcs and present formulas for counting the number of vertices, arcs, and galls in a binary level-1 network. These results improve on the results in [@CJSS05] which imply that the number of vertices in a binary level-1 network on $n$ leaves is linear in $n$ and that the number of hybrid vertices is at most $n-1$. In Section \[counting-softwired\], we formally define the softwired cluster system $\mathcal S(N)$ induced by a binary level-1 network $N$ and establish Theorem \[theo:countingClusters\]. In Section \[counting-triplets\], we define the triplet system $\mathcal R(N)$ induced by a binary level-1 network $N$ and establish Proposition \[prop:triplet-system-size\]. In Section \[sec:triplets-sn\], we establish in Proposition \[prop:samecut\] a relationship between the triplet system induced by a binary level-1 network $N$ and a certain partition of the leaf set of $N$ that will be crucial for showing Theorem \[theo:shapely-saturated-triplet\]. In Section \[sec:triplets-define\], we first formalize the notion of “uniquely determining” and then present the aforementioned examples for triplet systems. Starting in that section and continuing in Section \[sec:define-clusters\], we investigate saturated, 4-outwards, binary level-1 networks and establish Theorem \[theo:shapely-saturated-triplet\] and Theorem \[theo:shapely-saturated-cluster\], respectively.
Definitions and Notation {#notation}
========================
Throughout the paper, let $X$ denote a finite set of size $n\geq 2$. Also all graphs $G$ considered have non-empty finite sets of vertices and edges (or arcs in case $G$ is directed) and have no loops or multiple edges (or arcs in case $G$ is directed).
Suppose for the following that $G=(V,A)$ is a directed acyclic graph (DAG). If $v$ and $w$ are vertices of $G$ such that there exists an arc $a$ from $v$ to $w$ in $G$ then we denote that arc by $(v,w)$ and refer to $v$ as the [*tail of $a$*]{}, denoted by $tail(a)$, and to $w$ as the [*head of $a$*]{}, denoted by $head(a)$. Suppose $v\in V$ is a vertex of $G$. Then we denote by the out-degree of $v$ and by the [*in-degree*]{} of $v$ . The sum of the out-degree and the in-degree of $v$ is called the [*degree*]{} of $v$ denoted by . If $indeg(v)=1$ and $outdeg(v)=0$ then $v$ is called a [*leaf*]{} of $G$. The set of leaves of $G$ is denoted by $L(G)$. Every vertex in $V-L(G)$ is called an [*interior vertex*]{} of $G$. If $G$ has a unique vertex $\rho=\rho_G\in V$ with $indeg(\rho)=0$ and $outdeg(\rho)\geq 2$ then $\rho$ is called the [*root*]{} of $G$ and $G$ is called a [*rooted DAG*]{}. If $G$ is a rooted DAG with leaf set $X$ and $G'=(V',A')$ is a further rooted DAG with leaf set $X$ then we say that $G$ is [*equivalent*]{} to $G'$ .
A [*phylogenetic network $N$ on $X$*]{} is a rooted DAG whose set of leaves is $X$, and every interior vertex $v$ of $N$ except the root $\rho_N$ is either (i) a [ *vertex*]{} of $N$, that is, $indeg(v)=1$ and $outdeg(v)\geq 2$ or (ii) a [*hybrid vertex*]{} of $N$, that is, $indeg(v)\geq 2$ and $outdeg(v)\geq 1$. In case only the size of $X$ is of relevance to the discussion then we will simply call $N$ a [*phylogenetic network on $|X|$ leaves*]{} and if the set $X$ is of no relevance to the discussion then we will simply call a phylogenetic network $N$ on $X$ a [*phylogenetic network*]{}. We denote the set of hybrid vertices of a phylogenetic network $N$ by $H(N)$ and say that $N$ is [*binary*]{} if the root of $N$ as well as every vertex of $N$ has out-degree two and . graph $G$ is called [*biconnected*]{} if $G$ is connected and $G - v$ is connected for all $v \in V(G)$. a [*level-$k$ (phylogenetic) network*]{}, if every biconnected component of $U(N)$ contains at most $k$ hybrid vertices. Reflecting the fact that a cycle of length three in the underlying graph of a phylogenetic network is indistinguishable (from a triplet or cluster perspective) from a vertex, we follow common practice and will always assume that a cycle in the underlying graph of a level-1 network $N$ contains at least four vertices.
Note that a phylogenetic network $N$ for which $H(N)=\emptyset$ holds is simply a [*rooted phylogenetic tree*]{} on $X$ (sensu [@SS03]). Thus, level-0 networks are rooted phylogenetic trees. All phylogenetic trees considered in this article are rooted so we henceforth drop the “rooted” prefix.
We denote the class of all binary level-1 networks on $n\geq 2$ leaves by ${{\mathcal L}}_1(n)$. Alternatively, we will also use ${{\mathcal L}}_1(X)$ to denote that class if we want to emphasize the leaf set $X$ of the networks in ${{\mathcal L}}_1(n)$. Now, suppose that $N$ is a level-$k$ network, $k\geq 1$. Then we call $N$ [*proper*]{} if $N$ is not also a level-$l$ network for some $0\leq l\leq k-1$. Note that in case $k=1$ such a network must have at least three leaves and at least one hybridization vertex. In that case, we call a biconnected component of $U(N)$ with its original directions in $N$ restored a [*gall*]{} of $N$ and denote the set of galls of a level-1 network $N$ by $\mathcal G(N)$. If $N$ is binary, contains precisely one gall $C$, and every leaf of $N$ is adjacent with a vertex of $C$ then $N$ is called [*simple*]{}. Together with phylogenetic trees, such networks may be viewed as the building blocks of (proper) level-1 networks [@IKKSHB09]. For the convenience of the reader, we present examples of two simple level-1 networks on $X=\{x_1,\ldots, x_5\}$ in Fig. \[fig:simple\].
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
![\[fig:simple\] Two examples of simple level-1 networks on $X=\{x_1,\ldots, x_5\}$. Note that both networks are 4-outwards and saturated. ](FigExampleSkew1.eps "fig:") ![\[fig:simple\] Two examples of simple level-1 networks on $X=\{x_1,\ldots, x_5\}$. Note that both networks are 4-outwards and saturated. ](FigExampleNonSkew1.eps "fig:")
(i) (ii)
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
Counting arcs, vertices and galls {#arcs-vertices-galls}
=================================
In this section, we present some enumerative results concerning the number of vertices, arcs, and galls of a level-1 network. We start by introducing some relevant notation. Suppose $N$ is a phylogenetic network on $X$. Following [@ISS10], we say that a phylogenetic tree $T$ on $X$ is [*displayed*]{} by $N$ if there exists a subgraph $N'$ of $N$ that is a *subdivision* of $T$ For $N$ a level-1 network, we denote the number of galls of $N$ by $g(N)$, that is, we $g(N)=|\mathcal G(N)|$.
Counting arcs and vertices
--------------------------
In case $N$ is a binary level-0 network on $n\geq 2$ leaves, that is, $N$ is a phylogenetic tree on $n$ leaves, it is easy to see that $N$ has $2n-1$ vertices and $2(n-1)$ edges (see e.g.[@SS03 Proposition 2.1.3] for the corresponding result for unrooted binary phylogenetic trees). For the more general case that $N$ is a binary, proper, level-$k$ network on $n\geq 2$ (and thus on $n\geq 3$) leaves, $k\geq 1$, it was shown in [@I09 Lemma 4.5] that any such network can contain at most $2n-1+k(n-1)$ vertices and at most $2n-2+\frac{3}{2}k(n-1)$ arcs. Denoting for $n\geq 3$ the class of all proper level-1 networks in ${{\mathcal L}}_1(n)$ by ${{\mathcal L}}_1(n)^-$, the sizes of the vertex and arc of a network $N=(V,A)$ in $\in{{\mathcal L}}_1(n)^-$ can be at most $3n-2$ and $3.5(n-1)$, respectively. Moreover, if follows from [@I09 Lemma 4.4] that $|V|=2n+1=|A|$ holds in the special case that $N$ is simple. The next result indicates that the size of the vertex set of a simple level-1 network lends itself to providing lower bounds on the sizes of the vertex set and arc set of a general proper level-1 network, respectively.
\[lem:countbounds\] Let $n\geq 3$ and suppose $N = (V,A)\in{{\mathcal L}}_1(n)^-$. Then $2n + 1 \leq |V| \leq 3n-2$ and $2n + 1 \leq |A| \leq 3.5(n-1)$. bounds are tight if $n=3$ in which case $N$ must be a simple level-1 network.
Counting galls
--------------
We next establish a formula for counting the number of galls of a level-1 network. To this end, we require further terminology. Suppose $G=(V,A)$ is a rooted DAG. Then an arc $a \in A$ is called a [*cut arc*]{} of $G$ if the deletion of $a$ disconnects the underlying graph $U(G)$ of $G$. If $a$ is a cut arc of $G$ such that $head(a)$ is a leaf of $G$ then we call $a$ a [*trivial*]{} cut arc of $G$ and a [*non-trivial*]{} cut arc of $G$ otherwise. We denote the number of non-trivial cut arcs of a level-1 network $N$ by $c_N$.
Suppose $N$ is a level-1 network on $X$. For a gall $C$ of $N$, we call an arc of $N$ whose tail but not its head is a vertex of $C$ an [*outgoing arc*]{} of $C$. Note that our assumption implies that every gall must have at least three outgoing arcs. Moreover, if $N$ is binary then we call two distinct leaves $x$ and $y$ of $N$ a [*cherry*]{} of $N$ if $x$ and $y$ have a common parent. Note that that parent must be a vertex of $N$. In addition, if $N$ is a binary phylogenetic tree and $|X|=3$ then $N$ is called a [*triplet (on $X$)*]{}. Saying that a vertex $v$ of a rooted DAG $G$ is [*below*]{} a vertex $w$ of $G$ if $w$ lies on a directed path from the root of $G$ to $v$ but is distinct from $v$, we denote a triplet $t$ on $X=\{a,b,c\}$ for which the last common ancestor of $a$ and $b$ is below the root of $t$ by $ab|c$ (or equivalently $c|ab$). Finally, a collection $\mathcal R$ of triplets is called a [*triplet system (on $\bigcup_{t\in\mathcal R} L(t)$)*]{}.
\[theo:gall-number\] Let $n\geq 2$ and suppose $N\in{{\mathcal L}}_1(n)$. Then $g(N)\leq n - c_N - 2$ and this bound is tight if either $N$ is a phylogenetic tree or every gall of $N$ has exactly three outgoing arcs.
We prove the theorem by induction on $n\geq 2$. Suppose $N\in {{\mathcal L}}_1(n)$. Then the stated inequality clearly holds in the form of an equality for $n=2$ since in that case $N$ is a phylogenetic tree. It also holds for $n=3$ because in that case $N$ is either a triplet and so has one non-trivial cut arc but no gall, or $N$ is a simple level-1 network and so has precisely one gall but no non-trivial cut arcs.
Suppose that $N$ has $n\geq 4$ leaves and assume that $g(N')\leq n-1-c_{N'}-2 $ holds for all level-1 networks $N'\in {{\mathcal L}}_1(n-1)$. Clearly, $g(N)= n-c_N-2 $ holds in case $N$ is a phylogenetic tree as in that case $g(N)=0$ and every non-trivial cut arc of $N$ is an interior edge of $N$ of which there are $n-2$. So assume that $N\in {{\mathcal L}}_1(n)^-$. To see the stated bound for $g(N)$, we distinguish between the cases that (i) $N$ contains a gall $C$ whose outgoing arcs are all trivial cut arcs and (ii) that this is not the case, that is, $N$ contains a cherry.
Assume first that Case (i) holds. We distinguish the cases that $C$ has three outgoing arcs and that $C$ has at least four outgoing arcs. Assume first that $C$ has at least four outgoing arcs. Then there must exist a leaf $a$ of $N$ that is the head of an outgoing arc of $C$ but is not adjacent with the unique hybrid vertex of $C$. Consider the rooted DAG $N'$ obtained from $N$ by first removing $a$, its parent $a'\in V(N)$, and all arcs adjacent with $a'$ and then adding a new arc from the parent of $a'$ to the child of $a'$ contained in $V(C)$. Clearly, $N'$ is a binary level-1 network on $L(N) \setminus \{a\}$ and $g(N)=g(N')$ and $c_N=c_{N'}$ both hold. Since $|L(N')|= n-1$, we have $g(N)=g(N')=n-1 - c_{N'} -2=n-3 - c_N< n-c_N-2$, by induction hypothesis. Consequently, $g(N)<n - c_N -2$ holds in this case.
Next, assume that $C$ has exactly three outgoing arcs $a_1,a_2,a_3$. Let $N'$ be the rooted DAG obtained from $N$ by contracting $C$ as well as $a_1$, $a_2$, and $a_3$ into a new leaf $x$. Clearly, $N'$ is a binary level-1 network on $L(N)\cup\{x\}\setminus\{head(a_1),head(a_2), head(a_3)\}$ and $g(N')=g(N)-1$ and $c_{N'}=c_N-1$.
, assume that Case (ii) holds, that is, $N$ contains two leaves $x$ and $y$ that form a cherry. Let $N'$ denote the rooted DAG obtained from $N$ by first deleting $x$, its parent $p$, and all arcs incident with $p$ and then adding a new arc from the parent of $p$ to $y$. Clearly $N'$ is a binary level-1 network on $L(N)\setminus \{x\}$ and $g(N')=g(N)$ and $c_{N'}=c_N-1$ both hold. Consequently, $g(N)=g(N') \leq n-1 - c_{N'} -2= n-1-(c_N-1)-2=n-c_N-2$ holds by induction hypothesis. This concludes the proof of this case and thus the proof of the stated bound for $g(N)$.
It remains to establish that the stated bound for $g(N)$ is tight for a level-1 network $N\in {{\mathcal L}}_1(n)$ for which all of its galls have precisely three outgoing arcs. To see this one can again perform induction on $n\geq 2$ but this time assuming that $g(N')=n-c_{N'}-2$ holds for all level-1 networks $N'\in{{\mathcal L}}_1(n-1)$ for which every gall has precisely three outgoing arcs. it should be noted that the cases $n\in\{2,3\}$ and $N$ is a phylogenetic tree on $n$ leaves have already been observed above. We leave the details to the interested reader.
Counting clusters and triplets
==============================
In this section we establish enumerative results for computing the sizes of the so-called hardwired and softwired cluster system, respectively, that have both been introduced in the literature for phylogenetic network reconstruction . In addition, we establish that the corresponding result for triplets does not hold. We start with clusters.
Counting clusters {#counting-softwired}
-----------------
We call a non-empty subset of $X$ a [*cluster*]{} and refer to a set of clusters of $X$ as a [*cluster system on $X$*]{}, or just a cluster system if the set $X$ is clear from the context. Suppose for the following that $N$ is a phylogenetic network on $X$ and that $v\in V(N)$. Then we define the cluster $C_N(v)$ associated $v$ to comprise of all leaves of $N$ that are below $v$ and $C_N(v)=\{v\}$ in case $v$ is a leaf of $N$. Again, we simplify our notation by writing $C(v)$ rather than $C_N(v)$ if $N$ is clear from the context. Note that $C(\rho_N)=X$. Then the [*hardwired cluster system*]{} $\mathcal C(N)$ associated $N$ is the cluster system $\{C(v)\,:\, v\in V(N)\}$. Note that if $N\in{{\mathcal L}}_1(X)^-$ then Lemma \[lem:countbounds\] implies that $2n+1\leq |\mathcal C(N)|\leq 3n-2$ and if $N$ is phylogenetic tree then $|\mathcal C(N)|=|V(N)|=2n-1$ where $n$ denotes $|X|$ . Denoting by $\mathcal T(N)$ the set of phylogenetic trees on $X$ displayed by $N$, the [*softwired cluster system*]{} $\mathcal S(N)$ associated $N$ is defined as $\mathcal S(N)=\bigcup_{T\in\mathcal T(N)} \mathcal C(T)$.
To illustrate this definition consider the level-1 network $N_1$ on $X=\{x_1,\ldots,x_5 \}$ depicted in Fig. \[fig:softwired-clusters\](i). Then $\mathcal S(N_1)$ comprises the clusters $ X$, $\{x_2, x_3, x_4, x_5\}$, $\{x_3,x_4,x_5\}$, $\{x_4,x_5\}$, $\{x_2,x_3,x_4\}$, $\{x_3,x_4\}$, $\{x_1\}$, $\{x_2\}$, $\{x_3\}$, $\{x_4\}$, and $\{x_5\}$.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------ -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[fig:softwired-clusters\] Two networks $N_1$ and $N_2$ on $X=\{x_1,\ldots, x_5\}$. Note that both networks are 4-outwards, but neither is saturated.](FigExampleSkew2-label.eps "fig:") ![\[fig:softwired-clusters\] Two networks $N_1$ and $N_2$ on $X=\{x_1,\ldots, x_5\}$. Note that both networks are 4-outwards, but neither is saturated.](FigExampleNonSkew2-label.eps "fig:")
(i) (ii)
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------ -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The next result improves on this $O(n)$ observation by providing a formula for the size of the closely related cluster system $\mathcal S(N)^-:=\mathcal S(N)\setminus\{X\}$. To establish it, we require further terminology.
Suppose $N\in{{\mathcal L}}_1(X)$ and $X' \subseteq X$. Then we define the [*restriction $N|_{X'}$ of $N$ to $X'$*]{} to be the network in ${{\mathcal L}}_1(X')$ obtained from $N$ by applying the following operations in any order until no more can be applied[^2] : (i) suppressing vertices with in-degree and out-degree both equal to one; (ii) deleting vertices with out-degree zero that are not an element in $X$; (iii) collapsing multi-arcs into a single arc; (iv) if a gall $G$ has been created that has exactly two outgoing cut arcs $(u,v)$, $(u',v')$, then deleting these two cut arcs and all the arcs of $G$ and adding arcs $(r, v)$ and $(r, v')$ where $r$ is the unique vertex of $G$ whose children are $u$ and $u'$; (Note that if $N$ is a phylogenetic tree this definition specializes to the usual definition of “restriction” used in the tree literature.) We often write $N|_{X-x}$ as shorthand for $N|_{X - \{x\}}$.
\[theo:countingClusters\] Let $n\geq 2$ and suppose $N\in {{\mathcal L}}_1(n)$. Then $|\mathcal S(N)^-|=3n - 4 - c_N$.
We prove the theorem by induction on $n\geq 2$. Suppose $N\in{{\mathcal L}}_1(n)$. Then the stated equality holds if $n=2$ as then $N$ is a phylogenetic tree on two leaves and if $n=3$ because in that case $N$ is either simple and so $c_N=0$ holds or $N$ is a triplet. In the former case, $|\mathcal T(N)|= 2$ and both phylogenetic trees contained in $\mathcal T(N)$ are triplets. Thus, $|\mathcal S(N)^-|=5=3n - 4 - c_N$ holds in this case. In the latter case, $c_N=1$ follows and thus $|\mathcal S(N)^-|=4=3n - 4 - c_N$ in this case, too.
Now suppose $n>3$ and assume that . $X=L(N)$. We distinguish between the cases that every cut arc of $N$ is trivial and the case that $N$ contains at least one non-trivial cut arc.
Suppose first that every cut arc of $N$ is trivial. Then $c_N = 0$ and $N$ is simple. Note that since $n>3$, at least one of the two directed paths from the root $\rho_N$ to the hybrid vertex $h_N$ of $N$ must contain at least two vertices distinct from $\rho_N$ and $h_N$. Let $P_1$ denote such a path. Moreover, let $v\in V(P_1)$ denote the vertex on $P_1$ that is adjacent with $\rho_N$ and let $x\in X$ denote the leaf of $N$ that is adjacent with $v$. $X'=X-\{x\}$ and $N'=N|_{X'}$. Clearly $N'\in{{\mathcal L}}_1(n-1)$ and $c_{N'}=c_N=0$. Thus, $|\mathcal S(N')^-|=3(n-1) - 4$, by induction hypothesis. Observe that the definition of $S(N)$ implies that $S(N)^{-}$ contains exactly three clusters that $S(N')^{-}$ does not. Indeed, in case the other directed path from $\rho_N$ to $h_N$ also contains vertices distinct from $\rho_N$ and $h_N$, the three clusters missing from $S(N')^{-}$ are $\{x\}$, $C_N(v) \setminus \{h\}$ and $C_N(v)$, where $h$ is the leaf below $h_{N}$. Otherwise, the three clusters missing from $S(N')^{-}$ are $\{x\}$, $C_N(v) \setminus \{h\}$ and $C_{N}(v')$ where $v'$ is the child of $v$ that is not contained in $X$. Consequently, $|\mathcal S(N)^-|=|\mathcal S(N')^-| +3= 3(n-1) - 4-0+3=3n-4-c_N$ holds in this case.
Now suppose that $N$ has a non-trivial cut arc $e=(u,v)$. $Y_1=\{l\in X: l \mbox{ is below } v\}$ and $Y_2=X-Y_1$. Note that $2\leq |Y_1|<n$. Hence, $1\leq |Y_2|\leq n-2$. Consider the rooted DAG $N_1$ with leaf set $Y_1$ obtained from $N$ by deleting all vertices (plus their incident arcs) that are not below $v$ and the rooted DAG $N_2$ on $Y_2\cup\{v\}$ obtained from $N$ by deleting all vertices below $v$ (plus their incident arcs). Since $|Y_1|\geq 2$ it follows that $N_1\in{{\mathcal L}}_1(Y_1)$ and since $|Y_2\cup\{v\}|\geq 2$, we have that $N_2\in{{\mathcal L}}_1(Y_2\cup\{v\})$. Note that a phylogenetic tree $T$ is displayed by $N$ if and only if $T|_{Y_1}$ is displayed by $N_1$ and $T|_{Y_2\cup\{v\}}$ is displayed by $N_2$. Consequently, $\mathcal S(N)^-=\mathcal S(N_1)^-
\stackrel{\cdot}{\cup}\{C\in \mathcal S(N_2)^-\,:\, v\not \in C\}
\stackrel{\cdot}{\cup}\{C-\{v\}\cup Y_1\,:\, v\in C \mbox{ and }
C\in \mathcal S(N_2)^-\}$ must hold. Thus, $$|\mathcal S(N)^-|=|\mathcal S(N_1)^-|+ |\mathcal S(N_2)^-|.$$
Let $i=1,2$ and $n_i=|L(N_i)|$ and $c_i=c_{N_i}$. Then $|\mathcal S(N_i)^-|=3n_i - 4 - c_i$, by induction hypothesis. Consequently, $|\mathcal S(N)^-|=3n_1 - 4 -c_1 + 3n_2 - 4 - c_2$. Since $n_1 + n_2 = n + 1$ and $c_1 + c_2 = c_N - 1$ it follows that $
|\mathcal S(N)^-| =3(n + 1) - 8 - (c_N - 1) = 3n - 4 - c_N,
$ holds in this case, too.
Counting triplets
-----------------
In view of the close relationship between the cluster system $\mathcal C(T)$ induced by a phylogenetic tree $T$ on at least three leaves and the triplet system $\mathcal R(T)$ induced by $T$ (see e.g. [@DHKMS12] or [@vIK11b]) it is reasonable to hope that the companion result to Theorem \[theo:countingClusters\] might hold for the triplet system $\mathcal R(N)$ induced by a phylogenetic network $N$ on at least three leaves. Put differently, it should be possible to express the size of $\mathcal R(N)$ in terms of the number of galls and non-trivial cut arcs of $N$. As the next result , this is not the case. We start with defining the triplet system $\mathcal R(N)$.
Suppose $N\in{{\mathcal L}}_1(X)$ where $|X|\geq 3$ and $a$, $b$, and $c$ are distinct elements in $X$. Then the triplet $ab|c$ is said to be [*consistent*]{} with $N$ if there exist distinct vertices $v$ and $w$ in $N$ and directed paths in $N$ from $v$ to $c$ and $w$, respectively, and from $w$ to $a$ and $b$, respectively, such that any pair of those paths does not have an interior vertex in common. The triplet system $\mathcal R(N)$ is then the set of all triplets $t$ with $L(t)\subseteq X$ that are consistent with $N$.
To illustrate this definition consider the simple level-1 network $N_2$ on $X=\{x_1,\ldots,x_5\}$ depicted in Fig. \[fig:simple\](ii). Then $\mathcal R(N_2)$ comprises the sixteen triplets $x_3|x_1x_2$, $x_4|x_1x_2$, $x_5|x_1x_2$, $x_1|x_3x_4$, $x_4|x_1x_3$, $x_1|x_3x_5$, $x_5|x_1x_3$, $x_1|x_4x_5$, $x_2|x_3x_4$, $x_4|x_2x_3$, $x_2|x_3x_5$, $x_5|x_2x_3$, $x_2|x_4x_5$, $x_3|x_4x_5$, $x_5|x_3x_4$, $x_1|x_2x_3$.
\[prop:triplet-system-size\] For all $n\geq 6$ there exist distinct networks $N_1,N_2\in {{\mathcal L}}_1(n)$ with the same number of galls and non-trivial cut arcs but $|\mathcal R(N_1)|\not=|\mathcal R(N_2)|$.
Let $X'$ denote a finite set of size at least two and let $a$, $b$, $c$, and $d$ denote pairwise distinct elements not contained in $X'$. Consider the binary level-1 networks $N_1$ and $N_2$ on $X:=X'\cup\{a,b,c,d\}$ depicted in Fig. \[FigTripletSystemSize\] where the triangle marked $T$ denotes some binary phylogenetic tree on $X'$. As can be easily checked, $N_1$ and $N_2$ have the same number of leaves and both contain one gall and have $c_{T}+3$ non-trivial cut arcs. Moreover, for any $3$-set $Y\in {X\choose 3}$ there exists exactly one triplet on $Y$ that is contained in ${{\mathcal R}}(N_1)$ except for $Y=\{a,b,c\}$ for which $a|bc, c|ab\in {{\mathcal R}}(N_1)$ holds. Hence, $|{{\mathcal R}}(N_1)|=\binom{n}{3}+1$ where $n=|X|$. Similarly for every 3-subset $Y\in {X\choose 3}$ there exists exactly one triplet on $Y$ that is contained in ${{\mathcal R}}(N_2)$ except for $Y=\{a,c,x\}$ with $x \in X'\cup \{b\}$ for which $a|cx,c|ax\in{{\mathcal R}}(N_2)$ holds. Consequently, $|{{\mathcal R}}(N_2)|=\binom{n}{3}+1+|X'|>\binom{n}{3}+1= |{{\mathcal R}}(N_1)|$.
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Two binary level-1 networks $N_1$ and $N_2$ on $X'\cup\{a,b,c,d\}$ for which the respective number of leaves, galls, and non-trivial cut arcs are the same yet $|{{\mathcal R}}(N_2)|\not=|{{\mathcal R}}(N_1)|$ – see the proof of Proposition \[prop:triplet-system-size\] for details.[]{data-label="FigTripletSystemSize"}](FigTripletSystemSize1-label.eps "fig:") ![Two binary level-1 networks $N_1$ and $N_2$ on $X'\cup\{a,b,c,d\}$ for which the respective number of leaves, galls, and non-trivial cut arcs are the same yet $|{{\mathcal R}}(N_2)|\not=|{{\mathcal R}}(N_1)|$ – see the proof of Proposition \[prop:triplet-system-size\] for details.[]{data-label="FigTripletSystemSize"}](FigTripletSystemSize2-label.eps "fig:")
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Triplet systems and the partition $Cut(N)$ {#sec:triplets-sn}
==========================================
In this section, we start turning our attention to the question of how many triplets suffice to uniquely determine a binary level-1 network. Central to our arguments will be a special type of subsets of $X$ called SN-sets which were originally introduced in [@JNS06] and further studied in, for example, [@IKKSHB09; @vIK11a].
Suppose $|X|\geq 3$ and $\mathcal R$ is a triplet system on $X$. Then a subset $S\subseteq X$ is called an [*SN-set*]{} of $\mathcal R$ if there is no triplet $xy|z\in\mathcal R$ with $x, z \in S$ and $y \not \in S$. In addition, such a set $S$ is called [*non-trivial*]{} if $S\not=X$.[^3] , a SN-set $S$ for $\mathcal R$ is called [*maximal*]{} if there is no non-trivial SN-set that is a strict superset of $S$.
As it turns out, for a binary network $N$ (of any level) the SN-sets of $\mathcal R(N)$ are closely related to the cut arcs of $N$ in the sense that if $(u,v)$ is a cut arc of $N$, then $C_N(v)$ is an SN-set of $\mathcal R(N)$ because there cannot exist a triplet $xy|z \in \mathcal R(N)$ such that $x,z\in C_N(v)$ and $y\not\in C_N(v)$.
To illustrate, consider the network $N_1$ on $X=X'\cup\{a,b,c,d\}$ depicted in Fig. \[FigTripletSystemSize\]. Then $Cut(N_1)$ is the bipartition $\{\{a,b,c\},X' \cup \{d\}\}$.
We begin with an auxiliary observation which relies on the concept of compatibility of pairs of sets, whereby two non-empty finite sets $A$ and $B$ are called [*compatible*]{} if $A\cap B\in \{\emptyset,A,B\}$ holds and [*incompatible*]{} otherwise. More generally, a cluster system $\mathcal C$ is called [*compatible*]{} if any two clusters in $\mathcal C$ are compatible and [*incompatible*]{} otherwise (see e.g.[@SS03 Section 3.5] and [@DHKMS12] for more on such objects).
\[obs:twocut\] Suppose that $n\geq 3$ and that $N$ and $N'$ are two networks in ${{\mathcal L}}_1(n)$ such that $\mathcal R(N) \subseteq \mathcal R(N')$. Let $v\in V(N)$ and $v'\in V(N')$ denote two vertices that are heads of cut arcs of $N$ and $N'$, respectively. Then the induced clusters $C_N(v)$ and $C_{N'}(v')$ are compatible. In particular, if $C_{N}(v) \subsetneq C_{N'}(v')$ then $C_N(v)$ is not a maximal SN-set for $\mathcal R(N)$.
$C=C_N(v)$ and $C'=C_{N'}(v')$. Clearly, if $C=C'$ then $C$ and $C'$ are compatible. So suppose $C\not=C'$. Assume for the sake of contradiction that $C$ and $C'$ are not compatible, that is, $C\cap C'\not\in \{\emptyset,C,C'\}$. Choose elements $x \in C \setminus C'$, $y \in C \cap C'$ and $z \in C' \setminus C$. Then, out of the three possible triplets with leaf set $\{x,y,z\}$, only the triplet $xy|z$ can be contained in $\mathcal R(N)$. Hence, $xy|z\in\mathcal R(N')$ and, so, $C'$ cannot be an SN-set of $\mathcal R(N')$; a contradiction as the incoming arc of $v'$ is a cut arc of $N'$ and, so, $C'$ must be an SN-set of $\mathcal R(N')$. Thus, $C$ and $C'$ must be compatible.
To see the remainder of the observation, assume that $C\subsetneq C'$. Then since $\mathcal R(N) \subseteq \mathcal R(N')$ and $C'$ is an SN-set of $\mathcal R(N')$, it follows that $C'$ is also an SN-set of $\mathcal R(N)$. Since $C$ is also an SN-set of $\mathcal R(N)$, it cannot be a maximal SN-set of $\mathcal R(N)$.
The next result will be required by the induction argument that we will use in the proof of Theorem \[theo:shapely-saturated-triplet\]. relies on the facts that for any saturated network $N\in{{\mathcal L}}_1(X)$ (i) the partition $Cut(N)$ contains at least three elements and (ii) there exists a gall $B$ of $N$ such that the root of $N$ is a vertex of $B$.
-------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------
![The structure of networks $N$ and $N'$ considered in the proof of Proposition \[prop:samecut\].[]{data-label="fig:prop51"}](FigSection5a-label.eps "fig:") ![The structure of networks $N$ and $N'$ considered in the proof of Proposition \[prop:samecut\].[]{data-label="fig:prop51"}](FigSection5b-label.eps "fig:")
(i) (ii)
-------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------
\[prop:samecut\] Suppose that $|X|\geq 3$, that $N$ is saturated network in ${{\mathcal L}}_1(X)$ and that $N'$ is a network in ${{\mathcal L}}_1(X)$ such that $\mathcal R(N) \subseteq \mathcal R(N')$. Then $Cut(N)=Cut(N')$.
\
\
[**]{} Since both $Cut(N)$ and $Cut(N')$ are partitions of $X$, there exists some $C'\in Cut(N')$ distinct from $C$ such that $C\cap C'\not=\emptyset$. Since, in view of [@vIK11a Observation 3] recalled above, $C$ is a maximal SN-set of $\mathcal R(N)$, Thus, there exists a further element $C''\in Cut(N')$ distinct from $C$ and $C'$ such that $C\cap C''\not=\emptyset$ holds, too. Thus, $|C|\geq 2$.\
[**]{} Let $r\in V(N)$ denote the head of the cut arc $(r',r)$ of $N$ for which $C=C_N(r)$ holds and let $B_r$ denote the gall of $N$ that contains $r$ in its underlying cycle (which exists because $|C|\geq 2$ and $N$ is saturated). In view of the usual assumption that no gall in a phylogenetic network has two or fewer outgoing cut arcs, $B_r$ has at least three outgoing cut arcs $c_1, c_2$ and $c_3$ (see Fig. \[fig:prop51\](i)). Let $c_1$ denote the outgoing cut arc of $B_r$ whose tail is the hybrid vertex $h_{B_r}$ of $B_r$. Let $z\in C_N(h_{B_r})$, let $x\in C_N( head(c_2) )$ and let $y\in C_N( head(c_3) )$. Clearly, $\mathcal R(N)$ contains two distinct triplets $t$ and $t'$ on $\{x,y,z\}$.\
[**]{} Since $\mathcal R(N)\subseteq \mathcal R(N')$ we also have $t,t'\in \mathcal R(N')$. Since $Cut(N')$ is the partition of $X$ induced by the maximal SN-sets of $\mathcal R(N')$, it follows that either (i) there exists some element $A\in Cut(N')$ such that $x,y,z\in A$ or (ii) there exist distinct elements $C_x,C_y,C_z\in Cut(N')$ such that $a\in C_a$, for all $a\in \{x,y,z\}$.
Assume first that Case (i) holds. We claim that $C \subseteq A$. To see this, note that we were free to choose any two cut arcs $c_2$ and $c_3$ distinct from $c_1$ and subsequently we had a free choice of $z$, $x$, $y$. For any $Z:=\{x,y,z\}$ chosen this way - let us call this a valid choice - it is straightforward to see that there exist two triplets on $Z$ in $\mathcal{R}(N)$ and thus in $\mathcal{R}(N')$. Since $A$ is an SN-set of $\mathcal{R}(N')$ it follows that as soon as two of the three leaves of a triplet on $Z$ are contained in $A$, so too is the third. Now, let $\{x,y,z\}$ be our initial valid choice, so by assumption $\{x,y,z\} \subseteq A$. Simple case analysis shows that for any element $p\in C$, at least one of $\{x,y,p\}$, $\{x,p,z\}$ or $\{p,y,z\}$ is a valid choice. Hence, $p\in A$ which proves the claim. Since $C \not \in Cut(N')$ we have in fact $C\not=A$. But $C\subsetneq A$ cannot hold either because $C$ is a maximal SN-set for $\mathcal{R}(N)$ and $A$ is a maximal SN-set for $\mathcal{R}(N')$, and by Observation \[obs:twocut\] this cannot happen. Thus, Case (ii) must hold (see Fig. \[fig:prop51\](ii)).\
[**]{} Let $h\in V(N)$ denote the hybrid vertex of the topmost gall $K$ of $N$, that is, the gall of $N$ that contains the root of $N$ in its vertex set (which must exist because $N$ is saturated). Also note that because $C\in Cut(N)$ it follows that $(r',r)$ is a highest cut arc of $N$ and thus $r'$ is a vertex of $K$. Since $|Cut(N)|\geq 3$ there exist distinct elements $C_1,C_2\in Cut(N)-\{C\}$ such that $C_N(h)\in \{C,C_1,C_2\}$. Choose some $p\in C_1$ and some $q\in C_2$. Combined with the definition of $\mathcal R(N)$ it follows that $\mathcal R(N)$ must contain two triplets $t_1$ and $t_2$ on $\{x,p,q\}$, two triplets on $\{y,p,q\}$, and two triplets on $\{z,p,q\}$. Note that since $\mathcal R(N)\subseteq \mathcal R(N')$, those six triplets are also contained in $\mathcal R(N')$.\
[**]{} Using $x$, $y$, $z$, $p$ and $q$, we next analyze the structure of $Cut(N')$ (see Fig. \[fig:prop51\](ii)). Observe first that since $|Cut(N')| \geq 3$, the root of $N'$ must be contained in a gall $B'$ of $N'$. Let $h'\in V(N')$ denote the unique hybrid vertex of $B'$. Let $C_p,C_q\in Cut(N')$ be such that $p\in C_p$ and $q\in C_q$.
We claim that $C_p$ and $C_q$ are distinct elements in $Cut(N')-\{C_x,C_y,C_z\}$. To see this, note first that, since the sets $C_x$, $C_y$ and $C_z$ are pairwise distinct and $t,t'\in \mathcal R(N')$, it follows that one of $x$, $y$, and $z$ must be contained in $C_{N'}(h')$. Without loss of generality, assume that $z\in C_{N'}(h') = C_z$. Note next that $C_p \neq C_q$. Indeed, at least two elements of $\{x,y,z\}$ are not contained in $C_p$, because $C_x, C_y$ and $C_z$ are distinct. Suppose, without loss of generality, $x\not\in C_p$. If $C_p = C_q$, then only the triplet $pq|x$ will be contained in $\mathcal{R}(N')$, contradicting the fact that $t_1$ and $t_2$ are distinct triplets on $\{x,p,q\}$ contained in $\mathcal{R}(N')$. It remains to show that $C_p,C_q\not\in \{C_x,C_y,C_z\}$. Assume for contradiction that $p\in C_x$. Then only $xp|q$ is in $\mathcal{R}(N')$, because $q \not \in C_x$, contradicting the fact that both $t_1$ and $t_2$ are in $\mathcal{R}(N')$. Similarly, if $p$ is in $C_y$, then at most one of the two triplets on $\{y,p,q\}$ are in $\mathcal{R}(N')$, and if $p$ is in $C_z$, at most one of the two triplets on $\{z,p,q\}$ are in $\mathcal{R}(N')$. So $C_p \not \in \{C_x, C_y, C_z\}$. By a symmetrical argument, $C_q \not \in \{C_x, C_y, C_z\}$. This proves the claim.\
\
.
Triplet systems that ${{\mathcal L}}_1(X)$-define {#sec:triplets-define}
=================================================
As is well-known, every binary phylogenetic tree $T$ on $X$ is defined by the triplet set $\mathcal R(T)$ induced by $T$ where a a phylogenetic tree $S$ on $X$ is said to be [*defined*]{} by a triplet system $\mathcal R$ on $X$, if, up to equivalence, $S$ is the unique phylogenetic tree on $X$ for which $\mathcal R\subseteq \mathcal R(S)$ holds (see e.g.[@SS03]). In this context it is important to note that this uniqueness only holds within the space of phylogenetic trees because all networks $N\in{{\mathcal L}}_1(X)$ that display $T$ have the property that $\mathcal R(T) \subseteq \mathcal R(N)$. Combined with the fact that the network $N$ pictured in Fig. \[fig:new\](i) is, up to equivalence, the only binary level-1 network on $X=\{x_1,\ldots, x_4\}$ that is consistent with all five triplets depicted in Fig. \[fig:new\](ii) - a simple case analysis can be applied to verify this - and $|\mathcal R(N)|= 7$, it is natural to ask how many triplets suffice to “uniquely determine” a level-1 network. In this section we provide a partial answer to this question.
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![The binary level-1 $N$ on $X=\{x_1,\ldots, x_4\}$ depicted in (i) is uniquely determined by the five triplets pictured in (ii) but $|\mathcal R(N)|=7$.[]{data-label="fig:new"}](FigIntroLevel1.eps "fig:") ![The binary level-1 $N$ on $X=\{x_1,\ldots, x_4\}$ depicted in (i) is uniquely determined by the five triplets pictured in (ii) but $|\mathcal R(N)|=7$.[]{data-label="fig:new"}](FigIntro5Triplets.eps "fig:")
(i) (ii)
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
More precisely, saying that a network $N\in {{\mathcal L}}_1(X)$ is [*${{\mathcal L}}_1(X)$-defined by a triplet system $\mathcal R$ (on $X$)*]{} if, up to equivalence, $N$ is the unique network in ${{\mathcal L}}_1(X)$ such that $\mathcal R\subseteq \mathcal R(N)$ holds, we show that every 4-outwards network $N$ in ${{\mathcal L}}_1(X)$ that is also simple is ${{\mathcal L}}_1(X)$-defined by a triplet system of size at most $2|X|-1$. In addition, we show that if the requirement that $N$ is simple is replaced by the requirement that $N$ is saturated then $N$ is ${{\mathcal L}}_1(X)$-defined by $\mathcal R(N)$.
.
As the next result not all triplets in $\mathcal R(N)$ are required to ${{\mathcal L}}_1(X)$-define a network $N\in {{\mathcal L}}_1(X)$ in case To simplify its exposition, we say that a triplet system on $X$ [*${{\mathcal L}}_1(X)$-defines*]{} a network $N\in {{\mathcal L}}_1(X)$ if $N$ is [*${{\mathcal L}}_1(X)$-defined*]{} by it.
\[theo:defineSimpleTriplets\] Every simple network in ${{\mathcal L}}_1(X)$ with at least four leaves is ${{\mathcal L}}_1(X)$-defined by a triplet system of size at most $2|X|-1$.
We prove the theorem by induction on $|X|\geq 4$. Suppose $N$ is a simple network in ${{\mathcal L}}_1(X)$ where $n=|X|\geq 4$. $X=\{x_1,\ldots, x_n\}$. Assume without loss of generality that $x_1$ is the head of the outgoing arc of the unique gall $C$ of $N$ starting at the hybrid vertex $v_1$ of $C$. If $n=4$ then a straightforward case analysis implies that $N$ is ${{\mathcal L}}_1(X)$-defined by $\mathcal R(N)$. Note that $|\mathcal R(N)|=7=2n-1$ holds in this case.
Now assume that $n\geq 5$ and that for every set $Y$ with $4 \leq |Y|\leq n-1$ and every simple network $N'\in{{\mathcal L}}_1(Y)$ there exists a triplet system $\mathcal R$ of size at most $2|Y|-1$ such that $N'$ is ${{\mathcal L}}_1(Y)$-defined by it. Starting at $v_1$ and traversing the unique cycle $C$ in the underlying graph $U(N)$ of $N$ counter-clockwise let $v_1,v_2,\ldots, v_{i-1},v_i=\rho_N,
v_{i+1},\ldots, v_{n+1},v_1$ denote a circular ordering of the vertices of $C$. Without loss of generality assume that for all $2\leq j\leq i-1$ the head of the outgoing arc of $C$ starting at $v_i$ is $x_i$ and that for all $i+1\leq j\leq n+1$ the head of the outgoing arc of $C$ starting at $v_j$ is $x_{j-1}$. $X'=X-\{x_n\}$. We distinguish between the cases that (i) the root $\rho_N$ of $N$ equals $v_{n+1}$ and (ii) that this is not the case.
Case (i): Assume that $\rho_N=v_{n+1}$ and $N'=N|_{X'}$. Since $N'$ is clearly simple and $4\leq |L(N')|= n-1$ it follows by induction hypothesis that there exists a triplet system $\mathcal R'$ on $X'$ of size at most $2(n-1)-1$ such that $N'$ is ${{\mathcal L}}_1(X')$-defined by $\mathcal R'$. $t_1=x_1|x_{n-1}x_n$ and $t_2=x_n|x_1x_{n-1}$. We claim that $N$ is ${{\mathcal L}}_1(X)$-defined by $\mathcal R=\mathcal R'\cup\{t_1,t_2\}$. To see this, assume that $N_1$ is a network in ${{\mathcal L}}_1(X)$ for which $\mathcal R\subseteq \mathcal R(N_1)$ holds. We need to show that $N$ and $N_1$ are equivalent.
$N_1'=N_1|_{X'}$. By construction, $\mathcal R'\subseteq \mathcal R(N_1')$. Since $N'$ is ${{\mathcal L}}_1(X')$-defined by $\mathcal R'$ it follows that $N'$ and $N_1'$ must be equivalent. Consequently $N_1'$ must also be a simple network in ${{\mathcal L}}_1(X')$. Combined with the fact that $t_1,t_2\in\mathcal R\subseteq \mathcal R(N_1)$ it follows that $N$ and $N_1$ must be equivalent.
Case (ii) Assume that $\rho_N\not= v_{n+1}$. Then $i\in \{2,\ldots,n\}$. We distinguish between the cases that $i=n$, that is, $\rho_N=v_n$ and that $i\in \{2,\ldots,v_{n-1}\}$. In the former case the proof of the induction step is similar to the previous case but with $t_1$ replaced by $x_{n-1}|x_nx_1$. In the latter case the proof of that step is also similar to the previous case but this time with $t_2$ replaced by $x_{n-1}|x_nx_1$.
Combined with the definition of ${{\mathcal L}}_1(X)$-defining triplet systems, Theorem \[theo:defineSimpleTriplets\] immediately implies:
\[cor:defineSimpleTriplets\] Every simple network in ${{\mathcal L}}_1(X)$ with at least four leaves is ${{\mathcal L}}_1(X)$-defined by its induced triplet system.
An obvious problem with extending Corollary \[cor:defineSimpleTriplets\] to general networks in ${{\mathcal L}}_1(X)$ is that 4-outwards networks can have tree-like regions. a 4-outwards network $N$ contains a directed path of length 3 or more consisting solely of cut arcs. can then transform it into a new network $N'$ in ${{\mathcal L}}_1(X)$ for which $\mathcal R(N)\subseteq \mathcal R(N')$ holds by subdividing the first and last cut arc of that path by new vertices $u$ and $v$, respectively, and adding a new arc $(u,v)$. There are however more subtle situations possible which do not require adding vertices and arcs. Consider, for example, the two networks $N$ and $N'$ on $X=\{x_1,\ldots, x_5\}$ presented in Figures \[fig:simple\](ii) and \[fig:softwired-clusters\](ii), respectively. Then $\mathcal R(N')\subseteq \mathcal R(N)$ holds but $N$ and $N'$ are clearly not equivalent. Thus $N'$ is not ${{\mathcal L}}_1(X)$-defined by $\mathcal R(N')$ (although $N'$ is clearly encoded by $\mathcal R(N')$ as it is $4$-outwards [@GH12]). We therefore next turn our attention to
To establish our next result (Theorem \[theo:shapely-saturated-triplet\]), we require a construction from [@vIK11a] that allows us to associate a level-1 network $Collapse(N)$ to any level-1 network $N$ such that $Collapse(N)$ is either simple, or is a phylogenetic tree on two leaves. We next review this construction for networks in ${{\mathcal L}}_1(X)$.
. For each element $C\in Cut(N)$ choose some element $c_C\in C$ and $X^*=\{c_C\,:\,C\in Cut(N)\}$. Note that $|X^*|\geq 2$, but if $N$ is saturated $|X^*|\geq 3$ and if $N$ is saturated and 4-outwards $|X^*|\geq 4$. Clearly, if $|C_N(v)|\geq 2$ then $N_{x_v}$ is contained in ${{\mathcal L}}_1(C_N(v))$ and is an isolated vertex otherwise. That $Collapse(N)$ is a simple network in ${{\mathcal L}}_1(X^*)$ or a phylogenetic tree on two leaves is clear. Let $\mathcal R_{Collapse(N)}$ denote the triplet system on $X^*$ comprising all triplets $x_w | x_u x_v$ for which there $x_1 \in C_N(w)$, $x_2 \in C_N(u)$ and $x_3 \in C_N(v)$ such that $x_1|x_2x_3\in\mathcal R(N)$. It is straightforward to see that $\mathcal R(Collapse(N)) = \mathcal R_{Collapse(N)}$.
\[theo:shapely-saturated-triplet\] Every 4-outwards network in ${{\mathcal L}}_1(X)$ that is also saturated is ${{\mathcal L}}_1(X)$-defined by its induced triplet system.
We prove the theorem by induction on the number $g(N)$ of galls in a saturated, 4-outwards network $N\in{{\mathcal L}}_1(X)$. Suppose $N$ is such a network. $g=g(N)$. Then since $|X|\geq 2$ and $N$ is saturated we have $g\geq 1$. Hence, $|X|\geq 3$. In case $g=1$ the assumption that $N$ is saturated implies that $N$ is simple, and thus $|X| \geq 4$ because $N$ is 4-outwards. Corollary \[cor:defineSimpleTriplets\], $N$ is ${{\mathcal L}}_1(X)$-defined by $\mathcal R(N)$.
So assume that $g\geq 2$ and that every saturated, 4-outwards network $N\in {{\mathcal L}}_1(Y)$ with $g-1$ galls is ${{\mathcal L}}_1(Y)$-defined by a triplet system on $Y$, where $Y$ is a finite set of size at least two. Let $N'\in{{\mathcal L}}_1(X)$ denote a network for which $\mathcal R(N) \subseteq \mathcal R(N')$ holds. We need to show that $N$ and $N'$ are equivalent. To see this, we first analyze the networks $Collapse(N)$ and $Collapse(N')$.
Note first that, by Proposition \[prop:samecut\], $Cut(N') = Cut(N)$. Hence, we may assume without loss of generality that $X^*$ is the leaf set of both $Collapse(N)$ and $Collapse(N')$. Next note that $Collapse(N)$ is 4-outwards because $N$ is 4-outwards and $|Cut(N)|\geq 4$. Since a simple level-1 network is in particular saturated and $Collapse(N)$ has precisely one gall, the base case of the induction implies that $Collapse(N)$ is ${{\mathcal L}}(X^*)$-defined by $\mathcal R_1:=\mathcal R(Collapse(N))$. Since with $\mathcal R_2:=\mathcal R(Collapse(N'))$ we have $\mathcal R_{Collapse(N)}=\mathcal R_1
\subseteq \mathcal R_2=
\mathcal R_{Collapse(N')}$ and so $\mathcal R_1\subseteq \mathcal R_2$ holds it follows that $Collapse(N)$ and $Collapse (N')$ must be equivalent.
We next analyze the level-1 networks $N_v$ of $N$ with $v\in X^*$. Let $v\in X^*$ and let $C\in Cut(N)$ be such that $v\in C$. Note that if $|C|=1$ then $N_v$ is an isolated vertex and thus a rooted DAG with leaf set $\{v\}$. So assume that $|C|\geq 2$. Then since $N$ is a saturated, 4-outwards network in ${{\mathcal L}}_1(X)$, $N_v$ is a saturated, 4-outwards network in ${{\mathcal L}}(C)$. Since $N_v$ has at most $g-1$ galls the induction hypothesis implies that $N_v$ is ${{\mathcal L}}(C)$-defined by $\mathcal R(N_v)$. By assumption, $\mathcal R(N) \subseteq \mathcal R(N')$ and so $\mathcal R(N_v) \subseteq \mathcal R(N'_v)$. Thus $N'_v$ and $N_v$ must be equivalent. Combined with the observation that the networks $Collapse(N)$ and $Collapse (N')$ are equivalent it follows that $N$ and $N'$ are equivalent.
${{\mathcal L}}_1(X)$-defining cluster systems {#sec:define-clusters}
===============================================
In this section, we turn our attention to the companion question of Section \[sec:triplets-define\]. That is, whether some (not necessarily proper) subset of $\mathcal S(N)$ is sufficient to “uniquely determine” a 4-outwards network $N$ in ${{\mathcal L}}_1(X)$. We first present a formalization of the idea of “uniquely determining” to being ${{\mathcal L}}_1(X)$-defined for cluster systems. Subsequent to this, we then show that all 4-outwards networks $N\in{{\mathcal L}}_1(X)$ that are also simple are ${{\mathcal L}}_1(X)$-defined by a cluster system of size at most $|X|$ (Theorem \[theo:simple-clusters\] and Corollary \[cor:defineSimple\]). Replacing the requirement that $N$ is simple by the more general requirement that $N$ is saturated we also show that such networks are ${{\mathcal L}}_1(X)$-defined by their induced softwired cluster system (Theorem \[theo:shapely-saturated-cluster\]).
Let $N$ denote a phylogenetic network on $X$ and let $\mathcal S$ denote a cluster system on $X$. Then we say that $N$ [*displays $\mathcal S$ (in the softwired sense)*]{} if $\mathcal S\subseteq \mathcal S(N)$ holds. Furthermore, we say that a network $N\in {{\mathcal L}}_1(X)$ is [*${{\mathcal L}}_1(X)$-defined by a cluster system $\mathcal S$ on $X$*]{} if, up to equivalence, $N$ is the unique network in ${{\mathcal L}}_1(X)$ that displays $\mathcal S$. It should be noted that, as in the case of triplet systems, a binary phylogenetic tree $T$ on $X$ is not ${{\mathcal L}}_1(X)$-defined by its induced cluster system $\mathcal C(T)=\mathcal S(T)$. The reason is again that, by subdividing arcs of $T$ and adding new arcs joining the subdivision vertices, we can transform $T$ into a network $N$ in ${{\mathcal L}}_1(X)$ for which $\mathcal C(T)\subseteq \mathcal S(N)$ holds. Also it should be noted that a network in ${{\mathcal L}}_1(X)$ is not ${{\mathcal L}}_1(X)$-defined by its induced hardwired cluster system. Analogous to the triplet result presented in Section \[sec:triplets-define\], a 4-outwards networks $N\in {{\mathcal L}}_1(X)$ also need not be ${{\mathcal L}}_1(X)$-defined by $\mathcal S(N)$. Indeed, consider again the two 4-outwards networks $N_1$ and $N_2$ on $X=\{x_1,\ldots, x_5\}$ presented in Figures \[fig:softwired-clusters\](i) and \[fig:simple\](i), respectively. Then $N_1$ and $N_2$ are clearly not equivalent but $\mathcal S(N_1) \subseteq
\mathcal S(N_2)$.
\[theo:simple-clusters\] Let $X=\{x_1,\ldots, x_n\}$, $n\geq 4$, and suppose that $N$ is a simple network in ${{\mathcal L}}_1(X)$ such that, when starting at the hybrid vertex $v_1$ of $N$ and traversing the unique cycle $C$ of $U(N)$ counter-clockwise, the obtained vertex ordering for $C$ is $v_1, v_2, \ldots, v_{i-1}, v_i=\rho_N, v_{i+1},\ldots, v_{n+1}, v_1$ and $x_j$ is a child of $v_j$, for all $1\leq j\leq i-1$, and $x_j$ is a child of $v_{j+1}$, for all $i\leq j \leq n$. $N$ is ${{\mathcal L}}_1(X)$-defined by the cluster system $\mathcal S_d(X)$ where
1. $\mathcal S_d(N):=\bigcup_{2\leq j\leq n-1}\{\{x_1,x_2,\ldots,x_j\}\}
\cup\{ \{x_2,x_3,\ldots, x_n\}\}$ if $\rho_N=v_{n+1}$ .
2. $\mathcal S_d(N):=
\bigcup_{3\leq j\leq n-1}\{\{x_2,x_3\ldots, x_j\}\}
\cup \{\{x_1,x_2\},\{x_1,x_n\},\{x_1,x_2,x_3\}\}$ if $\rho_N=v_n$.
3. $$\begin{aligned}
\mathcal S_d(N)&:=&\bigcup_{3\leq j\leq i-1}\{\{x_2,x_3\ldots, x_j\}\}\cup
\bigcup_{n-1 \geq j\geq i}\{\{x_n,x_{n-1},\ldots, x_j\}\}\\
&\cup&
\{\{x_1,x_2\},\{x_1,x_n\},\{x_1,x_n,x_{n-1}\}\}\end{aligned}$$
if $\rho_N\not\in\{v_n,v_{n+1}\}$.
Let $N_1 \in \mathcal{L}_1(X)$ be a network such that $\mathcal{S}_d(N) \subseteq \mathcal{S}(N_1)$. We first claim that $N_1$ must be simple. Assume for the sake of contradiction that $N_1$ is not simple, that is, $N_1$ contains a non-trivial cut arc $(u,v)$. Then every cluster in $\mathcal{S}(N_1)$ must be compatible with $C= C_{N_1}(v)$, $2 \leq |C| < n$, and $C\in \mathcal{S}(N_1)$. We will derive a contradiction by showing that $\mathcal{S}_d(N)$, and thus also $\mathcal{S}(N_1)$, contains at least one cluster that is incompatible with $C$.
Case (i). We distinguish between the two alternatives that $x_1 \in C$ and that $x_1 \not\in C$. Assume first that $x_1 \in C$. Then since $2 \leq |C| < n$ we have for $C':=\{x_2, \ldots, x_n\}\in \mathcal{S}_d(N)$ that $C' \cap C\not=\emptyset$ and that $C' \cap (X\backslash C) \not=\emptyset$, that is, $C'\not\subseteq C$. Since $x_1\in C$ it follows that $C\subseteq C'$ cannot hold either and so $C$ and $C'$ are incompatible, as required. Now, suppose $x_1 \not \in C$. Then since $2 \leq |C|$ there exist $p,q\in\{2,\ldots, n\}$ with $p<q$, say, such that $x_p, x_q \in C$. Clearly, $x_q\not\in C':= \{x_1, \ldots, x_p\}\in \mathcal{S}_d(N)$. But then $C'$ and $C$ are again incompatible, as required.
A similar analysis holds for cases (ii) and (iii); we leave the details to the interested reader. Hence, $N_1$ must be simple, as claimed.
Let $h$ denote the unique hybrid vertex of $N_1$ and let $x$ denote the leaf of $N_1$ that is incident with $h$. For the remainder of the proof, we consider each of the three cases stated in the theorem separately. All three cases use the following observations: (a) if $N_1 |_{X-x}$ is a tree, then all clusters in $\mathcal{S}( N_1 |_{X-x} )$ are pairwise compatible; (b) If $\mathcal{S}_d(N)\subseteq \mathcal{S}(N_1)$, then $\mathcal{S}_d(N)|_{X-x}\subseteq \mathcal{S}(N_1)|_{X-x}$ where for any cluster system $\mathcal C$ of $X$ we $\mathcal C|_{X-x} = \{ C \setminus \{x\}\, :\,C \in \mathcal C\}$; (c) $\mathcal{S}(N_1 |_{X-x}) = \mathcal{S}(N_1)|_{X-x}$. For ease of presentation we will liberally make use of the assumption that $\mathcal{S}_d(N) \subseteq \mathcal{S}(N_1)$ without explicitly stating it.
*Case (i)*. First, we argue that $x \in \{x_1, x_2\}$. Assume for the sake of contradiction that $x \not\in \{x_1, x_2\}$. Then $C = \{x_1, x_2\}$ and $C' = \{x_2, \ldots, x_n\} \setminus \{x\}$ are incompatible and clearly contained in $\mathcal{S}_d(N)|_{X-x}\subseteq \mathcal{S}(N_1)|_{X-x}$. Hence, $\mathcal{S}(N_1)|_{X-x}$ is not compatible which is impossible because $x$ is incident with $h$ and so $N_1 |_{X-x}$ is a phylogenetic tree. So $x \in \{x_1, x_2\}$. In fact, similar reasoning implies that $x = x_2$ is also impossible as otherwise $\mathcal{S}_d(N)|_{X-x}$ would contain incompatible clusters $\{x_1, x_3\}$ and $\{x_3, \ldots, x_n\}$. So $x=x_1$. Since $\{x_1, x_2\} \subseteq \mathcal{S}_d(N)$ it follows that the other child of the parent of $x_2$ in $N_1$ is $h$. Combined with the fact that $\bigcup_{2 \leq j \leq n-1} \{ \{x_1, x_2, \ldots, x_j\}\} \subseteq \mathcal{S}_d(N)$ it follows that the other child of the parent of $x_k$ in $N_1$ is the parent of $x_{k-1}$, $3 \leq k \leq n-1$. Since $\{x_2, \ldots, x_n\} \in \mathcal{S}_d(N)$ it follows that the other child of the parent of $x_n$ in $N_1$ is the parent of $x_{n-1}$. Hence $N_1$ is equivalent to $N$.
*Case (ii)*. We claim that $x \in \{x_1, x_2, x_n\}$. The argument is similar to case (i) in that if $x \not \in \{x_1, x_2, x_n\}$ then $\mathcal{S}_d(N)|_{X-x}\subseteq \mathcal{S}(N_1)|_{X-x}$ contains incompatible clusters $\{x_1, x_2\}$ and $\{x_1, x_n\}$, leading to a contradiction of the fact that $N_1 |_{X-x}$ is a phylogenetic tree. In fact, similar arguments utilizing the facts that $\{x_1, x_2, x_3 \} \in \mathcal{S}_d(N)$ and that $n \neq 3$ imply that $x \neq x_2$ and $x \neq x_n$. So again $x = x_1$. Since $\{x_1, x_2\}$ and $\{x_1, x_n\}$ are contained in $\mathcal S_d(N)\subseteq \mathcal S(N_1)$ it follows that the other child of the parents of $x_2$ and $x_n$ in $N_1$, respectively, is $h$. In view of $\{x_1,x_2,x_3\} \subseteq\mathcal S_d(N)
\subseteq \mathcal S(N_1)$ we see that the other child of the parent of $x_3$ in $N_1$ must be the parent of $x_2$. Since $\bigcup_{3\leq j\leq n-1}\{\{x_2,x_3\ldots, x_j\}\}
\subseteq\mathcal S_d(N) $ similar arguments as in the previous case imply that $N$ and $N_1$ must be equivalent. *Case (iii)* Again the fact that $N_1 |_{X-x}$ is a phylogenetic tree implies that $x \in \{x_1, x_2, x_n\}$. However, $x = x_n$ cannot hold because $n-1 \neq 2$ and so $\{x_1, x_2\}$ and $\{x_1, x_{n-1}\}$ are distinct clusters that are both contained in $\mathcal{S}_d(N)|_{X-x}$ and thus in $ \mathcal{S}(N_1)|_{X-x}$. But then $ \mathcal{S}(N_1)|_{X-x}$ is incompatible which is impossible as $N_1 |_{X-x}$ is a phylogenetic tree. Similarly, $x \not= x_2$ as otherwise the two incompatible clusters $\{x_1, x_n\}$ and $\{x_n, x_{n-1}, \ldots, x_i\}$ are contained in $\mathcal{S}_d(N)|_{X-x}$. So $x = x_1$. Focussing as in case (ii) on $x_2$ and $x_n$ we see again that the common child of their respective parents is $h$. Since $\bigcup_{3\leq j\leq i-1}\{\{x_2,x_3\ldots, x_j\}\}\cup
\bigcup_{n-1 \geq j\geq i}\{\{x_n,x_{n-1},\ldots, x_j\}\}
\subseteq\mathcal S_d(N) $ the location of the remaining leaves of $N_1$ is forced. Hence, $N_1$ is equivalent to $N$.
As an immediate consequence of Theorem \[theo:simple-clusters\], we obtain the companion result for Theorem \[theo:defineSimpleTriplets\].
\[cor:defineSimple\] Every simple network in ${{\mathcal L}}_1(X)$ with at least four leaves is ${{\mathcal L}}_1(X)$-defined by a cluster system of size at most $|X|$.
We now prove the cluster equivalent of Theorem \[theo:shapely-saturated-triplet\] i.e.that requiring that a 4-outwards network in ${{\mathcal L}}_1(X)$ is also saturated guarantees that it is uniquely determined by its induced softwired cluster system.
\[theo:shapely-saturated-cluster\] Every 4-outwards network in ${{\mathcal L}}_1(X)$ that is also saturated is ${{\mathcal L}}_1(X)$-defined by its induced softwired cluster system.
Let $N$ and $N'$ be networks in ${{\mathcal L}}_1(X)$ such that $N$ is 4-outwards and saturated and $\mathcal S(N) \subseteq \mathcal S(N')$ holds. We need to show that $N'$ is equivalent with $N$. $\mathcal T = \mathcal T(N)$. Clearly, $\bigcup_{T \in \mathcal T} \mathcal S(T)
= \mathcal S(N)\subseteq \mathcal S(N')$. Combined with [@vIK11b Proposition 1] which implies that $\mathcal R(\mathcal T) \subseteq \mathcal R(N')$ and the fact that $ \mathcal R(N) = \mathcal R(\mathcal{T})$ it follows that $ \mathcal R(N) \subseteq \mathcal R(N')$. Since, by Theorem \[theo:shapely-saturated-triplet\], $N$ is ${{\mathcal L}}_1(X)$-defined by $\mathcal R(N)$ it follows that $N$ and $N'$ are equivalent.
In fact, due to the very general character of [@vIK11b Proposition 1], Theorem \[theo:shapely-saturated-cluster\] can easily be extended to prove that, whenever $\mathcal R(N)$ has been proven sufficient to uniquely determine (in our sense) a specified subfamily - any subfamily - of phylogenetic networks $N$, so too is $\mathcal S(N)$ where we canonically extend the notions of an induced triplet system and softwired cluster system to such networks.
Conclusions
===========
In this paper, we have presented enumerative results concerning the number of vertices, arcs, and galls of a binary level-1 network. By focusing on triplet systems and (softwired) cluster systems we have also investigated the question if subsets of those systems suffice to uniquely determine the binary level-1 network that induced them. As part of this, we have presented examples that illustrate that a level-1 network need not be uniquely determined by the triplet/cluster system it induces thus illustrating the difference between the notion of encoding and our formalization of uniquely determining. In addition, we have provided bounds on the size of such a system in case the network in question is simple and has at least four leaves. For the more general class of $4$-outwards, saturated, binary level-1 networks we have shown that any network in that class is uniquely determined by the triplet/softwired cluster system it induces. However, a number of open questions remain. For example for which binary level-1 networks are the aforementioned bounds sharp and are 4-outwards saturated binary level-1 networks characterizable by the fact that they are uniquely determined by their induced triplet/softwired cluster system?
We conclude with remarking that in [@HM13] [*trinets*]{}, that is, rooted directed acyclic graphs on just three leaves have recently been introduced in the literature for phylogenetic network reconstruction. In that paper it was also shown that any level-1 network is encoded by the trinet system that it induces. In addition, it was shown in [@vIM12] that the more general tree-sibling and level-2 networks are encoded by their induced trinet systems, a fact that is not shared in general for the triplet system or the softwired cluster system induced by such networks. Formalizing the idea of “uniquely determining” for trinet systems in a canonical way to [*${{\mathcal L}}_1(X)$-defining trinet systems*]{} it might be interesting to explore what kind of trinet systems ${{\mathcal L}}_1(X)$-define such networks.
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors thank the referee for a very careful reading of the paper. KTH and PG thank the London Mathematical Society (LMS) for its support in the context of its Computer Science Small Grant Scheme.
[^1]: School of Computing Sciences, University of East Anglia, UK,\
LIGM, Universit[é]{} Marne-la-Vall[é]{}e, France,\
Department of Knowledge Engineering (DKE), Maastricht University, The Netherlands
[^2]:
[^3]:
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is a classical result that Lebesgue measure on the unit circle is invariant under inner functions fixing the origin. In this setting, the distortion of Hausdorff contents has also been studied. We present here similar results focusing on inner functions with fixed points on the unit circle. In particular, our results yield information not only on the size of preimages of sets under inner functions, but also on their distribution with respect to a given boundary point. As an application, we use them to estimate the size of irregular points of inner functions omitting large sets. Finally, we also present a natural interpretation of the results in the upper half plane.'
author:
- 'Matteo Levi[^1], Artur Nicolau and Odí Soler i Gibert[^2]'
bibliography:
- 'bibliography.bib'
title: Distortion and Distribution of Sets under Inner Functions
---
***Keywords—*** Inner functions, boundary fixed points, angular derivatives, Hausdorff contents.
Introduction
============
Let ${\mathbb{D}}$ be the open unit disc of the complex plane. An analytic mapping $f\colon {\mathbb{D}}\rightarrow {\mathbb{D}}$ is called *inner* if $ \left|\lim_{r \to 1} f(r\xi)\right| = 1$ for almost every point (a.e.) $\xi$ of the unit circle ${\partial\mathbb{D}}$. Hence, an inner function $f$ induces a map defined at almost every point $\xi \in {\partial\mathbb{D}}$ by $f^\ast(\xi)= \lim_{r \to 1} f(r\xi)$, which we will denote by $f$ as well. This induced map lacks the regularity of the inner function itself and it is actually discontinuous at every point $\xi \in {\partial\mathbb{D}}$ where $f$ does not extend analytically. More concretely, fixed $\xi \in {\partial\mathbb{D}}$ where $f$ does not extend analytically and $\eta \in {\partial\mathbb{D}}$ there exists a sequence $\xi_n\to \xi$ such that $f(\xi_n) \to \eta$ (see page 77 of [@ref:GarnettBoundedAnalyticFunctions], and page 4 of [@ref:NoshiroClusterSets]). We are interested in studying certain invariance and distortion properties of measures and Hausdorff contents of sets in the unit circle under the action of inner functions.
Let $f\colon {\mathbb{D}}\rightarrow {\mathbb{D}}$ be an analytic mapping. We say that a point $p \in {\partial\mathbb{D}}$ is a *boundary Fatou point* of $f$ if $f(p)=\lim_{r \to 1} f(rp)$ exists and $f(p) \in {\partial\mathbb{D}}$. Hence, the set of boundary Fatou points of an inner function has full measure. For $0< \beta < 1$ and $p \in {\partial\mathbb{D}}$, let $\Gamma_{\beta} (p) = \{z \in {\mathbb{D}}: |z-p| < \beta (1-|z|) \}$ be the Stolz angle with opening $\beta$ and vertex at $p$. A holomorphic self map $f$ of the unit disc has finite angular derivative at $p \in {\partial\mathbb{D}}$ if there is a point $\eta \in {\partial\mathbb{D}}$ and $\beta > 0$ such that the non-tangential limit $$f'(p) \coloneqq \lim_{\Gamma_{\beta} (p) \ni z \to p} \frac{\eta-f(z)}{p-z}$$ exists and is finite. Observe that in this case $ \eta = f(p).$ We set $|f'(p)|=+\infty$ if the function $f$ does not have a finite angular derivative at the point $p \in {\partial\mathbb{D}}$. Observe that this is the case if $p$ is not a boundary Fatou point of $f$. With this convention, for any $p \in {\partial\mathbb{D}}$, the classical Julia-Carathéodory theorem gives $$\label{eq:ModulusDerivative}
\liminf_{z \to p} \frac{1-|f(z)|}{1-|z|} = |f'(p)| > 0,$$ in the sense that either the $\liminf$ is finite and equal to $|f'(p)|>0$ or both quantities are infinite. See for example Chapters IV and V of [@ref:ShapiroCompositionOperators].
We denote by $\lambda$ the normalized Lebesgue measure on ${\partial\mathbb{D}}$ and by $\lambda_z$ the harmonic measure from the point $z \in {\mathbb{D}},$ given by $$\lambda_z(E) = \int_E \frac{1-|z|^2}{|\xi - z|^2} \, d\lambda(\xi),$$ for any measurable set $E \subseteq {\partial\mathbb{D}}$. A classical result due to Löwner (see, for instance, page 12 of [@ref:Ahlfors]) says that Lebesgue measure is invariant under the action of any inner function fixing the origin. Hence, the following conformally invariant version of Löwner’s Lemma holds.
\[thm:LownerLemma\] Let $f:{\mathbb{D}}\to {\mathbb{D}}$ be an inner function and $z \in {\mathbb{D}}.$ Then, $$\lambda_z(f^{-1}(E)) = \lambda_{f(z)}(E)$$ for any measurable set $E \subseteq {\partial\mathbb{D}}.$
Observe that, if $z \in {\mathbb{D}}$ is a fixed point of $f$, Theorem \[thm:LownerLemma\] says that $\lambda_z$ is invariant under the action of $f.$ However, it may be the case that $f$ has no fixed points in ${\mathbb{D}}$ but only on ${\partial\mathbb{D}}.$ A point $p \in {\partial\mathbb{D}}$ is a fixed point for $f$ if $\lim_{r \to 1} f(rp) = p.$ Actually, the classical Denjoy-Wolff Theorem states that for any analytic self mapping $f$ on the unit disc which is not an elliptic automorphism, there exists a fixed point $p \in {\overline{\mathbb{D}}}$ of $f,$ called the Denjoy-Wolff fixed point of $f,$ such that the iterates $f^n = f \circ \overset{n)}{\ldots} \circ f$ tend to $p$ uniformly on compacts sets of ${\mathbb{D}}.$ Moreover, $p$ is the unique fixed point of $f$ in ${\overline{\mathbb{D}}}$ such that $0 < |f'(p)| \leq 1.$ See for example Chapter V of [@ref:ShapiroCompositionOperators]. We are interested in analogues of Theorem \[thm:LownerLemma\] when $z \in {\partial\mathbb{D}}.$ This situation occurs naturally when the Denjoy-Wolff fixed point of $f$ is on the unit circle. In this situation, instead of considering the harmonic measure from a point in the open unit disc, it is natural to measure sets with respect to boundary points. We will consider a measure introduced by Doering and Mañé in [@ref:DoeringMane]. Fix a point $p \in {\overline{\mathbb{D}}}$ and consider the positive measure $\mu_p$ on ${\partial\mathbb{D}}$ defined by $$\mu_p(E) = \int_E \frac{1}{|\xi -p|^2} \, d\lambda(\xi)$$ for any measurable set $E \subseteq {\partial\mathbb{D}}.$ Observe that for a point $p \in {\partial\mathbb{D}}$ the measure $\mu_p$ is not finite, while for $p \in {\mathbb{D}},$ it is just a scalar multiple of the harmonic measure given by $\mu_p = (1-|p|^2)^{-1} \lambda_p.$ A very natural interpretation of the measure $\mu_p$ when $p \in {\partial\mathbb{D}}$ is the following. Let $\omega_p \colon {\mathbb{D}}\rightarrow {\mathbb{H}}$ be the conformal map from the disc into the upper half-plane ${\mathbb{H}}$ such that $\omega_p(p) = \infty$ and $\omega_p(0) = i/2.$ Then, for any measurable set $E \subseteq {\partial\mathbb{D}},$ we have that $\mu_p(E) = |\omega_p(E)|,$ where we denote by $|A|$ the Lebesgue measure of a set $A \subseteq {\mathbb{R}}.$ Roughly speaking, for a point $p \in {\partial\mathbb{D}},$ the measure $\mu_p$ gives information about the size and the distribution of a set around the point $p.$ Sets having large $\mu_p$ measure are those that are highly concentrated around the point $p.$ In particular, if $E$ is an open neighbourhood of $p,$ then $\mu_p(E) = \infty.$ Our first result is the following analogue of Theorem \[thm:LownerLemma\].
\[thm:LownerMuP\] Let $f\colon {\mathbb{D}}\to {\mathbb{D}}$ be an inner function and let $p \in {\partial\mathbb{D}}$ be a boundary Fatou point of $f.$
(a) \[stm:LownerMuPFiniteDerivative\] Assume $|f'(p)| < \infty.$ Then $$\mu_p(f^{-1}(E)) = |f'(p)|\mu_{f(p)}(E)$$ for any measurable set $E \subseteq {\partial\mathbb{D}}.$
(b) If $|f'(p)| = \infty$ and $E \subseteq {\mathbb{D}}$ is a measurable set, then $\mu_p(f^{-1}(E)) = \infty$ if $\mu_{f(p)}(E) > 0$ and $\mu_p(f^{-1}(E)) = 0$ if $\mu_{f(p)}(E) = 0.$
As we can see, we still have a general relation between the measure of a set and its preimage under $f,$ independent from the set. Nonetheless, in this case, a distortion term appears and it is given by the size of the angular derivative at the point $p.$ If $p \in {\partial\mathbb{D}}$ is the Denjoy-Wolff fixed point of $f,$ this result was previously proved in [@ref:DoeringMane].
In [@ref:FernandezPestanaDistortionInnerFunctions], Fernández and Pestana studied the distortion of Hausdorff contents under inner functions. Fixed $z \in {\mathbb{D}}$ and $0 < \alpha < 1,$ consider the Hausdorff content defined as $$M_{\alpha}(\lambda_z)(E) = \inf \sum_j \lambda_z (I_j)^{\alpha},$$ where the infimum is taken over all collections of arcs $\{I_j \}$ of the unit circle such that $E \subseteq \bigcup I_j$. Thus $M_{\alpha}(\lambda_0) (E)$ is the standard Hausdorff content of $E,$ which is denoted by $M_\alpha (E)$. Observe that if $z \in {\mathbb{D}}$ and $\tau$ is the automorphism of ${\mathbb{D}}$ which interchanges $z$ and $0$, then $M_{\alpha}(\lambda_z)(E) = M_\alpha ({\tau}^{-1} (E))$ for any $E \subseteq {\partial\mathbb{D}}$. Fernández and Pestana proved the following result, analogous to Theorem \[thm:LownerMuP\] for Hausdorff contents, stated here in a conformally invariant way.
\[thm:ConformalFernandezPestana\] For any $0 < \alpha < 1$ there exists a constant $C_\alpha > 0$ such that, if $f\colon {\mathbb{D}}\to {\mathbb{D}}$ is an inner function and $z \in {\mathbb{D}},$ we have $$M_{\alpha}(\lambda_z)(f^{-1}(E)) \geq C_{\alpha} M_{\alpha}(\lambda_{f(z)})(E)$$ for any Borel set $E \subseteq {\partial\mathbb{D}}.$
It is also shown in [@ref:FernandezPestanaDistortionInnerFunctions] that there exists an inner function $f$ such that the preimage of a single point has Hausdorff dimension $1.$ Hence, the converse estimate in Theorem \[thm:ConformalFernandezPestana\] is false. It is worth mentioning that a related result for sets $E \subseteq {\mathbb{D}}$ was established in [@ref:Hamilton]. For $0 < \alpha < 1$ and $p \in {\partial\mathbb{D}},$ we define the $(p,\alpha)$-Hausdorff content of a Borel set $E \subseteq {\partial\mathbb{D}}$ as $$M_{\alpha}(\mu_p)(E) \coloneqq \inf \sum_j \mu_p (I_j)^{\alpha},$$ where the infimum is taken over all collections of arcs $\{I_j\}$ of the unit circle such that $E \setminus \{p\} \subseteq \bigcup I_j.$ Our second result is the following analogue of Theorem \[thm:ConformalFernandezPestana\] when $z \in {\partial\mathbb{D}}.$
\[thm:ContentsMuP\] Let $f\colon {\mathbb{D}}\to {\mathbb{D}}$ be an inner function and let $p \in {\partial\mathbb{D}}$ be a boundary Fatou point of $f.$
(a) \[stm:ContentsMuPFiniteDerivative\] Assume $|f'(p)| < \infty.$ Then for any $0 < \alpha < 1$ there exists a constant $C_\alpha > 0,$ independent of $f,$ such that $$M_{\alpha}(\mu_p)(f^{-1}(E)) \geq C_{\alpha} |f'(p)|^{\alpha} M_{\alpha}(\mu_{f(p)})(E)$$ for any Borel set $E \subseteq {\partial\mathbb{D}}.$
(b) Assume $|f'(p)| = \infty.$ Then we have that $M_{\alpha}(\mu_p)(f^{-1}(E)) = \infty$ for any Borel set $E \subseteq {\partial\mathbb{D}}$ such that $M_{\alpha}(\mu_{f(p)})(E) > 0.$
The proofs of Theorem \[thm:LownerMuP\] and Theorem \[thm:ContentsMuP\] are given in Section \[sec:BoundaryTheorems\]. In Section \[sec:Applications\] we give two applications of our results. The first one concerns a smoothness property of inner functions which omit large sets of the unit disc and it is inspired on a nice result in [@ref:FernandezPestanaDistortionInnerFunctions]. In the second application we obtain analogue results on distortion of sets in the real line under inner mappings of the upper half plane.
It is a pleasure to thank J. J. Donaire, J. L. Fernández, P. Gorkin and M. V. Melián for helpful discusions.
Boundary distortion theorems {#sec:BoundaryTheorems}
============================
In this section we prove our main results. We start with some elementary properties of the measure $\mu_p$ and the content $M_{\alpha}(\mu_p).$ Recall that a sequence of points $\lbrace p_n \rbrace \subseteq {\mathbb{D}}$ converges non-tangentially to a point $p \in {\partial\mathbb{D}}$ if $\lim p_n = p$ and there exists $\beta > 0$ such that $\{p_n\} \subseteq \Gamma_\beta (p).$
\[lemma:MeasuresConvergence\] Let $p \in {\partial\mathbb{D}}.$ For every sequence of points $\lbrace p_n \rbrace \subseteq {\mathbb{D}}$ converging non-tangentially to $p,$ we have $$\mu_{p_n}(E) \longrightarrow \mu_p(E),\quad \text{ as } n \to \infty,$$ for any measurable set $E \subseteq {\partial\mathbb{D}}.$
Let $\lbrace p_n \rbrace_n \subseteq {\mathbb{D}}$ be any sequence of points approaching $p,$ and write $\mu_n = \mu_{p_n}$ for every $n \geq 1.$ By Fatou’s Lemma, we have $$\liminf_n \mu_n(E) \geq \int_E \lim_n \frac{1}{|\xi -p_n|^2}\, d\lambda(\xi) = \mu_p(E),$$ from which it follows that the result is true when $\mu_p(E) = \infty.$ So assume $\mu_p(E) < \infty.$ Fix $\varepsilon > 0$ and consider an arc $I$ centred at $p$ and such that $\mu_p(E \cap I) < \varepsilon.$ Since $p_n \to p$ non-tangentially, there exists a constant $C > 0$ such that $|\xi - p_n| \geq C |\xi - p|$ for every $\xi \in {\partial\mathbb{D}}$ and every $n \geq 1.$ Hence, we have that $\mu_n(E \cap I) \leq C^{-2} \varepsilon$ for every $n.$ On the other hand, by dominated convergence, we have that $$\mu_n(E \cap ({\partial\mathbb{D}}\setminus I)) \longrightarrow \mu_p(E \cap ({\partial\mathbb{D}}\setminus I)),\quad \text{ as } n \to \infty,$$ from which the result follows.
Observe that the assumption on the non-tangential convergence of the sequence $\{p_n\}$ to $p$ only enters into play if $p \in \overline{E}.$ If $p \notin \overline{E},$ the result holds true for any approaching sequence. However, as the following example shows, Lemma \[lemma:MeasuresConvergence\] fails badly if $p_n$ approaches $p$ tangentially. Fix a point $p \in {\partial\mathbb{D}}$ and consider a sequence of points $\{\xi_n\} \subseteq {\partial\mathbb{D}}$ such that $|\xi_n - p| = 1/(2n)$ for every $n \geq 1.$ Consider as well the sequence of pairwise disjoint arcs $\{I_n\}$ such that $I_n$ is centred at $\xi_n$ and $\lambda(I_n) = 1/(4n^4)$ for every $n \geq 1.$ Now, let $E \coloneqq \bigcup_n I_n,$ $p_n = \left(1-\lambda(I_n)\right) \xi_n,$ and $\mu_n = \mu_{p_n},$ for every $n \geq 1.$ Since $(1-|p_n|)/|p-p_n| \leq 1/n^3 \longrightarrow 0,$ the sequence $\{p_n\}$ converges to $p$ tangentially. For $\xi \in I_n,$ we have $|p_n-\xi| \leq 2 \lambda(I_n)$ and $\mu_n(I_n) \geq (4\lambda(I_n))^{-1} = n^4.$ Now, on one hand we have $\mu_n(E) \geq \mu_n(I_n) \longrightarrow \infty,$ as $n\to \infty.$ On the other hand since $|p - \xi| \leq 1/n$ for any $\xi \in I_n$, we have $\mu_p (I_n) \leq n^2 \lambda(I_n) = 1/4n^2$ and we deduce $$\mu_p(E) = \sum_n \mu_p(I_n) < \infty.$$
For $0 < \alpha < 1$ and $z \in {\mathbb{D}}$ consider the $(z,\alpha)$-Hausdorff content of a Borel set $E \subseteq {\partial\mathbb{D}}$ defined as $$M_{\alpha}(\mu_z)(E) = \inf \sum_j \mu_z(I_j)^{\alpha},$$ where the infimum is taken over all collections of arcs $\{I_j\}$ such that $E \subseteq \bigcup I_j.$
\[lemma:BoundaryInteriorMuP\] Given $p \in {\partial\mathbb{D}}$ and $\beta > 0,$ let $\Gamma_\beta(p)$ be the Stolz angle of opening $\beta$ with vertex at $p.$ Then there exists a constant $C = C(\beta) > 0$ such that $$\mu_z(A) \leq C \mu_p(A)$$ for any measurable set $A \subseteq {\partial\mathbb{D}}$ and any $z \in \Gamma_\beta(p).$ Consequently, for any $0 < \alpha < 1$ we also have $M_\alpha(\mu_z)(A) \leq C^{\alpha} M_\alpha(\mu_p)(A)$ for any set $A \subseteq {\partial\mathbb{D}}$ and any $z \in \Gamma_\beta(p).$
Observe that there exists a constant $C = C(\beta) > 0$ such that $|\xi - z| \geq C |\xi - p|$ for any $z \in \Gamma_\beta(p)$ and any $\xi \in {\partial\mathbb{D}}.$ Hence, $\mu_z(A) \leq C^{-2} \mu_p(A)$ for any measurable set $A \subseteq {\partial\mathbb{D}}$ and any $z \in \Gamma_\beta(p).$ This last estimate also gives $M_\alpha(\mu_z)(A) \leq C^{-2 \alpha} M_\alpha(\mu_p)(A).$
The corresponding result to Lemma \[lemma:MeasuresConvergence\] for Hausdorff contents reads as follows.
\[lemma:ContentsNTConvergence\] Let $0 < \alpha < 1$ and $p \in {\partial\mathbb{D}}.$ For any sequence of points $\lbrace p_n \rbrace \subseteq {\mathbb{D}}$ converging non-tangentially to $p,$ we have $$\label{eq:ContentLimit}
\lim_{n \to \infty} M_\alpha(\mu_{p_n})(E) = M_\alpha(\mu_p)(E)$$ for any set $E \subseteq {\partial\mathbb{D}}.$
Write $\mu_n = \mu_{p_n}$ for every $n \geq 1.$ Assume that $M_\alpha(\mu_p)(E) < \infty.$ In this case, we split the proof of the result into two parts. First we show that $$\label{eq:ContentLimitSup}
\limsup_{n \to \infty} M_\alpha(\mu_{n})(E) \leq M_\alpha(\mu_p)(E),$$ and then we prove that $$\label{eq:ContentLimitInf}
\liminf_{n \to \infty} M_\alpha(\mu_{n})(E) \geq M_\alpha(\mu_p)(E),$$ from which follows immediately. To prove , given $\varepsilon >0,$ take a covering by open arcs $\{I_j\}$ of the set $E \setminus \{p\}$ such that $$\sum_j \mu_p(I_j)^\alpha \leq M_\alpha(\mu_p)(E) + \varepsilon.$$ Now, by Lemma \[lemma:BoundaryInteriorMuP\], for each interval $I_j$ and for every $n \geq 1$ we have that $$\mu_{n}(I_j) \leq C\mu_p(I_j).$$ Thus, by Lemma \[lemma:MeasuresConvergence\] and dominated convergence, we get that $$\sum_j \mu_{n}(I_j)^\alpha \longrightarrow \sum_j \mu_p(I_j)^\alpha,\quad \text{ as } n \to \infty.$$ By definition, $M_\alpha(\mu_{n})(E) \leq \sum_j \mu_{n}(I_j)^\alpha$ and, thus follows immediately.
We prove inequality considering two cases. Assume first that $p \not\in \overline{E}.$ Pick $\varepsilon > 0$ and a covering of $E$ by open arcs $\{I_j\},$ such that ${\operatorname{dist}}(I_j,p) \geq {\operatorname{dist}}(\overline{E},p)/2$ for every arc $I_j.$ Observe that, in this situation, there exists $n_0 > 0$ such that if $ n > n_0,$ we have that $$\mu_{n}(I_j) \geq (1-\varepsilon)^{1/\alpha} \mu_p(I_j)$$ for every arc $I_j$ in our covering. Thus, for any such covering of $E \setminus \{p\},$ if $ n > n_0$ we have that $$\label{eq:AlphaSumLimitInf}
\sum_j \mu_{n}(I_j)^\alpha \geq (1-\varepsilon) M_\alpha(\mu_p)(E).$$ Observe that the infimum of $\sum_j \mu_{n}(I_j)^\alpha$ when ranging over all coverings $\{I_j\}$ of $E \setminus \{p\}$ by open arcs satisfying that ${\operatorname{dist}}(I_j,p) \geq {\operatorname{dist}}(\overline{E},p)/2$ is, precisely, $M_\alpha(\mu_{n})(E).$ Hence, equation follows in the case that $p \not\in \overline{E},$ and therefore equation as well in this situation.
In the case that $p \in \overline{E},$ since we assumed that $M_\alpha(\mu_p) (E) < \infty,$ given $\varepsilon > 0$ we can choose $\delta = \delta(\varepsilon) > 0$ such that $M_\alpha(\mu_p)(E \cap I(p,\delta)) < \varepsilon,$ where $I(p,\delta)$ denotes the arc centred at $p$ of length $\delta.$ Let us denote $E_\delta = E \setminus I(p,\delta).$ Since $p \not \in \overline{E_\delta},$ we already know that $$\lim_{n \to \infty} M_\alpha(\mu_{n})(E_\delta) = M_\alpha(\mu_p)(E_\delta) \geq M_\alpha(\mu_p)(E) - \varepsilon.$$ Hence, for any given $\varepsilon > 0,$ we have $$\liminf_{n \to \infty} M_\alpha(\mu_{n})(E) \geq \lim_{n \to \infty} M_\alpha(\mu_{n})(E_\delta) \geq M_\alpha(\mu_p)(E) - \varepsilon.$$ This concludes the proof whenever $M_\alpha(\mu_p) (E) < \infty.$
Assume now that $M_\alpha(\mu_p) (E) = \infty.$ In this case, for any $N > 0$ we can find $\delta = \delta(N) > 0$ such that $M_\alpha(\mu_p)(E_\delta) > N,$ where again $E_\delta = E \setminus I(p,\delta).$ Since $p \not \in \overline{E_\delta},$ we have that $$\lim_{n \to \infty} M_\alpha(\mu_{n})(E_\delta) = M_\alpha(\mu_p)(E_\delta) > N.$$ Hence, there exists $n_0 > 0$ such that if $n > n_0,$ then $M_\alpha(\mu_{n})(E_\delta) > N.$ Using that $M_\alpha(\mu_{n})(E) \geq M_\alpha(\mu_{n})(E_\delta),$ we get in the case in which $M_\alpha(\mu_p) (E) = \infty$ as well.
We will use the following auxiliary result which is certainly well known. It is included because we have not found a precise reference.
\[lemma:FiniteDerivativeCones\] Let $f$ be a holomorphic self map of the unit disc. Let $\{p_n\}$ be a sequence of points in ${\mathbb{D}}$ converging non-tangentially to a point $p \in {\partial\mathbb{D}}.$ Assume that $|f'(p)| < \infty,$ then $\{f(p_n)\}$ also converges to $f(p) \in {\partial\mathbb{D}}$ non-tangentially.
Since $|f'(p)| < \infty$ we have that $f(p) \in {\partial\mathbb{D}}.$ Write $$\frac{1-|f(p_n)|}{|f(p)-f(p_n)|} = \frac{1-|f(p_n)|}{1-|p_n|} \frac{1-|p_n|}{|p-p_n|} \frac{|p-p_n|}{|f(p)-f(p_n)|}.$$ Also because $|f'(p)|<\infty$, by Julia-Carathéodory Theorem, the first and third terms converge respectively to $|f'(p)|$ and $|f'(p)|^{-1}$, and therefore $$\liminf_n \frac{1-|f(p_n)|}{|f(p)-f(p_n)|} = \liminf_n \frac{1-|p_n|}{|p-p_n|} > 0.$$
Note that the assumption of finite angular derivative is necessary in the above statement, even if we ask the function $f$ to be inner. In fact, it can be proved that there exist inner functions mapping a given Stolz angle to a tangential region (see [@ref:DonaireRadialBehaviourLittleBloch]).
We are now ready to prove our main results.
We can choose a sequence of points $\{p_n\}$ in ${\mathbb{D}}$ approaching $p$ non-tangentially such that $$\label{eq:AngularDerivativeCondition}
\lim_{n \to \infty} \frac{1-|f(p_n)|^2}{1-|p_n|^2} = |f'(p)| > 0.$$
By Theorem \[thm:LownerLemma\], we have that $$\label{eq:LownerPropertyMuPN}
\mu_{p_n}(f^{-1}(E)) = \frac{1-|f(p_n)|^2}{1-|p_n|^2} \mu_{f(p_n)}(E).$$
Lemma \[lemma:MeasuresConvergence\] gives that $\mu_{p_n}(f^{-1}(E)) \rightarrow \mu_p(f^{-1}(E))$ as $n \to \infty.$ If $|f'(p)| < \infty,$ Lemma \[lemma:FiniteDerivativeCones\] gives that $f(p_n)$ converges to $f(p)$ non-tangentially. Thus, Lemma \[lemma:MeasuresConvergence\] gives that $\mu_{f(p_n)}(E) \rightarrow \mu_{f(p)}(E)$ as $n \to \infty.$ Therefore, equations and give the statement . Assume now that $|f'(p)| = \infty.$ If $\mu_{f(p)}(E) = 0,$ we have $\lambda(E) = 0.$ Hence, by Theorem \[thm:LownerLemma\], we have that $\lambda(f^{-1}(E)) = 0$ and it follows that $\mu_p(f^{-1}(E)) = 0.$ Finally assume $\mu_{f(p)}(E) > 0.$ Observe that for any $n \geq 1$ we have $\mu_{f(p_n)}(E) > \lambda(E) /4 >0.$ Thus, since $|f'(p)| = \infty,$ the right-hand side of equation tends to infinity and, by Lemma \[lemma:MeasuresConvergence\], we deduce that $\mu_p (f^{-1}(E)) = \infty.$
We will use Theorem \[thm:ConformalFernandezPestana\] in the following form. For $z \in {\mathbb{D}}$ we have that $$\label{eq:GeneralisedInternalFP}
M_\alpha(\mu_z)(f^{-1}(E)) \geq C_\alpha \left(\frac{1-|f(z)|^2}{1-|z|^2}\right)^\alpha M_\alpha(\mu_{f(z)})(E)$$ for any Borel set $E \subseteq {\partial\mathbb{D}}.$ We can choose a sequence of points $\{p_n\}$ in ${\mathbb{D}}$ approaching $p$ non-tangentially such that $$\label{eq:AngularDerivativeCondition2}
\lim_{n \to \infty} \frac{1-|f(p_n)|^2}{1-|p_n|^2} = |f'(p)| > 0.$$ Assume $|f'(p)|< \infty.$ Applying Lemma \[lemma:ContentsNTConvergence\] and equation , we get $$\begin{split}
M_{\alpha}(\mu_p)(f^{-1}(E)) &= \lim_{r\to 1} M_{\alpha}(\mu_{p_n})(f^{-1}(E))\\
&\geq \limsup_{n\to \infty} C_{\alpha} \left(\frac{1-|f(p_n)|^2}{1-|p_n|^2}\right)^{\alpha} M_{\alpha}(\mu_{f(p_n)})(E)\\
&= C_{\alpha} |f'(p)|^{\alpha} \limsup_{n\to \infty} M_{\alpha}(\mu_{f(p_n)})(E).
\end{split}$$ By Lemma \[lemma:FiniteDerivativeCones\], $f(p_n)$ tends to $f(p)$ non-tangentially as $n \to \infty$ and hence, Lemma \[lemma:ContentsNTConvergence\] gives that $$\lim_{n \to \infty} M_{\alpha}(\mu_{f(p_n)})(E) = M_{\alpha}(\mu_{f(p)})(E),$$ which finishes the proof of part . Assume now $|f'(p)| = \infty.$ We can assume $f(p) \notin E.$ Since $M_{\alpha}(\mu_{f(p)})(E) > 0,$ there exists an arc $I$ centred at $f(p)$ such that $M_{\alpha}(\mu_{f(p)})(E \setminus I) > 0.$ Write $E^* = E \setminus I.$ Then there exists $n_0 >0$ such that $M_{\alpha}(\mu_{f(p_n)})(E^*) > M_{\alpha}(\mu_{f(p)})(E^*) /2$ if $n > n_0.$ Now, $$\begin{split}
M_{\alpha}(\mu_p)(f^{-1}(E^*)) &= \lim_{n\to \infty} M_{\alpha}(\mu_{p_n})(f^{-1}(E^*))\\
&\geq C_\alpha \limsup_{n\to \infty} \left(\frac{1-|f(p_n)|^2}{1-|p_n|^2}\right)^{\alpha} M_{\alpha}(\mu_{f(p_n)})(E^*) = \infty.
\end{split}$$ Hence $M_{\alpha}(\mu_p)(f^{-1}(E)) = \infty.$
Applications {#sec:Applications}
============
Omitted values
--------------
A classical result by Frostman says that any inner function $f$ can omit at most a set of logarithmic capacity zero, that is, ${\mathbb{D}}\setminus f({\mathbb{D}})$ has logarithmic capacity zero (see Chapter II of [@ref:GarnettBoundedAnalyticFunctions]). Conversely, given a relatively compact set $K$ of the unit disc of logarithmic capacity zero, the universal covering map $f\colon {\mathbb{D}}\rightarrow {\mathbb{D}}\setminus K$ is an inner function (see page 323 of [@ref:TsujiPotentialTheory]). Given a set $E \subseteq {\mathbb{D}},$ its *non-tangential closure* on ${\partial\mathbb{D}}$, denoted by $E^{NT},$ is the set of points $\xi \in {\partial\mathbb{D}}$ for which there exists a sequence $\{z_n\} \subseteq E$ such that $z_n \to \xi$ non-tangentially. We first state an auxiliary result which may have independent interest.
\[lemma:inclusion\] Let $f\colon {\mathbb{D}}\rightarrow {\mathbb{D}}$ be an inner function and let $E = {\mathbb{D}}\setminus f({\mathbb{D}})$ be the set of its omitted points. Then $$f^{-1}(E^{NT}) \subseteq \{\xi \in {\partial\mathbb{D}}\colon |f'(\xi)| = \infty\}.$$
Consider a point $\xi \in {\partial\mathbb{D}}$ such that the angular derivative of $f$ at $\xi$ exists and it is finite, and let $\zeta = f(\xi).$ In other words assume that $$\label{eq:FiniteAngularDerivative}
\lim_{\Gamma_\beta(\xi) \ni z \rightarrow \xi } \frac{\zeta - f(z)}{\xi-z} = A$$ is finite. We want to see that, in this situation, for any opening $\gamma > 1,$ there is $0 < s = s(\gamma) < 1$ such that the truncated cone $$\Gamma_{\gamma,s}(\zeta) = \left\{w \in {\mathbb{D}}\colon |\zeta-w| < \gamma (1-|w|), |\zeta-w| < s \right\}$$ does not intersect $E,$ that is, $\Gamma_{\gamma,s}(\zeta) \subseteq f({\mathbb{D}}).$ So fix $\gamma > 1$ and consider $\Gamma_{\gamma,s}(\zeta)$ with $0 < s < 1$ to be determined. Fix $w_0 \in \Gamma_{\gamma,s}(\zeta).$ We want to see that there is $z_0 \in {\mathbb{D}}$ such that $f(z_0) = w_0.$ By equation , we can express $$f(z) = \zeta + A (z-\xi) + o(|z-\xi|),$$ where $o(|z-\xi|)/|z-\xi| \rightarrow 0$ as $z \to \xi$ non-tangentially. Consider $\Gamma_{\beta,r}(\xi)$ with $\beta > 2\gamma$ and $0 < r < 1$ to be determined. Observe that there exists $0 < r_0 < 1$ such that, if $r < r_0$ and $0< s < |A|r/2,$ then for any $z \in \partial\Gamma_{\beta,r}(\xi)$ we have that $$\left|(f(z)-w_0) - (\zeta+A(z-\xi)-w_0)\right| < |\zeta+A(z-\xi)-w_0|.$$ Thus, by Rouché’s Theorem, the functions $f(z)-w_0$ and $g(z)-w_0 = \zeta+A(z-\xi)-w_0$ have the same number of zeroes in $\Gamma_{\beta,r}(\xi).$ But $g(z)$ is a degree $1$ polynomial and $g(\Gamma_{\beta,r}(\xi)) = \Gamma_{\beta,|A|r}(\zeta) \supseteq \Gamma_{\gamma,s}(\zeta),$ and thus $g(z)-w_0$ has a single zero on $\Gamma_{\beta,r}(\xi).$ Therefore, there is $z_0 \in \Gamma_{\beta,r}(\xi)$ such that $f(z_0) = w_0,$ which completes the proof.
As an application of Theorem \[thm:LownerMuP\] and Lemma \[lemma:inclusion\], we have the following result.
Let $f\colon {\mathbb{D}}\rightarrow {\mathbb{D}}$ be an inner function and let $E = {\mathbb{D}}\setminus f({\mathbb{D}})$ be the set of its omitted points. Let $p$ be a boundary Fatou point of $f$.
(a) Assume $|f'(p)| < \infty.$ Then for any $0 < \alpha < 1$ there exists a constant $C_\alpha > 0,$ independent of $f,$ such that $$\label{eq:InfiniteDerivativeContent}
M_\alpha(\mu_p)\left(\{\xi \in {\partial\mathbb{D}}\colon |f'(\xi)| = \infty\}\right) \geq C_\alpha |f'(p)|^\alpha M_\alpha(\mu_{f(p)})(E^{NT}).$$
(b) Assume $|f'(p)| = \infty.$ Then $M_\alpha(\mu_p)\left(\{\xi \in {\partial\mathbb{D}}\colon |f'(\xi)| = \infty\}\right) = \infty$ whenever $M_\alpha(\mu_{f(p)})(E^{NT}) > 0.$
Inner functions in the upper half plane
---------------------------------------
Let ${\mathbb{H}}= \{w \in \mathbb{C}\colon {\operatorname{Im}}(w) > 0 \}$ be the upper half plane. A holomorphic mapping $g\colon {\mathbb{H}}\rightarrow {\mathbb{H}}$ is an *inner function* of the upper half plane if $\lim_{y \to 0} g(x + iy) \in {\mathbb{R}}$ for a.e. $x \in {\mathbb{R}}.$ This natural definition agrees with conformal changes of coordinates: given $p \in {\partial\mathbb{D}}$ denote by $w_p$ the Möbius transformation mapping ${\mathbb{D}}$ onto ${\mathbb{H}},$ the point $p$ to $\infty$ and, say, the origin to $i/2.$ Then, $g$ is an inner function of the upper half plane if and only if $f = w_p^{-1} \circ g \circ w_p$ is an inner function of the unit disc ${\mathbb{D}}.$ Observe that $g(\infty) = \lim_{t \to +\infty} g(it) = \infty$ if and only if $f(p) = p.$ A holomorphic mapping $g$ from ${\mathbb{H}}$ into ${\mathbb{H}}$ has a finite angular derivative at $\infty$ if $$g'(\infty) = \lim_{t \to +\infty} \frac{it}{g(it)}$$ exists and is finite. Otherwise, we write $|g'(\infty)| = \infty.$ Observe that $g$ has a finite angular derivative at infinity if and only if $f = w_p^{-1} \circ g \circ w_p$ has a finite angular derivative at $p.$ Moreover, the identity $|g'(\infty)| = |f'(p)|$ holds in the sense that both quantities coincide when they are finite, and if one of them is infinite so is the other. This fact easily follows from the identity $$\frac{w}{g(w)} = \frac{p + z}{p + f(z)}\frac{p - f(z)}{p - z}.$$ Let $|A|$ denote the Lebesgue measure of a measurable set $A \subseteq {\mathbb{R}}$ and, for $0 < \alpha < 1,$ let $M_\alpha (A)$ denote its $\alpha$-Hausdorff content. We now state the versions of and in this setting.
\[cor:LownerMuP\] Let $g\colon {\mathbb{H}}\to {\mathbb{H}}$ be an inner function and assume that $g(\infty) = \infty.$
(a) \[stm:LownerMuPFiniteDerivativeUpperHalfPlane\] Assume $|g'(\infty)| < \infty.$ Then $$\label{first}
|g^{-1}(A)| = |g'(\infty)| |A|$$ for any measurable set $A \subseteq {\mathbb{R}}.$ Moreover, for any $0 < \alpha < 1$ there exists a constant $C_\alpha > 0,$ independent of $g,$ such that $$\label{formulacontent}
M_{\alpha} (g^{-1}(A)) \geq C_{\alpha} |g'(\infty)|^{\alpha} M_{\alpha}(A)$$ for any Borel set $A \subseteq {\mathbb{R}}.$
(b) \[stm:LownerMuPInfiniteDerivativeUpperHalfPlane\] If $|g'(\infty)| = \infty$ and $A \subseteq {\mathbb{R}}$ is a measurable set, then $|g^{-1}(A)| = \infty$ if $|A| > 0$ and $|g^{-1}(A)| = 0$ if $|A|= 0.$ Moreover, $M_{\alpha}(g^{-1}(A)) = \infty$ for any Borel set $A \subseteq {\mathbb{R}}$ such that $M_{\alpha}(A) > 0.$
Note that for any measurable set $A \subseteq {\mathbb{R}}$ we have $$\label{leb r}
|A| = \mu_p(w_p^{-1}(A)), \qquad p \in {\partial\mathbb{D}}.$$ Hence, $|g^{-1}(A)| = \mu_p(w_p^{-1}(g^{-1}(A))) = \mu_p(f^{-1} (w_p^{-1}(A))).$ Applying Theorem \[thm:LownerMuP\] and we deduce $|g^{-1}(A)| = |f'(p)| \mu_p(w_p^{-1}(A)) = |g'(\infty)| |A|$ which is . It follows from and $w_p$ being a Möbius map that $$\label{rel contents}
M_\alpha(\mu_p)(E) = M_\alpha(w_p(E)), \quad E \subseteq {\partial\mathbb{D}}.$$ Thus, the previous argument shows that holds. Part follows from similar considerations.
Matteo Levi: <span style="font-variant:small-caps;">Università di Bologna, Dipartimento di Matematica, Via Zamboni 33, 40126 Bologna, Italia</span>
*E-mail address*: `[email protected]`
Artur Nicolau: <span style="font-variant:small-caps;">Universitat Autònoma De Barcelona, Departament de Matemàtiques, Edifici C, 08193-Bellaterra, Catalunya</span>
*E-mail address*: `[email protected]`
Odí Soler i Gibert: <span style="font-variant:small-caps;">Universitat Autònoma De Barcelona, Departament de Matemàtiques, Edifici C, 08193-Bellaterra, Catalunya</span>
*E-mail address*: `[email protected]`
[^1]: The first author is partially supported by the 2015 PRIN grant *Real and Complex Manifolds: Geometry, Topology and Harmonic Analysis* of the Italian Ministry of Education (MIUR).
[^2]: The three authors are supported in part by the Generalitat de Catalunya (grant 2017 SGR 395) and the Spanish Ministerio de Ciencia e Innovación (projects MTM2014-51824-P, MTM2017-85666-P).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Learning both hierarchical and temporal representation has been among the long-standing challenges of recurrent neural networks. Multiscale recurrent neural networks have been considered as a promising approach to resolve this issue, yet there has been a lack of empirical evidence showing that this type of models can actually capture the temporal dependencies by discovering the latent hierarchical structure of the sequence. In this paper, we propose a novel multiscale approach, called the hierarchical multiscale recurrent neural network, that can capture the latent hierarchical structure in the sequence by encoding the temporal dependencies with different timescales using a novel update mechanism. We show some evidence that the proposed model can discover underlying hierarchical structure in the sequences without using explicit boundary information. We evaluate our proposed model on character-level language modelling and handwriting sequence generation.'
author:
- |
Junyoung Chung, Sungjin Ahn & Yoshua Bengio [^1]\
Département d’informatique et de recherche opérationnelle\
Université de Montréal\
`{junyoung.chung,sungjin.ahn,yoshua.bengio}@umontreal.ca`\
bibliography:
- './junyoung\_thesis.bib'
title: |
Hierarchical Multiscale\
Recurrent Neural Networks
---
Introduction {#sec:introduction}
============
One of the key principles of learning in deep neural networks as well as in the human brain is to obtain a hierarchical representation with increasing levels of abstraction [@bengio2009learning; @lecun2015deep; @schmidhuber2015deep]. A stack of representation layers, learned from the data in a way to optimize the target task, make deep neural networks entertain advantages such as generalization to unseen examples [@hoffman2013one], sharing learned knowledge among multiple tasks, and discovering disentangling factors of variation [@kingma2013auto]. The remarkable recent successes of the deep convolutional neural networks are particularly based on this ability to learn hierarchical representation for spatial data [@krizhevsky2012imagenet]. For modelling temporal data, the recent resurgence of recurrent neural networks (RNN) has led to remarkable advances [@mikolov2010recurrent; @graves2013generating; @Cho-et-al-EMNLP2014; @sutskever2014sequence; @vinyals2015show]. However, unlike the spatial data, learning both hierarchical and temporal representation has been among the long-standing challenges of RNNs in spite of the fact that hierarchical multiscale structures naturally exist in many temporal data [@schmidhuber1991neural; @mozer1993induction; @el1995hierarchical; @lin1996learning; @koutnik2014clockwork].
A promising approach to model such hierarchical and temporal representation is the multiscale RNNs [@schmidhuber1992learning; @el1995hierarchical; @koutnik2014clockwork]. Based on the observation that high-level abstraction changes slowly with temporal coherency while low-level abstraction has quickly changing features sensitive to the precise local timing [@el1995hierarchical], the multiscale RNNs group hidden units into multiple modules of different timescales. In addition to the fact that the architecture fits naturally to the latent hierarchical structures in many temporal data, the multiscale approach provides the following advantages that resolve some inherent problems of standard RNNs: (a) computational efficiency obtained by updating the high-level layers less frequently, (b) efficiently delivering long-term dependencies with fewer updates at the high-level layers, which mitigates the vanishing gradient problem, (c) flexible resource allocation (e.g., more hidden units to the higher layers that focus on modelling long-term dependencies and less hidden units to the lower layers which are in charge of learning short-term dependencies). In addition, the learned latent hierarchical structures can provide useful information to other downstream tasks such as module structures in computer program learning, sub-task structures in hierarchical reinforcement learning, and story segments in video understanding.
There have been various approaches to implementing the multiscale RNNs. The most popular approach is to set the timescales as hyperparameters [@el1995hierarchical; @koutnik2014clockwork; @bahdanau2016end] instead of treating them as dynamic variables that can be learned from the data [@schmidhuber1991neural; @schmidhuber1992learning; @chung2015gated; @chung2016character]. However, considering the fact that non-stationarity is prevalent in temporal data, and that many entities of abstraction such as words and sentences are in variable length, we claim that it is important for an RNN to dynamically adapt its timescales to the particulars of the input entities of various length. While this is trivial if the hierarchical boundary structure is provided [@sordoni2015hierarchical], it has been a challenge for an RNN to discover the latent hierarchical structure in temporal data without explicit boundary information.
In this paper, we propose a novel multiscale RNN model, which can learn the hierarchical multiscale structure from temporal data without explicit boundary information. This model, called a *hierarchical multiscale recurrent neural network* (HM-RNN), does not assign fixed update rates, but adaptively determines proper update times corresponding to different abstraction levels of the layers. We find that this model tends to learn fine timescales for low-level layers and coarse timescales for high-level layers. To do this, we introduce a binary boundary detector at each layer. The boundary detector is turned on only at the time steps where a segment of the corresponding abstraction level is completely processed. Otherwise, i.e., during the within segment processing, it stays turned off. Using the hierarchical boundary states, we implement three operations, UPDATE, COPY and FLUSH, and choose one of them at each time step. The UPDATE operation is similar to the usual update rule of the long short-term memory (LSTM) [@hochreiter1997long], except that it is executed sparsely according to the detected boundaries. The COPY operation simply copies the cell and hidden states of the previous time step. Unlike the leaky integration of the LSTM or the Gated Recurrent Unit (GRU) [@Cho-et-al-EMNLP2014], the COPY operation retains the whole states without any loss of information. The FLUSH operation is executed when a boundary is detected, where it first ejects the summarized representation of the current segment to the upper layer and then reinitializes the states to start processing the next segment. Learning to select a proper operation at each time step and to detect the boundaries, the HM-RNN discovers the latent hierarchical structure of the sequences. We find that the straight-through estimator [@hinton2012coursera; @bengio2013estimating; @courbariaux2016binarized] is efficient for training this model containing discrete variables.
We evaluate our model on two tasks: character-level language modelling and handwriting sequence generation. For the character-level language modelling, the HM-RNN achieves the state-of-the-art results on the Text8 dataset, and comparable results to the state-of-the-art on the Penn Treebank and Hutter Prize Wikipedia datasets. The HM-RNN also outperforms the standard RNN on the handwriting sequence generation using the IAM-OnDB dataset. In addition, we demonstrate that the hierarchical structure found by the HM-RNN is indeed very similar to the intrinsic structure observed in the data. The contributions of this paper are:
We propose for the first time an RNN model that can learn a latent hierarchical structure of a sequence without using explicit boundary information.
We show that it is beneficial to utilize the above structure through empirical evaluation.
We show that the straight-through estimator is an efficient way of training a model containing discrete variables.
We propose the [*slope annealing*]{} trick to improve the training procedure based on the straight-through estimator.
Related Work {#sec:related_work}
============
Two notable early attempts inspiring our model are @schmidhuber1992learning and @el1995hierarchical. In these works, it is advocated to stack multiple layers of RNNs in a decreasing order of update frequency for computational and learning efficiency. In @schmidhuber1992learning, the author shows a model that can self-organize a hierarchical multiscale structure. Particularly in @el1995hierarchical, the advantages of incorporating a priori knowledge, “*temporal dependencies are structured hierarchically*", into the RNN architecture is studied. The authors propose an RNN architecture that updates each layer with a fixed but different rate, called a hierarchical RNN.
LSTMs [@hochreiter1997long] employ the multiscale update concept, where the hidden units have different forget and update rates and thus can operate with different timescales. However, unlike our model, these timescales are not organized hierarchically. Although the LSTM has a self-loop for the gradients that helps to capture the long-term dependencies by mitigating the vanishing gradient problem, in practice, it is still limited to a few hundred time steps due to the leaky integration by which the contents to memorize for a long-term is gradually diluted at every time step. Also, the model remains computationally expensive because it has to perform the update at every time step for each unit. However, our model is less prone to these problems because it learns a hierarchical structure such that, by design, high-level layers learn to perform less frequent updates than low-level layers. We hypothesize that this property mitigates the vanishing gradient problem more efficiently while also being computationally more efficient.
A more recent model, the clockwork RNN (CW-RNN) [@koutnik2014clockwork] extends the hierarchical RNN [@el1995hierarchical] and the NARX RNN [@lin1996learning][^2]. The CW-RNN tries to solve the issue of using soft timescales in the LSTM, by explicitly assigning hard timescales. In the CW-RNN, hidden units are partitioned into several modules, and different timescales are assigned to the modules such that a module $i$ updates its hidden units at every $2^{(i-1)}$-th time step. The CW-RNN is computationally more efficient than the standard RNN including the LSTM since hidden units are updated only at the assigned clock rates. However, finding proper timescales in the CW-RNN remains as a challenge whereas our model *learns* the intrinsic timescales from the data. In the biscale RNNs [@chung2016character], the authors proposed to model *layer-wise* timescales adaptively by having additional gating units, however this approach still relies on the *soft* gating mechanism like LSTMs.
Other forms of Hierarchical RNN (HRNN) architectures have been proposed in the cases where the explicit hierarchical boundary structure is provided. In @ling2015character, after obtaining the word boundary via tokenization, the HRNN architecture is used for neural machine translation by modelling the characters and words using the first and second RNN layers, respectively. A similar HRNN architecture is also adopted in @sordoni2015hierarchical to model dialogue utterances. However, in many cases, hierarchical boundary information is not explicitly observed or expensive to obtain. Also, it is unclear how to deploy more layers than the number of boundary levels that is explicitly observed in the data.
While the above models focus on online prediction problems, where a prediction needs to be made by using only the past data, in some cases, predictions are made after observing the whole sequence. In this setting, the input sequence can be regarded as 1-D spatial data, convolutional neural networks with 1-D kernels are proposed in @kim2014convolutional and @kim2015character for language modelling and sentence classification. Also, in @chan2016listen and @bahdanau2016end, the authors proposed to obtain high-level representation of the sequences of reduced length by repeatedly merging or pooling the lower-level representation of the sequences.
Hierarchical RNN architectures have also been used to discover the segmentation structure in sequences [@fernandez2007sequence; @kong2015segmental]. It is however different to our model in the sense that they optimize the objective with explicit labels on the hierarchical segments while our model discovers the intrinsic structure only from the sequences without segment label information.
The COPY operation used in our model can be related to Zoneout [@krueger2016zoneout] which is a recurrent generalization of stochastic depth [@huang2016deep]. In Zoneout, an identity transformation is randomly applied to each hidden unit at each time step according to a Bernoulli distribution. This results in occasional copy operations of the previous hidden states. While the focus of Zoneout is to propose a regularization technique similar to dropout [@srivastava2014dropout] (where the regularization strength is controlled by a hyperparameter), our model learns (a) to dynamically determine when to copy from the context inputs and (b) to discover the hierarchical multiscale structure and representation. Although the main goal of our proposed model is not regularization, we found that our model also shows very good generalization performance.
Hierarchical Multiscale Recurrent Neural Networks {#sec:model}
=================================================
Motivation {#sec:motivation}
----------
![(a) The HRNN architecture, which requires the knowledge of the hierarchical boundaries. (b) The HM-RNN architecture that discovers the hierarchical multiscale structure in the data.[]{data-label="fig:motivation"}](motiv_1.pdf){width="1.\columnwidth"}
![(a) The HRNN architecture, which requires the knowledge of the hierarchical boundaries. (b) The HM-RNN architecture that discovers the hierarchical multiscale structure in the data.[]{data-label="fig:motivation"}](motiv_2.pdf){width="1.\columnwidth"}
(a)
(b)
To begin with, we provide an example of how a stacked RNN can model temporal data in an ideal setting, i.e., when the hierarchy of segments is provided [@sordoni2015hierarchical; @ling2015character]. In Figure \[fig:motivation\] (a), we depict a hierarchical RNN (HRNN) for language modelling with two layers: the first layer receives characters as inputs and generates word-level representations (C2W-RNN), and the second layer takes the word-level representations as inputs and yields phrase-level representations (W2P-RNN).
As shown, by means of the provided end-of-word labels, the C2W-RNN obtains word-level representation after processing the last character of each word and passes the word-level representation to the W2P-RNN. Then, the W2P-RNN performs an update of the phrase-level representation. Note that the hidden states of the W2P-RNN remains unchanged while all the characters of a word are processed by the C2W-RNN. When the C2W-RNN starts to process the next word, its hidden states are reinitialized using the latest hidden states of the W2P-RNN, which contain summarized representation of all the words that have been processed by that time step, in that phrase.
From this simple example, we can see the advantages of having a hierarchical multiscale structure: (1) as the W2P-RNN is updated at a much slower update rate than the C2W-RNN, a considerable amount of computation can be saved, (2) gradients are backpropagated through a much smaller number of time steps, and (3) layer-wise capacity control becomes possible (e.g., use a smaller number of hidden units in the first layer which models short-term dependencies but whose updates are invoked much more often).
*Can an RNN discover such hierarchical multiscale structure without explicit hierarchical boundary information?* Considering the fact that the boundary information is difficult to obtain (for example, consider languages where words are not always cleanly separated by spaces or punctuation symbols, and imperfect rules are used to separately perform segmentation) or usually not provided at all, this is a legitimate problem. It gets worse when we consider higher-level concepts which we would like the RNN to discover autonomously. In Section \[sec:related\_work\], we discussed the limitations of the existing RNN models under this setting, which either have to update all units at every time step or use fixed update frequencies [@el1995hierarchical; @koutnik2014clockwork]. Unfortunately, this kind of approach is not well suited to the case where different segments in the hierarchical decomposition have different lengths: for example, different words have different lengths, so a fixed hierarchy would not update its upper-level units in synchrony with the natural boundaries in the data.
The Proposed Model {#sec:proposed}
------------------
A key element of our model is the introduction of a parametrized boundary detector, which outputs a binary value, in each layer of a stacked RNN, and learns when a segment should end in such a way to optimize the overall target objective. Whenever the boundary detector is turned on at a time step of layer $\ell$ (i.e., when the boundary state is $1$), the model considers this to be the end of a segment corresponding to the latent abstraction level of that layer (e.g., word or phrase) and feeds the summarized representation of the detected segment into the upper layer ($\ell + 1$). Using the boundary states, at each time step, each layer selects one of the following operations: UPDATE, COPY or FLUSH. The selection is determined by (1) the boundary state of the current time step in the layer below $z_t^{\ell-1}$ and (2) the boundary state of the previous time step in the same layer $z_{t-1}^\ell$.
In the following, we describe an HM-RNN based on the LSTM update rule. We call this model a hierarchical multiscale LSTM (HM-LSTM). Consider an HM-LSTM model of $L$ layers ($\ell=1,\dots,L$) which, at each layer $\ell$, performs the following update at time step $t$: \_t\^, \_t\^, z\_t\^= f\_\^(\_[t-1]{}\^, \_[t-1]{}\^, \_t\^[-1]{}, \_[t-1]{}\^[+1]{}, z\_[t-1]{}\^, z\_t\^[-1]{}). Here, $\bh$ and $\bc$ denote the hidden and cell states, respectively. The function $f_\text{HM-LSTM}^\ell$ is implemented as follows. First, using the two boundary states $z_{t-1}^\ell$ and $z_t^{\ell-1}$, the cell state is updated by: \[eq:update\_rule\] \^\_t
\_t\^\_[t-1]{}\^+ \_t\^\_t\^& z\^\_[t-1]{} = 0 z\^[-1]{}\_[t]{} = 1\
\^\_[t-1]{} & z\^\_[t-1]{} = 0 z\^[-1]{}\_[t]{} = 0\
\_t\^\_t\^& z\^\_[t-1]{} = 1 ,
and then the hidden state is obtained by: \_t\^=
\_[t-1]{}\^&\
\_t\^(\_[t]{}\^) &
Here, $(\bff, \bi, \bo)$ are forget, input, output gates, and $\bg$ is a cell proposal vector. Note that unlike the LSTM, it is not necessary to compute these gates and cell proposal values at every time step. For example, in the case of the COPY operation, we do not need to compute any of these values and thus can save computations.
The COPY operation, which simply performs $(\bc_t^\ell, \bh_t^\ell) \law (\bc_{t-1}^\ell,\bh_{t-1}^\ell)$, implements the observation that an upper layer should keep its state unchanged until it receives the summarized input from the lower layer. The UPDATE operation is performed to update the summary representation of the layer $\ell$ if the boundary $z^{\ell-1}_t$ is detected from the layer below but the boundary $z^\ell_{t-1}$ was not found at the previous time step. Hence, the UPDATE operation is executed sparsely unlike the standard RNNs where it is executed at every time step, making it computationally inefficient. If a boundary is detected, the FLUSH operation is executed. The FLUSH operation consists of two sub-operations: (a) EJECT to pass the current state to the upper layer and then (b) RESET to reinitialize the state before starting to read a new segment. This operation implicitly forces the upper layer to absorb the summary information of the lower layer segment, because otherwise it will be lost. Note that the FLUSH operation is a [*hard*]{} reset in the sense that it completely erases all the previous states of the same layer, which is different from the [*soft*]{} reset or [*soft*]{} forget operation in the GRU or LSTM.
Whenever needed (depending on the chosen operation), the gate values ($\bff_t^\ell, \bi_t^\ell, \bo_t^\ell$), the cell proposal $\bg_t^\ell$, and the pre-activation of the boundary detector $\tilde{z}_t^\ell$ [^3] are then obtained by:
\^\_t\
\^\_t\
\^\_t\
\^\_t\
\^\_t
\
\
\
\
f\_(\^[()]{}\_t + \^[()]{}\_t + \^[()]{}\_t + \^[()]{} ), where \^[()]{}\_t U\_\^\^\_[t-1]{},\
\[eq:top\_down\] \^[()]{}\_t z\^\_[t-1]{}U\_[+1]{}\^\^[+1]{}\_[t-1]{},\
\^[()]{}\_t z\^[-1]{}\_t W\_[-1]{}\^ \_t\^[-1]{}. Here, we use $W_i^j \in \eR^{(4dim(\bh^\ell)+1) \times dim(\bh^{\ell-1})}, U_i^j\in\eR^{(4dim(\bh^\ell)+1) \times dim(\bh^\ell)}$ to denote state transition parameters from layer $i$ to layer $j$, and $\bb \in \eR^{4dim(\bh^\ell)+1}$ is a bias term. In the last layer $L$, the top-down connection is ignored, and we use $\bh_t^{0} = \bx_t$. Since the input should not be omitted, we set $z^0_t=1$ for all $t$. Also, we do not use the boundary detector for the last layer. The $\mathtt{hard\;sigm}$ is defined by [$\mathtt{hard\;sigm}(x) = \max\left(0,\min\left(1,\f{ax + 1}{2}\right)\right)$]{} with $a$ being the slope variable.
Unlike the standard LSTM, the HM-LSTM has a top-down connection from $(\ell+1)$ to $\ell$, which is allowed to be activated only if a boundary is detected at the previous time step of the layer $\ell$ (see Eq. \[eq:top\_down\]). This makes the layer $\ell$ to be initialized with more long-term information after the boundary is detected and execute the FLUSH operation. In addition, the input from the lower layer ($\ell - 1$) becomes effective only when a boundary is detected at the current time step in the layer ($\ell - 1$) due to the binary gate $z_t^{\ell-1}$. Figure \[fig:gating\_mechanism\_and\_output\_module\] (left) shows the gating mechanism of the HM-LSTM at time step $t$.
Finally, the binary boundary state $z^\ell_t$ is obtained by: z\^\_t = f\_(\^\_t). For the binarization function $f_\mathtt{bound}: \eR\rightarrow \{0,1\}$, we can either use a deterministic step function: \[eq:step\_func\] z\^\_t
1 & \^\_t > 0.5\
0 & ,
or sample from a Bernoulli distribution $z^\ell_t \sim\text{Bernoulli}(\tilde{z}^\ell_t)$. Although this binary decision is a key to our model, it is usually difficult to use stochastic gradient descent to train such model with discrete decisions as it is not differentiable.
\[sec:proposed\]
![Left: The gating mechanism of the HM-RNN. Right: The output module when $L=3$.[]{data-label="fig:gating_mechanism_and_output_module"}](one_step.pdf){height="0.5\columnwidth"}
![Left: The gating mechanism of the HM-RNN. Right: The output module when $L=3$.[]{data-label="fig:gating_mechanism_and_output_module"}](output_module.pdf){height="0.5\columnwidth"}
Computing Gradient of Boundary Detector
---------------------------------------
Training neural networks with discrete variables requires more efforts since the standard backpropagation is no longer applicable due to the non-differentiability. Among a few methods for training a neural network with discrete variables such as the REINFORCE [@williams1992simple; @mnih2014neural] and the straight-through estimator [@hinton2012coursera; @bengio2013estimating], we use the straight-through estimator to train our model. The straight-through estimator is a biased estimator because the non-differentiable function used in the forward pass (i.e., the step function in our case) is replaced by a differentiable function during the backward pass (i.e., the hard sigmoid function in our case). The straight-through estimator, however, is much simpler and often works more efficiently in practice than other unbiased but high-variance estimators such as the REINFORCE. The straight-through estimator has also been used in @courbariaux2016binarized and @vezhnevets2016strategic.
**The Slope Annealing Trick**. In our experiment, we use the slope annealing trick to reduce the bias of the straight-through estimator. The idea is to reduce the discrepancy between the two functions used during the forward pass and the backward pass. That is, by gradually increasing the slope $a$ of the hard sigmoid function, we make the hard sigmoid be close to the step function. Note that starting with a high slope value from the beginning can make the training difficult while it is more applicable later when the model parameters become more stable. In our experiments, starting from slope $a=1$, we slowly increase the slope until it reaches a threshold with an appropriate scheduling.
Experiments {#sec:experiments}
===========
We evaluate the proposed model on two tasks, character-level language modelling and handwriting sequence generation. Character-level language modelling is a representative example of discrete sequence modelling, where the discrete symbols form a distinct hierarchical multiscale structure. The performance on real-valued sequences is tested on the handwriting sequence generation in which a relatively clear hierarchical multiscale structure exists compared to other data such as speech signals.
Character-Level Language Modelling
----------------------------------
A sequence modelling task aims at learning the probability distribution over sequences by minimizing the negative log-likelihood of the training sequences: \[eq:sequence\_task\] \_-\^[N]{}\_[n=1]{}\^[T\^n]{}\_[t=1]{}p(x\^n\_tx\^n\_[<t]{};), where $\theta$ is the model parameter, $N$ is the number of training sequences, and $T^n$ is the length of the $n$-th sequence. A symbol at time $t$ of sequence $n$ is denoted by $x_t^n$, and $x^n_{<t}$ denotes all previous symbols at time $t$. We evaluate our model on three benchmark text corpora: (1) Penn Treebank, (2) Text8 and (3) Hutter Prize Wikipedia. We use the bits-per-character (BPC), $\eE[-\log_2 p(x_{t+1} \mid x_{\leq t})]$, as the evaluation metric.
#### Model
We use a model consisting of an input embedding layer, an RNN module and an output module. The input embedding layer maps each input symbol into $128$-dimensional continuous vector without using any non-linearity. The RNN module is the HM-LSTM, described in Section \[sec:model\], with three layers. The output module is a feedforward neural network with two layers, an output embedding layer and a softmax layer. Figure \[fig:gating\_mechanism\_and\_output\_module\] (right) shows a diagram of the output module. At each time step, the output embedding layer receives the hidden states of the three RNN layers as input. In order to adaptively control the importance of each layer at each time step, we also introduce three scalar gating units $g_t^\ell \in \eR$ to each of the layer outputs: \[eq:scaling\_unit\] g\^\_t = (\^), where $\bw^\ell\in\mathbb{R}^{\sum^L_{\ell=1} dim(\bh^\ell)}$ is the weight parameter. The output embedding $\bh_t^\text{e}$ is computed by: \[eq:output\_module\] \^\_t = (\^[L]{}\_[=1]{}g\^\_t W\_\^\^\_t), where $L=3$ and $\mathtt{ReLU}(x)=\max(0, x)$ [@nair2010rectified]. Finally, the probability distribution for the next target character is computed by the softmax function, [$\mathtt{softmax}(x_j)=\frac{e^{x_j}}{\sum^K_{k=1} e^{x_k}}$]{}, where each output class is a character.
#### Penn Treebank
We process the Penn Treebank dataset [@marcus1993building] by following the procedure introduced in @mikolov2012subword. Each update is done by using a mini-batch of $64$ examples of length $100$ to prevent the memory overflow problem when unfolding the RNN in time for backpropagation. The last hidden state of a sequence is used to initialize the hidden state of the next sequence to approximate the full backpropagation. We train the model using Adam [@kingma2014adam] with an initial learning rate of $0.002$. We divide the learning rate by a factor of $50$ when the validation negative log-likelihood stopped decreasing. The norm of the gradient is clipped with a threshold of $1$ [@mikolov2010recurrent; @pascanu2012difficulty]. We also apply layer normalization [@ba2016layer] to our models. For all of the character-level language modelling experiments, we apply the same procedure, but only change the number of hidden units, mini-batch size and the initial learning rate.
For the Penn Treebank dataset, we use $512$ units in each layer of the HM-LSTM and for the output embedding layer. In Table \[tab:ptb\_wiki\_bpc\] (left), we compare the test BPCs of four variants of our model to other baseline models. Note that the HM-LSTM using the step function for the hard boundary decision outperforms the others using either [*sampling*]{} or [*soft*]{} boundary decision (i.e., hard sigmoid). The test BPC is further improved with the slope annealing trick, which reduces the bias of the straight-through estimator. We increased the slope $a$ with the following schedule [$a=\min\left(5, 1+0.04\cdot N_{epoch}\right)$]{}, where $N_{epoch}$ is the maximum number of epochs. The HM-LSTM achieves test BPC score of $1.24$. For the remaining tasks, we fixed the hard boundary decision using the step function without slope annealing due to the difficulty of finding a good annealing schedule on large-scale datasets.
#### Text8
The Text8 dataset [@mahoney2009large] consists of $100$M characters extracted from the Wikipedia corpus. Text8 contains only alphabets and spaces, and thus we have total 27 symbols. In order to compare with other previous works, we follow the data splits used in @mikolov2012subword. We use $1024$ units for each HM-LSTM layer and $2048$ units for the output embedding layer. The mini-batch size and the initial learning rate are set to $128$ and $0.001$, respectively. The results are shown in Table \[tab:Text8\_bpc\]. The HM-LSTM obtains the state-of-the-art test BPC $1.29$.
#### Hutter Prize Wikipedia
The Hutter Prize Wikipedia (`enwik8`) dataset [@Hutter2012] contains $205$ symbols including XML markups and special characters. We follow the data splits used in @graves2013generating where the first 90M characters are used to train the model, the next 5M characters for validation, and the remainders for the test set. We use the same model size, mini-batch size and the initial learning rate as in the Text8. In Table \[tab:ptb\_wiki\_bpc\] (right), we show the HM-LSTM achieving the test BPC $1.32$, which is a tie with the state-of-the-art result among the neural models. Although the neural models, show remarkable performances, their compression performance is still behind the best models such as PAQ8hp12 [@mahoney2005adaptive] and decomp8 [@mahoney2009large].
#### Visualizing Learned Hierarchical Multiscale Structure
![Hierarchical multiscale structure in the Wikipedia dataset captured by the boundary detectors of the HM-LSTM.[]{data-label="fig:wiki_bound"}](wiki_boundary.pdf){width="1.1\columnwidth"}
In Figure \[fig:wiki\_bound\] and \[fig:ptb\_hidden\], we visualize the boundaries detected by the boundary detectors of the HM-LSTM while reading a character sequence of total length $270$ taken from the validation set of either the Penn Treebank or Hutter Prize Wikipedia dataset. Due to the page width limit, the figure contains the sequence partitioned into three segments of length $90$. The white blocks indicate boundaries $z^{\ell}_t=1$ while the black blocks indicate the non-boundaries $z^{\ell}_t=0$.
Interestingly in both figures, we can observe that the boundary detector of the first layer, $z^1$, tends to be turned on when it sees a space or after it sees a space, which is a reasonable breakpoint to separate between words. This is somewhat surprising because the model self-organizes this structure without any explicit boundary information. In Figure \[fig:wiki\_bound\], we observe that the $z^1$ tends to detect the boundaries of the words but also fires within the words, where the $z^2$ tends to fire when it sees either an end of a word or $2,3$-grams. In Figure \[fig:ptb\_hidden\], we also see flushing in the middle of a word, e.g., “tele-FLUSH-phone”. Note that “tele” is a prefix after which a various number of postfixes can follow. From these, it seems that the model uses to some extent the concept of *surprise* to learn the boundary. Although interpretation of the second layer boundaries is not as apparent as the first layer boundaries, it seems to segment at reasonable semantic / syntactic boundaries, e.g., “consumers may” - “want to move their telephones a” - “little closer to the tv set <unk>”, and so on.
Another remarkable point is the fact that we do not pose any constraint on the number of boundaries that the model can fire up. The model, however, learns that it is more beneficial to delay the information ejection to some extent. This is somewhat counterintuitive because it might look more beneficial to feed the fresh update to the upper layers at every time step without any delay. We conjecture the reason that the model works in this way is due to the FLUSH operation that poses an implicit constraint on the frequency of boundary detection, because it contains both a reward (feeding fresh information to upper layers) and a penalty (erasing accumulated information). The model finds an optimal balance between the reward and the penalty.
To understand the update mechanism more intuitively, in Figure \[fig:ptb\_hidden\], we also depict the heatmap of the $\ell^2$-norm of the hidden states along with the states of the boundary detectors. As we expect, we can see that there is no change in the norm value within segments due to the COPY operation. Also, the color of $\|\bh^1\|$ changes quickly (at every time step) because there is no COPY operation in the first layer. The color of $\|\bh^2\|$ changes less frequently based on the states of $z^1_t$ and $z^2_{t-1}$. The color of $\|\bh^3\|$ changes even slowly, i.e., only when $z^2_t=1$.
![The $\ell^2$-norm of the hidden states shown together with the states of the boundary detectors of the HM-LSTM.[]{data-label="fig:ptb_hidden"}](hidden_norm.pdf){width="1.1\columnwidth"}
A notable advantage of the proposed architecture is that the internal process of the RNN becomes more interpretable. For example, we can substitute the states of $z^1_t$ and $z^2_{t-1}$ into Eq. \[eq:update\_rule\] and infer which operation among the UPDATE, COPY and FLUSH was applied to the second layer at time step $t$. We can also inspect the update frequencies of the layers simply by counting how many UPDATE and FLUSH operations were made in each layer. For example in Figure \[fig:ptb\_hidden\], we see that the first layer updates at every time step (which is $270$ UPDATE operations), the second layer updates $56$ times, and only $9$ updates has made in the third layer. Note that, by design, the first layer performs UPDATE operation at every time step and then the number of UPDATE operations decreases as the layer level increases. In this example, the total number of updates is $335$ for the HM-LSTM which is $60\%$ of reduction from the $810$ updates of the standard RNN architecture.
Handwriting Sequence Generation
-------------------------------
We extend the evaluation of the HM-LSTM to a real-valued sequence modelling task using IAM-OnDB [@liwicki2005iam] dataset. The IAM-OnDB dataset consists of $12,179$ handwriting examples, each of which is a sequence of $(x,y)$ coordinate and a binary indicator $p$ for pen-tip location, giving us $(x_{1:T^n},y_{1:T^n}, p_{1:T^n})$, where $n$ is an index of a sequence. At each time step, the model receives $(x_t,y_t,p_t)$, and the goal is to predict $(x_{t+1},y_{t+1},p_{t+1})$. The pen-up ($p_t=1$) indicates an end of a stroke, and the pen-down ($p_t=0$) indicates that a stroke is in progress. There is usually a large shift in the $(x,y)$ coordinate to start a new stroke after the pen-up happens. We remove all sequences whose length is shorter than $300$. This leaves us $10,465$ sequences for training, $581$ for validation, $582$ for test. The average length of the sequences is $648$. We normalize the range of the $(x,y)$ coordinates separately with the mean and standard deviation obtained from the training set. We use the mini-batch size of $32$, and the initial learning rate is set to $0.0003$.
We use the same model architecture as used in the character-level language model, except that the output layer is modified to predict real-valued outputs. We use the mixture density network as the output layer following @graves2013generating, and use $400$ units for each HM-LSTM layer and for the output embedding layer. In Table \[tab:iamondb\_nll\], we compare the log-likelihood averaged over the test sequences of the IAM-OnDB dataset. We observe that the HM-LSTM outperforms the standard LSTM. The slope annealing trick further improves the test log-likelihood of the HM-LSTM into $1167$ in our setting. In this experiment, we increased the slope $a$ with the following schedule [$a=\min\left(3, 1+0.004\cdot N_{epoch}\right)$]{}. In Figure \[fig:iamondb\_bound\], we let the HM-LSTM to read a randomly picked validation sequence and present the visualization of handwriting examples by segments based on either the states of $z^2$ or the states of pen-tip location[^4].
![The visualization by segments based on either the given pen-tip location or states of the $z^2$.[]{data-label="fig:iamondb_bound"}](handwriting_boundary.pdf){width="1.\columnwidth"}
Conclusion {#sec:conclusion}
==========
In this paper, we proposed the HM-RNN that can capture the latent hierarchical structure of the sequences. We introduced three types of operations to the RNN, which are the COPY, UPDATE and FLUSH operations. In order to implement these operations, we introduced a set of binary variables and a novel update rule that is dependent on the states of these binary variables. Each binary variable is learned to find segments at its level, therefore, we call this binary variable, a boundary detector. On the character-level language modelling, the HM-LSTM achieved state-of-the-art result on the Text8 dataset and comparable results to the state-of-the-art results on the Penn Treebank and Hutter Prize Wikipedia datasets. Also, the HM-LSTM outperformed the standard LSTM on the handwriting sequence generation. Our results and analysis suggest that the proposed HM-RNN can discover the latent hierarchical structure of the sequences and can learn efficient hierarchical multiscale representation that leads to better generalization performance.
Acknowledgments {#sec:ack .unnumbered}
===============
The authors would like to thank Alex Graves, Tom Schaul and Hado van Hasselt for their fruitful comments and discussion. We acknowledge the support of the following agencies for research funding and computing support: Ubisoft, Samsung, IBM, Facebook, Google, Microsoft, NSERC, Calcul Québec, Compute Canada, the Canada Research Chairs and CIFAR. The authors thank the developers of Theano [@team2016theano]. JC would like to thank Arnaud Bergenon and Frédéric Bastien for their technical support. JC would also like to thank Guillaume Alain, Kyle Kastner and David Ha for providing us useful pieces of code.
[^1]: Yoshua Bengio is CIFAR Senior Fellow.
[^2]: The acronym NARX stands for Non-linear Auto-Regressive model with eXogenous inputs.
[^3]: $\tilde{z}_t^\ell$ can also be implemented as a function of $\bh_t^\ell$, e.g., $\tilde{z}_t^\ell=\mathtt{hard\;sigm}(U\bh_t^\ell)$.
[^4]: The plot function could be found at [blog.otoro.net/2015/12/12/handwriting-generation-demo-in-tensorflow/](blog.otoro.net/2015/12/12/handwriting-generation-demo-in-tensorflow/).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have studied the radial distribution of the early (E/S0) and late (S/Irr) types of satellites around bright host galaxies. We made a volume-limited sample of 4,986 satellites brighter than $M_r = -18.0$ associated with 2,254 hosts brighter than $M_r =-19.0$ from the SDSS DR5 sample. The morphology of satellites is determined by an automated morphology classifier, but the host galaxies are visually classified. We found segregation of satellite morphology as a function of the projected distance from the host galaxy. The amplitude and shape of the early-type satellite fraction profile are found to depend on the host luminosity. This is the morphology-radius/density relation at the galactic scale. There is a strong tendency for morphology conformity between the host galaxy and its satellites. The early-type fraction of satellites hosted by early-type galaxies is systematically larger than that of late-type hosts, and is a strong function of the distance from the host galaxies. Fainter satellites are more vulnerable to the morphology transformation effects of hosts. Dependence of satellite morphology on the large-scale background density was detected. The fraction of early-type satellites increases in high density regions for both early and late-type hosts. It is argued that the conformity in morphology of galactic satellite system is mainly originated by the hydrodynamical and radiative effects of hosts on satellites.'
author:
- |
H. B. Ann$^{1}$[^1], Changbom Park$^{2}$[^2], Yun-Young Choi$^{3}$[^3]\
$^{1}$ Division of Science Education, Pusan National University, Busan 609-735, Korea,\
$^{2}$ Korea Institute for Advanced Study, Dongdaemun-gu, Seoul 130-722, Korea\
$^{3}$ Astrophysical Research Center for the Structure and Evolution of the Cosmos, Sejong University, Seoul 143-747, Korea
date: 'Accepted . Received ; in original form.. '
title: 'Galactic Satellite Systems: Radial Distribution and Environment Dependence of Galaxy Morphology'
---
\[firstpage\]
galaxies: general – galaxies: formation – galaxies: interactions – methods: observational
Introduction
============
Morphology reflects the integral property of a galaxy, such as stellar populations, gas content, and dynamical structures. Its origin is one of the central problems in the study of galaxy formation and evolution. If a galaxy remains isolated after its formation, all of its physical properties would be entirely determined by the initial conditions of the proto-galactic cloud and by the subsequent internal evolution. But, it seems unlikely because galaxies are believed to form through a series of minor/major mergers. In fact, the isolated bright galaxies in high density regions are more likely to be recently merged ones and the morphology of galaxies contains imprints of interaction with environment in addition to the formation process (Park, Gott, & Choi 2008).
There is observational evidence that shows an intimate correlation between the morphology of the central galaxy and its neighbors (Wirth 1983; Hickson et al. 1984; Ramella et al. 1987; Osmond & Ponman 2004; Weinmann et al. 2006; Park et al. 2008). Recent analysis of the morphology of Sloan Digital Sky Survey (SDSS; York et al. 2000) galaxies by Park et al. (2007, 2008) showed that galaxy morphology does depend on the large-scale background density but the role of the nearest neighbor is more decisive. The critical roles of the closest neighbor in determining galaxy morphology appear as the galactic conformity (Weinmann et al. 2006) between a galaxy and its neighbors. Galaxy morphology also depends on luminosity in that galaxy morphology is more likely to be early type for brighter galaxies. Since bright galaxies mainly live in high density regions through the luminosity-density relation, it appears that early types are more prevalent at high densities. Satellite systems are good places to inspect the environmental dependence of galaxy morphology and to study the galaxy formation process since they are abundant and very localized systems with a size of less than 1 Mpc. Most of the previous studies of satellite galaxies were focused on the radial distribution of satellite galaxies (Sales & Lambas 2005; van den Bosch et al. 2005; Chen et al. 2006), dark matter halo (McKay et al. 2002; Prada et al. 2003; van den Bosch 2004), and angular distributions (Zaritsky et al. 1997; Sales & Lambas 2004; Zentner et al. 2005; Yang et al. 2006; Libeskind et al. 2007; Bailin et al. 2007; Kang et al. 2007; Sales et al. 2007). The morphology of satellite galaxies is also an observable parameter that is directly related to formation and evolution of galaxy.
The purpose of the present paper is to study the relation between the morphology of satellite galaxies and the local environment such as the host morphology and background density. We used large and homogeneous morphology samples made by both visual and automated classifications. We will see a tight correlation between the host and satellites morphologies. The satellite systems in our study are hosted by the typical bright galaxies, and are not in general large groups or clusters of galaxies. Our host sample is dominated by the $L_*$ galaxies, and their satellites are fainter by about two magnitudes.
Data
====
Isolated satellite systems
--------------------------
The basic source of data is the large-scale structure sample (LSS), DR4plus, from the New York University Value-Added Galaxy Catalogue (NYU-VAGC; Blanton et al. 2005) which is a subset of the SDSS Data Release 5 (DR5; Adelman-McCarthy et al. 2007). The primary sample of galaxies used here is a subset of the LSS-DR4plus, which includes Main galaxies (Strauss et al. 2002) with extinction corrected apparent Petrosian $r$-magnitudes in the range $14.5 \leq r_{\rm Pet} < 17.77$ and redshifts in the range $0.001 < z < 0.5$. Our survey region covers 4464 deg$^2$, which is shown in Figure 1 of Park et al. (2007). To this primary sample, we added the galaxies brighter than the bright limit ($r_{\rm Pet}=14.5$) of the sample. Various existing redshift catalogs are searched for the redshifts of the bright galaxies with no spectrum. The catalogs include RC3 (de Vaucouleurs et al. 1991), Catalog of Nearby Galaxies (Tully & Fisher 1988) and Updated Zwicky catalog [^4] (ZCAT 2000 Version). In case of no measured redshift even in these catalogs, we used the redshift taken from NASA/IPAC Extragalactic Database[^5] (NED) when available. We added 5,503 bright galaxies to the primary sample. The final data set consists of 370,789 galaxies with known redshift and photometry. Throughout this paper, we use a flat $\Lambda$CDM cosmology with density parameters $\Omega _m = 0.27$ and $\Omega_{\Lambda} = 0.73$.
To search for isolated satellite systems we take two steps. We first look for isolated galaxies in a volume-limited sample of galaxies brighter than the $r$-band absolute magnitude $M_r = -19.0
+ 5 {\rm log} h$ (hereafter we are going to drop the term $5 {\rm log} h$) and with redshifts between 0.02 and 0.04724. The lower redshift limit is chosen to make our sample as complete as possible since galaxies with $z < 0.02$ in the SDSS seems to be incomplete even though we supplemented bright galaxies (Park et al. 2007). The comoving space number density of galaxies is approximately constant in the radial direction at $z>0.02$, but drops significantly at $z<0.02$. The upper limit of $z=0.04724$ corresponds to the survey limits of $r=17.77$ for a galaxy with $M_r = -18.0$.
A target galaxy is isolated if the projected separation to its nearest neighbor galaxy is larger than the virial radii of both galaxies. The neighbors of a target galaxy with $M_r$ are those with absolute magnitude brighter than $M_r +1.0$, and velocity difference less than 1,000 km s$^{-1}$. We have also used the most influential neighbor instead of the nearest one for a comparison in the measurement of the projected separation $r_p$. Our results are basically the same for these two choices. The most influential neighbor is the neighbor which induces the highest local density at the location of the target galaxy. Given $r_p$ between them, we calculate the local mass density due to the neighbor with luminosity $L_n$ by $$\rho_{n}= 3\gamma_n L_n /4\pi r_p^3,$$ where we adopt the mass-to-light ratios $\gamma_n=2$ for early types and 1 for late types. This choice is based on the morphology-specific central stellar velocity dispersion and on the pairwise peculiar velocity difference of early and late-type galaxies with their neighbors (see Park et al. 2008 for more details). We define the virial radius of each galaxy as the radius where the mean density within the sphere centered at the galaxy given by equation (1) becomes the virialized density, which is set to $766{\bar\rho}$ (see section 3.1 of Park et al. 2008). The mean mass density is obtained from ${\bar{\rho}}= \sum_{all} \gamma_i L_i/V $ where the summation is over all galaxies in our full volume-limited sample of volume $V$ with the absolute magnitude constraint $M_r <-18.0$. An early or late-type galaxy with $M_r=-20$ has virial radius of 300 or 240 $h^{-1}$kpc, respectively. For those with $M_r = -20.5$, the virial radii are 350 and 280 $h^{-1}$kpc, respectively.
We found 8,883 isolated galaxies in our volume-limited sample. They are physically isolated ones in the sense that they are not hydrodynamically interacting with neighbors. In all previous studies isolation is determined by using a pre-selected fixed radius ignoring the physical size of individual galaxies involved. A blindly large radius of the isolation boundary results in too small sample size, while any fixed value in the right range results in contamination in the sample with interacting galaxies added.
Once the bright isolated galaxies are found, we search for satellites associated with them. We limit the satellite candidates only to galaxies with $M_r$ brighter than $-18.0$, a limit one magnitude fainter than that of the host candidates. This choice gives us a uniform and complete selection of satellites for host galaxies also uniformly and completely selected across our sample volume (see Figure 1 below). At each location of the isolated galaxies we search for galaxies with velocity difference less than 500 km s$^{-1}$, absolute magnitude more than one magnitude fainter (but brighter than $-18.0$), and the projected separation less than the smaller of 1 $h^{-1}$Mpc and $r_p$(neighbor)$- r_{vir}$(neighbor), where the letter is the difference between the host-neighbor separation and the neighbor’s virial radius. We used the Petrosian $g$-band absolute magnitude for satellite identification because $g$ magnitude is most similar to the $B$ magnitude that is used for some bright galaxies whose SDSS photometry is too poor to be used without correction.
Among the 8,883 isolated galaxies, 2,254 have satellites, and the total number of satellites belonging to these systems is 4,986. Figure 1 shows the distributions of the host galaxies (red points) and satellites (blue). The host distribution shows our sample boundaries in redshift. Some of the satellites lie beyond the redshift boundaries because we allowed 500 km$s^{-1}$ difference in radial velocity in the search for satellites. The median absolute magnitude of the hosts is $M_r = -20.47$, which is very close to that of the $L_*$ SDSS galaxies (Choi et al. 2007). Therefore, the host galaxies of our satellite systems are dominated by normal bright galaxies, and are not in general the central $cD$ galaxies holding the bright galaxies as satellites. The median absolute magnitude of our satellites is $-18.67$. So they are not dwarf galaxies, but subluminous bright galaxies typically 1.8 magnitude fainter than their hosts. Since both our hosts and satellites are selected uniformly in the absolute magnitude space, our study of satellite morphology is unbiased against host and satellite luminosity.
Morphology
----------
We classify the morphology of host galaxies by the visual inspection because visual classification is accurate for bright galaxies. For visual classification, ellipticals(E) and lenticulars(S0) as well as spiral(S) and irregulars(Irr) are distinguished, but for better statistics we categorized E and S0 galaxies as early types, and S and Irr galaxies as late types in the present analysis. We mainly employed the automated classifier of Park & Choi (2005) for satellites. This classifier divides galaxies into early and late types based on their location in the three-dimensional parameter space of $u-r$ color, $g-i$ color gradient, and the $i$-band concentration index. The classification boundaries in the parameter space are chosen by using a large training set of galaxies with known morphology. All of the satellites are visually checked. But the visual classification is used only as a complementary one, especially for relatively bright satellites or for those undergoing close interactions or mergers. This is because in most cases ($>$ 90%) the visual and automated classifications of satellites agree with each other, and because for the very faint galaxies close to the faint limit of the sample, it is not certain whether or not the visual classification is on average any better than the result of the automated classification.
PROPERTIES OF SATELLITE SYSTEMS
===============================
Morphology and radial distribution of satellites
------------------------------------------------
We measured the early-type fraction $f(E_{s})$ and surface number density $\Sigma(E{_s})$ of satellite galaxies as a function of projected distance ($r_{p}$) from the host galaxies. The top panel of Figure 2 shows the early-type fraction of satellites associated with our isolated early-type hosts $f(E_{s}|E_{h})$ (filled circles) and isolated late-type hosts $f(E_{s}|L_{h})$. The innermost bin is $r_{p} < 37.8 h^{-1}$ kpc, which corresponds to the fiber collision radius of $55^{\prime \prime}$ at the outer boundary ($z=0.04724$) of our volume-limited sample.
It can be noted that $f(E_{s}|E_{h})$ is significantly higher than $f(E_{s}|L_{h})$ at least out to about 350 $h^{-1}$kpc, which is roughly the virial radius of the typical early-type host galaxies analyzed in this study. This result means that the morphology of satellites tends to be similar to that of hosts. It demonstrates the morphology-radius relation at the galactic scales. A similar finding was reported by Weinmann et al. (2006) for galaxies in groups and clusters, and by Park et al. (2007, 2008) for galaxy pairs. For late-type hosts, the early-type satellite fraction increases very slowly as satellites approach their hosts. Some of this effect must be due to the morphology-luminosity relation. The early-type hosts are in general brighter than the late-type hosts, and correspondingly the satellites of our early-type hosts are also on average brighter than those of late-type hosts due to our satellite finding process, i.e. more than one magnitude fainter relative to the host. Because of the morphology-luminosity relation, the morphology of early-type hosts’ satellites is in general earlier than that of late-type hosts’ satellites even if there is no direct physical influence of the host on satellites. We do not think this is the main reason for the host-satellite morphology correlation we found because the satellite morphology is a very strong function of host-satellite separation in early-type host systems and the early-type satellite fractions for early and late-type hosts start to merge at $r_p \sim 1h^{-1}$Mpc.
Irrelevance of the morphology-luminosity relation to our findings can be also demonstrated by the early-type satellite fraction plot drawn for hosts with fixed luminosity. The middle and bottom panels of Figure 2 show the early-type satellite fractions for host galaxies brighter than $M_r=-20.5$ and fainter than $-21.0$, respectively. We allowed an overlap in $M_r$ to decrease the statistical fluctuations. We also divided satellites into a subset more than $\Delta M_g=1.9$ magnitude fainter than the host and a subset more than 1.0 but less than 1.9 magnitude fainter. Drawn are the four cases of early-type hosts ($E_h$) and satellites with $\Delta M_g >1.9$ (filled circles, solid line), $E_h$ and satellites with $1.0<\Delta M_g<1.9$ (open circles, dashed line), $L_h$ and satellites with $\Delta M_g >1.9$ (stars, solid line), and $L_h$ and satellites with $1.0<\Delta M_g<1.9$ (crosses, dashed line).
The satellites with smaller $\Delta M_g$ are on average brighter than those with larger $\Delta M_g$, and are more likely to be early-types in accordance with the morphology-luminosity relation. The mean level of $f(E_s )$ at very large $r_p$ is indeed higher for smaller $\Delta M_g$ satellites in both middle and bottom panels of Figure 2. Once we subtract the dependence of this asymptotic value on host and satellite luminosity from these figures, interesting dependence of $f(E_s )$ on $r_p$ and host morphology becomes evident. The fraction of early-type satellites associated with early-type hosts, $f(E_s | E_h)$, depends on $r_p$ more sensitively for fainter satellites (compare the open and filled circles). This is true for both relatively bright (middle panel) and faint (bottom panel) hosts. It can be also noted from the middle and bottom panels that the outer boundary of the region of early-type host influence is farther for brighter hosts. The net effects of the late-type hosts on satellite morphology seem insignificant.
The satellites of early-type hosts are likely to be deprived of their cold gas through the hydrodynamic and radiative interactions with the X-ray emitting hot gas of their host. The satellites of late-type hosts are in principle able to get cold gas from their hosts although the hot gas in the halo of late-type hosts can also remove the cold gas in their satellites. Based on a detailed study of morphology-environment relation of galaxy pairs Park et al. (2008) concluded that the galaxy morphology-local density relation is mainly due to the interaction between nearest neighbor galaxies. When galaxies are closer than their virial radii, they start to interact hydrodynamically and this causes the conformity in morphology of close galaxy pairs. The present results support their scenario, and this seems to be the origin of the morphology conformity in galactic satellite systems.
One major difference between our result and that of Park et al. is that the satellite morphology does not tend to be of late type as satellites approach late-type host. The galaxy pairs in Park et al.’s sample are dominated by those with similar luminosity and therefore their interaction can affect physical properties of both galaxies significantly. On the other hand, in the current analysis satellites are typically 1.8 magnitudes fainter than hosts, and the influence is largely lopsided from hosts to satellites. The slight rising tendency of $f(E_{s})$ very close to late-type hosts can be because satellites are suffering from cold gas stripping and ionization by the host halo gas, but can not actively catch the cold gas from their hosts as efficiently as companion galaxies having luminosity similar to the hosts.
The slopes of the surface density profiles also reflects the physical effects of their host galaxies on satellites. Figure 3 presents the satellite surface number density profiles for early (top panel) and late-type hosts. The ratio of two profiles in each panel gives the morphology fraction in the top panel of Figure 2.
It demonstrates the number density profile of satellites critically depends on both host and satellite morphology. The surface density profiles show why $f(E_{s}|E_{h})$ decreases more rapidly than $f(E_{s}|L_{h})$. It is due to the dramatic drop in the surface density of early-type satellites hosted by early-type hosts $\Sigma(E_{s}|E_{h})$, and to the slower drop of that of late-type satellites hosted by early-type hosts $\Sigma(L_{s}|E_{h})$. This makes late-type satellites dominant in the satellite systems of early-type hosts at $r_{p} > 100\sim 200$ $h^{-1}$ kpc, where the exact location depending on the host luminosity and the host-satellite magnitude difference (see Figure 2). In the systems hosted by late-type galaxies, late-type satellites are dominant at all $r_p$. The shapes of the surface density profiles of both early and late-type satellites, $\Sigma(E_{s}|L_{h})$ and $\Sigma(L_{s}|L_{h})$, are similar to each other, making their ratio roughly constant of $r_p$. At large separations, $r_p > 600 h^{-1}$kpc, the surface densities satellites belonging to early-type (upper panel) and late-type (lower panel) hosts approach roughly the same ratio, resulting in $f(E_s)\approx 0.2$. This seems the field value of galaxy morphology for galaxies having absolute magnitudes similar to those of the satellites in our sample.
The morphology fraction shown in Figure 2 is the result of projection of the three dimensional distribution on the sky. In order to get a rough idea on the central morphology fraction we try to deproject the profile as follows. We assume the radial number density of each of early and late-type satellites follows a power-low, $\rho(r)=\rho_o (r/r_{o})^{-\gamma}$. Then the projected density follows the form (Binney & Tremaine 1987), $$\Sigma(r_{p})=\rho_{o}r_{o}^{\gamma} (-{1\over2})!({{\gamma-3}\over{2}})!/
r_{p}^{\gamma-1}({{\gamma-2}\over{2}})!.$$ The parameters in the fraction are obtained from a least-square fit to the inner-most three points shown in Figure 3 for each case of host and satellite morphology. Only two parameters are free. We found the slope of the three-dimensional profile is $-1.8\sim-1.9$ at $r<200 h^{-1} kpc$ except for the late-type satellite associated with early-type host case, which has about -1.5. The true fraction of early-type satellites very close to early-type hosts is found to reach about 0.71 and 0.78 at $r=30$ and 10$h^{-1}$kpcs, respectively. On the other hand, the fraction for late-type hosts in 3D is nearly the same as that shown in Figure 2 because the slope of radial density profile is almost independent of satellite morphology in this case.
Background density dependence
-----------------------------
As argued in the previous sections, the morphology conformity in galactic satellite systems seems to be due to the local effects of hosts on their satellites. However, the galactic conformity can be affected by the global environment as well as local one. Park et al. (2008) showed that, even though the morphology of galaxies depends mainly on luminosity and the small-scale environment due to the nearest neighbor, it also depends on the large-scale background density. The dependence of galaxy morphology on the large-scale density was found even when both the luminosity of the target galaxy and the environment due to the nearest neighbor were fixed. This was explained by the dependence of the hot halo gas of galaxies on the large-scale density. In this section we look for a similar effect on galactic satellites.
We used the galaxy number density estimator defined by 20 nearest $L_*$ galaxies with $-20.0 > M_r > -21.0$ drawn from the full volume-limited sample $$\rho_{20}/{\bar{\rho}}= \sum_{i=1}^{20} W_i(|{\bf x}_i -
{\bf x}|)/{\bar\rho},$$ where $W(r)$ is a spline-kernel weight and $\bar{\rho}$ is the mean number density of the $L_*$ galaxies in the SDSS. This choice is the same as those used by Park et al. (2007). The median value of the effective Gaussian smoothing scale, corresponding to the adaptive spline smoothing, is $4.7 h^{-1}$ Mpc.
The top panel of Figure 4 shows the distributions of the large-scale density for early-type (solid line) and late-type (dotted) hosts. It can be seen that their distributions are nearly the same except for the highest densiy bin even though the early-type galaxies are in general preferentially located at higher densities. This may be because the isolation constraint on hosts excluded more early-type galaxies than late-types in high density regions. However, as can be seen in the bottom panel of Figure 4, our isolated host sample still respects the luminosity-density relation.
Figure 5 shows $f(E_{s})$ as a function of the projected separation from the hosts in different large-scale background density regions. We fixed the luminosity of host galaxies to $-20.5>M_{r}>-21.5$ to separate the luminosity effects from the background density effects on galaxy morphology. The large scale environment is divided into high and low density regions with $\rho_{20}/{\bar{\rho}} > 2.2$ and $< 2.2$, respectively, where $\rho_{20}/{\bar{\rho}} = 2.2$ is the median density for our isolated hosts.
It can be seen in Figure 5 that $f(E_{s})$ is higher in high density regions for both early and late-type hosts with fixed luminosity. The background density seems to play a definite role in determining the morphology of galactic satellites. The background density can directly affect the satellites, or affect them indirectly through the host whose properties depend statistically on the background density. Park et al. (2008) found that the early-type galaxies in high density regions have higher X-ray luminosity than those in low density regions even when their optical luminosity is the same. This means that the hot halo gas of early-type galaxies is hotter and denser at high densities. Taking into account this finding we interpret the background density dependence of the satellite morphology as due to the hydrodynamic and radiative effects of the hot gas of host galaxies on satellites. This is supported by the fact that, even though $f(E_s)$ is generally higher in high density regions in Figure 5, it is so only when satellites are close to their hosts and $f(E_s)$ at $r_p$ much larger than the host’s virial radius is rather independent of the background density. If the background density directly affect the morphology of satellites, the satellite morphology should depend on the background density at all host-satellite separations. It can be also noted that $f(E_{s})$ is higher for early-type hosts than for late-type hosts both in high and low density regions. Therefore, the conformity in morphology at galactic scales prevails in both high and low density environments.
The $f(E_{s})$ in high density environment decreases almost linearly with $r_{p}$ while $f(E_{s})$ in low density environment decreases nearly exponentially. It is surprising to see this background density dependence even if we fixed host morphology and luminosity. All $f(E_{s})$ seems to converge at $r_{p}>600$ $h^{-1}$ kpc. This is a scale a little larger than the virial radius of the host galaxies under consideration. The virial radii are about 280 and 350 $h^{-1}$kpc for late and early-type galaxies with $M_r = -20.5$. In the analysis of luminous galaxy pairs the dependence of galaxy morphology on the neighbor’s morphology appears at separations of $r_p \la r_{\rm vir}$ (Park et al. 2008). This was explained by the hydrodynamic interactions between the pairs within the virialized region.
Previous studies showed that the fraction of interlopers could be large at large $r_p$ (Prada et al. 2003) and that the interloper fraction depends on the color of the satellites, with interlopers being rare amongst the red satellites, but making up about half of the blue satellites. If our satellite samples were dominated by interlopers at large $r_p$, the difference in $f(E_s)$ in high and low density regions could be due simply to the interlopers which respect the morphology-density relation. However, Figure 5 shows that $f(E)$ converges to about 0.2 both in high and low density regions and both for early- and late-type host galaxies. Therefore, the satellites at large separations do not show the trends that are expected for the general background galaxies. This indicates our results are not significantly affected by interlopers.
Discussion and Conclusions
==========================
We have found the morphology-radius relation for galactic satellite systems. Early-type satellites are prevalent in the vicinity of early type hosts. The origin of the conformity in morphology is thought to be the hydrodynamic and radiative effects of hosts on satellites in addition to the tidal (gravitational) effects.
The satellite morphology is found to depend on the large-scale background density. In high density regions the early-type fraction of satellites decreases relatively slowly beyond the virial radius of the host galaxy. However, in low density regions the fraction of satellites with early morphological type drops sharply at separations of $r_p =50 \sim 200 h^{-1}$kpc for both early and late-type host systems. As we fixed the mass of host galaxies by fixing luminosity and morphology, this difference must be coming from non-gravitational effects. It is argued that the hot halo gas of the host galaxies is responsible for prevalence of early-type satellites in the vicinity of hosts, and that in high density regions the hot halo gas can be more confined by the ambient intergalactic medium and has higher density and temperature, which can better deplete the cold gas in satellites more efficiently.
The galactic conformity found from the present sample of satellite systems is not much affected by the detailed selection criteria of the satellites. The magnitude difference between host and satellites is not critical because we obtained similar results for the satellite systems defined by different magnitude differences. Using the most-influential neighbors instead of the nearest neighbors in identification of isolated hosts and satellites also did not make much difference. We also examined whether or not our results are affected by our isolation requirement for host galaxies, and found that all of our results qualitatively remain the same. We made exactly the same analysis for satellites defined for host galaxies which are not constrained to be isolated. In this analysis a galaxy becomes a satellite if it finds a host galaxy within $r_p=800 h^{-1}$kpc that is more than 2 magnitudes brighter and has velocity difference less than 500 km s$^{-1}$. If there is more than one such hosts, the closest one is chosen. Hosts are limited to $M_r <-20.0$, and satellites have $M_r < -18.0$. We found 8,353 satellites in 3,472 systems. We obtain basically the same results for these satellite systems as for the isolated ones but with much higher statistical significances. Therefore, our results are robust against various choices of parameters used to identify hosts and satellites. In the forthcoming paper we will study the shape and internal properties of satellites. Rather than dividing satellites into early and late types, we will adopt a new classification scheme that is more appropriate for the satellite galaxies. We found this is necessary because our satellite galaxies are fainter and located in the special environment given by the hosts compared to the normal bright galaxies for which the usual morphology classification schemes are developed.
Acknowledgments {#acknowledgments .unnumbered}
===============
H.B.A., Y.Y.C. and C.B.P. acknowledge the support of the Korea Science and Engineering Foundation (KOSEF) through the Astrophysical Research Center for the Structure and Evolution of the Cosmos (ARCSEC).
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. The Participating Institutions are the American Museum of Natural History, Astrophysical Institute Potsdam, University of Basel, University of Cambridge, Case Western Reserve University, University of Chicago, Drexel University, Fermilab, the Institute for Advanced Study, the Japan Participation Group, Johns Hopkins University, the Joint Institute for Nuclear Astrophysics, the Kavli Institute for Particle Astrophysics and Cosmology, the Korean Scientist Group, the Chinese Academy of Sciences (LAMOST), Los Alamos National Laboratory, the Max-Planck-Institute for Astronomy (MPIA), the Max-Planck-Institute for Astrophysics (MPA), New Mexico State University, Ohio State University, University of Pittsburgh, University of Portsmouth, Princeton University, the United States Naval Observatory, and the University of Washington.
Adelman-McCarthy, J. K., et al. 2007, ApJS, submitted Azzaro, M., Patiri, S. G., Prada, F., & Zentner, A. R. 2007, MNRAS, 376, 43 Bailin et al. 2007, astro-ph/07061350 , Binney, J., & Tremaine, S. 1987, Galactic Dynamics (Princeton: Princeton University Press), 41 Blanton, M. R., et al. 2005, ApJ, 129, 2562 Chen, J. et al. 2006, ApJ, 647, 86 Choi, Y.-Y., Park, C., & Vogeley, M. S. 2007, ApJ, 658, 884 de Vaucouleurs, G. et al. 1991, Third Reference Catalogue of Bright Galaxies, Volume 1-3, XII (Springer-Verlag) Hickson, P., Ninkov, Z., Huchra, J. P., & Mamon, G. A. 1984, Clusters and Groups of Galaxies, ed. F. Mardirossian, G. Giuricin, & M. Mezzetti (Kluwer Academic Publishers), 367 Kang, X. et al. 2007, MNRAS, 378, 1531 Libeskind et al. 2007, MNRAS, 374, 16 McKay, T. A. et al. 2002, ApJ, 571, 85 Osmond, J. P. F., & Ponman, T. J. 2004, MNRAS, 350, 1511 Park, C., & Choi, Y.-Y. 2005, ApJ, 635, L29 Park, C., Choi, Y.-Y., Vogeley, M. S., Gott, J. R., & Blanton, M. R. 2007, ApJ, 658, 898 Park, C., Gott, J. R. & Choi, Y.-Y. 2008, ApJ, 674, 784 Prada, F. et al. ApJ, 2003, 598, 260 Ramella, M. et al. 1987, A&A, 188, 1 Sales, L., & Lambas, D. G. 2004, MNRAS, 348, 1236 Sales, L., & Lambas, D. G. 2005, MNRAS, 356, 1045 Sales, L. et al. 2007, astro-ph/07062009 Strauss, M. A., et al. 2002, AJ, 124, 1810 Tully, R. B. & Fisher, J. R. 1988, Catalog of Nearby Galaxies, i by R. B. Tully, & J. R. Fisher, (Cambridge University Press), 224 van den Bosch, F. C., Yang, X., & Mo, H. J., & Norberg, P. 2005, MNRAS, 356, 1233 van den Bosch, F. C., Norberg, P., Mo, H. J., & Yang, X. 2004, MNRAS, 352, 1302 Weinmann, S. M., van den Bosch, F. C., Yang, X., & Mo, H. J. 2006, MNRAS, 366, 2 Wirth, A. 1983, ApJ, 274, 541 Yang, X. et al. 2006, MNRAS, 369, 1293 Zaritsky, D., Smith, R., Frenk, C., & White, S. D. 1997, ApJ, 478, 39 Zentner, A. R., Kravtsov, A. V., Gnedin, O. Y. & Klypin, A. A. ApJ, 629, 219 York, D., et al. 2000, AJ, 120, 1579
\[lastpage\]
[^1]: E-mail:[email protected]
[^2]: E-mail:[email protected]
[^3]: E-mail:[email protected]
[^4]: http://www.cfa.harvard.edu/\~huchra/zcat/zcom.htm
[^5]: http://nedwww.ipac.caltech.edu/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the adsorption behavior of Ga on (0001) GaN surfaces combining experimental specular reflection high-energy electron diffraction with theoretical investigations in the framework of a kinetic model for adsorption and [*ab initio*]{} calculations of energy parameters. Based on the experimental results we find that, for substrate temperatures and Ga fluxes typically used in molecular-beam epitaxy of GaN, *finite* equilibrium Ga surface coverages can be obtained. The measurement of a Ga/GaN adsorption isotherm allows the quantification of the equilibrium Ga surface coverage as a function of the impinging Ga flux. In particular, we show that a large range of Ga fluxes exists, where $2.5\pm 0.2$ monolayers (in terms of the GaN surface site density) of Ga are adsorbed on the GaN surface. We further demonstrate that the structure of this adsorbed Ga film is in good agreement with the laterally-contracted Ga bilayer model predicted to be most stable for strongly Ga-rich surfaces \[J. E. Northrup *et al.*, Phys. Rev. B **61**, 9932 (2000)\]. For lower Ga fluxes, a discontinuous transition to Ga monolayer equilibrium coverage is found, followed by a continuous decrease towards zero coverage; for higher Ga fluxes, Ga droplet formation is found, similar to what has been observed during Ga-rich GaN growth. The boundary fluxes limiting the region of 2.5 monolayers equilibrium Ga adsorption have been measured as a function of the GaN substrate temperature giving rise to a Ga/GaN adsorption phase diagram. The temperature dependence is discussed within an [*ab initio*]{} based growth model for adsorption taking into account the nucleation of Ga clusters. This model consistently explains recent contradictory results of the activation energy describing the critical Ga flux for the onset of Ga droplet formation during Ga-rich GaN growth \[B. Heying *et al.*, J. Appl. Phys. **88**, 1855 (2000); C. Adelmann *et al.*, J. Appl. Phys. **91**, 9638 (2002).\].'
author:
- Christoph Adelmann
- Julien Brault
- Guido Mula
- Bruno Daudin
- Liverios Lymperakis
- Jörg Neugebauer
title: 'Gallium adsorption on (0001) GaN surfaces'
---
Introduction
============
Due to its importance as a base material for the fabrication of optoelectronic devices in the blue and ultraviolet spectral region, GaN has been extensively studied in recent years. The comprehensive study of material and device properties[@Jain] has recently been complemented by an increasing number of studies concerning surface structures[@Hacke; @Fritsch; @Zywietz; @Smith; @Xue; @Munkholm; @Northrup1; @Wang; @Rapcewicz] and associated growth mechanisms, in particular for growth by plasma-assisted molecular-beam epitaxy (PAMBE) on the Ga-polar (0001) surface.[@Tarsa; @Widmann; @Xie; @Heying1; @Bourret; @Mula; @Adelmann]
One of the fundamental results of these studies is that GaN growth by PAMBE ought to be carried out under Ga-rich conditions in order to obtain a smooth surface morphology and optimized material properties. At low growth temperatures, however, this leads to Ga accumulation and droplet formation, which is detrimental to the GaN epilayer quality.[@Heying2; @Kruse] As a consequence, it was thought that optimum GaN growth conditions must be as close as possible to Ga/N stoichiometry. It has been recently observed that such Ga accumulation can be prevented when growing GaN at high temperatures and small Ga excess fluxes.[@Heying1; @Adelmann] In particular, it has been shown that, at high growth temperatures, a wide range of Ga fluxes exists, for which a finite amount of excess Ga is present on the GaN surface whose quantity is independent of the value of the Ga flux.[@Adelmann] Such conditions may provide a “growth window” for GaN PAMBE, i.e. a region, where the growth mechanisms and the surface morphology are independent of fluctuations of Ga flux and growth temperature.[@Heying1; @Adelmann]
However, the quantitative description of a GaN “growth diagram”, which describes the Ga surface coverage during growth as a function of Ga flux and growth temperature, has not yet been achieved: the results on the temperature dependence of the critical excess Ga flux at the onset of Ga droplet formation are contradictory, yielding activation energies of 2.8eV (Ref. ) and 4.8eV (Ref. ), respectively. This suggests that the underlying mechanisms of Ga accumulation are not yet understood.
To address this discrepancy, we have performed Ga adsorption measurements on (0001) GaN. We discuss the results in the framework of a lattice-gas growth model for adsorption, which is based on [*ab initio*]{} calculated parameters. This model explains the origin of the apparently contradicting parameters derived from previous experimental studies[@Heying1; @Adelmann] and gives a consistent description.
Experimental Procedure
======================
The adsorption experiments were performed in a MECA2000 molecular-beam epitaxy chamber equipped with a standard effusion cell for Ga evaporation. The chamber also contains a rf plasma cell to provide active nitrogen for GaN growth. The pseudo-substrates used were 2$\mu$m thick (0001) (Ga-polarity) GaN layers grown by MOCVD on sapphire. The substrate temperature $T_S$ was measured by a thermocouple in mechanical contact to the backside of the molybdenum sample holder and shielded from direct heating. Prior to all experiments, a 100 nm thick GaN layer was grown under Ga-rich conditions on the pseudosubstrates to remove the influence of a possible surface contamination layer.
Ga fluxes $\Phi$ have been calibrated to Ga effusion cell temperatures by reflection high-energy electron diffraction (RHEED) intensity oscillations during N-rich GaN growth at a substrate temperature of $T_S = 620$[$^\circ$]{}C. In these conditions, the growth rate is actually proportional to the impinging Ga flux. It is sound to assume that the Ga adatom sticking coefficient is unity at such a low substrate temperature, which permits an absolute calibration of the Ga flux *in terms of the GaN surface site density*.
The Ga surface coverage was assessed by analyzing the specularly-reflected RHEED intensity by a method described in Refs. and . This method uses the oscillatory transients in specular RHEED intensity, which are observed during Ga adsorption and desorption on/from (0001) GaN surfaces. It has been shown that the duration of these transients can be qualitatively related to the amount of adsorbing or desorbing Ga.[@Mula; @Zheng] In general, the relation between intensity and Ga coverage is unknown. Although tempting, the interpretation of these electron reflectivity transients in terms of RHEED oscillations is not obvious. Furthermore, it must been noted that the shape of the transients (albeit not their duration) depends on the diffraction conditions, notably the incidence angle. Of course, the modeling of electron reflection would allow to directly relate the RHEED intensity to the Ga coverage (and the surface structure) but this is beyond the scope of this work.
However, the total duration of the transients occurring during Ga desorption can be used to qualitatively estimate the amount of Ga adsorbed on the surface. In Ref. , an indirect method has been used to draw limited *quantitative* information from the desorption transients; below we will demonstrate a fully quantitative calibration relating the Ga desorption transient time to absolute Ga surface coverages. This allows us to circumvent the problem that the RHEED intensity cannot in general be easily related to adatom coverage. The experimental procedure is thus as follows: to assess the Ga quantity present after Ga adsorption for different impinging Ga fluxes, the Ga flux has been interrupted after a fixed adsorption time. The subsequent variation of the specular RHEED intensity due to Ga evaporation under vacuum has been recorded.
Results
=======
Ga adsorption
-------------
![\[Fig:OR\] Specular RHEED intensity during Ga desorption from a (0001) GaN surface. Beforehand, Ga adsorption has been carried out at Ga fluxes $\Phi$ as indicated. For $\Phi < 0.68$ML/s, the desorption transients correspond to equilibrium adsorption, i.e. they do not change as a function of the previous adsorption time. For $\Phi > 0.68$ML/s, this does not hold and the desorption transients depend on the adsorption time (here 1min). The substrate temperature is $T_S = 740$[$^\circ$]{}C.](Figure1.eps){width="8cm"}
Figure \[Fig:OR\] shows the variation of the specular RHEED intensity after 1min of Ga adsorption (substrate temperature $T_S = 740$[$^\circ$]{}C, Ga fluxes $\Phi$ as indicated) and subsequent interruption of the Ga flux (at $t = 0$). We observe oscillatory transients during Ga desorption from the (0001) GaN surface. To quantify the desorption process, we define a desorption time $t_{\rm des}$ as the time interval between the shuttering of the Ga flux and the last inflection point in RHEED intensity (Fig. \[Fig:OR\]). The desorption time depends on the amount of Ga present on the GaN surface after adsorption.[@Mula] The fundamental finding is that for Ga fluxes below $\Phi = 0.68$ monolayers (ML)/s, the desorption transients — and thus the desorption time $t_{\rm des}$ — are independent of the previous adsorption time, i.e. of the amount of nominally impinged Ga. This is consistent with the results in Ref. and is visualized in Fig. \[Fig:Sat\], which shows $t_{\rm des}$ as a function of the amount of nominally impinged Ga $\theta$ (defined as the product of the Ga flux and the adsorption time). The derivative of this curve gives the (coverage and flux dependent) Ga sticking coefficient. Data are shown in Fig. \[Fig:Sat\] for a substrate temperature of $T_S = 740$[$^\circ$]{}C and two different Ga fluxes: for $\Phi = 0.30$ML/s, the desorption time (and hence the amount of adsorbed Ga) monotonously increases until it saturates after $\theta \simeq 4$ML. At larger $\theta$, the adsorption has reached equilibrium and the coverage remains constant.
![\[Fig:Sat\] Ga desorption time as a function of the amount of nominally impinged Ga, $\theta$, for two different Ga fluxes, as indicated. $T_S = 740$[$^\circ$]{}C.](Figure2.eps){width="8.5cm"}
![\[Fig:Isotherme\] Ga desorption time as a function of impinging Ga flux $\Phi$ after equilibrium has been attained for regions 1 and 2, and after 1min of Ga adsorption in region 3. The substrate temperature is $T_S = 740$[$^\circ$]{}C.](Figure3.eps){width="8.5cm"}
For a higher Ga flux of $\Phi = 0.90$ML/s, we observe the same behavior for $\theta \lesssim 4$ML, but thereafter, we find no saturation. Instead, a continuous increase of the desorption time is observed, corresponding to Ga accumulation on the GaN surface. A more detailed analysis finds that the desorption transients for $\Phi \le 0.72$ML/s correspond to finite equilibrium Ga surface coverages. For higher Ga fluxes, no finite equilibrium coverages exist and Ga will infinitely grow and finally form macroscopic droplets on the surface.
These results of Fig. \[Fig:OR\] are summarized in Fig. \[Fig:Isotherme\], which shows the Ga desorption time — related to the amount of adsorbed Ga — as a function of the impinging Ga flux at a constant substrate temperature $T_S = 740$[$^\circ$]{}C. It can be regarded as a Ga adsorption isotherm. We can discriminate three different regions:(1) an S-shaped increase of the Ga coverage for $\Phi < 0.20$ML/s, (2) a constant Ga coverage up to $\Phi = 0.72$ML/s, independent of the Ga flux, and (3) Ga accumulation and no finite equilibrium Ga coverages for higher $\Phi$. It is worth noting that the transitions between the three regimes are discontinuous within the experimental precision of 1[$^\circ$]{}C of the Ga effusion cell ($\sim 7\times 10^{-3}$ML/s for fluxes around 0.5ML/s). In particular, no intermediate equilibrium coverages have been found between regimes 1 and 2. Therefore, the transition fluxes between the different regimes are well defined.
![\[Fig:DDP\] Ga adsorption phase diagram indicating the Ga surface coverage as a function of impinging Ga flux $\Phi$ and substrate temperature $T_S$. The definition of regions 1–3 follows Fig. \[Fig:Isotherme\]. The inset shows an Arrhenius plot of the data.](Figure4.eps){width="8.5cm"}
Reference Transition $E_{A}^{\rm{exp}}$ (eV) $\nu_{\rm{des}}^{\rm{exp}} (\text{Hz})$ $v$ (ML/s) $\alpha$ ([meV/K]{}) $\nu_{\rm{des}}^{\rm{ren}} (\text{Hz})$
----------- ------------------ ------------------------- ----------------------------------------- ------------ ---------------------- -----------------------------------------
This work $1\rightarrow 2$ 5.2 $3\times 10^{25}$ 0 $-2.3$ $4.7\times10^{25}$
This work $2\rightarrow 3$ 5.1 $2\times 10^{25}$ 0 $-2.2$ $1.5\times 10^{25}$
Ref. $1\rightarrow 2$ 3.7 $5\times 10^{17}$ 0.28 $-0.8$ $1.6\times 10^{18}$
Ref. $2\rightarrow 3$ 4.8 $1\times 10^{24}$ 0.28 $-1.9$ $4.8\times 10^{23}$
Ref. $2\rightarrow 3$ 2.8 $1 \times 10^{14}$ 1.1 $> - 0.1$ $5.5\times 10^{13}$
To fully assess the thermodynamics of the adsorption process, the variation of the adsorption isotherm has to be known as a function of substrate temperature. In the following we will restrict ourselves to the study of the variation of the transition fluxes between the different regimes. The result is plotted in Fig. \[Fig:DDP\]. We see that the transition fluxes vary exponentially with substrate temperature. The inset shows an Arrhenius plot of the data, yielding activation energies of $E_A^{(12)} = 5.2 \pm 0.1$eV and $E_A^{(23)} = 5.1 \pm 0.05$eV for the $1\rightarrow 2$ and $2\rightarrow 3$ transition, respectively. These values, the prefactors, and the corresponding values for the growth phase diagram in Refs. ($1\rightarrow 2$ corresponds to $B\rightarrow C$ and $2\rightarrow 3$ corresponds to $C\rightarrow D$) and (discussing the $2\rightarrow 3$ transition only) are summarized in Tab. \[Tab:FitResults\] and will be discussed in Sec. IV.
GaN surface structures
----------------------
The adsorption isotherm in Fig. \[Fig:Isotherme\] is given in terms of the Ga desorption time, which is only *qualitatively* related to the amount of adsorbed Ga. For a fully *quantitative* treatment of Ga adsorption, one must know the dependence of the Ga desorption rate $\Phi_{\rm des}$ on the Ga surface coverage $c$. This would allow the modeling of the Ga re-evaporation, as the desorption rate is given by
$$\label{Eq:RateEq} \Phi_{\rm des}(c) = \frac{{\rm d}c}{{\rm d}t}
\quad ,$$
with the initial condition $c(t=-t_{\rm des}) = c_{\rm eq}$, which denotes the amount of Ga adsorbed in equilibrium conditions, i.e., before Ga desorption sets in. After the time interval $t_{\rm des}$, the Ga coverage becomes zero. The knowledge of $t_{\rm des}$ would thus allow us to compute $c_{\rm eq}$ if $\Phi_{\rm des}(c)$ was known (which it is not). This requirement can be circumvented by considering that in equilibrium, the impinging Ga flux $\Phi$ must exactly balance the evaporation rate, hence $\Phi_{\rm des}(c_{\rm eq}) = - \Phi(c_{\rm eq}[t_{\rm des}])$. Integrating Eq. (\[Eq:RateEq\]), taking the first derivative with respect to $t_{\rm des}$, and using the above substitution leads to
$$\label{Eq:ModelAdsorptionIsotherm}
c_{\rm eq}(\Phi) = \int_0^{\Phi} \Phi'\frac{\partial t_{\rm des}}{\partial\Phi'} {\rm d}\Phi'\quad .$$
This expression allows the computation of $c_{\rm eq}$ from $t_{\rm des}$ as a function of $\Phi$ (which is known from the experimental data) and can be evaluated numerically. Note that $c_{\rm eq}$ only depends on the derivative of $t_{\rm des}(\Phi)$, which means that $t_{\rm des}$ does not necessarily have to denote the very end of Ga adsorption but can be taken after any time interval, as long as it is well-defined in the RHEED signal. Any further Ga desorption after the end of the chosen time interval will lead to a constant offset $t_{\rm des}$, which does not contribute to $c_{\rm eq}$ in Eq. (\[Eq:ModelAdsorptionIsotherm\]).
![\[Fig:Calib\] Calibrated Ga adsorption isotherm at $T_S = 740$[$^\circ$]{}C using Eq. (\[Eq:ModelAdsorptionIsotherm\]). The data are derived from Fig. \[Fig:Isotherme\]. As the method can only be applied to equilibrium coverages, only regions 1 and 2 are represented.](Figure5.eps){width="8.5cm"}
Using Eq. (\[Eq:ModelAdsorptionIsotherm\]), we can now calibrate the Ga adsorption isotherm in Fig. \[Fig:Isotherme\]. Since the relation $\Phi_{\rm des}(c_{\rm eq}) = - \Phi(t_{\rm des})$ implies steady-state conditions, only the data in regions 1 and 2 can be treated. Applying Eq. (\[Eq:ModelAdsorptionIsotherm\]) to the data in region 3 does not lead to the amount of Ga adsorbed after a finite time but to incorrect values because under these conditions $\vert\Phi_{\rm des}\vert < \vert\Phi\vert$. Since experimental data are available only for $\Phi > 0.032$ML/s, the isotherm has been extrapolated by an exponential for smaller fluxes. The contribution of the interval $0 < \Phi < 0.032$ML/s to the overall integral in Eq. (\[Eq:ModelAdsorptionIsotherm\]) is 0.04ML. This is only a minor correction, suggesting that the specific form of the extrapolation function is not important. The result of the calibration is plotted in Fig. \[Fig:Calib\]. We observe that in region 1, the Ga coverage increases from almost zero to a value of 0.98ML, close to 1ML. The coverage then increases abruptly to a value of 2.5ML in region 2. Typical systematical errors can be estimated to be of the order of $\pm 0.2$ML.
What is the detailed structure of such adsorbed Ga films? It must be kept in mind that the Ga fluxes have been calibrated by GaN RHEED oscillations and are hence given in terms of the GaN surface site density. A GaN coverage of 1ML thus indicates the adsorption of a Ga adatom on each GaN site. This suggests that Ga adsorbs in region 1 as a coherent (pseudomorphic) adlayer.
At the $1\rightarrow 2$ transition, the Ga coverage increases by about 1.5ML, i.e. by more than a pseudomorphic adlayer. This compares favorably to the laterally-contracted bilayer model, which has been calculated to be the most stable structure of Ga-rich (0001) GaN surfaces.[@Northrup1] It consists of two adsorbed Ga adlayers on top of a Ga-terminated GaN surface. The first layer is found to be pseudomorphic to the GaN surface but the second one has an in-plane lattice parameter of 2.75Å, about 13.8% smaller than that of GaN (3.189Å). The second layer thus contains 1.3ML of Ga in terms of the GaN surface site density, in good agreement with the experimental value of 1.5ML.
![\[Fig:RHEED\] **(a)** RHEED pattern (azimuth $\langle 11\bar{2}0 \rangle$) of a (0001) GaN surface with a Ga bilayer present at $T_S = 600$[$^\circ$]{}C. The white arrows indicate additional streaks due to the Ga film. **(b)** RHEED intensity profile in the $\langle 10\bar{1}0 \rangle$ direction evidencing the additional RHEED streaks due to the Ga bilayer.](Figure6.eps){width="8cm"}
The formation of a laterally-contracted bilayer structure in regime 2 is further corroborated by the observation of supplementary streaks in the RHEED pattern after Ga adsorption in regime 2 and rapid quenching of the sample down to substrate temperatures below about $T_S = 650$[$^\circ$]{}C \[see Fig. \[Fig:RHEED\](a)\]. At higher temperatures, the supplementary streaks are too weak to be detected unambiguously. The lattice parameter corresponding to these streaks is found to be $2.73 \pm 0.03$Å by a fit using pseudo-Voigt functions and assuming the bulk lattice parameter for the GaN layer. This value compares very favorably to that obtained by the *ab inito* calculations (2.75Å).[@Northrup1] Combined with the adsorption results, this strongly suggests that adsorption in region 2 leads to the formation of a laterally-contracted Ga bilayer.
The shape of the oscillation transients in Fig. \[Fig:OR\] suggests that, in region 3, Ga droplets are formed on top of this Ga bilayer. This Ga de-wetting transition may indicate that the attractive interaction energy of a Ga adatom with the surface is maximum in the second layer and lower in the third layer.[@Brochard] This conclusion is consistent with the first principles results which will be discussed in the next section. Ga thus grows in a Stranski-Krastanow mode on (0001) GaN.
Finally, the different adsorption regimes can be summarized as follows: (1) Ga coverage $c \le 1$ML, i.e. successive formation of a coherent Ga monolayer, (2) a Ga coverage of $c = 2.5$ML, forming a laterally-contracted Ga bilayer, and (3) Ga accumulation and droplet formation on top of a Ga bilayer. In the following, we will derive an ab initio based growth model which describes the temperature dependence of the transition fluxes between the different regimes.
Discussion
==========
![\[Fig:Comp\] Comparison between the experimental growth phase diagram in Ref. (open symbols) and the diagram derived from the the Ga adsorption data in this work (solid lines). The different regimes 1, 2, and 3 of the adsorption phase diagram are indicated, which correspond to regimes B, C, and D in Ref. , respectively.](Figure7.eps){width="8cm"}
An intuitive connection between adsorption and growth phase diagrams can be made by assuming that the *excess* Ga in Ga-rich growth conditions behaves as if it would be adsorbed on a GaN surface. Then, adding the GaN growth rate (0.28ML/s in the experiments of Ref. ) to the adsorption phase transition fluxes should reproduce the growth phase diagram of Ref. . This comparison is shown in Fig. \[Fig:Comp\]. It demonstrates good overall agreement, suggesting that the assumption is valid. However, it must be noted that the critical fluxes derived from adsorption measurements fall below those in the growth phase diagram, although differences are of the order of the experimental precision.
Yet, a closer look at the activation barriers and prefactors (Tab. \[Tab:FitResults\]), as derived from the experimental phase diagrams, poses a number of questions. First, why are the activation energies and prefactors in the adsorption and growth phase diagram so different, even though absolute values of transition fluxes are close? Second, why are the values for the same transition ($2\rightarrow 3$ for growth) in Refs. and so different? Finally, while the barrier of 2.8eV for the $2\rightarrow 3$ transition in the growth phase diagram in Ref. is close to the cohesive energy of bulk Ga and has thus been directly interpreted as a Ga desorption barrier, what is the physical origin of the 5.1eV activation energy and the meaning of a prefactor of 10$^{25}$Hz? The latter value is fundamentally different from prefactors typically observed/calculated for diffusion or desorption processes (which are of the order of $\sim 10^{13}$Hz).
Growth and adsorption model
---------------------------
To address the above questions, we have analyzed the data in terms of a simple growth model and in combination with first principles total energy calculations. In order to simplify the discussion of the growth model, we divide the problem in two parts: First, we derive how the density of Ga adatoms $\rho$ on the surface[@footnote] depends on parameters such as temperature $T$, Ga flux $\Phi$, and growth rate $v$. Second, we calculate the critical adatom density, at which nucleation occurs and the system undergoes a phase transition.
The adatom coverage is given by:
$$\label{eq:drhodta}
\frac{\rm{d} \rho (\mathbf{r},t)}{\rm{d}t} = D \nabla^2
\rho(\mathbf{r},t)+\Phi-\frac{1}{\tau_{\rm{inc}}}\rho
(\mathbf{r},t)-\frac{1}{\tau_{\rm{des}}}\rho
(\mathbf{r},t) \quad .$$
Here, the desorption time $\tau_{\rm des}$ is given by
$$\label{eq:des}
\tau_{\rm des}^{-1} = \nu_{\rm des} e^{-E_{\rm des}/k_B T}\quad ,$$
with $E_{\rm des}$ the desorption barrier and $k_B$ the Boltzmann constant. $D$ is the surface diffusion constant for Ga adatoms and $\mathbf{r}$ gives the lateral position on the surface. For step flow growth (as realized in Ga-rich growth conditions[@Adelmann]), incorporation occurs essentially at the step edges. Since these are moving, the incorporation rate ${\tau^{-1}_{\rm inc}(\mathbf{r},t)}$ is in general inhomogeneous and time dependent. For ideal step flow it will be zero on the terraces and $>0$ only at the step edges. The incorporation rate and the growth rate $v$ are directly related by:
$$v(t) = \frac{1}{A} \int \frac{1}{\tau_{\rm inc}(\mathbf{r},t)}\rho(\mathbf{r},t) d\mathbf{r} \quad .$$
Here the integration is performed over the total surface area $A$. Since in our experimental setup stationary conditions are realized, the explicit time dependence in the growth rate disappears (i.e. $v(t)=v$). Furthermore, in stationary conditions Eq. (\[eq:drhodta\]) also simplifies: $\mathrm{d}\rho/\mathrm{d}t=0$, i.e. the left hand side becomes zero. A further simplification in Eq. (\[eq:drhodta\]) can be made by taking into account that the phase transitions we are interested in occur exclusively under Ga-rich conditions. Under these conditions the surface steps on the surface are Ga-terminated[@StepsToBePublished] and the incorporation rate $\tau^{-1}_{\rm {inc}}$ at such a step will be small. This is in contrast to the conventional step flow picture where the sticking probability of an atom at the step edge is assumed to be close to one. The difference to the conventional model is due to the fact that we have two species with very different concentrations. For nitrogen (which is the minority species for very Ga-rich conditions) the sticking coefficient at the Ga-terminated step edges will be close to one and steps act as sinks to the nitrogen concentration which becomes highly inhomegenous. For Ga atoms, however, the sticking probablity is low. Thus, the effect of steps on Ga will be small and the Ga-adatom density is virtually homogeneous, i.e. $D\nabla^2 \rho(\mathbf{r}) = 0$.
Based on the above discussion, Eq. (\[eq:drhodta\]) can be written as:
$$\label{eq:drhodt}
0 = \Phi - v - \frac{1}{\tau_{\rm des}}\rho_0 \quad .$$
Here, $\rho_0$ denotes the equilibrium adatom density. We note that this equation holds for both growth ($v>0$) and adsorption ($v=0$). The solution is easily found to be
$$\label{eq:rho}
\rho_0 = \left({\Phi - v}\right){\tau_{\rm des}} \quad .$$
Nucleation occurs if the (stable) nuclei are in thermodynamic equilibrium with the lattice gas (which is described by the adatom density) on the surface. At low densities ($\rho_0\ll 1$), interactions in the lattice gas itself can be neglected and we obtain
$$\label{eq:rhocrit}
\frac{N_{\rm ad}}{N_{\rm sites}} = \rho_{\rm crit} = e^{-\Delta E_{\rm
nuc}/k_B T} \quad .$$
Here, $N_{\rm ad}$ is the number of adatoms in the lattice gas, $N_{\rm sites}$ is the total number of surface sites which can be occupied by the adatoms, and $\Delta E_{\rm nuc}$ is the energy the adatom gains if it is attached to a subcritical nucleus making the latter stable.
Based on the above model, we can directly obtain the critical Ga flux $\Phi_{\rm crit}$ at which the phase transitions occur. Combining Eqs. (\[eq:rho\]) and (\[eq:rhocrit\]) gives:
$$\label{eq:critFlux}
\Phi_{\rm crit} = v + \tau_{\rm des}^{-1} e^{-\Delta E_{\rm nuc}/k_B T} \quad .$$
This equation can be rewritten using Eq. (\[eq:des\]) as
$$\label{eq:critFlux2}
\Phi_{\rm crit}= v + \nu_{\rm des} e^{-(E_{\rm des}+\Delta E_{\rm nuc})/k_B T}$$
and applies both to the adsorption and growth phase diagram. The activation energy is thus expected to represent the total binding energy of a Ga atom in a critical Ga cluster.
First principles analysis
-------------------------
A comparison with the experimental results (see Tab. \[Tab:FitResults\]) shows that energy and prefactor are [*not*]{} constant \[as expected from Eq. (\[eq:critFlux2\])\] but vary largely with the growth conditions. As a general trend, one finds that the activation energy and the prefactor decrease with increasing growth rate (N flux). It is also interesting to note that only in the case of high growth rate ($v = 1.1$ML/s) the prefactor ($1\times 10^{14}$Hz) is close to the typical attempt frequencies observed/expected for desorption, i.e. in the $10^{13}$Hz range. For conditions where growth is slow ($v = 0.28$ML/s) or absent (adsorption), prefactors are found which are many orders of magnitude larger.
In order to identify the origin of these apparent discrepancies, we have explicitly calculated the desorption and the formation of small Ga clusters on the Ga bilayer surface employing density functional theory. In the following we will focus on the Ga-bilayer structure (which corresponds to the $2\rightarrow 3$ transition). Based on the almost identical energies of the $1 \rightarrow 2$ and $2\rightarrow 3$ transitions in the adsorption phase diagram we expect the mechanisms/energetics to be rather similar.
![\[Fig:Schema1\] Schematic top view of an adatom on the Ga bilayer structure. The white balls mark the positions of the Ga atoms in the second layer. The gray balls mark the positions of the Ga atoms in the contracted Ga epilayer and the black ball the Ga adatom in the T4 position. The dashed line shows the $1\times 1$ surface unit cell of the ideal bulk truncated GaN (0001) surface. The solid line shows the surface unit cell of the Ga bilayer structure.](Figure8.eps){width="8cm"}
![\[Fig:Schema2\] Schematic top view of a 2 atom island (dimer) on the Ga bilayer structure. The solid line shows the $\left( 2\sqrt3\times2\sqrt3\right)$ surface unit cell which has been used to describe this structure.](Figure9.eps){width="8cm"}
Specifically, we use soft Troullier-Martins pseudopotentials[@Troullier] and the Perdew-Burke-Ernzerhof generalized-gradient approximation (PBE-GGA) to describe exchange/correlation.[@PBE] The Ga $3d$ semicore states were described in the frozen core approximation (nlcc).[@Loui82] Details about the method can be found elsewhere.[@Bockstedtexx] The calculations have been performed with a plane wave basis set (energy cutoff: 50Ry). The Brillouin zone has been sampled by a $2\times 2 \times 1$ Monkhorst-Pack mesh for the $\left( 2\sqrt 3 \times 2\sqrt 3\right)$ unit cell and $4\times 4 \times 1$ for the $\left( \sqrt 3 \times \sqrt 3\right)$ unit cell.[@Monkhorstxx] The Ga bilayer with one up to four adatoms has been modeled by a slab consisting of 2 double layers and $\left(\sqrt 3 \times \sqrt 3\right)$ (to describe the adsorption of a single adatom, see Fig. \[Fig:Schema1\]) and larger $\left( 2\sqrt 3 \times 2\sqrt 3\right)$ (to describe adatom islands, see Fig. \[Fig:Schema2\]) unit cells. Increasing the slab thickness to 4 double layers changes the surface energy by less than 0.5meV per surface unit cell. Adatoms are added only on the upper surface of the slab. The lower surface has been passivated by pseudo-hydrogen to remove the electrically active surface states. The adatom(s) and the first two surface layers have been fully relaxed. Detailed convergence checks can be found in Refs. .
Based on these studies, we have calculated the desorption energy of an adatom and the binding energy of adatoms in small islands. The desorption energy (the energy needed to remove the adatom from the surface) is defined by
$$E_{\rm des} = -\left( E_{\rm tot}^{\rm adatom} - E_{\rm tot}^{\rm slab} - E_{\rm
tot}^{\rm atom}\right)\quad ,$$
where $E_{\rm tot}^{\rm adatom}$ is the total energy of the surface including the adatom, $E_{\rm tot}^{\rm slab}$ that of the free surface and $E_{\rm tot}^{\rm atom}$ that of the (spin-polarized[@spinPolarization]) Ga atom. Using this expression we find an adatom binding energy of 2.52eV for the $\left(\sqrt 3 \times \sqrt 3\right)$ structure and 2.41eV for the $\left(2\sqrt 3 \times 2\sqrt 3\right)$ unit cell. In the equilibrium configuration, the adatom sits on a three-fold coordinated hollow site. The Ga–Ga bond length between adatom and surface layer is 2.68Å, i.e. very close to the nearest neighbor distance in $\alpha$-Ga of 2.71Å.[@Wyckoff]
The binding energy of an adatom in an island consisting of $n_{\rm ad}$ adatoms is given by:
$$E_{\rm isl} = -\left[\frac{1}{n_{\rm ad}}\left( E_{\rm tot}^{\rm isl} - E_{\rm tot}^{\rm
slab} - E_{\rm tot}^{\rm atom}\right) - E_{\rm des}\right] \quad .$$
For an island consisting of 2 adatoms we find $E_{\rm isl}=0.15$eV. For larger islands consisting of 3 adatoms $E_{\rm isl}=0.30$eV and $E_{\rm isl}=0.34$eV for a 4 atom island. The numbers are the island formation energies as calculated in a $\left( 2\sqrt3 \times 2\sqrt3\right)$ cell. It is interesting to note here that all islands are unstable against the formation of Ga-droplets: the formation energy of a Ga atom in an island ($E_{\rm des} + E_{\rm isl} \simeq 2.75$eV) is smaller than the cohesive energy of bulk Ga of 2.8eV.[@Kittel] Therefore the islands act as nucleation centers for Ga droplet formation.
Interpetation of the results
----------------------------
![\[Fig:lnnuEA\] Natural logarithm of the experimental prefactors $\ln\left(\nu_{\rm{des}}^{\rm{exp}}\right)$ as a function of the activation energy $E_A^{\rm{exp}}$ for various data in this work and Refs. and (full circles; see Tab. \[Tab:FitResults\]). The solid line represents a linear fit.](Figure10.eps){width="8cm"}
We can now compare these energies with those obtained from the analysis of the experimental data (Tab. \[Tab:FitResults\]). As can be seen, the activation energy is close to our calculated desorption energy only for the high growth case (1.1ML/s). For lower growth rates or adsorption, activation energies are found which are way too large. A closer analysis of the prefactors and the activation energies shows a clear relation (see Fig. \[Fig:lnnuEA\]):
$$\label{eq:fit}
\ln\left(\nu_{\rm{des}}^{\rm{exp}}\right) = a_1 E_{\rm{A}}^{\rm{exp}} + a_0 \quad .$$
The supercript “exp” indicates experimental data. For the values given in Tab. \[Tab:FitResults\] we get $a_1 = 11.45$eV$^{-1}$ and $a_0 = -0.427$.
The observed relation between prefactor and activation barrier can be explained in terms of a temperature dependent activation energy. Since the experimentally accessible temperature range is rather small ($\simeq 50$K), we assume a linear dependence: $E_A = E_0 + \alpha (T-T_0)$ with $E_0$ the temperature independent contribution, $T_0$ the temperature offset, and $\alpha$ the linear temperature coefficient. Equation (\[eq:critFlux2\]) then becomes
$$\label{eq:critFluxlinEA}
\Phi'_{\rm crit} = v + \nu_{\rm des}e^{-\alpha/k_B} e^{-(E_0-\alpha T_0)/k_B
T} \quad .$$
A comparison between Eqs. (\[eq:fit\]) and (\[eq:critFluxlinEA\]) gives the following relations:
$$\label{eq:expEa}
E_{\rm{A}}^{\rm{exp}} = E_0 - \alpha T_0 \qquad \text{and} \qquad
\ln\left(\nu_{\rm{des}}^{\rm{exp}}\right) = \ln\left(\nu_{\rm des}\right) -
\frac{\alpha}{k_B} \quad .$$
This leads to
$$\begin{aligned}
a_0 = \ln(\nu_{\rm des}) - \frac{E_0}{k_BT_0} \qquad \text{and} \qquad a_1 =
\frac{1}{k_B T_0} \quad .\end{aligned}$$
Using these relations and $E_0=2.75$eV (as found from our first principles calculations), we obtain $T_0 = 740$[$^\circ$]{}C (which is close to the experimentally accessed temperature range, so the linear approximation for the temperature dependence of $E_A$ is well justified) and $\nu_{\rm des} = 3\times 10^{13}$Hz (which is close to typical attempt frequencies).
Using the above parameters and Eq. (\[eq:expEa\]), the linear temperature coefficient $\alpha = (E_0-E_A^{\rm{exp}})/T_0$ can be computed for all transitions. The result is shown in Tab. \[Tab:FitResults\]. Based on these values we can renormalize the prefactors following $\nu_{\rm{des}}^{\rm{ren}} = \nu_{\rm des}e^{-\alpha/k_B}$. The resulting numbers calculated with $\nu_{\rm des} = 3\times 10^{13}$Hz are listed in Tab. \[Tab:FitResults\]. The good agreement with the experimental prefactors $\nu_{\rm{des}}^{\rm{exp}}$ shows that the model consistently describes all previous experimental studies. The large variation in frequencies and activation energies can be explained by assuming that the N flux affects the temperature dependence: it is largest if no N flux is present and monotonously decreases with increasing N flux (growth rate).
It is interesting to note that, although the temperature dependence has a huge effect on the experimentally measured apparent activation barrier $E_A^{\rm{exp}} = E_0 - \alpha T_0$ (by almost a factor of two) and prefactor $\nu_{\rm{des}}^{\rm{exp}}$ (by up to 12 orders of magnitude), the actual change in the effective desorption energy $E_A = E_0 + \alpha (T-T_0)$ is small: in the case of adsorption ($v = 0$), the activation barrier changes from 2.84eV to 2.73eV within the experimentally measured temperature window (700[$^\circ$]{}C to 750[$^\circ$]{}C). In the case of GaN growth, the variation is even smaller. This can be intuitively understood by considering that the experimentally measured activation energy represents a linear projection to zero temperature even when its real temperature dependence deviates strongly from linear behavior outside our experimental temperature window.
Conclusion
==========
Based on a combination of experimental RHEED studies and first-principle growth models we have (i) quantitatively determined the Ga coverage on the GaN (0001) surface during adsorption as a function of Ga flux and substrate temperature and (ii) derived a model, which consistently describes the adsorption of Ga on GaN surfaces as well as the accumulation of Ga during Ga-rich GaN growth. This model resolves the discrepancy in previous measurements of the activation energy characterizing the critical Ga flux for the onset of Ga droplet fromation during GaN growth.[@Heying1; @Adelmann]
The model also explains the origin of the experimentally-observed unphysically high prefactors in terms of a temperature dependent desorption barrier. At the moment, we can only speculate about possible mechanisms which reduce the activation barrier at higher temperatures. A possible scenario emerges from our first principles calculations where we find that the number of atoms in the compressed Ga layer of the bilayer structure and thus the lateral lattice constant of the top Ga layer significantly changes with temperature. This change in the surface geometry is expected to have an important effect on the island formation energy and will be discussed in a forthcoming paper.[@LymperakisToBePublished]
The authors would like to thank O. Briot (University of Montpellier, France) for providing the GaN templates and F. Rieutord (CEA Grenoble, DRFMC/SI3M) for valuable discussions. E. Bellet-Amalric (CEA Grenoble, DRFMC/SP2M) is acknowledged for the analysis of Fig. \[Fig:RHEED\]. L. L. and J. N. like to thank the EU TMR program IPAM and J. N. the Deutsche Forschungsgemeinschaft SFB 296.
[99]{}
S. C. Jain, M. Willander, J. Narayan, and R. van Overstraeten, J. Appl. Phys. **87**, 965 (2000).
P. Hacke, G. Feuillet, H. Okumura, and S. Yoshida, Appl. Phys. Lett. **69**, 2507 (1996).
K. Rapcewicz, M. Buongiorno Nardelli, and J. Bernholc, Phys. Rev. B **56**, R12725 (1997).
T. Zywietz, J. Neugebauer, and M. Scheffler, Appl. Phys. Lett. [**73**]{}, 487 (1998).
A. R. Smith, V. Ramachandran, R. M. Feenstra, D. W. Greve, M.-S. Shin and M. Skowronski, J. Neugebauer, and J. E. Northrup, J. Vac. Sci. Technol. B **16**, 1641 (1998); A. R. Smith, R. M. Feenstra, D. W. Greve, M. S. Shin, M. Skowronski, J. Neugebauer, and J. E. Northrup, *ibid.* **16**, 2242 (1998).
J. Fritsch, O. F. Sankey, K. E. Schmidt, and J. B. Page, Phys. Rev. B **57**, 15360 (1998).
Q.-Z. Xue, Q.-K. Xue, R. Z. Bakhtizin, Y. Hasegawa, I. S. T. Tsong, T. Sakurai, and T. Ohno, Phys. Rev. B **59**, 12604 (1999); Q.-K. Xue, Q.-Z. Xue, R. Z. Bakhtizin, Y. Hasegawa, I. S. T. Tsong, T. Sakurai, and T. Ohno, Phys. Rev. Lett. **82**, 3074 (1999).
A. Munkholm, G. B. Stephenson, J. A. Eastman, C. Thompson, P. Fini, J. S. Speck, O. Auciello, P. H. Fuoss, and S. P. DenBaars, Phys. Rev. Lett. **83**, 741 (1999).
J. E. Northrup, J. Neugebauer, R. M. Feenstra, and A. R. Smith, Phys. Rev. B **61**, 9932 (2000).
F.-H. Wang, P. Krüger, and J. Pollmann, Phys. Rev. B **64**, 035305 (2001).
E. J. Tarsa, B. Heying, X. H. Wu, P. Fini, S. P. DenBaars, and J. S. Speck, J. Appl. Phys. [**82**]{}, 5472 (1997).
F. Widmann, B. Daudin, G. Feuillet, N. Pelekanos, and J. L. Rouvière, Appl. Phys. Lett. [**73**]{}, 2642 (1998).
M. H. Xie, S. M. Seutter, W. K. Zhu, L. X. Zheng, H. Wu, and S. Y. Tong, Phys. Rev. Lett. **82**, 2749 (1999).
B. Heying, R. Averbeck, L. F. Chen, E. Haus, H. Riechert, and J. S. Speck, J. Appl. Phys. [**88**]{}, 1855 (2000).
A. Bourret, C. Adelmann, B. Daudin, J.-L. Rouvière, G. Feuillet, and G. Mula, Phys. Rev. B **63**, 245307 (2001).
G. Mula, C. Adelmann, S. Moehl, J. Oullier, and B. Daudin, Phys. Rev. B **64**, 195406 (2001).
C. Adelmann, J. Brault, D. Jalabert, P. Gentile, H. Mariette, G. Mula, and B. Daudin, J. Appl. Phys. [**91**]{}, 9638 (2002).
B. Heying, I. Smorchkova, C. Poblenz, C. Elsass, P. Fini, S. DenBaars, U. Mishra, and J. S. Speck, Appl. Phys. Lett. [**77**]{}, 2885 (2000).
C. Kruse, S. Einfeldt, T. Böttcher, D. Hommel, D. Rudloff, and J. Christen, Appl. Phys. Lett. [**78**]{}, 3827 (2001).
L. X. Zheng, M. H. Xie, and S. Y. Tong, Phys. Rev. B **61**, 4890 (2001).
F. Brochard-Wyart, J.-M. di Meglio, D. Quéré, and P.-G. de Gennes, Langmuir **7**, 335 (1991).
Here, $\rho$ denotes the adatom density on top of the last completed Ga layer (a monolayer for the $1\rightarrow 2$ transition and a bilayer for the $2\rightarrow 3$ transition). This is contrasted by the Ga coverage $c$, which gives the *total* amount of Ga present on the surface (including completed Ga mono- or bilayers).
L. Lymperakis *et al.* (unpublished).
N. Troullier and J.L. Martins, Phys. Rev. B **43**, 1993 (1991).
J.P. Perdew, K. Burke and M. Ernzerhof, Phys. Rev. Lett. **77**, 3865 (1996); *ibid.* **80**, 891 (1998).
S.G. Louie, S. Froyen, M.L. Cohen, Phys. Rev. B **26**, 1738 (1982).
M. Bockstedte, A. Kley, J. Neugebauer and M. Scheffler, Comput. Phys. Commun. **107**, 187 (1997).
H.J. Monkhorst and J.D. Pack, Phys. Rev. B **13**, 5188 (1976).
C. D. Lee, R. M. Feenstra, J. E. Northrup, L. Lymperakis, J. Neugebauer, Mater. Res. Soc. Proc. **743**, L4.1 (2002)
C. D. Lee, R. M. Feenstra, J. E. Northrup, L. Lymperakis, J. Neugebauer, Appl. Phys. Lett. (in print).
M. Fuchs, J. L. F. Da Silva, C. Stampfl, J. Neugebauer, M. Scheffler, Phys. Rev. B **65**, 245212 (2002).
The spin-polarized correction given by\
`http://math.nist.gov/DFTdata/atomdata/tables/ptable.html`\
has been used.
R. W. G. Wyckoff, *Crystal Structures* (Wiley, New York, 1962), Vol. 1, 2$^\mathrm{nd}$ ed.
C. Kittel, *Introduction to Solid State Physics* (Wiley, New York, 1986).
L. Lymperakis *et al.* (unpublished).
| {
"pile_set_name": "ArXiv"
} |
**The birth of strange stars:**
**kinetics, hydrodynamics and phenomenology**
**of supernovae and GRBs**
J. E. Horvath
*Instituto de Astronomia, Geofísica e Ciências Atmosféricas*
*Rua do Matão 1226, 05508-900 São Paulo SP, Brazil*
[email protected]
Abstract
We present a short review of strange quark matter in supernovae and related explosions, with particular attention to the issue of the propagation of the combustion in the dense stellar environment. We discuss the instabilities affecting the flame and present some new results of application to the turbulent regime. The transition to the distributed regime and further deflagration-to-detonation mechanism are addressed. Finally we show that magnetic fields may be important for this problem, because they modify the flame through the dispersion relations which characterize the instabilities. A tentative classification of explosive phenomena according to the value of the average local magnetic field affecting the burning and the type of stellar system in which this conversion is taking place is presented. As a general result, we conclude that “short” conversion timescales are always favored, since the burning falls in either the turbulent Rayleigh-Taylor (or even the distributed) regime, or perhaps in the detonation one. In both cases the velocity is several orders of magnitude larger than $v_{lam}$, and therefore the latter is irrelevant in practice for this problem. Interesting perspectives for the study of this problem are still open and important issues need to be addressed.
Introduction
============
Intensive work in the 60´s and 70´s definitely established the concept of elementary constituents of nucleons (quarks and gluons). At increasing center-of-mass energy in experimental searches of the elementary components ([*partons*]{}) of protons and other hadrons revealed new physics in need of a theoretical framework to be developed. The theory of “new” strong interactions (as opposed to the “old” nuclear physics) was constructed in parallel, first focused on classification schemes (or, as is called today, flavor physics) and later on finding a theory to describe the dynamics. The development and success of gauge theories in the ’70s eventually leaded to a non-abelian version based on the $SU(3)_c$ symmetry group [@QCD] as a “natural” candidate for a theory of strong interactions. The fundamental quantum number carried by the elementary constituents (quarks) was named “color”, and consequently the dynamics involving quarks and gauge fields (gluons) become known as [*Quantum Chromodynamics*]{} (QCD for short).
It was considered by many somewhat puzzling that repeated efforts to find these entities as free particles (asymptotic states) failed. Subsequent work elaborated on a striking feature of the theory: that the interactions themselves preclude the appearance of the quarks and gluons outside ordinary hadrons, they remain confined inside them at low energies. Another property was soon demonstrated to hold when momentum transfer scales $Q$ became large enough. This is the so-called [*asymptotic freedom*]{}, and states that the colored particles behave as if they were free in the limit $Q \, \rightarrow \, 0$. Actually, there is an energy (or momentum) scale above which color quantum number is not confined any more, but how large the momentum transfer should be (or in other words, which is energy, as measured by the temperature or density of the ensemble allowing the deconfinement) is still a matter of debate. These developments mean that the early universe passed through a deconfinement $\rightarrow$ confinement phase transition along its cooling, although less certainty holds for the densities of the “natural” laboratories (neutron stars) in which [*compression*]{} would deconfine hadronic matter. The earliest calculations [@Baymetal] using reasonable models for both the confined and deconfined phases imprinted on successive researchers the uncertain conclusion that quarks and gluons (forming a state known as the quark-gluon plasma, or QGP) should appear at densities above a large threshold, say, $10 \times \rho_{0}$; with $\rho_{0}$ the nuclear saturation density.
From the starting of these calculations it has proved very difficult to reliably determine the transition points, and also the nature of the transition itself (at least when full numerical calculations were out of sight [@num]). Most of the times the conclusions had to be extracted from simultaneous extrapolations of both a quark model, expected to be valid for $\rho \rightarrow
\infty$, and an hadronic model valid around $\rho_{0}$ but uncertain much above it. Since there is no certainty in either one, the final result is always subject to reasonable doubts. The “induction” of a definite [*order*]{} of the transition because of the adopted functional forms of the thermodynamical quantities of both sides. Nevertheless these serious and honest attempts have proliferated until today, given that the transition is still elusive (the extensively studied finite temperature case still has some small uncertainty in the value of $T_{c}$ and a quite consensual assessment of the order, see [@T] for details). Recent analysis [@RHIC] of hadronic flows have added a lot of excitement to these topics, since it appears that the QGP was indeed produced in heavy ion collisions, but the asymptotic form is [*not*]{} reached, rather behaving as a glass-like system. Needless to say, this kind of studies attract a lot of attention and offers a concrete form to glimpse the deconfined state of hadronic matter, yet to be characterized and understood.
Stable strange quark matter?
=============================
While the study of the quark-gluon plasma occupied many studies in connection with the early Universe and compact stars, a much radical proposal emerged in the 80´s about it, which may be described as follows: it is true that the asymptotic freedom property guarantees that quarks and gluons will be the ground state of QCD at high densities/temperatures, but it says nothing about the ground state at lower densities or temperatures. The everyday experience strongly suggests that ordinary hadrons confine the quarks/gluons and thus constitute the “true” (in the sense of $\rho \rightarrow 0$ and $T \rightarrow 0$) ground state of hadronic matter. The emerging [*strange matter hypothesis*]{} came precisely to challenge this “common sense” statement: it says that the true ground state of hadronic matter is a particular form of the QGP , differing from the ordinary matter by the presence of a key quantum number (strangeness). This is counterintuitive to many people, but a careful look at the physical arguments shows no inconsistency whatsoever, at least in principle.
An argument for the SQM being the true ground state can be made as follows: as is well-known the quantity that determines which phase is preferred is the Gibbs free energy per particle $G/n$ as a function of the pressure (we impose $T=0$ hereafter as appropriate for highly degenerate hadronic matter, it is easy to see that the term $-TS$ in the free energy disfavors SQM at high temperatures). As $P$ is increased starting from the neighborhood of the nuclear matter saturation point $\rho_{0}$ the asymptotic freedom guarantees that there has to be a switch from nuclear matter (N) to elementary hadronic constituents, that is, the lighter quarks $u$ and $d$. The point at which this is supposed to happen will be labelled as $P_{c}$. Thus, the doubts stated above about the appearance of the QGP inside neutron stars may be now restated as whether the pressure at the center is larger or smaller than $P_{c}$.
However, this is where the concept of strangeness plays an important role. Strangeness is the flavor quantum number carried by $\Lambda$s and other heavy hadrons. At the elementary level, it is carried by a different quark $s$, with current mass in the ballpark of $\sim 100 \, MeV$, that is, light enough to be present at a few times the nuclear saturation density. While creating strangeness in hadrons costs energy (because strange hadrons are heavier than non-strange ones; for instance, the $\Lambda$s are heavier than the neutrons and so on); this trend is reversed inside the QGP. The reason is simply the Pauli exclusion principle: a new Fermi sea in the liquid (the one of the $s$ quark) allows a rearrangement of the energy, and sharing it lowers the energy per particle. However, the gain is not precisely known, but it is not impossible to imagine lowering the free energy per particle to a value that would be lower than the mass of the neutron $m_{n}$ even when $P \rightarrow 0$. If realized, this would preclude the (strange) QGP to decay into ordinary hadrons because this would [*cost*]{} energy and the SQM would have been created. Put it simply, the compression would liberate the elementary components that quickly create their own way of surviving. We stress that all these are bulk (i.e. large number) concepts, and it is central to the SQM hypothesis to reach a strangeness per baryon of the order one (and exactly one if the strange quark had no mass to deplete its relative abundance). This is not possible in a few-body system like a nucleus, because each weak decay creating a strangeness unit contributes roughly with a factor $G_{Fermi}^{2}$ to the amplitude, and thus the simultaneous decays are strongly suppressed; this is why it has been very difficult to produce even doubly strange nuclei, let alone higher multiplicity ones. However, once quarks roam free in the QGP they can easily decay by $u + d \, \rightarrow \, u + s$ because there is plenty of phase space for the products until equilibrium is reached. These bulk estimates have been always one way or another behind the idea of SQM.
As it stands, the SQM hypothesis is very bold. It conjectures that every hadron we see around us is in a metastable state, and if conditions for creation of a large net strangeness were met, the matter would not make back ordinary hadrons (technically it is said that SQM constitutes a non-topological soliton stabilized against decays by a conserved charge, the baryon number, see [@FOGA] for a thorough discussion of this case and related ones). The general idea of reaching extreme conditions and stabilizing the QGP is already apparent in the paper of Bodmer [@Bodmer], later reintroduced and refined in references [@CK; @Tera; @Bjo] and colorfully discussed in the paper of Witten [@Ed], which was fundamental to give a big boost to SQM research.
SQM as a theoretical construction is interesting, but finding it in nature would be infinitely more. Key questions of SQM such as whether it does exist or not, and whether it has been ever produced in the Universe are still unsolved. On the other hand, we begin for the first time to have the possibility of falsifying these basic questions mainly thanks to the new generation of space telescopes (HST, Chandra, XMM) and neutrino observatories (SNO, Kamioka, Icecube), to name just a few. These instruments may be used to look for exotic states in compact stars and their birth events.
Many applications of SQM in astrophysics were foreseen during the first decade after its official birth [@Ed] and early infancy [@FJ] Since astrophysical insight has shown to be essential in the determination of fundamental questions related to SQM, we shall focus briefly in a very definite (and important) astrophysical problem, trying to give an assessment and pointing on the uncertainties and possible directions that may be explored in the near future. We thus restrict our discussion to SQM in compact stars, and more specifically, on how a seed of SQM may grow and propagate throughout a just-born neutron star. This has been a popular choice for an energy source in GRBs and core-collapse supernovae, therefore it is important to establish its basic features with confidence to build on them.
SQM in protoneutron stars: effects in core-collapse supernovae
==============================================================
As a “natural” environment in which SQM might form, core collapse supernovae has received reasonable attention [@BHV; @Matt; @chinos; @Anand]. Despite of more than three decades of theoretical research and hard numerical modelling, the processes that cause the explosion of massive stars are still not understood ([@Bur]). If, as the more recent and detailed numerical simulations suggest, the neutrino-driven mechanism works on special conditions only, the current paradigm for explaining massive star explosions would have to be deeply revised. “Conventional” physics has now turned attention to the role of rotation inside the progenitor and magnetic fields [@SN1; @SN2], possibly relating this problem to the GRB one [@Massimo]. Although it is still too early for making definitive conclusions, investigations including the possible transition to deconfined QCD phases may be relevant to this problem. The first studies of SQM in supernovae ([@Mac; @BHV; @LBV; @DPL; @Anand]) showed that this hypothetic subnuclear energy source is more than adequate to contribute to the explosion, and that some observed characteristics in the neutrino emission of SN1987A may be naturally explained within this scenario ([@Hat; @SatSuz; @Mac]) (a second peak in the neutrino emission is naturally predicted in these models, and such signal has been tentatively associated to the late neutrinos from SN1987A detected by Kamiokande, which have to be otherwise interpreted as a statistical gap within the current paradigm).
From a wide perspective, supernovae are perhaps the only astrophysical events in which we could have the possibility of making a “multiwavelength” detection (neutrinos, various electromagnetic wavelengths, gravitational waves) of the process of SQM formation. However, these calculations are still in the infancy, and just bold expectations have been formulated. Some specific simulations [@FryerWoos] have addressed (negatively) a few questions posed in GRB models. In addition, a firmer detailed observational background would be needed, which imperatively needs the occurrence of a number of supernova explosions in the neighborhood of our galaxy, and thus is out of any human control (in turn the instrumentation must be improved greatly). Second, although the general picture of SQM formation in supernovae has been qualitatively constructed, no systematic calculations have been made. There are also many unresolved questions related to strong interactions at high densities, which introduce an uncomfortable degree of uncertainty in all conclusions. We shall attempt below to describe the basics of the SQM propagation problem, a subject that has been addressed in the literature over the years from the kinetic/energetics point of view [@Olinto; @MadsenOlesen; @Heiselberg; @Drago; @hindues], but has a high degree of complexity from the coupling to hydrodynamics, much in the same way thermonuclear supernovae do. We will be guided by the work done in the latter problem, even though most of our discussion is new (i.e. unpublished) for the specific problem of SQM propagation. We shall later attempt to sketch the effects of the magnetic fields for the propagation, which leads to a tentative classification of the different phenomenological events.
SQM combustion dynamics: early stages
=====================================
As discussed and agreed in the literature, a seed of SQM must become active or form following the standard bounce onto the former iron core. We shall not address this problem of the seed here, just assuming that by some of the proposed mechanisms [@AFO; @Mac] the seed of SQM is present within $\sim \,
seconds$ after the bounce (if the quick appearance is bypassed, a late conversion could ensue [@Bombaci] but without effects in supernovae). The neutron-SQM interface must then propagate outwards powered by the energy release of converted neutrons, much in the same way as a laboratory combustion.
It seems reasonable to assume the combustion to begin as a [*laminar*]{} deflagration, in which the diffusion of $s$ quarks set the scale for the flame length $l_{th}$. This has been actually the subject of early calculations [@varii; @Olinto; @MadsenOlesen; @Heiselberg], in which a plane front approximation was used to obtain the laminar velocity ${\bf u}_{lam}$ as a function of temperature, density and other relevant quantities. The result ${\bf u}_{lam} \leq \, 10^{4} cm \, s^{-1}$ suggested that a just-born NS would convert to a SS in $\sim \, 100 \, s$ or so. From the combustion theory point of view this is equivalent to decouple completely the kinetics of the burning from the hydrodynamics of the flow in the star. Nevertheless, the reasonable convergence of several approaches to the calculation of ${\bf u}_{lam}$ gives some confidence that the result is reliable within the approximations.
In a situation as such (a combustion starting around the center and propagating outwards), it has been known for many years [@Darrieus; @Landau] that small perturbations are unstable for all wavelengths at a linear level. In fact Horvath and Benvenuto [@HB] calculated the perturbation growth for this specific problem with the resulting condition
j\^[4]{} < 4 g \_[1]{}\^[2]{} \_[2]{}\^[2]{} [1]{}
where $j$ is the mass flux onto the flame, $\sigma$ the surface tension, $g$ the local gravitational acceleration and $\rho_{1},
\rho_{2}$ the densities of the “fuel” (neutron matter) and “ashes” (SQM) respectively. As it stands, this is impossible to satisfy for [*any*]{} deflagration (in particular, the laminar), because by its very definition $j^{2}=(P_{2}-P_{1})\rho_{2}\rho_{1}/(\rho_{2} - \rho_{1})$, and thus a deflagration which must obey $P_{2} < P_{1}$ and $\rho_{2}
< \rho_{1}$ making the r.h.s a negative number. This way, the flame wrinkles in a timescale $\leq$ the dynamical timescale $\tau_{dyn} \sim \, 10^{-3} \, s$ (as appropriate in a protoneutron star). Thus, the strong statement made by Landau and Darrieus is confirmed at the linear level.
Numerical calculations of the Landau-Darrieus instability beyond the linear level [@RoepkeHill] show the formation of [*cusps*]{}, leading to quadratic and higher-order terms in the dispersion relation and stabilizing the flame [@Zeld]. The flame acquires a cellular shape and accelerates, since the contact area between the fuel and ashes increases. The stationary, scale-invariant amplitude of this cusps leads to an acceleration of the flame, with velocity described in this regime as
\_[cell]{} = [**u**]{}\_[lam]{} [(1 + 0.4(1 - )\^[2]{})]{}
with $\mu = \rho_{2}/\rho_{1}$. The flame velocity is higher than in the laminar regime by a modest amount for all reasonable compression ratios $\mu$. An alternative cellular flame model has been developed by Blinnikov and Sasorov [@BlinSas]. They observe that the wrinkled flame can be described with a fractal model, which in the 2-D case yields
\_[cell]{} \_[lam]{} [( [l]{})\^[D\_[cell]{}-2]{}]{}
with $D_{cell}$ the fractal dimension of the surface and $l_{crit}$ a suitable minimum length. A calculation of the latter quantity finally yields
\_[cell]{} \_[lam]{} [( [l\_[max]{}]{})\^[0.6(1-)\^[2]{}]{}]{}
where we have imposed the radial distance $l_{max}$ as the maximum scale for which this theory is valid. Arguments related to the propagation of L-D unstable flames suggest that $l_{crit}$ may be identified with the [*Markstein*]{} length [@Markstein], or at least $\sim \, 100 l_{th}$. Eq. (4) leads to the same conclusion as before: there is a modest acceleration of the flame and stabilization at the small scales.
In Ref.[@HB], the extreme assumption that velocity of the flame can not become supersonic, we obtained a (small) maximum length for this regime to hold. This should be rather interpreted as the scale beyond which the above L-D description breaks down definitely, due to combined additional physical effects that we now address.
Since the gravitational pull is always being exerted onto the flame, one could have anticipated that the cell structure can not be scale-invariant indefinitely, and in fact disruption of the bubbles does occur [@NieHill95]. A [*turbulent cascade*]{} dominates the burning above certain length, which can be estimated from the point when the velocity of turbulent fluctuations ${\bf
u'}(l)$ becomes equal to ${\bf u}_{cell}$. This defines the so-called [*Gibson scale*]{} $l_{gib}$ [@Peters]. Imposing a Kolmogorov spectrum (it is now established that this is more accurate than the so-called Bolgiano-Obukhov spectrum for 3D, whereas the latter is relevant for 2D models) ${\bf u'}(l) \propto
l^{1/3}$, it can be shown that the scaling of $l_{gib}$ is
l\_[gib]{} \^[3]{}
where the fluctuations have to be normalized to the largest scales $L$ encountered in the system. Given that the turbulence itself can not become supersonic (the speed of sound is already $\sim \,
c$ in the problem), and using the former value ${\bf u}_{cell}
\geq 10^{4} cm \, s^{-1}$, we obtain for $l_{gib}$ the value of $\sim 10^{-4} cm$ as an extreme upper limit, and decreasing with time. This is, however, initially much larger than $l_{th}$ in the diffusive regime, and allows a classification of the burning into the [*flamelet*]{} regime: while the flame propagation is still determined by diffusion, the [*total burning*]{} is in turn controlled by turbulence in a turbulent region called the [*flame brush*]{}. In the flamelet regime, for all scales $\gg
l_{gib}$, the turbulent velocity ${\bf u}_{turb}$ and front width $l_{turb}$ are determined by the Kolmogorov spectrum at the larger scales. The important point to stress here is that the turbulent eddy turnover controls the transport and fuel consumption, quite unlikely a pure laminar regime [@Kerstein; @Clavin]. Diffusion processes do not play the dominant role once the flamelet regime is achieved, quickly after the start of the combustion. We note that if $l_{gib}$ decreases below the value of $l_{th}$ one can no longer talk of a laminar regime and the burning is likely described by the [*distributed*]{} regime, in which turbulent eddies disrupt the flame and dominate the burning on macroscopic and microscopic scales. We shall assume that the flamelet regime exists and proceed to describe the large-scale physics, keeping in mind the possibility of being bypassed in favor of the distributed regime.
SQM combustion dynamics: turbulent large-scale regime
=====================================================
While at the small scales L-D instability affects the flame eventually leading to the flame brush in the way described above, on still larger scales, buoyancy of hot burned fuel (SQM) dominates the dynamics of the process as a consequence of the Rayleigh-Taylor instability. In fact, one obtains essentially the R-T results by letting ${\bf u}_{lam} \rightarrow 0$ in the L-D analysis. The classical solution of this problem [@Chandra] is well-known and indicates that in the linear regime the perturbations grow exponentially with a time scale $\tau_{RT} =
(4\pi l/g)^{1/2} (\rho/\Delta \rho)^{1/2}$, with $\Delta \rho =
\rho_{1}-\rho_{2}$. After the modes attain amplitudes similar to the originally unperturbed, the merging/fragmentation of the bubbles and the shear [*Kelvin-Helmholtz*]{} instability between bubble surfaces give rise to a turbulent mixing layer. In models with a single bubble scale, the velocity is
\_[RT]{} [1]{} [1]{}
with $At$ the [*Atwood number*]{} and $l$ the radius of the tube in the experiment [@DaviesTaylor]. In astrophysical problems a single-scale expression is seldom enough to describe the intrinsic multi-scale system, and a 1-D model containing most of the relevant physics, the so-called [*Sharp-Wheeler*]{} model [@SW] is widely used to calculate evolution of the flames. In this picture the combustion front advances into the cold unburned fuel with a speed of
\_[SW]{} \~ [1]{} (1 - ) g t
It is clear that the bubble radius evolves linearly with the distance to the center. The Sharp-Wheeler speed eq.(7) can be identified with the effective speed of the burning provided the latter is completed inside the R-T mixing zone.
Fractal models have also been employed as an alternative description for the R-T regime [@TimWoos; @Ghezzi], with a generic prediction that can be summarized as
\_[R-T]{} = [**u**]{}\_[lam]{} [( [L]{} )]{}\^[n/2]{}
where $n=2(D-2)$ relates the index to the fractal index $D$, $L$ is the scale at which the turbulent velocities equal the R-T instability velocity, and $l_{min}$ is the smallest scale that can still deform the flame front (bounded from below by the Gibson scale defined above).
For both the Sharp-Wheeler model eq.(7) and the fractal model of eq.(8) the velocity increase is very large respect to the “kinetic” laminar models for the same problem. This is quite analogous to the carbon burning regime in type I supernova models, in which all the hydrodynamical aspects are being considered together with the reaction kinetics. It is important to remark that in all the cases the flames can be still defined quite properly, and that energetics determined by Hugoniot curves are still valid, as they should.
As performed in Type I SN studies, we plot in Fig. 1 the relevant velocities for the burning flame as a function of the scale. From a simple inspection of this figure, it is clear that the $n
\rightarrow SQM$ combustion should accelerate substantially when evolving at relatively low radii (certainly $\ll \, 1 km$, still deep in the stellar interior. Further evolution of the flame will be discussed below, ending with some of the expected consequences and phenomenological features.
SQM combustion dynamics: distributed regime and the transition to the detonations branch (DDT)
==============================================================================================
The evolution of the flames described below is now quite well established and substantiated by several numerical simulations. One may wonder about the final outcome of the burning process when the flame is well within the R-T stage. The possibility that turbulence disrupts the flame on microscopic scales, which would not be well-defined any more, can be adopted as a rough intuitive description of the [*distributed regime*]{}. In the latter mixed regions of fuel and ashes burn in regions that have a distribution of temperatures interact strongly with the turbulence. Alternatively, the combustion may reach the edge of the star without reaching the distributed regime.
More rigourously, the distributed regime can be characterized by the inequality $l_{gib} \, < \, l_{th}$. It is not clear whether this condition is achieved in the $n \rightarrow \, SQM$ conversion. As suggested above, it may be achieved directly in the early stages. However, and in spite that $l_{gib}$ decreases along the propagation, [**u’**]{} is clearly bounded from above by $c$. Therefore, $l_{gib}$ may be short for the distributed regime to be reached if it is not reached in the early stages, and this is a point that needs a detailed investigation.
One may nevertheless entertain the possibility of a distributed regime in the problem because it is one of the expected pathways to the [*detonation*]{} branch of the combustions. In these scenarios a detonation (self-propagated combustion mediated by a shock) can start, for example, by means of the Zel’dovich gradient mechanism [@Zel2]. For this to occur, a macroscopic region of the mixed fuel/ashes should be able to burn “at once” (i.e. allowing a supersonic phase velocity), which requires a very shallow temperature gradient $\nabla T$. It is not known how large the critical macroscopic region should actually be, detailed calculations show that its value for the WD carbon burning problem is $L_{c} \sim 10^{4} cm$, and it is likely much smaller in our problem. Estimations of the size of the distributed flames yield essentially $l_{dist} = \alpha l_{th} Ka$, where $\alpha$ is a pure number and $Ka$ is the [*Karlovitz number*]{}, used in turbulence studies as a measure of the quotient of diffusive to eddy turnover time. Physically, if $l_{dist}$ can stretch to reach the $L_{c}$ value, the system would satisfy at least a necessary condition for a transition to detonation (since the deflagrations come first, this is call in the literature as Deflagration-to-Detonation Transition, or DDT, [@Khlo]). This condition can be combined with the expression $l_{gib} =
l_{th}/Ka^{2}$ to yield the relation
\^[1/3]{} Ka [( [L\_[c]{}]{} )]{}\^[1/3]{}
converted into a bound on $\alpha$ when we observe that the distributed regime starts at $Ka \, > \, 1$. $L_{c}$ values larger than $\sim 10 \, cm$ would not allow the burning to become a detonation (DDT) by the Zel’dovich gradient mechanism. Thus, a necessary, but not sufficient, condition for the DDT can be established whenever $L_{c} \, \leq \, 10 \, cm$.
Another condition for DDT within the gradient mechanism is related to the hierarchy of time scales of mixing, burning and dynamical. Contrary to the WD explosion problem, we have already seen that $\tau_{dyn}$ is always much longer than $\tau_{burn}$ (identified with the weak interaction time scale $\tau_{W} \, \sim 10^{-8} \,
s)$ to create the strangeness). Therefore we have
\_[mix]{} \_[W]{} \_[dyn]{}
which yields, after substituting
\_[W]{}
using for the turbulent velocity fluctuations the estimate ${\bf
u'}(L) = (1/2) \sqrt {g_{eff}L}$ [@NiemWoos], we obtain an [*upper*]{} bound on $L_{c}$
L\_[c]{} 10\^[-5]{} (1 - ) cm
This is a small length over which to mix fluids, and would make the former condition on the Karlovitz number (eq. 9) irrelevant, likely leading to a DDT phenomenon immediately. Other mechanisms for DDT do exist, but is too difficult to discuss them in connection with our problem at this stage.
From all the above discussion we believe it is clear that the examination of the laminar diffusive regime is just a part of the whole very complex problem. The full evolution of the burning $n
\, \rightarrow \, SQM$ can be accurately described by using the so-called [*reactive Euler equations*]{}
+ [. ()]{} = 0
+ [. ()]{} + P = 0
+ [. ((E+P) [**u**]{})]{} = 0
+ [([**u**]{}.)X\_[i]{}]{} = R (T, , X\_[i]{})
with $X_{i}$ are the fraction of each quarks and the reaction rates $R (T, \rho, X_{i})$ have to be calculated at finite temperature for the dense environment (see, i.e. [@Anand]). Due to enormously disparate length scales, ranging from $\sim$ few $fm$ to perhaps $\sim km$, a model of the flame can facilitate the calculations, otherwise it is known that resolving the full structure demands a huge computational investment [@WoosEtAlnew].
Role of $n \rightarrow SQM$ conversion in supernovae
====================================================
In the original proposal [@Mac; @BHV] of a fast combustion mode in supernovae, a newtonian calculation was employed to estimate the dynamical quantities, in particular the energy that could be transferred to the outer layer of a stalled shock in a massive star. We have seen that a complex but quick sequence of phenomena affects the flame, even if initially starts as a slow laminar combustion. If energy can not be directly transferred to the outer layers, SQM formation may still be important because of the production of neutrinos by appropriate reactions in the deconfined phase. The binding energy of the strange star has to be released as well [@BomDat], much in the same way as the binding energy of the neutron star in the standard picture. Although new fresh neutrinos could in principle produce a late revival of the stalled shock wave, other features than the total released energy are essential such as spectral features of the neutrino emission, and more importantly (if the transition happens to be somewhat delayed) the exact time of its occurrence, since if it occurs too late there will be no way to explode the star by the shock reheating mechanism at all.
While it is still not clear whether the detonation mode is feasible, since it requires fast transport of heat to sustain the front and a working DDT mechanism (if it is not initiated “directly”), assuming the latter case, and since the conversion is not expected to be exothermic all the way down to zero pressure it is unavoidable that a detonation will become a standard shock wave beyond some radius (assuming the MIT Bag model for SQM this radius is the one for which $E - 3P= 4B$). This shock wave will propagate outwards and the question is whether or not it will be able to transfer its energy and complete the work unfinished by the unsuccessful prompt shock wave. In turn, a more moderate turbulent combustion (subsonic but still very fast) may be the final outcome instead of a detonation, and its propagation would mix the material on macroscopic scales due to the action of Landau-Darrieus and Rayleigh-Taylor instabilities. Its role in the reenergization of the stalled shock, possibly by neutrino transfer, has not been calculated as yet.
A better understanding of the previous sequence of combustion processes will also give information about the timescale of the conversion of the star, which is closely related to the different observational signals. These calculations also constitute an important task for the near future.
Delayed conversions, compact star structure and gamma-ray bursts
================================================================
Up to now we have considered the hydrodynamics of the reactive flows with the assumption of its occurrence well inside the first seconds after the prompt shock bounce. If the just-born protoneutron stars do not collapse to black holes due to accretion in the early stages [@LatPra], and within the SQM hypothesis, then pure strange stars, made up entirely of strange quark matter from the center to the surface, may be the compact remnants of supernovae. But even in the case of absolute stability, if the transition is [*not*]{} triggered during the supernova explosion, all observed “normal” neutron stars would be in a metastable state, which is quite difficult to imagine because of ISM contamination arguments [@AFO; @Jes; @MTH] and the mismatch $\tau_{conv} \, \ll \,\tau_{star}$ between the timescale in which favorable conditions for conversion occurs $\tau_{conv}$ and the lifetime of the star $\tau_{star}$. According to recent calculations the deconfinement transition is more likely to occur by heating and compression during the Kelvin-Helmholtz phase of proto-neutron star (PSN) evolution (see, for instance, [@LuBe]). If it did not happen there, once the PNS has cooled to temperatures below $\sim \, 1 MeV$, only accretion from a companion star or strangelet contamination would allow the transition (and many barriers may preclude its occurrence), even in the case where it is energetically favored. Thus, the existence of strange stars is determined not only by fundamental questions concerning the true ground state of dense matter but also by the exact physical conditions in the specific astrophysical environments together with the plausibility of the conversion mechanisms in these situations.
The SQM conversion has been repeteadly associated with gamma-ray bursts. Many works in the past have explored the idea that the conversion of NM into SQM in NSs may be an energy source for GRBs ([@AFO1; @MX; @Hae; @CD; @BomDat; @Chinos2; @Rach; @PacHaens; @Berezhiani]). These models mostly address spherically symmetric conversions of the whole NS rendering isotropic gamma emission. Accumulating observational evidence suggests that at least “long” GRBs are strongly asymmetric, jet-like outflows, a feature that needs some crucial ingredient in the SQM physics formation/propagation to proceed. The “short” burst subclass was not obviously asymmetric prior to HETE2 and SWIFT data, but now evidence has mounted for a substantial asymmetry (but not extreme) in them. The association of Type Ib/c with a few GRBs has reinforced the investigation of underlying explosion mechanisms, and the absence of a temporal break in most of the light curves (interpreted in terms of a collimated jet effect) is a puzzling feature that might be related to the total energy budget in a yet unclear resolution.
A new potentially important feature recently recognized in this class of models is that if a conversion to SQM actually begins near the center of an NS, the presence of a moderate magnetic field B ($ \sim 10^{13}$ G) will originate a prompt [*asymmetric*]{} gamma emission, which may be observed as a short, beamed GRB after the recovery of a fraction of the neutrino energy via $\nu {\bar\nu} \rightarrow e^{+}e{-} \rightarrow \gamma
\gamma$ [@Lug]. The basic physical effect is again related to the instabilities described in the former sections: the influence of the magnetic field expected to be present in NS interiors quenches the growth of the hydrodynamic instabilities in the equatorial direction of the star (parallel to the magnetic field) while it allows them to grow in the polar one. As a result, the flame will propagate much faster in the polar direction, and this will result in a strong (transitory) asymmetry in the geometry of the just formed core of hot SQM, which will resemble a cylinder orientated in the direction of the magnetic poles of the NS. While it lasts, this geometrical asymmetry gives rise to a bipolar emission of the thermal neutrino-antineutrino pairs produced in the process of SQM formation. This is because almost all the thermal neutrinos generated in the process of SQM formation will be emitted in a free streaming regime through the polar cap surface, and not in other directions due to the opacity of the matter surrounding the cylinder. The neutrino-antineutrino pairs annihilate into electron-positron pairs just above the polar caps of the NS, giving rise to a relativistic fireball, thus providing a suitable form of energy transport and conversion to gamma-emission that may be associated to short gamma-ray bursts. A unifying scheme in which SQM appearance produces spherical ejection phenomena to highly asymmetric gamma beaming, as a more or less continuous function of the magnetic field $B$ and the astrophysical system under examination may be possible, and is tentatively sketched in Table 1.
[**Table 1.**]{} Tentative classification of explosive events due to SQM in several stellar systems
Mag. field (G) Type II SN LMXB-HMXB$^{\ast}$ AIC(?)$^{\dagger}$
-------------------- ------------- -------------------------- --------------------- --
$0 < B < 10^{12}$ “normal” SN spherical,weak short GRB UV-X flash
$B \sim 10^{13}$ bipolar SN bipolar,strong short GRB bipolar UV-X flash
$ B \geq 10^{14}$ ? jet-like,weak short GRB jet-like UV-X flash
$B \gg 10^{15-16}$ – -no SQM formation- –
$\ast$ only if $NM \rightarrow SQM$ conversion is sometimes suppressed when a NS is formed.
$\dagger$ upper limit to the rate $\sim 10^{-4} yr^{-1}
galaxy^{-1}$ needs to be revised if SQM burning occurs modifying nucleosynthetic yields.
We are still very far from a thorough understanding of magnetic field effects, and a reliable simulation is even more challenging than simulating the $B=0$ reactive Euler equations (13-16). However, we believe that it is fair to say that magnetic fields are relevant for the physics of the conversion, even at moderate values. In summary, we may be witnessing an ultimate subnuclear energy source in action, powering SN-GRBs if SQM exists.
Conclusions
===========
We have presented a discussion of the main features of hypothetic $n \rightarrow SQM$ conversions inside neutron stars. We have shown that even if the initial state of the process could be a laminar deflagration, the hitherto ignored hydrodynamic instabilities (Landau-Darrieus and Rayleigh-Taylor) quickly take over and determine the propagation through the vast majority of the star, in a regime of turbulent deflagration [@HB; @Lug]. Models which ignore hydrodynamics altogether or just concentrate on the energy conditions to determine the combustion mode miss completely this important features. In particular, the association of long timescales (up to $10^{3}-10^{4} \, s$) of GRBs based on the identification of a laminar deflagration as the relevant timescale in the process is not tenable [@HZu]. Other proposed models differ in their kinetic aspects, for example, models in which energy is obtained by pairing quarks [@LH2; @LH3; @Hsu; @Ouy] typically operate on strong interaction timescales, and thus may be thought as an isocoric burning, i.e. much faster than the described instability scenario. Still other energy transfer mechanisms are possible [@Xu], and certainly the issue of neutrino transport from the reaction zone ahead has never been addressed in detail [@BenLug], although there is plenty of energy carried by them.
It is still possible that all these regimes could be bypassed in favor of a “prompt” detonation mode started at the very central region, for example, by the sudden conversion of a macroscopic small region, further sending a shock wave with $\sim$ half of the initial overpressure [@Mamaia]. Propagation of such a combustion mode is in principle possible [@Mac; @Tokareva], but more detailed studies have yet to be performed on this problem by coupling properly the energy transport to the structure of the flame front. Models treating the conversion much in the same way as a plain phase transition are even more remotely relevant to the actual physics.
Acknowledgements
================
Along several years of SQM research many people contributed to clarify and explain features of the physics and astrophysics of SQM to the author, and provided guidance in many respects. Among them we wish to acknowledge O.G. Benvenuto, H. Vucetich, G. Lugones, J.A. de Freitas Pacheco and I. Bombaci, colleagues and friends. The authors wish to thank the São Paulo State Agency FAPESP for financial support through grants and fellowships, and the partial support of the CNPq (Brazil).
[99]{}
see for example D.J. Gross, hep-ph/9210207 (1992); F. Wilczek, Int.J.Mod.Phys.A8:1359-1382,1993 and references therein.
G. Baym and S. A. Chin, [*Phys. Lett. B*]{}[**62**]{}, 241 (1976); J. C. Collins and M.J. Perry [*Phys.Rev.Lett.*]{}[**34**]{}, 1353 (1975).
see, for example, C.R. Allton et al., hep-ph/0206200 (2002) and references therein.
S. Ejiri et al., hep-lat/0209012 (2002) and references therein.
see for example, D. Karzeev, talk given at the XX International Symposium on Lattice Field Theory “Lattice 2002”, Cambridge, MA, June 24 - 29, (2002).
J. Frieman, A.V. Olinto, M. Gleiser and C.Alcock, [*Phys. Rev. D*]{}[**40**]{}, 3241 (1989).
A. R. Bodmer, [*Phys. Rev.D*]{}[**4**]{}, 1601 (1971).
S.A. Chin and A.K. Kerman, [*Phys. Rev. Lett.*]{}[**43**]{}, 1292 (1979).
H. Terazawa, Tokyo U. Report INS-336 (1979).
J. D. Björken and L. McLerran, [*Phys. Rev. D*]{}[**20**]{}, 2353 (1979).
E. Witten, [*Phys. Rev. D*]{}[**30**]{}, 272 (1984).
E.Farhi and R.L. Jaffe, [*Phys. Rev.D*]{}[**30**]{}, 2379 (1984).
O.G.Benvenuto, J.E.Horvath and H.Vucetich, [*Int. Jour. Mod. Phys. A*]{} [**4**]{}, 257 (1989).
N.A. Gentile; M.B. Aufderheide; G.J. Mathews; F.D. Swesty and G.M. Fuller, [*Astrophys. J.*]{}[**414**]{}, 701 (1993).
M. Takahara and K. Sato, [*Phys. Lett. B*]{}[**156**]{}, 17 (1985).
J.D. Anand, A.Goyal, V.K. Gupta and S. Singh, ,[*Astrophys. J.*]{}[**481**]{}, 954 (1997).
R.Buras, M. Rampp, H-Th. Janka and K. Kifonidis, [*Phys. Rev. Lett.*]{} [**90**]{}, 241101 (2003).
A. Burrows, L. Dessart, E. Livne, C.D. Ott and J. Murphy, astro/ph0702539 (2007)
A. Mezzacappa, J.M. Blondin, O.E.B. Messer, S.W. Bruenn,in [*AIP Conference Proceedings*]{}[**847**]{}, 179 (2006).
O.G.Benvenuto and J.E.Horvath, [*Phys. Rev. Lett.*]{} [**63**]{}, 716 (1989).
M. Della Valle, [*Chin. J. Astron. Astrophys. Supp.*]{}[**6**]{}, 315 (2006).
G. Lugones, O.G.Benvenuto and H. Vucetich, [*Phys. Rev. D*]{} [**50**]{}, 6100 (1994).
Z. Dai, Q. Peng and T. Lu,[*Astrophys. J.*]{}[**440**]{}, 815 (1995).
T. Hatsuda, [*Mod.Phys.Lett.A*]{}[**2**]{}, 805 (1987).
K. Sato and H. Suzuki, [*Phys.Rev.Lett.*]{}[**58**]{}, 2722 (1987).
C. Fryer and S.E. Woosley, [*Astrophys. J.*]{}[**501**]{}, 780 (1998).
G. Baym, E.W. Kolb, L. McLerran, T.P. Walker and R.L. Jaffe, [*Phys. Lett. B*]{}[**160**]{}, 181 (1985).
A.V. Olinto, [*Phys. Lett. B*]{}[**192**]{}, 71 (1987)
M.L. Olesen and J. Madsen, in [*Proceedings of the Workshop on Strange Quark Matter in Physics and Astrophysics*]{}, eds. J.Madsen e P.Haensel. [*Nuc.Phys.B Proc. Supp.*]{}[**24**]{}, 170 (1991).
H. Heiselberg, G. Baym and C.J. Pethick, in [*Proceedings of the Workshop on Strange Quark Matter in Physics and Astrophysics*]{}, eds. J.Madsen e P.Haensel. [*Nuc.Phys.B Proc. Supp.*]{}[**24**]{}, 144 (1991).
A. Drago, G. Pagliara and I. Parenti, [*Nuc. Phys. A*]{}[**782**]{}, 418 (2007)
A. Bhattacharyya, S. K. Ghosh, P. S. Joardar, R. Mallick and S. Raha, [*Phys. Rev. C*]{}[**74**]{}, 065804 (2006)
C.Alcock, E.Farhi and A.V.Olinto, [*Astrophys. J.*]{}[**310**]{}, 261 (1986).
I. Bombaci, I. Parenti and I. Vidaña, in [*Superdense QCD Matter and Compact Stars*]{}, 353. D. Blaschke and D. Sedrakian (eds)(Springer, Dordretch 2006)
G. Darrieus, unpublished (1938).
L.D. Landau, [*Acta Physicochim. URSS*]{}[**19**]{}, 77 (1944).
J.E. Horvath and O.G. Benvenuto, [*Phys. Lett. B*]{} [**213**]{}, 156 (1988).
F.K. Röpke and W. Hillebrandt, in [*Cosmic Explosions, on the 10th Anniversary of SN1993J*]{} Proceedings of IAU Colloquium 192.J.M. Marcaide and K W. Weiler (eds), 333. (Springer Proceedings in Physics 99. Berlin 2005)
Ya. B. Zel’dovich, [*J. Appl. Mech. Tech. Phys.*]{}[**7**]{}, 68 (1966).
S.I. Blinnikov and P.V. Sasorov, [*Phys. Rev. E*]{}[**53**]{}, 4827 (1996).
G. H. Markstein, [*J. Aeronaut. Sci.*]{}[**18**]{}, 199 (1951).
J. Niemeyer and W. Hillebrandt, [**Astrophys. J.**]{}[**452**]{}, 769 (1995).
N. Peters, in [*Proc. of 21st Int. Sym. Comb.*]{}, 1232 (Combustion Institute, Pittsburg 1988).
J.C. Niemeyer and A. Kerstein, [*New Astron.*]{}[**2**]{}, 239 (1997).
P. Clavin, [*Annu. Rev. Fluid. Mech.*]{}[**26**]{}, 321 (1994).
S. Chandrasekhar, [*Hydrodynamic and Hydromagnetic Stability*]{} (Oxford University Press, Oxford 1961).
R.M. Davies and G. Taylor, [*Proc. Roy. Soc. London*]{}[**200**]{}, 375 (1949).
D.H. Sharp, [*Physica D*]{}[**12**]{}, 3 (1984)
F. Timmes and S. E. Woosley, [*Astrophys. J.*]{}[**396**]{}, 649 (1992).
C.R. Ghezzi, E.M. de Gouveia Dal Pino and J.E. Horvath, [*Astrophys. J. Lett.*]{}[**548**]{}, L193 (2001).
Ya. B. Zel’dovich, V.B. Librovich, G.M. Makhviladze and G.I. Sivashinsky, [*Astronaut. Acta*]{}[**15**]{}, 313 (1970).
A.M. Khokhlov, E.S. Oran and J.C. Wheeler, [*Astrophys. J.*]{}[**478**]{}, 678 (1997).
J. C. Niemeyer and S.E. Woosley, [*Astrophys. J.*]{}[**4575**]{}, 740 (1997).
M. Zingale, A.S. Almgren, J.B. Bell, M.S. Day, C.A. Rendleman and S.E. Woosley, [*Journal of Physics: Conference Series*]{}[**46**]{}, 385 (2006).
I. Bombaci and B. Datta, [*Astrophys. J. Lett.*]{}[**530**]{}, L69 (2000).
J.M. Lattimer and M. Prakash, [*Phys. Repts.*]{}[**333**]{}, 121 (2000).
J. Madsen, [*Phys.Rev.Lett.*]{}[**61**]{}, 2909 (1988).
G.A.Medina-Tanco and J.E.Horvath, [*Astrophys.J.*]{} [**464**]{}, 354 (1996).
G.Lugones and O.G. Benvenuto, [*Mon. Not. R.A.S.*]{}[**304**]{}, L25 (1999).
C.Alcock, E.Farhi and A.V.Olinto, [*Phys. Rev. Lett.*]{}[**57**]{}, 2088 (1986).
F. Ma and B. Xie, [*Astrophys.J.Lett.*]{} [**462**]{}, L63 (1996).
P. Haensel, B. Paczynski and P. Amsterdamski, [*Astrophys.J.*]{} [**375**]{}, 209 (1991).
K.S. Cheng and Z.G. Dai, [*Phys. Rev. Lett.*]{}[**77**]{}, 1210 (1996).
X.Y. Wang, Z.G. Dai, T. Lu, D.M. Wei and Y.F. Huang, [*Astron. Astrophys.*]{}[**357**]{}, 543 (2000).
R. Ouyed and F. Sannino, [*Astron. Astrophys.*]{}[**387**]{}, 725 (2002).
B. Paczynski and P. Haensel, [*MNRAS*]{} [**362**]{}, L4 (2005).
Z. Berezhiani, I. Bombaci, A. Drago, F. Frontera and A. Lavagno, in [*Astronomical Society of the Pacific Conferences*]{}[**312**]{}, 407 (2004).
G. Lugones, C.R. Ghezzi, E.M. de Gouveia Dal Pino and J.E. Horvath, [*Astrophys. J. Lett.*]{}[**581**]{}, L101 (2002).
J.E. Horvath, in [*Proceedings of the Carpathian Summer School of Physics 2005*]{}, 188 (eds. S. Stoica, L. Trasche and R. Tribble). (Singapore: World Scientific 2006)
I. Tokareva, A. Nusser, V. Gurovich and V. Folomeev, [*Int. Jour. Mod. Phys. D*]{} [**14**]{}, 33 (2005) ; I. Tokareva and A. Nusser, [*Phys.Lett. B*]{}[**639**]{}, 232 (2006).
P. Haensel and J.L. Zdunik, in [*SWIFT and GRBs: Unveiling the Relativistic Universe*]{}, Venice, June 5-9, 2006. astro-ph/0701258.
G.Lugones e J.E. Horvath, [*Phys. Rev. D*]{}[**66**]{}, 074017 (2002).
G.Lugones e J.E. Horvath, [*Astron. Astrophys.*]{}[**403**]{}, 173 (2003).
D. K. Hong, S.D.H. Hsu and F. Sannino, [*Phys. Lett. B*]{}[**516**]{}, 362 (2001).
R. Ouyed, R. Raff and C. Vogt, [*Astrophys. J.*]{}[**632**]{}, 1001 (2005).
A. Chen and R.Xu, astro-ph/0605285 (2006).
O.G. Benvenuto and G. Lugones, [*MNRAS*]{}[**304**]{}, L25 (1999).
At a given instant the regimes dominating the burning are shown as a function of the lengthscale. Below $\sim \, 100 \, l_{th}$ the laminar flame ensues. Cells appear above that scale and produce a weakly-dependent velocity (as described by fractal models, for instance). Above $l_{gib}$ cellular stabilization fails and above a transition scale the buoyancy ultimate dominates the burning ${\bf u_{RT}} \propto \, l^{1/2}$. It should be kept in mind that the distributed regime may be directly reached, disrupting the flame that no longer follows the regimes of Fig.1
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '> In this paper, we investigate the usage of autoencoders in modeling textual data. Traditional autoencoders suffer from at least two aspects: scalability with the high dimensionality of vocabulary size and dealing with task-irrelevant words. We address this problem by introducing supervision via the loss function of autoencoders. In particular, we first train a linear classifier on the labeled data, then define a loss for the autoencoder with the weights learned from the linear classifier. To reduce the bias brought by one single classifier, we define a posterior probability distribution on the weights of the classifier, and derive the marginalized loss of the autoencoder with Laplace approximation. We show that our choice of loss function can be rationalized from the perspective of Bregman Divergence, which justifies the soundness of our model. We evaluate the effectiveness of our model on six sentiment analysis datasets, and show that our model significantly outperforms all the competing methods with respect to classification accuracy. We also show that our model is able to take advantage of unlabeled dataset and get improved performance. We further show that our model successfully learns highly discriminative feature maps, which explains its superior performance.'
author:
- |
Shuangfei Zhai Zhongfei (Mark) Zhang\
Computer Science Department, Binghamton University\
4400 Vestal Pkwy E, Binghamton, NY 13902\
[email protected] [email protected]\
bibliography:
- 'mybib.bib'
title: Semisupervised Autoencoder for Sentiment Analysis
---
Introduction
============
In machine learning, documents are usually represented as Bag of Words (BoW), which nicely reduces a piece of text with arbitrary length to a fixed length vector. Despite its simplicity, BoW remains the dominant representation in many applications including text classification. There has also been a large body of work dedicated to learning useful representations for textual data [@vsm; @lda; @lsa; @word2vec; @xavier]. By exploiting the co-occurrence pattern of words, one can learn a low dimensional vector that forms a compact and meaningful representation for a document. The new representation is often found useful for subsequent tasks such as topic visualization and information retrieval. In this paper, we investigate the application of one of the most popular representation learning methods, namely autoencoders [@bengio], to learn task-dependent representations for textual data. Our model differs from most of the existing work as it naturally incorporates label information into its objective function, which allow the learned representation to be directly coupled with the task of interest.
In this paper we focus on a specific class of task in text mining: Sentiment Analysis (SA). We further focus on a special case of SA as a binary classification problem, where a given piece of text is either of positive or negative attitude. This problem is interesting largely due to the emergence of online social networks, where people consistently express their opinions about certain subjects. Also, it is easy to obtain a large amount of clean labeled data for SA by crawling reviews from websites such as IMDB or Amazon. Thus, SA is an ideal benchmark for evaluating text classification models (and features).
Autoencoders have attracted a lot of attention in recent years as a building block of Deep Learning [@bengio]. They act as the feature learning methods by reconstructing inputs with respect to a given loss function. In a neural network implementation of autoencoders, the hidden layer is taken as the learned feature. While it is often trivial to obtain good reconstructions with plain autoencoders, much effort has been devoted on regularizations in order to prevent them against overfitting [@bengio; @dae; @cae]. However, little attention has been devoted to the loss function, which we argue is critical for modeling textual data. The problem with the commonly adopted loss functions (squared Euclidean distance and element-wise KL Divergence, for instance) is that they try to reconstruct all dimensions of input independently and undiscriminatively. However, we argue that this is not the optimal approach when our interest is text classification. The reason is two folds. First, it is well known that in natural language the distribution of word occurrences follows the power-law. This means that a few of the most frequent words will account for most of the probability mass of word occurrences. An immediate result is that the Autoencoder puts most of its effort on reconstructing the most frequent words well but (to a certain extent) ignores the less frequent ones. This may lead to a bad performance especially when the class distribution is not well captured by merely the frequent words. For sentiment analysis, this problem is especially severe because it is obvious that the truly useful features (words or phrases expressing a clear polarity) only occupy a small fraction of the whole vocabulary; and reconstructing irrelevant words such as ’actor’ or ’movie’ very well is not likely to help learn more useful representations to classify the sentiment of movie reviews. Second, explicitly reconstructing all the words in an input text is expensive, because the latent representation has to contain all aspects of the semantic space carried by the words, even if they are completely irrelevant. As the vocabulary size can easily reach the range of tens of thousands even for a moderate sized dataset, the hidden layer size has to be chosen very large to obtain a reasonable reconstruction, which causes a huge waste of model capacity and makes it difficult to scale to large problems.
In fact, the reasoning above applies to all the unsupervised learning methods in general, which we argue is one of the most important problems to address in order to learn task-specific representations. This naturally leads us to the semisupervised approach, where label information is introduced to guide the feature learning procedure. In particular, we propose a novel loss function for training autoencoders that are directly coupled with the classification task. We first train a linear classifier on BoW, then a Bregman Divergence [@bregman] is derived as the loss function of a subsequent autoencoder. The new loss function gives the autoencoder the information about directions along which the reconstruction should be accurate, and where larger reconstruction errors are tolerated. Informally, this can be considered as a weighting of words based on their correlations with the class label: predictive words should be given large weights in the reconstruction even they are not frequent words, and vice versa. Furthermore, to reduce the bias introduced by the linear classifier, we take a Bayesian view by defining a posterior distribution on the weights of the classifier. We then approximate the posterior with Laplace approximation and derive the marginalized loss function for the autoencoder. We show that our model successfully learns features that are highly discriminative with respect to class labels, and also outperform all the competing methods evaluated by classification accuracy. Moreover, the derived loss can also be applied to unlabeled data, which allows the model to learn further better representations.
Model
=====
Denoising Autoencoders {#intro:ae}
----------------------
Autoencoders learn functions that can reconstruct the inputs. They are typically implemented as a neural network with one hidden layer, and one can extract the activation of the hidden layer as the new representation. Mathematically, we are given a collection of data points $X = \{x_i\}, x_i \in R^d, i \in [1, m]$, the objective function of an autoencoder is thus: $$\label{eq:ae}
\begin{split}
\min &\sum_i D(\tilde{x}_i, x_i) \\
s.t. \quad &h_i = g(Wx_i + b), \tilde{x}_i = f(W'h_i + b')
\end{split}$$ where $W \in R^{k \times d}, b \in R^{k}, W' \in R^{d \times k}, b' \in R^{d}$ are the parameters to be learned; $D$ is a loss function, such as the squared Euclidean Distance $\|\tilde{x} - x\|_2^2$; $g$ and $f$ are predefined nonlinear functions, which we set as $g(x) = \max(0, x)$, $f(x) = (1 + exp(-x))^{-1}$ in this paper; $h_i$ is the learned representation; $\tilde{x}_i$ is the reconstruction. A common approach is to use tied weights by setting $W = W'$; this usually works better as it speeds up learning and prevents overfitting at the same time. For this reason, we always use tied weights in this paper.
Autoencoders transform an unsupervised learning problem to a supervised one by the self reconstruction criteria. This enables one to use all the tools developed for supervised learning such as back propagation to efficiently train the autoencoders. Moreover, thanks to the nonlinear functions $f$ and $g$, autoencoders are able to learn non-linear and possibly overcomplete representations, which give the model much more expressive power than their linear counter parts such as PCA (LSA) [@lsa].
In this paper, we adopt one of the most popular variants of autoencoders, namely Denoising Autoencoder. Denoising Autoencoder works by reconstructing the input from a noised version of itself. The intuition is that a robust model should be able to reconstruct the input well even in the presence of noises, due to the high correlation among features. For example, imagine deleting or adding a few words from/to a document, the semantics should still remain unchanged, thus the autoencoder should learn a consistent representation from all the noisy inputs. In the high level, Denoising Autoencoders are equivalent to ordinary autoencoders trained with dropout [@dropout], which has been shown as an effective regularizer for (deep) neural networks. Formally, let $q(\bar{x} | x)$ be a predefined noising distribution, and $\bar{x}$ be a noised sample of $x$: $\bar{x} \sim q(\bar{x} | x)$. The objective function takes the form of sum of expectations over all the noisy samples: $$\label{eq:daeloss}
\begin{split}
\min &\sum_i \mathrm{E}_{q(\bar{x}_i | x_i)}D(\tilde{x}_i, x_i) \\
s.t. \quad & h_i = g(W\bar{x}_i + b), \tilde{x}_i = f(W'h_i + b')
\end{split}$$ where we have slightly overloaded the notation to let $\tilde{x}_i$ denote the reconstruction calculated from the noised input $\bar{x}_i$. While the marginal objective function requires infinite many noised samples per data point, in practice it is sufficient to simulate it stochastically. That is, for each example seen in the stochastic gradient descent training, we randomly sample a $\bar{x}_i$ from $q(\bar{x}_i | x_i)$ and calculate the gradient with ordinary back propagation.
Loss Function as Bregman Divergence {#intro:bregman}
-----------------------------------
We then discuss the proper choice of the loss function $D$ in as a specific form of Bregman Divergence. Bregman Divergence [@bregman] generalizes the notion of distance in a $d$ dimensional space. To be concrete, given two data points $\tilde{x}, x \in R^d$ and a convex function $f(x)$ defined on $R^d$, the Bregman Divergence of $\tilde{x}$ from $x$ with respect to $f$ is: $$D_f(\tilde{x}, x) = f(\tilde{x}) - (f(x) + \nabla{f(x)}^T(\tilde{x} - x)).$$ Namely, Bregman Divergence measures the distance between two points $\tilde{x}, x$ as the deviation between the function value of $f$ and the linear approximation of $f$ around $x$ at $\tilde{x}$.
Two of the most commonly used loss functions for autoencoders are the squared Euclidean distance and element-wise KL divergence. It is not difficult to verify that they both fall into this family by choosing $f$ as the squared $\ell_2$ norm and the sum of element-wise entropy respectively. What the two loss functions have in common is that they make no distinction among dimensions of the input. In other words, each dimension of the input is pushed to be reconstructed equally well. While autoencoders trained in this way have been shown to work very well on image data, learning much more interesting and useful features than the original pixel intensity features, they are less appropriate for modeling textual data. The reason is two folds. First, textual data are extremely sparse and high dimensional, where the dimensionality is equal to the vocabulary size. To maintain all the information of the input in the hidden layer, a very large layer size must be adopted, which makes the training cost extremely large. Second, ordinary autoencoders are not able to deal with the power law of word distributions, where a few of the most frequent words account for most of the word occurrences. As a result, frequent words naturally gain favor to being reconstructed accurately, and rare words tend to be reconstructed with less precision. This problem is also analogous to the imbalanced classification setting. This is especially problematic when frequent words carry little information about the task of interest, which is not uncommon. Examples include stop words (*the, a, this, from*) and topic related terms (*movie, watch, actress*) in a movie review sentiment analysis task.
Semisupervised Autoencoder with Bregman Divergence {#model}
--------------------------------------------------
To address the problems mentioned above, we propose to introduce supervision to the training of autoencoders. To achieve this, we first train a linear classifier on Bag of Words, and then use the weight of the learned classifier to define a new loss function for the autoencoder. Now let us first describe our choice of loss function, and then elaborate the motivation later: $$\label{eq:df}
D(\tilde{x}, x) = (\theta^T(\tilde{x} - x))^2.$$ where $\theta \in R^d$ are the weights of the linear classifier, and we have omitted the bias for simplicity. Before we delve into more details, note that Equation is a valid distance, as it is non-negative and reaches zeros if and only if $\tilde{x} = x$. Moreover, the reconstruction error is only measured after projecting on $\theta$; this guides the reconstruction to be accurate only along directions where the linear classifier is sensitive to. Note also that Equation on the one hand uses label information ($\theta$ has been trained with labeled data), on the other hand no explicit labels are directly referred to (only requires $x_i$). Thus one is able to train an autoencoder on both labeled and unlabeled data with the loss function in Equation . This subtlety distinguishes our method from pure supervised or unsupervised learning, and allows us to enjoy the benefit from both worlds.
As a design choice, we consider SVM with squared hinge loss (SVM2) and $\ell_2$ regularization as the linear classifier, but other classifiers such as Logistic Regression can be used and analyzed similarly. Let us denote $\{x_i\}, x_i \in R^d$ as the collection of samples, and $\{y_i\}, y_i \in \{1, -1\}$ as the class labels; the objective function SVM2 is: $$\label{eq:svm}
L(\theta) = \sum_i (\max(0, 1 - y_i \theta^T x_i))^2 + \lambda \|\theta\|^2.$$ Here $\theta \in R^d$ is the weight; $\lambda$ is the weight decay parameter.
Equation is continuous and differentiable everywhere with respect to $\theta$, so the model can be easily trained with stochastic gradient descent. The next (and most critical) step of our approach is to transfer label information from the linear classifier to the autoencoder. With this in mind, we examine the loss induced by each sample as a function of the input, while with $\theta$ fixed: $$f(x_i) = (\max(0, 1 - y_i \theta^T x_i))^2$$ Note that $f(x_i)$ is defined on the input space $R^d$, which should be contrasted with $L(\theta)$ in Equation which is a function of $\theta$. We are interested in $f(x_i)$ because if we consider moving each input $x_i$ to $\tilde{x}_i$, $f(x_i)$ indicates the direction along which the loss is sensitive to. If we think of $\tilde{x}$ as the reconstruction of $x_i$ obtained from an autoencoder, a good $\tilde{x}_i$ should be in a way such that the deviation of $\tilde{x}_i$ from $x_i$ is small evaluated by $f(x_i)$. In other words, we would like $\tilde{x}_i$ to still be correctly classified by the pretrained linear classifier. Therefore, $f(x_i)$ should be a much better function to evaluate the deviation of two samples. if we can derive a Bregman Divergence from $f(x_i)$ and use it as the loss function of the subsequent autoencoder training, the autoencoder should be guided to give reconstruction errors that do not confuse the classifier. Note that $f(x_i)$ is a quadratic function of $x_i$ whenever $f(x_i) > 0$, so we only need to derive the Hessian matrix in order to achieve the Bregman Divergence. The Hessian follows as:
$$\label{eq:hessian}
H(x_i) =
\begin{cases}
\theta \theta^T, & \text{if}\ 1 - y_i\theta^Tx_i > 0 \\
0, & \text{otherwise}.
\end{cases}$$
Recall that for a quadratic function with Hessian matrix $H$, the Bregman Divergence is simply $ (\tilde{x} - x)^TH(\tilde{x} - x)$; then we have: $$\label{eq:df0}
D(\tilde{x}_i, x_i) =
\begin{cases}
(\theta^T(\tilde{x}_i - x_i))^2, & \text{if}\ 1 - y_i\theta^Tx_i > 0 \\
0, & \text{otherwise}
\end{cases}$$ In words, Equation says that we measure the reconstruction loss for difficult examples (those that satisfy $1 - y_i\theta^Tx_i > 0$) with Equation ; and there is no reconstruction loss at all for easy examples. This discrimination is undesirable, because in this case the Autoencoder would completely ignore easy examples, and there is no way to guarantee that the $\tilde{x}_i$ can be correctly classified. Actually, this split is just an artifact of the hinge loss and the asymmetrical property of Bregman Divergence. Hence, we perform a simple correction by ignoring the condition in Equation , which basically pretends that all the examples induce a loss. This directly yields the loss function as in Equation .
The Bayesian Marginalization
----------------------------
In principle, one may directly apply Equation as the loss function in place of the squared Euclidean distance and train an autoencoder. However, doing so might introduce a bias brought by one single classifier. As a remedy, we resort to the Bayesian approach, which defines a probability distribution over $\theta$. Although SVM2 is not a probabilistic classifier like Logistic Regression, we can borrow the idea of Energy Based Model [@bengio] and use $L(\theta)$ as the negative log likelihood of the following distribution:
$$p(\theta) = \frac{\exp(-\beta L(\theta))}{\int \exp(-\beta L(\theta)) d\theta}$$
where $\beta > 0$ is the temperature parameter which controls the shape of the distribution $p$. Note that the larger $\beta$ is, the sharper $p$ will be. In the extreme case, $p(\theta)$ is reduced to a uniform distribution as $\beta$ approaches $0$, and collapses into a single $\delta$ function as $\beta$ goes to positive infinity.
Given $p(\theta)$, we rewrite Equation as an expectation over $\theta$:
$$\label{eq:int}
\begin{split}
D(\tilde{x}, x) = \mathrm{E}_{\theta \sim p(\theta)} (\theta ^T(\tilde{x} - x))^2
= \int (\theta^T(\tilde{x} - x))^2 p(\theta) d\theta.
\end{split}$$
Obviously there is now no closed form expression for $D(\tilde{x}, x)$. To solve it one could use sampling methods such as MCMC, which provides unbiased estimates of the expectation but could be slow in practice. Instead, we use the Laplace approximation, which approximates $p(\theta)$ by a Gaussian distribution $\tilde{p}(\theta) = \mathcal{N}(\hat{\theta}, \Sigma)$. As estimating the full covariance matrix is prohibitive, we further constrain $\Sigma$ to be diagonal. The benefit of doing so is that the expectation can now be computed directly in closed form. To see this, by simply replacing $p(\theta)$ with $\tilde{p}(\theta)$ in Equation : $$\label{eq:int}
\begin{split}
D(\tilde{x}, x) = &\mathrm{E}_{\theta \sim \tilde{p}(\theta)} (\theta^T(\tilde{x} - x))^2\\
= &(\tilde{x} - x)^T \mathrm{E}_{\theta \sim \tilde{p}(\theta)}(\theta \theta^T) (\tilde{x} - x) \\
= &(\tilde{x} - x)^T (\hat{\theta} \hat{\theta}^T + \Sigma) (\tilde{x} - x) \\
= &(\hat{\theta}^T(\tilde{x} - x))^2 + ({\Sigma}^{\frac{1}{2}}(\tilde{x} - x))^T({\Sigma}^{\frac{1}{2}}(\tilde{x} - x)).
\end{split}$$ where $D$ now involves two parts, corresponding to the mean and variance term of the Gaussian distribution respectively. Now let us derive $\tilde{p}(\theta)$ for $p(\theta)$. In Laplace approximation, $\hat{\theta}$ is chosen as the mode of $p(\theta)$, which is exactly the solution to the SVM2 optimization problem. For $\Sigma$, we have: $$\label{eq:df1}
\begin{split}
\Sigma = &(diag(\frac{\partial^2{L(\theta)}}{\partial{\theta^2}}))^{-1} \\
= &\frac{1}{\beta}(diag(\sum_i{\mathbb{I}(1 - y_i \theta^Tx_i > 0)} x_i^2))^{-1}
\end{split}$$ Here we have overridden $diag$ but letting it denote a diagonal matrix induced either by a square matrix or a vector; $\mathbb{I}$ is the indicator function; $(\cdot)^{-1}$ denotes matrix inverse. Interestingly, the second term in Equation is now equivalent to the squared Euclidean distance after performing element-wise normalizing the input using all difficult examples. The effect of this normalization is that the reconstruction errors of frequent words are down weighted; on the other hand, discriminative words are given higher weights as they would occur less frequently in difficult examples. Note that it is important to use a relatively large $\beta$ in order to avoid the variance term dominating the mean term. In other words, we need to ensure $p(\theta)$ to be reasonable peaked around $\hat{\theta}$ to effective take advantage of label information.
Experiments {#experiment}
===========
Datasets
--------
We evaluate our model on six Sentiment Analysis benchmarks. The first one is the IMDB dataset [^1] [@mass], which consists of movie reviews collected from IMDB. The IMDB dataset is one of the largest sentiment analysis dataset that is publicly available; it also comes with an unlabeled set which allows us to evaluate semisupervised learning methods. The rest five datasets are all collected from Amazon [^2][@amazon], which corresponds to the reviews of five different products: books, DVDs, music, electronics, kitchenware. All the six datasets are already tokenized as either uni-gram or bi-gram features. For computational reasons, we only select the words that occur in at least $30$ training examples. We summarize the statistics of datasets in Table \[tb:datasets\].
-- --------------------------------------------------------------------- -- -- -- -- --
**IMDB & **books & **DVD & **music & **electronics & **kitchenware\
\# train & 25,000 & 10,000 & 10,000 & 18,000 & 6,000 & 6,000\
\# test & 25,000 & 3,105 & 2,960 & 2,661 & 2,862 & 1,691\
\# unlabeled & 50,000 & N/A & N/A & N/A & N/A & N/A\
\# features & 8,876 & 9,849 & 10,537 & 13,099 & 5,091 & 3,907\
% positive & 50 & 49.81 & 49.85 & 50.16 & 49.78 & 50.08\
************
-- --------------------------------------------------------------------- -- -- -- -- --
: Statistics of the datasets.[]{data-label="tb:datasets"}
Methods
-------
- Bag of Words (BoW). Instead of using the raw word counts directly, we take a simple step of data normalization: $$\label{eq:normalization}
x_{i,j} = \frac{\log(1 + c_{i,j})}{\max_j \log(1 + c_{i,j})}$$ where $c_{i, j}$ denotes the number of occurrences of the $j$*th* word in the $i$*th* document, $x_{i, j}$ denotes the normalized count. We choose this normalization because it preserves the sparsity of the Bag of Words features; also each feature element is normalized to the range $[0, 1]$. Note that the very same normalized Bag of Words features are fed into the autoencoders.
- Denoising Autoencoder (DAE) [@dae]. This refers to the regular Denoising Autoencoder defined in Equation with squared Euclidean distance loss: $D(\tilde{x}, x) = \|\tilde{x} - x\|_2^2$. This is also used in [@xavier] on the Amazon datasets for domain adaptation. We use ReLu $max(0, x)$ as the activation function, and Sigmoid as the decoding function.
- Denoising Autoencoder with Finetuning (DAE+) [@dae]. This denotes the common approach to continue training an DAE on labeled data by replacing the decoding part of DAE with a Softmax layer.
- Feedforward Neural Network (NN). This is the standard fully connected neural network with one hidden layer and random initialization. We use the same activation function as that in Autoencoders, i.e., ReLU.
- Logistic Regression with Dropout (LrDrop) [@lrdrop]. This is a model where logistic regression is regularized with the marginalized dropout noise. LrDrop differs from our approach as it uses feature noising as an explicit regularization. Another difference is that our model is able to learn nonlinear representations, not merely a classifier, and thus is potentially able to model more complicated patterns in data.
- Semisupervised Bregman Divergence Autoencoder (SBDAE). This corresponds to our model with Denoising Autoencoder as the feature learner. The training process is roughly equivalent to training on BoW followed by the training of DAE, except that the loss function of DAE is replaced with the loss function defined in Equation . We cross validate $\beta$ from the set $\{10^{4}, 10^{5}, 10^{6}, 10^{7}, 10^{8}\}$ (note that larger $\beta$ corresponds to weaker Bayesian regularization).
- Semisupervised Bregman Divergence Autoencoder with Finetuning (SBDAE+).
Note that except for BoW and LrDrop, all the other methods require a predefined dimensionality of representation. We use fixed sizes on all the datasets. For SBDAE and NN, a small hidden size is sufficient, so we use $200$. For DAE, we observe that it benefits from very large hidden sizes; however, due to computational constraints, we take $2000$. For BoW, DAE, SBDAE, we use SVM2 as the classifier. All the models are trained with mini-batch Stochastic Gradient Descent with momentum of $0.9$.
Results
-------
-- ----------------------------------------------------------------------------------------- -- -- -- -- -- --
**books & **DVD & **music & **electronics & **kitchenware & **IMDB & **IMDB + unlabled\
BoW & 10.76 & 11.82 & 11.80 & 10.41 & 9.34 & 11.48 &N/A\
DAE & 15.10 & 15.64 & 15.44 & 14.74 & 12.48 & 14.60 &13.28\
DAE+ & 11.40 & 12.09 & 11.80 & 11.53 & 9.23 & 11.48 &11.47\
NN & 11.05 & 11.89 & 11.42 & 11.15 & 9.16 &11.60 & N/A\
LrDrop & 9.53 & 10.95 & 10.90 & **9.81 & **8.69 &10.88 & 10.73\
SBDAE & **9.16 & **10.90 & **10.59 & 10.02 & 8.87 &**10.52 & **10.42\
SBDAE+ & **9.12 & **10.90 & **10.58 & 10.01 & 8.83 &**10.50 & **10.41\
**************************************
-- ----------------------------------------------------------------------------------------- -- -- -- -- -- --
We first summarize the results as in classification error rate in Table \[tb:results\]. First of all, our model consistently beats BoW with a margin, and it achieves the best results on four (larger) datasets out of six. On the other hand, DAE, DAE+ and NN all fail to outperform BoW, although they share the same architecture as nonlinear classifiers. This suggests that SBDAE be able to learn a much better nonlinear feature transformation function by training with a more informed objective (than that of DAE). Moreover, note also that finetuning on labeled set (DAE+) significantly improves the performance of DAE, which is ultimately on a par with training a neural net with random initialization (NN). However, finetuning offers little help to SBDAE, as it is already implicitly guided by labels during the training.
LrDrop is the second best method that we have tested. Thanks to the usage of dropout regularization, it consistently outperforms BoW, and achieves the best results on two (smaller) datasets. Compared with LrDrop, it appears that our model works better on large datasets ($\approx 10 K$ words, more than $10K$ training examples) than smaller ones. This indicates that in high dimensional spaces with sufficient samples, SBDAE benefits from learning a nonlinear feature transformation that disentangles the underlying factors of variation, while LrDrop is incapable of doing so due to its nature as a linear classifier.
As the training of the autoencoder part of SBDAE does not require the availability of labels, we also try incorporating unlabeled data after learning the linear classifier in SBDAE. As shown in Table \[tb:results\], doing so further improves the performance over using labeled data only. This justifies that it is possible to bootstrap from a relatively small amount of labeled data and learn better representations with more unlabeled data with SBDAE.
To gain more insights of the results, we further visualize the filters learned by SBDAE and DAE on the IMDB dataset in Table \[tb:topics\]. In particular, we show the top $5$ most activated and deactivated words of the first $8$ filters (corresponding to the first $8$ rows of $W$) of SBDAE and DAE, respectively. First of all, it seems very difficult to make sense of the filters of DAE as they are mostly common words with no clear co-occurrence pattern. By comparison, if we look at the filters from SBDAE, they are mostly sensitive to words that demonstrate clear polarity. In particular, all the $8$ filters seem to be most activated by certain negative words, and are most deactivated by certain positive words. In this way, the activation of each filter of SBDAE is much more indicative of the polarity than that of DAE, which explains the better performance of SBDAE over DAE. Note that this difference only comes from reweighting the reconstruction errors in a certain way, with no explicit usage of labels.
-------------- --------------- --------------- --------------- -------------- ----------- ---------- --------------
nothing disappointing badly save even dull excuse ridiculously
cannon worst disappointing redeeming attempt fails had dean
outrageously unfortunately annoying awful unfunny stupid failed none
lends terrible worst sucks couldn’t worst rest ruined
teacher predictable poorly convince worst avoid he attempt
first tears loved amazing excellent perfect years with
classic wonderfully finest incredible surprisingly ? terrific best
man helps noir funniest beauty powerful peter recommended
hard awesome magnificent unforgettable unexpected excellent cool perfect
still terrific scared captures appreciated favorite allows heart
long wasn’t probably to making laugh tv someone
worst guy fan the give find might yet
kids music kind and performances where found goes
anyone work years this least before kids away
trying now place shows comes ever having poor
done least go kind recommend although ending worth
find book trying takes instead everyone once interesting
before day looks special wife anything wasn’t isn’t
work actors everyone now shows comes american rather
watching classic performances someone night away sense around
-------------- --------------- --------------- --------------- -------------- ----------- ---------- --------------
Related Work and Discussion {#related}
===========================
Our work falls into the general category of learning representations for text data. In particular, there have been a lot of efforts that try to learn compact representations for either words or documents [@vsm; @lda; @lsa; @word2vec; @pv; @mass]. LDA [@lda] explicitly learns a set of topics, each of which is defined as a distribution on words; a document is thus represented as the posterior distribution on topics, which is a fixed-length, non-negative vector. Closely related are matrix factorization models such as LSA [@lsa] and Non-negative Matrix Factorization (NMF) [@nmf]. While LSA factorizes the doc-term matrix via Singular Value Decomposition, NMF learns non-negative basis and coefficient vectors. Similar to these efforts, our model also works directly on the doc-term matrix. However, thanks to the usage of autoencoder, the representation for documents are calculated instantly via direct matrix product, which eliminates the need of expensive inference. Our work also distinguishes itself from other work as a semisupervised representation learning model, where label information can be effectively leveraged.
Recently, there has also been an active thread of research on learning word representations. Notably, [@word2vec] shows that we can learn interesting word embeddings via very simple architecture on a large amount of unlabeled dataset. Moreover, [@pv] proposed to jointly learn representations for sentences and paragraphs together with words in a similar unsupervised fashion. While our work does not explicitly model the representations for words, it is straightforward to incorporate this idea by adding an additional linear layer at the bottom of the autoencoder.
From the perspective of machine learning methodology, our approach resembles the idea of layer-wise pretraining in deep Neural Networks [@bengio]. Our model differs from the traditional training procedure of autoencoders in that we effectively utilize the label information to guide the representation learning. Related idea has been proposed in [@semi-ae], where they train Recursive autoencoders on sentences jointly with prediction of sentiment. Due to the delicate recursive architecture, their model only works on sentences with given parsing trees, and could not generalize to documents. MTC [@mtc] is another work that models the interaction of autoencoders and classifiers. However, their training of autoencoders is purely unsupervised, the interaction comes into play by requiring the classifier to be invariant along the tangents of the learned data manifold. It is not difficult to see that the assumption of MTC would not hold when the class labels did not align well with the data manifold, which is a situation our model does not suffer from.
Conclusion
==========
In this paper, we have proposed a novel extension to autoencoders for learning task-specific representations for textual data. We have generalized the traditional autoencoders by relaxing their loss function to the Bregman Divergence, and then derived a discriminative loss function from the label information. Experiments on text classification benchmarks have shown that our model significantly outperforms Bag of Words, traditional Denoising Autoencoder, and other competing methods. We have also qualitatively visualized that our model successfully learns discriminative features, which unsupervised methods fail to do.
Acknowledgments
===============
This work is supported in part by NSF (CCF-1017828).
[^1]: http://ai.stanford.edu/ amaas/data/sentiment/
[^2]: http://www.cs.jhu.edu/ mdredze/datasets/sentiment/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the present paper we introduce the property AA of a subsemigroup of the endomorphism semigroup of an abelian variety, which holds for semigroup of endomorphisms of an abelian variety defined over a number field, and show that the orbit of any cycle under a semigroup with property AA in the Chow group $\otimes{\mathbb Q}$ has finite dimensional span.'
author:
- Bogdan Zavyalov
title: '[**Chow Groups of Abelian Varieties and Beilinson’s Conjecture**]{}'
---
Introduction {#introduction .unnumbered}
============
One of the most celebrated conjectures in 20th century algebraic geometry is a conjecture of Beilinson, which in particular predicts that for any variety $X$ defined over a number field the group ${\operatorname{CH}}(X)\otimes{\mathbb Q}$ has finite dimension. This conjectural picture is completely opposite to the one in complex setting, e. g., Mumford proved that the Chow group of a smooth complex surface posessing a nonzero holomorphic 2-form is infinitely generated in any reasonable sense. The most naïve way to produce a counterexample to this conjecture is to build a variety $X$ over a number field together with a cycle on it such that its orbit w. r. t. [*all*]{} the endomorphisms of $X$ in the Chow group $\otimes{\mathbb Q}$ contains infinitely many linearly independent elements.
The goal of the present paper is to show that this attempt fails at least for $X$ an abelian variety. Although this fact is arithmetic in its nature, we show indeed that orbit of any cycle under any subsemigroup with the property AA (to be defined in the paper) of endomorphism semigroup (w. r. t. composition) has finite dimensional span in ${\operatorname{CH}}\otimes{\mathbb Q}$. This has some interesting applications even in the case of algebraically closed field. The proof of the fact that property AA holds for the entire endomorphism semigroup over number field is due to Mordell–Weil theorem.
The first part of the paper is a reminder of Fourier–Mukai transform developed by Beauville in his paper [@Beau]. The prove itself is given in the second part. The Fourier–Mukai transform allows us to reduce the case of an arbitrary cycle on $A$ to the case of the Poincaré line bundle on $A \times \widehat{A}$, and next we deal with the case of divisors explicitly.
We would like to thank Marat Rovinsky for suggesting the problem and numerous helpful discussions. Also, we would like to thank Rodion Déev for reviewing the first draft of the paper and Dmitrii Pirozhkov without whose assistance this paper might have not been written at all. The research was supported in part by Dobrushin stipend.
1. For any abelian variety $A$ and a $k$-rational point $a\in A(k)$ we will denote by $t_a: A \to A$ the translation of $A$ by $a$. We will denote the dual of the abelian variety $A$ by $\widehat{A}$ and the Poincaré line bundle by $L \in {\operatorname{Pic}}(A \times \widehat{A}).$ Moreover, we will denote the first Chern class of $L$ by $l$.
2. For any smooth projective variety $X$ over a field $k$ we will denote by ${\operatorname{CH}}(X):= \bigoplus_i {\operatorname{CH}}^i(X)$ the direct sum of all its Chow groups. The cup product enhances the abelian group ${\operatorname{CH}}(X)$ with a structure of a graded ring. For any proper map $f:X {\rightarrow}Y$ we will denote by $f_*$ the pushforward map on Chow rings (resp. groups) $f_*: {\operatorname{CH}}(X) {\rightarrow}{\operatorname{CH}}(Y)$ (resp. $f_*: {\operatorname{CH}}^i(X) {\rightarrow}{\operatorname{CH}}^i(Y)$). For any flat morphism $g: X {\rightarrow}Y$ we will denote by $g^*$ the pullback map on Chow rings (resp. groups) $g^*: {\operatorname{CH}}(Y) {\rightarrow}{\operatorname{CH}}(X)$ (resp. $g^*: {\operatorname{CH}}^i(Y) {\rightarrow}{\operatorname{CH}}^i(X)$).
3. All the products of varieties $X \times Y$ will actually mean $X \times_{{\operatorname{Spec}}k} Y$. By $p_{i_1, \dots, i_n}: \prod_{j:=1}^m X_j {\rightarrow}\prod_{k:=1}^{n} X_{i_k}$ we will always denote the corresponding projection map.
4. For any variety $A$ over a field $k$ we will denote by $\operatorname{End}(A)$ the semigroup of its endomorphisms over ${\operatorname{Spec}}(k)$ w. r. t. composition. For an abelian variety $A$ we will denote by $\operatorname{End}_0(A)$ its ring of group endomorphisms over a base field. Note that $\operatorname{End}(A)$ has also a group structure w. r. t. addition on $A$, but we will not deal with it (unless the opposite is mentioned explicitly).
5. Subscript $_{{{\mathbb Q}}}$ will always mean tensoring by ${{\mathbb Q}}$. For example, for any abelian group (resp. ring) $R$, $R_{{{\mathbb Q}}}$ will be a ${{\mathbb Q}}$-vector space (resp. ${{\mathbb Q}}$-algebra) $R\otimes_{{\mathbb Z}} {{\mathbb Q}}$.
Prelimanaries
=============
In this section we will provide the reader with all the well-known facts that will be important in the paper.
Picard Group of Abelian Varieties
---------------------------------
All the results in this section are well-known and could be found in any textbook on abelian varieties, possibly except for Proposition $\ref{split},$ which is not always formulated in such a manner.
Let $A$ be an abelian variety of dimension $g$ over a field $k$. We will need several facts about the structure of its Picard group.
Basic fact from algebraic geometry states that there is a short exact sequence $$0 {\rightarrow}{\operatorname{Pic}}^0(A) {\rightarrow}{\operatorname{Pic}}(A) {\rightarrow}{\operatorname{NS}}(A) {\rightarrow}0,$$ where ${\operatorname{Pic}}^0(A)$ is a group of algebraically trivial cycles, and ${\operatorname{NS}}(A)$ is a group of cycles modulo algebraic equivalence. Note that ${\operatorname{NS}}(A)$ is finitely generated ${\mathbb Z}$-module for any smooth projective variety. Also, the semigroup of endomorphisms $G:=\operatorname{End}(A)$ acts on each term of this short exact sequence making it into a short exact sequence of $G$-modules. The crucial fact is that after tensoring by ${{\mathbb Q}}$ it splits as a short exact sequence of $G_0:=\operatorname{End}_0(A)$-modules.
A divisor $D \in {\operatorname{Pic}}(A)$ is called [*symmetric*]{}, if $[-1]^*D=D.$, and [*antisymmetric*]{}, if $[-1]^*D=-D.$ Let us denote the group of symmetric (resp. antisymmetric) divisors by ${\operatorname{Pic}}^+(A)$ (resp. ${\operatorname{Pic}}^{-}(A)$).
The following lemmata give us a very good characterization of symmetric and antisymmetric divisors.
[\[sym\]]{} Let $A$ be an abelian variety over a field $k$, $D \in {\operatorname{Pic}}(A)$. Then the following are equivalent:
1. $D$ is symmetric,
2. $[n]^*D=n^2D$ for all $[n] \in {\mathbb Z}$.
[@Milne Cor. 5.4]
[\[antisym\]]{} Let $A$ be an abelian variety over a field $k$, $D \in {\operatorname{Pic}}(A)$. Then the following are equivalent:
1. $D$ is antisymmetric,
2. $[n]^*D=nD$ for all $[n] \in {\mathbb Z}$,
3. For every $f,g \in \operatorname{End}_0(A)$ we have $(f+g)^*(D)=f^*(D) + g^*(D)$,
4. $D \in {\operatorname{Pic}}^0(A)$.
Additionally, if $D$ is antisymmetric, then for each $a\in A(k)$ one has $\ t_a^*(D)=D$.
[@Milne Cor. 5.4. + Remark 8.5]
These lemmata allows us to construct a splitting of the short exact sequence $$\begin{gathered}
\label{very_good_sequence}\\ 0 {\rightarrow}{\operatorname{Pic}}^0(A)_{{{\mathbb Q}}} {\rightarrow}{\operatorname{Pic}}(A)_{{{\mathbb Q}}} {\xrightarrow}{r} NS(A)_{{{\mathbb Q}}} {\rightarrow}0 \\ \end{gathered}$$
[\[split\]]{} There is a canonical splitting of \[very\_good\_sequence\] as $G_0$-modules given by a map $ \phi: {\operatorname{Pic}}(A)_{{{\mathbb Q}}} {\rightarrow}{\operatorname{Pic}}^0(A)_{{{\mathbb Q}}}$ defined as $\phi(D)=(D-[-1]^*D)/2$. Moreover, ${\operatorname{Pic}}(A)_{{{\mathbb Q}}}\simeq {\operatorname{Pic}}^+(A)_{{{\mathbb Q}}} \oplus {\operatorname{Pic}}^-(A)_{{{\mathbb Q}}}$ as $G_0$-modules as well as ${\operatorname{Pic}}^+(A)_{{{\mathbb Q}}} \simeq {\operatorname{Pic}}^0(A)_{{{\mathbb Q}}}, \ {\operatorname{Pic}}^-(A)_{{{\mathbb Q}}} \simeq {\operatorname{Pic}}^0(A)_{{{\mathbb Q}}}$
First of all, note that the lemma \[antisym\] says that a natural inclusion of ${\operatorname{Pic}}^-(A)_{{{\mathbb Q}}}$ into ${\operatorname{Pic}}^0(A)_{{{\mathbb Q}}}$ gives us an isomorphism of them as $G_0$-modules. Secondly, we have two projections $p_{\pm}:{\operatorname{Pic}}(A)_{{{\mathbb Q}}} \to {\operatorname{Pic}}^{\pm}(A)_{{{\mathbb Q}}}$, namely $p_{\pm}(D)=\frac{D\pm [-1]^*D}{2}$. It is easy to check that these morphisms are morphisms of $G_0$-modules, so it provides us with a decomposition ${\operatorname{Pic}}(A)_{{{\mathbb Q}}}\simeq {\operatorname{Pic}}^+(A)_{{{\mathbb Q}}} \oplus {\operatorname{Pic}}^-(A)_{{{\mathbb Q}}}$. This implies that $\phi$ is a $G_0$-module section of the short exact sequence. Finally, $\ker r = {\operatorname{Pic}}^0(A)_{{{\mathbb Q}}} = {\operatorname{Pic}}^-(A)_{{{\mathbb Q}}}$ and ${\operatorname{Pic}}^+(A)_{{{\mathbb Q}}} \cap {\operatorname{Pic}}^-(A)_{{{\mathbb Q}}} = 0.$ Thus, by exactness of \[very\_good\_sequence\] and the fact that all morphisms are morphisms of $G_0$-modules, we conclude that ${\operatorname{Pic}}^+(A)_{{{\mathbb Q}}} \simeq {\operatorname{NS}}(A)_{{{\mathbb Q}}}.$
In particular this statement claims that $[n]^*({\alpha})=n^2{\alpha}$ for every ${\alpha}\in {\operatorname{NS}}(A)_{{{\mathbb Q}}}.$ Also, the proof actually shows that there is a morphism of $G_0$-modules $\psi: {\operatorname{NS}}(A)_{{{\mathbb Q}}} \to {\operatorname{Pic}}(A)_{{{\mathbb Q}}}$ defined by $\psi({\alpha})=(D+[-1]^*D)/2,$ where $D \in {\operatorname{Pic}}(A)_{{{\mathbb Q}}}$ is any representative of ${\alpha}\in {\operatorname{NS}}(A)_{{{\mathbb Q}}}.$
Fourier–Mukai Tranform
----------------------
[\[FM\]]{} Let $A$ be an abelian variety. The Fourier–Mukai transform is a map $F: {\operatorname{CH}}(A)_{{{\mathbb Q}}} \to {\operatorname{CH}}(\widehat{A})_{{{\mathbb Q}}}$ defined by $F(\alpha)= p_{2*}(p_1^*(\alpha) \cdot {\operatorname{exp}}(l))$ for every ${\alpha}\in {\operatorname{CH}}(A)_{{{\mathbb Q}}}$. By ${\operatorname{exp}}$ we undestand the map defined by formal power series ${\operatorname{exp}}(t)=1+t+\frac{t^2}{2}+ \dots$. Note that ${\operatorname{exp}}(l)$ is well-defined because $A$ is of finite dimension.
By $\widehat{F}$ we will denote the Fourier–Mukai tranform on the dual of abelian variety $\widehat{F} \colon {\operatorname{CH}}(\widehat{A})_{{{\mathbb Q}}} \to {\operatorname{CH}}(A)_{{{\mathbb Q}}}.$
The following theorem is essential in the proof of the Main Theorem. Namely, it will allow us to reduce the general case to the case of divisors.
[\[inv\]]{} Let $A$ be an abelian variety of dimension $g$. Then the following formula holds $$\widehat{F}\circ F = (-1)^g[-1]^*$$
[@Beau F2 p.647]
The Main Theorem
================
Before proving the Main Theorem we have to define some class of subsemigroups of the endomorphism semigroup of abelian variety for which it holds. The definition of such subsemigroups will be a little bit technical, but we will provide the reader with some interesting examples of such subsemigroups.
Recall that the semigroup of endomorphisms $G=\operatorname{End}(A)$ of any abelian variety is a semidirect product of the semigroup of its group endomorphisms $G_0=\operatorname{End}_0(A)$ and translations. If we denote the latter one by $T$, then $G=T \rtimes G_0$.
We say that a subsemigroup $H \subset G$ has [*property AA*]{}, if there is a finite number of $k$-rational points $a_1, \dots , a_n$ such that any element of $h \in H$ can be written as $h=f \circ t_{a_n}^{l_n} \circ \dots \circ t_{a_1}^{l_1},$ where $f \in G_0$ and $l_i \in \mathbb Z$.
1. The subsemigroup $G_0$ has property AA for trivial reasons.
2. Every finitely generated subgroup $T_0 \subset T$ has property AA.
3. Fix some finitely generated subgroup $T_0 \subset T$. Then a subsemigroup $T_0 \rtimes G_0$ has property AA.
4. (the most interesting example)\[vajno\] Let $A$ be an abelian variety over a number field $k$. Then the semigroup of all endomorphisms of $A$ satisfies the property AA. The Mordell–Weil theorem ([@Serre Chapter 4]) claims that for any abelian variety over number field $A(k)$ is a finitely generated group. It reduces this example to the third one.
[\[mtheorem\]]{}\[The Main Theorem\] Let $A$ be an abelian variety over a field $k$. Fix some subsemigroup $H \subset G$ that satisfies the property AA. Then for any cycle ${\alpha}\in {\operatorname{CH}}^p(A)_{{{\mathbb Q}}}$ its orbit under the action of the semigroup $H$ spans a finite-dimensional ${{\mathbb Q}}$-vector space.
We are going to prove the theorem in three steps. Firstly, we will prove it for $H=G_0$ and $p=1$. Secondly, we will reduce the case of any subsemigroup with property AA and $p=1$ to the case $H=G_0.$ Finally, we will reduce the general case to the case $p=1$.
**Step 1.** Assume that $p=1$ and $H=G_0.$ We will prove that for every finite-dimensional vector space $V \subset {\operatorname{Pic}}(A)_{{{\mathbb Q}}}$ its orbit under the action of $G_0$ spans finite-dimensional vector space.
According to Proposition \[split\], we have a split short exact sequence of $G_0$-modules $$0 \to {\operatorname{Pic}}^0(A)_{{{\mathbb Q}}} \to {\operatorname{Pic}}(A)_{{{\mathbb Q}}} \to {\operatorname{NS}}(A)_{{{\mathbb Q}}} \to 0.$$
Since this short exact sequence of $G_0$-modules splits and ${\operatorname{NS}}(A)_{{{\mathbb Q}}}$ is of finite dimension, we can assume that $V \subset {\operatorname{Pic}}^0(A)_{{{\mathbb Q}}}$. From now on we will consider $G_0$ as a group under addition instead of considering it as semigroup under composition. Lemma \[antisym\] tells us that the action of $G_0$ is linear (w.r.t. to this group structure) on ${\operatorname{Pic}}^0(A)_{{{\mathbb Q}}}$. Combined with the fact that $G_0$ is a finitely generated group ([@Milne Prop. 10.5 + Lemma 10.6]), we conclude that the orbit of the vector subspace $V$ under the action of $G_0$ spans a finite-dimensional vector space.
**Step 2.** Now assume that $p=1$, but semigroup $H$ is any subsemigroup of $G$ with the property AA. According to the definition, we can choose a finite number of $k$-rational points $a_1, \dots, a_n \in A(k)$ such that any element of $g \in H$ can be written in the following form $g=f \circ t_{a_n}^{l_n} \circ \dots \circ t_{a_1}^{l_1}$ with $f \in G_0$. We know [@Milne Lemma 8.8] that ${\beta}_i:=t_{a_i}^*({\alpha})-{\alpha}\in {\operatorname{Pic}}^0(A)_{{{\mathbb Q}}}$ for any $0<i \leq n.$ Since all $t_{a_i}$ commute and $t_a^*({\beta})-{\beta}=0$ for any ${\beta}\in {\operatorname{Pic}}^0(A)_{{{\mathbb Q}}}$(Lemma \[antisym\]), we conclude that $(t_{a_n}^{l_n})^* \circ \dots \circ (t_{a_1}^{l_1})^* ({\alpha})= {\alpha}-l_1{\beta}_1 - \dots -l_n{\beta}_n.$ In particular, the vector space $V$ generated by the orbit of $\alpha$ under the action of endomorphisms of the form $t_{a_n}^{l_n} \circ \dots \circ t_{a_1}^{l_1}$ is of finite dimension. Therefore, since $H$ has property AA, it suffices to show that the orbit of $V$ under the action of $G_0$ spans finite-dimensional vector space. But it has been already done in the Step 1.
**Step 3.** Finally, we are going to reduce the general case to the case of a divisor on the abelian variety $A\times \widehat{A}.$
Theorem \[inv\] says that ${\alpha}= (-1)^g[-1]^*\widehat{F}(F({\alpha})).$ Let $F({\alpha})=\sum_{q=1}^{g} \eta_q$ with $\eta_q \in {\operatorname{CH}}^q(\widehat{A})_{{{\mathbb Q}}}.$ Again, choose finite number of $k$-rational points $a_1, \dots, a_n \in A(k)$ such that any element of $g \in H$ can be written in the following form $g=f \circ t_{a_n}^{l_n} \circ \dots \circ t_{a_1}^{l_1}$ with $f \in G_0$. Choose such a representation for each $g \in H$ and denote by $g':=f \circ t_{-a_n}^{l_n} \circ \dots \circ t_{-a_1}^{l_1}.$ Also we will denote by $\overline g$ the morphism $g \times {\operatorname{Id}}_{\widehat{A}}: A\times \widehat{A} \to A\times \widehat{A}$. Now, for any $g \in H$ we have $$g^*({\alpha})=(-1)^g(g^* \circ [-1]^* \circ \widehat{F} \circ F )({\alpha})=(-1)^g(g^* \circ [-1]^* \circ \widehat{F})(\sum_{q=1}^g \eta_q)= (-1)^g\sum_{q=1}^{g}g^*([-1]^* \circ \widehat{F}(\eta_q))$$
This expression allows us to conclude that in order to prove the Theoreom for a cycle ${\alpha}$ it is sufficient to prove it for every $[-1]^* \circ \widehat{F}(\eta_q)$. From now on we are going to prove the Theorem in this case.
Note that for any $g \in H$ we have that $g^*\circ [-1]^*=[-1]^*g'$. Also the base change formula applied to the following Cartesian diagram asserts that $f^* \circ p_{1*}=p_{1*}\circ(\overline f)$ for every $f \in G$. $$\begin{CD}
A \times \widehat{A} @>\overline f>> A \times \widehat{A}\\
@Vp_1VV @Vp_1VV \\
A @>f>> A
\end{CD}$$
We conclude that
$$\begin{split}
g^*([-1]^* \circ \widehat{F}(\eta_q))=g^* \circ [-1]^*(p_{1*}(p_2^*(\eta_q) \cdot {\operatorname{exp}}(l)))=[-1]^*\circ g'^*(p_{1*}(p_2^*(\eta_q) \cdot {\operatorname{exp}}(l)))= \\
=[-1]^*\circ p_{1*}(\overline g'^*(p_2^*(\eta_q) \cdot {\operatorname{exp}}(l)))=[-1]^* \circ p_{1*}(p_2^*(\eta_q) \cdot {\operatorname{exp}}(\overline g'^*(l))) =\\
=[-1]^* \circ p_{1*}(p_2^*(\eta_q) \cdot (\sum_{i=1}^{\infty}\overline g'^*(l^n)))=\sum_{i=1}^{\infty} ([-1]^* \circ p_{1*}(p_2^*(\eta_q) \cdot \overline g'^*(l^n)))
\end{split}$$
The sum is actually finite because $A\times \widehat{A}$ is a variety of dimension $2g$, so $l^n=0$ for $n>2g$. Since cup-product, pullbacks and pushforwards are linear maps, it is enough to show that the orbit of each $l^n$ under the action of elements of the form $\overline g'^*$ spans a finite-dimensional vector space. Moreover, since $\overline g'^*(l^n)=\overline g'^*(l)^n$ it is enough to show it only for $l$. Every element of the form $\overline g'$, by definition, can be written in the form $$\overline g'=(f \times {\operatorname{Id}}) \circ (t^{k_n}_{-a_n} \times {\operatorname{Id}}) \circ \dots \circ (t^{k_1}_{-a_1} \times {\operatorname{Id}})=(f \times {\operatorname{Id}}) \circ t^{k_n}_{(-a_n,0)} \circ \dots \circ t^{k_1}_{(-a_1,0)}$$
Since $f\in \operatorname{End}_0(A)$, we have $f\times {\operatorname{Id}}\in \operatorname{End}_0(A\times \widehat{A}).$ Therefore, the subset $$H'=\{f \in \operatorname{End}(A\times \widehat{A})| \exists g \in H \colon f = \overline g'\}$$ is a subsemigroup and has property AA. Therefore, we can apply the Step 2 to $H'$ and $l$ and finish the proof.
Let $A$ be an abelian variety over any field $k$. Fix some cycle ${\beta}\in {\operatorname{CH}}^p(A)_{{{\mathbb Q}}}$ and a $k$-rational point $x \in A(k),$ then the set $\{{\alpha}, t_x^*({\alpha}), (t^*_x)^2({\alpha}), \dots\}$ spans finite dimensional vector space.
Apply Theorem \[mtheorem\] to $H={\mathbb Z}t_x^*$ and ${\alpha}={\beta}$.
Let $A$ be an abelian variety over any field $k$. Fix some cycle ${\beta}\in {\operatorname{CH}}^p(A)_{{{\mathbb Q}}}$, then the orbit of ${\beta}$ under the action of the semigroup $\operatorname{End}_0(A)$ spans finite dimensional vector space.
Apply Theorem \[mtheorem\] to $H=\operatorname{End}_0(A)$ and ${\alpha}={\beta}$.
Let $A$ be an abelian variety over number $k$. Fix some cycle ${\beta}\in {\operatorname{CH}}^p(A)_{{{\mathbb Q}}}$. Then the orbit of ${\beta}$ under the action of the group of all the endomorphism semigoup $\operatorname{End}(A)$ spans a finite-dimensional vector space.
According to Example \[vajno\], $\operatorname{End}(A)$ has property AA. Thus we can apply Theorem \[mtheorem\] to $H=\operatorname{End}(A)$ and ${\alpha}={\beta}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose an approach for calculating $K\rightarrow\pi\pi$ decays to the next-to-leading order in chiral expansion. A detailed numerical study of this approach is being performed.'
author:
- |
Ph. Boucaud$^{a}$, V. Giménez$^{b}$, C.-J. D. Lin$^{c}$[^1], V. Lubicz$^{d}$, G. Martinelli$^{e}$, M. Papinutto$^{f}$,\
F. Rapuano$^{e}$ and C.T. Sachrajda$^{c}$ (The SPQ$_{\tiny{\mathrm{CD}}}$R Collaboration)\
$^{a}$ Université de Paris Sud, L.P.T. (Bât. 210), Centre d’Orsay, 91405 Orsay-Cedex, France\
$^{b}$ Dep. de Física Teòrica, IFIC, Dr. Moliner 50, E-46100, Burjassot, València, Spain\
$^{c}$ Dept. of Physics and Astronomy, Univ. of Southampton, Southampton SO17 1BJ, England\
$^{d}$ Dip. di Fisica, Universitá di Roma Tre, Via della Vasca Navale 84, I-00146 Roma, Italy\
$^{e}$ Dip. di Fisica, Universitá di Roma “La Sapienza”, Piazzale A. Moro 2, I-00185 Roma, Italy\
$^{f}$ Dip. di Fisica, Universitá di Pisa and INFN - Pisa, Via Buonarroti 2, I-56100 Pisa, Italy
bibliography:
- 'refs.bib'
title: '$K\rightarrow\pi\pi$ matrix elements beyond the leading-order chiral expansion [^2]'
---
ł
\#1[[O]{}ł(\#1)]{}
\#1[Eq. (\[\#1\])]{} \#1\#2[Eqs. (\[\#1\]) and (\[\#2\])]{} \#1[Table \[\#1\]]{} \#1\#2[Tables \[\#1\] and \[\#2\]]{} \#1[Section \[\#1\]]{} \#1\#2[Sections \[\#1\] and \[\#2\]]{} \#1[Figure \[\#1\]]{} \#1\#2[Figures \[\#1\] and \[\#2\]]{} \#1\#2\#3[Figure \[\#1\], \[\#2\] and \[\#3\]]{} \#1\#2\#3[Figure \[\#1\], \[\#2\] and \[\#3\]]{} \#1[Ref. ]{}
3[${\mathrm{SU}}(3)_{\mathrm{L}}\times {\mathrm{SU}}(3)_{\mathrm{R}}$]{} 3[${\mathrm{SU}}(3)_{\mathrm{L}}$]{} 3[${\mathrm{SU}}(3)_{\mathrm{R}}$]{}
Introduction
============
Controlling the effects of final state interactions (FSI) is one of the main barriers towards a high-precision theoretical prediction for $K\rightarrow\pi\pi$ decays. This is particularly difficult for lattice QCD because of the analytic continuation to Euclidean space. The finite-volume techniques developed in Refs. [@Lellouch:2000pv; @Lin:2001ek] can exactly take into account the FSI effects, but their numerical implementation is very demanding. In the foreseeable future, the most practical approach, which is also reliable and systematically improvable, to the study of non-leptonic kaon decays remains the combination of lattice QCD and chiral perturbation theory ($\chi$PT).
Apart from a study of the CP-conserving, $\Delta I = 3/2$, $K\rightarrow\pi\pi$ decay in Ref. [@Aoki:1998ev], all the existing lattice results for matrix elements of the form $\langle\pi\pi|Q|K\rangle$ [@Martinelli:2001pm] are obtained by simulating matrix elements of the kind $\langle\pi|Q|K\rangle$, then using lowest-order $\chi$PT [@Bernard:1985wf] to relate them to the desired matrix elements. In this procedure, the effects of higher-order chiral corrections due to FSI, which could be very large, are completely missing[^3].
Here we present the status of an on-going project in which all the relevant matrix elements for $\epsilon^{\prime}/\epsilon$ and the $\Delta I = 1/2$ rule are being calculated by computing matrix elements of the kind $\langle\pi\pi|Q|K\rangle$, at some “unphysical” kinematics, on the lattice, and then using the chiral expansion to next-to-leading order (NLO) to obtain them at the “physical” kinematics. In this talk, we focus[^4] on the study of the chiral behaviour of $\Delta I = 3/2$, $K\rightarrow\pi\pi$ decay amplitudes associated with the operators $$\begin{aligned}
Q_{4} &=&
(\bar{s}_{\alpha}d_{\alpha})_{L}(\bar{u}_{\beta}u_{\beta}
- \bar{d}_{\beta}d_{\beta})_{L} +
(\bar{s}_{\alpha}u_{\alpha})_{L}(\bar{u}_{\beta}d_{\beta})_{L} ,\nonumber\\
\label{eq:OpDef}
Q_{7} &=& \frac{3}{2}(\bar{s}_{\alpha}d_{\alpha})_{L}
\sum_{q=u,d,s,c}e_{q}(\bar{q}_{\beta}q_{\beta})_{R} , \\
Q_{8} &=& \frac{3}{2}(\bar{s}_{\alpha}d_{\beta})_{L}
\sum_{q=u,d,s,c}e_{q}(\bar{q}_{\beta}q_{\alpha})_{R} ,\nonumber\end{aligned}$$ where $\alpha$, $\beta$ are colour indices and $e_{q}$ is the electric charge of $q$. $(\bar{\psi}_{1}\psi_{2})_{L,R}$ means $\bar{\psi}_{1}\gamma_{\mu}(1\mp\gamma_{5})\psi_{2}$.
Choice of “unphysical” kinematics {#sec:kin}
=================================
The matrix elements $\langle\pi^{+}\pi^{0}|Q_{i}|K^{+}\rangle$ ($i = 4, 7$ and 8) are computed in the unphysical kinematics such that $K^{+}$ and one of the pions are always at rest, while the other pion might carry non-zero spatial momentum. We denote these matrix elements by $$\begin{aligned}
\langle\pi^{+}(\vec{p}_{\pi})\pi^{0}(\vec{0})|Q_{i}|
K^{+}\rangle_{\mathrm{unphys}}
&,& \mbox{ }\mbox{ }\mathrm{if}\mbox{ }\pi^{0}
\mbox{ }\mathrm{is}\mbox{ }\mathrm{at}\mbox{ }\mathrm{rest} ,
\nonumber\\
\langle\pi^{+}(\vec{0})\pi^{0}(\vec{p}_{\pi})|Q_{i}|
K^{+}\rangle_{\mathrm{unphys}}
&,& \mbox{ }\mbox{ }\mathrm{if}\mbox{ }\pi^{+}
\mbox{ }\mathrm{is}\mbox{ }\mathrm{at}\mbox{ }\mathrm{rest} .\nonumber\end{aligned}$$ Notice that the spatial momentum $\vec{p}_{\pi}$ in the above matrix elements might be zero as well. In this case, both pions are at rest. The correlators used to extract these matrix elements are discussed in Ref. [@Boucaud:2001tx]. A technical difficulty arising at this stage is that the final states $|\pi^{+}(\vec{p}_{\pi})\pi^{0}(\vec{0})\rangle$ and $|\pi^{+}(\vec{0})\pi^{0}(\vec{p}_{\pi})\rangle$ are not purely $I=2$, because Bose-Einstein statistics does not rule out the $I=1$ components. In order to eliminate these components, we take the symmetric combination $$\begin{aligned}
& &\langle\pi^{+}\pi^{0}|Q_{i}|K^{+}\rangle_{\mathrm{unphys}}^{\mathrm{sym}}
\nonumber\\
&=& \frac{1}{2}\bigg (
\langle\pi^{+}(\vec{p}_{\pi})\pi^{0}(\vec{0})|Q_{i}|
K^{+}\rangle_{\mathrm{unphys}}\nonumber\\
& & \mbox{ }+
\langle\pi^{+}(\vec{0})\pi^{0}(\vec{p}_{\pi})|Q_{i}|
K^{+}\rangle_{\mathrm{unphys}}
\bigg ) .\end{aligned}$$
Chiral expansion at NLO {#sec:ChPT}
=======================
In order to relate $\langle\pi^{+}\pi^{0}|Q_{i}|K^{+}\rangle_{\mathrm{unphys}}^{\mathrm{sym}}$ to their counter parts at physical kinematics, denoted as $\langle\pi^{+}\pi^{0}|Q_{i}|K^{+}\rangle_{\mathrm{phys}}$, we resort to NLO chiral expansion. Under the chiral $SU(3)_{\mathrm{L}}\otimes SU(3)_{\mathrm{R}}$ transformation, $Q_{4}$ is in the (27,1) representation and $Q_{7,8}$ are in the (8,8) representation. Therefore the leading term in the chiral expansion for $Q_{4}$ ($Q_{7,8}$) is of ${\mathcal{O}}(p^{2})$ (${\mathcal{O}}(p^{0})$). Chirally expanding $Q_{4}$ ($Q_{7,8}$) to ${\mathcal{O}}(p^{4})$ (${\mathcal{O}}(p^{2})$) requires the knowledge of counterterm operators in the relevant chiral representation at this order, as well as one-loop calculations in $\chi$PT using the leading-order operators.
There is only one $\chi$PT representative for operators in (27,1) $\big ( (8,8) \big )$ representation at ${\mathcal{O}}(p^{2})$ (${\mathcal{O}}(p^{0})$). At ${\mathcal{O}}(p^{4})$, there are thirty four $\chi$PT representatives for (27,1) operators, labelled as ${\mathcal{O}}^{(27)}_{i}$ ($i=1,2,\ldots, 34$) in Ref. [@Kambor:1990tz]. For the purpose of this work, we only need ${\mathcal{O}}^{(27)}_{2}$, ${\mathcal{O}}^{(27)}_{4}$, ${\mathcal{O}}^{(27)}_{5}$, ${\mathcal{O}}^{(27)}_{7}$, ${\mathcal{O}}^{(27)}_{22}$ and ${\mathcal{O}}^{(27)}_{24}$ [^5]. The seven ${\mathcal{O}}(p^{2})$ chiral representatives for (8,8) operators can be found in Eq. (9) of Ref. [@Cirigliano:1999pv]. Here we label them as ${\mathcal{O}}^{(8,8)}_{i}$ ($i=1,2,\ldots, 7$).
The one-loop $\chi$PT calculations at physical kinematics, in infinite volume, are reported in Ref. [@Golterman:1997wb] for (27,1), and Ref. [@Cirigliano:1999pv] for (8,8). In our work, the chiral logarithms at the chosen unphysical kinematics are much more complicated than in these references, because of the energy-momentum injection at the operator. The resulting amplitudes depend upon three kinematical variables, $\bar{M}_{K}$ , $\bar{M}_{\pi}$ (kaon and pion masses in the simulation) and $\bar{E}_{\pi}$ (energy of the pion carrying a non-zero spatial momentum in the simulation), while they only depend upon $M_{K}$ and $M_{\pi}$ (physical kaon and pion masses) at the physical kinematics. Furthermore, our numerical simulations are performed in a finite volume in the quenched approximation. This requires one-loop calculations in finite-volume quenched $\chi$PT (q$\chi$PT). So far, we have only finished the one-loop (unquenched) $\chi$PT calculations in infinite volume as an exercise. These calculations have been performed in a general way, such that the four-momenta of $K^{+}$, $\pi^{+}$ and $\pi^{0}$ are used as the kinematical variables, and energy-momentum conservation is only implemented at the ${\mathcal{O}}(p^{2})$ strong vertices. The desiredkinematics is imposed at the end of the calculations.
The results for the NLO chiral expansion for the amplitudes of interest are $$\begin{aligned}
& &\ampl_{\mathrm{phys}}^{(27)} =
\frac{-6\sqrt{2}}{f_{K}f^{2}_{\pi}} \times
\bigg \{ \mbox{ }
\alpha M^{2}_{K}-
\alpha M^{2}_{\pi}
\nonumber\\
&+& (\beta_{4}-\beta_{5}+4\beta_{7}+ 2 \beta_{22})
M^{4}_{K}\nonumber\\
&+& (-4\beta_{2}+3\beta_{4}+\beta_{5}-2\beta_{7}\nonumber\\
& &\mbox{ }-2\beta_{22}
+16\beta_{24}) M^{2}_{K}M^{2}_{\pi}\nonumber\\
&+& (4\beta_{2}-4\beta_{4}-2\beta_{7}-16\beta_{24})
M^{4}_{\pi} \mbox{ } \bigg \} \nonumber\\
&+& \frac{\alpha}{f_{K}f^{2}_{\pi}}\times
(\mathrm{one}\mbox{ }
\mathrm{loop})^{(27)}_{\mathrm{phys}} , \nonumber\\
& &\ampl^{(8,8)}_{\mathrm{phys}} = \frac{2\sqrt{2}}{f_{K}f^{2}_{\pi}}
\times {\bigg \{ }
{\gamma}\nonumber\\
&+& [-({\delta_{2}}+\delta_{3})+
2({\delta_{4}}+\delta_{5})+
4{\delta_{6}}]
{M^{2}_{K}}\nonumber\\
&+& [({\delta_{1}}+\delta_{2})+
4({\delta_{4}}+\delta_{5})+
2{\delta_{6}}]
{M^{2}_{\pi}}
{\bigg \} }\nonumber\\
\label{eq:PhysAmp}
&+& \frac{{\gamma}}{f_{K}f^{2}_{\pi}}
({\mathrm{one}\mbox{ }
\mathrm{loop}})^{(8,8)}_{\mathrm{phys}},\end{aligned}$$ for the physical kinematics, and & & (\^[(27)]{})\^\_ =\
& &(|[E]{}\_|[M]{}\_+ + )\
&+& 4 \_[2]{} |[M]{}\^[4]{}\_ + (4\_[4]{}+2\_[7]{})|[E]{}\_|[M]{}\^[3]{}\_\
&+& (\_[4]{}-\_[5]{}+\_[7]{})|[M]{}\^[3]{}\_|[M]{}\_[K]{}\
&+& (\_[4]{}-\_[5]{}+\_[7]{}+2\_[22]{}) |[E]{}\_|[M]{}\^[2]{}\_|[M]{}\_[K]{}\
&+& (-4\_[2]{}+8\_[24]{})|[M]{}\^[2]{}\_|[M]{}\^[2]{}\_[K]{}\
&+& (-2\_[5]{}+4\_[7]{}+4\_[22]{}) |[E]{}\_|[M]{}\_|[M]{}\^[2]{}\_[K]{}\
&+& (\_[4]{}+2\_[7]{}) (|[M]{}\_+|[E]{}\_)|[M]{}\^[3]{}\_[K]{}\
&+& (-16\_[24]{})|[E]{}\^[2]{}\_|[M]{}\^[2]{}\_ + 2\_[22]{}|[E]{}\^[2]{}\_ |[M]{}\_ |[M]{}\_[K]{}\
&+& 8\_[24]{}|[E]{}\^[2]{}\_|[M]{}\^[2]{}\_[K]{}\
&+& \[([ ]{})\^[(27)]{}\]\^\_,\
& & (\^[(8,8)]{})\^\_ =\
&+& \[4([\_[4]{}]{}+\_[5]{}) + 2[\_[6]{}]{}\][|[M]{}\^[2]{}\_]{}+ \[-([\_[1]{}]{}+\_[2]{})\] [|[E]{}\_|[M]{}\_]{}\
&+&\[([\_[1]{}]{}+\_[2]{}) - ([\_[2]{}]{}+\_[3]{})\] [(|[M]{}\_+|[E]{}\_)|[M]{}\_[K]{}]{}\
&+& \[2([\_[4]{}]{}+\_[5]{})+ 4[\_[6]{}]{}\] [|[M]{}\^[2]{}\_[K]{}]{} [} ]{}\
\[eq:UnphysAmp\] &+& \[([ ]{})\^[(8,8)]{}\]\^\_ , for the unphysical kinematics described in Sec. \[sec:kin\]. Here $\alpha$ ($\gamma$) is the coupling constant accompanying the ${\mathcal{O}}(p^{2})$ (${\mathcal{O}}(p^{0})$) operator, and $\beta_{i}$ ($\delta_{i}$) are the coupling constants associated with ${\mathcal{O}}(p^{4})$ $\big({\mathcal{O}}(p^{2})\big )$ counterterm operators for (27,1) $\big ( (8,8) \big )$ chiral representation. We are able to reproduce the $(\mathrm{one}\mbox{ }\mathrm{loop})^{(27)}_{\mathrm{phys}}$ and $(\mathrm{one}\mbox{ }\mathrm{loop})^{(8,8)}_{\mathrm{phys}}$ results in Refs. [@Golterman:1997wb] and [@Cirigliano:1999pv]. We have also checked that all the amplitudes in Eqs. (\[eq:PhysAmp\]) and (\[eq:UnphysAmp\]) are independent of the renormalisation scale $\mu_{\chi}$ in $\chi$PT with the same $\mu_{\chi}$ dependence in each NLO coupling.
By having enough data points for the amplitudes in Eq. (\[eq:UnphysAmp\]) at different values of $\bar{E}_{\pi}$, $\bar{M}_{\pi}$ and $\bar{M}_{K}$, one can determine the coupling constants up to NLO chiral expansion and construct the physical amplitudes in Eq. (\[eq:PhysAmp\]). We are currently performing a detailed numerical study (clover action, $\beta=6.0$) of this strategy, as well as the one-loop calculations in finite-volume q$\chi$PT. Figure 1 shows an example of our numerical work.
\[fig:O8\_plot\]
[^1]: Presenter at the conference.
[^2]: SHEP 01/20. Support from PPARC under grant PPA/G/S/1998/00529 is acknowledged. We thank Sebastien Descotes and Jonathan Flynn for discussions.
[^3]: In Ref. [@Aoki:1998ev], the decay amplitude is also obtained at the precision of leading-order chiral expansion.
[^4]: For details of other aspects of this work, please refer to Ref. [@Boucaud:2001tx].
[^5]: Notice that ${\mathcal{O}}^{(27)}_{22}$ and ${\mathcal{O}}^{(27)}_{24}$ are not included in Ref. [@Golterman:2000fw]. ${\mathcal{O}}^{(27)}_{24}$ also contributes to $\langle\pi|Q|K\rangle$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The radiative decay of $J/\psi$ into a pure gauge tensor glueball is studied in the quenched lattice QCD formalism. With two anisotropic lattices, the multipole amplitudes $E_1(0)$, $M_2(0)$ and $E_3(0)$ are obtained to be $0.114(12)(6)$ GeV, $-0.011(5)(1)$ GeV, and $0.023(8)(1)$ GeV, respectively. The first error comes from the statistics, the $Q^2$ interpolation, and the continuum extrapolation, while the second is due to the uncertainty of the scale parameter $r_0^{-1}=410(20)$ MeV. Thus, the partial decay width $\Gamma(J/\psi\rightarrow \gamma G_{2^{++}})$ is estimated to be $1.01(22)(10)$ keV, which corresponds to a large branch ratio $1.1(2)(1)\times 10^{-2}$. The phenomenological implication of this result is also discussed.'
author:
- |
Yi-Bo Yang,$^{1}$ Long-Cheng Gui,$^{1}$ Ying Chen,$^{1,}$[^1] Chuan Liu,$^{2}$ Yu-Bin Liu,$^{3}$ Jian-Ping Ma$^{4}$, and Jian-Bo Zhang$^{5}$\
(CLQCD Collaboration)
title: 'Lattice Study of Radiative $J/\psi$ Decay to a Tensor Glueball'
---
Glueballs are exotic hadron states made up of gluons. Their existence is permitted by QCD but has not yet been finally confirmed by experiment. In contrast to the scalar glueball, whose possible candidate can be $f_0(1370)$, $f_0(1500)$, or $f_0(1710)$, the experimental evidence for the tensor glueball is more obscure. Quenched lattice QCD studies predict the tensor glueball mass to be in the range 2.2-2.4 GeV [@prd56; @prd60; @prd73], which is also supported by a recent $2+1$ flavor full-QCD lattice simulation [@Gregory:2012]. In this mass region, Mark III [@Baltrusaitis:1986] and BES [@Bai:1996] have observed a narrow tensor meson $\xi(2230)$ \[now as $f_J(2220)$ in PDG[@PDG2012]\] in the $J/\psi$ radiative decays with a large production rate, whose features favor the interpretation of a tensor glueball. However, it was not seen in the inclusive $\gamma$ spectrum [@Kopke:1989] by the Crystal Ball Collaboration and in $p\bar{p}$ annihilations to pseudoscalar pairs [@Barnes:1993; @Hasan:1992; @Hasan:1996; @Bardin:1987; @Sculli:1987; @Evangelista:1997; @Evangelista:1998; @Buzzo:1997]. So, the existence of $\xi(2230)$ \[$f_J(2220)$\] needs confirmation by new experiments, especially by the BESIII experiment with the largest $J/\psi$ sample.
It is well known that the production of glueballs is favored in $J/\psi$ decays because of the gluon-rich environment. The radiative decay is of special importance, owing to its cleaner background. So, the production rate of the tensor glueball in the decay can be an important criterion for its identification. The decay has been studied only in a few theoretical works [@Li1981; @Li:1987; @Tenzo:1988; @Melis:2004]. In these works, the tree-level perturbative QCD approach is employed. Under certain assumptions, the helicity amplitudes of the decay are related to the coupling of the two gluons to the tensor glueball. This coupling has been determined with the quenched lattice QCD [@prd73; @Meyer2009]. Based on results of Refs. [@Li1981; @Li:1987; @Tenzo:1988] the branch ratio is estimated as $2\times
10^{-3}$ [@Li:2009], but the theoretical uncertainties are not under control.
In fact, the decay can be investigated directly from the numerical lattice QCD studies [@dudek06; @Gui:2013], which provide first principles calculations from the QCD Lagrangian, especially in quenched lattice QCD. Quenched lattice QCD can be taken as a theory which only consists of heavy quarks and gluons. In this theory amplitudes of the decay do not have an absorptive part because of masses of states. Hence, the amplitudes can be directly calculated in the theory in Euclidian space. It should be noted that it is still a challenging task for the full-QCD lattice study of the decay because glueballs can be mixed with states of light quark pairs. Nevertheless, the study of the decay in quenched QCD will give important information about nonperturbative properties of glueballs.
At the lowest order of QED, the amplitude for the radiative decay $J/\psi\rightarrow\gamma
G_{2^{++}}$ is given by $$M_{r,r_\gamma,r_G}=\epsilon_{\mu}^*(\vec{q},r_\gamma)\langle
G(\vec{p}_f,r_G)|j^{\mu}(0)|J/\psi(\vec{p}_i,r)\rangle,$$ where $\vec{q}=\vec{p}_i-\vec{p}_f$ is the momentum of the real photon, and $r$, $r_\gamma$, and $r_G$ are the quantum numbers of the polarizations of $J/\psi$, the photon, and the tensor glueball, respectively. $\epsilon_\mu(\vec{q},r_\gamma)$ is the polarization vector of the photon, and $j^\mu$ is the electromagnetic current operator. The hadronic matrix element appearing in the above equation can be obtained directly from a lattice QCD calculation of corresponding three-point functions. On the other hand, these matrix elements can be expressed (in Minkowski space-time) in terms of multipole form factors as follows: $$\begin{aligned}
\label{multipole}
&&\langle G(\vec{p}_f, r_G) | j^{\mu}(0) | J/\psi(\vec{p}_i,r)\rangle = \alpha_1^\mu
E_1(Q^2) + \alpha_2^{\mu}M_2(Q^2)\nonumber\\
&& + \alpha_3^\mu E_3(Q^2) + \alpha_4^\mu C_1(Q^2)+ \alpha_5^\mu C_2(Q^2)\end{aligned}$$ where $\alpha_i^{\mu}$ are Lorentz-covariant kinematic functions of $p_i$ and $p_f$ (and specific polarizations of the states), whose explicit expressions can be derived exactly [@Dudek2009; @Yang:2012], and $E_1(Q^2)$, $M_2(Q^2)$, $E_3(Q^2)$, $C_1(Q^2)$, and $C_2(Q^2)$ are the form factors which depend only on $Q^2=-(p_i-p_f)^2$. Since $C_1(Q^2)$ and $C_2(Q^2)$ vanish at $Q^2=0$, we focus on the extraction of the first three which are involved in the calculation of the decay width $\Gamma(J/\psi\rightarrow \gamma G_{2^{++}})$ as $$\label{tensor_width}
\Gamma=\frac{4\alpha|\vec{p}_\gamma|}{27M_{J/\psi}^2}[|E_1(0)|^2+|M_2(0)|^2+|E_3(0)|^2],$$ where $\alpha$ is the fine structure constant, and $\vec{p}_\gamma$ is the photon momentum with $|\vec{p}_\gamma|=(M_{J/\psi}^2-M_G^2)/(2M_{J/\psi})$.
We use the tadpole-improved gauge action [@prd56] to generate gauge configurations on anisotropic lattices with the aspect ratio $\xi=a_s/a_t=5$, where $a_s$ and $a_t$ are the spatial and temporal lattice spacings, respectively. Two lattices $L^3\times T=8^3\times 96(\beta=2.4)$ and $12^3\times 144(\beta=2.8)$ are applied to check the effect of the finite lattice spacings. The relevant input parameters are listed in Table \[tab:lattice\], where $a_s$ values are determined from $r_0^{-1}=410(20)$ MeV. Since glueball relevant study needs quite a large statistics, the spatial extensions of both lattices are properly chosen to be $\sim 1.7$ fm according to the study of the finite volume effect study of Ref. [@prd73], which is a compromise of the computational resource requirement and negligible finite volume effects both for glueballs [@prd73] and charmonia. In the practice, we generated 5000 configurations for each lattice. The charm quark propagators are calculated using the tadpole-improved clover action for anisotropic lattices [@chuan1; @chuan2] with the bare charm quark masses set by the physical mass of $J/\psi$, $M_{J/\psi}=3.097$ GeV, through which the spectrum of the $1S$ and $1P$ charmonia are well reproduced [@Yang:2012]. In practice, disconnected diagrams due to the charm and quark-antiquark annihilation are expected to be unimportant according to the Okubo-Zweig-Iizuka rule and therefore are neglected in the calculation of relevant two-point and three-point functions.
$\beta$ $\xi$ $a_s/r_0$ $a_s$(fm) $La_s$(fm) $L^3\times T$ $N_{\rm conf}$
--------- ------- ----------- -------------- -------------- ------------------ ----------------
2.4 5 0.461(4) 0.222(2)(11) $\sim 1.78$ $8^3\times 96$ 5000
2.8 5 0.288(2) 0.138(1)(7) $\sim 1.66 $ $12^3\times 144$ 5000
: \[tab:lattice\] The input parameters for the calculation. Values for the coupling $\beta$, anisotropy $\xi$, the lattice size, and the number of measurements are listed. $a_s/r_0$ is determined by the static potential, and $a_s$ is estimated by $r_0^{-1}=410(20)$ MeV.
The calculations in this Letter are performed in the rest frame of the tensor glueball. One of the key issues in our calculation is to construct optimal interpolating field operators which couple dominantly to the pure gauge tensor glueball. This is realized by applying completely the same scheme as that in the calculations of the glueball spectrum [@prd60; @prd73]. On the cubic lattice, a tensor ($J=2$) state corresponds to the $E$ and $T_2$ irreducible representations of the lattice symmetry group $O$. So, we build the $E$ and $T_2$ operators from a set of prototype Wilson loops. By using different gauge-link smearing techniques, an operator set $\{\phi_\alpha^{(i)},
\alpha = 1,2,\ldots, 24\}$ of 24 different gluonic operators is constructed for each component of the $T_2^{++}$ and $E^{++}$ representations, where the superscript $i$ labels the three components of $T_2$ and two components of $E$. Finally, for each component, an optimal operator $\Phi^{(i)}(t)=\sum v_{\alpha}\phi_\alpha^{(i)}(t)$ for the ground state tensor glueball is obtained with the combinational coefficients $v_{\alpha}$ determined by solving the generalized eigenvalue problem $$\label{eigen} \tilde{C}^{(i)}(t_D){\bf v}^{(R)} = e^{-t_D\tilde{m}(t_D)}\tilde{C}^{(i)}(0){\bf
v}^{(R)},$$ at $t_D=1$, where $\tilde{C}^{(i)}(t)$ is the correlation matrix of the operator set $$\tilde{C}_{\alpha\beta}(t) =\frac{1}{N_t} \sum\limits_{\tau}\langle
0|{\phi}^{(i)}_\alpha(t+\tau){\phi}^{(i)}_\beta(\tau)|0\rangle.$$ In addition, the glueball two-point functions are normalized as $$\begin{aligned}
\label{glb_two}
C^{i}(t)&=&\frac{1}{T}\sum\limits_{\tau}\langle
\Phi^{(i)}(t+\tau)\Phi^{(i)\dagger}(\tau)\rangle \nonumber\\
&\approx& \frac{|\langle 0|\Phi^{(i)}(0)|T_i\rangle|^2}{2M_TV_3}e^{-M_Tt}\approx e^{-M_Tt},\end{aligned}$$ where $|T_i\rangle$ refers to $i$th component of the $T_2^{++}$ and $E^{++}$ glueball states. We are assured that $C^i(t)$ can be well described by a single exponential $C(t)=We^{-M_Tt}$, with $W$ usually deviating from one by a few percents. It should be noted that the $SO(3)$ rotational symmetry is broken on the lattice with a finite lattice spacing, and consequently the masses of $T_2$ and $E$ glueballs are not necessarily the same, even though they converge to the same tensor glueball mass in the continuum limit when the rotational invariance is restored. However, with the two lattice spacings we used in this Letter, we observe that the difference of the two masses is not distinguishable within errors, which implies that the effects of the rotational symmetry breaking are not important. So, in the following, we neglect this symmetry breaking and assume that the five components of the $T_2$ and $E$ and that of the corresponding spin-two state can be connected by a normal transformation.
We calculate the three-point functions in the rest frame of the tensor glueball with $J/\psi$ moving with a definite momentum $\vec{p}_f=2\pi \vec{n}/La_s$, where $\vec{n}$ ranges from $(0,0,0)$ to $(2,2,2)$. In order to increase the statistics additionally, for each configuration, we calculate $T$ charm quark propagators $S_F(\vec{x},t;\vec{0},\tau)$ by setting a point source on each time slice $\tau$, which permits us to average over the temporal direction when calculating the three-point functions $$\begin{aligned}
\Gamma_{i,\mu,j}^{(3)}(\vec{q};t_f,t) &=&
\frac{1}{T}\sum\limits_{\tau=0}^{T-1}\sum\limits_{\vec{y}} e^{-i\vec{q}\cdot \vec{y}} \langle
\Phi^{(i)}(t_f+\tau)\nonumber\\
&&\times J_\mu (\vec{y},t+\tau)O_{V,j}(\vec{0},\tau)\rangle\nonumber\\
&=&\sum\limits_{T,V,r}\frac{e^{-M_T (t_f-t)}e^{-E_V(\vec{q}) t}}{2M_T V_3 2E_V(\vec{q})}\nonumber\\
&&\times\langle 0|\Phi^{(i)}(0)|T_i\rangle \langle
T_i|J_\mu(0)|V(\vec{q},r)\rangle\nonumber\\
&&\times \langle V(\vec{q},r)|O_{V,j}^{\dagger}(0)|0\rangle,\end{aligned}$$ where $J_\mu(x) =\bar{c}(x)\gamma_\mu c(x)$ is the vector current operator, $O_{V,j}=\bar{c}\gamma_j c$ is the conventional interpolation field for $J/\psi$, and the summation in the last equality is over all the possible states with different polarizations. In the rest frame of the tensor glueball, the momentum of the initial $J/\psi$ is the same as that of the current operator, say, $\vec{p}_V=\vec{q}$. The vector current $J_\mu (x)$, which is conserved in the continuum limit, is no longer conserved on the lattice and requires a multiplicative renormalization. The renormalization constant of spatial components of the vector current is determined to be $Z_V^{(s)}=1.39(2)$ for $\beta=2.4$ and $Z_V^{(s)}=1.11(1)$ for $\beta=2.8$ [@Gui:2013] using a nonperturbative scheme [@dudek06].
The matrix elements $\langle T_i|J_\mu(0)|V(\vec{q},r)\rangle$ can be extracted from the above three-point functions along with the two-point function of the glueball $C^i(t)$ and that of $J/\psi$ $\Gamma_j^{(2)}(\vec{q},t)$, $$\Gamma_j^{(2)}(\vec{q},t)=\sum\limits_{\vec{x}} e^{-i\vec{q}\cdot\vec{x}}\langle
0|O_{V,j}(\vec{x},t)O_{V,j}^\dagger(\vec{0},0)|0\rangle,$$ which provide the information of $M_T$, $E_V(\vec{q})$, and the other two matrix elements. According to Eq. (\[glb\_two\]), one has approximately $$\langle 0|\Phi^{(i)}(0)|T_i\rangle\approx \sqrt{2M_TV_3}.$$
![The effective energy plot for $J/\psi$ with different spatial momenta. From top to bottom are the plateaus for momentum modes, $\vec{p}=2\pi\vec{n}/L$, with $\vec{n}=(2,2,2)$, $(2,2,1)$, $(2,2,0)$, $(2,1,1)$, $(2,1,0)$, $(2,0,0)$, $(1,1,1)$, $(1,1,0)$, $(1,0,0)$, and $(0,0,0)$. \[psi\_plat\]](fig1.ps){height="4.5cm"}
$M_T$ and $E_V(\vec{q})$ can be determined precisely from the two-point functions. Figure \[psi\_plat\] shows the nice effective energy plateaus of $J/\psi$ for typical momentum modes at $\beta=2.4$. We also check the dispersion relation of $J/\psi$ and find the largest deviation of squared speed of light $c^2$ from one is less than 4%. The matrix elements $\langle
V(\vec{q},r)|O_{V,j}^{\dagger}(0)|0\rangle$ are included implicitly in the three-point and two-point functions and can be canceled out by taking a ratio $R_{i,\mu,j}(\vec{q},t)$ for some specific $\{i,\mu,j\}$ combinations $$\begin{aligned}
\label{eq_amp}
R_{i,\mu,j}(\vec{q},t)&=&\Gamma^{(3)}_{i,\mu,j}(\vec{q},t_f,t)\times\nonumber\\
&&\frac{\sqrt{4V_3 M_TE_V(\vec{q})}}{C^{i}(t_f-t)}\sqrt{\frac{\Gamma^{(2)}_j(\vec{q},t_f-t)}
{\Gamma^{(2)}_j(\vec{q},t)\Gamma^{(2)}_j(\vec{q},t_f)}},\nonumber\\\end{aligned}$$ which is expected to suppress the contamination from excited states of vector charmonia and should be insensitive to the variation of $t$ in a time window. As such, the desired matrix elements can be derived by the fit form $$R_{i,\mu,j}(\vec{q},t)=\sum\limits_{r}\langle
T_i|J_{\mu}(0)|V(\vec{q},r)\rangle\epsilon_j(\vec{q},r)+\delta f(t),$$ where $\vec{\epsilon}(\vec{q},r)$ is the polarization vector of $J/\psi$ and $\delta f(t)=a
e^{-mt}$ accounts for the residual contamination from excited states.
In the data analysis, the 5000 configurations are divided into 100 bins and the average of 50 measurements in each bin is taken as an independent measurement. For the resultant 100 measurements, the one-eliminating jackknife method is used to perform the fit for the matrix elements ($M_T$ and $E_V$ determined from two-point functions are used as known parameters). Generally speaking, the time separations $t$ and $t_f-t$ should be kept large for the saturation of the ground state, but we have to fix $t_f-t=1$ because of the rapid damping of the glueball signal with respect to the noise. Fortunately this is justified to some extent by the optimal glueball operators, which couple almost exclusively to the ground state. The second step of the data analysis is to extract the form factors $E_1(Q^2)$, $M_2(Q^2)$, and $E_3(Q^2)$ at different $Q^2=2E_V(\vec{q})M_T-M_V^2-M_T^2$ according to Eq. (\[multipole\]). Since the matrix elements are measured from the same configuration ensemble, we carry out a correlated data fitting to get these three form factors simultaneously with a covariance matrix constructed from the jackknife ensemble described above. The symmetric combinations of the indices $\{i,\mu,j\}$ and the momentum $\vec{q}$ which gives the same $Q^2$ are averaged to increase the statistics. In order to get the form factor at $Q^2=0$, we carry out a correlated polynomial fit to the three form factors from $Q^2=-0.5$ to 2.7 ${\rm GeV}^2$, $$\label{inter}
F_i(Q^2)=F_i(0)+a_iQ^2+b_iQ^4,$$ where $F_i$ refers to $E_1$, $M_2$ or $E_3$. Figure \[form\] shows the results of $F_i(Q^2)$ for $\beta=2.4$ (upper panel) and $\beta=2.8$ (lower panel), where the data points are the calculated values with jackknife errors, and the curves are the polynomial fits with jackknife error bands. The corresponding interpolated $F_i(0)$’s are listed in Table \[twobeta\]. Note that the renormalization constant $Z_V^{(s)}$ of the spatial components of the vector current is applied to the final numerical values. We also fit the form factors by functions either linear in $Q^2$ in the range $-0.5\,{\rm GeV^2}<Q^2<1.0\,{\rm GeV^2}$, or by adding a $Q^6$ term to Eq. (\[inter\]) in the range $-0.5\,{\rm GeV^2}<Q^2<2.7\,{\rm GeV^2}$. The resultant $F_i(0)$’s are consistent with that of Eq. (\[inter\]) within errors.
![\[form\]The extracted form factors $E_1(Q^2)$, $M_2(Q^2)$, and $E_3(Q^2)$ in the physical units. The upper panel is for $\beta=2.4$ and the lower one for $\beta=2.8$. The curves with error bands show the polynomial fit with $F_i(Q^2)=F_i(0)+a_iQ^2+b_iQ^4$.](fig2a.eps "fig:"){height="4cm"} ![\[form\]The extracted form factors $E_1(Q^2)$, $M_2(Q^2)$, and $E_3(Q^2)$ in the physical units. The upper panel is for $\beta=2.4$ and the lower one for $\beta=2.8$. The curves with error bands show the polynomial fit with $F_i(Q^2)=F_i(0)+a_iQ^2+b_iQ^4$.](fig2b.eps "fig:"){height="4cm"}
The last step is the continuum extrapolation using the two lattice systems. After performing a linear extrapolation in $a_s^2$, the continuum limits of the three form factors are determined to be $E_1(0)=0.114(12)$ GeV, $M_2(0)=-0.011(5)$ GeV, and $E_3(0)=0.023(8)$ GeV, respectively. Considering the uncertainty of the scale parameter $r_0^{-1}=410(20)$ MeV, which also introduces $5\%$ error, the final result of the form factors is $$\begin{aligned}
E_1(0)&=&0.114(12)(6)\,{\rm GeV}\nonumber\\
M_2(0)&=&-0.011(5)(1)\,{\rm GeV}\nonumber\\
E_3(0)&=&0.023(8)(1)\,{\rm GeV}.\end{aligned}$$ Note that there is a pattern $|E_1(0)|\gg|M_2(0)|\sim |E_3(0)|$; hence, the decay width is dominated by the value of $E_1(0)$. For the continuum value of the tensor glueball mass, we get $M_G=2.37(3)(12)$ GeV (the second error is due to the uncertainty of $r_0$), which is compatible with that in Ref. [@prd73]. Thus, according to Eq. (\[tensor\_width\]), we finally get the decay width $\Gamma(J/\psi\rightarrow \gamma G_{2^{++}})=1.01(22)(10)$ keV. With the total width of $J/\psi$, $\Gamma_{\rm tot}=92.9(2.8)$ keV [@PDG2012], the corresponding branching ratio is $$\Gamma(J/\psi\rightarrow \gamma G_{2^{++}})/\Gamma_{\rm tot}=1.1(2)(1)\times 10^{-2}.$$
$\beta$ $M_T$(GeV) $E_1$ (GeV) $M_2$ (GeV) $E_3$ (GeV)
---------- ------------ ------------- ------------- -------------
2.4 2.360(20) 0.142(07) -0.012(2) 0.012(2)
2.8 2.367(25) 0.125(10) -0.011(4) 0.019(6)
$\infty$ 2.372(28) 0.114(12) -0.011(5) 0.023(8)
: \[twobeta\] The tensor glueball masses $M_T$ as well as the form factors $E_1(0)$, $M_2(0)$ and $E_3(0)$ for the two lattices with $\beta=2.4$ and 2.8. The last row gives the continuum extrapolation. The uncertainty of the scale parameter $r_0$ has not been incorporated yet.
The determined branching ratio is rather large. We admit that the calculation is carried out in the quenched approximation, whose systematical uncertainty cannot be estimated easily without unquenched calculations. A recent full-QCD lattice study of the mass spectrum of glueballs in Ref. [@Gregory:2012] indicates that there is no substantial correction of the masses of the scalar and tensor glueball. Based on this fact, if the form factors also show similar behavior as the masses, the unquenching effects might not change our result drastically. Of course, a full-QCD lattice calculation would be very much welcome.
In experiments, the narrow state $f_J(2220)$ observed by Mark III and BES in the $J/\psi$ decay was once interpreted as a candidate for the tensor glueball. But, the analysis of the processes $p\bar{p}\rightarrow \pi\pi (K\bar{K})$ yields no indication of the narrow $f_J(2220)$ and sets an upper bound for the branch ratios ${\rm Br}(f_J\rightarrow p\bar{p}){\rm Br}(f_J\rightarrow \pi\pi,
K\bar{K})$ (see the review article Ref. [@Crede:2009] and the references therein). Combining this with the results of Mark III and BES, a lower bound for the branching ratio is obtained to be ${\rm Br}[J/\psi\rightarrow \gamma f_J(2220)]>2.5\times 10^{-3}$ [@PDG2012], which seems compatible with our result. However, BESII with substantially more statistics does not find the evidence of a narrow structure around 2.2 GeV of the $\pi\pi$ invariant mass spectrum in the processes $J/\psi\rightarrow \gamma \pi\pi$ [@Ablikim06], and [*BABAR*]{} does not observe it in $J/\psi \rightarrow \gamma (K^+K^-, K_S K_S)$ [@Sanchez:2010]. Recently, based on 225 million $J/\psi$ events, the BESIII Collaboration performs a partial wave analysis of $J/\psi\rightarrow \gamma\eta\eta$ and also finds no evident narrow peak for $f_J(2220)$ in the $\eta\eta$ mass spectrum [@Ablikim:2012]. So the existence of $f_J(2220)$ is still very weak. Another possibility also exists that the tensor glueball is a broad resonance and readily decays to light hadrons. Our result motivates a serious joint analysis of the radiative $J/\psi$ decay into tensor objects in $VV$, $PP$, $p\bar{p}$, and $4\pi$ final states (where $V$ and $P$ stand for vector and pseudoscalar mesons, respectively), among which $VV$ channels may be of special importance since they are kinematically favored in the decay of a tensor meson.
To summarize, we have carried out the first lattice study on the $E_1$, $M_2$, and $E_3$ multipole amplitudes for $J/\psi$ radiatively decaying into the pure gauge tensor glueball $G_{2^{++}}$ in the quenched approximation. With two different lattice spacings, the amplitudes are extrapolated to their continuum limits. The partial decay width and branch ratio for $J/\psi\rightarrow \gamma
G_{2^{++}}$ are predicted to be $1.01(22)(10)$ keV and $1.1(2)(1)\times 10^{-2}$, respectively, which imply that the tensor glueball can be copiously produced in the $J/\psi$ radiative decays if it does exist. To date, the existence of $f_J(2220)$ needs confirmation and a broad tensor glueball is also possible. Hopefully, the BESIII data will be able to clarify the situation.
This work is supported in part by the National Science Foundation of China (NSFC) under Grants No.10835002, No.11075167, No.11021092, No.11275169, and No.10975076. Y. C. and C. L. also acknowledge the support of the NSFC and DFG (CRC110).
[99]{} C.J. Morningstar and M. Peardon, Phys. Rev. D [**56**]{}, 4043 (1997). C.J. Morningstar and M. Peardon, Phys. Rev. D [**60**]{}, 034509 (1999). Y. Chen [*et al.*]{}, Phys. Rev. D [**73**]{}, 014516 (2006). E. Gregory, A. Irving, B. Lucini, C. McNeile, A. Rago, C. Richards, and E. Rinaldi, J. High Energy Phys. 10 (2012) 170. R.M. Baltrusaitis [*et al.*]{}, Phys. Rev. Lett. [**56**]{}, 107 (1986). J.Z. Bai [*et al.*]{}(BES Collaboration), Phys. Rev. Lett. [**76**]{}, 3502 (1996). J. Beringer [*et al.*]{} (Particle Data Group), Phys. Rev. D [**86**]{}, 010001 (2012). L. Kopke and N. Wermes, Phys. Rep. [**174**]{}, 67 (1989). P.D. Barnes [*et al.*]{}, Phys. Lett. B [**309**]{}, 469 (1993). A. Hasan [*et al.*]{}, Nucl. Phys. [**B378**]{}, 3 (1992). A. Hasan and D.V. Bugg, Phys. Lett. B [**388**]{}, 376 (1996). G. Bardin [*et al.*]{}, Phys. Lett. B [**195**]{}, 292 (1987). J. Sculli, J.H. Christenson, G.A. Kreiter, P. Nemethy, and P. Yamin, Phys. Rev. Lett. [**58**]{}, 1715 (1987). C. Evangelista [*et al.*]{} (JETSET Collaboration), Phys. Rev. D [**56**]{}, 3803 (1997). C. Evangelista [*et al.*]{} (JETSET Collaboration), Phys. Rev. D [**57**]{}, 5370(1998). A. Buzzo [*et al.*]{} (JETSET Collaboration), Z. Phys. C [**76**]{}, 475 (1997). B.A. Li and Q.X. Shen, Phys. Lett. [**126B**]{}, 125 (1983). B.A. Li, Q.X. Shen, and K.-F. Liu, Phys. Rev. D [**35**]{}, 1070 (1987). K. Ishikawa, I. Tanaka, K.-F. Liu and B.A. Li, Phys. Rev. D [**37**]{}, 3216 (1988). M. Melis, F. Murgia, and J. Parisi, Phys. Rev. D [**70**]{}, 034021 (2004). H.B. Meyer, J. High Energy Phys. 01 (2009) 071. G. Li, Y. Chen, B.-A. Li, and K.-F. Liu (unpublished). L.-C. Gui, Y. Chen, G. Li, C. Liu, Y.-B. Liu, J.-P. Ma, Y.-B. Yang, and J.-B. Zhang, Phys. Rev. Lett. [**110**]{} 021601 (2013). J.J. Dudek, R.G. Edwards, and D.G. Richards, Phys. Rev. D [**73**]{}, 074507 (2006). Y.-B. Yang, Y. Chen, L.-C. Gui, C. Liu, Y.-B. Liu, Z. Liu, J.-P. Ma, and J.-B. Zhang, Phys. Rev. D [**87**]{}, 014501 (2013). J.J. Dudek, R.G. Edwards, and C.E. Thomas, Phys. Rev. D [**79**]{}, 094504 (2009) C. Liu, J. Zhang, Y. Chen, and J.P. Ma, Nucl. Phys. [**B624**]{}, 360 (2002). S. Su, L. Liu, X. Li, and C. Liu, Int. J. Mod. Phys. A [**21**]{}, 1015 (2006); Chin. Phys. Lett. [**22**]{}, 2198 (2005). V. Crede and C.A. Meyer, Prog. Part. Nucl. Phys. [**63**]{}, 74 (2009). M. Ablikim [*et al.*]{}(BES Collaboration), Phys. Lett. B [**642**]{}, 441 (2006). P. del Amo Sanchez [*et al.*]{} ([*BABAR*]{} Collaboration), Phys. Rev. Lett. [**105**]{}, 172001 (2010). M. Ablikim [*et al.*]{}(BES Collaboration), Phys. Rev. D [**87**]{}, 092009 (2013).
[^1]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Summarizing basic facts from abstract topological modules over Colombeau generalized complex numbers we discuss duality of Colombeau algebras. In particular, we focus on generalized delta functionals and operator kernels as elements of dual spaces. A large class of examples is provided by pseudodifferential operators acting on Colombeau algebras. By a refinement of symbol calculus we review a new characterization of the wave front set for generalized functions with applications to microlocal analysis.'
author:
- |
Claudia Garetto\
Dipartimento di Matematica\
Università di Torino, Italia\
`[email protected]`\
\
Günther Hörmann[^1]\
Institut für Mathematik\
Universität Wien, Austria\
`[email protected]`
title: '[**On duality theory and pseudodifferential techniques for Colombeau algebras: generalized delta functionals, kernels and wave front sets**]{}'
---
[10]{}
H. A. Biagioni. . Number 1421 in Lecture Notes in Math. Springer-Verlag, Berlin, 1990.
C. Garetto. Pseudo-differential operators in algebras of generalized functions and global hypoellipticity. , 80(2):123–174, 2004.
C. Garetto. Pseudodifferential operators with generalized symbols and regularity theory. , 2004.
C. Garetto. Topological structures in [C]{}olombeau algebras: topological $\wt{\C}$-modules and duality theory. , , 2005. To appear.
C. Garetto. Topological structures in [C]{}olombeau algebras: investigation of the duals of ${\Gc(\Om)}$, ${\G(\Om)}$ and ${\GS(\R^n)}$. , , 2005. To appear.
C. Garetto, T. Gramchev, and M. Oberguggenberger. Pseudodifferential operators with generalized symbols and regularity theory. , 2003.
C. Garetto and G. Hörmann. Microlocal analysis of generalized functions: pseudodifferential techniques and propagation of singularities. , 2005. To appear.
M. Grosser, M. Kunzinger, M. Oberguggenberger, and R. Steinbauer. , volume 537 of [ *Mathematics and its Applications*]{}. Kluwer, Dordrecht, 2001.
L. H[ö]{}rmander. ourier integral operators [I]{}. , 127:79–183, 1971.
L. H[ö]{}rmander. , volume III. Springer-Verlag, 1985. Second printing 1994.
G. H[ö]{}rmann and M. V. de Hoop. Microlocal analysis and global solutions of some hyperbolic equations with discontinuous coefficients. , 67:173–224, 2001.
G. H[ö]{}rmann and M. Oberguggenberger. Elliptic regularity and solvability for partial differential equations with [C]{}olombeau coefficients. , 2004(14):1–30, 2004.
G. Hörmann, M. Oberguggenberger, and S. Pilipovic. Microlocal hypoellipticity of linear partial differential operators with generalized functions as coefficients. , 2005. To appear.
M. Kunzinger, M. Oberguggenberger, R. Steinbauer, and J. Vickers. Generalized flows and singular [ODE]{}s on manifolds. , 80:221–241, 2004.
F. Lafon and M. Oberguggenberger. Generalized solutions to symmetric hyperbolic systems with discontinuous coefficients: the multidimensional case. , 160:93–106, 1991.
M. Nedeljkov, S. Pilipovi[ć]{}, and D. Scarpal[é]{}zos. . Pitman Research Notes in Mathematics 385. Longman Scientific [&]{} Technical, 1998.
M. Oberguggenberger and M. Kunzinger. Characterization of [C]{}olombeau generalized functions by their point values. , 203:147–157, 1999.
D. Scarpalézos. Topologies dans les espaces de nouvelles fonctions généralisées de [C]{}olombeau. ${\widetilde{\C}}$-modules topologiques. Université Paris 7, 1992.
D. Scarpalézos. Some remarks on functoriality of [C]{}olombeau’s construction; topological and microlocal aspects and applications. , 6(1-4):295–307, 1998.
D. Scarpalézos. Colombeau’s generalized functions: topological structures; microlocal properties. a simplified point of view. , 25:89–114, 2000.
[^1]: Supported by FWF grant P16820-N04
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
\
National Centre for Radio Astrophysics - Tata Institute of Fundamental Research (NCRA-TIFR), S. P. Pune University Campus, Post Bag 3, Ganeshkhind, Pune 411007, India\
E-mail:
bibliography:
- 'ms.bib'
title: 'Parsec-scale Nuclear Radio Structures in Seyfert Galaxies'
---
Introduction
============
Radio outflows of extents ranging from tens to hundreds of parsecs on one end, and ten to twenty kiloparsecs (kpc) on the other, have been detected in Seyfert and LINER galaxies [@Ulvestad81; @Baum93; @Thean00; @Gallimore06; @Kharb16; @Baldi18]. The origin of these radio outflows is not unequivocally clear. [@Condon82] and [@Baum93] have suggested that these outflows could be powered by starburst superwinds, while [@Colbert96] have pointed out to the distinct morphological differences between Seyfert outflows and those in starburst galaxies, supporting an AGN origin for the former. [@Gallimore06] have concluded from a VLA study of a complete sample that most kiloparsec-scale radio structures (KSRs) in Seyferts and LINERs are AGN-driven, but the starburst superwind contribution cannot be ruled out. It is worth noting here that while Narrow-line Seyfert 1 galaxies (NLS1s) differ from regular Seyferts in their emission-line and (possibly) black hole properties, they exhibit KSRs similar to the ones observed in Seyferts and LINERs [@Richards15; @Berton18].
One of the best ways to probe the AGN role is by studying the radio outflows in Seyferts and LINERs on parsec-scales via the technique of Very Long Baseline Interferometry (VLBI). VLBI arrays have angular resolutions ranging from a few milli-arcseconds (mas, e.g., VLBA[^1]) to tens and hundreds of mas (e.g., MERLIN[^2]). For the nearby galaxies, this translates to spatial scales of a few parsecs to a few sub-kiloparsecs (sub-kpc), respectively. Continuous tracing and accounting for the radio contribution going from parsec to sub-kpc to kpc-scales, can go a long way in settling the AGN versus the starburst superwind debate for Seyfert and LINER outflows.
VLBI observations of several Seyfert galaxies have revealed the presence of weak radio cores and one-sided or two-sided radio jets in them [@Mundell00; @Orienti10]. The brightness temperatures ($T_b$) of the radio cores are typically of the order of $10^6-10^{11}$ K; they exhibit flat or inverted spectral indices [@Ho08]. However, steep spectrum radio “cores" have also been reported in several Seyfert galaxies [@Roy00; @Kharb10; @Bontempi12]. These “cores” could be contaminated by the presence of steep spectrum jet emission. Alternately, the real radio cores may have failed detection at the observing frequencies and could show up at higher radio frequencies. Overall, the VLBI results are consistent with Seyferts and LINERs harbouring low luminosity AGN [@Falcke00; @Middelberg07; @Panessa13].
Probing the Parsec-scale - Kiloparsec-scale Radio Connection
------------------------------------------------------------
VLBI jets in several Seyfert galaxies reveal bends and wiggles [@Roy00; @Lal04; @Hada13]. Jet precession has been invoked to explain these morphological peculiarties [e.g., @Veilleux93; @Middelberg05]. Jet precession in turn could arise due to accretion disk warping, jet instabilities, or the presence of binary black holes. Mrk6 is one such Seyfert galaxy showing an S-shaped radio jet in MERLIN observations [@Kukula96]. It is a peculiar galaxy in that it shows two sets of KSRs aligned nearly perpendicular to each other [@Kharb06]. Our two-frequency (1.6, 5 GHz) VLBA observations of Mrk6 detected an inverted spectrum radio core at the higher frequency for the first time in this source, and resolved the knots in the S-shaped MERLIN jet into elongated jet-like features [@Kharb14a]. The precessing jet model of [@Hjellming81] could fit the parsec-scale emission as well as the bright edges of the north-south oriented KSR. Two episodes of AGN jet activity with precessing jets could explain the entire complex structure observed in Mrk6.
A precessing radio jet could also explain the radio emission on parsec-, sub-kpc- and kpc-scales in the Seyfert / LINER galaxy with an ongoing nuclear starburst, NGC6764 [@Kharb10]. Interestingly, the precession model best-fit values of jet inclination ($\sim18^\circ$) and jet speed ($\sim0.028c$) could explain the observed jet-to-counterjet surface brightness ratio ($R_J\sim1.2$) in NGC6764. We examined the starburst-wind contribution to the radio emission in NGC6764 and another Seyfert + starburst composite galaxy, NGC3079 [see @Irwin03]. We found that only about $25-30\%$ of the total radio flux density appears to arise in clear equatorial emission that could be attributable to stellar winds, in both these sources. VLBI observations of NGC3079 have revealed multiple misaligned jet-like features, which are not all consistently along one position angle with respect to the KSR; this lead [@Kondratko05] to invoke a wide-angle parsec-scale outflow in NGC3079 [see also @Mukherjee18]. In Mrk6 and NGC6764 and possibly NGC3079, precessing jets launched from their black hole $-$ accretion disk systems, could be powering the KSRs and dissipating on sub-kpc or kpc-scales. This could also be the case in the Seyfert galaxies NGC1320 and NGC2992: newly acquired eMERLIN data on these sources reveal $\sim20-30$ parsec-scale core-jet structures in them, that are misaligned to their KSRs (Kharb et al. 2018a, in preparation). Jet precession has also been invoked from direct multi-epoch VLBI observations in the Seyfert / LINER galaxies III Zw 2 and M81 by [@Brunthaler05] and [@Marti11], respectively.
In Figure \[fig1\] we have plotted the parsec-to-kpc-scale jet misalignment angles in Mrk6, NGC6764, NGC1320, and NGC2992, along with others from the literature [@Baum93; @Middelberg05]. While there could be large uncertainties in deriving the KSR position angles due to its diffuse structure (of the order of $10^\circ-20^\circ$), we note that the distribution of misalignment angles in Seyferts does not show any clear preference, unlike what is observed in blazars which show a bimodal distribution in misalignment angles [@Conway93; @Kharb10]. This latter result has been attributed to the presence of low-pitch helical jets in a sub-population of blazars by [@Conway93]. Seyfert outflows on the other hand, show the entire range of misalignment angles. Could continuously bending outflows, that are observationally sampled on different spatial scales in different sources, produce this misalignment angle distribution ? Clearly, many more Seyfert and LINER galaxies need to be examined on several different spatial scales to answer this question.
Changes in the jet propagation direction from parsec to kpc-scales have often been observed in jetted AGN, as has episodic AGN activity [e.g., @Saikia09]. Assuming that jet re-orientation is more likely than re-orientation of large-scale galactic disks producing starburst superwinds, the existence of multiple misaligned KSRs would support an AGN origin for them. Interestingly, secondary misaligned KSRs have been revealed in polarized emission in the Seyfert galaxies NGC2992 and NGC3079 in the Continuum Halos in Nearby Galaxies - an EVLA Survey (CHANG-ES) by [@Irwin17]. These secondary KSRs have not been observed in total intensity images, unlike the case of Mrk6. Nevertheless, they suggest that episodic AGN activity may be a common phenomena in Seyfert and LINER galaxies.
Could Jet Precession Point to the Presence of Binary Black Holes?
-----------------------------------------------------------------
About 1% of SDSS AGN show double-peaked emission lines in their optical spectra [@Wang09]. The presence of double-peaked narrow emission lines in these double-peaked AGN (DPAGN) has been suggested to arise due to (i) binary black holes with associated narrow-line regions (NLRs), (i) jet-NLR interaction or (iii) disky NLRs. In order to search for binary black holes in Seyferts, we have been observing Seyfert DPAGN with phase-referenced VLBI. Dual frequency VLBA observations have detected a single weak $\sim$0.7 mJy radio core of size ($8\times6$ pc) in the Seyfert DPAGN KISSR1494 [@Kharb15b]. The brightness temperature of this core (T$_b\sim1.4\times10^7$ K) and steep radio spectrum ($\alpha\leq-1.5\pm0.5$; $S_\nu\propto\nu^\alpha$), are consistent with optically thin synchrotron emission. However, the core is not “compact” and its brightness is not centrally concentrated; it gets resolved out in images with different weighting schemes. And yet it does not resemble a typical parsec-scale jet component. We have suggested this component to be the base of a tenuous synchrotron-emitting coronal wind. This could explain the double-peaked narrow lines as also originating in a wind-driven NLR.
Similar dual-frequency phase-referenced VLBA observations of the Seyfert DPAGN KISSR1219, have revealed a one-sided $\sim70$ parsec radio jet with a steep radio spectrum [$\alpha\leq-1.0\pm0.2$; @Kharb17a]. Using the jet-to-counterjet surface brightness ratio in KISSR1219 and an orientation angle consistent with its Seyfert type 2 classification ($\theta\gtrsim50^\circ$ based on the typical dusty torus half opening angle), we concluded that the jet speed had decreased from $\gtrsim0.55c$ on parsec-scales to $\gtrsim0.25c$ on kpc-scales. The radio jet was likely pushing the NLR clouds in opposite directions, giving rise to the double-peaked emission lines. This may have resulted in jet deceleration and eventual dissipation.
New VLBA observations of the Seyfert DPAGN KISSR434 have revealed an intriguing $\sim150$ parsec C-shaped curved jet, again suggesting jet-NLR interaction as the source of double-peaked narrow emission lines (Kharb et al. 2018b, in preparation). However, the curved jet itself could arise due to precession in a binary black hole system [e.g., see @Rubinur17; @Rubinur18]. Interestingly, we have found suggestions of a binary black hole system through multi-frequency VLBA observations of the Seyfert / luminous infrared galaxy (LIRG) NGC7674 [@Kharb17b]. This galaxy possesses a Z-shaped $\sim$0.7 kpc radio jet [@Momjian03]. The dual VLBA cores have T$_b\sim2 - 6\times10^7$ K and inverted spectral indices, consistent with being the synchrotron self-absorbed bases of two jets launched from two accreting supermassive black holes. Their projected separation of $\sim$0.65 mas (=0.35 parsec) makes them the closest separation binary black hole pair to be imaged with VLBI.
Summary
=======
Radio outflows are frequently observed in Seyfert and LINER galaxies, inspite of their “radio-quiet" AGN status. These can span extents ranging from tens of parsecs to $10-20$ kpc or more. Nuclear starburst wind contributions to the radio emission cannot be completely ruled out, in at least some Seyferts and LINERs. However, sensitive or phase-referenced VLBI observations that can probe faint parsec-scale radio emission, more often than not reveal weak radio cores and wiggly radio jets in Seyfert and LINER galaxies. In individual sources with multi-scale data on parsec, sub-kpc and kpc-scales, an AGN jet, which could be curved in many cases, can connect the emission on different spatial scales, making the case for an AGN origin for the radio outflows. Curved jets could suggest jet-ISM interaction or precession, which in turn could suggest the presence of binary black holes or accretion disk instabilities.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank the anonymous referee for a positive response to our paper. This conference has been organized with the support of the Department of Physics and Astronomy “Galileo Galilei”, the University of Padova, the National Institute of Astrophysics INAF, the Padova Planetarium, and the RadioNet consortium. RadioNet has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 730562.
[^1]: Very Long Baseline Array
[^2]: Multi-Element Radio Linked Interfermeter Network
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study multi-photon resonances in a strongly-driven three-level quantum system, where one level is periodically swept through a pair of levels with constant energy separation $E$. Near the multi-photon resonance condition $n\hbar\omega = E$, where $n$ is an integer, we find qualitatively different behavior for $n$ even or odd. We explain this phenomenon in terms of families of interfering trajectories of the multi-level system. Remarkably, the behavior is insensitive to fluctuations of the energy of the driven level, and survives deep into the strong dephasing regime. The setup can be relevant for a variety of solid state and atomic or molecular systems. In particular, it provides a clear mechanism to explain recent puzzling experimental observations in strongly-driven double quantum dots.'
author:
- 'J. Danon and M. S. Rudner'
title: 'Multi-level interference resonances in strongly-driven three-level systems'
---
=1
The advent of intense microwave and laser sources has opened a range of new possibilities for investigating the strong-driving regime of both natural and artificial (solid-state) atoms and molecules [@PhysRevA.76.042514; @lang:natphys; @Oliver:science]. In this regime, the amplitude of an applied ac driving field may greatly exceed both the driving field photon energy, $\hbar\omega$, as well as the separation between energy levels of the system. High-order multi-photon processes and multi-level coherences may then become important [@scullyzubairy; @sun:natcomm], leading to interesting dynamical effects which go well beyond the canonical Rabi oscillations of weakly driven two-level systems.
While the dynamics of strongly-driven two-level systems have been studied extensively, both theoretically [@PhysRevA.75.063414; @lzs_review; @PhysRevB.87.235318] and experimentally [@Oliver:science; @PhysRevB.86.121303], [*multi*]{}-level systems offer new avenues to explore. Intriguing and potentially useful phenomena such as amplitude spectroscopy [@berns:nature], population inversion [@Sun:APL; @deGraaf:PRL; @reilly:natcomm], and microwave-induced cooling [@Valenzuela08122006] have been realized in a variety of systems.
Recently a new type of multi-photon resonance was discovered in experiments on spin-blockaded double quantum dots (DQDs) subjected to large-amplitude modulations of a nearby gate electrode [@Laird2009; @PhysRevLett.112.227601]. The resonances show a striking asymmetry, with current [*enhanced*]{} when the electron Zeeman splitting matches an odd-integer multiple of the driving field photon energy, $E_Z = (2n + 1)\hbar\omega$, and [*suppressed*]{} for even-integer resonances $E_Z = 2n\hbar\omega$. Such a dramatic even/odd effect does not occur in two-level systems, and appears to be a robust feature of the multi-level DQDs. Analytical [@PhysRevB.84.241305; @PhysRevB.89.115409] and numerical [@PhysRevB.86.125428; @stehlik:v1] investigations have accounted for the existence of multi-photon resonances, but crucially could not explain the even/odd asymmetry (though a Fano-like origin was speculated [@PhysRevB.84.241305]).
Motivated by this puzzle, we look at the dynamics of strongly-driven multi-level systems. For the case of three levels we find multi-photon resonances with characteristics which differ markedly from those of familiar two-level resonances. We connect our model to the experiments of [@Laird2009; @PhysRevLett.112.227601] and show that it captures all relevant features of the data, including the striking even/odd asymmetry.
![(Color online) Spectrum of (a,b) the two-level Hamiltonian (\[eq:2lh\]) and (c,d) the three-level Hamiltonian (\[eq:3lh\]). In (a,c) the energy levels are plotted as a function of detuning and in (b,d) as a function of time assuming strong driving (thin blue lines). We have set: (b) $A = 5 \varepsilon \gg q$, and (d) $A = 3E \gg q_{1,2}$, $\varepsilon = 0$. In (b,d) we added pairs of possible paths in time bringing the system from $\ket{1}$ to $\ket{S}$ (thick red lines). The paths shown in (d) illustrate why we expect resonances at $E=2n\omega$.[]{data-label="fig:fig1"}](levels2){width="82mm"}
To highlight the key qualitative differences between two-level and multi-level resonances, we first briefly review the phenomenology of multi-photon resonances in a two-level system. We consider a system with basis states $\{ \ket{1}, \ket{S} \}$, its dynamics governed by the Hamiltonian $$H_2(t) = \left( \begin{array}{cc} 0 & q \\ q & -{\varepsilon}(t) \end{array} \right), \quad {\varepsilon}(t) = \varepsilon_0 - A \cos \omega t.
\label{eq:2lh}$$ Here we focus on the case of strong driving, $A \gg q$ and $A > |{\varepsilon}_0|$. Figures \[fig:fig1\]a,b show the instantaneous spectrum of this system, plotted versus detuning ${\varepsilon}$, and time $t$.
The relevant features of the driven system’s dynamics can be understood heuristically in terms of families of interfering trajectories (Fig. \[fig:fig1\]b). For strong driving, transitions take place at relatively well defined points in time $\{t_p\}$ when the two levels are nearly degenerate. Two paths taking the system from state $\ket{1}$ to $\ket{S}$ are indicated by the dashed and solid red lines. In the illustration, the transitions occur at times $t_2$ and $t_4 = t_2 + T$, where $T = 2\pi/\omega$ is the driving period, and the interference phase corresponds to the difference of shaded areas shown, $\Phi = |\Phi_1| - |\Phi_2|$. When ${\varepsilon}_0 = n\omega$ (we set $\hbar = 1$), we have $\Phi = 2n\pi$, and for integer $n$ the interference is constructive. In this case all paths featuring transitions at “even” times $t_{2p}$ mutually interfere constructively, as do all paths with transitions at odd times $t_{2p+1}$. This provides a resonant response. Additional structure results from interferences between these two groups of trajectories, which are sensitive to the individual phases $\Phi_{1,2}$. In the case of sinusoidal driving, the resulting $\Phi_{1,2}$ give rise to the characteristic “Bessel staircase” of modulated resonance intensities, with the Bessel function $J_n(A/\omega)$ controlling the strength of the $n$-photon resonance line [@lzs_review]. The intensities of these two-level multi-photon resonances are thus highly sensitive to both the amplitude and frequency of driving, exhibiting sequences of peaks and nodes as $A/\omega$ is varied.
We now turn our attention to strong driving in a [*multi-level*]{} system. To clearly demonstrate the essential physics of multi-level resonances, we focus on the case of three levels. We assume that the driving field couples strongly to one level, $\ket{S}$, while the energy separation between the other levels $\ket{1}$ and $\ket{2}$ is unaffected (see Figs. \[fig:fig1\]c,d). The state $\ket{S}$ therefore acts as a “shuttle,” mediating population transfer between $\ket{1}$ and $\ket{2}$. This situation is described by the generic Hamiltonian $$\begin{aligned}
H_3(t) = \left( \begin{array}{ccc} E/2 & 0 & q_2 \\ 0 & -E/2 & q_1 \\ q_2 & q_1 & -{\varepsilon}(t) \end{array} \right),
\label{eq:3lh}\end{aligned}$$ written in the basis $\{\ket{2}, \ket{1}, \ket{S}\}$. Here, $E$ is the energy splitting between states $\ket{1}$ and $\ket{2}$, $q_{1,2}$ are the coupling matrix elements, and ${\varepsilon}(t)= \varepsilon_0 - A \cos \omega t$ as before.
Two-level resonances between $\ket{S}$ and $\ket{1}$ or $\ket{2}$, analogous to those described above, can occur whenever the corresponding static detuning ${\varepsilon}_0 \pm E/2$ matches the $n$-photon energy $n\omega$. Such resonances do not present qualitatively new physics.
More interestingly, we investigate the existence of resonances associated with the energy splitting $E$. Such resonances must occur via the strongly modulated level $\ket{S}$, thereby constituting a true multi-level phenomenon.
How could such resonances arise? In the spirit of the discussion above, in Fig. \[fig:fig1\]d we illustrate a characteristic pair of interfering trajectories, in this case from $\ket{1}$ to $\ket{S}$. For large driving amplitude $A \gg {\varepsilon}_0, E$, the interference phase is given by $\Phi_1 + \Phi_2 + \Phi_3 \approx E\,(t_4 - t_1) = 3 E \, (T/2)$. Importantly, this phase is controlled only by the splitting $E$ and the driving half-period $T/2=\pi/\omega$, and [*not*]{} by the driving amplitude or waveform. There exist many such paths, where the last two transitions take place approximately at the same time $t_{p>1}$, all contributing to the full transition amplitude at the same (fourth) order in the couplings $q_{1,2}$. Constructive interference for this series of paths is achieved when $\Phi_1 = \pi E / \omega = 2\pi n$, suggesting the existence of resonances at $E = 2n\omega$, i.e. at [*even*]{} multiples of $\omega$. Similar considerations for transitions from $\ket{1}$ to $\ket{2}$ reveal a series of processes depending on the [*full*]{} driving period $T$, predicting additional resonances at [*all*]{} multiples of the photon energy, $E = n\omega$. Thus we expect this system to display resonances for driving frequencies commensurate with the splitting $E$, showing very different behavior for $E$ an even or odd multiple of $\omega$. Further, in sharp distinction with the two-level case discussed above, these resonances are only weakly sensitive to the driving amplitude $A$ and detuning $\varepsilon_0$. Indeed, the interference phase $\Phi_1$ only changes appreciably when $\varepsilon_0$ is varied on the order of $A$, or vice versa.
We now begin our detailed analysis, which is based on a perturbative treatment in terms of the small parameters $q_{1,2}^2/(A\omega)$ that characterize the strong driving limit. To most clearly exhibit the effect, and to allow us to arrive at analytic results, we focus on a regime of strong dephasing where coherences between $\ket{1}$ and $\ket{S}$ and between $\ket{2}$ and $\ket{S}$ are rapidly lost, on a time scale shorter than the driving period. In contrast, we allow coherences between $\ket{1}$ and $\ket{2}$ to be long-lived on this timescale. The dephasing is modeled by Gaussian white-noise fluctuations on each of the unperturbed energy levels via $$\begin{aligned}
\delta H_3 (t)= \sum_{\alpha} \xi_\alpha(t) \ket{\alpha}\bra{\alpha},\quad \alpha \in \{1, 2, S\},\end{aligned}$$ with $\overline{\xi_\alpha (t) \xi_\beta (t')} = \Gamma_\alpha \delta(t-t')\delta_{\alpha\beta}$, where the overbar indicates averaging over noise realizations. Within this model we calculate the rates of interlevel transitions, working up to fourth order in the couplings $q_{1,2}$.
Strong dephasing is particularly relevant for the experiments in Refs. [@Laird2009; @PhysRevLett.112.227601], where the level corresponding to $\ket{S}$ exhibits strong lifetime broadening due to coupling to a nearby reservoir (see discussion below). The multi-level resonances survive deep into the strong-dephasing regime, where the quasi-two-level resonances at ${\varepsilon}_0 \pm E/2 = n\omega$ are completely washed out.
The first analytical step is to transform to a modified interaction picture via $\ket{\psi_R(t)} = e^{iR(t)}\ket{\psi(t)}$, with $R(t) = \sum_{\alpha} \phi_\alpha(t)\ket{\alpha}\bra{\alpha}$. The phases $\phi_\alpha$ are given by $\phi_\alpha(t) = -\int_0^t d\tau\, \tilde{{\varepsilon}}_\alpha(\tau)$, with $\tilde{\varepsilon}_{1,2}(\tau) = \mp\frac12 E + \xi_{1,2}(\tau)$ and $\tilde{\varepsilon}_S(\tau) = {\varepsilon}(\tau) + \xi_S(\tau)$. States in this interaction picture evolve according to $i\frac{d}{dt}\ket{\psi_R} = \tilde{H}_3(t)\ket{\psi_R}$, with ${\tilde{H}_3(t) = -\dot{R} + e^{iR(t)}(H_3+\delta H_3)e^{-iR(t)}}$. This yields $$\begin{aligned}
\label{eq:H3R}\tilde{H}_{3}(t) = \left( \begin{array}{ccc} 0 & 0 & q_2 e^{-i\phi_{S2}(t)} \\ 0 & 0 & q_1 e^{-i\phi_{S1}(t)} \\ q_2 e^{i\phi_{S2}(t)} & q_1 e^{i\phi_{S1}(t)} & 0 \end{array} \right),\end{aligned}$$ where $\phi_{\alpha\beta}(t) \equiv \phi_\alpha(t) - \phi_\beta(t)$.
The transition rate between states $\ket{\alpha}$ and $\ket{\beta}$ is calculated as the time-derivative of the transition probability, $$\begin{aligned}
W_{\alpha\to\beta} = \frac{d}{dt} \overline{\big|{\braket{\beta|U(t)|\alpha}}\big|^2},\end{aligned}$$ where $U(t)$ evolves the system between times $0$ and $t$. We expand the time-evolution operator in powers of $q_{1,2}$ as $U(t) = 1 + U^{(1)}(t) + U^{(2)}(t) + \dots$, with $$\begin{aligned}
U^{(m)}(t) & = (-i)^m\int_0^t dt_1\cdots\int_0^{t_{m-1}}\!\!\!\!\!\!\!dt_m\ \tilde{H}_{3}(t_1)\cdots\tilde{H}_{3}(t_m). \nonumber\end{aligned}$$ Working up to third order in $q_{1,2}$ gives access to the transition [*rates*]{} up to fourth order in the couplings.
For illustration we now evaluate $W_{1\to S}$ to lowest (second) order; other rates are obtained similarly. We write $$\begin{aligned}
W_{1\to S}^{(2)} & = \frac{d}{dt} \overline{\big|{\braket{S|U^{(1)}(t)|1}}\big|^2} \nonumber\\
& = q_1^2 \frac{d}{dt} \int_0^t dt_1\int_0^t dt_2\ \overline{e^{i[\phi_{S1}(t_1)-\phi_{S1}(t_2)]}},\end{aligned}$$ and use $\overline{\exp\{i\xi(t)\}} = \exp\{-\tfrac{1}{2}\overline{\xi(t)^2}\}$. The result is simplified under the assumption $\Gamma_S \gg \omega,\Gamma_{1,2}$, giving $$\begin{aligned}
W_{1\to S}^{(2)} & = \frac{q_1^2\Gamma_S}{(\frac12 E - \varepsilon_0 + A \cos \omega t)^2 + \tfrac{1}{4}\Gamma_S^2}.
\label{eq:w212}\end{aligned}$$ Moving to the strong driving limit $A \gg \Gamma_S$ and assuming $A>|{\varepsilon}_0 - \tfrac{1}{2}E|$, the transition rate displays sharp bursts, well-separated in time, occurring whenever the levels cross, i.e. when $A\cos\omega t \approx \varepsilon_0 - \frac12 E$. Averaging these bursts over one period yields $$\begin{aligned}
W_{1\to S}^{(2)} & \approx \frac{2q_1^2}{\sqrt{A^2 - (\frac12 E - \varepsilon_0)^2}}.
\label{eq:w122nd}\end{aligned}$$ Similarly, we find $W_{2\to S}^{(2)} \approx 2q_2^2/[A^2 - (\tfrac{1}{2}E + \varepsilon_0)^2]^{1/2}$ in the same limit, and identical rates for the reverse processes $W^{(2)}_{S\to 1}$ and $W^{(2)}_{S\to 2}$.
Multi-level interference resonances first arise at fourth order, $$\begin{aligned}
W^{(4)}_{\alpha\to\beta} = \frac{d}{dt}& \bigg\{ \overline{ \big|{\braket{\beta|U^{(2)}(t)|\alpha}}\big|^2} \nonumber\\
& +\ 2\,{\rm Re}\overline{ \braket{\alpha|U^{\dagger(3)}(t)|\beta}\braket{\beta|U^{(1)}(t)|\alpha}} \bigg\}.\label{eq:w4}\end{aligned}$$ Due to the form of $\tilde{H}_3$, the rates $W^{(4)}_{1\leftrightarrow 2}$ only involve the first term in Eq. (\[eq:w4\]), while the rates $W^{(4)}_{1,2 \leftrightarrow S}$ involve only the last. Proceeding along similar lines as above, we assume $\Gamma_S \gg \omega,\Gamma_{1,2}$ and work in the strong-driving limit $A\gg \Gamma_S$.
After some algebra we find analytic approximations for the rates in two important cases, valid for times $t \gg \Gamma_{1,2}^{-1}$. First, at zero detuning, $\varepsilon_0 = 0$, we find $W^{(4)}_{1,2 \leftrightarrow S} \approx - g_0 \bar{W}$ and $W^{(4)}_{1\leftrightarrow 2} \approx (\tfrac{1}{2} g_0 + h_0)\bar{W}$, with $\bar{W} = 2\pi q_1^2q_2^2/(A^2\omega)$ and $$\begin{aligned}
g_0 & = \frac{2\cos(n\pi)\sinh(\tfrac{1}{2}\Gamma')+e^{\Gamma'}-\cos 2n\pi}{\cosh\Gamma'-\cos 2n\pi}, \nonumber\\
h_0 & = \frac{\sin^2(\tfrac{1}{2}n\pi )\sinh(\Gamma')}{\cosh\Gamma'-\cos 2n\pi}.\nonumber\end{aligned}$$ Here we use the (continuum-valued) dimensionless energy splitting $n = E/\omega$ and dephasing rate $\Gamma' = (\Gamma_1+\Gamma_2)\pi/\omega$. Second, for arbitrary detuning but [*integer*]{} $n$, we find $W^{(4)}_{1,2 \leftrightarrow S} \approx -g_i \bar{W}$ and $W^{(4)}_{1\leftrightarrow 2} \approx (\tfrac{1}{2} g_i + h_i)\bar{W}$, with $$\begin{aligned}
g_i & = \frac{\cos(n d_-)[\sinh (\tfrac{\Gamma'd_-}{2\pi}) + \sinh (\frac{\Gamma'd_+}{2\pi})]+e^{\Gamma'}-1}{(\cosh\Gamma'-1)(1-\delta^2)}, \nonumber\\
h_i & = \frac{\sin^2(\tfrac{1}{2}nd_+)\coth(\tfrac{1}{2}\Gamma')}{1-\delta^2},\nonumber\end{aligned}$$ where $d_\pm = \pi \pm 2 \sin^{-1}\delta$ and $\delta = \varepsilon_0/A$.
![(Color online) The rate $W^{(4)}_{1\to S}$ at (a) zero detuning as a function of $n = E/\omega$, and (b) for integer $n=1,2,3$ as a function of $\delta$. (c,d) The same for the rate $W^{(4)}_{1\to 2}$. In all plots we used $\Gamma'/\pi = 0.3$. In (d) the curves are offset in steps of $\tfrac{3}{2}$. []{data-label="fig:fig3"}](ws){width="82mm"}
In Fig. \[fig:fig3\] we plot the rates as a function of $n$ for $\delta = 0$ (a,c) and as a function of $\delta$ for $n = 1,2,3$ (b,d), in all plots setting $\Gamma'/\pi = 0.3$. The rates display resonant features at integer $n$. Moreover, the resonances for even and odd $n$ are qualitatively different, as anticipated above. The negative sign of $W^{(4)}_{1\to S}$ indicates that this fourth-order contribution provides a suppression of the large (second-order) background transition rate $W^{(2)}_{1 \to S}$, see Eq. (\[eq:w122nd\]). As long as $\Gamma_1+\Gamma_2 > W^{(2)}_{1 \to S}$ the total rate $W^{(2)}_{1 \to S} + W^{(4)}_{1 \to S}$ is positive. For $W^{(2)}_{1 \to S} > \Gamma_1+\Gamma_2$, lifetime broadening of $\ket{1}$ and $\ket{2}$ due to driving-induced transitions to $\ket{S}$ becomes dominant. To capture this effect, higher terms in the perturbation expansion must be included.
We now connect our results to the experimental observations of Refs. [@Laird2009; @PhysRevLett.112.227601], in which current through spin-blockaded DQDs was measured in the presence of strong ac driving. In the two-electron regime, the low-energy electronic subspace of the DQD is spanned by five states: a “(1,1)” spin-singlet and a spin triplet with a single electron in each dot, and a “(0,2)” spin-singlet with double occupancy of the right dot (the left dot being empty). Current flow is mediated by the (0,2) singlet state, which is the only state with direct coupling to the drain lead.
In spin blockade, current is limited by the lifetimes of the (1,1) triplet states, which to zeroth approximation do not couple to the (0,2) singlet. Finite coupling between spin triplet and singlet levels occurs via spin-orbit, hyperfine, and/or inhomogeneous Zeeman coupling. Away from singlet-triplet degeneracy points, ac driving (e.g. applied to one of the gate electrodes controlling the DQD potential) can provide the energy necessary to stimulate triplet-singlet transitions [@katja:science; @laird]. When the driving frequency and level splittings are in resonance, such coupling is expected to lift the blockade and produce an [*enhancement*]{} of current. The striking even/odd effect observed in Refs. [@Laird2009; @PhysRevLett.112.227601] thus clearly does not fit in this simple picture. Furthermore, only a smooth modulation of resonance intensity with $A/\omega$ was observed, in stark contrast to the nodal structure expected for conventional multi-photon resonances as described in the introduction.
As we will now show, the multi-level multi-photon resonances described above account for all of the main features of the experimental data. The three-level model certainly does not provide a complete representation of dynamics in the full five-dimensional low-energy subspace of a spin-blockaded DQD, but it nonetheless captures the essential physics at play near the resonances. To make the connection explicit, state $\ket{S}$ in our model represents the (0,2) singlet state of the DQD, while $\ket{1}$ represents the triplet state ${T_+}$, with both electron spins pointing up, and $\ket{2}$ represents a particular superposition of the (1,1) singlet and $T_0$ states, which is determined by Zeeman energy inhomogeneities in the DQD [@footnote_fourthlevel].
Using all contributions to the transition rates up to fourth order, we compute the steady-state current via a master equation for the time-dependent level occupation probabilities $\{p_\alpha\}$, $$\begin{aligned}
\frac{dp_1}{dt} = & -p_1(W_{1\to S}+W_{1\to 2}) + p_2W_{2\to 1} \nonumber\\
\label{eq:Master}& + \frac{1}{2}(p_1W_{1\to S}+p_2W_{2\to S}),\end{aligned}$$ where $p_2 = 1-p_1$. To eliminate $p_S$, we assumed that the decay of $\ket{S}$ and the consecutive reloading of $\ket{1}$ or $\ket{2}$ (with equal probabilities) happens instantaneously on the time scale of the dynamics of $p_{1,2}$. We solve for $dp_1^{\rm (eq)}/dt=0$, with the steady-state current following as $I/e = p_1^{\rm (eq)}W_{1\to S} + p_2^{\rm (eq)}W_{2 \to S}$.
![(Color online) Calculated current through a driven double quantum dot in spin blockade, normalized to the background current $I_{\rm bg}$. (a) The current at $\delta = 0$ as a function of $\omega$ and $E$. (b) Slow modulation of the resonances: the current as a function of $\delta$ for fixed $n=1,2,3$. In all plots we used $q_1^2/A = 5$ MHz, $q_2^2/A = 50$ MHz, and $\Gamma_{1,2} = 100$ MHz. []{data-label="fig:fig4"}](plots){width="82mm"}
To compare with the data presented in Fig. 2d of Ref. [@PhysRevLett.112.227601], we set $\delta = 0$ and assume that $\omega, E \sim$ GHz. We set $q_1^2/A = 5$ MHz and $q_2^2/A = 50$ MHz, i.e. $q_2^2 / q_1^2 = 10$ [@PhysRevB.81.201305], and choose $\Gamma_{1,2} = 100$ MHz. In Fig. \[fig:fig4\]a we plot the resulting steady-state current, normalized to $I_{\rm bg}$, the off-resonant “background” current (i.e. the current due to “direct” second-order transitions $W_{1,2 \to S}^{(2)}$ associated with repeated sweeps through the $S$-$T_+$ level crossing; in the experiment $I_{\rm bg} \sim 15$ pA). The model reproduces all important features of the data: a resonant response of current along all $n$-photon lines, alternating between enhancement for odd $n$ and suppression for even $n$. At even $n$, the negative contributions $W_{1,2 \to S}^{(4)}$ suppress escape from $\ket{1}$ and $\ket{2}$ to $\ket{S}$, resulting in a reduction of current relative to the background. The rate $W_{1 \leftrightarrow 2}^{(4)}$ is largest for odd $n$, where it efficiently mixes $\ket{1}$ and $\ket{2}$ and thus enhances the escape rate out of the most strongly blocked state, $\ket{1}$, thereby increasing the total current. Including a second unpolarized (1,1) level [@footnote_fourthlevel], split from $\ket{1}$ by $E'$, would yield another fan of current peaks and dips at $E'=n\omega$, reproducing the “doubled” line shape of Fig. 2d of Ref. [@PhysRevLett.112.227601].
We finally investigate the detuning-dependence of the current, which in the experiment showed a strikingly slow modulation (on the scale of $\varepsilon_0\sim A$) with qualitatively distinct shapes for each of the resonances, see Fig. 3b of [@PhysRevLett.112.227601]. In Fig. \[fig:fig4\]b we plot the current as a function of $\delta$ at fixed $n = 1,2,3$, using the same parameters as for Fig. \[fig:fig4\]a. The detuning-dependence of $I$ agrees well with the experimental observations. Here it arises from the weak dependence of the interference phases $\Phi_n$ on $\varepsilon_0$, as explained above.
To summarize, we investigated multi-photon resonances in a strongly-driven three-level quantum system. We identified new resonant responses which crucially depend on the multi-level structure of the system. We further revealed how these resonances provide a mechanism to explain recent puzzling experimental observations in strongly-driven double quantum dots. Interestingly, the behavior survives deep into the regime of strong dephasing on one of the levels. Detailed explorations of the fully-coherent regime and the role of decoherence are interesting directions for further study.
We thank C. M. Marcus and K. Flensberg for helpful discussions. MR acknowledges support by the Villum Foundation.
[10]{}
M. Mark, F. Ferlaino, S. Knoop, J. G. Danzl, T. Kraemer, C. Chin, H.-C. Nägerl, and R. Grimm, Phys. Rev. A **76**, 042514 (2007).
F. Lang, P. [v. d. Straten]{}, B. Brandstätter, G. Thalhammer, K. Winkler, P. S. Julienne, R. Grimm, and J. [Hecker Denschlag]{}, Nat. Phys. **4**, 223 (2008).
W. D. Oliver, Y. Yu, J. C. Lee, K. K. Berggren, L. S. Levitov, and T. P. Orlando, Science **310**, 1653 (2005).
M. O. Scully and M. S. Zubairy, *Quantum Optics* (Cambridge University Press, Cambridge, UK, 1997).
G. Sun, X. Wen, B. Mao, J. Chen, Y. Yu, P. Wu, and S. Han, Nat. Comm. **1**, 51 (2010).
S. Ashhab, J. R. Johansson, A. M. Zagoskin, and F. Nori, Phys. Rev. A **75**, 063414 (2007).
S. N. Shevchenko, S. Ashhab, and F. Nori, Phys. Rep. **492**, 1 (2010).
H. Ribeiro, J. R. Petta, and G. Burkard, Phys. Rev. B **87**, 235318 (2013).
J. Stehlik, Y. Dovzhenko, J. R. Petta, J. R. Johansson, F. Nori, H. Lu, and A. C. Gossard, Phys. Rev. B **86**, 121303 (2012).
D. M. Berns, M. S. Rudner, S. O. Valenzuela, K. K. Berggren, W. D. Oliver, L. S. Levitov, and T. P. Orlando, Nature **455**, 51 (2008).
Guozhu Sun, Xueda Wen, Yiwen Wang, Shanhua Cong, Jian Chen, Lin Kang, Weiwei Xu, Yang Yu, Siyuan Han, and Peiheng Wu, Appl. Phys. Lett. [**94**]{}, 102502 (2009). S. E. de Graaf, J. J. Leppakangas, A. Adamyan, A. V. Danilov, T. Lindstrom, M. Fogelstrom, T. Bauch, G. Johansson, and S. E. Kubatkin, Phys. Rev. Lett. [**111**]{}, 137002 (2013). J. I. Colless, X. G. Croot, T. M. Stace, A. C. Doherty, S. D. Barrett, H. Lu, A. C. Gossard, and D. J. Reilly, Nat. Comm. **5**, 3716 (2014).
S. O. Valenzuela, W. D. Oliver, D. M. Berns, K. K. Berggren, L. S. Levitov, and T. P. Orlando, Science **314**, 1589 (2006).
A. Ferrón, D. Domínguez, and M. J. Sánchez, Phys. Rev. B **82**, 134522 (2010).
L. Du and Y. Yu, Phys. Rev. B **82**, 144524 (2010).
E. A. Laird, C. Barthel, E. I. Rashba, C. M. Marcus, M. P. Hanson, and A. C. Gossard, Semicond. Sci. Technol. **24**, 064004 (2009).
J. Stehlik, M. D. Schroer, M. Z. Maialle, M. H. Degani, and J. R. Petta, Phys. Rev. Lett. **112**, 227601 (2014).
E. I. Rashba, Phys. Rev. B **84**, 241305 (2011).
G. Széchenyi and A. Pályi, Phys. Rev. B **89**, 115409 (2014).
M. P. Nowak, B. Szafran, and F. M. Peeters, Phys. Rev. B **86**, 125428 (2012).
J. Stehlik, M. D. Schroer, M. Z. Maialle, M. H. Degani, and J. R. Petta, arXiv:1312.3875v1 (2013).
K. C. Nowack, F. H. L. Koppens, Y. V. Nazarov, and L. M. K. Vandersypen, Science **318**, 1430 (2007).
E. A. Laird, C. Barthel, E. I. Rashba, C. M. Marcus, M. P. Hanson, and A. C. Gossard, Phys. Rev. Lett. **99**, 246601 (2007).
Including a fourth level representing the second unpolarized (1,1) level would produce to first approximation another, equivalent series of resonances at $E' = n\omega$, where $E'/E = g_1/g_2$ is set by the ratio of the effective g-factors in the two dots. Adding the fifth level, $T_-$, would not significantly change the physics.
S. Nadj-Perge, S. M. Frolov, J. W. W. van Tilburg, J. Danon, Y. V. Nazarov, R. Algra, E. P. A. M. Bakkers, and L. P. Kouwenhoven, Phys. Rev. B **81**, 201305 (2010).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Recently, Tsupko *et al.* have put forward the very interesting proposal to use the shadows of high-redshift supermassive black holes (SMBHs) as standard rulers. This would in principle allow us to probe the expansion history within a redshift range which would otherwise be challenging to access. In this short note, we critically examine this proposal, and identify a number of important issues which had been previously overlooked. These range from difficulties in obtaining reliable SMBH mass estimates and reaching the required angular resolution, to an insufficient knowledge of the accretion dynamics of high-redshift SMBHs, to weak lensing effects leading to typical angular deflections much larger than the SMBH shadow angular sizes. While these issues would currently appear to prevent high-redshift SMBH shadows from being used as robust standard rulers, we hope that our pointing them out early will help in making such a probe theoretically mature by the time it will be experimentally feasible.'
author:
- Sunny Vagnozzi
- Cosimo Bambi
- Luca Visinelli
bibliography:
- 'BH.bib'
title: Concerns regarding the use of black hole shadows as standard rulers
---
Introduction {#sec:intro}
============
One of the most important breakthroughs in 21st century cosmology has been the ability to probe the expansion history of the universe and the relation between distance and redshift far beyond our local neighbourhood. These determinations usually rely on objects (or classes of objects) with well-known intrinsic properties, such as so-called *standard candles* [@Perlmutter:1998np; @Riess:1998cb], *standard sirens* [@Schutz:1986gp; @Holz:2005df], *standard rulers* [@Eisenstein:1998tu; @Eisenstein:2005su], and *standard clocks* [@Jimenez:2001gg; @Heavens:2014rja]. Here, we shall mostly be concerned with the concept of a standard ruler (SR), an object of known intrinsic size. The distance to a SR can be then determined by comparing its observed angular size to its known physical size. The archetype of SRs (which is more precisely a statistical SR) is represented by the scale imprinted by Baryon Acoustic Oscillations (BAOs) [@Eisenstein:2005su] which are set up by the interplay between radiation pressure and gravity of the strongly coupled photon-baryon fluid in the early universe. BAOs imprint a scale corresponding to the sound horizon at baryon drag in the distribution of matter, resulting in a preferred clustering scale for tracers of the large-scale structure. A statistical analysis of a given large-scale structure tracer at a given redshift allows one to extract this preferred scale, and hence the distance to the redshift in question.
The use of BAOs as SRs has revolutionised our understanding of dark energy and cosmic acceleration and has been instrumental in establishing the $\Lambda$CDM concordance cosmological model [@Aubourg:2014yra]. Nonetheless, there are plenty of theoretical and observational reasons to believe that $\Lambda$CDM might not be the end of the story, ranging from considerations over the theoretical implausibility of a cosmological constant of the observed magnitude [@Weinberg:1988cp], to mismatches between cosmological parameters estimated from independent probes (such as the “$H_0$ tension”, see e.g. [@DiValentino:2016hlg; @Bernal:2016gxb; @Vagnozzi:2017ovm; @Renk:2017rzu; @Mortsell:2018mfj; @Vagnozzi:2018jhn; @Nunes:2018xbm; @Yang:2018euj; @Guo:2018ans; @Aylor:2018drw; @Poulin:2018cxd; @DiValentino:2019exe; @Pan:2019gop; @Vagnozzi:2019ezj; @Visinelli:2019qqu; @Cai:2019bdh; @Pan:2019hac; @DiValentino:2019ffd; @Escudero:2019gvw; @DiValentino:2019jae]) suggesting that the $\Lambda$CDM description of the dark sectors of the Universe might be incomplete. Anticipated improvements in BAO measurements from future surveys such as DESI [@Aghamousa:2016zmz] and Euclid [@Laureijs:2011gra] will be crucial towards either further strengthening the case for $\Lambda$CDM, or conclusively finding evidence for new physics.
Regardless of the success of BAOs in mapping the late-time expansion history, it is desirable to find novel and independent standard rulers, which might be used to either cross-validate existing BAO distance measurements or, more intriguingly, allow us to probe a new redshift window otherwise not accessible to BAOs. A wide variety of novel standard rulers have been proposed in the literature, including (but not limited to): double-lobed radio sources [@Buchalter:1997vz; @Carlberg:1998rk], X-ray gas mass fractions from galaxy clusters [@Allen:2002sr; @Mantz:2007qh], ultra-compact radio sources [@1993Natur.361..134K; @1994ApJ...425..442G], Minkowski functionals of the large-scale structure density field [@Park:2009ja; @Blake:2013noa], dust time lags [@Hoenig:2014jca; @Honig:2016oyn], strongly-lensed systems [@Paraficz:2009xj; @Agnello:2015ala], the so-called “linear point” in the large-scale structure two-point correlation function [@Anselmi:2015dha; @Anselmi:2018hdn], the cosmic homogeneity scale [@Ntelis:2018ctq; @Nesseris:2019mlr], velocity-induced acoustic oscillations at Cosmic Dawn [@Munoz:2019fkt; @Munoz:2019rhi], and light echos [@Kervella:2008ne; @Bond:2008ax]. However, it is fair to say that none has (yet) even gone close to achieving the same level of maturity and reliability of BAOs.
Recently, a very interesting possibility for a new SR making use of black hole (BH) shadows has been proposed by Tsupko *et al.* in [@Tsupko:2019pzg]. A BH shadow is the apparent (*i.e.* gravitationally lensed) image of the photon sphere, the region in the vicinity of the BH along which photons travel in unstable circular orbits. More precisely, the proposal advocated by [@Tsupko:2019pzg] makes us of measurements of the angular sizes of supermassive black hole (SMBH) shadows (whose evolution as a function of redshift is in principle known, if the SMBH mass is known) for SMBHs located at cosmological distances. A very interesting follow-up in [@Qi:2019zdk] examined the cosmological implications of this SR, finding that such a probe can potentially lead to exquisite constraints on the expansion history at very high redshift ($z \gtrsim 10$), as well as on cosmological parameters such as $\Omega_m$. On the other hand, at low redshifts SMBH shadows might allow for precise constraints on the Hubble constant $H_0$, thus possibly providing more insight into the $H_0$ tension. Therefore, it appears that the use of SMBH shadows as standard rulers can provide an extremely successful cosmological probe.
In this short note, we wish to advocate a more cautious approach on the subject, despite the promising results of Tsupko *et al.* [@Tsupko:2019pzg] being formally correct. In particular, our goal is to point out a number of rather important practical issues and difficulties overlooked by [@Tsupko:2019pzg], which render the use of SMBH shadows as standard rulers more problematic than what has been originally thought. While we certainly do not want to discourage astrophysicists and cosmologists from thinking about using SMBH shadows as standard rulers, given the huge potential therein, we believe that at the same time it is important to point out the associated difficulties as early in the process as possible, in order to allow such a probe to reach a high level of theoretical maturity by the time it will be experimentally feasible.
The rest of this paper is organized as follows. In Sec. \[sec:shadow\] we review of the concept of a BH shadow, how its angular size evolves with redshift, and how it can be used as a standard ruler in an expanding universe. In Sec. \[sec:issue\] we discuss why we find such a probe to be problematic, identifying six independent concerns. We provide concluding remarks in Sec. \[sec:conclusions\]. Throughout the paper, we work in Planck units with $G = c = \hbar = 1$.
Black hole shadows as standard rulers {#sec:shadow}
=====================================
Black holes are unique regions of space-time, and might hold the key towards the unification of Quantum Mechanics and General Relativity (GR) [@Hawking:1976ra; @Mathur:2005ai; @Dvali:2011aa; @Giddings:2017jts; @Giddings:2019jwy]. They represent the final state of continuous gravitational collapse of matter and are defined by their event horizon, a one-way causal space-time boundary from which nothing can escape [@Einstein:1916vd; @Schwarzschild:1916uq; @Penrose:1964wq]. Observationally speaking, BHs are ubiquitous in a wide range of environments (for a recent review on astrophysical BHs see [@Bambi:2019xzp]). Of particular interest are so-called supermassive BHs (SMBHs), with masses in the range $ \left ( 10^{5}-10^{10} \right ) \,M_{\odot}$. It is believed that most sufficiently massive galaxies harbor SMBHs at their centres [@LyndenBell:1969yx; @Kormendy:1995er].
The so-called BH shadow is an important feature resulting from the combination of an event horizon (or, more precisely, of a photon sphere, around which photons orbit the BH on unstable circular orbits) and the strong gravitational lensing in the vicinity of a BH. More formally, the BH shadow constitutes a closed curved on the sky which separates capture orbits from scattering orbits, see [@Dokuchaev:2019jqq] for a review. In particular, for a BH surrounded by a geometrically thick, optically thin emission region, the shadow should be visible as a dark region on the sky, surrounded by a bright emission ring (see e.g. [@Luminet:1979nyg; @Lu:2014zja; @Cunha:2018acu; @Gralla:2019xty; @Narayan:2019imo]). For a Schwarzschild BH, the radius of the shadow $r_{\rm sh}=3\sqrt{3}M \approx 5.2M$ is equal neither to the Schwarzschild radius $r_{\rm s}=2M$ nor to the photon sphere radius $r_{\rm ph}=3M$, but is actually slightly larger than both due to the fact that the shadow is the gravitationally lensed image of the photon sphere [@Luminet:1979nyg].
Very-long-baseline interferometry (VLBI) has been argued to be a promising technique to image the shadows of SMBHs [@Falcke:1999pj]. A very successful example is represented by the Event Horizon Telescope (EHT) [@Fish:2016jil], a global network of radio telescopes which in 2019 imaged the shadow of the SMBH M87\* [@Akiyama:2019cqa; @Akiyama:2019brx; @Akiyama:2019sww; @Akiyama:2019bqs; @Akiyama:2019fyp; @Akiyama:2019eap]. The shadow of M87\* appears to be broadly consistent with that of GR Kerr BH [@Kerr:1963ud], although the possibility that M87\* might be a more complex object (either a non-Kerr BH or a BH mimicker) cannot yet be excluded. In fact, a number of works have examined the possibility of using M87\*’s shadow as a probe of fundamental physics, and possibly of deviations from GR [@Moffat:2019uxp; @Nokhrina:2019sxv; @Abdikamalov:2019ztb; @Held:2019xde; @Wei:2019pjf; @Shaikh:2019fpu; @Tamburini:2019vrf; @Davoudiasl:2019nlo; @Ovgun:2019yor; @Bambi:2019tjh; @Nemmen:2019idv; @Churilova:2019jqx; @Safarzadeh:2019imq; @Firouzjaee:2019aij; @Konoplya:2019nzp; @Kawashima:2019ljv; @Contreras:2019nih; @Bar:2019pnz; @Jusufi:2019nrn; @Vagnozzi:2019apd; @Banerjee:2019cjk; @Roy:2019esk; @Ali:2019khp; @Long:2019nox; @Zhu:2019ura; @Contreras:2019cmf; @Dokuchaev:2019pcx; @Wang:2019tto; @Konoplya:2019goy; @Roy:2019hqf; @Pavlovic:2019rim; @Biswas:2019gia; @Wang:2019skw; @Nalewajko:2019mxh; @Tian:2019yhn; @Cunha:2019ikd; @Banerjee:2019nnj; @Shaikh:2019hbm; @Vrba:2019vqh; @Kumar:2019pjp; @Allahyari:2019jqz; @Li:2019lsm; @Jusufi:2019ltj; @Rummel:2019ads; @Kumar:2020hgm].
The proposal put forward by Tsupko *et al.* in [@Tsupko:2019pzg] is to use SMBH shadows as standard rulers, by computing the angular size $\alpha_{\rm sh}(z)$ of the shadow of a Schwarzschild BH at arbitrary redshift. The issue of computing the size of a BH shadow at cosmological distances is actually highly non-trivial. The main difficulties in performing an analytical calculation are first of all that of finding an adequate description of a BH embedded in an expanding universe, and next that of computing light ray trajectories in the strong gravity regime. Usually the problem is approached by exploiting constants of motion which are either conserved or approximately conserved. However, the Friedmann-Lemaître-Robertson-Walker (FLRW) space-time does not possess a time-like Killing vector, implying that energy is not conserved, which complicates the analytical computation of BH shadows therein. The issue of embedding a BH solution in an expanding universe has been tackled in recent years, for instance within the so-called Einstein-Straus model [@Einstein:1945id; @Einstein:1946zz] or within the McVittie metric [@McVittie:1933zz; @Nolan:1998xs; @Nolan:1999kk; @Nolan:1999wf] (see also [@Carrera:2008pi; @Gibbons:2009dr; @Nandra:2011ug; @Nandra:2011ui]). Other works focused on computing the shadow of a Schwarzschild BH embedded in a de Sitter universe [@Stuchlik:1999qk; @Bakala:2007pw; @Stuchlik:2018qyz]. More progress was made in [@Perlick:2018iye], where the authors computed the size of a Schwarzschild BH shadow as seen by a comoving observer in an expanding universe with a cosmological constant.
A later study in [@Bisnovatyi-Kogan:2018vxl] proposes an approximate method for computing the size of the shadows of Schwarzschild BHs in an expanding FLRW universe as seen by a comoving observer. The key observation made in [@Bisnovatyi-Kogan:2018vxl] (see also [@Bisnovatyi-Kogan:2019wdd]) is that for BHs located at cosmological distances (*i.e.* well within the Hubble flow) the observer is typically very far from the BH event horizon, and the expansion of the universe is slow enough that it can be neglected near the BH. Within these approximations, one can compute the size of the BH shadow in a FLRW universe with arbitrary energy content, by first neglecting the expansion of the universe as light rays propagate near the BH, and then neglecting the strong BH gravity as light rays propagate towards the distant observer. Under these approximations, which are most certainly reasonable for SMBHs located at cosmological distances (but not applicable to SMBHs situated in the local universe, such as M87\*), the expression for the angular size of a Schwarzschild BH shadow at redshift $z$ is [@Bisnovatyi-Kogan:2018vxl]: $$\begin{aligned}
\alpha_{\rm sh}(z) \simeq \frac{3\sqrt{3}M}{D_A(z)}\,.
\label{eq:ash}\end{aligned}$$ Here, $D_A(z)$ is the angular diameter distance to redshift $z$, which depends on the energy content of the universe (photons, baryons, dark matter, dark energy, and neutrinos) as a function of time through the Hubble expansion rate $H(z)$ at redshift $z$, as: $$\begin{aligned}
D_A(z) = \frac{1}{(1+z)}\int_{0}^{z}\frac{dz'}{H(z')}\,.
\label{eq:da}\end{aligned}$$ For sufficiently small $z \ll 1$ one finds that $D_A(z) \approx z/H_0$, with $H_0$ the Hubble constant (but in this regime neglecting the strong BH gravity might not be justified, and peculiar velocities become important).
The expression in Eq. reflects the fact that the shadow of a Schwarzschild BH is an object of known intrinsic physical size, so that the influence of gravity on the propagation of photons can be neglected. The apparent angular size of the BH is related to its intrinsic physical size through the angular diameter distance at redshift $z$. It is worth remaking once more that this approximation is valid only for observers sufficiently far from the BH [@Bisnovatyi-Kogan:2018vxl]. Within this regime, the validity of Eq. (\[eq:ash\]) has been checked in [@Bisnovatyi-Kogan:2018vxl] against the full computation performed in [@Perlick:2018iye].
Interestingly, given the well-known fact that in an Universe with a cosmological constant the angular diameter distance $D_A(z)$ reaches a maximum at $z_{\max} \approx 1.5$ before continuously decreasing, the angular size of SMBH shadows dramatically increases for $z \gg z_{\rm max}$. We show this in Fig. \[fig:alphash\], where we plot the angular size of SMBH shadows as a function of redshift, for various values of the SMBH mass as reported in the caption. The angular size is reported in $\mu$as, and we consider SMBHs with masses up to $10^{11}M_{\odot}$, with heavier SMBHs leading to larger shadows as is obvious from Eq. (\[eq:ash\]). The heaviest SMBHs known to us fall just short of the $10^{11}M_{\odot}$ threshold. For example, TON618 is the heaviest SMBH known, and weighs about $6.6 \times 10^{10}M_{\odot}$ [@Shemmer:2004ph]. The next-to-heaviest SMBHs known are Homberg 15A [@2019ApJ...887..195M], IC 1101 [@2017MNRAS.471.2321D], and S5 0014+81 [@2009MNRAS.399L..24G; @2010MNRAS.405..387G], all with masses $\approx 4 \times 10^{10}M_{\odot}$. Therefore, $10^{11}M_{\odot}$ can be considered to be a loose more-than-optimistic upper limit for the heaviest SMBHs existing in Nature, and by extension the red curve in Fig. \[fig:alphash\] gives a rough upper limit to the size of how large the size of a SMBH shadow can be at any given redshift.
![Angular size of supermassive black hole shadows $\alpha_{\rm sh}$ (in $\mu$as) as a function of redshift $z$, given by Eq. (\[eq:ash\]). The angular sizes are computed for various SMBH masses: $M=10^7\,M_{\odot}$ (magenta curve), $10^8\,M_{\odot}$ (yellow curve), $10^9\,M_{\odot}$ (green curve), $10^{10}\,M_{\odot}$ (blue curve), and $10^{11}\,M_{\odot}$ (red curve). The horizontal black dashed line denotes the angular resolution of $1\,\mu{\rm as}$, which approximately corresponds to the current sensitivity of the Event Horizon Telescope. We have furthermore also included a very optimistic forecast sensitivity for an angular resolution of $0.1\,\mu{\rm as}$ (gray dotted line). $10^{11} M_{\odot}$ represents a more-than-optimistic rough optimistic upper limit to the heaviest SMBH existing in Nature, and hence the red curve gives a rough upper limit to the size of a SMBH shadow one might ever hope to observe at any given redshift.[]{data-label="fig:alphash"}](alpha_z.pdf){width="50.00000%"}
The proposal put forward by Tsupko *et al.* in [@Tsupko:2019pzg] is that an independent determination of the mass $M$ and redshift $z$ of SMBHs at cosmological distances, whose shadow angular size has been measured, leads to an indirect measurement of the angular diameter distance $D_A(z)$ through Eq. . As we see from Fig. \[fig:alphash\], given that the current angular resolution of the EHT is of ${\cal O}(\mu{\rm as})$, such a technique could in principle allow us to probe the distance-redshift relation within the redshift window $z \gtrsim \mathcal{O}(10)$ for SMBHs with masses $M \gtrsim 10^{10}\,M_{\odot}$, corresponding to the heaviest SMBHs known. This redshift window is extremely intriguing, and is well beyond the region that is currently accessible by conventional distance ladder methods, for example by the use of Supernovae Type Ia (SNeIa) or BAOs, which in the most optimistic cases can reach redshifts $z \lesssim 2-3$. Future 21-cm measurements might instead probe the same redshift window as SMBH shadows (see e.g. [@Sprenger:2018tdb; @Brinckmann:2018owf; @Munoz:2019fkt; @Munoz:2019rhi; @Munoz:2019hjh]). As shown in [@Qi:2019zdk], assuming that SMBH shadows could be used as a standard ruler, a combination of SMBH shadows and SNeIa measurements would lead to exquisite constraints on $\Omega_m$ and $H_0$ [@Qi:2019zdk]. In the next Section, we will advocate a more cautious approach towards the problem, highlighting a number of issues which were overlooked in the original proposal of Tsupko *et al.* [@Tsupko:2019pzg], and which appear to prevent SMBH shadows from becoming, at least at present, a reliable standard ruler.
Issues with the use of black hole shadows as standard rulers {#sec:issue}
============================================================
In this Section, we discuss in more detail the difficulties which have been overlooked on the road towards using SMBH shadows as standard rulers.
Reliably determining black hole masses {#subsec:mass}
--------------------------------------
Assuming that Eq. (\[eq:ash\]) is valid (see the later Sec. \[subsec:modeldependence\] for concerns on the matter), it is clear that in order to obtain a reliable distance measurement, an equally precise determination of the mass of the SMBH in question is required. Ideally, independent determinations of the SMBH mass should agree between each other. Unfortunately, this is far from being the case even with current SMBH mass determinations.
For instance, aside from the EHT-based determination of M87\*’s mass, there are essentially two main ways to determine this quantity: either using stellar dynamics measurements (e.g. [@Gebhardt:2011yw]) or gas dynamics observations (e.g. [@Walsh:2013uua]). These two methods to determine M87\*’s mass disagree by about a factor of $2$, and a similar level of disagreement is present for most SMBH mass estimates at low redshift. There are preliminary indications that incorporating non-Keplerian components in the modelling of the gas orbits might solve this discrepancy [@Jeter:2018eoh], however the situation is extremely far from being settled. Overall, it is clear that current SMBH mass determinations come with a significant ($\gtrsim 100\%$) systematic uncertainty budget, which directly translates into an equally large uncertainty budget on the inferred distance if SMBH shadows are used as standard rulers. It is impossible to do precision cosmology with such a large systematic uncertainty budget floating around.
Another possibility, especially useful at high redshifts, is reverberation mapping [@1982ApJ...255..419B; @2001sac..conf....3P]. However, present uncertainties obtained through this method are huge, again $\gtrsim 100\%$. Moreover, the uncertainty is dominated by systematics in our understanding of the so-called broad emission-line region form factor (see e.g. [@Denney:2008gk; @Shen:2013pea; @Campitiello:2019otf]). Until these broad emission-line regions are better understood, it will not be possible to improve this uncertainty budget, thus calling into question whether it will even be feasible to obtain precise measurements for SMBH masses at high redshift. The problem of reliably determining SMBH masses does not depend on the SMBH redshift, and we therefore expect it to be a significant limitation over the whole redshift window.
Reaching the required angular sensitivity {#subsec:sensitivity}
-----------------------------------------
From Fig. \[fig:alphash\], we see that in order to realistically resolve high-redshift SMBH shadows, a better than $0.1\,\mu{\rm as}$ angular resolution is required. Note that the red curve in Fig. \[fig:alphash\] is very optimistic, since we do not know of any SMBH as heavy as $10^{11}M_{\odot}$, whereas only a handful of SMBHs with masses of order $10^{10}M_{\odot}$ are known. Most known SMBHs have masses of order $10^9M_{\odot}$ (for instance, M87\* has a mass of about $6.5 \times 10^9M_{\odot}$).
An angular resolution of better than $0.1\,\mu{\rm as}$ requires an improvement of over an order of magnitude compared to the current angular resolution of the EHT. While the EHT (as well as planned surveys/space observatories) do plan to improve their sensitivity by both including multiple space-based telescopes, as well as moving to different frequencies, even the most optimistic setup does not seem to be able to achieve the required sensitivity of $0.1\,\mu{\rm as}$ or better (see e.g. [@Kardashev:2015xua; @2019arXiv190309539F; @2019ApJ...881...62P]). While we cannot exclude that future VLBI technology will be able to reach such a sensitivity, this target appears very futuristic at present.
In [@Tsupko:2019pzg], it was suggested that the target resolution might be reached by using VLBI technology in the optical band (recall that the EHT is currently observing at $1.3\,{\rm mm}$). However, there are reasons to be skeptical about high-redshift optical VLBI. In fact, the presence of dust in galactic nuclei strongly limits the capabilities of optical observations, which thus do not appear to be a plausible solution to the issue of increasing the angular sensitivity.
A perhaps more plausible alternative is that of using X-ray interferometry (XRI) techniques which, employing a constellation of satellites, may reach the necessary resolution in a relatively distant future [@Uttley:2019ngm] (the build and launch of constellation sub-$\mu$as XRI facilities can be expected indicatively no earlier than 2060). However, XRI facilities aim at observing the direct image of SMHBs with optically thick disks, in which case the shadow does not correspond to the apparent image of the photon sphere, but to the inner edge of the accretion disk, which should also strongly depend on the black hole spin parameter (see further discussions below in Sec. \[subsec:highredshift\]). More generally, the issue of what is the most appropriate electromagnetic wavelength to use is closely related to the emission mechanisms of the accreting material, which are far from being well understood, as we will discuss below in Sec. \[subsec:highredshift\]. In addition, XRI projects are expected to be able to image the shadows of SMBHs located near us, not at cosmological distances. In summary, the issue of reaching a sensitivity of $0.1\,\mu{\rm as}$ or better appears to be a severe limitation for most of the redshift range under consideration, unless a substantial population of high-redshift SMBHs with masses $>10^{10}M_{\odot}$ exists and can be observed.
Do we understand high-redshift black holes well enough? {#subsec:highredshift}
-------------------------------------------------------
Another possible concern is that the key expression for $\alpha_{\rm sh}(z)$, Eq. (\[eq:ash\]), might be modified in the presence of accretion flow which inevitably surrounds the SMBH. One might in fact worry that the observed size of the shadow would depend strongly on the shape and inclination of the accretion disk. More generally, the observed shadow might depend on the details of the accretion flow themselves (in fact, such a concern was recently raised in [@Gralla:2019xty], see also a partial response in [@Narayan:2019imo]), making SMBH shadows unsuitable for cosmological studies unless the accretion details were sufficiently understood. Fortunately, it is known that for advection dominated accretion flow (ADAF) [@Narayan:1994et; @Narayan:1994is; @Yuan:2014gma], the BH shadow is indeed the apparent image of the photon sphere, whose size is thus insensitive to the details of the accretion flow (see e.g. [@Narayan:2019imo]). The ADAF model is believed to be a valid description of the accretion flow around M87\* and SgrA\*, and in fact for several low-redshift SMBHs.
Is this still the case at high redshift? Unfortunately, things appear to be significantly more complicated. In fact, observations of SMBHs at redshifts as high as $z \sim 7-8$ (see e.g. [@Mortlock:2011va; @DeRosa:2013iia; @2015Natur.518..512W]) suggest that objects as massive as $M \approx \left ( 10^{9}-10^{10} \right ) \,M_{\odot}$ were in place less than $1\,{\rm Gyr}$ after the Big Bang [@Fan:2005eq]. This challenges the conventional picture of SMBH growth [@Volonteri:2010wz], which would require significantly longer timescales to build up so massive objects. It is not clear what the solution to this conundrum is, although a possibility very seriously considered in the literature is that the process of accretion around SMBHs at high redshift is significantly modified (see e.g. [@Volonteri:2005fj; @Madau:2014pta; @Alexander:2014noa]). In several of the scenarios advocated to explain the anomalously large population of high-redshift SMBHs, the details of the accretion flow are substantially different from the standard ADAF scenario, see for instance [@Haehnelt:1997js; @Nulsen:1999mt; @Barausse:2012fy; @Volonteri:2014lja; @Pacucci:2015efa; @Pacucci:2015rwa; @Pacucci:2015wea]. This implies that the resulting shadows of high-redshift SMBHs might be significantly affected by the details of the accretion flow, making them unsuitable for cosmological studies until the details of accretion onto high-redshift SMBHs is better understood.
On completely general grounds, one would in fact expect much higher accretion rates around high-redshift SMBHs, which would lead to an optically thick accretion flow. In this case, we expect the shadow to corresponds to the apparent image of the inner edge of the accretion disk, ranging from the innermost stable circular orbit (ISCO) for sources accreting at $\sim 10$% of the Eddington limit to the marginally bound orbit near the Eddington limit, and this clearly modifies Eq. (\[eq:ash\]). While this can in principle be imaged by XRI as we discussed in Sec. \[subsec:sensitivity\], the main issue is that the resulting angular size is extremely sensitive to both the SMBH spin and its inclination angle, and can vary by up to a factor of $\approx 10$. Thus, a reliable use of the angular sizes of high-redshift SMBHs with optically thick accretion flow requires a simultaneous precise measurement of both the BH spin and inclination angle, which appears to be extremely challenging at present.
Overall, it is more than fair to state that there is yet no general consensus regarding the formation and accretion dynamics of high-redshift SMBHs. This is of course a very active field of research, and there is all the reason to hope that improvements in future surveys will shed significantly more light on these issues (see for instance [@Paliya:2019oyn]). Only once the picture becomes clearer may we seriously start investigating realistic shadows of SMBHs at high-redshift (to the best of our knowledge, no such study exists in the literature). This issue makes it very premature to even consider using the shadows of SMBHs at redshift $z \gtrsim 7$ (even assuming they can be detected).
Weak lensing {#subsec:weaklensing}
------------
Another potentially important issue is that of weak lensing (WL). Recall that WL is the deflection of photons by intervening matter along their path from a source to us, in the limit where the deflection only causes small modifications to the photon’s path but not visually striking phenomena such as multiple images (see for instance [@Bartelmann:1999yn; @Refregier:2003ct; @Mandelbaum:2017jpr] for reviews). Photons coming from SMBHs at high redshift will inevitably encounter several lenses (or, more precisely, gravitational potentials) along the line-of-sight to us. To understand whether WL is a concern in this case, we need to estimate the typical angular deflections of photons coming from high-redshift SMBHs, and compare these to the typical shadow angular sizes plotted in Fig. \[fig:alphash\].
Consider a photon belonging to the boundary of a high-redshift SMBH shadow travelling to us and encountering several gravitational potentials along its way, and let us focus on one chosen gravitational potential. Denoting by $\Psi_i$ the depth of the gravitational potential at the point of closest approach on the non-deflected path, General Relativity predicts that the photon will be deflected by an angle $\delta_i \sim 4\Psi_i$. Typical gravitational potentials have a depth of $\Psi_i \sim 2 \times 10^{-5}$, leading to a typical deflection angle of $\delta_i \sim 10^{-4}$. How many such potentials does a high-redshift photon encounter on its path to us? The typical comoving size of gravitational potentials is $\sim 300\,{\rm Mpc}$ (twice the BAO scale), whereas the comoving distance to $z \sim 10$ is $\chi \approx 10000\,{\rm Mpc}$. This leads us to expect that a typical photon from a high-redshift SMBH will encounter about 30 gravitational potentials along its path to us. Assuming uncorrelated potentials, this gives a total rms deflection angle of about $\delta_{\rm tot} \sim \sqrt{30}\delta_i \approx 5 \times 10^{-4}$, or approximately $2\,{\rm arcmin}$.
Therefore, we expect a typical photon coming from the boundary of a high-redshift SMBH shadow to experience WL deflections of ${\cal O}({\rm arcmin})$. On the other hand, we see from Fig. \[fig:alphash\] that even in the most optimistic cases the angular sizes of SMBH shadows are of ${\cal O}(\mu{\rm as})$, *i.e.* 8 orders of magnitude smaller than the typical WL deflections. [^1] This makes it clear that WL represents a severe limitation towards the use of SMBH shadows as standard rulers, as the typical WL-induced deflections are much larger than the typical angular sizes of high-redshift SMBH shadows. From Fig. \[fig:alphash\] we easily see that, in order to have a shadow with angular size $\gg {\cal O}(\mu{\rm as})$ at $z \sim 10$, it is necessary to consider SMBHs with masses $M \gg 10^{11}M_{\odot}$, which are not believed to exist in Nature.
One could object that, in the same way WL surveys have been able not only to extract the WL signal but also elevate it to a precision cosmological probe, one could remove the lensing contamination to the SMBH shadow signal. However, there is a key difference between the two pictures. In the case of WL surveys, it is thanks to a statistical approach making use of images of over a million galaxies that one is able to reconstruct the weak lensing power spectra at relatively high signal-to-noise. For the SMBH shadow scenario, we will at best only be able to detect a handful of sources, making such a statistical lensing reconstruction procedure completely unfeasible.
Alternatively, one could hope to use knowledge of the foreground weak lensing convergence field along the line-of-sight to a given SMBH as reconstructed by future weak lensing surveys to undo the effect of WL. This would essentially provide a direct measurement of the projected mass overdensity along the line-of-sight. However, for the reconstructed convergence field to be useful, its resolution should be better than ${\cal O}(\mu{\rm as})$, in order to exactly target the SMBH shadow line-of-sight. This target precision is way beyond the expectations of the next-generation (or even next-to-next-generation) weak lensing surveys. In conclusion, WL appears to be an extremely severe contaminant towards the use of SMBH shadows at $z \gtrsim 1$ as a cosmological probe, and prospects for undoing the WL effect at the precision required (${\cal O}(\mu{\rm as})$ or better) are basically inconceivable in the foreseeable future.
Model-dependence {#subsec:modeldependence}
----------------
One more potential concern regarding the use of SMBH shadows as standard rulers is the model-dependence of the shadow angular size, or more precisely the model-dependence of Eq. (\[eq:ash\]). In fact, a reliable standard ruler (or standard candle/siren/clock for that matter) should be as model-independent as possible, *i.e.* the interpretation of the resulting measurement should not depend (or only depend weakly) on the assumption of any specific model. The expression for $\alpha_{\rm sh}(z)$ in Eq. (\[eq:ash\]) is valid only for Schwarzschild (*i.e.* non-rotating) BHs in GR. The extension to rotating (Kerr) BHs in GR is in principle not too problematic, since for Kerr BHs the main difference with respect to Schwarzschild BHs is the fact that the shadow becomes less circular (particularly at high observation angles, see for instance Fig. 1 in [@Bambi:2019tjh]), whereas its angular size remains roughly unchanged (although it does shrink slightly). The main problem appears when one considers theories of gravity beyond GR, many of which have been invoked in the literature to address the issues of cosmic acceleration, cosmic inflation, or dark matter (see e.g. [@Hu:2007nk; @Boehmer:2007kx; @Saridakis:2009bv; @Clifton:2011jh; @Capozziello:2011et; @Chamseddine:2013kea; @Myrzakulov:2015qaa; @Cai:2015emx; @Rinaldi:2016oqp; @Sebastiani:2016ras; @Capozziello:2017rvz; @Nojiri:2017ncd; @Vagnozzi:2017ilo; @Dutta:2017fjw; @Casalino:2018tcd]). While the Kerr solution persists as a solution to various theories beyond GR, in several other well-motivated theories this is not the case. As a result, the sizes of the shadows of beyond-GR BHs can deviate from the $r_{\rm sh}=3\sqrt{3}M$ predicted from GR. Essentially the same problem can occur when moving beyond BHs and considering so-called “BH mimickers” (including for instance horizon-less compact objects).
The literature on the shadows of BHs beyond GR and BH mimickers is too vast to be summarized here. Important works in this direction (studying for instance BH shadows in theories such as Chern-Simons gravity, brane-world gravity, dilaton gravity, scalar-vector-tensor gravity, or Einstein-Born-Infeld gravity, and shadows of BH mimickers such as superspinars, gravastars, and so on) can be found in e.g. [@Bambi:2008jg; @Amarilla:2010zq; @Amarilla:2011fx; @Amarilla:2013sj; @Nedkova:2013msa; @Tinchev:2013nba; @Wei:2013kza; @Grenzebach:2014fha; @Papnoi:2014aaa; @Sakai:2014pga; @Wei:2015dua; @Moffat:2015kva; @Ghasemi-Nodehi:2015raa; @Atamurotov:2015xfa; @Cunha:2015yba; @Amir:2016cen; @Dastan:2016vhb; @Tretyakova:2016ale; @Mureika:2016efo; @Sharif:2016znp; @Alhamzawi:2017iyn; @Cunha:2016wzk; @Singh:2017vfr; @Tsukamoto:2017fxq; @Eiroa:2017uuq; @Kumar:2017vuh; @Hennigar:2018hza; @Vetsov:2018mld; @Shaikh:2018lcc; @Shaikh:2018kfv; @Mizuno:2018lxz; @Amir:2018pcu; @Ovgun:2018tua; @Ayzenberg:2018jip; @Okounkova:2018abo; @Wang:2018prk; @Haroon:2019new; @Kumar:2019ohr; @Ovgun:2019jdo; @Das:2019sty] (see for instance [@Amarilla:2015pgp] for a review). For many of the solutions studied, the size of the BH shadow can deviate appreciably from $3\sqrt{3}M$. Even in the highly idealized case where we are able to measure a SMBH mass to high accuracy (an issue which is in itself problematic as per our earlier discussion, see Sec. \[subsec:mass\]), if the true underlying model of gravity is such that the angular size of the SMBH shadow is not given by Eq. (\[eq:ash\]) but by the same equation rescaled by a factor of $\beta$, incorrectly interpreting the observed angular size as being that of a GR BH directly translates into a biased determination of the angular diameter distance by the same factor of $\beta$ (note that $\beta$ can be both $\gtrsim 1$ or $\lesssim 1$). Another potentially important concern is that, at least for certain models of dark matter, the details of the dark matter halo by which BHs are surrounded could significantly affect the size of the shadow, as shown in a few studies (see e.g. [@Jusufi:2019nrn; @Haroon:2019new; @Xu:2018wow; @Hou:2018bar; @Haroon:2018ryd; @Hou:2018avu; @Konoplya:2019sns; @Jusufi:2019ltj]).
It should be remarked that in most alternative theories the BH shadow size does not deviate too much from the GR predictions. Theories where such a deviation is substantial ($\gtrsim 100\%$) are few and arguably more exotic. Still, this model-dependence underlying the use of SMBH shadows as standard rulers should be kept in mind, and is potentially an important concern. Assuming that we will be able to detect the shadows of high-redshift SMBHs, a possible way to address this concern would be to independently show that such SMBHs are indeed GR SMBHs. We leave open the question as to what would be the best way to do so.
Decrease in flux at high redshift {#subsec:flux}
---------------------------------
One final concern is that detecting the shadows of high-redshift SMBHs, despite their angular size increasing with respect to their low-redshift counterparts, might be more challenging than naïvely expected. In fact, the SMBH shadow angular size increasing at high redshift is not the only relevant factor. What’s perhaps more important is the fact that the observed flux decreases dramatically as $(1+z)^2$, *i.e.* a factor of ${\cal O}(100)$ at $z \sim 10$.
This effect could in principle be counteracted if the luminosity function of active galactic nuclei (AGNs, extremely luminous objects resulting from the accretion of matter onto SMBHs at the centers of galaxies) peaked at a higher luminosity as one moves up in redshift. However, the exact opposite occurs in reality, as one could expect [@Boyle:1988zz]. In fact, at high redshift the AGN luminosity function first peaks at $z \approx 1.5$ before declining rapidly [@Wyithe:2002ij; @Silverman:2007qa; @Fiore:2011iv; @Georgakakis:2015rfa; @Kulkarni:2018ebj]. Moreover, the bright-end slope also steepens, meaning that high-luminosity AGNs become increasingly rarer.
Together, the two effects \[evolution of the AGN luminosity function at high redshift, and flux decreasing as $(1+z)^2$\] conspire to seriously complicate the detection prospects of SMBHs at high redshift, in spite of the fact that their angular size increases at sufficiently high redshift. This difficulty should be taken into account in realistic forecasts for the use of SMBH shadows as standard rulers.
Conclusions {#sec:conclusions}
===========
In this short note, we have critically examined the intriguing proposal put forward by Tsupko *et al.* [@Tsupko:2019pzg] of using the shadows of high-redshift supermassive black holes as standard rulers. This is a very interesting proposal which rests upon the fact that for sufficiently high redshift in an expanding dark energy-dominated universe, the angular sizes of SMBH shadows increase with increasing redshift (see Fig. \[fig:alphash\]). If feasible, such a probe could potentially lead to exquisite constraints on cosmological parameters [@Qi:2019zdk], potentially also shedding light on the persisting $H_0$ tension.
Here, we have critically examined the feasibility of such a proposal, and found several limitations and concerns regarding the use of high-redshift SMBH shadows as a cosmological probe. These issues identified include: difficulties in obtaining reliable measurements of SMBH masses, currently limited by $>100\%$ systematics, and the determination of which is crucial for the proposal in question, see Eq. (\[eq:ash\]); reaching an angular sensitivity of $0.1\,\mu{\rm as}$ or better, which remains challenging even when considering alternative techniques such as X-ray interferometry; an insufficient knowledge of the accretion dynamics of high-redshift SMBHs, and consequently of our understanding of how the shadows of the latter should appear; the fact that weak lensing by gravitational potentials along the line-of-sight leads to typical ${\cal O}({\rm arcmin})$ deflections, which are several orders of magnitude larger than the typical angular sizes of high-redshift SMBHs and appear challenging to undo; the model-dependence of the key equation for the angular size of SMBH shadows at high redshift, Eq. (\[eq:ash\]), which can be modified if the underlying theory of gravity is not General Relativity; and finally the fact that the flux coming from high-redshift SMBHs decreases dramatically compared to their low-redshift counterparts.
In conclusion, we have found a number of critical issues which appear to undermine the very interesting possibility put forward by Tsupko *et al.* of using SMBH shadows as standard rulers [@Tsupko:2019pzg], at least at present. While we of course do not want to discourage astrophysicists and cosmologists from further considering this probe, we believe it is important to highlight any shortcomings thereof as early as possible, in order for these to be thoroughly addressed by the time the proposal will be experimentally feasible. We leave the issue of proposing possible solutions to the issues identified to future work.
S.V. acknowledges support from the Isaac Newton Trust and the Kavli Foundation through a Newton-Kavli Fellowship, and acknowledges a College Research Associateship at Homerton College, University of Cambridge. C.B. acknowledges support by the Innovation Program of the Shanghai Municipal Education Commission, Grant No. 2019-01-07-00-07-E00035, and the National Natural Science Foundation of China (NSFC), Grant No. 11973019. L.V. is supported through the research program “The Hidden Universe of Weakly Interacting Particles” with project number 680.92.18.03 (NWO Vrije Programma), which is partly financed by the Dutch Research Council.
[^1]: We are not focusing on low-redshift SMBHs, whose shadows can be larger, but which at the same time are not in the Hubble flow, implying that their use is limited by difficulties in estimating the host galaxy peculiar velocity.
| {
"pile_set_name": "ArXiv"
} |
---
address: |
Department of Mathematics, University of Illinois\
Urbana-Champaign, Illinois 61801, USA
author:
- 'M. J. Bergvelt'
title: '$H_D$-Quantum Vertex Algebras'
---
[1]{},
[1]{}
We discuss a class of quantum vertex algebras where not only the commutativity of vertex algebra is broken by a braiding map $S^{(\tau)}$, but also the translation covariance is broken by a translation map $S^{(\gamma)}$. The new class of quantum vertex operators satisfy a Braided Jacobi Identity containing both the braiding and the translation maps.
What is a Vertex Algebra?
=========================
According to Borcherds [@MR1653021], [@MR1865087] a **vertex algebra** can be thought of as a commutative, associative, unital *singular* algebra with infinitesimal translation symmetry. Usually this made precise as follows. We start with a *State Space* $V$ (a vector space) with a distinguished element 1 (called the *vacuum*, playing the r\^ ole of unit in $V$). $V$ has an action of the Hopf algebra $H_D=\mathbb{C}[D]$, where $D$ is the infinitesimal translation operator. Then for each $a\in V$ we have a (singular) operator $Y(a,z)$ of left multiplication by $a$: $$Y(a,z)\colon V\to V((z)).$$ We have the following axioms:
- **Vacuum:** $Y(1,z)=1_V$, and $Y(a,z)1=e^{ZD}a$.
- **Translation Covariance:** $[D,Y(a,z)]=Y(Da,z)=\partial_z Y(a,z)$.
- **Commutativity:** For all $a, b\in V$ there is an $N$ such that $$(z_1-z_2)^N[Y(a,z_1),Y(b,z_2)]=0.$$
In this formulation in the multiplication $Y(a,z)b$ the left factor $a$ is treated differently than the right factor $b$; for instance $a$ has the variable $z$ associated to it, but $b$ has no variable attached. We find it useful to reformulate the theory so that both factors are treated symmetrically. We attach to the two factors $a$ and $b$ two variables, $z_1,z_2$, and define a singular multiplication $$\begin{aligned}
X_{z_1,z_2}\colon V{\otimes}V&\to V[[z_1,z_2]][(z_1-z_2){^{-1}}],\\
a{\otimes}b&\to X_{z_1,z_2}(a{\otimes}b)=e^{z_2D}Y(a,z_1-z_2)b.
\end{aligned}$$ Then one derives from the axioms for $Y$ the following properties of $X_{z_1,z_2}$:
- **Vacuum:** $X_{z_1,z_2}(a{\otimes}1)=e^{z_1 D}a$, $X_{z_1,z_2}(1{\otimes}a)=e^{z_2 D}a$.
- **Translation Covariance:** $e^{\gamma D}X_{z_1,z_2}(a{\otimes}b)=
X_{z_1+\gamma,z_2+\gamma}(a{\otimes}b)$.
- **Commutativity:** $X_{z_1,z_2}(a{\otimes}b)=X_{z_2,z_1}(b{\otimes}a)$.
Relation of $X_{z_1,z_2}$ to $Y(a,z)$ is given by expansions. Let $i_{z_1;z_2}$ be the [expansion]{} of a rational function in $z_1-z_2$ in the region ${\lvertz_1\rvert}>{\lvertz_2\rvert}$. Then $$\begin{aligned}
i_{z_1;z_2}X_{z_1,z_2}(a{\otimes}b)&= Y(a,z_1)Y(b,z_2)1,\\
i_{z_2;z_1}X_{z_1,z_2}(a{\otimes}b)&= Y(b,z_2)Y(a, z_1)1,\\
i_{z_2;z_3}X_{z_2+z_3,z_2}(a{\otimes}b)&= Y(Y(a,z_3)b,z_2)1.
\end{aligned}\label{eq:expVertexOps}$$ From this one easily derives the axioms for vertex algebras in terms of $Y$, starting with the properties of $X_{z_1,z_2}$.
Quantum Vertex Algebras via Deformation.
========================================
As a motivation for our definition of a quantum vertex algebra consider a commutative algebra $M$ , with a multiplication $$m\colon M{\otimes}M\to M,$$ so that we have in particular $m(a{\otimes}b)=m(b{\otimes}a)$. Now a quantization of $(M,m)$ could be defined by introducing a formal variable $t$, and a deformed multiplication $$m_t\colon M_t{\otimes}M_t\to M_t,\quad M_t=M[[t]],$$ where $m=m_t\mod t$. In general $m_t$ will not longer be commutative. We could require that there is a *Braiding Map*: $$S\colon M_t{\otimes}M_t\to M_t{\otimes}M_t,$$ such that $m_t \circ S(a{\otimes}b)= m_t(b{\otimes}a)$. So the braiding describes the failing of $m_t$ to be commutative. Now if the commutative algebra $M$ has some symmetry, say via a group $G$ acting on $M$ (so that $m(ga{\otimes}gb)=gm(a{\otimes}b)$), then we can expect that after quantisation the deformed multiplication is no longer $G$-symmetric. Instead, we can require that there is for each $g\in G$ a map $$S^g\colon M_t{\otimes}M_t \to M_t{\otimes}M_t,$$ such that $$g m_t\circ S^g(a{\otimes}b)=m_t(ga{\otimes}gb).$$ So a deformation of a commutative algebra with symmetry $(M,m,G)$ would be a quintuple $(M_t,m_t,S, S^g, G)$, satisfying a complicated system of axioms we don’t want to write down here, see [@AngBergv].
Now a vertex algebra is a singular analog of a commutative algebra with as symmetry the Hopf algebra $H_D$ of infinitesimal translations. So if we quantize we can expect that the commutativity and translation covariance are no longer exact, and that we need extra structures to describe the broken symmetries.
Introduce a quantum variable $t$ and deformed singular multiplication $$\begin{aligned}
X_{z_1,z_2}\colon V^{{\otimes}2}&\to V[[z_1,z_2]][z_1{^{-1}},(z_1-z_2){^{-1}}][[t]].
\end{aligned}$$ The commutativity and translation covariance are supposed to be not longer exact. The extra structure we need is
- a *Braiding* map: $$S^{(\tau)}_{z_1,z_2}\colon V^{{\otimes}2}\to
V^{{\otimes}2}[z_1^{\pm 1},z_2^{\pm 1},(z_1-z_2){^{-1}}][[t]].$$
- a *Translation* map: $$S^{(\gamma)}_{z_1,z_2}\colon
V^{{\otimes}2}\to V^{{\otimes}2}[z_1^{\pm
1},z_2,(z_1-z_2){^{-1}},(z_1+\gamma)^{\pm1}, z_2+\gamma][[t]].$$
The deformed multiplication and the braiding and translation maps are supposed to have the following properties:
- Braided Commutativity: $$X_{z_1,z_2} (a{\otimes}b)=X_{z_2,z_1}\circ S^{(\tau)}_{z_2,z_1}(b{\otimes}a).$$
- (Broken) Translation covariance: $$e^{\gamma D}X_{z_1,z_2}\circ S^{(\gamma)}_{z_1,z_2}(a{\otimes}b)=X_{z_1+\gamma,z_2+\gamma}(a{\otimes}b).$$
- Plus a bunch of other axioms (Yang-Baxter, hexagon, …), see [@AngBergv])
This defines a **$H_D$-quantum vertex algebra**.
Vertex Operators and the Braided Jacobi Identity.
=================================================
We define a 1-variable vertex operator as usual: $$Y(a,z)b=X_{z,0}(a{\otimes}b),$$ i.e., by evaluating the second variable of $X$ at zero. (Note that in general in an $H_D$-quantum vertex algebra we can not evaluate the *first* variable at 0.) Then the relation between $Y$ and $X$ is given by a variant of , where we need to insert braiding and translation matrices in appropriate places. More generally, one shows that there exists $$X_{z_1,z_2,z_3}\colon V^{{\otimes}3}\to
V[[z_k]][z_i{^{-1}},(z_i-z_j){^{-1}}][[t]],\quad {1\le i<j \le 3}, i\le k\le 3$$ such that $$i_{z_1;z_2,z_3}X_{z_1,z_2,z_3}=X_{z_1,0}(1{\otimes}X_{z_2,z_3}).$$ Then we have the following expansions: if $A=a{\otimes}b{\otimes}c$, then $$\begin{aligned}
i_{z_1;z_2}X_{z_1,z_2,0}(A)&=Y(a,z_1)Y(b,z_2)c, \\
i_{z_2;z_1}X_{z_1,z_2,0}(A)&=Y_{z_2}(1\otimes Y_{z_1})
i_{z_2;z_1}S^{(\tau), 12}_{z_2,z_1}
(b\otimes a\otimes c),\\
i_{z_2;z_3}X_{z_2+z_3,z_2,0}(A)&=Y_{z_2}(Y_{z_3}\otimes 1)
i_{z_2;z_3}S^{(z_2),12}_{z_3,0} (a\otimes b\otimes
c).
\end{aligned}\label{eq:ExpansQVertex}$$ Here we write $Y_z(a{\otimes}b)$ for $Y(a,z)b$. From these expansions one derives the *Braided Jacobi Identity* for $H_D$-quantum vertex algebras: $$\begin{gathered}
i_{z_1;z_2}\delta(z_1-z_2,z_3)a(z_1)b(z_2)c-
i_{z_2;z_1}\delta(z_1-z_2,z_3)Y_{z_2}(1\otimes Y_{z_1})
S^{(\tau),12}_{z_2,z_1}(b\otimes a\otimes c)\\
=i_{z_2;z_3}\delta(z_1,z_2+z_3)Y_{z_2}(Y_{z_3}\otimes
1)S^{(z_2),12}_{z_3,0}(a\otimes b\otimes c).
\end{gathered}$$ Here we write $a(z)$ for $Y(a,z)$.
Example: Hall-Littlewood polynomials.
=====================================
The main inspiration for our construction came from the theory of quantum vertex operators for Hall-Littlewood polynomials introduced by Jing, [@MR1112626]. These quantum vertex operators occur naturally in a $H_D$-quantum vertex algebra $V_{L,t}$ which is a deformation of the lattice vertex algebra based on the lattice $L=\mathbb{Z}\alpha$, with pairing $L{\otimes}L\to \mathbb{Z}$, $a\alpha{\otimes}b\alpha\mapsto
ab$. $V_{L,t}$ is in a sense generated by $e^\alpha$ and the braiding and translation maps in this case are given by $$S^{(\tau)}_{z_1,z_2}(e^{\alpha}\otimes
e^{\alpha})=-\frac{1-tz_2/z_1}{1-tz_1/z_2}e^\alpha{\otimes}e^\alpha,\quad
S^{(\gamma)}_{z_1,z_2}(e^{\alpha}\otimes
e^{\alpha})=\frac{1-tz_2/z_1}{1-t\frac{z_2+\gamma}{z_1+\gamma}}e^\alpha{\otimes}e^\alpha,$$ An effective method to do calculations in $V_{L,t}$ and similar $H_D$-quantum vertex algebras is given by the theory of bicharacters, see [@Anguelova:thesis].
Conclusion and Outlook.
=======================
The $H_D$-quantum vertex algebras introduced above are generalizations of the quantum vertex operators of Etingof-Kazhdan [@MR2002i:17022]. In their theory the vertex operators are translation covariant, so that the translation maps $S^{(\gamma)}$ are the identity.
Vertex algebras are commutative algebras with translation covariance and singularities in the product of vertex operators of the form $(z_1-z_2)^{-N}$. In our $H_D$-quantum vertex algebras the commutativity and translation covariance is broken via nontrivial braiding and translation maps $S^{(\tau)}$ and $S^{(\gamma)}$, but the singularities are essentially still of the same type $(z_1-z_2)^{-N}$. Now in examples of quantum vertex operators, see e.g., [@q-alg/9706023], one sees that in practice the product can have singularities of the form $$\frac{1}{z_1-p^kq^\ell z_2}.$$ This means that in such quantum vertex algebras one needs to *extend* the symmetry algebra $H_D=\mathbb{C}[D]$ by adding (group like) operators $T_p, T_q$ that act like $$T_p f(z)=f(pz),\quad T_q f(z)=f(qz).$$ Note that $H_{p,q}=\mathbb{C}[T_p^{\pm 1},T^{\pm 1}_q,d]$ is non commutative.
Now the basic formalism of vertex algebras (and $H_D$-quantum vertex algebras) is very much based on $H_D$. For instance, the delta distribution $\delta(z_1,z_2)$, which is ubiquitous in the theory, is the difference of two expansions of the basic singularity $\frac
1{z_1-z_2}$, and expansions are given by the *exponential operators* canonically associated to $H_D$.
Replacing the commutative Hopf algebras $H_D$ by the non commutative $H_{p,q}$ changes the basic framework of vertex algebras drastically: for instance, one needs a new theory of Dirac delta distributions adapted to $H_{p,q}$, [@AngBergvChiral].
Acknowledgments {#acknowledgments .unnumbered}
===============
The results in this paper are joint work with Iana Anguelova, see [@AngBergv].
\[2\][ [\#2](http://www.ams.org/mathscinet-getitem?mr=#1) ]{} \[2\][\#2]{}
[Ang06]{}
I. I. Anguelova and M. J. Bergvelt, *[$H_D$]{}-quantum vertex algebras and bicharacters*, preprint, arXiv:math.QA/0706.1528.
, *[$H_t$]{}-quantum vertex algebras and deformed chiral algebras*, in preparation.
I. I. Anguelova, *Bicharacter constructions of quantum vertex algebras*, Ph.D. thesis, University of Illinois, Urbana-Champaign, 2006.
Richard E. Borcherds, *Vertex algebras*, Topological field theory, primitive forms and related topics (Kyoto, 1996), Progr. Math., vol. 160, Birkhäuser Boston, Boston, MA, 1998, pp. 35–77.
, *Quantum vertex algebras*, Taniguchi Conference on Mathematics Nara ’98, Adv. Stud. Pure Math., vol. 31, Math. Soc. Japan, Tokyo, 2001, pp. 51–74.
Pavel Etingof and David Kazhdan, *Quantization of [L]{}ie bialgebras. [V]{}. [Q]{}uantum vertex operator algebras*, Selecta Math. (N.S.) **6** (2000), no. 1, 105–130.
Edward Frenkel and Nikolai Reshetikhin, *[Towards Deformed Chiral Algebras]{}*, Proceedings of the Quantum Group Symposium at the XXIth International Colloquium on Group Theoretical Methods in Physics, Goslar 1996, 1997, arXiv:q-alg/9706023.
Nai Huan Jing, *Vertex operators and [H]{}all-[L]{}ittlewood symmetric functions*, Adv. Math. **87** (1991), no. 2, 226–248.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the nonlinear observability of a system’s states in view of *how well* they are observable and what control inputs would improve the convergence of their estimates. We use these insights to develop an observability-aware trajectory-optimization framework for nonlinear systems that produces trajectories well suited for self-calibration. Common trajectory-planning algorithms tend to generate motions that lead to an unobservable subspace of the system state, causing suboptimal state estimation. We address this problem with a method that reasons about the quality of observability while respecting system dynamics and motion constraints to yield the optimal trajectory for rapid convergence of the self-calibration states (or other user-chosen states). Experiments performed on a simulated quadrotor system with a GPS-IMU sensor suite demonstrate the benefits of the optimized observability-aware trajectories when compared to a covariance-based approach and multiple heuristic approaches. Our method is $\thicksim$80x faster than the covariance-based approach and achieves better results than any other approach in the self-calibration task. We applied our method to a waypoint navigation task and achieved a $\thicksim$2x improvement in the integrated RMSE of the global position estimates and $\thicksim$4x improvement in the integrated RMSE of the GPS-IMU transformation estimates compared to a minimal-energy trajectory planner.'
author:
-
-
bibliography:
- 'references.bib'
title: 'Observability-Aware Trajectory Optimization for Self-Calibration with Application to UAVs'
---
Introduction {#sec:intro}
============
State estimation is a core capability for autonomous robots. For any system, it is desirable to estimate the state at any point in time as accurately as possible. Accurate state estimation is crucial for robust control strategies and serves as the foundation for higher-level planning and perception. In addition to the states directly used for system control, such as position, velocity, and attitude, more recent work also estimates internal states that calibrate the sensor suite of the system [@ROB:ROB21466]. These so-called *self-calibration states* include all information needed to calibrate the sensors against each other – such as the position and attitude of one sensor with respect to another – as well as their intrinsic parameters such as measurement bias.
In general, these states could be estimated *a priori* using offline calibration techniques. The advantages of including self-calibration states in the *online* state estimator are threefold: i) the same implementation of the self-calibrating estimator can be used for different vehicles ii) the state estimator can compensate for errors in the initialization or after collisions and iii) the platform does not require any offline calibration routine because it can self-calibrate itself while operating.
![Root Mean Squared Error (RMSE) convergence of EKF self-calibration (accelerometer and gyroscope biases $\bb_a$, $\bb_\omega$ and position of the GPS sensor $\bp^p_i$) state estimates for a) figure-eight trajectory, b) star trajectory, c) optimal trajectory from our method. We introduced additional yaw motion for a) and b) trajectories in order to improve state estimation of these heuristics.[]{data-label="fig:cover"}](cover_transpose_pretty.eps){width="0.99\columnwidth"}
Including self-calibration states in the estimator has many advantages, but it comes at an important cost: the dimensionality of the state vector increases while the number of measurements remains unchanged. This cost often leads to the requirement of engineered system inputs to render all states observable, i.e. the system needs a tailored trajectory that might require extra time or energy compared to an observability-unaware trajectory. This requirement holds true not only for the initial self-calibration but also throughout the mission. Usually, the self-calibration motion is executed by an expert operator who controls the vehicle and continuously checks if the states have converged to reasonable values. During the mission, the expert operator must take care to excite the system sufficiently to keep all states observable. More importantly, in autonomous missions, trajectory-planning algorithms that minimize energy use may generate trajectories that lead to an unobservable subspace of the system state.
In this paper, we present a framework that optimizes trajectories for self-calibration. The resulting trajectories avoid unobservable subspaces of the system state during the mission. We develop a cost function that explicitly addresses the quality of observability of system states. Our method takes into account motion constraints and yields an optimal trajectory for fast convergence of the self-calibration states or any other user-chosen states. The presented theory applies to any (non)linear system and it is not specific to a particular realization of the state estimator such as KF, EKF, or UKF. While past approaches have focused on analyzing the environment to compute *where* to move to obtain informative measurements for state estimation [@bry11icra], we assume the presence of accurate measurements[^1] and focus on *how* to move to generate motions that render the full state space observable.
In order to evaluate our method, we conduct several experiments using a simulated Unmanned Aerial Vehicle (UAV) with GPS-IMU sensor suite, as this is a common scenario that illuminates the problem of self-calibration. An example of the performance of the self-calibration framework is presented in , where an optimized trajectory outperforms common calibration heuristics usually chosen by experts in terms of speed and accuracy of the state convergence.
The key contributions of our approach are:
- we present a method that is able to predict the quality of state estimation based on the vehicle’s ego-motion rather than on the perceived environment; the method takes into account system dynamics, measurements and the nonlinear observability analysis.
- Our method is carried out on the nonlinear continuous system without making any state-estimator-specific assumptions.
- We demonstrate a full self-calibration-based trajectory optimization framework that is readily adjustable for any dynamical system and any set of states of the system.
- We show that the observability-aware trajectory optimization can be also used for the waypoint navigation task which results in more accurate state estimation.
Related Work {#sec:rw}
============
Previous work on improving state estimation of a system has mainly focused on analyzing the environment to choose informative measurements [@bryson04acra; @julian2012distributed]. With the arrival of robust visual-inertial navigation solutions (e.g. Google Tango[^2]), sufficiently accurate measurements can be obtained most of the time during a mission without special path reasoning. These results give us the opportunity to shift focus towards other aspects of the trajectory, in particular, its suitability for self-calibration. In [@Maye06102015], the authors find the best set of measurements from a given trajectory to calibrate the system. Unobservable parameters are locked until the trajectory has sufficient information to make them observable. The analysis is performed on the linearized system and analyzes a given trajectory rather than generating an optimal convergence trajectory. @s150203154 analyzed a specific marine system and developed a trajectory to calibrate it based on heuristics. The approach is not generally applicable to other systems. Other approaches analyze the final covariance of the system when simulating it on a test trajectory: @4177715 maximize the inverse of the covariance at the final time step and use this cost in an optimization procedure. Similarly, in [@achtelik13icra] the authors sample a subset of the state space with a Rapidly Exploring Random Tree approach [@bry11icra] and optimize for the final covariance of the system. These approaches are sample-based techniques discretizing their environment and state space. The discretization and linearization steps induce additional errors and may lead to wrong results similar to the well-known rank issue when analyzing a system in its linearized instead of the nonlinear form [@Hesch13112013].
From the nonlinear observability analysis described in [@hermann1977nonlinear], @krener2009measures develop a measure of observability rather than only extracting binary information on the observability of a state. @6580494 make use of this measure to generate trajectories that optimize the convergence of states that are directly visible in the sensor model. We make use of this definition and extend the approach to analyze the quality of observability of states that are not directly visible in the sensor model. This way, we can also generate motions leading to trajectories that optimize the convergence of, e.g., IMU biases.
Trajectory optimization has been successfully used in many different applications including robust perching for fixed-wing vehicles [@moore2014robust], locomotion for humanoids [@kuindersma2015optimization] and manipulation tasks for robotic arms [@levine2015learning]. Since we evaluate our system using a model of an Unmanned Aerial Vehicle (UAV), we present the related work on the trajectory optimization in this area. @mellinger2011minimum use trajectory optimization of differentially flat variables of a quadrotor to obtain minimum-snap trajectories. @richter2013polynomial presented an extension of this approach to generate fast quadrotor paths in cluttered environments using an unconstrained QP. In [@mueller14iros], the authors generate risk-aware trajectories using the formulation developed by @van2011lqg with a goal of safe quadrotor landing. @hausman2015cooperative use the framework of trajectory optimization to generate controls for multiple quadrotors to track a mobile target. Finally, @moore2014robust use LQR-trees to optimize for a trajectory that leads to robust perching for a fixed-wing vehicle. In this work, we use trajectory optimization to generate paths for self-calibration, which are evaluated on a simulated quadrotor system. Our work augments other trajectory-optimization-based approaches by providing an observability-aware cost function that can be used in combination with other optimization objectives.
Problem Formulation and Fundamentals
====================================
Motion and Sensor Models
------------------------
We assume the following nonlinear system dynamics, i.e. the motion model: $$\begin{aligned}
\dot{\bx} &= f(\bx,\bu,\mathbf{\delta}),
$$ where $\bx$ is the state, $\bu$ are the control inputs and $\mathbf{\delta}$ is noise caused by non-perfect actuators and modelling errors.
For sensory output we use the nonlinear sensor model: $$\begin{aligned}
\bz &= h(\bx, \mathbf{\epsilon}),
$$ where $\bz$ is the sensor reading and $\mathbf{\epsilon}$ is the sensor noise.
It is often the case that there are certain elements in the state vector $\bx$, such as sensor biases, that stay constant according to the motion model, i.e. their values are independent of the controls $\bu$ and other states $\bx$. In this paper, we will call these entities self-calibration states $\bx_{sc}$.
Nonlinear Observability Analysis {#sec:nonlinear-observability-analysis}
--------------------------------
The observability of a system is defined as the possibility to compute the initial state of the system given the sequence of its inputs $\bu(t)$ and measurements $\bz(t)$. A system is said to be *globally observable* if there exist no two points $\bx_0(0)$, $\bx_1(0)$ in the state space with the same input-output $\bu(t)$-$\bz(t)$ maps for any control inputs. A system is *weakly locally observable* if there is no point $\bx_1(0)$ with the same input-output map in a neighborhood of $\bx_0(0)$ for a specific control input.
Observability of linear as well as nonlinear systems can be determined by performing a *rank test* where the system is observable if the rank of the observability matrix is equal to the number of states. In the case of a nonlinear system, the nonlinear observability matrix is constructed using the Lie derivatives of the sensor model $h$. Lie derivatives are defined recursively with zero-noise assumption. The 0-th Lie derivative is the sensor model itself, i.e.: $$\begin{aligned}
L_0^h = h(\bx),\end{aligned}$$ the next Lie derivative is constructed as: $$\begin{aligned}
L_{i+1}^h = \frac{\partial}{\partial t}{L}_{i}^h = \frac{\partial {L}_{i}^h}{\partial \mathbf{x}}\frac{\partial \mathbf{x}}{\partial t} = \frac{\partial {L}_{i}^h}{\partial \mathbf{x}}f(\mathbf{x},\mathbf{u}).\end{aligned}$$ One can observe that Lie derivatives with respect to the sensor model are equivalent to the respective time derivatives of the sensory output $\bz$: $$\begin{aligned}
\dot{\bz} = \frac{\partial}{\partial t}\bz(t) = \frac{\partial}{\partial t}h(\mathbf{x}(t)) = \frac{\partial h}{\partial \mathbf{x}}\frac{\partial \mathbf{x}}{\partial t} = \frac{\partial h}{\partial \mathbf{x}}f(\mathbf{x},\mathbf{u}) = {L}_{1}^h.\end{aligned}$$ By continuing to compute the respective Lie derivatives one can form the matrix: $$\begin{aligned}
O(\bx,\bu) &=
\begin{bmatrix}
\nabla L_0^h&\nabla {L}_{1}^h&\nabla {L}_{2}^h&\dots
\end{bmatrix}^T,\end{aligned}$$ where $\nabla L_0^h = \frac{\partial L_0^h}{\partial \bx}$ and $\dot{\bz}=\frac{\partial \bz}{\partial t}$.
The matrix $O(\bx,\bu)$ formed from the sensor model and its Lie derivatives is the nonlinear observability matrix. Following @hermann1977nonlinear, if the observability matrix has full column rank, then the state of the nonlinear system is weakly locally observable. Unlike linear systems, nonlinear observability is a local property that is input– and state-dependent.
It is worth noting that the observability of the system is a binary property and does not quantify how *well* observable the system is. This limits its utility for gradient-based methods. We address this issue in the next section.
Quality of Observability {#sec:quality}
========================
Following @krener2009measures and according to the definition presented in , we introduce the notion of quality of observability.
A state is said to be *well* observable if the system output changes significantly when the state is marginally perturbed [@WeissPhD2012]. A state with this property is robust to measurement noise and it is highly distinguishable within some proximity where this property holds. Conversely, a state that leads to a small change in the output, even though the state value was extensively perturbed, is defined as *poorly* observable. In the limit, the measurement does not change even if we move the state value through its full range. In this case, the state is unobservable [@hermann1977nonlinear].
Taylor Expansion of the Sensor Model
------------------------------------
In order to model the variation of the output in relation to a perturbation of the state, we approximate the sensor model using the $n$-th order Taylor expansion about a point $t_0$:
$$\begin{aligned}
h_{t_0}(\bx(t),\bu(t)) &= \sum_{i=0}^n \frac{(t-t_0)^i}{i!} h^i(\bx(t_0),\bu(t_0)),\end{aligned}$$
where $h_{t_0}$ represents the Taylor expansion of $h$ about $t_0$ with the following Taylor coefficients $h^0, h^1, \ldots, h^n$: $$\begin{aligned}
h^0(\bx(t_0),\bu(t_0)) &= h(\bx(t_0),\bu(t_0)) = L^h_0(\bx(t_0),\bu(t_0)) \\
h^1(\bx(t_0),\bu(t_0)) & = \frac{\partial}{\partial t} (h(\bx(t_0),\bu(t_0))) = L^h_1(\bx(t_0),\bu(t_0)) \\
& \vdots \\
h^n(\bx(t_0),\bu(t_0)) & = L^h_n(\bx(t_0),\bu(t_0)). \end{aligned}$$
Using this result, one can also approximate the state derivative of the sensor model $\frac{\partial}{\partial \bx}h(\bx(t),\bu(t))$. For brevity, we introduce the notation $\delta{t} = t-t_0$, $h_{t_0}(t) = h_{t_0}(\bx(t),\bu(t))$, and we omit the arguments of the Lie derivatives: $$\begin{aligned}
\frac{\partial}{\partial \bx}h_{t_0}(t) &= \sum_{i=0}^n
\frac{\delta{t}^i}{i!}\nabla L^h_i.
$$ This result in matrix form is: $$\begin{aligned}
\frac{\partial}{\partial \bx}h_{t_0}(t) &= \begin{bmatrix}I & \delta{t}I & \frac{\delta{t}^2}{2}I & \dots & \frac{\delta{t}^n}{n!}I\end{bmatrix} O(\bx(t),\bu(t)),
\label{eq:taylor-final}\end{aligned}$$ where $O(\bx(t),\bu(t))$ is the nonlinear observability matrix.
describes the Jacobian of the sensor model $h$ with respect to the state $\bx$ around the time $t_0$. Using this Jacobian, we are able to predict the change of the measurement with respect to a small perturbation of the state. This prediction not only incorporates the sensor model but it also implicitly models the dynamics of the system via high order Lie derivatives. Hence, in addition to showing the effect of the states that directly influence the measurement, also reveals the effects of the varying control inputs and the states that are not included in the sensor model. This will prove useful in .
Observability Gramian
---------------------
In addition to the change in the output with respect to the state perturbation, one needs to take into account the fact that different states can have different influence on the output. Thus, a large effect on the output caused by a small change in one state can swamp a similar effect on the output caused by a different state and therefore, *weaken* its observability. In order to model these interactions, following [@krener2009measures], we employ the local observability Gramian: $$\begin{aligned}
W_o({0},{T}) &= \int_{{0}}^{{T}} \Phi({0}, t)^T H(t)^T H(t) \Phi({0}, t) dt,
\label{eq:gramian}\end{aligned}$$ where $\Phi({0}, t)$ is the state transition matrix (see [@krener2009measures] for details), $H(t)$ is the Jacobian of the sensor model $H(t) = \frac{\partial}{\partial \bx}h(t)$ and the trajectory spans the time interval $t \in [0, T]$.
Since a nonlinear system can be approximated by a linear time-varying system by linearizing its dynamics about the current trajectory, one can also use the local observability Gramian for nonlinear observability analysis. If the rank of the local observability Gramian is equal to the number of states, the original nonlinear system is locally weakly observable [@hermann1977nonlinear].
@krener2009measures introduced measures of observability that are based on the condition number or the smallest singular value of the local observability Gramian. Unfortunately, the local observability Gramian is difficult to compute for any nonlinear system. In fact, it can only be computed in closed form for certain simple nonlinear systems. In order to solve this problem, the local observability Gramian can be approximated numerically by simulating the sensor model for small state perturbations, resulting in the *empirical local observability Gramian* [@krener2009measures]: $$\begin{aligned}
W_o \approx \frac{1}{4\epsilon^2} \int_{{0}}^{{T}} \begin{bmatrix} \Delta{\bz_1^T(t)}\\ \vdots \\\Delta{\bz_n^T(t)} \end{bmatrix} [ \Delta{\bz_1(t)} \, \dots \, \Delta{\bz_n(t)}] dt,
\label{eq:gramian-approx}\end{aligned}$$ where $\Delta{\bz_i} = \bz^{+i} - \bz^{-i}$ and $\bz^{\pm i}$ is the simulated measurement when the state $\bx^i$ is perturbed by a small value $\pm \epsilon$. The empirical local observability Gramian in converges to the local observability Gramian in for $\epsilon \rightarrow 0$.
The main disadvantage of this numerical approximation is that it cannot approximate the local observability Gramian for the states that do not appear in the sensor model. As $\epsilon \rightarrow 0$ this approximation replaces the state transition matrix $\Phi({0}, t)$ with the identity matrix. This relieves the burden of finding an analytical solution for $\Phi({0}, t)$, however, it also eliminates any effects on the local observability Gramian caused by the states that are not in the sensor model. Thus, it becomes difficult to reason about the observability of these states using this approximation. We address this problem in the following section.
Measure of Observability {#sec:measure}
------------------------
In order to present the hereby proposed measure of observability concisely, we introduce the following notation: $$\begin{aligned}
K_{t_0} (t) &= \frac{\partial}{\partial \bx}h_{t_0}(t) = \frac{\partial}{\partial \bx}h_{t_0}(\bx(t), \bu(t)).\end{aligned}$$
Following the definition of the local observability Gramian, we use the Taylor expansion of the sensor model to approximate the local observability Gramian: $$\begin{aligned}
W_o({0},{T}, \Delta{t}) &\approx \int_{{0}}^{{T}} K_t (t + \Delta{t})^T K_t (t + \Delta{t}) dt,
\label{eq:gramian-final}\end{aligned}$$ where $\Delta{t}$ is a fixed horizon that enables us to see the effects of the system dynamics. In order to measure the quality of observability we use the smallest singular value of the approximated local observability Gramian $W_o({0},{T}, \Delta{t})$.
In contrast to the empirical local observability Gramian, our formulation is able to capture input-output dependencies that are not visible in the sensor model. We achieve this property by incorporating higher order Lie derivatives that are included in the observability matrix. Intuitively, at each time step, we use the Taylor expansion of the sensor model about the current time step $t$ to approximate the Jacobian of the measurement in a fixed time horizon $\Delta{t}$. We use this approximation to estimate the local observability Gramian which is integrated over the entire trajectory.
In order to measure the observability of a subset of the states, one can use the smallest singular value of the submatrix of the local observability Gramian that includes only the states of interest. We use this property to focus on different self-calibration states of the system.
Trajectory Representation and Optimization {#sec:trajectory}
==========================================
Differentially Flat Trajectories
--------------------------------
In order to efficiently represent trajectories, we consider differentially flat systems [@van1997real]. A system is differentially flat if all of its inputs $\bx, \bu$ can be represented as a function of flat outputs $\mathbf{y}$ and their finite derivatives $\mathbf{\tilde{y}}$, i.e.: $$\begin{aligned}
\bx &= \mathbf{\zeta}(\mathbf{y},\dot{\mathbf{y}}, \ddot{\mathbf{y}}, ..., \overset{(n)}{\mathbf{y}}) = \mathbf{\zeta}(\mathbf{\tilde{y}})\\
\bu &= \mathbf{\psi}(\mathbf{y},\dot{\mathbf{y}}, \ddot{\mathbf{y}}, ..., \overset{(m)}{\mathbf{y}}) = \mathbf{\psi}(\mathbf{\tilde{y}}).\end{aligned}$$ For the rest of this paper, we express robot trajectories as flat outputs because this is the minimal representation that enables us to deduce the state and controls of the system over time.
Constrained Trajectory Representation using Piecewise Polynomials {#sec:constrained}
-----------------------------------------------------------------
Similar to @mueller14iros, we represent a trajectory by a $k$-dimensional, $d$-degree piecewise polynomial, composed of $q$ pieces: $$\begin{aligned}
\vy(t)=
\begin{cases}
P_1 {\mathbf{t}}(t) &\text{if } t_0 \le t < t_1 \\
\vdots\\
P_q {\mathbf{t}}(t) &\text{if } t_{q-1} \le t \le t_q,
\end{cases}\end{aligned}$$ where $P_i$ is the $k \times (d+1)$ matrix of polynomial coefficients for the $i$th polynomial piece, and ${\mathbf{t}}$ is the time vector, i.e.: $$\begin{aligned}
{\mathbf{t}}(t) = \begin{bmatrix}t^0 & t^1 & \dots & t^d\end{bmatrix}^T.
$$
We formulate constraints on the initial and final positions and derivatives of a trajectory as a system of linear equations [@mueller14iros]. For example: $$\begin{aligned}
\vc_1 &= \vy({0}) = P_1 {\mathbf{t}}({0}) \\
\vc_2 &= \dot{\vy}({0}) = P_1 \dot{{\mathbf{t}}}({0}) \\
\vdots \\
\vc_f &= \overset{(n)}{\vy}({T}) = P_q \overset{(n)}{{\mathbf{t}}}({T}),\end{aligned}$$ where $\vc_1, \vc_2, ..., \vc_f$ are the trajectory constraints and $\dot{{\mathbf{t}}}$ is the trivial derivative $\dot{\vy}(t) = P_i\dot{{\mathbf{t}}}(t)$. In matrix form, the initial constraints appear as: $$\begin{aligned}
\label{eq:constraints}
\begin{bmatrix} \mathbf{c}_1 & \mathbf{c}_2 & \dots \end{bmatrix} = P_1 \begin{bmatrix} \mathbf{t}({0}) & \mathbf{\dot{t}}({0}) & \dots \end{bmatrix}\end{aligned}$$ and the final constraints are defined simlarly.
In addition to start and end constraints, a physically plausible trajectory must be continuous up to the $\beta$-th derivative: $$\begin{aligned}
P_1 {\mathbf{t}}(t_1) = P_2 {\mathbf{t}}(t_1) &\: \hdots\: P_1 \overset{(\beta)}{{\mathbf{t}}}(t_1) = P_2 \overset{(\beta)}{{\mathbf{t}}}(t_1)\\
& \enspace \, \vdots \\
P_{q-1} {\mathbf{t}}(t_{q-1}) = P_q {\mathbf{t}}(t_{q-1}) &\: \hdots\: P_{q-1} \overset{(\beta)}{{\mathbf{t}}}(t_{q-1}) = P_q \overset{(\beta)}{{\mathbf{t}}}(t_{q-1}).\end{aligned}$$ To compactly express the evaluation of a polynomial and its first $\beta$ derivatives at a point in time, we define the time matrix: $$\begin{aligned}
{\mathbf{T}}_i = \begin{bmatrix} {\mathbf{t}}(t_i) & \dot{{\mathbf{t}}}(t_i) & \dots & \overset{(\beta)}{{\mathbf{t}}}(t_i) \end{bmatrix}.\end{aligned}$$ We may thus express the smoothness constraints as a banded linear system: $$\begin{aligned}
\label{eq:ppconstraints}
\begin{bmatrix} P_1 & \dots & P_q \end{bmatrix}
\begin{bmatrix}
{\mathbf{T}}_1 & & & \\
-{\mathbf{T}}_1 & {\mathbf{T}}_2 & & \\
& {\hspace{{.2ex}}\raisebox{{0.2ex}}{\rotatebox{{13}}{$\ddots$}}}& {\hspace{{.2ex}}\raisebox{{0.2ex}}{\rotatebox{{13}}{$\ddots$}}}& \\
& & -{\mathbf{T}}_{q-2} & {\mathbf{T}}_{q-1} \\
& & & -{\mathbf{T}}_{q-1} \\
\end{bmatrix} = 0.\end{aligned}$$ If we add equations in the form of for the initial and final constraints, the resulting linear system completely expresses the problem constraints.
With an appropriately high polynomial degree $d$, forms an underdetermined system. Therefore, we can use the left null space of the constraint matrix as the optimization space. Any linear combination of the left null space may be added to the particular solution to form a different polynomial that still satisfies the trajectory constraints. For the particular solution, we use the minimum-norm solution from the Moore-Penrose pseudoinverse, which gives smooth trajectories.
The described change of variables significantly reduces the dimensionality of the optimization problem and eliminates all equality constraints. The only remaining constraints are nonlinear inequalities related to the physical limits of the motor torques, etc., required to execute the path.
Numerical Stability for Constrained Trajectory
----------------------------------------------
The linear system presented in the previous section tends to be ill-conditioned. Its condition number grows exponentially with the polynomial degree [@pan2015bad] and it is exacerbated by longer time intervals in the polynomial pieces. Using the formulation from [@mellinger2011minimum], we scale the time duration of the problem such that each polynomial piece lies in a time interval of $\leq 1$ second. This produces a better-conditioned matrix whose solution can easily be converted back into a solution to the original problem.
The Optimization Objective {#sec:objective}
--------------------------
The goal of this work is to find a trajectory that will provide an optimal convergence of the self-calibration parameters of a nonlinear system. In order to achieve this goal, we aim at minimizing the cost function of the following form: $$\begin{aligned}
\argmin_{\tilde{\mathbf{y}}({0}), ..., \tilde{\mathbf{y}}({T})} o(\mathbf{\tilde{y}}(t)),\end{aligned}$$ where $o(\mathbf{\tilde{y}}(t))$ is the observability-dependent cost that is directly related to the convergence of the self-calibration states in the estimator.
In case of the hereby presented measure of quality of observability, $o(\mathbf{\tilde{y}}(t))$ is: $$\begin{aligned}
o(\mathbf{\tilde{y}}(t)) = \sigma_{min}(W_o({0},{T}, \Delta{t})),\end{aligned}$$ where $\sigma_{min}(W_o({0},{T}, \Delta{t}))$ is the minimum singular value of the approximated local observability Gramian $W_o$ described in .
To the best of our knowledge, the only other cost function that reflects the convergence of the states of the system is based on the EKF covariance. Minimizing the trace of the covariance results in minimizing the uncertainty about the state for all of its individual dimensions [@beinhofer2013robust] and yields better results than optimizing its determinant (i.e. mutual information) [@hausman2015cooperative]. Therefore, we employ the covariance-trace cost function that integrates the traces of the covariance submatrices that are responsible for the self-calibration states. We use this method as one of the baselines for our approach.
As described in , introducing the new constrained trajectory representation enables us to pose trajectory optimization as an unconstrained optimization problem and reduce its dimensionality. However, in order to ensure the physical plausibility of the trajectory, we still need to optimize it subject to physical limits of the system. We represent the physical inequality constraints as nonlinear functions of the differentially flat variables.
For optimization we use the implementation of the Sequential Quadratic Programming (SQP) method with nonlinear inequality constraints from Matlab Optimization Toolbox.
Example Application to UAVs with IMU-GPS State Estimator {#sec:application}
========================================================
We demonstrate the presented theory on a simulated quadrotor with a 3-DoF position sensor (e.g. GPS) and a 6-DoF inertial measurement unit (IMU). This is a simple, widely popular sensor suite, but it presents a challenging self-calibration task, as there is limited intuition for what kind of trajectory would make the states well observable. Although we present experiments for the quadrotor, we emphasize that the presented theory can be applied to a variety of nonlinear systems.
EKF for IMU-GPS Sensor Suite
----------------------------
As a realization of the state estimator of the quadrotor, we employ the popular Extended Kalman Filter (EKF). The EKF continuously estimates state values by linearizing the motion and sensor model around the current mean of the filter. It recursively fuses all controls $\bu_{1:k}$ and sensor readings $\bz_{1:k}$ up to time $k$ and maintains the state posterior probability: $$\begin{aligned}
p(\mathbf{x}_k \mid \mathbf{z}_{1:k}, \mathbf{u}_{1:k}) = \mathcal{N}(\hat{\mathbf{x}}_k, \Sigma_k)\end{aligned}$$ as a Gaussian with mean $\hat{\mathbf{x}}_k$ and covariance $\Sigma_k$. In particular, we use the indirect formulation of an iterated EKF [@lynen2013robust] where the state prediction is driven by IMU measurements. We choose this state estimator due its ability to work with various sensor suites and proven robustness in the quadrotor scenario.
The state consists of the following: $$\begin{aligned}
\bx^T &= [{\bp^{i}_w}^T, {\bv^{i}_w}^T, {\bq^{i}_w}^T, {\bb_\omega}^T, {\bb_a}^T, {\bp^{p}_i}^T],
\label{eq:state_core}\end{aligned}$$ where $\bp^{i}_w$, $\bv^{i}_w$ and $\bq^{i}_w$ are the position, velocity and orientation (represented as a quaternion) of the IMU in the world frame, $\bb_w$ and $\bb_a$ are the gyroscope and accelerometer biases, and ${\bp^{p}_i}^T$ is is the relative position between the GPS module and the IMU.
The state is governed by the following differential equations: $$\begin{aligned}
\dot{\bp}^{i}_w &= \bv^{i}_w \nonumber \\
\dot{\bv}^{i}_w &= \bC_{(\bq^{i}_w)}^T (\mathbf{a}_m - \bb_a - \bn_a) - \mathbf{g} \nonumber \\
\dot{\bq}^{i}_w &= \frac{1}{2} \Omega (\mathbf{\omega}_m - \bb_\mathbf{\omega} - \bn_\omega) \bq^{i}_w \nonumber \\
\dot{\bb_w} &= \bn_{\bb_\mathbf{\omega}}, \dot{\bb_a} = \bn_{\bb_a},
\label{eq:random-process}\end{aligned}$$ where $\bC_{(\bq)}$ is the rotation matrix obtained from the quaternion $\bq$, $\Omega(\mathbf{\omega})$ is the quaternion multiplication matrix of $\mathbf{\omega}$, $\mathbf{a}_m$ is the measured acceleration, and $\mathbf{\omega}_m$ is the angular velocity with white Gaussian noise $\bn_a$ and $\bn_\omega$. Since the IMU biases can change over time, they are modeled as random processes where $\bn_{\bb_w}$ and $\bn_{\bb_a}$ are assumed to be zero-mean Gaussian random variables.
Starting from the initial state defined in , we define the error state as: $$\begin{aligned}
\tilde{\bx}^T &= [\Delta{\bp^i_w}^T, \Delta{\bv^{i}_w}^T, {\delta\Theta^{i}_w}^T, \Delta{\bb_\omega}^T, \Delta{\bb_a}^T, \Delta{\bp^{p}_i}^T],
$$ where $\tilde{\bx}$ is the error between the real state value $\bx$ and the state estimate $\hat{\bx}$. For quaternions the error state is defined as: $\delta\bq = \bq \otimes \hat{\bq} \approx {[1 \quad \frac{1}{2}\delta\Theta^T]}^T$.
In this setup, the self-calibration error states $\tilde{\bx}_{sc}$ are the gyroscope and accelerometer biases $\bb_\omega, \bb_a$ and position of the GPS sensor in the IMU frame $\bp^{p}_i$: $$\begin{aligned}
\tilde{\bx}^T_{sc} &= [\Delta{\bb_\omega}^T, \Delta{\bb_a}^T, \Delta{\bp^{p}_i}^T].\end{aligned}$$
Using the IMU-GPS state vector in , the system dynamics in , and assuming the connection between the IMU and the GPS sensor is rigid, we define the GPS sensor model as: $$\begin{aligned}
\bz_{gps} &= h(\bx, \bn_{\bz_{gps}}) = \bp_w^i + \bC_{(\bq^{i}_w)}^T \bp_i^p + \bn_{\bz_{gps}},\end{aligned}$$ where $\bn_{\bz_{gps}}$ is white Gaussian measurement noise.
The nonlinear observability analysis in [@kelly2011visual] and [@WeissPhD2012] shows that the system is fully observable with appropriate inputs.
Differentially Flat Outputs and Physical Constraints of the System
------------------------------------------------------------------
As shown by @mellinger2011minimum the quadrotor dynamics are differentially flat. This means that a quadrotor can execute any smooth trajectory in the space of flat outputs as long as the trajectory respects the physical limitations of the system. The flat outputs are $x,y,z$ position and yaw $\theta$: $$\begin{aligned}
\mathbf{y} = [x,y,z,\theta]^T.\end{aligned}$$ The remaining extrinsic states, i.e. roll and pitch angles, are functions of the flat outputs and their derivatives. In order to ensure that trajectories are physically plausible we place inequality constraints on 3 entities: the thrust-to-weight ratio ($\leq 1.5$), angular velocity ($\leq \pi \frac{rad}{s}$), and angular acceleration ($\leq 5\pi\frac{rad}{s^2}$). These values are rough estimates for a small-size quadrotor.
Experimental Results {#sec:experiments}
====================
Experimental Setup
------------------
We evaluate the proposed method in simulation using the quadrotor described in . The simulation environment enables extensive testing with ground truth self-calibration states that would not be possible for a real robot. We represent trajectories as degree-6 piecewise polynomials with continuity up to the 4th derivative. In all experiments, we require trajectories with zero velocity and acceleration at the beginning and end points.
The quadrotor has a GPS sensor that is positioned $\bp^p_i = [0.1\,\,0.1\,\,0.1]^T$m away from the IMU and produces measurements with standard deviation of 0.2m. The accelerometer and the gyroscope have initial biases of $\bb_a = [0.05\,\,0.05\,\,0.05 ]^T$ m/s$^2$ and $\bb_\omega = [0.01\,\,0.01\,\,0.01]^T$ rad/s respectively. These are common values for real quadrotor systems that we have used. The initial belief is that all the self-calibration states are zero. Thus, a bad self-calibration trajectory will fail to converge the state estimate of the system.
Evaluation of Various Self-Calibration Routines
-----------------------------------------------
![Self-calibration task: results obtained for optimizing for different objectives. Top row: random and optimized results when optimizing for $\bp^p_i$. Bottom row: random and optimized results when optimizing for $\bb_a$. The left and right column show the final RMSE for $\bb_a$ and $\bp^p_i$ respectively.[]{data-label="fig:ba-vs-pip"}](bad_random_4up_final_pretty.eps){width="0.99\columnwidth"}
![Self-calibration task: final RMSE values for the accelerometer bias $\bb_a$ and the GPS position in the IMU frame $\bp^p_i$ obtained using optimization (green) and 3 different heuristics: star and figure eight trajectories from and randomly sampled trajectories that are close of the physical limits of the system.[]{data-label="fig:scatter-all"}](good_random_2up_final_pretty.eps){width="0.99\columnwidth"}
To evaluate the influence of choosing different states to construct the local observability Gramian, we compared two optimization objectives: i) the local observability Gramian constructed using the position states with the $\bp^p_i$ states, and ii) the position states with the $\bb_a$ states. Initial tests showed that $\bb_\omega$ converges quickly for almost any trajectory, so we did not include it in the evaluation. For the self-calibration task we require trajectories to start and end at the same position. We generated random trajectories by sampling a zero-mean Gaussian distribution for each optimization variable, i.e. each component of the left null space of the piecewise polynomial constraint matrix (). We then used each random trajectory as an initial condition for nonlinear optimization to produce an optimized trajectory.
shows the optimization results using both objectives. The optimized trajectories significantly outperformed the randomly generated ones. The two self-calibration states $\bp^p_i$ and $\bb_a$ are co-related in our system, i.e. optimizing for one state also leads to improved performance on the other state. However, one can observe that the trajectories optimized for $\bb_a$ yield improved $\bb_a$ final RMSE values compared to those of trajectories optimized for $\bp_i^p$, and analogously the trajectories optimized for $\bp^p_i$ yield better $\bp_i^p$ results than those optimized for $\bb_a$. Due to the small differences between results in the case of the accelerometer bias $\bb_a$ and the larger difference for the position of the GPS sensor $\bp^p_i$, we chose to conduct further experiments using the $\bp^p_i$ objective.
shows results from the same experiment for a number of differently constructed trajectories. *PL-random* is a more competitive set of random trajectories generated by choosing larger random null space polynomial weights and discarding trajectories that violated the quadcopter’s physical limits. The remaining trajectories are therefore likely to contain velocities and accelerations that are near the physical limits, which should lead to better observability. *Figure 8* and *star* are the heuristic trajectories presented in , and *our method* are trajectories generated from our optimization framework using the *PL-random* trajectories as initial conditions. While the *star* trajectory and some of the *PL-random* trajectories perform well on $\bb_a$, our approach outperforms all other methods on $\bp_i^p$.
![Self-calibration task: statistics collected over 50 runs of the quadrotor EKF using 6 different trajectories: ours - optimized trajectory using the hereby defined observability cost; trace - optimized trajectory using the covariance-trace cost function; PL-random - randomly sampled trajectory that is very close to the physical limits of the system; star, figure 8 - heuristics-based trajectories presented in ; random - randomly sampled trajectory that satisfies the constraints. Top left: GPS position integrated RMSE, top right: GPS position final RMSE, bottom left: accelerometer bias integrated RMSE, bottom right: accelerometer bias final RMSE.[]{data-label="fig:bars"}](bars_integrated_pretty.eps){width="0.99\columnwidth"}
In order to more extensively test the different self-calibration strategies, we collected statistics over 50 EKF simulations for a single representative trajectory from each strategy. summarizes our results in terms of the RMSE integrated over the entire trajectory and the final RMSE for accelerometer bias $\bb_a$ and GPS position $\bp^p_i$. Results show that our approach outperforms all baseline approaches in terms of the final and integrated RMSE of the GPS position $\bp^p_i$. The only method that is able to achieve a similar integrated RMSE value for the GPS position is the covariance-trace-based optimization described in . However, it takes approximately 13 hours for the optimizer to find that solution, versus approximately 10 minutes with our method. The main reason for this is that in order to estimate the trace of the covariance of the EKF, one needs to perform matrix inversion at every time step, which is more computationally expensive than the integration of the local observability Gramian and the one singular value decomposition used by our approach. The integrated RMSE of the accelerometer bias $\bb_a$ also suggests that our approach is able to make this state converge faster than in other methods. Nevertheless, a few other trajectories such as covariance-trace-based and PL-random were able to perform well in this test. This is also visible in the case of the final RMSE of the accelerometer bias $\bb_a$ where the first four methods yield similar results. While our method is slightly worse than the covariance-trace-based and the two heuristics-based approaches, given the standard deviation of the measurement (0.2m) and the final RMSE values of the bias being below 0.005 $\frac{m}{s^2}$ (which is less than 10% of the initial bias RMSE), we consider the trajectories from all 4 methods to have converged this estimate.
Evaluation of an Example Waypoint Trajectory
--------------------------------------------
In addition to the self-calibration task, we applied our method to a waypoint navigation task. With minor extensions, the piecewise polynomial constraint matrix formulation in can satisfy position and derivative constraints along the path in addition to start and endpoint constraints. We compare a trajectory optimized using our method to a minimum snap trajectory computed using the method from [@mellinger2011minimum].
shows both of the optimized trajectories. The trajectory optimized using our method is much more complex than a simple min-snap trajectory because it aims to yield well-observable states. The results in show that our trajectory yields 4x better GPS sensor position estimates and 2x better position estimates than the min-snap trajectory. We note that even though the observability-aware trajectory is longer and more complex, which makes the state estimation harder, the resulting estimates are still significantly better than the min-snap trajectory. This result supports the intuition that sensor calibration can have significant influence on the estimation of other system states.
![Waypoint navigation task: minimum snap trajectory (red) vs. optimized trajectory using our method (green). Position waypoints are shown in black. []{data-label="fig:waypoint"}](waypoints_newangle_pretty.eps){width="0.99\columnwidth"}
Discussion
----------
The presented results indicate that our approach is able to outperform other baselines at the task of estimating the position of the GPS sensor in the IMU frame. However, it yields comparable results with other methods regarding the accelerometer bias. Even though for this simple system one can think of heuristics that performs reasonably well, these may not generalize for more complex systems.
The main advantage of the presented framework lies in its generality – it is applicable to any nonlinear system – and the fact that it can be combined with other objectives as long as they can be represented in the cost function.
The waypoint navigation task presents another useful application of our technique, and shows that the influence of the GPS-IMU position estimate on the quality of position estimation can be significant (see ).
min-snap our method
----------------------- ------------------- -------------------
$\bp^i_w$ $\int$ RMSE $6.06 \pm 1 m$ $3.69 \pm 0.31 m$
$\bp^p_i$ $\int$ RMSE $3.40 \pm 0.83 m$ $0.85 \pm 0.22 m$
: Waypoint navigation task: statistics of the integrated RMSE values for position $\bp^i_w$ and GPS sensor position in the IMU frame $\bp^p_i$ collected over 50 runs. Both trajectories take 50 seconds.[]{data-label="tab:waypoints"}
Conclusion {#sec:conclusion}
==========
We introduced an observability-aware trajectory optimization framework that is applicable to any nonlinear system and produces trajectories that are well suited for self-calibration. In contrast to existing approaches, our method moves the focus from *where* to go during a mission to *how* to achieve the goal while staying well-observable. The presented results performed for a simulated quadrotor system with a GPS-IMU sensor suite demonstrate the benefits of the optimized observability-aware trajectories compared to other heuristics and a covariance-based approach. For the self-calibration task we were able to achieve almost 2x better final RMSE values for the GPS-IMU position state than all the baseline approaches and comparable converged values for the accelerometer and gyroscope biases. Our method runs $\thicksim$80x faster than the only other generic baseline approach that is applicable to other systems, and it achieved better results.
The presented method was also applied to a waypoint navigation task and achieved almost 2x better integrated RMSE of the position estimate and more than 4x better integrated RMSE of the GPS-IMU position estimate than the minimum snap trajectory.
In the future, we plan to test this method on multi-sensor fusion systems where the observability of the states is of even greater importance and the self-calibration states have bigger influence on the other states. The next steps also include evaluating the optimized trajectories on real UAVs and other robotic systems.
[^1]: Advanced sensors and state estimators are nowadays able to obtain accurate measurements in a large variety of different environments [@Li2013b]
[^2]: <https://www.google.com/atap/project-tango/>
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Matthew Baker, Barry Mazur, Ken Ribet'
bibliography:
- 'AMSNotices.bib'
title: 'Robert F. Coleman 1954 – 2014'
---
> IN MEMORIAM: [*Robert Coleman Professor of Mathematics, UC Berkeley* ]{}
{width=".4\textwidth"}
Biography
=========
Robert F. Coleman, an extraordinarily original and creative mathematician who has had a profound influence on modern number theory and arithmetic geometry, died of a sudden heart attack in El Cerrito, CA on the morning of March 24, 2014. He is survived by his wife Tessa, his sister Rosalind and brother Mark, his nephew Jeffrey and niece Elise, and his service dog Julep. The depth and importance of his mathematical ideas, his congeniality, the joy radiating from his playful disposition, and his sheer inexhaustible energy—all this in the face of Multiple Sclerosis, a condition that did not deter him from full engagement with life—made Robert an inspiration to his friends, family, students, and colleagues. Robert also worked toward making civic structures and laws more appropriate for people with disabilities, and his activism is yet another reason that Robert was so widely admired.
Robert was born on Nov. 22, 1954 in Glen Cove, NY. He displayed an early talent for mathematics, winning an Intel Science Talent Search Award in 1972 as a high school student. He earned a mathematics degree from Harvard University and subsequently completed Part III of the mathematical tripos at Cambridge, where he did research under the supervision of John Coates. By the time he entered graduate school at Princeton, Robert had essentially already written his doctoral dissertation, but his formal thesis advisor was Kenkichi Iwasawa. His dissertation, entitled “Division Values in Local Fields," is considered a landmark contribution to local class field theory. After completing his Ph.D., Robert returned to his alma mater, Harvard University, as a Benjamin Peirce Assistant Professor and Research Associate. He came to UC Berkeley as an Assistant Professor in 1983, and was promoted unusually quickly to Associate and then Full Professor. He taught and did mathematics at Berkeley until his untimely death. He had 12 Ph.D. students, published 63 research papers, and received numerous honors and recognition for his work, including a MacArthur “Genius" Fellowship in 1987.
{width=".4\textwidth"}
Overview of Robert’s mathematics
================================
Robert Coleman’s primary mathematical love was number theory, with a particular interest in $p$-adic aspects of the subject[^1], for which he had an amazing intuition. We will briefly summarize here some of Robert’s key contributions to mathematics; these will be elaborated upon in subsequent sections.
- [**Coleman maps:**]{} Classical analytic number theory, as developed in the works of Dirichlet and Dedekind, and in Minkowski’s Geometry of Numbers, makes a deep connection between arithmetic and analysis in algebraic number theory, with the theory of [*the regulator*]{} as the vital bridge. In the $p$-adic analytic number theory of number fields, elliptic curves, and modular forms, “Coleman maps’" provide the corresponding $p$-adic bridge. We will elaborate on this theme in Section \[reg\].
- [**Coleman integration:**]{} There is a serious obstacle to transferring one of our most useful tools in complex analysis—integration over paths—from ${\C}$ to $\C_p$. Among other obstructions, one of the fundamental problems is the totally disconnected topology of the $p$-adic fields. Robert Coleman could somehow visualize paths and structures in $p$-adic geometry which no one else in the world saw as keenly or as profoundly[[^2]]{}. Thanks to Robert’s vision, we can now integrate $p$-adic differential one-forms over paths in a manner analogous to the classical complex theory. Robert called the key new idea “analytic continuation along Frobenius”. In his 1985 Annals of Math paper “Torsion Points on Curves and $p$-adic Abelian Integrals", he wrote:
> Rigid analysis was created to provide some coherence in an otherwise totally disconnected $p$-adic realm. Still, it is often left to Frobenius to quell the rebellious outer provinces.
Applications of Coleman integration include the celebrated “method of Chabauty–Coleman” for finding rational points on curves, a new proof of the Manin-Mumford conjecture, and new results in $p$-adic Hodge theory. We will elaborate on these themes in Section \[p-int\].
- [**Coleman Families:**]{} The theme of “interpolation" of classical modular forms of integral weight to get “continuous families” of $p$-adic weights is one that might be traced back (with appropriate translation of language) to Ramanujan. The emergence of the interpolated “$p$-adic modular eigenforms" is one of the most important topics of contemporary number theory. Following pioneering work of J.-P. Serre, Robert produced [*$p$-adic families*]{} of $p$-adic Banach spaces of (overconvergent) $p$-adic modular forms, together with a completely continuous operator (the Atkin–Lehner operator) whose Fredholm spectrum gives one a good handle on the locus of eigenforms. Robert’s construction also opened the way for the “eigencurve”, about which we will elaborate in Section \[eigen\].
All three of these directions of Robert’s research dovetail together, and remain thoroughly contemporary in their importance. This point was illustrated in a comment made by Kevin Buzzard[^3] on the day that he heard of Robert’s death:
> Two or three years ago, I remember noting one afternoon \[after hearing three talks at Imperial College\] that even though the topics were covering a lot of modern algebraic number theory ($(\phi,\Gamma)$-modules and their applications, $p$-adic modular forms, and computational methods for rational points on algebraic varieties), Robert’s work had been a critical component of all three lectures I heard that day: the first was on the Coleman homomorphism, the second on Coleman families, and the third used Coleman’s explicit Chabauty. A wonderful and genuinely coincidental moment of synchronicity which only underlined how much Robert has influenced modern number theory.
Robert made many other important contributions to number theory, too many to discuss in any depth in this article. For example, he discovered and filled in a gap in Manin’s proof of the Mordell conjecture over function fields and, with Voloch, verified some previously unchecked compatibilities in Dick Gross’s theory of companion forms, thus completing Gross’s influential work. His work on semistable models for modular curves of arbitrary level was recently refined and applied to the Local Langlands Program by Jared Weinstein.
The challenges of MS
====================
Robert Coleman was an avid tennis player and world traveler when he was struck with a severe case of Multiple Sclerosis in 1985. According to his friend and collaborator Bill McCallum:
> \[Robert\] was on a visit to Japan and had a couple of strange incidents: being unable to hit the ball in tennis, stumbling on the stairs in the subway. The Japanese doctors didn’t know what was going on. When he got back to Berkeley he was diagnosed with MS. From that point he had a shockingly rapid descent that caught his doctors by surprise; for many patients MS is a long slow decline. He went from perfectly healthy to the verge of death within a matter of weeks. His decline was arrested by some experimental and aggressive chemotherapy treatment…But he never bounced back the way MS patients sometimes do.
Robert fought MS bravely and with a great sense of humor, and despite the severity of his illness he remained an active faculty member at UC Berkeley until his retirement in 2013. Those that knew him were consistently amazed by his optimism and by the way he continued to travel despite the challenges of his MS. His former Ph.D. student Harvey Stein called it “truly amazing" how happily Robert lived his life, carrying on “as if \[his MS\] wasn’t even relevant". His friend and colleague Jeremy Teitelbaum said:
> Of course Robert was brilliant, but what really impressed me about him was how relentless he was when it came to mathematics. He simply never gave up in the pursuit of an idea…. When his MS had made his hands shake and his voice slur, he brought that same determination to bear. I watched him spend hours typing slides for his lectures so that he could keep teaching—things that should have taken an hour would take half a day, but he did not give up.
Robert had a mischievous and impish sense of humor, and he surrounded himself with colorful and funny people. For many years Robert’s closest companion was his guide dog Bishop, who would join Robert everywhere. Bishop eventually passed away and Robert found a new canine companion named Julep. (Memorial fund donations to “Paws With A Cause" can be made in honor of Robert’s special relationship with his service dogs.)
{width=".4\textwidth"} {width=".4\textwidth"}
Dan Sparks, a graduate student of Robert’s, recalled Julep’s second birthday party:
> Several other dogs came to the party. The usual birthday stuff happened: Julep got presents and toys, as well as a nice big slice of cake. The cake was made special for the dogs from roast beef, with mashed potato icing.
and went on to reflect:
> …for all the amusing anecdotes (which reflect his unyielding sense of humor)… perhaps the most memorable thing about him is how optimistic and positive he was…His positive energy was completely contagious, and the difficulties that he overcame put my own concerns into perspective. The net result of that was the greatest contribution that Robert Coleman made to my life.
Robert purchased an off-road wheelchair from John Castellano in 1989, which he used to go mountain biking with John, his wife Marisa, and Bas Edixhoven. John and Marisa would take turns towing Robert up hills using bungee cords attached to their bikes. According to Marisa, “Sometimes we would tandem tow. And then Robert would fly downhill. Robert may have seemed meek in some ways, but he was a thrill seeker on those downhills!”
{width=".4\textwidth"} {width="40.00000%"}
Activism on behalf of people with disabilities
==============================================
Almost immediately after the onset of his illness, Robert began working as an activist and spokesperson for people with disabilities[^4]. A resident of El Cerrito, CA, he became an organizer of the Albany–El Cerrito Access Group, which argued successfully for a range of better resources for the disabled, including easier access to buildings and public spaces. In particular, Robert managed to effect the installation of a long ramp at an elementary school playground in El Cerrito that allowed a wheelchair user to ride to the top of a high slide and then take the slide back down to the ground (leaving the wheelchair behind).
Robert seemed to have a special ability to identify situations that could be turned into class action suits. He worked closely with a local law firm that had filed several such cases in the past, and succeeded repeatedly in getting offending local businesses to cease their egregious behavior. The settlements from these lawsuits were put to good use; for example, Robert’s group funded the installation of audible signals for the blind at major intersections in El Cerrito. The group also paid for improvements to local libraries for patrons with impaired vision.
The Potato Society
==================
For many years, Robert hosted a wine and cheese gathering in his office every Friday afternoon code-named “Potatoes”. Bill McCallum explains the origin of the name:
> In Fall of 1978 or 1979 I purchased a case of Australian wine (Jacob’s Creek Cabernet?) from the Wine and Cheese Cask and had it delivered to my cubicle in the math department at Harvard, meaning to take it home. Instead, a group of us started drinking it when the cheap Egri Bikaver the department served for Friday afternoon wine and cheese ran out… We started having our own wine and cheese in my cubicle area. Greg Anderson, Robert Coleman, Robert Indik and Brad Osgood were regulars. In Fall of 1981 we moved into to Brad’s office, and started going out to One Potato Two Potato for dinner afterwards. When Robert moved to Berkeley he continued the tradition there (it died out at Harvard after we all left).
Robert’s student Harvey Stein recalls:
> Sometimes Potato Society meetings were a little more exotic. Hendrik Lenstra, returning from the Netherlands, once brought what must have been a gallon sized bucket of herring, along with a bottle of jenever. I spent much time filleting fish that night.
And as Ken Ribet describes it:
> We always had a great time at Potato Society. We could usually begin by opening up a couple of the bottles that were left over from the previous week. After Society members began to stream in, there were baguettes, new bottles of wine and chunks of cheese on the tables. We always managed to drag in extra people from the department who were walking near Robert’s office. Our group often included postdocs, graduate students and staff members. There’d occasionally be an undergrad, but that was exceptional. We’d always have a good view of the Golden Gate Bridge and the setting sun. Depending on the season, the sun would be north of the bridge, right at the bridge or further south. Life stopped for a couple of hours late in the day every Friday.
Coleman maps {#reg}
============
Robert Coleman exploded onto the mathematical scene with his Ph.D. thesis, which was published as “Division Values in Local Fields” [@Coleman_DivisionValues] and continued in “The Arithmetic of Lubin-Tate Division Towers” [@Coleman_LubinTate]. In the introduction to [@Coleman_DivisionValues], Coleman writes:
> In his work on cyclotomic fields, Kummer observed that various formal operations on power series had number theoretic applications. Perhaps the most striking of these was Kummer’s idea of taking logarithmic derivatives of $p$-adic numbers. After a long period of neglect, various refinements and generalizations of Kummer’s idea have recently been used by Iwasawa and Wiles to study explicit reciprocity laws, and by Coates and Wiles to study the arithmetic of elliptic curves with complex multiplication… The aim of the present paper is to begin a deeper and more systematic study of the local analytic theory which underlies these relations between power series and $p$-adic numbers.
The main result of Coleman’s paper is a general theorem on the interpolation of division values in Lubin–Tate formal groups which has found numerous applications to local class field theory, Iwasawa theory, and the arithmetic of $p$-adic $L$-functions. We now describe this result, and its relation to Iwasawa theory from [@Coleman_LubinTate], in the special setting of the (formal group associated to the) multiplicative group ${\mathbf G}_m$ following Pierre Colmez’s exposition in [@Colmez_ColemanGeneralization]. We also attempt to place these results into the larger context of $p$-adic representations and $p$-adic $L$-functions.
To begin, fix a compatible system $\{ \epsilon_n \}$ of roots of unity in $\overline{\Q_p}$ with $\epsilon_1 \neq 1$ and $\epsilon_{n+1}^p = \epsilon_n$ for all $n \geq 1$. Let $K_n = \Q_p(\epsilon_n)$ and $K_\infty = \bigcup_{n \in {\mathbf N}} K_n$. Let $\Gamma = {\rm Gal}(K_\infty / \Q_p)$ and let $\Lambda = {\mathbf Z}_p [[\Gamma]]$ be the completed group algebra of $\Gamma$. The [*Coleman norm map*]{} is a homomorphism $E: \varprojlim {\mathcal O}_{K_n}^* \to \Lambda$ from the projective limit of the groups ${\mathcal O}_{K_n}^*$ with respect to the norm maps to the completed group algebra $\Lambda$. To define the map $E$, Coleman first proves that given an element $u = (u_n)$ of $\varprojlim {\mathcal O}_{K_n}^*$, there is a unique invertible power series ${\rm Col}_u(T) \in \Z_p[[T]]$ such that ${\rm Col}_u(\epsilon_n - 1) = u_n$ for all $n \geq 1$. (This is the interpolation theorem referred to above in the present context.) The logarithmic derivative of ${\rm Col}_u(T)$ again has coefficients in $\Z_p$ and there is a unique measure $\mu_u$ on $\Z_p$ such that $$\int_{\Z_p} (1 + T)^x \mu_u = (1+T) \frac{d}{dT} \log \left( {\rm Col}_u(T) \right).$$ Restricting this measure to $\Z_p^*$ and pulling it back to $\Gamma$ via the cyclotomic character gives the Coleman map $E$, which turns out to be nearly an isomorphism. As Colmez writes:
> The measure which is used to define the Kubota–Leopoldt $p$-adic zeta function is the image of the cyclotomic units via this map, so the \[Coleman norm map $E$\] can be thought of as a machine producing $p$-adic $L$-functions out of compatible systems of units.
All of this can be interpreted in terms of the $p$-adic representation $\Q_p(1)$ associated to the cyclotomic character, and much later Cherbonnier and Colmez [@ColmezCherbonnier] used the theory of $(\phi,\Gamma)$-modules introduced by J.-M. Fontaine[^5] to extend Coleman’s map to arbitrary $p$-adic representations.
Coleman subsequently realized that his norm map, combined with $p$-adic analytic properties of the dilogarithm function $$\label{eq:dilog}
\ell_2(z) = \sum_{k=1}^\infty \frac{z^k}{k^2},$$ allow one to give a new explicit formula for the Hilbert norm residue symbol of local class field theory. More specifically, in “The Dilogarithm and the Norm Residue Symbol” [@Coleman_NormResidueSymbol], he presented a new formula for the norm residue symbol of exponent $p^n$ attached to a cyclotomic extension of $\Q_p$ containing the $p^n$-th roots of unity. This was the first concrete connection between the dilogarithm series, which plays an important role in Thurston’s work on hyperbolic geometry and the work of Bloch and Milnor on algebraic K-theory, and class field theory. (Such a connection is quite natural in view of the fact that in both settings the Steinberg relation between $x$ and $1-x$ plays a central role.) Earlier explicit formulas for the norm residue symbol took on different forms for $p=2$ and odd primes, but in Coleman’s formula all primes are treated equally. Coleman relates being a local norm to being an exact differential, and is led by this to study a class of differential equations whose solutions can be expressed in terms of logarithms and dilogarithms. He is also able to give a new proof of the non-degeneracy of the norm residue symbol from his explicit formula.
Having been led by local class field theory to the study of p-adic analytic properties of the dilogarithm, one might naturally ask whether this function has a natural $p$-adic analytic continuation. Over $\C$, the series given in (\[eq:dilog\]) converges only in the open complex unit disc but can be analytically continued to a multi-valued function on $\C - \{ 0 \}$. Since the power series in (\[eq:dilog\]) also converges on the open $p$-adic unit disc in $\C_p$, Coleman had the idea to investigate whether the $p$-adic dilogarithm admits a similar natural extension. The primary difficulty is that analytic continuation does not work, even in the context of Tate’s rigid analytic spaces: the open disc is a maximal analytic domain for $\ell_2$. Coleman’s brilliant idea in “Dilogarithms, Regulators, and $p$-adic $L$-functions” [@Coleman_Dilogs] is to introduce what he called the “Dwork principle” of analytic continuation along Frobenius.[^6] Analytic extension for functions like polylogarithms, which satisfy differential equations with unipotent monodromy, is done by looking within the (vast) collection of locally analytic extensions at those global functions which satisfy certain rigidity assumptions on their twists by Frobenius. Coleman shows that these Frobenius conditions allow one to prove uniqueness of the resulting extension.
Having extended the range of definition of the $p$-adic dilogarithm (and, more generally, all polylogarithms), Coleman derives a new formula—analogous to classical formulas for Dirichlet $L$-functions over the complex numbers—for the value of Kubota–Leopold $p$-adic $L$-functions at a positive integer $k \geq 2$ in terms of the $k^{\rm th}$ polylogarithm function. He also uses the extended $p$-adic dilogarithm to produce a regulator map on $K_3(\C_p)$, analogous to Bloch’s regulator map on $K_3(\C)$ in terms of the complex dilogarithm. This uses a construction similar to Coleman’s explicit formula for the Hilbert norm residue symbol, and a similar method gives a $p$-adic regulator map for Tate elliptic curves. In the remarks at the end of the paper, Coleman writes: “In a subsequent paper we intend to show how the ideas in this paper lead to a theory of $p$-adic abelian integrals.” We turn to this theory next.
Coleman integration {#p-int}
===================
The theory laid out in [@Coleman_Dilogs] can be viewed as a theory of $p$-adic Abelian integrals on ${\mathbf P}^1$.[^7] This approach, and its applications to $p$-adic $L$-functions, was extended to arbitrary curves with good reduction by Coleman and de Shalit in “$p$-adic regulators on curves and special values of $p$-adic $L$-functions” [@ColemandeShalit_Regulators]. Specifically, if $C$ is a smooth complete curve over $\overline{\Q}_p$ whose Jacobian $J$ has good reduction, Coleman and de Shalit show how to define $p$-adic integrals for arbitrary meromorphic differential forms on $C$. As an application, if $W = H^0(C,\Omega^1_C)$ is the space of holomorphic differentials on $C$ and $T = {\rm Hom}(W,\overline{\Q}_p)$ is the tangent space of $J$ at the origin, they define a $p$-adic regulator pairing $r_{p,C} : K_2(\overline{\Q}_p(C)) \to T$ whose value at the Steinberg symbol $\{ f,g \}$ is the linear functional $$\label{eq:padicreg}
r_{p,C}(\{ f,g \})(\omega) = \sum_{i=1}^t \int_{p_i}^{q_i} {\rm Log}(g) \cdot \omega,$$ where “Log” denotes a fixed branch of the $p$-adic logarithm and ${\rm div}(f) = \sum_{i=1}^t (q_i) - (p_i).$
In the special case where $C=E$ is an elliptic curve with complex multiplication, the pairing (\[eq:padicreg\]) is related to a special value of the $p$-adic $L$-function of $E$; this gives a $p$-adic analogue of a theorem of Bloch. As Coleman and de Shalit write in their paper:
> The fragmentary evidence relating special values of classical $L$-functions to regulators on $K$-groups instigated very general and powerful conjectures of Bloch, Beilinson, and Deligne. Recently Soul[é]{} and Schneider have begun looking for $p$-adic conjectures. The relatively down-to-earth results of this paper, together with the earlier examples mentioned above, strongly support such analogues. Not less important, perhaps, is the indication that rigid analysis and $p$-adic integration ought to play some role in the proof of the $p$-adic conjectures.
{width=".4\textwidth"}
For a more restrictive class of differential forms (differentials of the second kind[^8]), but for varieties of any dimension having good reduction at $p$, Coleman developed in “Torsion points on curves and $p$-adic Abelian integrals” [@Coleman_AbelianIntegrals] a full-fledged theory of $p$-adic Abelian integrals of one-forms. The construction is again based on the Dwork principle of “analytic continuation along Frobenius”. Coleman establishes the basic properties of these $p$-adic integrals, including an important functoriality result which he uses to show that $p$-adic Abelian integrals of holomorphic one-forms satisfy an addition law. Once a branch of the logarithm has been fixed, the methods developed in [@Coleman_Dilogs; @ColemandeShalit_Regulators; @Coleman_AbelianIntegrals] can be combined to extend the theory of $p$-adic Abelian integrals to closed meromorphic one-forms on arbitrary varieties with good reduction, as well as to curves with bad reduction. In the latter case, one does not always have a single-valued primitive for holomorphic differentials: there are periods which arise. Coleman’s theory of $p$-adic integration of one-forms was extended by Zarhin, Colmez, Besser, Vologodsky, and others (c.f. [@Breuil_SeminarBourbaki]), culminating in the recent book “Integration of one-forms on $p$-adic analytic spaces” by Vladimir Berkovich [@Berkovich_Integration], which treats varieties of any dimension with no assumptions on the reduction type.
Going back to the paper [@Coleman_AbelianIntegrals], as a consequence of his addition law Coleman shows that if a curve $C$ over $\C_p$ with good reduction at $p$ is embedded in its Jacobian $J$, the torsion points of $J(\C_p)$ lying on $C$ are the common zeros of certain $p$-adic Abelian integrals of the first kind on $C$. These integrals can be expanded as power series in one variable (with respect to some local parameter on $C$), and the number of common zeros can be studied using Newton polygons and other classical tools from $p$-adic analysis. Using this idea, in “Ramified Torsion Points on Curves” [@Coleman_RTPC] Coleman was able to give an entirely new proof of the Manin–Mumford conjecture, originally proved by Raynaud: if $C$ is a curve of genus at least 2 over a field of characteristic zero, then there are only finitely many torsion points of $J$ lying on $C$. Unlike Raynaud’s proof, Coleman’s argument can sometimes be used to explicitly calculate this finite set of torsion points. For example, together with Tamagawa and Tzermias, Coleman showed in “The cuspidal torsion packet on the Fermat curve” [@ColemanTamagawaTzermias] that if $C$ is the Fermat curve $X^m + Y^m + Z^m = 0$, embedded in its Jacobian $J$ via a [*cusp*]{} (i.e., a point $(x,y,z)$ with $xyz=0$), then the set of torsion points of $J$ lying on $C$ is precisely the set of cusps for all $m \geq 4$. Coleman’s theory of $p$-adic integration plays a crucial role in the proof. Poonen [@Poonen_ComputingTorsionPoints] presented an effective algorithm to calculate torsion points on curves by combining one of Coleman’s main results in [@Coleman_RTPC] with ideas of Buium from [@Buium_pjets].
Perhaps the biggest splash made by the theory of $p$-adic Abelian integrals was Coleman’s highly influential paper “Effective Chabauty” [@Coleman_Chabauty], in which he resurrected an old idea due to Chabauty and showed that it led to effective bounds for the number of rational points on an algebraic curve $C$ over a number field $K$, provided that the Mordell–Weil rank of the Jacobian of $C$ is not too large. Specifically, Coleman proved the following theorem (which, for simplicity, we state only for $K={\mathbf Q}$):
[**Theorem:**]{} Let $C / {\mathbf Q}$ be an algebraic curve of genus $g \geq 2$, and assume that the rank $r$ of $J({\mathbf Q})$ is less than $g$. Then for any prime $p > 2g$ at which $C$ has good reduction, $$\# C({\mathbf Q}) \leq \# \overline{C}({\mathbf F}_p) + 2g-2.$$
Recall that the Mordell Conjecture, proved by Faltings (and independently by Vojta shortly thereafter), asserts that if $C$ is an algebraic curve over ${\mathbf Q}$ of genus at least 2, then the set $C({\mathbf Q})$ of rational points on $C$ is finite. At the present time, however, we do not know an effective algorithm (even in theory) to compute this finite set. The techniques of Faltings and Vojta do lead in principle to an upper bound for the size of $C({\mathbf Q})$, but the bound obtained is very far from sharp and is hard to write down explicitly. By contrast, when the Chabauty–Coleman method applies (i.e., when $r<g$), the bound is not only completely explicit, it is sometimes even sharp![^9] And even when Coleman’s bound is not sharp, the method of proof can often be used (together with explicit computations) to find the set of rational points on $C$ exactly. This has given rise to a whole industry in computational number theory, and many Diophantine equations have been solved using this method and its generalizations and refinements. For example, in book 6 of the [*Arithmetica*]{}, Diophantus of Alexandria poses a question which comes down to finding the positive rational solutions to $y^2 = x^6 + x^2 + 1$, which describes a genus 2 curve $C$. Diophantus provides the solution $(1/2, 9/8)$ and essentially asks whether there are any other positive rational solutions. In his 1998 Ph.D. thesis, Joe Wetherell used the method of Chabauty–Coleman to prove that there are no other solutions, thus resolving Diophantus’ ancient problem. In 2007, Poonen, Schaefer, and Stoll [@PoonenSchaeferStoll] found all primitive integral solutions to the generalized Fermat equation $x^2 + y^3 = z^7$, combining the method of Chabauty–Coleman with the Frey–Serre–Ribet approach to Fermat’s Last Theorem and the Wiles–Taylor modularity theorem. The complete list of solutions with $xyz \neq 0$ is $$% (\pm 1,-1,0),( \pm 1,0,1), \pm (0,1,1),
(\pm 3,-2,1),(\pm 71,-17,2),(\pm 2213459,1414,65),(\pm 15312283,9262,113),(\pm 21063928,-76271,17).$$
More recently, Michael Stoll developed a refinement of the method of Chabauty–Coleman, showing that there is a bound depending only on $g$ and $[K:Q]$ for the number of $K$-rational points on a hyperelliptic curve $C$ of genus $g$ over a number field $K$ with the property that the Mordell–Weil rank of its Jacobian is at most $g-3$. And in another recent preprint, Poonen and Stoll prove that for $g \geq 3$, a positive fraction of hyperelliptic curves of odd degree $2g+1$ over ${\mathbf Q}$ have only one rational point, the point at infinity. They also prove a lower bound on this fraction that tends to 1 as the genus tends to infinity. Their method combines a refinement of the Chabauty–Coleman method, based on an idea of McCallum, with the Bhargava–Gross equidistribution theorem for nonzero 2-Selmer group elements. So the method of Chabauty–Coleman is alive and well.[^10]
To round off our discussion of Coleman integration, we mention that it has deep connections with $p$-adic Hodge theory which have presumably only begun to be understood. One of the first such connections was made by Coleman himself in his 1984 paper “Hodge–Tate periods and $p$-adic abelian integrals” [@Coleman_HodgeTatePeriods]. In that work, Coleman gives a new description of one of the maps involved in the Hodge–Tate decomposition of an abelian variety $A$ with good reduction. Using this, he provides a formula for the Hodge–Tate periods[^11] involving $p$-adic Abelian integrals on $A$. Pierre Colmez built on this work in [@Colmez_Periods], constructing $p$-adic periods for differentials of the second kind on arbitrary Abelian varieties $A$ over a local field $K$ with values in Fontaine’s period ring $B^+_{\rm dR}$ via $p$-adic integration.
Coleman continued to elucidate connections between $p$-adic integration and $p$-adic Hodge theory throughout his career. For example, in the 1999 paper “The Frobenius and monodromy operators for curves and Abelian varieties” [@ColemanIovita_FrobeniusMonodromy], written with Adrian Iovita, they show that if $A$ is an abelian variety over a local field $K$ having split semistable reduction, then the $p$-adic integration pairing $T_p(A) \times H^1_{\rm dR}(A) \to B^+_{\rm dR}$ defined by Colmez in [@Colmez_Periods] induces a natural isomorphism between the integral part of $H^1_{\rm dR}(A)$ and Fontaine’s “monodromy module” $D_{\rm st}(\Hom(V_p(A),\Q_p)$ which is compatible with the actions of the Frobenius and monodromy operators on both sides, as well as with the natural filtrations. (Here $T_p(A)$ is the $p$-adic Tate module of $A$ and $V_p(A)=T_p(A) \otimes_{\Z_p} \Q_p$.) As a concrete corollary, Coleman and Iovita obtain a proof of Fontaine’s conjecture that if $A$ is an Abelian variety over a local field $A$, then $T_p(A)$ is crystalline if and only if $A$ has good reduction.
Coleman families {#eigen}
================
The classical work of Ramanujan on arithmetic properties of the Fourier coefficients of modular forms unearthed striking congruences which contain important number-theoretic information (see [@Ramanujan_Congruences], as well as the discussion in [@Serre_Ramanujan]). Perhaps the most famous of these is his congruence modulo the prime $691$ relating the Fourier coefficients of the cuspidal modular form $$\Delta(q)\ =\ q\prod_{n=1}^{\infty}(1-q^n)^{24}\ =\ \sum_{n=1}^{\infty}\tau(n)q^n$$ to the Fourier coefficients of the Eisenstein series $$E_{12}(q):= {\frac{1}{2}}\zeta(-11) \ + \ \sum_{n=1}^{\infty}\{\sum_{d \mid n}d^{11}\}q^n.$$ Concretely, Ramanujan discovered that, for every positive integer $n$, $\tau(n)$ is congruent modulo 691 to the sum of the eleventh powers of all positive divisors of $n$.
The Fourier coefficients of the family of Eisenstein series of even integral weights $k =4,6,\dots,$ given explicitly by $$E_{k}(q):= {\frac{1}{2}}\zeta(1-k) \ + \ \sum_{n=1}^{\infty}\{\sum_{d\mid n}d^{k-1}\}q^n,$$ already reveal a wealth of congruences between members of that family of different weights. Indeed, for any odd prime $p$, if $k \equiv k' \pmod{p-1}$ then Fermat’s Little Theorem tells us that the non-constant Fourier coefficients of $E_{k}(q)$ are congruent mod $p$ to the corresponding non-constant Fourier coefficients of $ E_{k'}(q)$, and the classical Kummer Congruence for Bernoulli numbers shows that the constant terms are congruent mod $p$ as well. An appropriate elaboration of this observation allows us to construct a family of “$p$-adic Eisenstein series" that vary over a $p$-adic analytic parameter space corresponding to their $p$-adic weights. The Fourier coefficients in this family vary $p$-adically analytically as a function of the weight, and for any $p$-adic weight $k$ that happens to be a rational integer at least 2, the member of the family of that weight $k$ is an honest classical Eisenstein series of weight $k$. Any two members of this family whose weights are appropriately close in this $p$-adic parameter space will have the property that their corresponding Fourier coefficients will be congruent modulo an appropriate power of $p$.
The work of Haruzo Hida (see e.g. [@Hida_GaloisReps; @Hida_OrdinaryHecke; @Hida_Book]) was a major step forward in structurally organizing the “wealth of congruences” we’ve referred to. Consider modular eigenforms on $\Gamma_0(pN)$ with $N$ not divisible by $p$. Such a modular form is said to be of [*slope zero*]{} if its eigenvalue for the Atkin–Lehner operator $U_p$ is a $p$-adic unit. Hida showed that any such [*cuspidal*]{} eigenform fits into a $p$-adically varying family of [*$p$-adic*]{} cuspidal modular eigenforms on $\Gamma_0(pN)$ (with nebentypus character depending on the weight) that have similar properties to the $p$-adic family of Eisenstein series described above. That is, the Fourier coefficients in Hida’s families vary $p$-adically analytically over their parameter space, and for integral weights $k\ge 2$ the members of the family are classical cuspidal modular eigenforms. These [*Hida families*]{} are of finite degree over their parameter weight space, and provide congruences modulo powers of $p$ for $p$-adically nearby members. Again, there are a wealth of such congruences.
The existence of such “$p$-adic continuity” among cuspidal eigenforms might seem more surprising than in the Eisenstein family, in that Eisenstein series over the complex numbers fit into a [*continuous spectrum*]{} while cuspidal modular eigenforms are in the discrete spectrum in the classical harmonic analysis of the $L^2$ space of functions on the upper half plane relative to the action of an appropriate congruence subgroup of ${\rm SL}_2({\Z})$. The $p$-adic setting, then, offers a type of coherence—via congruences modulo powers of $p$—that the classical one does not.
Robert Coleman’s great contribution in [@Coleman_OverconvergentModularForms; @Coleman_BanachSpaces] was to extend the construction of such $p$-adic families to cuspidal eigenforms on $\Gamma_0(pN)$ that are merely assumed to be of finite slope; that is, not in the kernel of the $U_p$-operator. [*Coleman families*]{} have all the properties—save one—that Hida families have. That one difference, though, is crucial, and is why Coleman’s contribution required a new vision, new techniques, and a strikingly original application of those techniques. Coleman families—in contrast with the Eisenstein families (which are of degree one over the parameter weight space) or Hida families (which are algebraic, of finite degree over the weight space)—project only $p$-adically analytically to the parameter weight space. Coleman obtains his families by performing a Fredholm analysis in a family of (infinite-dimensional) $p$-adic Banach spaces of $p$-adic modular forms (over $p$-adic weight space). To achieve this was a real advance, and it led to the construction in [@ColemanMazur_Eigencurve] of the [*eigencurve*]{}, a curve which is (very likely) of infinite degree over the weight space that parametrizes all $p$-adic eigenforms of finite slope. The eigencurve is nowadays a basic tool for understanding the arithmetic of modular forms.
Concluding remarks
==================
Robert Coleman was a kind, brave, and brilliant man whose influence on mathematics and on his friends and loved ones will long outlive his fragile body. He focused on what is important in mathematics and in life, from beginning to end, and had the imagination, originality, and courage to work on problems which lay just at the edge of the possible. He was unafraid to dream the great dreams of his subject.
{width=".4\textwidth"}
A conference on $p$-adic Methods in Number Theory will take place in Berkeley, CA in May 2015 in honor of Robert’s mathematical legacy.
[^1]: In the early twentieth century, Hensel and others introduced, for each prime number $p$, a field $\Q_p$ of $p$-adic numbers which serves as an “arithmetic” analogue of the real numbers. There is also a larger algebraically closed field $\C_p$ corresponding to the complex numbers. It was soon realized that not only do $p$-adic numbers have many important applications, but many facts from real and complex analysis and complex algebraic geometry have quite precise $p$-adic analogues.
[^2]: This was long before the invention of Berkovich spaces, which provide a natural way to embed $\C_p$ in a path-connected space.
[^3]: Every personal reminiscence in this article whose source is not identified is drawn from the comments section in the first author’s blog post <http://mattbakerblog.wordpress.com/2014/03/25/robert-f-coleman-1954-2014/>
[^4]: We thank Janet Abelson, Mayor of El Cerrito, CA, for briefing us on this topic.
[^5]: Fontaine associates to any $p$-adic representation $V$ of the absolute Galois group of a local field $K$ of characteristic zero a so-called $(\phi,\Gamma)$-module, which is a certain module $D(V)$ equipped with a Frobenius operator $\phi$, a left inverse of $\phi$, and a continuous action of the $\Z_p$-cyclotomic Galois group of $K$ commuting with the action of $\phi$. This gives an equivalence between suitable categories of Galois representations and modules, and translates the study of $p$-adic representations V into the “linear algebra” of its corresponding module $D(V)$.
[^6]: In the Acknowledgments to the paper, Coleman thanks Bernard Dwork “who showed us that although one Frobenius may be best, all are good.”
[^7]: The first such integral to appear in the literature was Kummer’s $p$-adic logarithm, which he used in his work on explicit reciprocity laws and Fermat’s Last Theorem.
[^8]: A meromorphic one-form on a variety $V$ is called a [*differential of the second kind*]{} if it is the sum of a holomorphic one-form and an exact one-form.
[^9]: David Grant gives the example $y^2=x(x-1)(x-2)(x-5)(x-6)$ with $p=7$.
[^10]: Minhyong Kim has developed a partly conjectural non-abelian analogue of the method of Chabauty–Coleman which appears to hold great promise for the future; it is related to iterated Coleman integrals and the Grothendieck section conjecture.
[^11]: A curious feature of $p$-adic Hodge theory, which may be surprising to the uninitiated, is that the numbers which one routinely calls “periods” are almost never defined via integration of differential forms.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Evidence for the energy threshold of creating the quark-gluon plasma in nucleus-nucleus collisions, the so-called onset of deconfinement, has been found by the energy scan program of the NA49 experiment at the CERN SPS. In this paper we review the experimental and theoretical status of this phenomenon. First, the basic, qualitative ideas are presented for non-experts. Next, the latest experimental results are compared to a statistical model within which the onset of deconfinement and its signals had been predicted. Finally, alternative interpretations and open questions are discussed.'
---
[**Onset of deconfinement in nucleus-nucleus collisions:**]{}
[**Review for pedestrians and experts**]{}
Marek Gazdzicki$^{a,b}$, Mark Gorenstein$^{c,d}$ and Peter Seyboth$^{e,b}$
[*$^{a}$Institut für Kernphysik, University of Frankfurt, Frankfurt, Germany\
*]{} [*$^{b}$Jan Kochanowski University, Kielce, Poland\
*]{} [*$^{c}$Bogolyubov Institute for Theoretical Physics, Kiev, Ukraine\
*]{} [*$^{d}$Frankfurt Institute for Advanced Studies, Frankfurt, Germany\
*]{} [*$^{e}$Max-Planck-Institut fuer Physik, Munich, Germany\
*]{}
Introduction
============
One of the important issues of contemporary physics is the understanding of strong interactions and in particular the study of the properties of strongly interacting matter in equilibrium. What are the phases of this matter and what do the transitions between them look like ? These questions motivate broad experimental and theoretical efforts since more than 40 years. The study of high energy collisions between two atomic nuclei give us the unique possibility to address these issues in well controlled laboratory experiments. In particular, the advent of the quark model of hadrons and the development of the commonly accepted theory of strong interactions, quantum chromodynamics (QCD), naturally led to expectations that matter at very high densities may exist in a state of quasi-free quarks and gluons, the quark-gluon plasma (QGP) [@qgp0; @qgp1; @qgp2].
Experimental searches for QGP signals started at the Super Proton Synchrotron (SPS) of the European Organization for Nuclear Research (CERN) and the Alternating Gradient Synchrotron (AGS) of Brookhaven National Laboratory (BNL) in the mid 1980s. Today they are pursued also at much higher collision energies at the Relativistic Heavy Ion Collider (RHIC) at BNL. Soon experiments on nucleus-nucleus collisions at the Large Hadron Collider (LHC) in CERN will join the world effort at energies 20 times higher than at RHIC. Most probably that the QGP is formed at the early stage of heavy ion collisions at the top SPS energy and at RHIC energies. Unambiguous evidence of the QGP state was however missing. This may be attributed to the difficulty of obtaining unique and quantitative predictions of the expected QGP signals from the theory of strong interactions.
For this reason the NA49 Collaboration at the CERN SPS has searched over the past years for signs of the onset of QGP creation in the energy dependence of hadron production properties. This search was motivated by a statistical model [@GaGo] showing that the onset of deconfinement should lead to rapid changes of the energy dependence of numerous experimentally detectable properties of the collisions, all appearing in a common energy domain. The predicted features have recently been observed [@evidence] and dedicated experiments now continue detailed studies in the energy region of the onset of deconfinement.
It is thus time for a summary. In this paper we review the experimental and theoretical status of the onset of deconfinement. First, the basic qualitative ideas are presented for non-experts. Next, a quantitative model within which the onset of deconfinement and its signals were predicted is reexamined and compared with the latest experimental results. Finally, alternative interpretations and open questions are discussed.
Onset of Deconfinement for Pedestrians
======================================
Phase transitions are fascinating physical phenomena. Small changes in temperature or pressure lead to dramatic changes in macroscopic properties of matter. Common examples from our daily life are transitions between solids, liquids and gases like boiling and freezing of water.
![\[water\_phases\] Phases of water. When adding heat (energy) at constant pressure water is transformed from solid to liquid and then from liquid to vapor as indicated by the dashed arrow. ](water_phases_3.pdf){width="1.0\linewidth"}
The well known phase diagram of water is shown in Fig. \[water\_phases\], where the regions of existence of the various phases of water are depicted in a diagram of pressure and temperature. When adding heat to water one increases its temperature moving through its different phases and crossing their boundaries, as indicated by the dashed arrow in Fig. \[water\_phases\] for the example of constant atmospheric pressure. Dependence of the water temperature on the amount of added heat, called the heating curve of water, is shown in Fig. \[water\_heating\_curve\]. In pure phases, such as ice, water or vapor, the temperature increases monotonically with added heat. The two regions of constant temperature (steps) signal the ice-water and water-vapor phase transitions. In these mixed phase regions added heat is used for the phase transformation instead of the increase of temperature as in the pure phase regions.
![\[water\_heating\_curve\] Heating curve of water at fixed atmospheric pressure. It corresponds to the trajectory in the phase diagram of water indicated by the dashed arrow in Fig. \[water\_phases\]. ](water_heating_curve.pdf){width="1.0\linewidth"}
Properties of water and other substances surrounding us and the transitions between their various phases are determined by electromagnetic interactions of atoms and molecules. On the other hand, the properties of atomic nuclei which are built from nucleons (protons and neutrons) are determined by strong interactions. Naturally, the question arises whether strongly interacting matter also exists in distinct phases. What are their properties? At which temperatures do the transitions between them take place? What do these transitions look like?
Since more than 40 years it is known that hadrons (i.e. mesons and baryons; all strongly interacting particles observed in nature are called hadrons) consist of more elementary particles, the quarks and gluons. However, isolated quarks or gluons were never observed. They seem to be always confined in the interior of hadrons. But could a different phase of strongly interacting matter exist in which quarks and gluons are deconfined ?
There are 3 parameters which describe the thermodynamical properties of a system. In non-relativistic systems they are temperature, particle number density, and pressure. The equation of state connects them, e.g., the pressure is a well defined function of temperature and particle density for a specific substance. In experiments on water one can most easily fix temperature and pressure to define the point on the phase diagram in Fig. \[water\_phases\]. Unlike in water, the number of particles is not conserved in strongly interacting relativistic matter. Instead of particle number density the baryonic number, i.e. the difference between the number of baryons and anti-baryons, is conserved. In calculations it is convenient to use the equivalent variables baryonic number density or baryonic chemical potential. The phase diagram of strongly interacting matter emerging from theoretical considerations and experimental results is shown in Fig. \[sim\_phases\] in terms of the commonly used variables temperature and baryonic chemical potential. Laboratory experiments (see discussion below) can create strongly interacting matter with different temperatures $T$ and baryonic chemical potentials $\mu_B$. The functional dependence of the pressure on $T$ and $\mu_B$, i.e. the equation of state of strongly interacting matter, remains the subject of intensive experimental and theoretical studies.
![\[sim\_phases\] Phases of strongly interacting matter. With increasing collision energy the matter created at the early stage of nucleus-nucleus collisions changes its properties as indicated by the arrow. At low energy it is in the confined phase (hadrons), at sufficiently high energy in the deconfined phase (QGP). $M$ is the critical point of the nuclear liquid-gas phase transition. The shaded band shows the 1$^{st}$ order phase boundary between the hadron and QGP phase which is expected to end in a critical end point $E$. At $E$ the sharp phase transition turns into a rapid crossover indicated as the dotted line. ](sim_phases_6.pdf){width="1.0\linewidth"}
A transition between the deconfined and the confined phase of strongly interacting matter probably took place during the expansion and cooling of the early Universe, about 1 microsecond after the Big Bang. Cosmological signatures of this transition are difficult to identify today. However, extremely dense strongly interacting matter fills the interior of neutron stars. Arguments in favor of the existence of quark matter in the center of such stars were advanced already in the 1960s, soon after formulation of the quark hypothesis. One of the pioneering papers [@qgp1] argued: “A neutron has a radius of about 0.5–1 fm (1 fm = 10$^{-15}$ m), and so has a density of about 8$\cdot$10$^{14}$ g$\cdot$cm$^{-3}$, whereas the central density of a neutron star can be as much as $10^{16}-10^{17}$ g$\cdot$cm$^{-3}$. In this case, one must expect the hadrons to overlap, and their individuality to be confused. Therefore, we suggest that matter at such densities is a quark soup.” The creation of matter in a deconfined phase, i.e. in the QGP phase, may be the only possibility to ‘see’ quarks and gluons moving freely in a large volume. Cosmological and astrophysical objects with the required properties are, unfortunately, difficult to investigate. Systematic study of the properties of strongly interacting matter requires a method to create it under well controlled conditions in the laboratory. The study of collisions of two heavy nuclei gives us this possibility. Such a collision produces a droplet of strongly interacting matter of high energy density, the so-called fireball. It is natural to expect that with increasing collision energy the fireball energy density also increases. Thus like in the case of water heating and observing successive transitions between its phases, we hope that with increasing collision energy we can detect anomalies in the energy dependence of hadron production properties and thus discover successive transitions between various phases of strongly interacting matter created at the early stage of collisions. The arrow in Fig. \[sim\_phases\] schematically traces the position of the initially created fireball on the phase diagram when the energy of nucleus-nucleus collisions is increasing. At sufficiently high collision energy this matter droplet may reach the QGP phase (see Fig. \[sim\_phases\]). Unfortunately the life time of the fireball is very short, about $10^{-22}$ seconds. It quickly expands, cools down (see Fig. \[sim\_phases1\]) and finally decays into hadrons and a few light nuclei. These decay products are measured in detectors surrounding the collision point.
![\[sim\_phases1\] Parameters of strongly interacting matter created at the early stage of nucleus-nucleus interactions are shown by the full circles for central Pb+Pb (Au+Au) collisions at the top AGS energy ($\sqrt{s}_{NN} \approx 5.5$ GeV), intermediate SPS energy ($\sqrt{s}_{NN} \approx 7.6$ GeV) and top SPS energy ($\sqrt{s}_{NN} \approx 17$ GeV). The created fireball expands and cools along trajectories indicated by solid lines and decouples at the freeze-out points (full squares, triangles and star). ](sim_phases+datapoints+trajectories_v1.png){width="1.0\linewidth"}
The first phase transition of strongly interacting matter was observed studyingg collisions at very low energies [@liquide] (the energy per nucleon-nucleon pair in the center of mass system $\sqrt{s}_{NN} < 2$ GeV). This transition between a nuclear liquid and a nuclear gas happens at a temperature of about 6$\cdot$10$^{10}$ K (5 MeV). The phase transition line and critical point $M$ lie at large $\mu_B$ and small $T$ inside the hadron phase region of the phase diagram as shown in Figs. \[sim\_phases\] and \[sim\_phases1\].
Emerging results from the study of high energy collisions of nuclei confirm the existence of the second phase transition in strongly interacting matter which was suggested by QCD. It is the so–called deconfinement phase transition. Let us briefly present how one looked for this phase transition, and explain the main results.
![\[sim\_heating\_curve\] Heating curves of strongly interacting matter. Hadron production properties (see text for details) are plotted as a function of collision energy for central Pb+Pb (Au+Au) collisions (upper set of points) and p+p interactions (lower set of points) [@evidence] ](kpip_4pi_data.pdf "fig:"){width="0.5\linewidth"} ![\[sim\_heating\_curve\] Heating curves of strongly interacting matter. Hadron production properties (see text for details) are plotted as a function of collision energy for central Pb+Pb (Au+Au) collisions (upper set of points) and p+p interactions (lower set of points) [@evidence] ](kap_slope_sqrts.pdf "fig:"){width="0.5\linewidth"}
The search for the phase of deconfined strongly interacting matter, the QGP, already has a long history. It received a boost from the first acceleration of oxygen and sulfur nuclei at the CERN SPS in 1986 ($\sqrt{s}_{NN} \approx 20$ GeV) and of lead nuclei in 1994 ($\sqrt{s}_{NN} \approx 17$ GeV). Measurements from an array of experiments indicated that the critical energy density was probably exceeded and matter with unusual properties appeared to be formed in the early stage of the collisions [@cernpr]. A key problem was the identification of experimental signatures of QGP creation. Several signatures of the formation of a transient QGP state during the early stage of the collision had been proposed in the past [@Rafelski; @Satz]. However, the uniqueness of these signatures came under renewed scrutiny and they were found not to be specific for the creation of QGP (see Appendix 8.1 for details).
In the mid 1990s a study of results from experiments at CERN and the AGS at BNL (maximum energy $\sqrt{s}_{NN} \approx 5.5$ GeV) raised [@GaRo1; @GaRo2] intriguing questions concerning the energy dependence of hadron production between top AGS and SPS energies. In response to these questions a statistical model of the early stage of the collision process was proposed [@GaGo] in which an equation of state with a 1$^{st}$ order phase transition was assumed. In this model the onset of deconfinement led to the prediction of a non-monotonic collision energy dependence of several hadron production properties. In particular, the model predicted a sharp maximum in the ratio of multiplicities of strange hadrons (the hadrons which contain $s$ and $\overline{s}$ quarks) to pions (the lightest hadron) at the beginning of the transition region, at about $\sqrt{s}_{NN}
\approx 7.5$ GeV. This prediction triggered an extension of the experimental program at the SPS, the energy scan program [@na49scan]. Within this program head–on (central) collisions of two lead nuclei (Pb+Pb) were registered at several lower SPS energies ($\sqrt{s}_{NN} = $ 6.3, 7.6, 8.7 and 12.3 GeV) by the NA49 experiment. Other heavy ion experiments at the SPS (NA45, NA50, NA57 and NA60) participated in selected runs of this program [@othersinscan]. Final results, obtained mainly by the NA49 collaboration, confirm the qualitative expectations and the quantitative predictions of the model: rapid changes in properties of hadron production occur within a narrow energy range, $\sqrt{s}_{NN}=$ 7–12 GeV [@evidence].
The most dramatic effect is seen in the energy dependence of the ratio of total particle yields of kaons and pions, $\langle K^+
\rangle/\langle \pi^+ \rangle$, in central Pb+Pb collisions which is plotted in Fig. \[sim\_heating\_curve\] $left$. Following a fast threshold rise the ratio passes through a sharp maximum in the SPS range and then seems to settle to a plateau value at higher energies. Kaons are the lightest strange hadrons and due to approximate isospin symmetry the $\langle K^+ \rangle$ yield counts about half of the strange (anti-)quarks produced in the collisions and contained in the reaction products (see Appendix 8.2 for details). Thus Fig. \[sim\_heating\_curve\] $left$ demonstrates that the fraction of strangeness carrying particles in the produced matter passes through a sharp maximum at the SPS in nucleus–nucleus collisions. This feature is not observed for proton–proton reactions as shown by the open dots in Fig. \[sim\_heating\_curve\] $left$.
A second important result is the stationary value of the apparent temperature $T$ of $K^+$ mesons in central Pb+Pb collisions at SPS energies as shown in Fig. \[sim\_heating\_curve\] $right$. In the fireball picture the apparent temperature is related to the local random motion of the particles and their collective expansion velocity in the direction transverse to the collision axis.
Presently the sharp maximum and the following plateau in the energy dependence of the $\langle K^+ \rangle/\langle \pi^+
\rangle$ ratio has only been reproduced by the statistical model of the early stage. In this model it reflects the decrease in the ratio of strange to non-strange number of degrees of freedom when deconfinement sets in. The stationary value of the apparent temperature of $K^+$ mesons was predicted [@van-hove; @Hu:95; @GoGaBu] as a consequence of the constant pressure and temperature at the early stage of nucleus–nucleus collisions in the SPS energy range due to the coexistence of hadronic and deconfined phases.
[**These results serve as evidence that the deconfinement phase transition in Pb+Pb collisions starts in the SPS energy range.**]{} The exciting and rich physics which can be studied in nucleus-nucleus collisions at the CERN SPS energies motivates ongoing and future experimental programs at the CERN SPS [@Gazdzicki:2006fy; @proposal], BNL RHIC [@rhic_low], FAIR SIS [@cbm] and JINR NICA [@mpd].
Statistical Model of the Early Stage (SMES)
===========================================
The experimental search for the onset of deconfinement performed by the experiment NA49 at the CERN SPS was motivated by the predictions of the Statistical Model of the Early Stage [@GaGo], which treats the creation of the fireball in nucleus–nucleus collisions in a statistical model approach. The model does not attempt a description of the subsequent (hydro–)dynamical evolution of the fireball. First, this section reviews the main assumptions of the originally formulated model and presents results obtained from analytical and numerical calculations as well as comparison to the experimental data available 12 years ago, when the model was formulated. Next, later extensions of the SMES are discussed which address collective flow at freeze-out [@GoGaBu; @bleicher] and event-by-event fluctuations [@GaGoMo; @GoGaZo]. The subsequent Sec. 4 shows the comparison of the predictions of the model with the most recent experimental results.
Main Assumptions
----------------
1\. The basic assumption of the SMES is that the production of new degrees of freedom during the early stage of A+A collisions is a statistical process. Thus formation of all microscopic states allowed by conservation laws is equally probable. As particle creation from energy does not produce net charges, only states with total baryon, flavor and electric charge quantum numbers equal to zero are considered. Presence of the colliding nucleons is assumed to affect the properties of the observed final state only via their interactions with the statistically produced particles during the expansion of the system. This issue is further discussed in point 5 below. Consequently, the properties of the state produced at the early stage are entirely defined by the available energy and the volume in which production takes place. In central A+A collisions this volume is chosen as the Lorentz contracted volume occupied by the colliding nucleons (participant nucleons) from a single nucleus: $$\label{volume}
V = \frac {V_0} {\gamma}~,$$ where $V_0 = \frac{4}{3} \pi r_0^3 A_p$ and $\gamma = \sqrt{s}_{NN}/(2
m_N)$, $m_N$ is the nucleon mass and $A_p$ is the number of participant nucleons from a single nucleus. The $r_0$ parameter is taken to be 1.30 fm in order to fit the mean baryon density in the nucleus, $\rho_0 = 0.11$ fm$^{-3}$.
2\. Only a fraction, $\eta$, of the total energy in A+A collision is transformed into the energy of new degrees of freedom created in the early stage. This is because a part of the energy is carried by the net baryon number which is conserved during the collision. The released (inelastic) energy can be expressed as: $$\label{energyin}
E = \eta (\sqrt{s}_{NN} - 2 m_N)~A_p~.$$ The parameter $\eta$ is assumed to be independent of the collision energy and the system size for A+A collisions. The value of $\eta$ used for the numerical calculations is 0.67 [@Ba:94].
3\. The elementary particles of strong interactions are quarks and gluons. The deconfined state is considered to be composed of $u$, $d$ and $s$ quarks and the corresponding anti-quarks each with internal number of degrees of freedom equal to 6 (3 color states $\times$ 2 spin states). The contribution of $c$, $b$ and $t$ quarks can be neglected due to their large masses. The internal number of degrees of freedom for gluons is 16 (8 color states $\times$ 2 spin states). The masses of gluons and non-strange (anti-)quarks are taken to be 0, the strange (anti-)quark mass is taken to be 175 MeV [@Le:96]. The properties of equilibrated matter is characterized by an equation of state (EoS). For the case of colored quarks and gluons the model assumes the ideal gas EoS modified by a bag constant $B$ (see, e.g., [@qgp2; @bag]): $$p = p^{id} - B~,~~~
\varepsilon = \varepsilon^{id} + B~,$$ where $p$ and $\varepsilon$ denote pressure and energy density, respectively, and the superscript $^{id}$ marks the quantities for the ideal gas. This equilibrium state is called the Quark Gluon Plasma or Q–state.
4\. The model uses an effective parametrization of the confined state, denoted as W–state (White–state). The non-strange degrees of freedom which dominate the entropy production are taken to be massless bosons, as suggested by the original analysis of entropy production in N+N and A+A collisions [@Ga:95-97]. Their internal number of degrees of freedom was fitted to the data [@Ga:95-97] and appeared to be about 3 times lower than the internal number of effective degrees of freedom in the QGP. The internal number of degrees of freedom for a QGP is 16 + (7/8)$\cdot$36 $\cong$ 48 and therefore the internal number of non-strange degrees of freedom for low energy collisions is taken to be 48/3 = 16. The mass of strange degrees of freedom is assumed to be 500 MeV, equal to the kaon mass. The internal number of strange degrees of freedom is estimated to be 14 as suggested by the fit to the strangeness and pion data at the AGS. Also for the W-state the ideal gas EoS is selected. Clearly, this description of the confined state should only be treated as an effective parametrization. The numerical parameters are fixed by fitting A+A data at the AGS and the parametrization is then used for extrapolation to higher collision energies where the transition between the confined and deconfined state is expected.
5\. It is assumed that the matter created at the early stage expands, hadronizes and freezes-out. Within the original SMES formulation these later stage were not modelled. It was, however, postulated that during these stages the total entropy and total number of $s$ and $\overline{s}$ quarks created in the early stage are conserved. The only process which changes the entropy content of the produced matter during the expansion is assumed to be the interaction with the baryonic subsystem. It was argued that this leads to an entropy transfer to baryons which corresponds to the effective absorption of about 0.35 $\pi$–mesons per baryon [@GaGoMo:98]. Thus the final hadronic state has non–zero baryonic number and electric charge.
Analytical Formulas
-------------------
In this section the simplified version of the model (massless particles) will be discussed which allows to perform calculations analytically. Subsequently in Sec. 4 quantitative results from numerical calculations using finite masses will be presented and compared to measured data. All chemical potentials have to be equal to zero, as only systems with all conserved charges equal to zero are considered. Thus the temperature $T$ remains the only independent thermodynamical variable. It is convenient to define the EoS in terms of the pressure function $p=p(T)$ as the entropy and energy densities can be calculated from the thermodynamical relations: $$\label{termid}
s(T)~=~\frac{dp}{dT}~,~~~~\varepsilon (T)~=~T\frac{dp}{dT}~-~p~.$$ In the case of an ideal gas the pressure of the particle species ‘$j$’ is given by: $$\label{pressi}
p^j(T)~=~\frac{g^j}{2\pi^2}\int_{0}^{\infty}k^2dk~\frac{k^2}
{3(k^2+m_j^2)^{1/2}}~\left[\exp\left(\frac{\sqrt{k^2+m_j^2}}{T}\right)
~\pm ~1\right]^{-1}~,$$ where $g^j$ is the internal number of degrees of freedom (degeneracy factor) for the $j$–th species, $m_j$ is the mass of the particle, ‘–1’ appears in Eq. (\[pressi\]) for bosons and ‘+1’ for fermions. The pressure $p(T)$ for an ideal gas of several particle species is additive: $p(T)=\sum_j p^j(T)$. The same is valid for the entropy and energy densities of Eq. (\[termid\]).
In order to be able to perform analytical calculations of the system entropy and illustrate the model properties it is assumed that all degrees of freedom are massless. In this simplified case the pressure Eq. (\[pressi\]) is equal to: $$\label{presso}
p^j(T)~=~\frac{\sigma^j}{3}~T^4~,$$ where $\sigma^j$ is the so called Stephan–Boltzmann constant, equal to $\pi^2g^j/30$ for bosons and $\frac{7}{8}\pi^2 g^j/30$ for fermions. The total pressure in the ideal gas of several massless species can then be written as $p(T)=\pi^2 g T^4/90$ with the effective number of degrees of freedom $g$ given by $$\label{g}
g~=~g^{b}+\frac{7}{8}~g^{f},$$ where $g^b$ and $g^f$ are internal degrees of freedom of all bosons and fermions, respectively. The $g$ parameter is taken to be $g_W$ for the W–state and $g_Q$ for the Q–state, with $g_Q > g_W$.
The pressure, energy and entropy densities then follow as: $$\label{wmatter}
p_W(T)=\frac{\pi^2g_W}{90}~T^4~,~~~
\varepsilon_W(T)=\frac{\pi^2g_W}{30}~T^4~,~~~
s_W(T)=\frac{2\pi^2g_W}{45}~T^3~,\\$$ $$\label{qmatter}
p_Q(T)=\frac{\pi^2g_Q}{90}~T^4-B~,~~~
\varepsilon_Q(T)=\frac{\pi^2g_Q}{30}~T^4+B~~~,~
s_Q(T)=\frac{2\pi^2g_Q}{45}~T^3~,$$ for the pure W– and Q–state, respectively. Note the presence of the non–perturbative bag terms in addition to the ideal quark–gluon gas expressions for the pressure and energy density of the Q–state.
The 1$^{st}$ order phase transition between W– and Q–state is defined by the Gibbs criterion $$\label{ptr}
p_W(T_c)~=~p_Q(T_c)~,$$ from which the phase transition temperature can be calculated as: $$\label{tcr}
T_c~=~\left[\frac{90B}{\pi^2(g_Q-g_W)}\right]^{1/4}~.$$ At $T=T_c$ the system is in the [*mixed*]{} phase with the energy and entropy densities given by $$\label{mixed}
\varepsilon_{mix}=(1-\xi)\varepsilon_W^c~+~\xi
\varepsilon_Q^c~,~~~~
s_{mix}=(1-\xi)s^c_W~+~\xi s_Q^c~,$$ where $(1-\xi)$ and $\xi$ are the relative volumes occupied by the W– and Q–state, respectively. From Eqs. (\[wmatter\], \[qmatter\]) one finds the energy density discontinuity (‘latent heat’) $$\label{lheat}
\Delta \varepsilon ~\equiv ~ \varepsilon_Q(T_c)-
\varepsilon_W(T_c)~\equiv ~\varepsilon_Q^c-
\varepsilon_W^c~=~4B~.$$
The early stage energy density is an increasing function of the collision energy and is given by (see Eqs. (\[volume\], \[energyin\])): $$\label{endensity}
\varepsilon ~ \equiv~ \frac {E}
{V} ~=~\frac{\eta~\rho_0~(\sqrt{s}_{NN}-2m_N)~\sqrt{s}_{NN}}{2 m_N}~.$$
![\[ept4\] Energy density and pressure divided by $T^4$ as a function of temperature $T$. The bag constant $B$ was adjusted to 600 MeV/fm$^3$ to obtain a critical temperature $T_c~=~200$ MeV. ](ept4.pdf){width="0.7\linewidth"}
There is a remarkable equivalence (see Appendix C of Ref. [@GaGo]) of the Gibbs criterion (i.e. the pure phase corresponds to the larger pressure $p_W$ or $p_Q$ and the mixed phase to equal pressures $p_W=p_Q$) and the maximum entropy criterion, $$\label{extremum}
s(\varepsilon)~=~\max~\{~s_W(\varepsilon),~
s_Q(\varepsilon),~s_{mix}(\varepsilon)~\}~,$$ for an arbitrary EoS $p=p(T)$ with a 1$^{st}$ order phase transition. For $\varepsilon
< \varepsilon_W^c$ or $\varepsilon > \varepsilon_Q^c$ the system consists of pure W– or Q–state, respectively, with entropy density given by the following equations: $$\label{entropyw}
s_W(\varepsilon)~=~\frac{4}{3} \left(\frac{\pi^2 g_W}{30}\right)^{1/4}~
\varepsilon^{3/4}~,$$ $$\label{entropyq}
s_Q(\varepsilon)~=~\frac{4}{3}\left(\frac{\pi^2 g_Q}{30}\right)^{1/4}~
(\varepsilon -B)^{3/4}~.$$ For $\varepsilon_W^c<\varepsilon<\varepsilon_Q^c$ the system is in the mixed phase (Eq. \[mixed\]) and its entropy density can be expressed as: $$\label{entropym}
s_{mix}(\varepsilon)~=~ \frac{\varepsilon_Q^cs_W^c - \varepsilon_W^cs_Q^c}
{4B}~+~\frac{s_Q^c - s_W^c}{4B}~\varepsilon~
\equiv~a~+~b~\varepsilon ~.$$ The ratio of the total entropy of the created state to the number of nucleons participating in A+A collisions is: $$\label{Entropy}
\frac{S}{2A_p}~=~ \frac{V~s}{2A_p}~=~\frac{
m_N~s}{\rho_0\sqrt{s}_{NN}}~,$$ and is independent of the number of participant nucleons. The entropy density $s$ in Eq. (\[Entropy\]) is given by the general expressions Eq. (\[extremum\]) with $\varepsilon$ defined by Eq. (\[endensity\]). For small $\sqrt{s}_{NN}$ the energy density Eq. (\[endensity\]) corresponds to that of the pure W–state and one finds $$\label{Entropyw}
\left(\frac{S}{2A_p}\right)_{W}~=~C~g_W^{1/4}~F~,$$ where $$\label{c}
C~=~\frac{2}{3}\left
(\frac{\pi^2m_N}{15\rho_0}\right)^{1/4}~\eta^{3/4}~,~~~~
F~ =~ \frac { (\sqrt{s}_{NN} - 2 m_N)^{3/4} } { (\sqrt{s}_{NN})^{1/4} }~.$$ Thus for low collision energies, where the W–state is created, the entropy per participant nucleon is proportional to $F$. For high $\sqrt{s}_{NN}$ the pure Q–state is formed and Eq. (\[Entropy\]) leads to $$\begin{aligned}
\label{Entropyq}
\left(\frac{S}{2A_p}\right)_Q~&=&~C~g_Q^{1/4}~F~\left(1~-~
\frac{2m_N B}{\eta \rho_0 (\sqrt{s}_{NN}-2m_N)\sqrt{s}_{NN}}
\right)^{3/4} \\
\nonumber
& \cong&~C~g_Q^{1/4}~F~\left(1-\frac{3m_NB}{2\eta\rho_0
F^4}\right)~.\end{aligned}$$ For large values of $F$ the entropy per participant nucleon in the Q–state is also proportional to $F$. The slope is, however, larger than the corresponding slope for the W–state by a factor $(g_Q/g_W)^{1/4}$. In the interval of $F$ in which the mixed phase is formed the energy dependence of the entropy per participant nucleon is given by: $$\label{Entropym}
\left(\frac{S}{2A_p}\right)_{mix}~=~ \frac{C_1}{\sqrt{s}_{NN}}~+~
C_2~(\sqrt{s}_{NN}-2 m_N)~,$$ where $$\label{c1}
C_1~=~\frac{m_N}{\rho_0}~a~,~~~~ C_2~=~\eta ~b~.$$ Equation (\[Entropym\]) gives approximately a quadratic increase with $F$ of the entropy per participant nucleon in the mixed phase region.
![\[temp\] [*Left:*]{} The early stage (initial) temperature of the fireball as a function of $F$. [*Right:*]{} The fraction of volume occupied by the QGP as a function of $F$. ](temp.pdf "fig:"){width="0.5\linewidth"} ![\[temp\] [*Left:*]{} The early stage (initial) temperature of the fireball as a function of $F$. [*Right:*]{} The fraction of volume occupied by the QGP as a function of $F$. ](ksi.pdf "fig:"){width="0.5\linewidth"}
Let us now turn to strangeness. The model defines $g^s_W$ and $g^s_Q$ as the numbers of internal degrees of freedom of (anti-)strangeness carriers in the W– and Q–state, respectively. The total entropy of the considered state is given by the sum of entropies of strange and non-strange degrees of freedom. Provided that all particles are massless the fraction of entropy carried by strange (and anti-strange) particles is proportional to the number of strangeness degrees of freedom: $$\label{strentr}
S_s ~=~ \frac {g^s} {g}~ S~.$$ Equation (\[strentr\]) is valid for both W– and Q–state. Note that all degeneracy factors are calculated according to the general relation Eq. (\[g\]). For massless particles of the $j$–th species the entropy is proportional to the particle number: $$S_j ~= ~4 N_j~.$$ Thus the number of strange and anti-strange particles can be expressed as $$N_s + N_{\overline{s}} ~=~ \frac {S}{4}~ \frac {g^s} {g}~,$$ and the strangeness to entropy ratio is equal to $$\label{strent}
\frac { N_s + N_{\overline{s}} } {S}
~ =~ \frac {1}{4}~ \frac {g^s} {g}~.$$ One concludes therefore that the strangeness to entropy ratio for the ideal gas of massless particles is dependent only on the ratio of strange to all degrees of freedom, $g^s/g$. This ratio is expected to be equal to $g_Q^s/g_Q\cong 0.22$ in the Q–state and $g_W^s/g_W\cong 0.5$ in the W–state. Consequently the phase transition from the W– to the Q–state should lead to a decrease of the strangeness to entropy ratio by a factor of about 2. This simple picture will be modified because of the large value of the mass of strange degrees of freedom in the W–state ($m_W^s\cong500$ MeV) compared to the temperature $T$. In this case the left hand side of Eq. (\[strent\]) is a strongly increasing function of $T$.
In order to demonstrate properties of the EoS the ratios of $\varepsilon/T^4$ and $p/T^4$ are plotted in Fig. \[ept4\] as functions of the temperature. The bag constant $B$ = 600 MeV/fm$^3$ was adjusted such that the resulting critical temperature $T_c$ is equal to 200 MeV.
Quantitative Calculations
-------------------------
We next turn to the results from numerical calculations based on the model using finite strangeness carrier masses. For the number of non-strange degrees of freedom $g_Q^{ns}$ and $g_W^{ns}$ one gets as in the simplified model: $$\label{gnonstrange}
g_Q^{ns}~=~2\cdot8~+~\frac{7}{8}\cdot 2\cdot 2\cdot 3\cdot 2 ~=~37~;
~~~~~~g_W^{ns}~=~16~.$$ The strange degrees of freedom are now considered to have realistic masses $m_Q^s$ and $m_W^s$. Equation (\[pressi\]) is used with $$\label{gstrange}
g_Q^s~=~2\cdot 2\cdot 3 ~=~12~,~~m_Q^s~\cong ~ 175~ \mbox{MeV}~;
~~~~~~ g_W^s~=~14~,~~m_W^s~\cong~ 500~ \mbox{MeV}~.$$ Note that there is no factor ‘7/8’ in the above expression for $g_Q^s$ as Eq. (\[pressi\]) with Fermi momentum distribution was taken. The contributions of strange degrees of freedom to the entropy and energy densities are calculated using the thermodynamical relations Eq. (\[termid\]).
A convenient variable to study collision energy dependence is the Fermi–Landau variable $F$ defined in Eq. (\[c\]). The dependence of the early stage temperature $T$ (initial temperature of the fireball) on $F$ in the SMES is shown in Fig. \[temp\] [*left*]{}. Outside the transition region $T$ increases approximately linearly with $F$. Inside the transition region $T$ is constant ($T = T_c$ = 200 MeV). The transition region begins at $F= 2.23$ GeV$^{1/2}$ ($p_{LAB} = 30A$ GeV) and ends at $F= 2.90$ GeV$^{1/2}$ ($p_{LAB} = 64A$ GeV). The fraction of the volume occupied by the Q–state, $\xi$, increases rapidly in the transition region, as shown in Fig. \[temp\] [*right*]{}.
![\[spb\] [*Left:*]{} The entropy per participant nucleon as a function of $F$ (solid line). Dashed line indicates the dependence obtained assuming that there is no transition to the QGP. [*Right:*]{} Ratio of produced entropy in pion units per participant nucleon, $\langle S_{\pi} \rangle/\langle N_P
\rangle$, as a function of $F$. Experimental data on central collisions of two identical nuclei are indicated by closed circles. These data correspond to the status of 1998 [@GaRo1; @GaRo2] and should be compared with the model predictions shown by the solid line. The open boxes show results obtained for nucleon–nucleon interactions. ](spb.pdf "fig:"){width="0.5\linewidth"} ![\[spb\] [*Left:*]{} The entropy per participant nucleon as a function of $F$ (solid line). Dashed line indicates the dependence obtained assuming that there is no transition to the QGP. [*Right:*]{} Ratio of produced entropy in pion units per participant nucleon, $\langle S_{\pi} \rangle/\langle N_P
\rangle$, as a function of $F$. Experimental data on central collisions of two identical nuclei are indicated by closed circles. These data correspond to the status of 1998 [@GaRo1; @GaRo2] and should be compared with the model predictions shown by the solid line. The open boxes show results obtained for nucleon–nucleon interactions. ](pipb.pdf "fig:"){width="0.5\linewidth"}
The number of non-strange and strange degrees of freedom and their masses are given by Eqs. (\[gnonstrange\], \[gstrange\]). They are different in the confined and deconfined phases. Thus, one expects abrupt changes of the pion multiplicity (entropy) (see Fig. \[spb\] [*left*]{}) and the multiplicity of strange particles (see Fig. \[str\] [*left*]{}) as a function of collision energy in the energy range where a transition from confined to deconfined matter takes place at the early stage of A+A collisions. The comparison of these predictions with experimental results is discussed in Secs. 4.1 and 4.2.
![\[str\] [*Left:*]{} The ratio of the total number of $s$ and $\overline{s}$ quarks and anti-quarks to the entropy (solid line) as a function of $F$. The dashed line indicates the corresponding ratio calculated assuming absence of the phase transition to the QGP. [*Right:*]{} The ratio $E_S$ of strange particle to pion production as a function of $F$. Experimental data on central collisions of two identical nuclei are plotted as closed circles. These data should be compared with the model predictions shown by the solid line. The open boxes show results obtained for nucleon–nucleon interactions, scaled be a factor 3.6 to match A+A data at AGS energy. The plotted data show the status of 1998 as compiled in [@GaRo1; @GaRo2]. ](str.pdf "fig:"){width="0.4\linewidth"} ![\[str\] [*Left:*]{} The ratio of the total number of $s$ and $\overline{s}$ quarks and anti-quarks to the entropy (solid line) as a function of $F$. The dashed line indicates the corresponding ratio calculated assuming absence of the phase transition to the QGP. [*Right:*]{} The ratio $E_S$ of strange particle to pion production as a function of $F$. Experimental data on central collisions of two identical nuclei are plotted as closed circles. These data should be compared with the model predictions shown by the solid line. The open boxes show results obtained for nucleon–nucleon interactions, scaled be a factor 3.6 to match A+A data at AGS energy. The plotted data show the status of 1998 as compiled in [@GaRo1; @GaRo2]. ](es.pdf "fig:"){width="0.4\linewidth"}
Extensions of the SMES
----------------------
Several extensions of the SMES were developed over the past 10 years which allow to predict signals of the onset of deconfinement related to the collective flow at freeze–out [@GoGaBu; @bleicher] and to event-by-event fluctuations [@GaGoMo; @GoGaZo]. These extensions are briefly presented in this subsection.
### Collective Flow at Freeze-out
The collective flow of matter at freeze-out depends on the properties of the early stage as well as on the expansion dynamics itself. Within the SMES the collision energy dependence of the early stage properties is predicted. In particular, in the energy range in which the mixed phase is created the pressure and temperature are constant and at the end of the mixed phase domain the pressure to energy density ratio reaches its minimum (the softest point of the EoS). From general hydrodynamic considerations this is expected to lead to a reduction of the buildup of transverse [@GoGaBu] and longitudinal [@bleicher] collective flow at freeze-out. The corresponding signals are discussed in Secs. 4.3 and 4.4.
### Event-by-Event Fluctuations
Up to this point only quantities averaged over many collisions (events) were considered. Next an extension of the SMES is reviewed which leads to predictions of fluctuations from event to event.
The key additional assumption is that when the collision energy is fixed, the energy, which is used for particle production (inelastic energy) can still fluctuate. These dynamical energy fluctuations lead to dynamical fluctuations of macroscopic properties $X$ of the matter, like its entropy and strangeness content [@GaGoMo]. The relation between them is given by the EoS. For example, different values of the energy of the early equilibrium state lead to different, but uniquely determined, entropies. Since the EoS shows an anomalous behavior in the phase transition region, this anomaly should also be visible in the ratio of entropy to energy fluctuations [@GaGoMo].
According to the first and the second principles of thermodynamics the entropy change $\delta S$ is given as $T\delta S = \delta E + p \delta V$. For central A+A collisions, one expects $ \delta V\cong 0$. Within the SMES the ratio of entropy to energy fluctuations can then be calculated and expressed as a simple function of the $p/\varepsilon$ ratio [@GaGoMo]: $$\label{R}
R_e ~\equiv ~\frac{(\delta S)^2/S^2}{(\delta E)^2/E^2}~=~
\left(1~+~\frac{p}{\varepsilon}\right)^{-2}~.$$ Within the SMES model, confined matter (which is modelled as an ideal gas) is created at the early collision stage below a collision energy of 30$A$ GeV. In this domain, the ratio $p/\varepsilon$, and consequently the $R_e$ ratio, are approximately independent of the collision energy and equal about 1/3 and 0.56, respectively. The SMES model assumes that the deconfinement phase transition is of the first order. Thus, there is a mixed phase region, corresponding to the energy interval 30$A$–60$A$ GeV. At the end of this region the $p/\varepsilon$ ratio reaches a minimum (the ‘softest point’ of the EoS [@Hu:95]). Thus in the transition energy range the $R_e$ ratio increases and reaches its maximum, $R_{e}\approx 0.8$, at the end of the transition domain. Further on, in the pure deconfined phase, which is represented by an ideal quark-gluon gas under bag pressure, the $p/\varepsilon$ ratio increases and again approaches its asymptotic value 1/3 at the highest SPS energy of 160$A$ GeV. The numerically calculated predictions of the SMES are plotted in Fig. \[rf\] [*left*]{}. The early stage energy and entropy fluctuations entering Eq. (\[R\]) are not directly observable, however, as argued in Ref. [@GaGoMo], they can be inferred from the experimentally accessible information on the final state energy and multiplicity fluctuations.
![\[rf\] The collision energy dependence of the fluctuation signals of the onset of the deconfinement calculated within the SMES. [*Left:*]{} The [*shark fin*]{} in of the ratio of entropy to energy fluctuations $R_e$ (\[R\]) (see Ref. [@GaGoMo]). [*Right:*]{} The [*tooth*]{} structure in the ratio of strangeness and entropy fluctuations $R_{s/e}$ (\[Rse\]) (see Ref. [@GoGaZo]). ](fig1f.pdf "fig:"){width="0.5\linewidth"} ![\[rf\] The collision energy dependence of the fluctuation signals of the onset of the deconfinement calculated within the SMES. [*Left:*]{} The [*shark fin*]{} in of the ratio of entropy to energy fluctuations $R_e$ (\[R\]) (see Ref. [@GaGoMo]). [*Right:*]{} The [*tooth*]{} structure in the ratio of strangeness and entropy fluctuations $R_{s/e}$ (\[Rse\]) (see Ref. [@GoGaZo]). ](fig3f.pdf "fig:"){width="0.5\linewidth"}
In Ref. [@GoGaZo] the energy dependence of dynamical strangeness fluctuations caused by dynamical energy fluctuations was studied within the SMES model. Defining $\overline{N}_s$ as the total number of strange quark–anti-quark pairs created in an A+A collision one calculates the fluctuation ratio as: $$R_s~=~\frac{(\delta \overline{N}_s)^2/\overline{N}_s^2}{(\delta
E)^2/E^2}~.
\label{Rs}$$ For $T\rightarrow \infty$ the system is in the QGP phase. Strange (anti-)quarks can be considered as massless and the bag constant can be neglected. Then $\varepsilon \propto T^4$ and $n_s\propto
T^3$ and consequently $d\varepsilon /\varepsilon = 4 \cdot dT/T$ and $dn_s/n_s=3\cdot dT/T$, which results in $R_s=(3/4)^2\cong
0.56$. In the confined phase, $T<T_c$, the energy density is still approximately proportional to $T^4$ due to the dominant contributions of non-strange hadron constituents. However, the dependence of the strangeness density on $T$ is governed by the exponential factor, $n_s\propto \exp(-m_S)$, as $T<<m_S = m_W^s \cong 500$ MeV. Therefore, at small $T$ one finds $d\varepsilon /\varepsilon = 4\cdot dT/T$ and $dn_s/n_s = m_S\cdot
dT/T^2$, so that the ratio $R_s=m_S/(4T)$ decreases with $T$. The strangeness density $n_s$ is small and goes to zero at $T\rightarrow 0$, but the fluctuation ratio $R_s$ Eq. (\[Rs\]) is large and increases to infinity in the zero temperature limit. One finds a non-monotonic energy dependence of $R_e$ with a maximum at the boundary between the mixed phase and the QGP [@GaGoMo]. A pronounced minimum-structure is expected in the dependence of $R_s$ on the collision energy [@GoGaZo]. It is located at $\sqrt{s}_{NN}~=$ 7–12 GeV ($30A-60A$ GeV), where the mixed phase is created at the early stage of A+A collision.
Both entropy and strangeness fluctuation measures, $R_e$ and $R_s$, show anomalous behavior in the transition region: a maximum is expected for $R_e$ and a minimum for $R_s$. Consequently, an even stronger anomaly is predicted for the ratio: $$R_{s/e}~\equiv~\frac{R_s}{R_e}~=\frac{(\delta
\overline{N}_s)^2/\overline{N}_s^2}{(\delta
\overline{N}_-)^2/\overline{N}_-^2}~, \label{Rse}$$ shown in Fig. \[rf\] [*right*]{}. Experimental measurements of $R_{s/e}$ may be easier than the measurements of $R_e$ and $R_s$ because the ratio $R_{s/e}$ requires measurements of particle multiplicities only, whereas both $R_e$ and $R_s$ involve also measurements of particle energies.
These predictions are discussed in Sec. 4.5 in the context of existing experimental data.
Signals of the Onset of Deconfinement
======================================
Next the predictions of the SMES model reviewed in Sec. 3 will be related to directly measurable qunatities and compared with available experimental results. In particular, their significance as evidence for the onset of deconfinement will be discussed in detail.
The Kink
--------
The majority of all particles produced in high energy interactions are pions. Thus, pions carry basic information on the entropy created in the collisions. On the other hand, entropy production should depend on the form of matter present at the early stage of collisions. Deconfined matter is expected to lead to a final state with higher entropy than that created by confined matter. Consequently, it is natural to expect that the onset of creation of deconfined matter should be signaled by an enhancement of entropy and thus pion production. This simple intuitive argument can be quantified within the SMES.
Equilibration between newly created matter and baryons is assumed to take place during the evolution of the system. It was argued that this equilibration causes transfer of entropy from the produced matter to baryons. The analysis of the pion suppression effect at low collision energies indicates that this transfer corresponds to the effective absorption of about 0.35 pion per participant nucleon [@GaGoMo:98]. It is further assumed that there are no other processes which change the entropy content of the state produced during the early stage.
For the comparison with the model it is convenient to define the quantity: $$\langle S_{\pi} \rangle = \langle \pi \rangle + \kappa
\langle K + \overline{K} \rangle + \alpha \langle N_P \rangle~,$$ where $\langle \pi \rangle$ is the measured total multiplicity of final state pions and $\langle K + \overline{K} \rangle$ is the multiplicity of kaons and anti-kaons. The factor $\kappa$ = 1.6 is the approximate ratio between mean entropy carried by a single kaon to the corresponding pion entropy at chemical freeze–out. The term $\alpha \langle N_P \rangle$ with $\alpha$ = 0.35 is the correction for the discussed partial transfer of the entropy to baryons. The quantity $\langle S_{\pi} \rangle$ can thus be interpreted as the early stage entropy measured in pion entropy units. The conversion factor between $S$ and $\langle S_{\pi} \rangle$ is chosen to be 4 ($\approx$ entropy units per pion at chemical freeze–out).
The dependence of the entropy per participant nucleon on $F$ is shown in Fig. \[spb\] [*left*]{}. Outside the transition region the entropy increases approximately proportional to $F$, but the slope in the Q–state region is larger than the slope in the W–state region.
The number of baryons which take part in the collision ($2A_p$ in the model calculations) is identified now with the experimentally measured number of participant nucleons, $\langle N_P \rangle$. The fraction of energy carried by the produced particles is taken to be $\eta= 0.67$ [@Ba:94] and is assumed to be independent of the size of the colliding nuclei and the collision energy.
The comparison made in 1998 between the ratio $\langle S_{\pi}
\rangle/\langle N_P \rangle$ calculated from available measurements and the model is shown in Fig. \[spb\] [*right*]{}. The parametrization of the W–state was chosen to fit the AGS data and, therefore, the agreement with low energy A+A data is not surprising. On the other hand the description of high energy (SPS) results obtained by the NA35 and NA49 Collaborations is essentially parameter free.
![\[pions\] Energy dependence of the mean pion multiplicity per participant nucleon measured in central Pb+Pb and Au+Au collisions (full symbols), compared to the corresponding results from $p+p(\bar{p})$ reactions (open symbols). The compilation of data is from Ref. [@evidence]. ](edep_pinp.pdf){width="1.0\linewidth"}
The pion multiplicity is proportional to the initial entropy, and the $\langle \pi \rangle/\langle N_P \rangle$ ratio can thus be calculated outside the transition region as, $$\label{npi}
\frac{\langle \pi \rangle}{\langle N_P \rangle}~\propto~
g^{1/4}~F~,$$ where $g=g_W^{ns}=16$ for the initial state in the confined phase and $g=g_Q=47.5$ for the initial state in the deconfined phase at $T >> m^s_Q$. Therefore, the $\langle \pi \rangle/\langle N_P \rangle$ ratio increases linearly with Fermi’s energy measure $F$ outside the transition region, and the slope parameter is proportional to $g^{1/4}$ [@Ga:95-97]. In the transition region, a steepening of the increase of pion production with energy is predicted, because of the activation of the partonic degrees of freedom.
The recent compilation of data [@evidence] on pion multiplicity produced in central Pb+Pb (Au+Au) collisions and $p+p(\bar{p})$ interactions is shown in Fig. \[pions\] which displays the mean pion multiplicity $\langle \pi\rangle = 1.5\, (\langle
\pi^-\rangle + \langle \pi^-\rangle )$ per wounded nucleon as a function of $F$. The results from $p+p(\bar{p})$ interactions are shown by the open symbols. Up to the top SPS energy the mean pion multiplicity in $p+p$ interactions is approximately proportional to $F$. A fit of $\langle \pi \rangle/\langle N_P\rangle =b \cdot F$ yields a value of $b \cong 1.063$ GeV$^{-1/2}$. For central Pb+Pb and Au+Au collisions (filled symbols in Fig. \[pions\]) the energy dependence is more complicated. Below 40$A$ GeV ($\sqrt{s}_{NN}=$ 8.7 GeV) the ratio $\langle \pi \rangle/\langle
N_P\rangle $ is lower in A+A collisions than in $p+p(\bar{p})$ interactions (pion suppression) while at higher energies this ratio is larger in A+A collisions than in $p+p(\bar{p})$ interactions (pion enhancement). A linear fit, $\langle \pi \rangle/\langle N_W\rangle =a+b \cdot F$ for $F < 1.85$ GeV$^{1/2}$ gives $a \cong -0.45$ and $b \cong
1.03$ GeV$^{-1/2}$. The slope parameter fitted in the range $F
> 3.5$ GeV$^{1/2}$ is $b \cong 1.33$. This is shown by the solid line in Fig. \[pions\] (the lowest data point at the top RHIC energy was excluded from the fit). Thus, in the region 15$A$–40$A$ GeV between the highest AGS and the lowest SPS energy the slope increases by a factor of about 1.3. This agrees with the SMES result: [$$\begin{aligned}
\label{gg}
\left(\frac{g_Q}{g_W^{ns}}\right)^{1/4}~=~\left(\frac{47.5}{16}\right)^{1/4}
~\cong~1.31~,
\end{aligned}$$]{} where the increase is caused by the creation of a transient state of deconfined matter in the early stage of the collisions at energies higher than $\sqrt{s}_{NN}~\approx$ 7.6 GeV (30$A$ GeV).
The Horn
--------
The enhanced production of strangeness was considered by many authors as a potential signal of QGP formation [@Rafelski; @Ka:86; @Ma:86]. The line of arguments is the following. One estimates that the strangeness equilibration time in the QGP is comparable to the duration of the collision process ($< 10 $ fm/c) and about 10 times shorter than the corresponding equilibration time in hadron matter. It is further assumed that in the early stage of the fireball the strangeness density is much below the equilibrium density, e.g., it is given by the strangeness obtained from the superposition of nucleon–nucleon interactions. Thus it follows that during the expansion of the matter the strangeness content increases rapidly and approaches its equilibrium value provided matter is in the QGP state. In hadron matter the modification of the initial strangeness content is less significant due to the long equilibration time. This leads to the expectation that strangeness production should rapidly increase when the energy threshold for the production of deconfined matter is crossed from below.
In the SMES the role of strangeness is different. This is because statistical production of particles is postulated and therefore also strange particles are assumed to be produced in equilibrium. Consequently possible secondary processes do not modify its value. At $T = T_c$ the strangeness density is lower in the QGP than in confined matter. Thus, a suppression of strangeness production is expected to occur when crossing the transition energy range from below. The low level of strangeness production in N+N interactions as compared to the higher strangeness yield per participant nucleon in central A+A collisions (called strangeness enhancement) can be understood as mostly due to the effect of strict strangeness conservation (canonical suppression) imposed on the strange and anti-strange degrees of freedom [@danos]. This constraint has an important effect for small statistical systems such as the confined matter in the early stage of N+N collisions.
We are interested in the collision energy region between the AGS and SPS. At ‘low’ collision energies (when a pure W–state is formed) the strangeness to entropy ratio increases with $F$. This is due to the fact that the mass of the strange degrees of freedom is significantly higher than the system temperature. At $T = T_c$ the ratio is higher in the W–state than in the Q–state region. Therefore the ratio decreases in the mixed phase region to the level characteristic for the Q–state. In the Q–state, due to the low mass of strange quarks compared to the system temperature, only a weak dependence of the ratio on $F$ is observed. The $F$ dependence of the strangeness/entropy ratio calculated in the SMES is shown in Fig. \[str\] [*left*]{}.
The comparison between the model and the data on strangeness production is performed under the assumption that the strangeness content created during the early stage is preserved till the hadronic freeze–out. This simplifies the model calculations by neglecting possible gluon contribution to strangeness production during hadronization of the QGP.
Total strangeness production can be studied (in the AGS and SPS energy range) using the experimental ratio: $$\label{esexp}
E_S ~=~ \frac {\langle \Lambda \rangle + \langle K + \overline{K} \rangle}
{\langle \pi \rangle}~,$$ where $\langle \Lambda \rangle$ is the mean multiplicity of $\Lambda$ hyperons (see Appendix 8.2). Within the SMES model $E_S$ of Eq. (\[esexp\]) is calculated as: $$\label{esmodel}
E_S ~=~ \frac {(N_s + N_{\overline{s}})/\zeta}
{ (S - S_s)/4 - \alpha \langle N_P \rangle }~,$$ where $\zeta$ = 1.36 is the experimentally estimated ratio between total strangeness production and strangeness carried by $\Lambda$ hyperons and $K$ + $\overline{K}$ mesons [@Bi:92] and $S_s$ is the fraction of the entropy carried by the strangeness carriers. The comparison between the calculations and the data available in 1998 is shown in Fig. \[str\] [*right*]{} [@GaGo]. The good description of the AGS data is again a consequence of the parametrization of the W-state: $g_W^s = 14$, $m_W^s = 500$ MeV which was based on these data. As in the case of the pion multiplicity, the description of the strangeness results at the SPS (from the NA35 and NA49 Collaborations) can be considered as being essentially parameter free [^1]. The agreement with the SPS data is obtained assuming creation of globally equilibrated QGP in the early stage of A+A collisions. The characteristic non–monotonic energy dependence of the $E_S$ ratio is a signature of the phase transition.
The entropy and strangeness production in central A+A collisions satisfies well the conditions needed for thermodynamical treatment. Therefore one expects that the measures of the entropy per participant nucleon, $\langle S_{\pi} \rangle/\langle N_P \rangle$, and the ratio of strangeness to entropy production, $E_S$, are independent of the number of participants for large enough values of $\langle N_P \rangle$. The energy dependence of the strangeness to entropy production ratio is a crucial signal of deconfinement. The temperature dependence of the multiplicity of a particle is strongly dependent on its mass. In the Boltzmann approximation one finds: $$\label{N}
\langle N_i \rangle ~=~ \frac{g^iV}{2 \pi^2}~ \int
\limits_{0}^{\infty} p^2 dp
~\exp\left(-~\frac{\sqrt{p^2+m_i^2}}{T}\right)~ =~ \frac{g_iV}{2
\pi^2}~ m_i^2 T~ K_{2}\left(\frac{m_i}{T} \right)~,$$ where $K_2$ is the modified Hankel function. For light particles ($m_l/T\ll 1$) one finds from Eq. (\[N\]), $\langle N_l \rangle \propto T^3$, whereas for heavy particles ($m_h/T\gg 1$) Eq. (\[N\]) leads to $\langle N_h \rangle \propto
T^{3/2}\exp(-~m_h/T)$. Within the SMES the strangeness to entropy production ratio increases steeply at low collision energies, when confined matter is produced. This is due to the low temperature at the early stage and the high mass of the carriers of strangeness (the kaon mass). Thus, $m_K\gg T$ and total strangeness production is proportional to $T^{3/2} \exp(-~m_K/T)$. On the other hand, the total entropy is approximately proportional to $T^3$. Therefore, the strangeness to pion production ratio is approximately $T^{-3/2} \exp(-m_K/T)$ in the confined phase and strongly increases with the collision energy. When the transition to deconfined matter occurs, the mass of the strangeness carriers is significantly reduced ($m_s \cong
175$ MeV, the strange quark mass). Due to the lower mass ($m_s <
T$) the strangeness yield becomes approximately proportional to the entropy (both are proportional to $T^3$), and the strangeness to entropy (or pion) production ratio becomes independent of energy in the QGP. This leads to a ‘jump’ in the energy dependence from the larger value for confined matter to the value for deconfined matter. Thus, within the SMES, the non-monotonic energy dependence of the strangeness to entropy production ratio followed by a plateau at the deconfined value is a direct consequence of the onset of deconfinement taking place at $\sqrt{s}_{NN}~\approx~$ 7.6 GeV (about 30$A$ GeV) [@GaGo].
The $E_S$ ratio was the first observable used to establish the energy dependence of the strangeness to entropy production ratio from the data. This ratio is closely proportional (see Appendix 8.2) to the $\langle K^+ \rangle/\langle \pi^+ \rangle$ ratio, which with time became better measured experimentally. The energy dependence of both ratios is plotted in Fig. \[strangeness\] for central Pb+Pb (Au+Au) collisions and $p+p$ interactions as function of collision energy.
![\[strangeness\] [*Left:*]{} Energy dependence of the $\langle K^+ \rangle /\langle \pi^+\rangle$ ratio measured in central Pb+Pb and Au+Au collisions (upper set of symbols) compared to the corresponding results from $p+p$ reactions (lower set of symbols). [*Right:*]{} Energy dependence of the relative strangeness production as measured by the $E_S$ ratio (see text) in central Pb+Pb and Au+Au collisions (upper set of symbols) compared to results from $p+p$ reactions (lower set of symbols). The compilation of data is from Ref. [@evidence]. The dashed-dotted line in the figure shows the predictions of the SMES [@GaGo]. ](kpip_4pi_data.pdf "fig:"){width="0.5\linewidth"} ![\[strangeness\] [*Left:*]{} Energy dependence of the $\langle K^+ \rangle /\langle \pi^+\rangle$ ratio measured in central Pb+Pb and Au+Au collisions (upper set of symbols) compared to the corresponding results from $p+p$ reactions (lower set of symbols). [*Right:*]{} Energy dependence of the relative strangeness production as measured by the $E_S$ ratio (see text) in central Pb+Pb and Au+Au collisions (upper set of symbols) compared to results from $p+p$ reactions (lower set of symbols). The compilation of data is from Ref. [@evidence]. The dashed-dotted line in the figure shows the predictions of the SMES [@GaGo]. ](es_sqrts.pdf "fig:"){width="0.5\linewidth"}
For $p+p$ interactions the ratios show a monotonic increase with energy. In contrast, a very different behavior is observed for central Pb+Pb (Au+Au) collisions. The steep threshold rise of the ratios characteristic for confined matter changes at high energy into a constant value at the level expected for deconfined matter. In the transition region (at low SPS energies) a sharp maximum is observed caused by the higher strangeness to entropy production ratio in confined matter than in deconfined matter. As seen in Fig. \[strangeness\] the measured dependence is consistent with that predicted within the SMES [@GaGo].
The Step
--------
The energy density at the early stage increases with increasing collision energy. At low and high energies, when a pure confined or deconfined phase is produced, this leads to an increase of the initial temperature and pressure. This, in turn, results in an increase of the transverse expansion of the produced matter and consequently a flattening of the transverse mass spectra of final state hadrons. One may expect an ‘anomaly’ [@van-hove; @Hu:95; @GoGaBu] in the energy dependence of the average hadron transverse momenta in the mixed phase region where the temperature and pressure are approximately constant.
The experimental data on spectra of the transverse mass $m_T= (m^2+p_T^2)^{1/2}$ are usually parameterized by a simple exponential dependence: $$\label{T*}
\frac{dN}{m_Tdm_T}~\cong~C~\exp\left(-~\frac{m_T}{T^*}\right)~.$$ The inverse slope parameter $T^{*}$ is sensitive to both the thermal and collective motion in the transverse direction. Hydrodynamical transverse flow with collective velocity $v_T$ modifies the Boltzmann $m_T$-spectrum of hadrons. At low transverse momenta, it leads to the result ($T_{kin}$ is the kinetic freeze-out temperature): $$\label{T*1}
T^*_{low-p_T} ~\cong ~T_{kin}~+~\frac{1}{2} m~v_T^2~.$$ Such a linear mass dependence of $T^*$ is supported by the data for hadron spectra at small $p_T$. However, for $p_T~\gg~m$ the hydrodynamical transverse flow leads to the mass-independent blue-shifted ‘temperature’: $$\label{T*2}
T^*_{high-p_T} ~=~T_{kin}~\cdot~ \sqrt{\frac{1+v_T}{1-v_T}}~.$$ Note that a simple exponential fit Eq. (\[T\*\]) neither works for light $\pi$-mesons, $T^{*}_{low-p_{T}}(\pi) <
T^{*}_{high-p_{T}}(\pi)$, nor for heavy (anti-)protons and (anti-)lambdas, $T^{*}_{low-p_{T}}(p,\Lambda)
> T^{*}_{high-p_{T}}(p,\Lambda)$ (see e.g., Refs. [@Teaney:2000cw; @Gorenstein:2001ti]).
Kaons are the best suited among measured hadron species for observing the effect of the modification of the EoS due to the onset of deconfinement in hadron transverse momentum spectra. The arguments are the following. First, the kaon $m_{T}$–spectra are only weakly affected by hadron re-scattering and resonance decays during the post-hydrodynamic hadron cascade at SPS and RHIC energies [@Teaney:2000cw]. Second, a simple one parameter exponential fit Eq. (\[T\*\]) is quite accurate for kaons in central A+A collisions at all energies. This simplifies the analysis of the experimental data. Third, high quality data on $m_T$-spectra of $K^+$ and $K^-$ mesons in central Pb+Pb (Au+Au) collisions are available over the full range of relevant energies.
![\[slopes\] Energy dependence of the inverse slope parameter $T^*$ of the transverse mass spectra of $K^+$ ([*left*]{}) and $K^-$ mesons ([*right*]{}) measured at mid-rapidity in central Pb+Pb and Au+Au collisions. The $K^{\pm}$ slope parameters are compared to those from $p +p$ reactions (open circles). The compilation of data is from Ref. [@evidence]. ](kap_slope_sqrts.pdf "fig:"){width="0.5\linewidth"} ![\[slopes\] Energy dependence of the inverse slope parameter $T^*$ of the transverse mass spectra of $K^+$ ([*left*]{}) and $K^-$ mesons ([*right*]{}) measured at mid-rapidity in central Pb+Pb and Au+Au collisions. The $K^{\pm}$ slope parameters are compared to those from $p +p$ reactions (open circles). The compilation of data is from Ref. [@evidence]. ](kam_slope_sqrts_data.pdf "fig:"){width="0.5\linewidth"}
![\[mt\] Energy dependence of the mean transverse mass, $\langle m_T\rangle ~-~m$, measured at mid-rapidity in central Pb+Pb and Au+Au collisions for $\pi^{\pm}$ ([*left*]{}), $K^{\pm}$ ([*middle*]{}), and $p$ and $\bar{p}$ ([*right*]{}). Results for positively (negatively) charged hadrons are shown by full (open) symbols. The compilation of data is from Ref. [@evidence]. ](fig8_meanmt.pdf){width="1.0\linewidth"}
The experimental results on the energy dependence of the inverse slope parameter of $K^+$ and $K^-$ transverse mass spectra for central Pb+Pb (Au+Au) collisions are shown in Fig. \[slopes\]. The striking features of the data can be summarized and interpreted [@GoGaBu] as follows. The $T^{*}$ parameter increases strongly with collision energy up to the SPS energy point at $\sqrt{s}_{NN}~=$ 7.6 GeV (30$A$ GeV). This is the energy region where the creation of confined matter at the early stage of the collisions is expected. Increasing collision energy leads to an increase of the early stage temperature and pressure. Consequently the transverse momenta of produced hadrons, measured by the inverse slope parameter, increase with collision energy. This rise is followed by a region of approximately constant value of the $T^{*}$ parameter in the SPS energy range $\sqrt{s}_{NN}=$ 7.6 - 17.2 GeV (30$A$ - 158$A$ GeV), where one expects the transition between confined and deconfined matter with the creation of mixed phases. The resulting modification of the equation of state ‘suppresses’ the hydrodynamical transverse expansion and leads to the observed plateau structure in the energy dependence of the $T^*$ parameter [@GoGaBu]. At higher energies (RHIC data), $T^{*}$ again increases with the collision energy. The EoS at the early stage becomes again stiff and the early stage pressure increases with collision energy, resulting in a resumed increase of $T^{*}$. As also shown in Fig.\[slopes\] the parameter $T^*$ in $p+p$ interactions appears to increase smoothly and does not show the step-like structure.
For the transverse mass spectra of pions and protons the inverse slope parameter depends on the transverse mass interval used in the fit. The mean transverse mass $\langle m_T\rangle$ provides an alternative characterization of the $m_T$-spectra. The energy dependence of $\langle m_T\rangle~-~m$ for pions, kaons and (anti-)protons is shown in Fig.\[mt\]. These results demonstrate that the approximate energy independence of $\langle m_T\rangle~-~m$ in the SPS energy range is a common feature for all abundantly produced particle species.
The Dale
--------
This subsection summarizes the analysis of Ref. [@bleicher] of the longitudinal pion spectra within Landau’s hydrodynamical model [@Landau:53; @Landau:55]. The interest in this model was revived by the remarkable observation that the rapidity distributions at all investigated energies can be well described by a single Gaussian (see [@Roland:2004] and references therein). Moreover, the energy dependence of the width can also be described reasonably well by the same model.
The main physics assumptions of Landau’s picture are as follows. The collision of two Lorentz-contracted nuclei leads to a complete stopping of the colliding nuclei and full thermalization of the created hadronic matter. This establishes the volume and energy density for the initial conditions of hydrodynamic expansion at each collision energy. Assuming for simplicity the equation of state in the form $p=c_s^2\varepsilon$ ($c_s$ denotes the speed of sound, and $c_s^2=1/3$ for an ideal massless particle gas) the pion rapidity spectrum is given by [@Shuryak:1972zq; @Carruthers:dw]: $$\frac{dn}{dy}=\frac{Ks_{\rm NN}^{1/4}}{\sqrt{2\pi
\sigma_y^2}}\,\exp\left(-\frac{y^2}{2\sigma_y^2}\right)
\label{eq1}$$ with $$\sigma_y^2=\frac{8}{3}\frac{c_s^2}{1-c_s^4}\,{\rm ln}(\sqrt
{s}_{\rm NN}/{2m_{\rm N}})\quad, \label{eq2}$$ where $K$ is a normalization factor converting entropy to pion density [^2]. The above prediction was compared with the experimental data on rapidity distributions of negatively charged pions produced in central Pb+Pb (Au+Au) collisions at various energies. Figure \[rapwidth\] [*left*]{} shows the measured width $\sigma_y$ of the rapidity spectra [@na49_blume; @Roland:2004; @klay; @brahms] as a function of the beam rapidity. The full line shows a linear fit through the data points. The dotted line indicates the Landau model predictions with $c_s^2=1/3$.
![\[rapwidth\] Comparison of the Landau hydrodynamic model with rapidity distributions of charged particles [@bleicher]. [*Left:*]{} The root mean square width $\sigma_y$ of the rapidity distributions of negatively charged pions in central Pb+Pb (Au+Au) reactions as a function of the beam rapidity $y_p$. The dotted line indicates the Landau model prediction with $c_s^2=1/3$, while the full line shows a linear fit through the data points. Data (full symbols) are taken from [@na49_blume; @Roland:2004; @klay; @brahms]. The statistical errors given by the experiments are smaller than the symbol sizes. Systematic errors are not available. [*Right:*]{} Speed of sound as a function of beam energy for central Pb+Pb (Au+Au) reactions as extracted from the data using Eq. (\[eq3\]). The statistical errors (not shown) are smaller than 3%. ](rw.pdf "fig:"){width="0.5\linewidth"} ![\[rapwidth\] Comparison of the Landau hydrodynamic model with rapidity distributions of charged particles [@bleicher]. [*Left:*]{} The root mean square width $\sigma_y$ of the rapidity distributions of negatively charged pions in central Pb+Pb (Au+Au) reactions as a function of the beam rapidity $y_p$. The dotted line indicates the Landau model prediction with $c_s^2=1/3$, while the full line shows a linear fit through the data points. Data (full symbols) are taken from [@na49_blume; @Roland:2004; @klay; @brahms]. The statistical errors given by the experiments are smaller than the symbol sizes. Systematic errors are not available. [*Right:*]{} Speed of sound as a function of beam energy for central Pb+Pb (Au+Au) reactions as extracted from the data using Eq. (\[eq3\]). The statistical errors (not shown) are smaller than 3%. ](c02.pdf "fig:"){width="0.5\linewidth"}
The model only roughly reproduces the measured dependence. At low AGS energies and at the top RHIC energy, the experimental points are under-predicted, while in the SPS energy regime Landau’s model over-predicts the width of the rapidity distributions.
These deviations can be attributed to the changes in the EoS, which can be effectively parameterized by allowing the speed of sound to be dependent on collision energy. By inverting Eq. (\[eq2\]) one can express $c_s^2$ in the medium as a function of the measured width of the rapidity distribution: $$c_s^2~=~-~\frac{4}{3}\frac{{\rm ln}(\sqrt {s}_{\rm NN}/{2
m_p})}{\sigma_y^2} ~+~\sqrt{\left[\frac{4}{3}\frac{{\rm ln}(\sqrt
{s}_{\rm NN}/{2 m_{\rm N}})}{\sigma_y^2}\right]^2+1}\quad.
\label{eq3}$$
The energy dependence of the sound velocities extracted from the data using Eq. (\[eq3\]) is presented in Fig. \[rapwidth\] [*right*]{}. The sound velocities exhibit a clear minimum (usually called the softest point) around a beam energy of $\sqrt{s}_{NN}~=~$7.6 GeV (30$A$ GeV).
As discussed previously the weakening of the transverse and longitudinal expansion is expected within the SMES at low SPS energies due to the onset of deconfinement which softens the EoS at the early stage. Generally, a softening of the equation of state was predicted as a signal for the mixed phase at the transition energy from hadronic to partonic matter [@Hung:1994eq; @Rischke:1995pe; @Brachmann:1999mp]. Therefore, we conclude that the data on rapidity spectra of negatively charged pions are indeed compatible with the assumption of the onset of deconfinement at the low SPS energies.
The Shark Fin and the Tooth
---------------------------
As discussed in Sec. 3.4 the event-by-event fluctuations of the energy used for particle production should lead to fluctuations which are sensitive to the onset of deconfinement.
The NA49 Collaboration looked for the [*shark fin*]{} structure in the energy dependence of the scaled variance of multiplicity distributions in central Pb+Pb collisions [@na49-fin]. The predicted [@GaGoMo] increase of the scaled variance of multiplicity distribution in the NA49 acceptance by about 0.01 due to the onset of deconfinement is smaller than the systematic error on the measurement. Therefore these data can neither support nor disprove the [*shark fin*]{} prediction.
The [*tooth*]{} structure in the energy dependence of $R_{s/e}$ shown in Fig. \[rf\] [*right*]{} might be seen in the event-by-event fluctuations of the $K$/$\pi$ ratio. The energy dependence of the fluctuations of this ratio in central Pb+Pb collisions was studied by NA49 using the so-called $\sigma_{dyn}$ measure [@na49-tooth]. The ‘dynamical’ $K/\pi$ fluctuations increase significantly with decreasing energy below 40$A$ GeV. It is unclear whether this increase is related to the rapid increase of the $R_{s/e}$ measure predicted due to the onset of deconfinement at energies below 30$A$ GeV.
Alternative Approaches
=======================
Several other analyzes of the energy dependence of hadron production properties in central Pb+Pb and Au+Au collisions within various theoretical approaches support the hypothesis that the onset of deconfinement is located at low SPS energies. In particular such a result was obtained from studies of hadron yields within a non-equilibrium hadron gas model [@rafelski] and from an analysis of the time evolution of the relative strangeness yield using the momentum integrated Boltzmann equation [@nayak]. Moreover, a simultaneous analysis of the two-pion correlation function and the transverse mass spectra found a plateau in the averaged phase-space density at SPS energies which may be associated with the onset of deconfinement [@sinyukov].
Recently a parity violating signal was observed in three-particle azimuthal correlations at RHIC [@star_parity]. Such an effect was predicted [@parity] since metastable domains may form in a QGP where parity and time-reversal symmetries are locally violated. The effect is expected to disappear when no QGP is produced in the collisions. It can therefore serve as an another indicator for the onset of deconfinement.
Numerous models have been developed to explain hadron production in reactions of heavy nuclei without explicitly invoking a transient QGP phase. The simplest one is the statistical hadron gas model [@Ha:94] which assumes that the hadrochemical freeze-out creates a hadron gas in equilibrium. The temperature, the baryon chemical potential, and the hadronization volume are free parameters of the model and are fitted to the data at each energy. In this formulation, the hadron gas model cannot make any prediction about the energy dependence of hadron production so that an extension of the model was proposed, in which the values of the temperature and baryon chemical potential evolve smoothly with collision energy [@Cl:01]. By construction (fit to the energy dependence), the prevailing trend in the data is reproduced by the model but important details are not, e.g. the decrease of the $\langle K^+ \rangle /\langle \pi^+\rangle$ ratio between $\sqrt{s}_{NN}~=$ 7.6 and 12.3 GeV (30$A$ and 80$A$ GeV) is not well described. The measured ratio of strangeness to pion yield in central Pb+Pb collisions at $\sqrt{s}_{NN}~=$ 17.2 GeV (158$A$ GeV) is about 25% lower than the expectation for the fully equilibrated hadron gas [@Cl:01; @Be:98]. Two strategies are followed in order to improve the quality of the hadron gas model fits. First, additional parameters have been introduced which allow for deviations from equilibrium [@Be:03; @Be:05; @rafelski]. Obviously the non-equilibrium hadron gas models [@Be:03; @Be:05; @rafelski] with all the parameters fitted separately to the data at each energy describe the experimental results significantly better. Secondly the equilibrium model was extended to include hypothetical high mass resonance states [@pbm]. Again by adding additional free parameters (mass dependence of the resonance state density and their branching ratios) the fit quality can be improved. Interestingly, the energy dependence of the parameters obtained within the extended hadron gas models is interpreted [@rafelski; @pbm] as an indication for the onset of deconfinement at $\sqrt{s}_{NN}~\approx~$ 7.6 GeV (30$A$ GeV).
Dynamical models of nucleus-nucleus collisions, such as RQMD [@RQMD], UrQMD [@URQMD] and HSD [@HSD] treat the initial nucleon-nucleon interactions within a string-hadronic framework. In addition these models include effects such as string-string interactions and hadronic re-scattering which are expected to be relevant in nucleus-nucleus collisions. RQMD [@RQMD; @RQMD1], UrQMD [@URQMD; @URQMD1; @Bratkovskaya:2004kv] and HSD [@Bratkovskaya:2004kv] models, like the hadron gas model, fail to describe the rapid change of hadron production properties with collision energy in the low SPS energy range.
It was shown that the maximum in relative strangeness production can be reproduced by invoking an unusually long lifetime of the fireball at low SPS energies which decreases with collision energy [@tomasik]. This assumption is however difficult to justify by dynamical models of the collision process [@URQMD; @HSD], and conflicts with the measured energy dependence of the two-pion correlation function [@hbt; @hbt1].
The step-like structure in the energy dependence of the inverse slope parameter of the transverse mass spectra was obtained within the hydrodynamical model by introduction of a rapid change of the freeze-out conditions at low SPS energies [@ivanov]. However, this assumption does not explain the increase of the inverse slope parameter suggested by the RHIC results.
In summary, one is led to conclude that models which do not invoke the onset of deconfinement at low SPS energies cannot explain comprehensively and consistently the energy dependence of hadron production properties in central Pb+Pb (Au+Au) collisions.
Open problems
=============
Open questions related to the onset of deconfinement are discussed in this section. First the theoretical questions are addressed, then the experimental issues are discussed.
Theoretical problems
--------------------
Quantum chromodynamics, the commonly accepted theory of strong interactions, was developed to model the interactions of quarks and gluons as well as their color neutral composites, the hadrons. Thus, in principle, this theory should be able to predict whether and via which observables the onset of deconfinement can be experimentally observed in nucleus-nucleus collisions. In QCD the strength of the strong force between two quarks increases with their distance. This property of the theory has as a consequence that, in general, predictions are either very difficult or impossible to calculate. Presently, there are no quantitative QCD predictions yet concerning the onset of deconfinement in nucleus-nucleus collisions.
The SMES model [@GaGo] assumes statistical particle production at the early stage of nucleus-nucleus collisions. The data on nucleus-nucleus collisions at RHIC, in particular results on anisotropic flow, seem to require a large degree of equilibration at the early stage of collisions [@heinz]. Thus the assumption of statistical particle production received independent experimental support.
The additional assumptions of the SMES, which lead to the [*kink*]{} and [*horn*]{} predictions, concern entropy and strangeness conservation during expansion and freeze-out. In fact, the model predictions remain at least qualitatively unchanged if one assumes an approximate proportionality of the final state entropy and strangeness to their early stage values. There is no easy way to prove or disprove these weaker requirements.
The predictions concerning the [*step*]{} and [*dale*]{} require assumptions concerning the equation of state of strongly interacting matter. In particular, they rely on a general feature of the EoS, i.e. the existence of the softest point when the transition between QGP and confined matter occurs. Thus, the qualitative predictions are independent of the nature of the transition (e.g. cross-over or 1$^{st}$ order phase transition, full or partial energy stopping and thermalization). However, the quantitative predictions are sensitive and a consistent description of the hydrodynamical evolution has not been achieved yet. In particular, the bag model equation of state with a 1$^{st}$ order phase transition to hadron gas leads to a significant over-prediction of transverse flow [@hama]. Further studies are needed.
The predictions concerning the [*tooth*]{} and [*shark fin*]{} are derived assuming that the early stage volume fluctuations can be neglected for central collisions. It is unclear, to which extent this condition is consistent with recent attempts to describe multiplicity distributions and high transverse momentum spectra in p+p interactions by a statistical model with volume fluctuations [@volume]. Further studies are needed.
Experimental issues
-------------------
The experimental results indicating the onset of deconfinement were obtained mainly by the NA49 experiment at the CERN SPS. Clearly, a confirmation of these measurements is necessary.
A beam energy scan program at the BNL RHIC has recently started. Pilot results at $\sqrt{s}_{NN}=$ 9.2 and 20 GeV are in agreement with the corresponding NA49 data [@rhic_sps]. New RHIC data being taken by the STAR experiment in 2010 will allow a more conclusive verification of the NA49 results.
At the CERN SPS the new experiment NA61 started a two dimensional system size and beam energy scan in 2009, which will continue over the next 4 years. The measurements aim to verify the existence of the onset of deconfinement in collisions of medium size nuclei. Moreover, they will allow to study the expected disappearance of the signals in collisions of light nuclei.
Summary and Conclusions
=======================
In this review we present the experimental and theoretical status of the evidence for the threshold of quark-gluon plasma creation in high energy nucleus-nucleus interactions. The location in energy of this so-called onset of deconfinement, as well as key experimental signals were predicted by the statistical model of the early stage of the collision process [@GaGo]. These signals were searched for and observed within the energy scan program of the NA49 Collaboration at the CERN SPS. Together with measurements at lower (LBL, JINR, SIS, BNL AGS) and higher (BNL RHIC) energies the properties of hadron production in heavy ion collisions were established in a broad energy range. Their energy dependence led to the conclusion that the predicted signals of the onset of deconfinement appear in a common energy domain covered by the SPS at CERN. These features of the data serve as strong experimental evidence for the existence of the onset of deconfinement and thus for the quark-gluon plasma itself.
Quantitative model predictions, discussed in this review, are derived within the statistical approach to particle production in high energy collisions. The use of this approach has a two-fold justification. First, it naturally includes the concept of phases of strongly interacting matter and the transition between them. Second, it is successful in describing numerous features of the experimental data. The relation between the statistical approach and the commonly accepted theory of strong interactions, QCD, remains unclear. This is because QCD is difficult or impossible to evaluate in the energy region relevant for multi-particle production in general and for the phase transitions of strongly interacting matter in particular.
New experimental programs have started at the CERN SPS and BNL RHIC which are devoted to the study of nucleus-nucleus collisions in the energy region where the NA49 experiment found evidence for the onset of deconfinement. The STAR experiment at RHIC will provide a necessary confirmation of these results. The new CERN experiment NA61 will address the questions how this observed phenomenon depends on the volume of matter and what the properties of the transition region are.
Appendices
==========
Strangeness Enhancement and $J/\psi$ Suppression
-------------------------------------------------
The idea of strangeness enhancement as a quark-gluon plasma signal in nucleus-nucleus (A+A) collisions was formulated a long time ago [@Rafelski]. It was based on the estimate that the strangeness equilibration time in the QGP is of the same order ($\approx 10$ fm/c) as the expected life time of the fireball formed in A+A collisions. Thus in the case of QGP creation strangeness is expected to approach its QGP equilibrium value. This equilibrium value is significantly higher than the level of strangeness production in nucleon–nucleon (N+N) collisions. Strangeness production in secondary hadronic interactions was estimated to be negligibly small. Therefore, if QGP is not formed, strangeness yields would be expected to be much lower than those predicted by equilibrium QGP calculations. Thus at that time a simple and elegant signature of QGP creation appeared: a transition to QGP should be signaled by an increase of the strangeness production level to the QGP equilibrium value.
The actual study has been done in the following way. The strangeness to pion ratio quantified by the ratios, $$\label{strange}
E_S~=~\frac{\langle \Lambda\rangle ~+~\langle
K+\overline{K}\rangle} {\langle \pi \rangle}~~~~ {\rm or}~~~~
\frac {\langle K^+\rangle } {\langle \pi^+\rangle }~ ,$$ was measured and analyzed. One expected that the ratios should increase [*strongly*]{} in A+A collisions if the QGP was formed. To reveal the specific increase of the strangeness/pion ratio in A+A collisions due to QGP formation the strangeness enhancement factor was introduced: $$\label{senh}
R_S(\sqrt{s}_{NN})
~\equiv~\frac{E_S^{AA}(\sqrt{s}_{NN})}{E_S^{NN}(\sqrt{s}_{NN})}
\approx \frac{\left(\langle K^+\rangle /\langle \pi^+\rangle
\right)_{AA}}{\left(\langle K^+\rangle /\langle \pi^+\rangle
\right)_{NN}} ~,$$ where superscripts $^{AA}$ and $^{NN}$ correspond respectively to A+A and N+N collisions at the same N+N c.m.energy $\sqrt{s}_{NN}$. The confrontation of this expectation with data was for the first time possible in 1988 when results from $^{32}$S and $^{28}$Si beams at the SPS and the AGS became available. Experiment NA35 reported that in central S+S collisions at 200$A$ GeV the strangeness to pion ratio is indeed about 2 times higher than in N+N interactions at the same energy per nucleon [@na35]. But an even larger enhancement ($R_S$ is about 5) was measured by E802 in Si+A collisions at the AGS. The data on central Au+Au collisions at low AGS energies 2$A$–10$A$ GeV completed the picture: strangeness enhancement is observed at all energies, and it is stronger at lower energies, i.e. the ratio $R_s$ Eq.(\[senh\]) [*increases*]{} monotonically with [*decreasing*]{} $\sqrt{s}_{NN}$. Figure \[enhancement\] shows a compilation of recent data [@mitrovski].
![\[enhancement\] Energy dependence of strangeness enhancement in central Pb+Pb (Au+Au) collisions [@mitrovski]. ](enhancement-vs-energy.png){width="1.0\linewidth"}
Moreover, the enhancement factor $R_s$ Eq.(\[senh\]) should evidently go to infinity at the threshold energy of strange hadron production in N+N collisions. At low AGS energies one does not expect the creation of a QGP and therefore the substantial strangeness enhancement is clearly of a different origin. The low level of strangeness production in p+p interactions as compared to the strangeness yield in central A+A collisions can be understood to a large extent in the statistical model as due to the effect of exact strangeness conservation. The canonical ensemble treatment of strangeness conservation leads to additional suppression factors imposed on strange hadron production in small systems such as created in p+p collisions. In any case, the AGS measurements indicating a strangeness enhancement $R_s$ larger than that at the SPS show clearly that the simple concept of strangeness enhancement as a signal of QGP production does not work.
The standard picture of $J/\psi$ production in collisions of hadrons and nuclei assumes a two step process: the creation of a $c\overline{c}$ pair in hard parton collisions at the very early stage of the reaction and subsequent formation of a bound charmonium state. It was proposed [@Satz] to use the $J/\psi$ as a probe for deconfinement in the study of A+A collisions. The argument was that in a QGP color screening dissolves initially created $J/\psi$ mesons into $c$ and $\overline{c}$ quarks which at hadronization form open charm hadrons. As the initial yield of $J/\psi$ is believed to have the same A–dependence as the yield of Dell–Yan lepton pairs, the measurement of a weaker A–dependence of the final $J/\psi$ yield ($J/\psi$ suppression) would signal charmonium absorption and therefore creation of QGP.
Production of charmonium states $J/\psi$ and $\psi^{\prime}$ was measured in A+A collisions at the CERN SPS over a period of 15 years by the NA38, NA50 and NA60 Collaborations [@jpsi]. The A-dependence of $J/\psi$ production in p+A is weaker than A$^1$ (approximately A$^{0.9}$). It was suggested that this $J/\psi$ suppression is due to absorption on nucleons in the target nucleus. The data on oxygen and sulfur collisions on nuclei at 200$A$ GeV also indicated considerable suppression. To improve the fit of the data a new source of $J/\psi$ absorption was introduced: the absorption on hadronic secondaries (‘co-movers’). Finally in central Pb+Pb collisions at 158$A$ GeV the measured suppression was found to be significantly stronger than expected in the models including nuclear and co-mover suppression. This ‘anomalous’ $J/\psi$ suppression was interpreted as evidence of QGP creation in Pb+Pb collisions at the CERN SPS. The uncertainties in estimates of the $J/\psi$ absorption by target nucleons and co-movers make $J/\psi$ suppression a problematic QGP signal. An essential part of the $J/\psi$ yield comes from decays of excited charmonium states like $\psi^{\prime}$ and $\chi$. All of them have different melting temperatures and absorption cross-sections.
Alternative approaches have been developed, namely the statistical [@Ga1] and the statistical coalescence [@Br1; @Go:00] models of $J/\psi$ production, which reproduce the A-dependence of the $J/\psi$ yield at SPS energies reasonably well. They are based on different physics pictures than the one leading to $J/\psi$ suppression as the signal of quark-gluon plasma creation. Specifically, the statistical model [@Ga1] assumes statistical production of $J/\psi$ mesons at hadronization, whereas in the coalescence model statistical coalescence of $c$ and $\overline{c}$ quarks at hadronization is assumed [@Br1; @Go:00]. In both models the $J/\psi$ yield is neither related to the J/psi suppression in the hadron gas nor in the quark-gluon plasma.
Main carriers of $s$ and $\bar{s}$ quarks
------------------------------------------
![\[strangeness\_carriers\] Main carriers of $s$ and $\bar{s}$ quarks. $K^-$ and $\bar{K}^0$ as well as $\Lambda$ and $\bar{\Lambda}$ yields are sensitive to the strangeness content and baryon density. $K^+$ and $K^0$ yields are sensitive mainly to strangeness content. ](strangeness_carriers.pdf){width="1.0\linewidth"}
The distribution of $s$ and $\bar{s}$ quarks among the most abundantly produced hadrons is considered here for the case of nucleus-nucleus collisions at the SPS energies. The sketch presented in Fig. \[strangeness\_carriers\] illustrates the following discussion.
The colliding nuclei have net numbers of $s$ and $\bar{s}$ quarks equal to zero. As strangeness is conserved in strong interactions the numbers of $s$ and $\bar{s}$ quarks in the final state have to be equal.
Kaons are the lightest strange hadrons. The isospin partners $K^+$ and $K^0$ mesons carry $\bar{s}$ quarks, whereas $K^-$ and $\bar{K}^0$ carry $s$ quarks. The (approximate) symmetry of the initial state and isospin conservation in strong interactions imply that: $$\langle K^+\rangle \approx \langle K^0\rangle$$ and $$\langle K^-\rangle \approx \langle \bar{K}^0\rangle .$$ The $\overline{s}$-quarks are also carried by the lightest anti-baryon, $\bar{\Lambda}$. Its fraction is however small (less than 5%) at the AGS and SPS energies due to suppression of the anti-baryon yield by the high net-baryon density. Consequently, $K^+$ and $K^0$ mesons carry each about half of all the anti-strange quarks produced in A+A collisions at AGS and SPS energies. Thus, their yields are nearly proportional to the total number of produced $s$ and $\bar{s}$ quarks.
This is not the case for $K^-$ and $\bar{K}^0$ mesons. A significant fraction of $s$-quarks (about 50% in central Pb+Pb collisions at 158 A$\cdot$GeV) is carried by hyperons. In addition this fraction strongly depends on collision energy. Consequently, the fraction of $s$-quarks carried by anti-kaons, $K^-$ and $\overline{K}^0$, is also dependent on collision energy and cannot be used easily to quantify strangeness production. In the $E_S$ ratio all main carriers of strange and anti-strange quarks are included. The neglected contribution of $\overline{\Lambda}$ and other hyperons and anti-hyperons is about 10% at SPS energies. Both the $\langle K^+\rangle /\langle
\pi^+\rangle$ and $E_S$ ratios are approximately, within 5% at SPS energies, proportional to the ratio of total multiplicity of $s$ and $\overline{s}$ quarks to the multiplicity of pions. It should be noted that the $\langle K^+\rangle/\langle \pi^+\rangle$ ratio is expected to be similar (within about 10%) for $p+p$, $n+p$, and $n+n$ interactions at 158$A$ GeV, whereas the $E_S$ ratio is independent of the isospin of nucleon-nucleon interactions.
Onset of deconfinement and critical point
------------------------------------------
![\[cpod\] Onset of deconfinement and critical point. ](cpod_v2.pdf){width="1.0\linewidth"}
This appendix discusses relations between the onset of deconfinement, the critical point of strongly interacting matter and the possibilities of their experimental study in relativistic nucleus-nucleus collisions. The two sketches presented in Fig. \[cpod\] should help to understand the basic ideas.
The onset of deconfinement refers to the beginning of the creation of a deconfined state of strongly interacting matter (ultimately a quark-gluon plasma) at the early stage of nucleus-nucleus collisions when increasing the collision energy. With increasing collision energy the energy density of matter created at the early stage of A+A collisions increases. Thus, if there are two phases[^3] of matter separated by the transition region (solid and dotted lines) as indicated in Fig. \[cpod\] [*left*]{} the early stage (the red point) first has to hit and then move above the transition region. Therefore, the existence of the onset of deconfinement is the most straightforward consequence of the existence of two phases of strongly interacting matter, i.e. confined matter and QGP. The experimental observation of the onset of deconfinement required a one dimensional scan in collision energy with heavy ions as performed by NA49. All signals of the onset of deconfinement discussed in this paper relate to the difference in properties of confined matter and QGP. They are not sensitive to the structure of the transition region.
Discovery of the onset of deconfinement implies the existence of QGP and of a transition region between confined and QGP phases. Numerous possibilities concerning the structure of the transition region are under discussion (see e.g., Ref. [@kapusta]). The most popular one [@ssr], sketched in Fig. \[cpod\], claims that a 1$^{st}$ order phase transition (thick gray line) separates both phases in the high baryonic chemical potential domain. In the low baryonic chemical potential domain a rapid crossover is expected (dotted line). The end point of the 1$^{st}$ order phase transition line is the critical point.
The characteristic signatures of the critical point can be observed if the freeze-out point (blue square in Fig. \[cpod\] [*right*]{}) is located close to the critical point. The analysis of the existing experimental data [@Be:05] indicates that the location of the freeze-out point in the phase diagram depends on the collision energy and the mass of the colliding nuclei. This dependence is schematically indicated in Fig. \[cpod\] [*right*]{}. Thus the experimental search for the critical point requires a two-dimensional scan in collision energy and size of the colliding nuclei. The NA61 experiment [@Gazdzicki:2006fy; @proposal] at the CERN SPS started this scan in 2009. It should be completed within several years. Note, that a two dimensional scan is actually required for any study of the structure of the transition region, independent of the hypothesis tested.
The transition region can be studied experimentally in nucleus-nucleus collisions only at $T$,$\mu_B$ values which correspond to collision energies higher than the energy of the onset of deconfinement. This important conclusion is easy to understand when looking at Fig \[cpod\]. Signals of the critical point can be observed provided the freeze-out point is close to it (see Fig. \[cpod\] $right$). On the other hand, by definition the critical point is located on the transition line. Furthermore, the energy density at the early stage of the collision is, of course, higher than the energy density at freeze-out. Thus, the condition that the freeze-out point is near the critical point implies that the early stage of the system is above (or on) the transition line. This in turn means that the optimal energy range for the search for the critical point lies above the energy of the onset of deconfinement (see Fig. \[cpod\] $left$). This general condition limits the search for the critical point to the collision energy range $E_{LAB} > 30A$ GeV.
[99]{}
D. D. Ivanenko and D. F. Kurdgelaidze, Astrophysics [**1**]{}, 251 (1965) \[Astrofiz. [**1**]{}, 479 (1965)\],\
N. Itoh, Prog. Theor. Phys. [**44**]{}, 291 (1970).
N. Cabibbo and G. Parisi, Phys. Lett. B [**59**]{}, 67 (1975),\
J. C. Collins and M. J. Perry, Phys. Rev. Lett. [**34**]{}, 1353 (1975). E. V. Shuryak, Phys. Rept. [**61**]{} (1980) 71. M. Gazdzicki and M. I. Gorenstein, Acta Phys. Polon. B [**30**]{}, 2705 (1999) \[arXiv:hep-ph/9803462\].
C. Alt [*et al.*]{} \[NA49 Collaboration\], Phys. Rev. C [**77**]{}, 024903 (2008) \[arXiv:0710.0118 \[nucl-ex\]\]. A report for non-experts may be found in: M. Gazdzicki, R. Stock and P. Seyboth, CERN Courier, Vol.43, No.3, 17 (2003). J. Pochodzalla [*et al.*]{}, Phys. Rev. Lett. [**75**]{}, 1040 (1995). http://cern.web.cern.ch/CERN/Announcements/2000/NewStateMatter/
J. Rafelski and B. Muller, Phys. Rev. Lett. [**48**]{}, 1066 (1982) \[Erratum-ibid. [**56**]{}, 2334 (1986)\];\
P. Koch, B. Müller, and J. Rafelski, Phys. Rep. [**142**]{}, 321 (1986). T. Matsui and H. Satz, Phys. Lett. B [**178**]{}, 416 (1986). M. Gazdzicki and D. Roehrich, Z. Phys. C [**65**]{}, 215 (1995). M. Gazdzicki and D. Roehrich, Z. Phys. C [**71**]{}, 55 (1996) \[arXiv:hep-ex/9607004\]. P. Seyboth \[NA49 Collaboration\], Addedndum-1 to the NA49 Proposal, CERN-SPSC-97-26;\
M. Gazdzicki, arXiv:nucl-th/9701050. M. Gazdzicki, J. Phys. G [**30**]{}, S161 (2004) \[arXiv:hep-ph/0305176\].
L. Van Hove, Phys. Lett. B [**118**]{}, 138 (1982). C. M. Hung and E. V. Shuryak, Phys. Rev. Lett. [**75**]{}, 4003 (1995) \[arXiv:hep-ph/9412360\];\
C. M. Hung and E. V. Shuryak, Phys. Rev. C [**57**]{}, 1891 (1998) \[arXiv:hep-ph/9709264\]. M. I. Gorenstein, M. Gazdzicki and K. A. Bugaev, Phys. Lett. B [**567**]{}, 175 (2003) \[arXiv:hep-ph/0303041\]. M. Gazdzicki [*et al.*]{} \[NA61/SHINE Collaboration\], PoS C [**POD2006**]{}, 016 (2006).
N. Antoniou [*et al.*]{} \[NA61/SHINE Collaboration\], CERN-SPSC-2006-034.
G. S. F. Stephans, [*critRHIC: The RHIC low energy program*]{}, arXiv:nucl-ex/0607030. P. Senger, T. Galatyuk, D. Kresan, A. Kiseleva and E. Kryshen, PoS C [**POD2006**]{} (2006) 018. A. N. Sissakian, A. S. Sorin and V. D. Toneev, arXiv:nucl-th/0608032.
M. Bleicher, arXiv:hep-ph/0509314,\
H. Petersen and M. Bleicher, PoS C [**POD2006**]{}, 025 (2006) \[arXiv:nucl-th/0611001\].
M. Gazdzicki, M. I. Gorenstein and S. Mrowczynski, Phys. Lett. B [**585**]{}, 115 (2004) \[arXiv:hep-ph/0304052\]. M. I. Gorenstein, M. Gazdzicki and O. S. Zozulya, Phys. Lett. B [**585**]{}, 237 (2004) \[arXiv:hep-ph/0309142\]. J. Baechler [*et al.*]{} \[NA35 Collaboration.\], Phys. Rev. Lett. [**72**]{}, 1419 (1994). H. Leutwyler, Phys. Lett. B [**378**]{}, 313 (1996) \[arXiv:hep-ph/9602366\]. J. Cleymans, R. V. Gavai and E. Suhonen, Phys. Rept. [**130**]{}, 217 (1986). M. Gazdzicki, Z. Phys. C [**66**]{}, 659 (1995);\
M. Gazdzicki, J. Phys. G [**23**]{}, 1881 (1997) \[arXiv:nucl-th/9706036\]. M. Gazdzicki, M. I. Gorenstein and S. Mrowczynski, Eur. Phys. J. C [**5**]{}, 129 (1998) \[arXiv:nucl-th/9701013\]. J. I. Kapusta and A. Mekjian, Phys. Rev. D [**33**]{}, 1304 (1986). T. Matsui, B. Svetitsky and L. D. McLerran, Phys. Rev. D [**34**]{}, 2047 (1986);\
T. Matsui, B. Svetitsky and L. D. McLerran, Phys. Rev. D [**34**]{}, 783 (1986) \[Erratum-ibid. D [**37**]{}, 844 (1988)\]. J. Rafelski and M. Danos, Phys. Lett. B [**97**]{}, 279 (1980). H. Bialkowska, M. Gazdzicki, W. Retyk and E. Skrzypczak, Z. Phys. C [**55**]{}, 491 (1992). D. Teaney, J. Lauret and E. V. Shuryak, Phys. Rev. Lett. [**86**]{}, 4783 (2001) \[arXiv:nucl-th/0011058\]. M. I. Gorenstein, K. A. Bugaev and M. Gazdzicki, Phys. Rev. Lett. [**88**]{}, 132301 (2002) \[arXiv:hep-ph/0112197\];\
K. A. Bugaev, M. Gazdzicki and M. I. Gorenstein, Phys. Lett. B [**544**]{}, 127 (2002) \[arXiv:hep-ph/0206109\];\
K. A. Bugaev, M. Gazdzicki and M. I. Gorenstein, Phys. Rev. C [**68**]{}, 017901 (2003) \[arXiv:hep-ph/0211337\]. L. D. Landau, Izv. Akad. Nauk Ser. Fiz. [**17**]{}, 51 (1953). S. Z. Belenkij and L. D. Landau, Nuovo Cim. Suppl. [**3S10**]{}, 15 (1956) \[Usp. Fiz. Nauk [**56**]{}, 309 (1955)\]. C. Alt [*et al.*]{} \[NA49 Collaboration\], Phys. Rev. C [**79**]{}, 044910 (2009) \[arXiv:0808.1237 \[nucl-ex\]\]. E. V. Shuryak, Yad. Fiz. [**16**]{}, 395 (1972). P. Carruthers, [*In \*New York 1973, Annals Of The New York Academy Of Sciences, Vol.229\*, New York 1974, 91-123 and Cornell Univ Ithaca - CLNS-219 (73,REC.APR) 51p*]{}
C. Blume, J. Phys. G [**31**]{}, S57 (2005). J. L. Klay [*et al.*]{} \[E-0895 Collaboration\], Phys. Rev. C [**68**]{}, 054905 (2003) \[arXiv:nucl-ex/0306033\]. I. G. Bearden [*et al.*]{} \[BRAHMS Collaboration\], Phys. Rev. Lett. [**94**]{}, 162301 (2005) \[arXiv:nucl-ex/0403050\]. C. M. Hung and E. V. Shuryak, Phys. Rev. Lett. [**75**]{}, 4003 (1995) \[arXiv:hep-ph/9412360\]. D. H. Rischke, Y. Pursun, J. A. Maruhn, H. Stoecker and W. Greiner, Heavy Ion Phys. [**1**]{}, 309 (1995) \[arXiv:nucl-th/9505014\]. J. Brachmann, A. Dumitru, H. Stoecker and W. Greiner, Eur. Phys. J. A [**8**]{}, 549 (2000) \[arXiv:nucl-th/9912014\]. C. Alt [*et al.*]{} \[NA49 Collaboration\], Phys. Rev. C [**78**]{}, 034914 (2008) \[arXiv:0712.3216 \[nucl-ex\]\]. C. Alt [*et al.*]{} \[NA49 Collaboration\], Phys. Rev. C [**79**]{}, 044910 (2009) \[arXiv:0808.1237 \[nucl-ex\]\].
J. Letessier and J. Rafelski, Eur. Phys. J. A [**35**]{}, 221 (2008) \[arXiv:nucl-th/0504028\]. J. K. Nayak, J. Alam, P. Roy, A. K. Dutt-Mazumder and B. Mohanty, Acta Phys. Slov. [**56**]{}, 27 (2006).
S. V. Akkelin and Yu. M. Sinyukov, Phys. Rev. C [**73**]{}, 034908 (2006). \[arXiv:nucl-th/0505045\]. B. Abelev [*et al.*]{} \[STAR Collaboration\], Phys. Rev. Lett. [**103**]{}, 251601 (2009).
D. Kharzeev, R. D. Pisarski and M. H. G. Tytgat, Phys. Rev. Lett. [**81**]{}, 512 (1998) \[arXiv:hep-ph/9804221\]. R. Hagedorn, CERN report CERN-TH-7190-94 and Proceedings of NATO Advanced Study Workshop on Hot Hadronic Matter: Theory and Experiment, Divonne-les-Bains, Switzerland, 27 Jun - 1 Jul 1994, edited by J. Letessier, H. Gutbrod and J. Rafelski, \[Hot Hadronic Matter, v. [**346**]{}, 13 (1994)\];\
J. Cleymans and H. Satz, Z. Phys. C [**57**]{}, 135 (1993);\
J. Sollfrank, M. Gaździcki, U. Heinz and J. Rafelski, Z. Phys. C [**61**]{}, 659 (1994);\
P. Braun–Munzinger, J. Stachel, J. Wessels and N. Xu, Phys. Lett. B [**365**]{}, 1 (1996);\
G. D. Yen, M. I. Gorenstein, W. Greiner, S.N. Yang, Phys. Rev. C [**56**]{}, 2210 (1997);\
F. Becattini and U. W. Heinz, Z. Phys. C [**76**]{}, 269 (1997);\
G. D. Yen and M. I. Gorenstein, Phys. Rev. C [**59**]{}, 2788 (1999).
J. Cleymans and K. Redlich, Phys. Rev. C [**60**]{}, 054908 (1999) \[arXiv:nucl-th/9903063\];\
P. Braun-Munzinger, J. Cleymans, H. Oeschler and K. Redlich, Nucl. Phys. A [**697**]{}, 902 (2002) \[arXiv:hep-ph/0106066\]. F. Becattini, M. Gazdzicki and J. Sollfrank, Eur. Phys. J. C [**5**]{}, 143 (1998) \[arXiv:hep-ph/9710529\]. F. Becattini, M. Gazdzicki, A. Keranen, J. Manninen and R. Stock, Phys. Rev. C [**69**]{}, 024905 (2004). F. Becattini, J. Manninen and M. Gazdzicki, Phys. Rev. C [**73**]{}, 044905 (2006). A. Andronic, P. Braun-Munzinger and J. Stachel, Nucl. Phys. A [**834**]{}, 237C (2010) \[arXiv:0911.4931 \[nucl-th\]\]. H. Sorge, H. Stocker and W. Greiner, Nucl. Phys. A [**498**]{}, 567 (1989). S. A. Bass [*et al.*]{}, Prog. Part. Nucl. Phys. [**41**]{}, 225 (1998) \[arXiv:nucl-th/9803035\].
W. Cassing, E. L. Bratkovskaya and S. Juchem, Nucl. Phys. A [**674**]{}, 249 (2000). \[arXiv:nucl-th/0001024\]. F. Wang, H. Liu, H. Sorge, N. Xu and J. Yang, Phys. Rev. C [**61**]{}, 064904 (2000) \[arXiv:nucl-th/9909001\]. H. Weber, E. L. Bratkovskaya and H. Stoecker, Phys. Lett. B [**545**]{}, 285 (2002). E. L. Bratkovskaya [*et al.*]{}, Phys. Rev. C [**69**]{}, 054907 (2004). B. Tomasik and E. E. Kolomeitsev, Eur. Phys. J. C [**49**]{}, 115 (2007). S. Kniege [*et al.*]{} \[NA49 Collaboration\], J. Phys. G [**30**]{}, S1073 (2004). C. Alt [*et al.*]{} \[NA49 Collaboration\], Phys. Rev. C [**77**]{}, 064908 (2008) \[arXiv:0709.4507 \[nucl-ex\]\].
Yu. B. Ivanov and V. N. Russkikh, Eur. Phys. J. A [**37**]{}, 139 (2008) \[arXiv:nucl-th/0607070\]. U. W. Heinz, J. Phys. Conf. Ser. [**50**]{}, 230 (2006) \[arXiv:nucl-th/0504011\]. M. Gazdzicki, M. I. Gorenstein, F. Grassi, Y. Hama, T. Kodama and O. J. Socolowski, Braz. J. Phys. [**34**]{}, 322 (2004) \[arXiv:hep-ph/0309192\]. V. V. Begun, M. Gazdzicki and M. I. Gorenstein, Phys. Rev. C [**78**]{}, 024904 (2008) \[arXiv:0804.0075 \[hep-ph\]\]. B. I. Abelev [*et al.*]{} \[STAR Collaboration\], Phys. Rev. C [**81**]{}, 024911 (2010) \[arXiv:0909.4131 \[nucl-ex\]\]. T. Alber [*et al.*]{} \[NA35 Collaboration.\], Z. Phys. C [**64**]{}, 195 (1994). M. Mitrovski \[NA49 Collaboration\], Ph.D. Thesis, University of Frankfurt, 2007.
C. Baglin [*et al.*]{} \[NA38 Collaboration\], Phys. Lett. B [**220**]{}, 471 (1989);\
M. C. Abreu [*et al.*]{} \[NA50 Collaboration\], Phys. Lett. B [**477**]{}, 28 (2000);\
R. Arnaldi [*et al.*]{} \[NA60 Collaboration\], Phys. Rev. Lett. [**99**]{}, 132302 (2007).
M. Gazdzicki and M. I. Gorenstein, Phys. Rev. Lett. [**83**]{}, 4009 (1999) \[arXiv:hep-ph/9905515\]. P. Braun-Munzinger and J. Stachel, Phys. Lett. B [**490**]{}, 196 (2000) \[arXiv:nucl-th/0007059\]. M. I. Gorenstein, A. P. Kostyuk, H. Stoecker and W. Greiner, Phys. Lett. B [**524**]{}, 265 (2002) \[arXiv:hep-ph/0104071\]. M. G. Alford, K. Rajagopal and F. Wilczek, Phys. Lett. B [**422**]{}, 247 (1998) \[arXiv:hep-ph/9711395\];\
L. McLerran and R. D. Pisarski, Nucl. Phys. A [**796**]{}, 83 (2007) \[arXiv:0706.2191 \[hep-ph\]\].
J. I. Kapusta and E. S. Bowman, arXiv:0908.0726 \[nucl-th\].
M. A. Stephanov, K. Rajagopal and E. V. Shuryak, Phys. Rev. Lett. [**81**]{}, 4816 (1998) \[arXiv:hep-ph/9806219\].
[^1]: The $E_S$ value resulting from a QGP can be estimated in a simple way. Assuming that $m_s =
0$, and neglecting the small ($< 5 \%$) effect of pion absorption at the SPS, one gets from Eq. (\[strent\]) and Eq. (\[esmodel\]) $E_S
\approx (g^s_Q/1.36)/g^{ns}_Q \approx 0.21$, where $g^s_Q = (7/8)
\cdot 12$ is the effective number of degrees of freedom of $s$ and $\overline{s}$ quarks and $g^{ns}_Q = 16 + (7/8) \cdot 24$ is the corresponding number for $u, \overline{u}, d, \overline{d}$ quarks and gluons. Moreover, we use the approximation that the pion entropy at freeze–out is equal to the mean entropy of $q$, $\overline{q}$ and $g$ in a QGP.
[^2]: There are two issues related to derivation of Eq. (\[eq2\]) which need clarification by future study. First, Eq. (\[eq2\]) is obtained assuming that $c_s$ depends only on the early stage energy density and its dependence on decreasing energy density during expansion is neglected. Second, the Landau model assumes stopping and thermalization of the total energy in the c.m. system, whereas only a fraction of the inelastic energy is stopped and thermalized in the SMES model.
[^3]: The discussed two phase diagram is the simplest one which allows to introduce the concepts of the onset of deconfinement and the transition region. There are numerous suggestions of phase diagrams with a significantly richer structure (see e.g., Ref. [@diagram]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Influence of hole shape on extraordinary optical transmission was investigated using hole arrays consisting of rectangular holes with different aspect ratio. It was found that the transmission could be tuned continuously by rotating the hole array. Further more, a phase was generated in this process, and linear polarization states could be changed to elliptical polarization states. This phase was correlated with the aspect ratio of the holes. An intuitional model was presented to explain these results.
PACS numbers:[78.66.Bz,73.20.MF, 71.36.+c]{}
author:
- 'Xi-Feng Ren, Pei Zhang, Guo-Ping Guo[^1], Yun-Feng Huang, Zhi-Wei Wang, Guang-Can Guo'
date: 'Received: date / Revised version: date'
title: Polarization properties of subwavelength hole arrays consisting of rectangular holes
---
introduction
============
In metal films perforated with a periodic array of subwavelength apertures, it has long been observed that there is an unusually high optical transmission[@1]. It is believed that metal surface plays a crucial role and the phenomenon is mediated by surface plasmon polaritons (SPPs) and there is a process of transforming photon to SPP and back to photon[@4; @crucial; @ebbesen5]. This phenomenon can be used in various applications, for example, sensors, optoelectronic device, etc[@williams; @brolo; @nahata; @luo; @shinada; @ebbeson07]. Polarization properties of nanohole arrays have been studied in many works[@Elli04; @Gordon04; @Altew05]. Recently, orbital angular momentum of photons was explored to investigate the spatial mode properties of surface plasmon assisted transmission [@ren061; @ren062]. It is also showed that entanglement of photon pairs can be preserved when they respectively travel through a hole array [@ren062; @Alt; @energy]. Therefore, the macroscopic surface plasmon polarizations, a collective excitation wave involving typically $10^{10}$ free electrons propagating at the surface of conducting matter, have a true quantum nature. However, the increasing use of EOT requires further understanding of the phenomenon.
The polarization of the incident light determines the mode of excited SPP which is also related to the periodic structure. For the manipulation of light at a subwavelength scale with periodic arrays of holes, two ingredients exist: shape and periodicity[@4; @crucial; @ebbesen5; @Elli04; @klein; @Ruan; @sarra]. Influence of unsymmetrical periodicity on EOT was discussed in [@renapl]. Influence of the hole shape on EOT was also observed recently[@klein; @sarra], in which the authors mainly focused on the transmission spectra. In this work, we used rectangle hole arrays to investigate the influence of hole shape on the polarization properties of EOT. It is found that linear polarization states could be changed to elliptical polarization states and a phase could be added between two eigenmode directions. The phase was changed when the aspect ratio of the rectangle holes was varied. The hole array was also rotated in the plane perpendicular to the illuminate beam. The optical transmission was changed in this process. It strongly depended on the rotation angle, in other words, the angle between polarization of incident light and axis of hole array, as in the case with unsymmetrical hole array structure[@renapl].
experimental results and modeling
=================================
Relation between transmission efficiency and photon polarization
----------------------------------------------------------------
Fig. 1(a) is a scanning electron microscope picture of part of our hole arrays. The hole arrays are produced as follows: after subsequently evaporating a $3$-$nm$ titanium bonding layer and a $135$-$nm$ gold layer onto a $0.5$-$mm$-thick silica glass substrate, a focused ion beam etching system is used to produce rectangle holes ($100nm\times 100nm$, $100nm\times 150nm$, $100nm\times 200nm$, $100nm\times 300nm$ respectively) arranged as a square lattice ($520nm$ period). The area of the hole array is $10\mu m\times 10\mu m$.
Transmission spectra of the hole arrays were recorded by a silicon avalanche photodiode single photon counter couple with a spectrograph through a fiber. White light from a stabilized tungsten-halogen source passed though a single mode fiber and a polarizer (only vertical polarized light can pass), then illuminated the sample. The hole arrays were set between two lenses of $35 mm$ focal length, so that the light was normally incident on the hole array with a cross sectional diameter about $10\mu m$ and covered hundreds of holes. The light exiting from the hole array was launched into the spectrograph. The hole arrays were rotated anti-clockwise in the plane perpendicular to the illuminating light, as shown in Fig. 1(b). Transmission spectra of the hole arrays for rotation angle $\theta=0\textordmasculine $ and $90\textordmasculine
$ were given in Fig. 2. There were large difference between the two cases, which was also observed in [@klein].
Further, the typical hole array($100nm\times 300nm$ holes) was rotated anti-clockwise in the plane perpendicular to the illuminating light(see Fig.1 (b)). Transmission efficiencies of $H$ and $V$ photons(702nm wavelength) were measured with rotation angle $\theta=0\textordmasculine , 30\textordmasculine ,
45\textordmasculine, 60\textordmasculine$, and $90\textordmasculine
$ respectively, as shown in Fig. 3. They were varied with $\theta$. To explain the results, we gave a simple model. For our sample, photons with $702nm$ wavelength will excite the SPP eigenmodes $(0,\pm 1)$ and $(\pm 1, 0)$. Since the SPPs were excited in the directions of long (L) and short (S) edges of rectangle holes, we suspected that this two directions were eigenmode-directions for our sample. The polarization of illuminating light was projected into the two eigenmode-directions to excite SPPs. After that, the two kinds of SPPs transmitted the holes and irritated light with different transmission efficiencies $T_{L}$ and $T_{S}$ respectively. For light whose polarization had an angle $\theta$ with the $S$ direction, the transmission efficiency $T_{\theta}$ will be
$$T_{\theta }=T_{S}\cos^2 (\theta)+T_{L}\sin^2 (\theta).$$
This equation was also given in the works[@sarra; @renapl]. Due to the unequal values of $T_{L}$ and $T_{S}$, the whole transmission efficiency was varied with angle $\theta$. So if we know the transmission spectra for enginmode-directions (here L and S), we can calculate out the transmission spectra (including the heights and locations of peaks) for any $\theta$. The theoretical calculations were also given in Fig. 3, which agreed well with the experimental data. The similar results were also observed when the hole arrays ($100nm\times 150nm$ and $100nm\times 200nm$) were used. With this model, the transmission efficiency can be continuously tuned in a certain range.
Influence of hole shape on photon polarization
----------------------------------------------
To investigate the polarization property of the hole array, we used the method of polarization state tomography. Experimental setup was shown in Fig. 4. White light from a stabilized tungsten-halogen source passed though single mode fiber and $4 nm$ filter (center wavelength 702 nm) to generate 702nm wavelength photons. Polarization of input light was controlled by a polarizer, a HWP (half wave plate, 702nm) and a QWP (quarter wave plate, 702nm). The hole array was set between two lenses of $35 mm$ focal length. Symmetrically, a QWP, a HWP and a polarizer were combined to analyze the polarization of transmitted photons. For arbitrary input states, the output states were measured in the four bases: $H$, $V$, $1/\sqrt{2}(|H\rangle+|V\rangle)$, and $1/\sqrt{2}(|H\rangle+i|V\rangle)$. With these experimental data, we could get the density matrix of output states, which gave the full polarization characters of transmitted photons. For example, in the case of $\theta=0\textordmasculine $, for input state $1/\sqrt{2}(|H\rangle+e^{I*0.5\pi}|V\rangle)$, four counts (8943, 31079, 3623 and 21760) were recorded when we used the four detection bases. The density matrix was calculated as: $$\left( \begin{matrix}
0.223 & -0.410 - 0.043i\\
-0.410 + 0.043i & 0.777 \\ \end{matrix} \right),$$ which had a fidelity of 0.997 with the pure state $0.472|H\rangle+0.882e^{I*0.967\pi}|V\rangle$. Compared this state with the input state, we found that not only the ratio of $|H\rangle$ and $|V\rangle$ was changed, but also a phase $\varphi=0.467\pi$ was added between them. The similar phenomenon was also observed when the input state was $1/\sqrt{2}(|H\rangle+|V\rangle)$ and in this case $\varphi=0.442\pi$. We also considered the cases for $\theta=30\textordmasculine , 45\textordmasculine,
60\textordmasculine$, and $90\textordmasculine $. The experimental density matrices had the fidelities all larger than 0.960 with the theoretical calculations, where $\varphi=(0.462\pm 0.053)\pi$. It can be seen that the phase $\varphi$ was hardly influenced by the rotation.
To study the dependence of phase $\varphi$ with the hole shape, we performed the same measurements on other hole arrays which were shown in Fig. 1. It was found that $\varphi$ was changed with the aspect ratio of the rectangle holes. Fig. 5 gave the relation between $\varphi$ and aspect ratio. The phases are $0$, $(0.227\pm
0.032)\pi$, $(0.357\pm 0.020)\pi$ and $(0.462\pm 0.053)\pi$ for aspect ratio $1$, $1.5$, $2.0$ and $3.0$ respectively. As mentioned above, period is another important parameter in the EOT experiments. Since no similar result was observed for hole arrays with symmetrical periods, a special quadrate hole array(see Fig. 1 of [@renapl]) was also investigated to show the influence of the hole period. We found that even the periods were different in two directions, there was no birefringent phenomenon($\varphi=0$).
This birefringent phenomenon might be explained with the propagating of SPPs on the metal surface. As we know, the interaction of the incident light with surface plasmon is made allowed by coupling through the grating momentum and obeys conservation of momentum $$\overrightarrow{k}_{sp}=\overrightarrow{k}_{0}\pm
i\overrightarrow{G}_{x}\pm j\overrightarrow{G}_{y},$$ where $\overrightarrow{k}_{sp}$ is the surface plasmon wave vector, $\overrightarrow{k}_{0}$ is the component of the incident wave vector that lies in the plane of the sample, $\overrightarrow{G}_{x}$ and $\overrightarrow{G}_{y}$ are the reciprocal lattice vectors, and i, j are integers. Usually, $G_{x}=G_{y}=2\pi/d$ for a square lattice, and relation $\overrightarrow{k}_{sp}*d=m\pi$ was satisfied, where $m$ was the band index[@Teje]. While for our rectangle hole arrays, the length of holes in $L$ direction was changed form $150nm$ to $300nm$, which was not as same as it in $S$ direction. Though $G_{x}=G_{y}=2\pi/d$ for our rectangle hole array, the time for surface plasmon polariton propagating in the $L$ direction must be influenced by the aspect ratio of hole shape, which could not be same as that in the $S$ direction. A phase difference $\varphi$ was generated between the two directions, leading the birefringent phenomenon. Due to the absorption or scattering of the SPPs and scattering at the hole edges, it is hard to give the accurate value of the phase or the exact relation between the phase and aspect ratio of holes. Even so, $\varphi$ could be controlled by changing the hole shape. As a contrast, there was no birefringent phenomenon observed when the quadrate hole array(see Fig. 1 of [@renapl]) was used. The reason was that phase $G_{x}*d_{x}$ always equal to $G_{y}*d_{y}$, even $G_{x}\neq G_{y}$ for the quadrate hole array.
conclusion
==========
In conclusion, rectangle hole array was explored to study the influence of hole shape on EOT, especially the properties of photon polarization. Because of the unsymmetrical of the hole shape, a birefringent phenomenon was observed. The phase was determined by the hole shape, which gave us a potential method to control this birefringent process. It was also found that the transmission efficiency can be tuned continuously by rotating the hole array. These results might be explained using an intuitional model based on surface plasmon eigenmodes.
This work was funded by the National Fundamental Research Program, National Natural Science Foundation of China (10604052), Program for New Century Excellent Talents in University, the Innovation Funds from Chinese Academy of Sciences, the Program of the Education Department of Anhui Province (Grant No.2006kj074A). Xi-Feng Ren also thanks for the China Postdoctoral Science Foundation (20060400205) and the K. C. Wong Education Foundation, Hong Kong.
[0]{} T.W. Ebbesen, H. J. Lezec, H. F. Ghaemi, T. Thio, and P. A. Wolff, Nature 391, 667 (1998).
H. Raether, [*Surface Plasmons on Smooth and Rough Surfaces and on Gratings*]{}, Vol. 111 of Springer Tracts in Modern Physics, Springer, Berlin, (1988).
D. E. Grupp, H. J. Lezec, T. W. Ebbesen, K. M. Pellerin, and Tineke Thio, Appl. Phys. Lett. 77 1569 (2000).
M. Moreno, F. J. Garc¨ªa-Vidal, H. J. Lezec, K. M. Pellerin, T. Thio, J. B. Pendry, and T. W. Ebbesen, Phys. Rev. Lett. 86, 1114 (2001).
S. M. Williams, K. R. Rodriguez, S. Teeters-Kennedy, A. D. Stafford, S. R. Bishop, U. K. Lincoln, and J. V. Coe, J. Phys. Chem. B. 108, 11833 (2004). A. G. Brolo, R. Gordon, B. Leathem, and K. L. Kavanagh, Langmuir. 20, 4813 (2004). A. Nahata, R. A. Linke, T. Ishi, and K. Ohashi, Opt. Lett. 28, 423 (2003). X. Luo and T. Ishihara, Appl. Phys. Lett. 84, 4780 (2004). S. Shinada, J. Hasijume and F. Koyama, Appl. Phys. Lett. 83, 836 (2003). C. Genet and T. W. Ebbeson, Nature, 445, 39 (2007).
J. Elliott, I. I. Smolyaninov, N. I. Zheludev, and A. V. Zayats, Opt. Lett. 29, 1414 (2004).
R. Gordon, A. G. Brolo, A. McKinnon, A. Rajora, B. Leathem, and K. L. Kavanagh, Phys. Rev. Lett. 92, 037401 (2004).
E. Altewischer, C. Genet, M. P. van Exter, and J. P. Woerdman, Opt. Lett. 30, 90 (2005).
X. F. Ren, G. P. Guo, Y. F. Huang, Z. W. Wang, and G. C. Guo, Opt. Lett. 31, 2792, (2006).
X. F. Ren, G. P. Guo, Y. F. Huang, C. F. Li, and G. C. Guo, Europhys. Lett. 76, 753 (2006). E. Altewischer, M. P. van Exter and J. P. Woerdman Nature 418 304 (2002).
S. Fasel, F. Robin, E. Moreno, D. Erni, N. Gisin and H. Zbinden, Phys. Rev. Lett. 94 110501 (2005).
K. J. Klein Koerkamp, S. Enoch, F. B. Segerink, N. F. van Hulst and L. Kuipers, Phys. Rev. Lett. 92 183901 (2004). Zhichao Ruan and Min Qiu, Phys. Rev. Lett. 96 233901 (2006). M. Sarrazin, J. P. Vigneron, Opt. Commun. 240 89 (2004) . X. F. Ren, G. P. Guo, Y. F. Huang, Z. W. Wang, and G. C. Guo, Appl. Phys. Lett. 90, 161112 (2007). F. L. Tejeira, S. G. Rodrigo, L. M. Moreno, F. J. G. Vidal, E. Devaux, T. W. Ebbesen, J. R. Krenn, I. P. Radko, S. I.Bozhevolnyi, M. U. Gonzalez, J. C. Weeber, and A. Dereux, Nature Physics 3, 324 (2007).
[^1]: E-mail: :[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The notion of regular cell complexes plays a central role in topological combinatorics because of its close relationship with posets. A generalization, called totally normal cellular stratified spaces, was introduced in [@1009.1851v5; @1106.3772] by relaxing two conditions; face posets are replaced by acyclic categories and cells with incomplete boundaries are allowed. The aim of this article is to demonstrate the usefulness of totally normal cellular stratified spaces by constructing a combinatorial model for the configuration space of graphs. As an application, we obtain a simpler proof of Ghrist’s theorem on the homotopy dimension of the configuration space of graphs. We also make sample calculations of the fundamental group of ordered and unordered configuration spaces of two points for small graphs.'
author:
- 'Mizuki Furuse, Takashi Mukouyama, and Dai Tamaki'
title: '**Totally Normal Cellular Stratified Spaces and Applications to the Configuration Space of Graphs**'
---
Dedicated to Professor Yuli Rudyak on the occasion of his 65th birthday.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We evaluate the Poynting vector generated by a heavy quark moving through a thermal state of ${\cal N}=4$ gauge theory using AdS/CFT. A significant diffusion wake is observed as well as a Mach cone. We discuss the ratio of the energy going into sound modes to the energy coming in from the wake.'
author:
- 'Steven S. Gubser'
- 'Silviu S. Pufu'
- Amos Yarom
bibliography:
- 'fourier.bib'
title: ' Sonic booms and diffusion wakes generated by a heavy quark in thermal AdS/CFT '
---
Introduction
============
The phenomenon of jet-splitting observed at the Relativistic Heavy Ion Collider (RHIC) [@Adler:2005ee; @Adams:2005ph] motivates a study of the stress tensor generated by a quark moving through a thermal state of ${\cal N}=4$ super-Yang-Mills theory (SYM). Although SYM is significantly different from quantum chromodynamics (QCD), it is widely hoped that comparisons between the two will lead to insights not readily extracted from conventional techniques such as perturbation theory and lattice simulations. The advantage of studying SYM is that powerful string theory methods exist to calculate gauge-singlet observables, such as the stress tensor, at strong coupling, starting from gravitational calculations in five-dimensional anti-de Sitter space (AdS${}_5$). These methods rely on the fact that SYM is a conformal field theory (CFT).
The basic string theory setup is well reviewed in earlier works [@Friess:2006aw; @Friess:2006fk; @Yarom:2007ni; @Gubser:2007nd; @Yarom:2007ap; @Gubser:2007xz; @Chesler:2007an]. Briefly, an infinitely massive, fundamentally charged quark that is dragged at constant velocity through an infinite, static, thermal plasma of SYM can be thought of as having a string trailing behind it down into AdS${}_5$-Schwarzschild, generating a drag force as computed in [@Herzog:2006gh; @Gubser:2006bz] (see also [@Liu:2006ug; @Casalderrey-Solana:2006rq]). This string perturbs the metric $g_{\mu\nu}$ of AdS${}_5$-Schwarzschild, and those perturbations can be translated into the expectation value $\langle T^{mn} \rangle$ of the stress tensor in the boundary gauge theory. Although it would perhaps be preferable to study an energetic gluon propagating through the thermal plasma, the infinitely heavy quark case is a textbook exercise in AdS/CFT because the quark has no dynamics apart from its constant-velocity motion: thus its trajectory can be regarded as part of the boundary conditions specified in the gravitational calculation. Infalling boundary conditions are specified at the horizon of AdS${}_5$-Schwarzschild, corresponding to causal physics in the boundary gauge theory.
In the first work on jet-splitting in AdS/CFT [@Friess:2006fk], Fourier coefficients $\langle T^{mn}(\vec{k}) \rangle$ were computed numerically, and high-$k$ and low-$k$ asymptotics were given. Subsequent work [@Yarom:2007ap; @Gubser:2007nd; @Yarom:2007ni; @Gubser:2007xz] has considerably refined the high-$k$ estimates. The position-space stress tensor can in principle be recovered from its Fourier coefficients using $$\langle T^{mn}(t,\vec{x}) \rangle =
\int {d^3 k \over (2\pi)^3} e^{i k_1 (x^1 - vt) + i k_2 x^2 + i k_3 x^3}
\langle T^{mn}(\vec{k}) \rangle \,. \label{ThreeFourier}$$ Here and below, $v$ is the quark’s velocity directed along the positive $x^1$ direction. Evidence of a sonic boom was already exhibited in [@Friess:2006fk], and in [@Gubser:2007xz; @Chesler:2007an] it was made more concrete by evaluating the full position-space energy density $\langle T^{00}(t,\vec{x}) \rangle$. A match to linearized hydrodynamics with $\eta/s = 1/4\pi$ was remarked upon in [@Gubser:2007xz; @Chesler:2007an]. What AdS/CFT provides beyond hydrodynamics is an all-scales description of dissipation, from the near-field of the quark all the way out to the linearized hydro regime.
A crucial question for application to RHIC phenomenology is the relative strength of the sonic boom and the diffusion wake. In the lab frame, the diffusion wake is a structure primarily visible in the Poynting vector, and it is concentrated within a parabolic region behind the quark. It is closely related to the well-known laminar wake of non-relativistic hydrodynamics. The forward flow in the diffusion wake should lead to enhanced particle production approximately collinear with the hard parton, as compared to vacuum fragmentation. The sonic boom, in contrast, describes high-angle emission. Intriguingly, data from the PHENIX experiment [@Adler:2005ee] favors suppression of the diffusion wake to such an extent that earlier authors [@Casalderrey-Solana:2006sq] have suggested turning it off altogether in a linearized hydro analysis. Other authors have proposed a model [@Renk:2006mv] in which it is assumed that $75\%$ percent of energy lost goes into a sonic boom. Data from STAR does not show as clear a minimum in associated particle production in the direction of the away-side parton [@Adams:2005ph]. In the string theory calculation, the relative strength of the sonic boom and the diffusion wake is fixed. Because other first-principles estimates of the relative strength of these two effects are lacking, it is clearly of interest to work out the string theory predictions in some detail. All the pieces are already in place except for sufficiently precise high-$k$ estimates of $\langle T^{0i}(\vec{k}) \rangle$ and a robust numerical scheme for performing the Fourier transforms. In this letter, we fill in these last gaps and extract explicit results for the Poynting vector. We also discuss one way of quantifying the relative strength of sonic and diffusive modes which suggests that, in the rest frame of the plasma, the sonic boom dissipates $1+v^2$ times as much energy as the diffusion wake feeds in.
Definitions
===========
The only scale in the problem is the temperature, so we will consistently work with dimensionless quantities $$\begin{aligned}
\vec{X} &\equiv& \pi T \vec{x} \\
\vec{K} &\equiv& \vec{k} / \pi T \,.
\end{aligned}$$ Instead of computing the expectation value of the complete stress tensor, we focus on a rescaled, subtracted quantity $$\begin{aligned}
&{\cal T}^{mn}(\vec{X}) \equiv
{\sqrt{1-v^2} \over (\pi T)^4 \sqrt{g_{YM}^2 N}}
\Big( \langle T^{mn}(0,\vec{x}) \rangle -
\langle T^{mn} \rangle_{\rm bath} \Big) \label{CurlyT} \\
&\langle T^{mn} \rangle_{\rm bath} \equiv
{\pi^2 \over 8} (N^2-1) T^4 {\rm diag}\{3,1,1,1\} \,.
\end{aligned}$$ On the right hand side of (\[CurlyT\]) we have set $t=0$, which leads to no loss of information because the computations are all done in a steady state approximation. The powers of $\pi T$ render ${\cal T}^{mn}$ dimensionless. The factor of $1/\sqrt{g_{YM}^2 N}$ cancels out an overall scaling with $\sqrt{g_{YM}^2 N}$ of all disturbances due to the trailing string. This scaling arises because it’s how the string tension depends on $g_{YM}$ and $N$. The quantity $\langle T^{mn} \rangle_{\rm bath}$ is just the stress tensor of the SYM plasma in the absence of the quark.
Our focus is on the first row of ${\cal T}^{mn}$, namely $${\cal T}^{0m} \equiv \begin{pmatrix} {\cal E} & {\cal S}_1 &
{\cal S}_2 & {\cal S}_3 \end{pmatrix} \,,$$ where ${\cal E}$ and $\vec{\cal S}$ are the rescaled, bath-subtracted energy density and Poynting vector. (We work in signature $-$+++, so there is no distinction between ${\cal S}^i$ and ${\cal S}_i$.) We further define $$\begin{aligned}
E &\equiv& {\cal E} - {\cal E}_{\rm Coulomb} \\
\vec{S} &\equiv& \vec{\cal S} - \vec{\cal S}_{\rm Coulomb} \,,
\end{aligned}$$ where ${\cal E}_{\rm Coulomb}$ and ${\cal S}_{\rm Coulomb}$ characterize the Coulombic near-field of the quark. Thus $E$ and $\vec{S}$ exclude not only the contribution of the bath in the absence of the quark, but also the contribution of the quark in the absence of the bath. The quantities which we eventually plot are $E$, $S_1$, and $S_\perp$ as functions of $X_1$ and $X_\perp$. Here $$X_\perp \equiv \sqrt{X_2^2 + X_3^2} \,,$$ and similarly for $K_\perp$ and $S_\perp$.
Asymptotics and subtractions
============================
The key to a robust numerical evaluation of stress-tensor components in position space is to make a split $$\begin{aligned}
{\cal E} &=& {\cal E}^{\rm UV} + {\cal E}^{\rm IR} +
{\cal E}^{\rm res} \\
\vec{\cal S} &=& \vec{\cal S}^{\rm UV} + \vec{\cal S}^{\rm IR} +
\vec{\cal S}^{\rm res} \,,
\end{aligned}$$ where ${\cal E}^{\rm UV}$ and $\vec{\cal S}^{\rm UV}$ are analytical approximations to ${\cal E}$ and $\vec{\cal S}$ for large $K$ which are valid to order $K^{-3}$, ${\cal E}^{\rm IR}$ and $\vec{\cal S}^{\rm IR}$ are analytical approximations for small $K$ valid to order $K^0$, and the remainders ${\cal E}^{\rm res}$ and $\vec{\cal S}^{\rm res}$ are uniformly bounded and integrable, so that they can be passed through a fast Fourier transform (FFT) with controllable errors. The full expressions for the analytical approximations are too long to be presented in full, but the leading terms are $$\begin{aligned}
{\cal E}^{\rm UV} &= -{2+v^2 \over 24} \sqrt{K_1^2 (1-v^2) +
K_\perp^2 + \mu_{\rm UV}^2} \nonumber \\
&\ {} +
{2 v^2 K_1^2 (1-v^2) + (2+v^2) \mu_{\rm UV}^2 \over
48 \sqrt{K_1^2 (1-v^2) + K_\perp^2 + \mu_{\rm UV}^2}} +
{\cal O}(K^{-1}) \label{EUV}\\
{\cal S}_1^{\rm UV} &= -{v \over 8} \sqrt{K_1^2 (1-v^2) +
K_\perp^2 + \mu_{\rm UV}^2} \nonumber \\
&\ {} +
{2 v K_1^2 (1-v^2) + 3 v \mu_{\rm UV}^2 \over
48 \sqrt{K_1^2 (1-v^2) + K_\perp^2 + \mu_{\rm UV}^2}} +
{\cal O}(K^{-1}) \label{S1UV} \\
{\cal S}_2^{\rm UV} &= {v K_1 K_2 (1-v^2) \over
24 \sqrt{K_1^2 (1-v^2) + K_\perp^2 + \mu_{\rm UV}^2}} +
{\cal O}(K^{-1}) \label{SpUV}\\
{\cal E}^{\rm IR} &=
-{1\over 2\pi} {3 i v K_1 (1+v^2) - 3 v^2 K_1^2 \over K^2 -
3 v^2 K_1^2 - i v K^2 K_1} \nonumber \\
&\ {}
+{1\over 2\pi} {3 i v K_1 (1+v^2) - 3 v^2 K_1^2 \over K^2 -
3 v^2 K_1^2 - i v K^2 K_1 + \mu_{\rm IR}^2} \label{EIR} \\
{\cal S}_1^{\rm IR} &=
- {1\over 2\pi} {i (1+v^2) K_1 + v K^2 - 2 v^3 K_1^2 \over
K^2 - 3 v^2 K_1^2 - i v K^2 K_1} \nonumber \\
&\ {} + {2 v\over \pi} {1 + i K_1/4 v \over K^2 - 4 i v K_1}
+ \hbox{(regulators)} \label{S1IR} \\
{\cal S}_2^{\rm IR} &=
-{1\over 2 \pi} {i (1+v^2) K_2 + v^3 K_1 K_2 \over K^2 -
3 v^2 K_1^2 - i v K^2 K_1} \nonumber \\
&\ {} +
{1\over 2\pi} {i K_2\over K^2 - 4 i v K_1} + \hbox{(regulators)}
\label{SpIR}\,.
\end{aligned}$$ Expressions for ${\cal S}_3^{\rm UV}$ and ${\cal S}_3^{\rm IR}$ may be deduced by rotational invariance around the $K_1$ axis. The omitted terms in (\[S1IR\]) and (\[SpIR\]) denoted “$\hbox{(regulators)}$” are analogous to the second term in (\[EIR\]).
The strategy for obtaining UV subtractions was explained in [@Gubser:2007xz]: we solve for the metric perturbations at large $K$ using an iterative Green’s function approximation, as first introduced in [@Yarom:2007ap], then “soften” the resulting power series through the introduction of a dimensionless parameter $\mu_{\rm UV}$ as in (\[EUV\])–(\[SpUV\]). To obtain the IR subtractions, we draw upon results of [@Friess:2006fk; @Gubser:2007xz] to find series expansions at small $K$ for components of ${\cal T}^{mn}$, then partially resum the series, as in (\[EIR\])–(\[SpIR\]), to eliminate divergences on the Mach cone and the $X_1$ axis, then add Pauli-Villars-style regulators involving another dimensionless parameter $\mu_{\rm IR}$.
The terms in (\[EIR\])–(\[SpIR\]) with cubic denominators do not appear to admit analytical Fourier transforms; however they can easily be Fourier transformed in the $K_\perp$ directions, leaving a final one-dimensional FFT to be performed numerically. Some of the subleading terms in our subtractions do not by themselves satisfy appropriate conservation properties, so the conservation of the full stress tensor has to be restored by combining numerical and analytical results.
Conservation
============
The stress tensor must be conserved except at the location of the quark, and the failure of conservation there matches the drag force on the quark [@Friess:2006fk]. In terms of our rescaled variables, and in the rest frame of the plasma, $$\begin{aligned}
{\partial \over \partial X^m} {\cal T}^{mn} &=& {\cal J}^n =
-{\cal F}_{\rm drag}^n \delta(X^1-vX^0) \\
\label{dTJ}
{\cal F}_{\rm drag}^m &\equiv&
-{1 \over 2\pi} \begin{pmatrix} v^2 & v & 0 & 0 \end{pmatrix} \,.
\end{aligned}$$ After using the co-moving ansatz, (\[dTJ\]) becomes $$\begin{aligned}
&i K_m {\cal T}^{mn}(\vec{K}) = -{\cal F}_{\rm drag}^n \\
&K_m = \begin{pmatrix} -v K_1 & K_1 & K_2 & K_3 \end{pmatrix} \,.
\end{aligned}$$ Now consider a split of the IR asymptotic values of $\mathcal{T}^{mn}$ $${\cal T}_{\rm IR}^{0m} = {\cal T}^{0m}_{\rm sound} +
{\cal T}^{0m}_{\rm diffuse} \label{TSplit}$$ where we define ${\cal T}^{0m}_{\rm sound}$ as all the terms in (\[EIR\])–(\[SpIR\]) with cubic polynomials in the denominator; ${\cal T}^{0m}_{\rm diffuse}$ by definition comprises all the remaining terms. With this split, $$\begin{aligned}
\lim_{K \to 0} i K_m {\cal T}_{\rm sound}^{0m} &=&
{1 + v^2 \over 2\pi} =
-\left( 1 + {1 \over v^2} \right) {\cal F}_{\rm drag}^0
\label{SoundForce}\\
\lim_{K \to 0} i K_m {\cal T}_{\rm diffuse}^{0m} &=&
-{1 \over 2\pi} =
{1 \over v^2} {\cal F}_{\rm drag}^0 \,. \label{DiffuseForce}
\end{aligned}$$ The $m=0$ term in the left hand side of (\[SoundForce\]) is, formally, the time derivative of the total energy in sound modes. Adding in the $m=1,2,3$ terms accounts for energy flow into sound modes at large $X$, and the sum over all $m$ is the total rate of energy loss by the quark into sound modes. Likewise, (\[DiffuseForce\]) is the energy loss into diffusion modes. But the relative sign between (\[SoundForce\]) and (\[DiffuseForce\]) shows that while sound modes carry energy away from the quark, the diffusion wake actually feeds energy in toward the quark.
Energy loss to sound modes, as measured by (\[SoundForce\]), is greater by a factor of $1+1/v^2$ than the total energy loss, and $1+v^2$ times the energy fed in by the diffusion wake [^1]. These ratios should be regarded with some caution because the argument leading to them is essentially algebraic; however, integrating the energy flux across the diffusion wake confirms (\[DiffuseForce\]).
Instead of calculating in the rest frame of the plasma, as we have done, one may instead perform computations analogous to (\[SoundForce\]) and (\[DiffuseForce\]) in the rest frame of the quark. Then, using also the IR asymptotics of $\mathcal{T}_{11}$ obtained in [@Friess:2006fk], one can show that the energy lost by the quark into sound modes precisely cancels the energy fed in through diffusion modes. Some intuition regarding this result may be gained by noting that in the quark rest frame, the zero component of the drag force vanishes.
Numerical results
=================
Our numerical analysis was mostly done using $128^3$ grids for the three-dimensional FFT’s of ${\cal E}^{\rm res}$ and $\vec{\cal S}^{\rm res}$, with wave-numbers $\vec{K}$ ranging from $-10$ to $10$ in each component. The one-dimensional FFT’s of the terms in ${\cal E}^{\rm IR}$ and $\vec{\cal S}^{\rm IR}$ with cubic denominators were performed on finer grids, usually with 1944 points, with $K_1$ ranging from $-20$ to $20$. We chose $\mu_{\rm UV} = \mu_{\rm IR} = 1$. Conservation was checked by comparing $\dot{\cal E}$ to $\nabla \cdot \vec{\cal S}$ in position space. This is a stringent check because derivatives of numerically known functions tend to be noisy. Conservation worked to within a few percent of $\dot{\cal E}$ except when $\dot{\cal E}$ became very small. A summary of our numerical results for $v=0.75$ is shown in figure \[BIGFIG\]. A more complete presentation of our numerical results is available [^2].
-0.5in{width="7.5in"}
Conclusions
===========
The Poynting vector describing the disturbances of a thermal plasma of ${\cal N}=4$ super-Yang-Mills theory from a moving quark exhibits a sonic boom and a diffusion wake, as expected on general grounds. Using a natural split of the energy flow into sonic and diffusive modes, one finds—in the rest frame of the plasma—that energy is lost from the quark through sound modes and fed in toward the quark through the diffusion wake. These two effects stand in the ratio $1+v^2 : -1$. Thus the sonic boom is a strong effect, but the diffusion wake is also significant.
The work of S. Gubser and S. Pufu was supported in part by the Department of Energy under Grant No. DE-FG02-91ER40671, and by the Sloan Foundation. A. Yarom is supported in part by the German Science Foundation and by the Minerva foundation.
[^1]: Another way of splitting ${\cal T}^{0m}$ into sound and diffusion modes is to associate ${\cal E}$ and the longitudinal component of the Poynting vector $\vec{\cal S} \cdot \hat{K}$ with the sound mode, and to associate the transverse part of $\vec{\cal S}$ (orthogonal to $\hat{K}$) with the diffusion mode. This has the advantage that it decouples the equations of linearized hydrodynamics. This decomposition nearly agrees with the split discussed in (\[TSplit\]), but additional non-local terms are added to the longitudinal modes and subtracted from the transverse modes which lead to non-zero but canceling values of $\vec{\cal S}^{\rm longitudinal}$ and $\vec{\cal S}^{\rm transverse}$ far ahead of the quark. If this split is followed, then formally one finds zero net energy going into the transverse mode. We thank D. Teaney for discussions related to these points.
[^2]: [http://physics.princeton.edu/$\sim$ssgubser/papers/fourier/details.pdf]{}.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For an integral convex polytope $\Pc \subset \RR^N$ of dimension $d$, we call $\delta(\Pc)=(\delta_0, \delta_1, \ldots, \delta_d)$ the $\delta$-vector of $\Pc$ and $\vol(\Pc)=\sum_{i=0}^d\delta_i$ its normalized volume. In this paper, we will establish the new equalities and inequalities on $\delta$-vectors for integral simplices whose normalized volumes are prime. Moreover, by using those, we will classify all the possible $\delta$-vectors of integral simplices with normalized volume 5 and 7.'
address: 'Akihiro Higashitani, Department of Pure and Applied Mathematics, Graduate School of Information Science and Technology, Osaka University, Toyonaka, Osaka 560-0043, Japan'
author:
- Akihiro Higashitani
title: |
Ehrhart polynomials of integral simplices\
with prime volumes
---
[^1]
Introduction {#introduction .unnumbered}
============
One of the most fascinating problems on enumerative combinatorics is to characterize the $\delta$-vectors of integral convex polytopes.
Let $\Pc \subset \RR^N$ be an [*integral*]{} convex polytope of dimension $d$, which is a convex polytope any of whose vertices has integer coordinates. Let $\partial \Pc$ denote the boundary of $\Pc$. Given a positive integer $n$, we define $$i(\Pc,n) = |n\Pc \cap \ZZ^N|, \;\;\;\;\;
i^*(\Pc,n)=|n (\Pc \setminus \partial \Pc) \cap \ZZ^N|,$$ where $n\Pc = \{ n\alpha : \alpha \in \Pc \}$ and $|X|$ is the cardinality of a finite set $X$. The enumerative function $i(\Pc,n)$ has the following fundamental properties, which were studied originally in the work of Ehrhart [@Ehrhart]:
- $i(\Pc,n)$ is a polynomial in $n$ of degree $d$;
- $i(\Pc,0) = 1$;
- (loi de réciprocité) $i^*(\Pc,n)=( - 1 )^d i(\Pc, - n)$ for every integer $n > 0$.
This polynomial $i(\Pc,n)$ is called the [*Ehrhart polynomial*]{} of $\Pc$. We refer the reader to [@BeckRobins Chapter 3], [@HibiRedBook Part II] or [@StanleyEC pp. 235–241] for the introduction to the theory of Ehrhart polynomials.
We define the sequence $\delta_0, \delta_1, \delta_2, \ldots $ of integers by the formula $$\begin{aligned}
\label{delta}
(1 - \lambda)^{d + 1} \sum_{n=0}^{\infty} i(\Pc,n) \lambda^n
= \sum_{i=0}^{\infty} \delta_i \lambda^i. \end{aligned}$$ Then, from a fundamental result on generating functions ([@StanleyEC Corollary 4.3.1]), we know that $\delta_i = 0$ for $i > d$. We call the integer sequence $$\delta(\Pc)= (\delta_0, \delta_1, \ldots, \delta_d),$$ which appears in , the [*$\delta$-vector*]{} of $\Pc$. The $\delta$-vector has the following properties:
- $\delta_0=1$, $\delta_1 = |\Pc \cap \ZZ^N| - (d + 1)$ and $\delta_d = |(\Pc \setminus \partial \Pc) \cap \ZZ^N|.$ Hence, $\delta_1 \geq \delta_d$. In particular, when $\delta_1=\delta_d$, $\Pc$ must be a simplex.
- Each $\delta_i$ is nonnegative ([@StanleyDRCP]).
- If $(\Pc \setminus \partial \Pc) \cap \ZZ^N$ is nonempty, then one has $\delta_1 \leq \delta_i$ for every $1 \leq i \leq d - 1$ ([@HibiLBT]).
- The leading coefficient $(\sum_{i=0}^d\delta_i)/d!$ of $i(\Pc,n)$ is equal to the usual volume of $\Pc$ ([@StanleyEC Proposition 4.6.30]). In particular, the positive integer $\vol(\Pc) = \sum_{i=0}^d\delta_i$ is said to be the [*normalized volume*]{} of $\Pc$.
Recently, the $\delta$-vectors of integral convex polytopes have been studied intensively. For example, see [@HigashiSS; @Staple1; @Staple2].
There are two well-known inequalities on $\delta$-vectors. Let $s = \max\{ i : \delta_i \neq 0 \}$. One is $$\begin{aligned}
\label{Stanley}
\delta_0 + \delta_1 + \cdots + \delta_i
\leq \delta_s + \delta_{s-1} + \cdots + \delta_{s-i},
\;\;\;\;\; 0 \leq i \leq \lfloor s/2 \rfloor, \end{aligned}$$ which is proved by Stanley [@StanleyJPAA], and another one is $$\begin{aligned}
\label{Hibi}
\delta_d + \delta_{d-1} + \cdots + \delta_{d-i}
\leq \delta_1 + \delta_2 + \cdots + \delta_i + \delta_{i+1},
\;\;\;\;\; 0 \leq i \leq \lfloor (d-1)/2 \rfloor, \end{aligned}$$ which appears in the work of Hibi [@HibiLBT Remark (1.4)].
On the classification problem on $\delta$-vectors of integral convex polytopes, the above inequalities and characterize the possible $\delta$-vectors completely when $\sum_{i=0}^d\delta_i \leq 3$ ([@HHN Theorem 0.1]). Moreover, when $\sum_{i=0}^d\delta_i = 4$, the possible $\delta$-vectors are determined completely by and together with an additional condition ([@HHNan Theorem 5.1]). Furthermore, by the proofs of [@HHN Theorem 0.1] and [@HHNan Theorem 5.1], we know that all the possible $\delta$-vectors can be realized as the $\delta$-vectors of integral simplices when $\sum_{i=0}^d \delta_i \leq 4$. However, unfortunately, this is no longer true when $\sum_{i=0}^d \delta_i = 5$. (See [@HHNan Remark 5.2].) Hence, for the further classifications of $\delta$-vectors with $\sum_{i=0}^d\delta_i \geq 5$, it is natural to study $\delta$-vectors of integral simplices at first. Even for non-simplex cases, since every convex polytope can be triangulated into finitely many simplices and we can compute the $\delta$-vecotor of an integral convex polytope from its triangulation, investigating $\delta$-vectors of integral simplices is an essential and important work.
In this paper, in particular, we establish some new constraints on $\delta$-vectors for integral simplices whose normalized volumes are prime numbers. The following theorem is our main result of this paper.
\[main\] Let $\Pc$ be an integral simplex of dimension $d$ and $\delta(\Pc)=(\delta_0,\delta_1,\ldots,\delta_d)$ its $\delta$-vector. Suppose that $\vol(\Pc)=\sum_{i=0}^d \delta_i=p$ is an odd prime number. Let $i_1,\ldots,i_{p-1}$ be the positive integers such that $\sum_{i=0}^d\delta_it^i=1+t^{i_1}+\cdots+t^{i_{p-1}}$ with $1 \leq i_1 \leq \cdots \leq i_{p-1} \leq d$. Then,
- [*(cf. [@HigashiSS Theorem 2.3])*]{} one has $$\begin{aligned}
i_1+i_{p-1}=i_2+i_{p-2}=\cdots=i_{(p-1)/2}+i_{(p+1)/2} \leq d+1; \end{aligned}$$
- one has $$\begin{aligned}
i_k+i_{\ell} \geq i_{k+ \ell} \; \text{ for } \; 1 \leq k \leq \ell \leq p-1
\; \text{ with } \; k+ \ell \leq p-1. \end{aligned}$$
We give a proof of Theorem \[main\] in Section \[sec:review\].
Now, we remark that the part (a) of Theorem \[main\] is not a new result in some sense. In [@HigashiSS Theorem 2.3], the author proved that for an integral simplex $\Pc$ with prime normalized volume, if $i_1+i_{p-1}=d+1$, then $\Pc$ is shifted symmetric, i.e., we have $i_1+i_{p-1}=i_2+i_{p-2}=\cdots=i_{(p-1)/2}+i_{(p+1)/2}$. Moreover, since every integral simplex with prime normalized volume is either a simplex with $i_1+i_{p-1}=d+1$ or a pyramid at height 1 over such simplex and taking such a pyramid does not change the normalized volume and the polynomial $1+\sum_{j=1}^{p-1}t^{i_j}$, we also obtain the equalities $i_1+i_{p-1}=i_2+i_{p-2}=\cdots=i_{(p-1)/2}+i_{(p+1)/2}$ on the case where $i_1+i_{p-1}<d+1$. On the other hand, in this paper, we give an another proof of this statement. More precisely, we give an elementary proof of Theorem \[main\] (a) in terms of some abelian groups associated with integral simplices.
In addition, as an application of Theorem \[main\], we give a complete characterization of the possible $\delta$-vectors of integral simplices when $\sum_{i=0}^d\delta_i=5$ and 7.
\[delta5\] Given a finite sequence $(\delta_0,\delta_1,\ldots,\delta_d)$ of nonnegative integers, where $\delta_0=1$ and $\sum_{i=0}^d\delta_i=5$, there exists an integral simplex $\Pc \subset \RR^d$ of dimension $d$ whose $\delta$-vector coincides with $(\delta_0,\delta_1,\ldots,\delta_d)$ if and only if $i_1,\ldots,i_4$ satisfy $i_1+i_4=i_2+i_3 \leq d+1$ and $i_k+i_{\ell} \geq i_{k+\ell}$ for $1 \leq k \leq \ell \leq 4$ with $k+\ell \leq 4$, where $i_1,\ldots,i_4$ are the positive integers such that $\sum_{i=0}^d\delta_it^i=1+t^{i_1}+\cdots+t^{i_4}$ with $1 \leq i_1 \leq \cdots \leq i_4 \leq d$.
\[delta7\] Given a finite sequence $(\delta_0,\delta_1,\ldots,\delta_d)$ of nonnegative integers, where $\delta_0=1$ and $\sum_{i=0}^d\delta_i=7$, there exists an integral simplex $\Pc \subset \RR^d$ of dimension $d$ whose $\delta$-vector coincides with $(\delta_0,\delta_1,\ldots,\delta_d)$ if and only if $i_1,\ldots,i_6$ satisfy $i_1+i_6=i_2+i_5=i_3+i_4 \leq d+1$ and $i_k+i_{\ell} \geq i_{k+\ell}$ for $1 \leq k \leq \ell \leq 6$ with $k+\ell \leq 6$, where $i_1,\ldots,i_6$ are the positive integers such that $\sum_{i=0}^d\delta_it^i=1+t^{i_1}+\cdots+t^{i_6}$ with $1 \leq i_1 \leq \cdots \leq i_6 \leq d$.
By virtue of Theorem \[main\], the “Only if” parts of Theorem \[delta5\] and \[delta7\] are obvious. A proof of the “If” part of Theomre \[delta5\] is given in Section \[sec:classify5\] and that of Theorem \[delta7\] is given in Section \[sec:classify7\]. Moreover, in Section \[sec:towards\], we note some problems towards the classification of Ehrhart polynomials of integral convex polytopes with general normalized volumes.
[10]{} M. Beck and S. Robins, “Computing the Continuous Discretely,” Undergraduate Texts in Mathematics, Springer, 2007. E. Ehrhart, “Polynômes Arithmétiques et Méthode des Polyèdres en Combinatoire,” Birkhäuser, Boston/Basel/Stuttgart, 1977. T. Hibi, “Algebraic Combinatorics on Convex Polytopes,” Carslaw Publications, Glebe NSW, Australia, 1992. T. Hibi, A lower bound theorem for Ehrhart polynomials of convex polytopes, [*Adv. in Math.*]{} [**105**]{} (1994), 162 – 165. T. Hibi, A. Higashitani and N. Li, Hermite normal forms and $\delta$-vectors, [*J. Comb. Theory Ser. A*]{} [**119**]{} (2012), 1158–1173.
T. Hibi, A. Higashitani and Y. Nagazawa, Ehrhart polynomials of convex polytopes with small volume, [*European J. Combinatorics*]{} [**32**]{} (2011), 226–232.
A. Higashitani, Shifted symmetric $\delta$-vectors of convex polytopes, [*Discrete Math.*]{} [**310**]{} (2010), 2925–2934. G. Károlyi, A compactness argument in the additive theory and the polynomial method, [*Discrete Math.*]{} [**302**]{} (2005), 124–144.
M. B. Nathanson, “Additive number theory. Inverse problems and the geometry of sumsets,” Graduate Texts in Mathematics 165, Springer-Verlag, New York, 1996.
R. P. Stanley, “Enumerative Combinatorics, Volume 1,” Wadsworth & Brooks/Cole, Monterey, Calif., 1986. R. P. Stanley, Decompositions of rational convex polytopes, [*Annals of Discrete Math.*]{} [**6**]{} (1980), 333 – 342. R. P. Stanley, On the Hilbert function of a graded Cohen–Macaulay domain, [*J. Pure and Appl. Algebra*]{} [**73**]{} (1991), 307 – 314. A. Stapledon, Inequalities and Ehrhart $\delta$-vectors, [*Trans. Amer. Math. Soc.*]{} [**361**]{} (2009), 5615–5626. A. Stapledon, Additive number theorem and inequalities in Ehrhart theory, arXiv:0904.3035v1.
[^1]: [**2010 Mathematics Subject Classification:**]{} Primary 52B20; Secondary 52B12.\
Integral simplex, Ehrhart polynomial, $\delta$-vector, Cauchy–Davenport theorem\
The author is supported by JSPS Research Fellowship for Young Scientists.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We address the question whether the condition on a fusion category being solvable or not is determined by its fusion rules. We prove that the answer is affirmative for some families of non-solvable examples arising from representations of semisimple Hopf algebras associated to exact factorizations of the symmetric and alternating groups. In the context of spherical fusion categories, we also consider the invariant provided by the $S$-matrix of the Drinfeld center and show that this invariant does determine the solvability of a fusion category provided it is group-theoretical.'
address: 'Facultad de Matemática, Astronomía y Física. Universidad Nacional de Córdoba. CIEM – CONICET. Ciudad Universitaria. (5000) Córdoba, Argentina'
author:
- 'Melisa Esca\~ nuela Gonz'' alez and Sonia Natale'
date: 'December 17, 2015'
title: On fusion rules and solvability of a fusion category
---
[^1]
Introduction
============
Throughout this paper we shall work over an algebraically closed field $k$ of characteristic zero. Let $G$ be a finite group. An important invariant of $G$ is given by its *character table*, defined as the collection $$\{\chi_i(g_j)\}_{0\leq i, j \leq n},$$ where $\epsilon = \chi_0, \dots, \chi_n$ are the irreducible characters of $G$ over $k$ and $e = g_0, \dots, g_n$, are representatives of the conjugacy classes of $G$. Several structural properties of $G$ can be read off from its character table. For instance, the character table of $G$ allows to determine the lattice of normal subgroups of $G$ and to decide if the group $G$ is nilpotent or solvable. See [@isaacs pp. 23]. It is known, however, that the character table of a finite solvable group $G$ does not determine its derived length [@mattarei], [@mattarei-2].
In particular, if $G$ and $\Gamma$ are finite groups with the same character table, then $G$ is solvable if and only if $\Gamma$ is solvable. In addition, the knowledge of the character table of a finite group $G$ is equivalent to the knowledge of the structure constants, in the canonical basis consisting of isomorphism classes of irreducible representations, of the Grothendieck ring of the fusion category ${\operatorname{Rep}}G$ of finite dimensional representations of $G$ over $k$, so-called the *fusion rules* of ${\operatorname{Rep}}G$.
The notions of nilpotency and solvability of a group $G$ have been extended to general fusion categories in [@gel-nik], [@ENO2]. Let ${{\mathcal C}}$ be a fusion category over $k$. Then ${{\mathcal C}}$ is *nilpotent* if there exists a series of fusion subcategories $$\label{cs-nil}{\operatorname{Vect}}= {{\mathcal C}}_0 \subseteq {{\mathcal C}}_{1} \subseteq \dots \subseteq {{\mathcal C}}_n = {{\mathcal C}},$$ and a series of finite groups $G_1, \dots, G_n$, such that ${{\mathcal C}}_i$ is a $G_i$-extension of ${{\mathcal C}}_{i-1}$, for all $i = 1, \dots, n$. On the other side, ${{\mathcal C}}$ is *solvable* if there exist a sequence of fusion categories ${\operatorname{Vect}}= {{\mathcal C}}_0,
\dots, {{\mathcal C}}_n = {{\mathcal C}}$, $n \geq 0$, and a sequence of cyclic groups of prime order $G_1, \dots, G_n$, such that, for all $1\leq i \leq n$, ${{\mathcal C}}_i$ is a $G_i$-equivariantization or a $G_i$-extension of ${{\mathcal C}}_{i-1}$. See Subsection \[nilp-solv\]. Some features related to nilpotency and solvability have been extended as well from the context of finite groups to that of fusion categories; remarkably, an analogue of Burnside’s $p^aq^b$-theorem was established for fusion categories in [@ENO2].
It is apparent from the definition of nilpotency of a fusion category ${{\mathcal C}}$ given in [@gel-nik] that this property depens only upon the Grothendieck ring of ${{\mathcal C}}$, that is, it is determined by its fusion rules. In this paper we address the question whether the solvability of a fusion category ${{\mathcal C}}$ is also determined by its fusion rules.
Since a solvable fusion category has nontrivial invertible objects and a simple group has no nontrivial one-dimensional representation, then no solvable fusion category can have the same fusion rules as a simple finite group. We show that if ${{\mathcal C}}$ is a fusion category with the same fusion rules as a dihedral group, then ${{\mathcal C}}$ is solvable. On the other hand, if ${{\mathcal C}}$ has the fusion rules of a symmetric group ${\mathbb{S}}_n$, $n \geq 5$, then ${{\mathcal C}}$ is not solvable; Theorem \[dihedral-fr\] and Corollary \[fr-sn\].
We study some families of examples of non-solvable fusion categories arising from representations of semisimple Hopf algebras associated to exact factorizations of the symmetric group ${\mathbb{S}}_n$ and the alternating group ${\mathbb A}_n$. For a wide class of such fusion categories ${{\mathcal C}}$, we show that ${{\mathcal C}}$ cannot have the fusion rules of any solvable fusion category. See Theorems \[jp-kp\], \[dual-jn-kn\], \[bn\*\] and \[bn\].
In the context of braided fusion categories, the solvability of a fusion category ${{\mathcal C}}$ is related to the existence of Tannakian subcategories of ${{\mathcal C}}$; it is known that if ${{\mathcal C}}$ is a non-pointed integral solvable braided fusion category, then it must contain a nontrivial Tannakian subcategory [@witt-wgt Lemma 5.1].
We show that if $\tilde {{\mathcal C}}$ is a non-pointed braided fusion category which has the same fusion rules as a solvable fusion category ${{\mathcal C}}$, then $\tilde {{\mathcal C}}$ contains a nontrivial Tannakian subcategory. See Theorem \[e-tann\].
For a spherical fusion category ${{\mathcal C}}$ we study a somehow stronger invariant, analogous to the character table of a finite group, consisting of the $S$-matrix of the Drinfeld center ${{\mathcal Z}}({{\mathcal C}})$ of ${{\mathcal C}}$. Indeed, the $S$-matrix of a modular category ${{\mathcal D}}$ is usually named the ’character table’ of ${{\mathcal D}}$ in the literature; see for instance [@ganchev]. A celebrated formula due to Verlinde, and valid for any modular category, implies that the $S$-matrix of ${{\mathcal Z}}({{\mathcal C}})$ determines its fusion rules. We call two spherical fusion categories $S$-equivalent if their Drinfeld centers have ’the same’ $S$-matrix; see Subsection \[s-equiv\].
We prove in Theorem \[s-equiv-gt\] that the $S$-matrix of the Drinfeld center does determine the solvability of a group-theoretical fusion category. That is, if ${{\mathcal C}}$ and ${{\mathcal D}}$ are $S$-equivalent spherical fusion categories and ${{\mathcal C}}$ is group-theoretical, then ${{\mathcal C}}$ is solvable if and only if ${{\mathcal D}}$ is solvable. We also show that being group-theoretical is a property invariant under $S$-equivalence, that is, it is a property determined by the $S$-matrix of the Drinfeld center; see Theorem \[s-gpttic\].
The paper is organized as follows. Section \[preli\] contains the main notions and facts on fusion categories that will be needed in the rest of the paper. In Section \[g-crossed\] we study the notion of Grothendieck equivalence of fusion categories and its connection with solvability, and prove some results on the fusion rules of dihedral and symmetric groups. In Section \[nsol-fr\] we consider examples of non-solvable fusion categories arising from exact factorizations of the symmetric and alternating groups. The case of braided fusion categories is studied in Section \[solv-fr-bfc\]. Finally, in Section \[s-char-tbl\] we study the notion of $S$-equivalence of spherical fusion categories.
Preliminaries {#preli}
=============
The category of finite dimensional vector spaces over $k$ will be denoted by ${\operatorname{Vect}}$. A fusion category over $k$ is a semisimple rigid monoidal category over $k$ with finitely many isomorphism classes of simple objects, finite-dimensional Hom spaces, and such that the unit object ${\textbf{1}}$ is simple. Unless otherwise stated, all tensor categories will be assumed to be strict. We refer the reader to [@ENO], [@DGNOI], for the main notions on fusion categories used throughout.
Fusion categories
-----------------
Let ${{\mathcal C}}$ be a fusion category over $k$. The Grothendieck group $K_0({{\mathcal C}})$ is a free abelian group with basis ${\operatorname{Irr}}({{\mathcal C}})$ consisting of isomorphism classes of simple objects of ${{\mathcal C}}$. For an object $X$ of ${{\mathcal C}}$, let us denote by $[X]$ its class in $K_0({{\mathcal C}})$.
The tensor product of ${{\mathcal C}}$ endows $K_0({{\mathcal C}})$ with a ring structure with unit element $[{\textbf{1}}]$ and such that, for all objects $X$ and $Y$ of ${{\mathcal C}}$, $[X][Y] = [X\otimes Y]$. Let $X, Y \in {\operatorname{Irr}}({{\mathcal C}})$. Then one can write $$X Y = \sum_{Z \in {\operatorname{Irr}}({{\mathcal C}})}N^Z_{X, Y} \, Z,$$ where $N^Z_{X, Y}$ are non-negative integers, for all $X, Y, Z \in {\operatorname{Irr}}({{\mathcal C}})$. The collection of numbers $\{N^Z_{X, Y}\}_{X, Y, Z}$ are called the *fusion rules* of ${{\mathcal C}}$ and they determine the ring structure of $K_0({{\mathcal C}})$. They are given by the formula $$N^Z_{X, Y} = \dim {\operatorname{Hom}}_{{\mathcal C}}(Z, X \otimes Y),$$ for all $X, Y, Z \in {\operatorname{Irr}}({{\mathcal C}})$. In the terminology of [@gel-nik Subsection 2.1], the pair $(K_0({{\mathcal C}}),{\operatorname{Irr}}({{\mathcal C}}))$ is a unital *based* ring.
A *fusion subcategory* of ${{\mathcal C}}$ is a full tensor subcategory ${{\mathcal D}}$ such that ${{\mathcal D}}$ is replete and stable under direct summands. Fusion subcategories of ${{\mathcal C}}$ are in bijective correspondence with subrings of $K_0({{\mathcal C}})$ spanned by a subset of ${\operatorname{Irr}}({{\mathcal C}})$, that is, based subrings of $K_0({{\mathcal C}})$.
The Frobenius-Perron dimension of a simple object $X \in {{\mathcal C}}$ is, by definition, the Frobenius-Perron eigenvalue of the matrix of left multiplication by the class of $X$ in the basis ${\operatorname{Irr}}({{\mathcal C}})$ of the Grothendieck ring of ${{\mathcal C}}$ consisting of isomorphism classes of simple objects. The Frobenius-Perron dimension of ${{\mathcal C}}$ is the number ${\operatorname{FPdim}}{{\mathcal C}}=
\sum_{X \in {\operatorname{Irr}}({{\mathcal C}})} ({\operatorname{FPdim}}X)^2$.
We shall indicate by ${\mathrm{cd}}({{\mathcal C}})$ the set of Frobenius-Perron dimensions of simple objects of ${{\mathcal C}}$. If $1 = d_0, d_1, \dots, d_r$ are distinct positive real numbers and $n_1, \dots, n_r$ are natural numbers, we shall say that ${{\mathcal C}}$ is of *type* $(d_0, n_0; d_1, n_1; \dots; d_r, n_r)$ if ${{\mathcal C}}$ has $n_i$ isomorphism classes of simple objects of Frobenius-Perron dimension $d_i$, for all $i = 0, \dots, r$.
The group of invertible objects of ${{\mathcal C}}$ will be denoted by $G({{\mathcal C}})$. Thus $G({{\mathcal C}})$ coincides with the subset of elements $Y$ of ${\operatorname{Irr}}({{\mathcal C}})$ such that ${\operatorname{FPdim}}Y = 1$. Thus, if ${{\mathcal C}}$ is of type $(1, n_0; d_1, n_1; \dots; d_r, n_r)$, then $n_0 = |G({{\mathcal C}})|$.
The category ${{\mathcal C}}$ is called *integral* if ${\operatorname{FPdim}}X \in \mathbb Z$, for all simple object $X \in {{\mathcal C}}$, and it is called *weakly integral* if ${\operatorname{FPdim}}{{\mathcal C}}\in \mathbb Z$.
Recall that a right module category over a fusion category ${{\mathcal C}}$ is a finite semisimple $k$-linear abelian category ${\mathcal{M}}$ endowed with a bifunctor $\otimes: {\mathcal{M}}\times {{\mathcal C}}\to {\mathcal{M}}$ satisfying the associativity and unit axioms for an action, up to coherent natural isomorphisms. The module category ${\mathcal{M}}$ is called indecomposable if it is not equivalent as a module category to a direct sum of non-trivial module categories. If ${\mathcal{M}}$ is an indecomposable module category over ${{\mathcal C}}$, then the category ${{\mathcal C}}^*_{\mathcal M}$ of ${{\mathcal C}}$-module endofunctors of ${\mathcal{M}}$ is also a fusion category.
Two fusion categories ${{\mathcal C}}$ and ${{\mathcal D}}$ are *Morita equivalent* if ${{\mathcal D}}$ is equivalent to ${{\mathcal C}}^*_{\mathcal M}$ for some indecomposable module category $\mathcal M$. If ${{\mathcal C}}$ and ${{\mathcal D}}$ are Morita equivalent fusion categories, then ${\operatorname{FPdim}}{{\mathcal C}}= {\operatorname{FPdim}}{{\mathcal D}}$.
By [@ENO2 Theorem 3.1], the fusion categories ${{\mathcal C}}$ and ${{\mathcal D}}$ are Morita equivalent if and only if its Drinfeld centers are equivalent as braided fusion categories.
A fusion category ${{\mathcal C}}$ is *pointed* if all its simple objects are invertible. If ${{\mathcal C}}$ is a pointed fusion category, then there exist a finite group $G$ and a 3-cocycle $\omega$ on $G$ such that ${{\mathcal C}}$ is equivalent to the category ${{\mathcal C}}(G, \omega)$ of finite-dimensional $G$-graded vector spaces with associativity constraint defined by $\omega$. A fusion category Morita equivalent to a pointed fusion category is called *group-theoretical*.
Nilpotent and solvable fusion categories {#nilp-solv}
----------------------------------------
Let $G$ be a finite group. A $G$-grading on a fusion category ${{\mathcal C}}$ is a decomposition ${{\mathcal C}}=
\oplus_{g\in G} {{\mathcal C}}_g$, such that ${{\mathcal C}}_g \otimes {{\mathcal C}}_h \subseteq {{\mathcal C}}_{gh}$ and ${{\mathcal C}}_g^* \subseteq {{\mathcal C}}_{g^{-1}}$, for all $g, h \in G$. A $G$-grading is *faithful* if ${{\mathcal C}}_g \neq 0$, for all $g \in G$. The fusion category ${{\mathcal C}}$ is called a *$G$-extension* of a fusion category ${{\mathcal D}}$ if there is a faithful grading ${{\mathcal C}}= \oplus_{g\in G} {{\mathcal C}}_g$ with neutral component ${{\mathcal C}}_e \cong {{\mathcal D}}$.
If ${{\mathcal C}}$ is any fusion category, there exist a finite group $U({{\mathcal C}})$, called the *universal grading group* of ${{\mathcal C}}$, and a canonical faithful grading ${{\mathcal C}}=
\oplus_{g \in U({{\mathcal C}})}{{\mathcal C}}_g$, with neutral component ${{\mathcal C}}_e = {{\mathcal C}}_{ad}$, where ${{\mathcal C}}_{ad}$ is the *adjoint* subcategory of ${{\mathcal C}}$, that is, the fusion subcategory generated by $X\otimes
X^*$, $X \in {\operatorname{Irr}}({{\mathcal C}})$.
In fact, $K_0({{\mathcal C}})_{ad} = K_0({{\mathcal C}}_{ad})$ is a based subring of $K_0({{\mathcal C}})$ and $K_0({{\mathcal C}})$ decomposes into a direct sum of indecomposable based $K_0({{\mathcal C}})_{ad}$-bimodules $K_0({{\mathcal C}})=\displaystyle{\oplus_{g\in U({{\mathcal C}})}}K_0({{\mathcal C}})_g$, with $K_0({{\mathcal C}})_e=K_0({{\mathcal C}})_{ad}$. Then the group structure on $U({{\mathcal C}}) := U(K_0({{\mathcal C}}))$ is defined by the following property: $gh = t$ if and only if $X_g X_h\in K_0({{\mathcal C}})_t$, for all $X_g\in K_0({{\mathcal C}})_g$, $X_h\in K_0({{\mathcal C}})_h$, $g,h,t\in U({{\mathcal C}})$; see [@gel-nik Theorem 3.5].
A fusion category ${{\mathcal C}}$ is (cyclically) *nilpotent* if there exists a sequence of fusion categories ${\operatorname{Vect}}= {{\mathcal C}}_0 \subseteq {{\mathcal C}}_1 \dots \subseteq {{\mathcal C}}_n
= {{\mathcal C}}$, and finite (cyclic) groups $G_1, \dots, G_n$, such that for all $i = 1, \dots, n$, ${{\mathcal C}}_i$ is a $G_i$-extension of ${{\mathcal C}}_{i-1}$.
On the other side, ${{\mathcal C}}$ is *solvable* if it is Morita equivalent to a cyclically nilpotent fusion category, that is, if there exists a cyclically nilpotent fusion category ${{\mathcal D}}$ and an idecomposable right module category ${\mathcal{M}}$ over ${{\mathcal D}}$ such that ${{\mathcal C}}$ is equivalent to the fusion category ${{\mathcal D}}^*_{\mathcal{M}}$ of ${{\mathcal D}}$-linear endofunctors of ${\mathcal{M}}$.
Consider an action of a finite group $G$ on a fusion category ${{\mathcal C}}$ by tensor autoequivalences $\rho: \underline
G \to \underline {\operatorname{Aut}}_{\otimes} \, {{\mathcal C}}$. The *equivariantization* of ${{\mathcal C}}$ with respect to the action $\rho$, denoted ${{\mathcal C}}^G$, is a fusion category whose objects are pairs $(X, \mu)$, such that $X$ is an object of ${{\mathcal C}}$ and $\mu = (\mu^g)_{g \in G}$, is a collection of isomorphisms $\mu^g:\rho^gX \to X$, $g \in G$, satisfying appropriate compatibility conditions.
The forgetful functor $F: {{\mathcal C}}^G \to {{\mathcal C}}$, $F(X, \mu) = X$, is a dominant tensor functor that gives rise to a central exact sequence of fusion categories ${\operatorname{Rep}}G \to {{\mathcal C}}^G \to {{\mathcal C}}$ [@indp-exact], where ${\operatorname{Rep}}G$ is the category of finite-dimensional representations of $G$.
The category ${{\mathcal C}}^G$ is integral (respectively, weakly integral) if and only if so is ${{\mathcal C}}$. See [@tensor-exact Proposition 4.9], [@indp-exact Proposition 2.12].
According to [@ENO2 Definition 1.2], a fusion category ${{\mathcal C}}$ is solvable if and only if there exists a sequence of fusion categories ${\operatorname{Vect}}= {{\mathcal C}}_0,
\dots, {{\mathcal C}}_n = {{\mathcal C}}$, $n \geq 0$, and a sequence of cyclic groups of prime order $G_1, \dots, G_n$, such that, for all $1\leq i \leq n$, ${{\mathcal C}}_i$ is a $G_i$-equivariantization or a $G_i$-extension of ${{\mathcal C}}_{i-1}$.
It is shown in [@ENO2 Proposition 4.1] that the class of solvable fusion categories is stable under taking extensions and equivariantizations by solvable groups, Morita equivalent categories, tensor products, Drinfeld center, fusion subcategories and components of quotient categories.
In view of [@ENO2 Proposition 4.5 (iv)], every nontrivial solvable fusion category has nontrivial invertible objects.
Suppose that the finite group $G$ acts on the fusion category ${{\mathcal C}}$ by tensor autoequivalences. Let $Y \in {\operatorname{Irr}}{{\mathcal C}}$. The stabilizer of $Y$ is the subgroup $G_Y = \{g\in G:\, \rho^g(Y) \cong Y \}$. Let $\alpha_Y : G_Y \times G_Y \to k^*$ be the 2-cocycle defined by the relation $$\label{alfa} \alpha_Y(g, h)^{-1} {\operatorname{id}}_Y = c^g \rho^g(c^h)({\rho^{g, h}_{2_Y}})^{-1}(c^{gh})^{-1}: Y \to Y,$$ where, for all $g \in G_Y$, $c^g: \rho^g(Y) \to Y$ is a fixed isomorphism [@fusionrules-equiv Subsection 2.3].
Then the simple objects of ${{\mathcal C}}^G$ are parameterized by pairs $(Y, U)$, where $Y$ runs over the $G$-orbits on ${\operatorname{Irr}}({{\mathcal C}})$ and $U$ is an equivalence class of an irreducible $\alpha_Y$-projective representation of $G_Y$. We shall use the notation $S_{Y, U}$ to indicate the isomorphism class of the simple object corresponding to the pair $(Y, U)$. The dimension of $S_{Y, U}$ is given by the formula $$\label{dim-equiv}{\operatorname{FPdim}}S_{Y, U} = [G:G_Y] \dim U {\operatorname{FPdim}}Y.$$
\[simple-p\] Let $p$ be a prime number. Suppose that the group ${{\mathbb Z}}_p$ acts on a fusion category ${{\mathcal C}}$ by tensor autoequivalences. Assume in addition that $G({{\mathcal C}}^{{{\mathbb Z}}_p})$ is of order $p$ and $G({{\mathcal C}}) \neq \{{\textbf{1}}\}$. Then ${{\mathcal C}}^{{{\mathbb Z}}_p}$ has a simple object of Frobenius-Perron dimension $p$.
Let $Y$ be an invertible object of ${{\mathcal C}}$ and let $U$ be an irreducible $\alpha_Y$-projective representation of the subgroup $G_Y \subseteq {{\mathbb Z}}_p$. Since $G_Y$ is cyclic, then $\alpha_Y = 1$ in $H^2(G_Y, k^*)$ and $\dim U = 1$. Then the Frobenius-Perron dimension of the simple object $S_{Y, U}$ is given by ${\operatorname{FPdim}}S_{Y, U} = [G:G_Y] {\operatorname{FPdim}}Y = [G:G_Y]$. Moreover, if $Y = {\textbf{1}}$, then $G_Y = {{\mathbb Z}}_p$. Therefore, letting $U_0 = \epsilon, U_1, \dots, U_{p-1}$ the non-isomorphic representations of ${{\mathbb Z}}_p$, we get that ${\textbf{1}}= S_{{\textbf{1}}, U_0}, S_{{\textbf{1}}, U_1}, \dots, S_{{\textbf{1}}, U_{p-1}}$ are all the non-isomorphic invertible objects of ${{\mathcal C}}^{{{\mathbb Z}}_p}$. Hence for all invertible object $Y \neq {\textbf{1}}$ of ${{\mathcal C}}$, we must have $[G: G_Y] = p$ and the simple object $S_{Y, U}$ has Frobenius-Perron dimension $p$. This proves the lemma.
\[cyc-nilp\] Let $p$ be a prime number. Suppose that ${{\mathcal C}}$ is a solvable fusion category such that $G({{\mathcal C}}) \cong {{\mathbb Z}}_p$ and ${{\mathcal C}}$ has no simple objects of Frobenius-Perron dimension $p$. Then ${{\mathcal C}}$ is cyclically nilpotent.
The proof is by induction on ${\operatorname{FPdim}}{{\mathcal C}}\geq p$. If ${\operatorname{FPdim}}{{\mathcal C}}= p$ there is nothing to prove. Suppose ${\operatorname{FPdim}}{{\mathcal C}}> p$. Since ${{\mathcal C}}$ is solvable then, for some prime number $q$, ${{\mathcal C}}$ must be a ${{\mathbb Z}}_q$-extension or a ${{\mathbb Z}}_q$-equivariantization of a fusion category ${{\mathcal D}}$. If the second possibility holds, then the assumption that $G({{\mathcal C}}) \cong {{\mathbb Z}}_p$ implies that $q = p$. Moreover, since ${{\mathcal D}}$ is also solvable, then ${{\mathcal D}}_{pt} \neq {\operatorname{Vect}}$. By Lemma \[simple-p\], ${{\mathcal C}}$ must have a simple object of dimension $p$, which contradicts the assumption.
Therefore ${{\mathcal C}}$ must be a ${{\mathbb Z}}_q$-extension of a fusion subcategory ${{\mathcal D}}$. In particular, ${{\mathcal D}}$ cannot have simple objects of dimension $p$ and since ${{\mathcal D}}$ is solvable, then ${{\mathcal D}}_{pt} \neq {\operatorname{Vect}}$, whence $G({{\mathcal D}}) = G({{\mathcal C}}) \cong {{\mathbb Z}}_p$. By induction, ${{\mathcal D}}$ and then also ${{\mathcal C}}$, is cyclically nilpotent. This finishes the proof of the proposition.
\[g-central\] Let ${{\mathcal C}}$ be a fusion category and let $G$ be a finite group acting on ${{\mathcal C}}$ by tensor autoequivalences. Then the forgetful functor $U: {{\mathcal Z}}({{\mathcal C}}^G) \to {{\mathcal C}}^G$ induces an injective ring homomorphism $K_0(G) \to Z(K_0({{\mathcal C}}^G))$. In particular, the group $\widehat G$ is isomorphic to a subgroup of the center of $G({{\mathcal C}}^G)$.
By [@ENO2 Proposition 2.10], the Drinfeld center ${{\mathcal Z}}({{\mathcal C}})$ contains a Tannakian subcategory ${{\mathcal E}}\cong {\operatorname{Rep}}G$ such that ${{\mathcal E}}$ embeds into ${{\mathcal C}}$ under the forgetful functor $U: {{\mathcal Z}}({{\mathcal C}}) \to {{\mathcal C}}$. As a consequence we obtain the lemma.
Braided fusion categories
-------------------------
A braided fusion category is a fusion category ${{\mathcal C}}$ endowed with a braiding, that is, a natural isomorphism $c_{X,Y} : X \otimes Y \rightarrow Y \otimes X$, $X, Y \in {{\mathcal C}}$, subject to the so-called hexagon axioms.
If ${{\mathcal D}}$ is a fusion subcategory of a braided fusion category ${{\mathcal C}}$, the *M" uger centralizer* of ${{\mathcal D}}$ in ${{\mathcal C}}$ will be denoted by ${{\mathcal D}}'$. Thus ${{\mathcal D}}'$ is the full fusion subcategory generated by all objects $X \in {{\mathcal C}}$ such that $c_{Y, X}c_{X, Y} =
{\operatorname{id}}_{X \otimes Y}$, for all objects $Y \in {{\mathcal D}}$.
The centralizer ${{\mathcal C}}'$ of ${{\mathcal C}}$ is called the *M" uger (or symmetric) center* of ${{\mathcal C}}$. The category ${{\mathcal C}}$ is called *symmetric* if ${{\mathcal C}}' = {{\mathcal C}}$. If ${{\mathcal C}}$ is any braided fusion category, its M" uger center ${{\mathcal C}}'$ is a symmetric fusion subcategory of ${{\mathcal C}}$. The category ${{\mathcal C}}$ is called *non-degenerate* (respectively, *slightly degenerate*) if ${{\mathcal C}}' \cong {\operatorname{Vect}}$ (respectively, if ${{\mathcal C}}' \cong {\operatorname{sVect}}$, where ${\operatorname{sVect}}$ denotes the category of super-vector spaces).
For a fusion category ${{\mathcal C}}$, the Drinfeld center of ${{\mathcal C}}$ will be denoted ${{\mathcal Z}}({{\mathcal C}})$. It is known that ${{\mathcal Z}}({{\mathcal C}})$ is a braided non-degenerate fusion category of Frobenius-Perron dimension ${\operatorname{FPdim}}{{\mathcal Z}}({{\mathcal C}}) = ({\operatorname{FPdim}}{{\mathcal C}})^2$.
Let $G$ be a finite group. The fusion category ${\operatorname{Rep}}G$ of finite dimensional representations of $G$ is a symmetric fusion category with respect to the canonical braiding. A braided fusion category ${{\mathcal E}}$ is called Tannakian, if ${{\mathcal E}}\cong {\operatorname{Rep}}G$ for some finite group $G$ as braided fusion categories.
Every symmetric fusion category is equivalent, as a braided fusion category, to the category ${\operatorname{Rep}}(G, u)$ of representations of a finite group $G$ on finite-dimensional super-vector spaces, where $u \in G$ is a central element of order 2 which acts as the parity operator [@deligne]. In particular, if ${{\mathcal C}}$ is symmetric, then it is equivalent to the category of representations of a finite group as a fusion category.
Let $G$ be a finite group. A *$G$-crossed braided fusion category* is a fusion category ${{\mathcal D}}$ endowed with a $G$-grading ${{\mathcal D}}= \oplus_{g \in G}{{\mathcal D}}_g$ and an action of $G$ by tensor autoequivalences $\rho:\underline G \to \underline {\operatorname{Aut}}_{\otimes} \, {{\mathcal D}}$, such that $\rho^g({{\mathcal D}}_h)
\subseteq
{{\mathcal D}}_{ghg^{-1}}$, for all $g, h \in G$, and a $G$-braiding $c: X \otimes Y \to
\rho^g(Y) \otimes X$, $g \in G$, $X \in {{\mathcal D}}_g$, $Y \in {{\mathcal D}}$, subject to compatibility conditions. The $G$-braiding $c$ restricts to a braiding in the neutral component ${{\mathcal D}}_e$.
If ${{\mathcal D}}$ is a $G$-crossed braided fusion category, then the equivariantization ${{\mathcal D}}^G$ under the action of $G$ is a braided fusion category containing ${\operatorname{Rep}}G$ as a Tannakian subcategory. Furthermore, the group $G$ acts by restriction on ${{\mathcal D}}_e$ by braided tensor autoequivalences. The equivariantization ${{\mathcal D}}_e^G$ coincides with the centralizer ${{\mathcal E}}'$ of the Tannakian subcategory ${{\mathcal E}}$ in ${{\mathcal D}}^G$. See [@mueger-crossed].
Let ${{\mathcal E}}$ be Tannakian subcategory of a braided fusion category ${{\mathcal C}}$ and let $G$ be a finite group such that ${{\mathcal E}}\cong {\operatorname{Rep}}G$ as symmetric categories. Let also $A \in {{\mathcal C}}$ be the algebra corresponding to the algebra $k^G \in {\operatorname{Rep}}G$ of functions on $G$ with the regular action. The de-equivariantization ${{\mathcal C}}_G$ of ${{\mathcal C}}$ with respect to ${\operatorname{Rep}}G$ is the fusion category ${{\mathcal C}}_A$ of right $A$-modules in ${{\mathcal C}}$. This is a $G$-crossed braided fusion category such that ${{\mathcal C}}\cong ({{\mathcal C}}_G)^G$. The neutral component of ${{\mathcal C}}_G$ with respect to the associated $G$-grading, denoted by ${{\mathcal C}}_G^0$, coincides with the de-equivariantization ${{{\mathcal E}}'}_G$ of the centralizer of ${{\mathcal E}}$ by the group $G$.
It was shown in [@witt-wgt Proposition ] that if ${{\mathcal E}}\cong {\operatorname{Rep}}G
\subseteq {{\mathcal C}}$ is a Tannakian subcategory, then ${{\mathcal C}}$ is weakly integral (respectively, integral or weakly group-theoretical) if and only if ${{\mathcal C}}^0_G$ is weakly integral (respectively, integral, weakly group-theoretical). In addition, ${{\mathcal C}}$ is solvable if and only if ${{\mathcal C}}^0_G$ is solvable and $G$ is solvable.
Let ${{\mathcal C}}$ be a braided fusion category. Then the subcategory ${{\mathcal C}}_{ad} \cap {{\mathcal C}}_{pt}$ is symmetric.
Suppose first that ${{\mathcal C}}$ is non-degenerate. Then ${{\mathcal C}}_{ad} =
{{\mathcal C}}_{pt}'$, by [@DGNOI Corollary 3.27]. Therefore ${{\mathcal C}}_{ad} \cap {{\mathcal C}}_{pt} =
{{\mathcal C}}_{pt}' \cap {{\mathcal C}}_{pt}$ is a symmetric subcategory.
Next, for an arbitrary braided fusion category ${{\mathcal C}}$, let ${{\mathcal Z}}({{\mathcal C}})$ be the Drinfeld center of ${{\mathcal C}}$. Since ${{\mathcal Z}}({{\mathcal C}})$ is non-degenerate, then the category ${{\mathcal Z}}({{\mathcal C}})_{ad} \cap {{\mathcal Z}}({{\mathcal C}})_{pt}$ is symmetric. The braiding of ${{\mathcal C}}$ induces a canonical embedding of braided fusion categories ${{\mathcal C}}\to {{\mathcal Z}}({{\mathcal C}})$. We may therefore identify ${{\mathcal C}}$ with a fusion subcategory of ${{\mathcal Z}}({{\mathcal C}})$. Observe that ${{\mathcal C}}_{ad} \subseteq {{\mathcal Z}}({{\mathcal C}})_{ad}$ and ${{\mathcal C}}_{pt} \subseteq {{\mathcal Z}}({{\mathcal C}})_{pt}$. Hence ${{\mathcal C}}_{ad} \cap {{\mathcal C}}_{pt} \subseteq {{\mathcal Z}}({{\mathcal C}})_{ad} \cap {{\mathcal Z}}({{\mathcal C}})_{pt}$, and then ${{\mathcal C}}_{ad} \cap {{\mathcal C}}_{pt}$ is symmetric, as claimed.
\[cent-cpt\] Let ${{\mathcal C}}$ be a braided fusion category such that ${{\mathcal C}}_{ad} = {{\mathcal C}}$. Then ${{\mathcal C}}_{pt}' = {{\mathcal C}}$.
Let ${{\mathcal B}}\subseteq {{\mathcal C}}$ be any fusion subcategory. By [@DGNOI Proposition 3.25] we have $({{\mathcal B}}_{ad})' = ({{\mathcal B}}')^{co} = {{\mathcal A}}$, where ${{\mathcal A}}\subseteq {{\mathcal C}}$ denotes the projective centralizer of ${{\mathcal B}}$. Letting ${{\mathcal B}}= {{\mathcal C}}_{pt}$ we find that ${{\mathcal C}}= ({{\mathcal B}}_{ad})'$ equals the projective centralizer of ${{\mathcal C}}_{pt}$. By [@DGNOI Lemma 3.15], the projective centralizer of a fusion subcategory ${{\mathcal B}}$ is a graded extension of the centralizer ${{\mathcal B}}'$. Since ${{\mathcal C}}= {{\mathcal C}}_{ad}$, this implies that ${{\mathcal C}}= {{\mathcal C}}_{pt}'$, as claimed.
Grothendieck equivalence of fusion categories {#g-crossed}
=============================================
Let ${{\mathcal C}}$ and $\tilde{{\mathcal C}}$ be fusion categories. A *Grothendieck equivalence* between ${{\mathcal C}}$ and $\tilde{{\mathcal C}}$ is a bijection $f:{\operatorname{Irr}}{{\mathcal C}}\rightarrow {\operatorname{Irr}}\tilde{{\mathcal C}}$ such that $$f({\textbf{1}})={\textbf{1}}, \quad \text{and } N_{f(X),f(Y)}^{f(Z)}=N_{X,Y}^Z,$$ for all $X, Y, Z \in {\operatorname{Irr}}{{\mathcal C}}$.
We shall say that ${{\mathcal C}}$ and $\tilde{{\mathcal C}}$ are *Grothendieck equivalent* if there exist a Grothendieck equivalence between them.
\[biy\] Suppose $f: {\operatorname{Irr}}{{\mathcal C}}\rightarrow {\operatorname{Irr}}\tilde{{\mathcal C}}$ is a Grothendieck equivalence. Then the map $f$ extends canonically to a ring isomorphism $f: K_0({{\mathcal C}})\rightarrow K_0(\tilde {{\mathcal C}})$.
In particular, $f$ induces a bijection between the lattices of fusion subcategories of ${{\mathcal C}}$ and $\tilde {{\mathcal C}}$. If ${{\mathcal D}}$ is a fusion subcategory of ${{\mathcal C}}$, we shall denote by $f({{\mathcal D}})$ the corresponding fusion subcategory of $\tilde{{\mathcal C}}$, that is, $f({{\mathcal D}})$ is the fusion subcategory whose simple objects are $f(X)$, $X \in {\operatorname{Irr}}{{\mathcal D}}$. Note that $f$ restricts to a Grothendieck equivalence $f: {\operatorname{Irr}}{{\mathcal D}}\to {\operatorname{Irr}}f({{\mathcal D}})$.
\[groth-eq\] Let ${{\mathcal C}}$ and $\tilde{{\mathcal C}}$ be fusion categories and suppose that $\mathit{f:{\operatorname{Irr}}{{\mathcal C}}\rightarrow {\operatorname{Irr}}\tilde{{\mathcal C}}}$ is a *Grothendieck equivalence*. Then the following hold:
\(i) If $X\in K_0({{\mathcal C}})$, then *FPdim*$(f(X))$ = *FPdim*$(X)$. Hence, if ${{\mathcal D}}$ is a fusion subcategory of ${{\mathcal C}}$, then *FPdim*$(f({{\mathcal D}}))$= *FPdim*$({{\mathcal D}})$.
\(ii) $X\in {\operatorname{Irr}}{{\mathcal C}}$ is invertible if and only if $f(X)\in {\operatorname{Irr}}\tilde {{\mathcal C}}$ is invertible.
\(iii) If $X \in {\operatorname{Irr}}{{\mathcal C}}$, then $f(X^*) = f(X)^*$.
\(iv) $f({{\mathcal C}}^{(n)})=\tilde {{\mathcal C}}^{(n)}$, for all $n \geq 0$. In particular, ${{\mathcal C}}$ is nilpotent if and only if $\tilde {{\mathcal C}}$ is nilpotent.
\(v) $f$ induces a group isomorphism $f: U({{\mathcal C}}) \to U(\tilde {{\mathcal C}})$ such that $f({{\mathcal C}}_g) = \tilde {{\mathcal C}}_{f(g)}$.
\(i) By Remark \[biy\] we know that $f$ extends to a ring isomorphism $ f:K_0({{\mathcal C}})\rightarrow K_0(\tilde {{\mathcal C}})$. By [@ENO Lemma 8.3] $\text{FPdim}:K_0({{\mathcal C}})\rightarrow \mathbb{R}$ is the only ring homomorphism such that ${\operatorname{FPdim}}(X)>0$ for any $0\neq X\in{{\mathcal C}}$, so ${\operatorname{FPdim}}(f(X)) = {\operatorname{FPdim}}(X)$, for all $X \in {\operatorname{Irr}}{{\mathcal C}}$.
\(ii) This follows from (i), since the invertible objects of a fusion category are exactly those objects with Frobenius-Perron dimension 1.
\(iii) Since $f({\textbf{1}}) = {\textbf{1}}$, then $N_{f(X), f(X^*)}^{\textbf{1}}= N_{X, X^*}^{\textbf{1}}= 1$. Therefore $f(X^*) = f(X)^*$.
\(iv) It follows from (iii) and the fact that $f$ preserves fusion rules that $f({{\mathcal C}}_{ad})=\tilde {{\mathcal C}}_{ad}$. Then $f$ induces by restriction a Grothendieck equivalence ${\operatorname{Irr}}{{\mathcal C}}_{ad} \to {\operatorname{Irr}}\tilde {{\mathcal C}}_{ad}$. An inductive argument implies that $f({{\mathcal C}}^{(n)})=\tilde {{\mathcal C}}^{(n)}$, for all $n \geq 0$.
\(v) By definition $U({{\mathcal C}})=U(K_0({{\mathcal C}}))$ [@gel-nik] and $K_0({{\mathcal C}})$ decomposes into a direct sum of indecomposable based $K_0({{\mathcal C}})_{ad}$-bimodules $K_0({{\mathcal C}})=\displaystyle{\oplus_{g\in U({{\mathcal C}})}}K_0({{\mathcal C}})_g$, with $K_0({{\mathcal C}})_e=K_0({{\mathcal C}})_{ad}$. This decomposition is unique up to a permutation of $U({{\mathcal C}})$. By Remark \[biy\], $f$ extends to a ring isomorphism $f:K_0({{\mathcal C}})\rightarrow K_0(\tilde {{\mathcal C}})$ and by (iv) $f$ restricts to a ring isomorphism $K_0({{\mathcal C}})_{ad} = K_0({{\mathcal C}}_{ad}) \cong K_0(\tilde {{\mathcal C}}_{ad})= K_0(\tilde {{\mathcal C}})_{ad}$. So for all $g\in U({{\mathcal C}})$, $f(K_0({{\mathcal C}})_g) = K_0(\tilde{{\mathcal C}})_{\tilde g}$, for a unique $\tilde g\in U(\tilde{{\mathcal C}})$. Letting $f(g) = \tilde g$, we obtain a group isomorphism $f: U({{\mathcal C}}) \to U(\tilde {{\mathcal C}})$ such that $f(K_0({{\mathcal C}}_g)) = K_0(\tilde {{\mathcal C}}_{f(g)})$. This implies (v).
\[grading\] Let $G$ be a finite group. Observe that any $G$-grading on a fusion category ${{\mathcal C}}$ with neutral component ${{\mathcal D}}$ is uniquely determined by a $G$-grading on the Grothendieck ring $K_0({{\mathcal C}})$ with neutral component $K_0({{\mathcal D}})$. In particular, if ${{\mathcal C}}$ and $\tilde{{\mathcal C}}$ are Grothendieck equivalent, then ${{\mathcal C}}$ is $G$-graded with neutral component ${{\mathcal D}}$ if and only if $\tilde {{\mathcal C}}$ is $G$-graded with neutral component $\tilde {{\mathcal D}}$, such that $\tilde
{{\mathcal D}}$ and ${{\mathcal D}}$ are Grothendieck equivalent.
Our first theorem concerns fusion categories with dihedral fusion rules.
\[dihedral-fr\] Let $n$ be a natural number and let ${{\mathcal C}}$ be a fusion category. Suppose that ${{\mathcal C}}$ is Grothendieck equivalent to the category ${\operatorname{Rep}}D_n$, where $D_n$ is the dihedral group of order $2n$. Then ${{\mathcal C}}$ is solvable.
It follows from [@naidu-rowell Theorem 4.2] that a fusion category Grothendieck equivalent to the representation category of a dihedral group is group-theoretical. Then ${{\mathcal C}}$ is group-theoretical, that is, it is Morita equivalent to a pointed fusion category ${{\mathcal C}}(\Gamma, \omega)$, where $\Gamma$ is a group and $\omega$ is a 3-cocycle on $\Gamma$.
Suppose first that $n$ is odd. Then the order of $\Gamma$ is equal to $2n$ and, since $n$ is odd, $\Gamma$ is solvable. Then ${{\mathcal C}}$ is solvable too.
If $n$ is even, then the center of $D_n$ is of order $2$ and $D_n/Z(D_n) \cong D_{n/2}$. Therefore, the category ${\operatorname{Rep}}D_n$ is a ${{\mathbb Z}}_2$-extension of ${\operatorname{Rep}}D_{n/2}$; see [@gel-nik Example 3.2]. Since ${{\mathcal C}}$ is Grothendieck equivalent to ${\operatorname{Rep}}D_n$, then it is a ${{\mathbb Z}}_2$-extension of a fusion subcategory ${{\mathcal D}}_1$, where ${{\mathcal D}}_1$ is Grothendieck equivalent to ${\operatorname{Rep}}D_{n/2}$. Continuing this process, we find that the category ${{\mathcal C}}$ is obtained by a sequence of ${{\mathbb Z}}_2$-extensions from a fusion subcategory ${{\mathcal D}}$ such that ${{\mathcal D}}$ is Grothendieck equivalent to ${\operatorname{Rep}}D_{m}$, with $m$ an odd natural number. By the above, ${{\mathcal D}}$ is solvable and therefore so is ${{\mathcal C}}$. This finishes the proof of the theorem.
The following consequence of Proposition \[cyc-nilp\] gives some restrictions that guarantee that the solvability of a fusion category is a Grothendieck invariant.
\[cor-cyc\] Let $p$ be a prime number. Suppose that ${{\mathcal C}}$ is a solvable fusion category such that $G({{\mathcal C}}) \cong {{\mathbb Z}}_p$ and ${{\mathcal C}}$ has no simple objects of Frobenius-Perron dimension $p$. If ${{\mathcal C}}$ is Grothendieck equivalent to a fusion category $\tilde {{\mathcal C}}$, then $\tilde {{\mathcal C}}$ is solvable.
By Proposition \[cyc-nilp\], ${{\mathcal C}}$ is cyclically nilpotent. Therefore $\tilde {{\mathcal C}}$ is cyclically nilpotent, whence solvable.
\[rmk-ansn-2\] For all $n \geq 2$, the alternating group ${\mathbb A}_{n}$ has no irredubible representation of degree $2$[^2]. In addition, if $n \geq 5$ (${\operatorname{Rep}}{\mathbb{S}}_{4}$ is of type $(1, 2; 2, 1; 3, 2)$), the symmetric group ${\mathbb{S}}_{n}$ has no irredubible representation of degree $2$ neither. In fact, if $V$ were such a representation, then the restriction $V\vert_{{\mathbb A}_{n}}$ would not be irreducible. Hence, since ${\mathbb A}_{n}$ has no nontrivial one-dimensional representations (because $n \geq 5$), then $V\vert_{{\mathbb A}_{n}}$ would be trivial. This is impossible, because the kernel of the restriction functor ${\operatorname{Rep}}{\mathbb{S}}_{n} \to {\operatorname{Rep}}{\mathbb A}_{n}$ is the pointed subcategory ${\operatorname{Rep}}{{\mathbb Z}}_2 \subseteq {\operatorname{Rep}}{\mathbb{S}}_{n}$.
\[fr-sn\] Let $n\geq 5$ be a natural number and let ${{\mathcal C}}$ be a fusion category. Suppose that ${{\mathcal C}}$ is Grothendieck equivalent to ${\operatorname{Rep}}{\mathbb{S}}_n$. Then ${{\mathcal C}}$ is not solvable.
The category ${\operatorname{Rep}}{\mathbb{S}}_n$ is not solvable. On the other hand, the group ${\mathbb{S}}_n$ has two non-equivalent representations of degree one and no irreducible representation of degree two, in view of Remark \[rmk-ansn-2\]. Hence $G({{\mathcal C}}) \cong {{\mathbb Z}}_2$ and ${{\mathcal C}}$ has no simple objects of Frobenius-Perron dimension $2$. The result is thus obtained as a consequence of Proposition \[cor-cyc\].
\[fr-simple\] Let $G$ be a non-abelian finite simple group. If ${{\mathcal C}}$ is a fusion category Grothendieck equivalent to ${\operatorname{Rep}}G$, then ${{\mathcal C}}_{pt} = {\operatorname{Vect}}$ and therefore ${{\mathcal C}}$ is not solvable.
Examples of non-solvable fusion rules {#nsol-fr}
=====================================
Abelian extensions
------------------
Consider an abelian exact sequence of Hopf algebras $$\label{abel-es}
k \longrightarrow k^\Gamma \overset{i}\longrightarrow H
\overset{\pi}\longrightarrow kF
\longrightarrow k,$$ where $\Gamma$ and $F$ are finite groups. Then gives rise to actions by permutations $\Gamma \overset{\lhd}\longleftarrow
\Gamma \times F \overset{\rhd}\longrightarrow F$ such that $(\Gamma, F)$ is a matched pair of groups. Moreover, $H \cong k^\Gamma {}^\tau\#_\sigma kF$ is a bicrossed product with respect to normalized invertible $2$-cocycles $\sigma: F \times F \to k^\Gamma$, $\tau: \Gamma \times \Gamma \to
k^F$, satisfying suitable compatibility conditions. See [@ma-ext].
The multiplication and comultiplication of $k^\Gamma {}^\tau\#_\sigma kF$ are determined in the basis $\{e_{s}\#x /s\in\Gamma, x\in F\}$, by the formulas $$\begin{aligned}
(e_s \# x)\otimes(e_t \# y) & = \delta_{t,s\lhd x} \, \sigma_s(x, y) \, e_s \# xy, \\
\Delta(e_s \# x) & = \sum_{gh=s} \tau_x(g, h) \, e_g \# (h \rhd x) \otimes e_h \#x,\end{aligned}$$ for all $s, t \in \Gamma$, $x, y \in F$, where $\sigma_s(x,y)=\sigma(x,y)(s)$ and $\tau_x(s, t)=\tau(s,t)(x)$. See [@ma-ext]. The exact sequence is called *split* if $\sigma$ and $\tau$ are the trivial 2-cocycles.
For all $s\in\Gamma$, the restriction of the map $\sigma_s:F \times F \to k^{\times}$ to the stabilizer subgroup $F_s = F \cap sFs^{-1}$ is a 2-cocycle on $F_s$.
The irreducible representations of $H \cong k^\Gamma {}^\tau\#_\sigma kF$ are classified for pairs $(s,U_s)$, where $s$ is a representative of the orbits of the action of $F$ in $\Gamma$ and $U_s$ is an irreducible representation of the twisted group algebra $k_{\sigma_s}F_s$, that is, a projective irreducible representation $F_s$ with cocycle $\sigma_s$. Given a pair $(s,U_s)$, the corresponding irreducible representation is given by $$\label{irrepns-bcpt}W_{(s,U_s)} = {\operatorname{Ind}}_{k^\Gamma \otimes kF_s}^H s \otimes U_s.$$ Observe that $\dim W_{(s,U_s)} = [F: F_s] \dim U_s$. See [@MW].
\[2-morita\] Recall that every matched pair $(\Gamma, F)$ gives rise to a group structure, denoted $F \bowtie \Gamma$, on the product $F \times \Gamma$ in the form $$(x, s) (y, t) = (x(s \rhd y), (s\lhd y) t),$$ $x, y \in F$, $s, t \in \Gamma$, where $\Gamma \overset{\lhd}\longleftarrow
\Gamma \times F \overset{\rhd}\longrightarrow F$ are the associated compatible actions.
The group $F \bowtie \Gamma$ has a canonical exact factorization into its subgroups $F = F \times \{e\}$ and $\Gamma = \{e\} \times \Gamma$; that is, $F \bowtie \Gamma = F \Gamma$ and $F\cap \Gamma = \{e\}$.
Conversely, every finite group $G$ endowed with an exact factorization $G = F \Gamma$ into its subgroups $F$ and $\Gamma$ gives rise to canonical actions by permutations $\Gamma \overset{\lhd}\longleftarrow
\Gamma \times F \overset{\rhd}\longrightarrow F$ making $(\Gamma, F)$ into a matched pair of groups.
Suppose $H \cong k^\Gamma {}^\tau\#_\sigma kF$ is an abelian extension of $k^\Gamma$ by $kF$. It follows from [@gp-ttic Theorem 1.3] that the category ${\operatorname{Rep}}H$ is Morita equivalent to the pointed fusion category ${{\mathcal C}}(F\bowtie\Gamma, \omega)$, where $\omega$ is a $3$-cocycle on $F\bowtie\Gamma$ arising from the pair $(\sigma, \tau)$ in an exact sequence due to G. I. Kac. In particular, there are equivalences of braided fusion categories $${{\mathcal Z}}({\operatorname{Rep}}H) \cong {\operatorname{Rep}}D(H) \cong {\operatorname{Rep}}D^\omega(F \bowtie \Gamma),$$ where $D^\omega(F \bowtie \Gamma)$ is the twisted Drinfeld double of $F\bowtie \Gamma$ [@dpr]. Note that ${\operatorname{Rep}}H$ is solvable if and only if the group $F \bowtie \Gamma$ is solvable.
\[ddoble\] Let $G$ be a finite group. Then the Drinfeld double $D(G)$ fits into a split cocentral abelian exact sequence $$k \longrightarrow k^G \longrightarrow D(G) \longrightarrow kG
\longrightarrow k.$$ This exact sequence is associated to the adjoint action $\lhd: G \times G \to G$, $h\lhd g = g^{-1}hg$, and to the trivial action $\rhd:
G \times G \to G$.
The following lemma describes the group of invertible objects of the category ${\operatorname{Rep}}H$, when $H$ is an abelian extension.
\[inv-abel\] Suppose $H$ fits into an exact sequence . Then there is an exact sequence $$1\longrightarrow \widehat F \overset{\pi^*}\longrightarrow G({\operatorname{Rep}}H)
\overset{i^*}\longrightarrow \Gamma_0 \longrightarrow 1,$$ where $\widehat F$ denotes the group of one-dimensional characters of $F$ and $\Gamma_0 = \{s\in
\Gamma^F:\, [\sigma_s] = 1 \textrm{ in } H^2(\Gamma, (k^F)^\times) \}$.
The group $G({\operatorname{Rep}}H)$ can be identified with the group $G(H^*)$ of group-like elements in the dual Hopf algebra $H^*$. In addition, $H^*$ fits into an abelian extension $$\label{abel-dual}
k \longrightarrow k^F \overset{i^*}\longrightarrow H^*
\overset{\pi^*}\longrightarrow k\Gamma
\longrightarrow k.$$ The lemma follows from [@ext-ty Lemma 2.2].
\[inv-split\] Keep the notation in Lemma \[inv-abel\]. Note that the dual exact sequence is associated to the actions $F \overset{\lhd'}\longleftarrow F \times \Gamma \overset{\rhd'}\longrightarrow
\Gamma$ defined in the form $x \lhd' s = (s^{-1} \rhd x^{-1})^{-1}$ and $x \rhd' s = (s^{-1} \lhd
x^{-1})^{-1}$, for all $x\in F$, $s\in \Gamma$ [@ma-ext Exercise 5.5].
Hence the exact sequence of groups of Lemma \[inv-abel\] induces the transpose of the action $\lhd'$ of $\Gamma_0$ on the abelian group $\widehat F$.
Clearly, is split if and only if is split and, if this is the case, the exact sequence of groups in Lemma \[inv-abel\] is split as well.
Therefore, in the case where $H$ is a split abelian extension, the group $G({\operatorname{Rep}}H)$ is isomorphic to the semidirect product $\widehat F \rtimes \Gamma_0$ with respect to the action $\lhd'$ of $\Gamma_0$ on $\widehat F$.
\[inv-doble\] Let $G$ be a finite group. Then the group of invertible objects of ${\operatorname{Rep}}D(G)$ is isomorphic to the direct product $G/[G, G]
\times Z(G)$.
This is a consequence of Lemma \[inv-abel\], in view of Example \[ddoble\] and Remark \[inv-split\]. In fact, $\widehat G \cong G/[G, G]$ and the actions $\rhd ': G \times G \to G$ and $\lhd ': G \times G \to G$ in Remark \[inv-split\] are given in this case by $h \rhd ' g=hgh^{-1}$ and $g \lhd ' h=g$, for all $g, h\in G$. Then $G_0=\{g\in G | h \rhd ' g=g, \forall h\in G \}=Z(G)$. The Corollary follows from the fact that the action $\lhd '$ is the trivial one.
Examples associated to the symmetric group
------------------------------------------
Let $n \geq 2$ be a natural number. The symmetric group ${\mathbb{S}}_n$ has an exact factorization ${\mathbb{S}}_n = \langle z \rangle \Gamma$, where $\Gamma = \{\sigma\in
{\mathbb{S}}_n:\, \sigma(n) = n\} \cong {\mathbb{S}}_{n-1}$ and $z = (1 2 \dots n)$, so that $\langle z \rangle \cong C_n$. This exact factorization induces mutual actions by permutations ${\mathbb{S}}_{n-1} \overset{\lhd}\longleftarrow
{\mathbb{S}}_{n-1} \times C_n \overset{\rhd}\longrightarrow C_n$ that make $({\mathbb{S}}_{n-1}, C_n)$ into a matched pair of groups. The actions $\lhd, \rhd$ are determined by the relations $$\label{rel-mp} \sigma c = (\sigma \rhd c) (\sigma \lhd c),$$ for all $\sigma \in \Gamma$, $c \in \langle z \rangle$.
Suppose $n$ is odd, so that $\langle z \rangle \subseteq {\mathbb A}_n$. Relations imply that the subgroup $\Gamma_+ = \Gamma \cap {\mathbb A}_n \cong
{\mathbb A}_{n-1}$ is stable under the action $\lhd$ of $\langle z \rangle$. Therefore the actions $\lhd, \rhd$ induce by restriction a matched pair $({\mathbb A}_{n-1}, C_n)$.
\[even-notstbl\] Let $\sigma \in \Gamma$. It follows from that $\sigma z = z^r (\sigma \lhd z)$, for some $0 \leq r \leq n-1$. Since $\sigma \lhd z \in \Gamma$ then $(\sigma \lhd z)(n) = n$, implying that $r = b(n)$, where $b = \sigma z$.
Suppose that $n\geq 4$ is even and $\sigma \in \Gamma \cap {\mathbb A}_n$. Since $z$ is an odd permutation and $\sigma \lhd z = z^{-r}\sigma z$, then $\sigma \lhd z$ is even if and only if $r$ is odd. Letting $\sigma = (12\dots (n-1))\in \Gamma \cap {\mathbb A}_n$, we find that $r = b(n) = \sigma z(n) = 2$; so that $\sigma \lhd z$ is an odd permutation. This shows that the subgroup $\Gamma_+ = \Gamma \cap {\mathbb A}_n \cong {\mathbb A}_{n-1}$ is not stable under the action $\lhd$ of $\langle z \rangle$ in this case.
Let us consider the associated Hopf algebras $J_n= k^{{\mathbb{S}}_{n-1}} \# kC_n$ and $K_n= k^{{\mathbb A}_{n-1}} \# kC_n$.
The categories ${\operatorname{Rep}}J_n$, ${\operatorname{Rep}}K_n$ are Morita equivalent to the categories ${\operatorname{Rep}}{\mathbb{S}}_n$ and ${\operatorname{Rep}}{\mathbb A}_n$, respectively; see Remark \[2-morita\]. In particular, ${\operatorname{Rep}}J_n$ and ${\operatorname{Rep}}K_n$ are not solvable, for all $n \geq 5$.
Observe that $J_n^*$ is a split abelian extension of $k^{C_n}$ by $k{\mathbb{S}}_{n-1}$ associated to the actions $\lhd'$ and $\rhd'$ in Remark \[inv-split\].
Suppose $n \geq
5$. It follows from [@qt-bicrossed Theorem 5.2], that the Hopf algebras $J_n$, $K_n$, $J_n^*$, $K_n^*$ admit no quasitriangular structure. In particular, the fusion categories ${\operatorname{Rep}}J_n$, ${\operatorname{Rep}}K_n$, ${\operatorname{Rep}}J_n^*$ and ${\operatorname{Rep}}K_n^*$ admit no braiding.
In addition, there are equivalences of braided fusion categories $${\operatorname{Rep}}D(J_n) \cong {\operatorname{Rep}}D({\mathbb{S}}_n), \quad {\operatorname{Rep}}D(K_n) \cong {\operatorname{Rep}}D({\mathbb A}_n).$$ It follows from Corollary \[inv-doble\] that there are group isomorphisms $G(D({\mathbb{S}}_n)^*) \cong {\mathbb{S}}_n/[{\mathbb{S}}_n, {\mathbb{S}}_n] \times Z({\mathbb{S}}_n) \cong {{\mathbb Z}}_2$, for all $n \geq 3$, and similarly, $G(D({\mathbb A}_n)^*) = 1$, for all $n \geq 5$.
\[tann-sn\] The pointed subcategory ${\operatorname{Rep}}D({\mathbb{S}}_n)_{pt}\cong {\operatorname{Rep}}{{\mathbb Z}}_2$ is a Tannakian subcategory of ${\operatorname{Rep}}D({\mathbb{S}}_n)$.
It follows from the description of the irreducible representations in , that the one-dimensional representations of $D({\mathbb{S}}_n)$ are parameterized by pairs $(s, U_s)$, where $s \in {\mathbb{S}}_n$ is a central element and $U_s$ is a one-dimensional representation of ${\mathbb{S}}_n$. Since $Z({\mathbb{S}}_n) = \{e\}$ and ${\mathbb{S}}_n$ has two non-isomorphic one dimensional representations trivial one and the sign representation ${\operatorname{Sg}}$, then ${\operatorname{Rep}}D({\mathbb{S}}_n)_{pt}\cong {\operatorname{Rep}}{{\mathbb Z}}_2$. Moreover, the unique nontrivial element of ${\operatorname{Rep}}D({\mathbb{S}}_n)_{pt}$ corresponds to the pair $(e, {\operatorname{Sg}})$. We have $$s_{(e,{\operatorname{Sg}}),(e,{\operatorname{Sg}})}= \frac{|{\mathbb{S}}_n|}{|{\mathbb{S}}_n|^2}\sum_{g\in {\mathbb{S}}_n} {\operatorname{Sg}}(e){\operatorname{Sg}}(e) =\frac{|{\mathbb{S}}_n|}{|{\mathbb{S}}_n|}=1,$$ and $$\theta_{(e,{\operatorname{Sg}})}= \frac{{\operatorname{Sg}}(e)}{\deg {\operatorname{Sg}}}= 1,$$ where $(s_{X, Y})_{X, Y \in {\operatorname{Irr}}(D({\mathbb{S}}_n))}$ and $\theta$ denote the $S$-matrix and the ribbon structure of ${\operatorname{Rep}}D({\mathbb{S}}_n)$, respectively. See for instance [@NNW Section 3.1]. This shows that ${\operatorname{Rep}}D({\mathbb{S}}_n)_{pt}$ is a Tannakian subcategory, as claimed.
\[secjnkn\] Let $n$ be an odd natural number. Then there is a central exact sequence of Hopf algebras $$\label{sec-jn}
k \longrightarrow k^{{{\mathbb Z}}_2} \longrightarrow J_n \overset{\pi}\longrightarrow K_n
\longrightarrow k,$$ where the map $\pi: J_n \to K_n$ is induced by the inclusion ${\mathbb A}_{n-1}
\subseteq {\mathbb{S}}_{n-1}$.
The map $\pi$ is defined in the form $\pi = j \otimes {\operatorname{id}}: J_n = k^{{\mathbb{S}}_{n-1}} \# kC_n
\to K_n = k^{{\mathbb A}_{n-1}} \# kC_n$, where $j: k^{{\mathbb{S}}_{n-1}}
\to k^{{\mathbb A}_{n-1}}$ is the canonical Hopf algebra map. Then $\pi$ is a surjective Hopf algebra map.
Since the index of $K_n$ in $J_n$ is $2$, then $J_n^{{\operatorname{co}}\pi} \cong k{{\mathbb Z}}_2 \cong k^{{{\mathbb Z}}_2}$ is a necessarily central Hopf subalgebra; see [@ssld Corollary 1.4.3].
\[RepJn\] *(i)* ${\operatorname{Rep}}J_n$ is a ${{\mathbb Z}}_2$-extension of ${\operatorname{Rep}}K_n$, for all odd natural number $n \geq 1$.
*(ii)* ${\operatorname{Rep}}J_n$ is a ${{\mathbb Z}}_2$-equivariantization of a fusion category ${{\mathcal D}}$, for all even natural number $n\geq 4$.
*(i)* This is an immediate consequence of Lemma \[secjnkn\]. That is, since the sequence is a central exact sequence of Hopf algebras, then ${\operatorname{Rep}}(J_n)$ is an ${{\mathbb Z}}_2$-graded fusion category with trivial component $({\operatorname{Rep}}(J_n))_0 = {\operatorname{Rep}}(K_n)$ [@gel-nik Theorem 3.8]. Therefore ${\operatorname{Rep}}J_n$ is a ${{\mathbb Z}}_2$-extension of ${\operatorname{Rep}}K_n$.
*(ii)* Suppose that $n \geq 4$ is even. We first claim that ${\operatorname{Rep}}J_n$ is not a ${{\mathbb Z}}_2$-extension of any fusion category. To see this, first note that it follows from [@jm Lemma 3.4] that $G(J_n) = G(k^{{\mathbb{S}}_{n-1}}) \cong {{\mathbb Z}}_2$, for all $n \geq 2$. Suppose that $K$ is a central Hopf subalgebra of $J_n$ such that $K \cong k^{{{\mathbb Z}}_2}$, then $K$ must necessarily coincide with $kG(J_n)$. Observe that the action $\lhd: {\mathbb{S}}_{n-1} \times C_n \to {\mathbb{S}}_{n-1}$ gives rise to an action by algebra automorphism $$\rightharpoonup: C_n \times k^{{\mathbb{S}}_{n-1}} \to k^{{\mathbb{S}}_{n-1}} \ \ \text{such that} \ \ x \rightharpoonup e_\sigma = e_{\sigma \lhd x^{-1}},$$ for all $x\in C_n$ and $\sigma \in {\mathbb{S}}_{n-1}$. If $\epsilon\neq\varphi\in G(J_n) = G(k^{{\mathbb{S}}_{n-1}})$, we have $\varphi(\sigma)={\operatorname{Sg}}(\sigma)$, for all $\sigma\in{\mathbb{S}}_{n-1}$ and $\varphi = \sum_{\sigma\in{\mathbb{S}}_{n-1}}{\operatorname{Sg}}(\sigma)e_\sigma$.
Then $G(J_n)\subseteq Z(J_n)$ if and only if $z\rightharpoonup\varphi=\varphi$, if and only if ${\mathbb A}_{n-1}$ is stable under the action $\lhd$ of $\langle z \rangle$. This contradicts the observation in Remark \[even-notstbl\]. Hence ${\operatorname{Rep}}J_n$ is not a ${{\mathbb Z}}_2$-extension of any fusion category, as claimed.
Let ${{\mathcal E}}= {\operatorname{Rep}}D({\mathbb{S}}_n)_{pt}$. By Lemma \[tann-sn\], ${{\mathcal E}}\cong {\operatorname{Rep}}{{\mathbb Z}}_2$ is a Tannakian subcategory of ${\operatorname{Rep}}D({\mathbb{S}}_n)$. Since ${\operatorname{Rep}}J_n$ is not a ${{\mathbb Z}}_2$-extension of any fusion category, it follows from [@ENO2 Propositions 2.9 and 2.10] that ${\operatorname{Rep}}J_n$ must be a ${{\mathbb Z}}_2$-equivariantization of a fusion category ${{\mathcal D}}$. We thus obtain (ii).
\[not-ext\] Let $n \geq 5$ be an odd natural number and let $q$ be a prime number. Then the following hold:
\(i) The category ${\operatorname{Rep}}J_n$ is not a ${{\mathbb Z}}_q$-equivariantization of any fusion category.
\(ii) Suppose that ${\operatorname{Rep}}J_n$ is a ${{\mathbb Z}}_q$-extension of a fusion category $\tilde
{{\mathcal D}}$. Then $q = 2$ and $\tilde {{\mathcal D}}= {\operatorname{Rep}}K_n$.
\(iii) The category ${\operatorname{Rep}}K_n$ is neither a ${{\mathbb Z}}_q$-equivariantization nor a ${{\mathbb Z}}_q$-extension of any fusion category.
Let ${{\mathcal C}}$ be one of the categories ${\operatorname{Rep}}J_n$ or ${\operatorname{Rep}}K_n$. If ${{\mathcal C}}$ is a ${{\mathbb Z}}_q$-extension of a fusion category $\tilde {{\mathcal D}}$, then the Drinfeld center ${{\mathcal Z}}({{\mathcal C}})$ must contain a Tannakian subcategory equivalent to ${\operatorname{Rep}}{{\mathbb Z}}_q$ which maps to the trivial fusion subcategory ${\operatorname{Vect}}\subseteq {{\mathcal C}}$ under the forgetful functor $U:{{\mathcal Z}}({{\mathcal C}}) \to {{\mathcal C}}$. In this case, the category $\tilde {{\mathcal D}}$ is canonically determined by the corresponding Tannakian subcategory. Dually, if ${{\mathcal C}}$ is a ${{\mathbb Z}}_q$-equivariantization, then ${{\mathcal Z}}({{\mathcal C}})$ must contain a Tannakian subcategory equivalent to ${\operatorname{Rep}}{{\mathbb Z}}_q$ which embeds into ${{\mathcal C}}$ under the forgetful functor. See [@ENO2 Propositions 2.9 and 2.10].
Since $n \geq 5$, then the group of invertible objects of the Drinfeld center of ${{\mathcal C}}$ coincides with ${{\mathbb Z}}_2$ if ${{\mathcal C}}= {\operatorname{Rep}}J_n$, and it is trivial if ${{\mathcal C}}= {\operatorname{Rep}}K_n$. Since ${\operatorname{Rep}}{{\mathbb Z}}_q$ is a pointed fusion category, we get (iii).
Suppose that ${{\mathcal C}}= {\operatorname{Rep}}J_n$. Then ${{\mathcal C}}$ is a ${{\mathbb Z}}_2$-extension of ${\operatorname{Rep}}K_n$. This implies that the pointed subcategory of ${{\mathcal Z}}({{\mathcal C}})$ is a Tannakian subcategory which maps to the trivial subcategory of ${{\mathcal C}}$ under the forgetful functor. Hence, for every prime number $q$, ${{\mathcal C}}$ is not a ${{\mathbb Z}}_q$-equivariantization of any fusion category and we get (i). On the other hand, if it is ${{\mathbb Z}}_q$-extension, then $q = 2$ and the corresponding Tannakian subcategory of ${{\mathcal Z}}({{\mathcal C}})$ coincides with the pointed subcategory ${{\mathcal Z}}({{\mathcal C}})_{pt}$. Thus we obtain (ii). This finishes the proof of the lemma.
\[inv-sdp\] Let $p$ be an odd prime number. Then the group $G(J_p^*)$ is a semidirect product $\widehat{C_p} \rtimes \langle \sigma \rangle$ where $\sigma
\in {\mathbb{S}}_{p-1}$ is a $(p-1)$-cycle and the action of $\langle \sigma \rangle$ on $\widehat{C_p}$ is induced by the action $\lhd': C_p
\times {\mathbb{S}}_{p-1} \to C_p$. Moreover, the subgroup $G(K_p^*) \subseteq G(J_p^*)$ is the semidirect product $\widehat{C_p} \rtimes \langle \sigma^2 \rangle$.
The subgroup ${\mathbb{S}}_{p-1}^{C_p}$ of invariants of ${\mathbb{S}}_{p-1}$ under the action $\rhd'$ of $C_p$ coincides with the subgroup of invariants under the action $\lhd$. It follows from [@jm Corollary 5.2] that ${\mathbb{S}}_{p-1}^{C_p}$ is cyclic generated by a $(p-1)$ cycle $\sigma$, i.e. ${\mathbb{S}}_{p-1}^{C_p}=\langle \sigma \rangle$ and therefore $G(J_p^*)\cong \widehat{C_p} \rtimes \langle \sigma \rangle$. It follows from this that the invariant subgroup ${\mathbb A}_{p-1}^{C_p}$ is also cyclic generated by $\sigma^2$. This implies the lemma, in view of Remark \[inv-split\].
\[triv-cent\] Let $p$ be an odd prime number. Then the Hopf algebras $J_p$, $K_p$ satisfy the following properties:
\(i) ${\mathrm{cd}}(J_p) = {\mathrm{cd}}(K_p) = \{1, p \}$.
\(ii) The groups $G(J_p^*)$ and $G(K_p^*)$ have trivial centers.
Part (i) follows from the description of irreducible representations of crossed products in [@MW].
We next show (ii). Recall that $C_p = \langle z \rangle$, where $z = (12\dots p)$ and the actions $\lhd$, $\rhd$ are determined by the relation $s c = (s\rhd c) (s \lhd c)$ in ${\mathbb{S}}_p$. So that the actions $\lhd'$, $\rhd'$ are determined by $cs = (c \rhd'
s) (c\lhd' s)$ in ${\mathbb{S}}_p$, for all $s \in {\mathbb{S}}_{p-1}$, $c \in C_p$.
It follows from the proof of [@jm Lemma 3.2] that $z \lhd' a_i = z^i$, for all $i = 1,
\dots, p-1$, where $a_i = (p-1, p-i)$. In addition, the stabilizer of $z$ under the action $\lhd'$ coincides with the subgroup $F_z = \{a \in {\mathbb{S}}_{p-1}:\,
a(p-1) = p-1\} \cong {\mathbb{S}}_{p-2}$. These imply that, for all $i = 1, \dots, p-1$, the stabilizer of $z^i$ coincides with the subgroup $F_{z^i} = \{a \in {\mathbb{S}}_{p-1}:\, a(p-i) = p-i\}$. In particular, $C_p$ has no nontrivial fixed points under the action $\lhd'$.
On the other hand, the nontrivial powers of the $(p-1)$-cycle $\sigma \in
{\mathbb{S}}_{p-1}$ have no fixed point in $\{1, \dots, p-1\}$. Hence no nontrivial power of $\sigma$ acts trivially on $C_p$ under the action $\lhd'$.
By Lemma \[inv-sdp\], $G(J_p^*) = \widehat{C_p} \rtimes
\langle \sigma \rangle$ is a semidirect product with respect to the action induced by $\lhd'$, where $\sigma$ is a $(p-1)$-cycle in ${\mathbb{S}}_{p-1}$. Then the center of $G(J_p^*)$ consists of all pairs $(e, x)$, where $x \in
\langle \sigma \rangle$ acts trivially on $C_p$ under the action $\lhd'$.
Similarly, $G(K_p^*) = \widehat{C_p} \rtimes
\langle \sigma^2 \rangle$ is a semidirect product with respect to the action induced by $\lhd'$ and the center of $G(K_p^*)$ consists of all pairs $(e, x)$, where $x \in
\langle \sigma^2 \rangle$ acts trivially on $C_p$ under the action $\lhd'$.
Thus we obtain that the centers of the groups $G(J_p^*)$ and $G(K_p^*)$ are both trivial.
\[jp-kp\] Let $\tilde{{\mathcal C}}$ be a fusion category. Suppose that $\tilde {{\mathcal C}}$ is Grothendieck equivalent to one of the categories ${\operatorname{Rep}}J_p$ or ${\operatorname{Rep}}K_p$. Then $\tilde {{\mathcal C}}$ is not solvable.
Suppose on the contrary that $\tilde {{\mathcal C}}$ is solvable and Grothendieck equivalent to ${{\mathcal C}}$, where ${{\mathcal C}}= {\operatorname{Rep}}J_p$ or ${\operatorname{Rep}}K_p$. It follows from Proposition \[groth-eq\], that $G(\tilde {{\mathcal C}}) \cong G({{\mathcal C}})$. By Proposition \[triv-cent\], the groups of invertible objects of ${\operatorname{Rep}}J_p$ and ${\operatorname{Rep}}K_p$ have trivial center. Then the center of $G(\tilde {{\mathcal C}})$ is trivial as well. It follows from Lemma \[g-central\] that, for every prime number $q$, the category $\tilde {{\mathcal C}}$ is not a ${{\mathbb Z}}_q$-equivariantization of any fusion category. Therefore $\tilde {{\mathcal C}}$ must be a ${{\mathbb Z}}_q$-extension of a fusion subcategory $\tilde {{\mathcal D}}$, and $\tilde {{\mathcal D}}$ is also a solvable fusion category. Hence ${{\mathcal C}}$ is a ${{\mathbb Z}}_q$-extension of a fusion subcategory ${{\mathcal D}}$ such that $\tilde {{\mathcal D}}$ is Grothendieck equivalent to ${{\mathcal D}}$. It follows from Proposition \[RepJn\] and Lemma \[not-ext\] that ${{\mathcal C}}= {\operatorname{Rep}}J_p$, $q = 2$ and ${{\mathcal D}}= {\operatorname{Rep}}K_p$. Applying the same argument to the solvable fusion category $\tilde {{\mathcal D}}$ we get a contradiction. This shows that $\tilde {{\mathcal C}}$ cannot be solvable and finishes the proof of the theorem.
Fusion rules of ${\operatorname{Rep}}K_5$
-----------------------------------------
In this subsection we determine explicitly the fusion rules of the category ${\operatorname{Rep}}K_p$ in the case $p = 5$. It follows from [@MW] that simple objects of the category ${\operatorname{Rep}}K_5$ are parameterized by pairs $(s, \rho)$, where $s$ runs over a set of representatives of the orbits of $C_5$ on ${\mathbb A}_4$ and $\rho$ is an irreducible representation of the stabilizer $F_s \subseteq C_5$. The dimension of the simple object $S_{s, \rho}$ corresponding to the pair $(s,
\rho)$ is given by the formula $\dim S_{s, \rho} = [C_5: F_s]$.
The $C_5$-action on ${\mathbb{S}}_4$ is explicitly determined in [@jm Table 1]. We have in this case that there are $10$ fixed points and the remaining 2 orbits consist of $5$ distinct elements each. Furthermore, there are exactly $4$ distinct fixed points $\sigma$ such that $\sigma = \sigma^{-1}$ and both nontrivial orbits contain elements of order $2$. In view of [@jm Theorem 4.8], ${\operatorname{Rep}}K_5$ has $5$ invertible objects of order $2$ and the $5$-dimensional simple objects are self-dual.
Let us denote by $Y, Y'$ the simple objects corresponding to the nontrivial orbits $\mathcal O, \mathcal O'$, respectively. By [@jm Table 1], we have $$\begin{aligned}
& \mathcal O = \{(123), (243), (132), (13)(24), (234)\}, \\
& \mathcal O' =
\{(124), (143), (134), (12)(34), (142)\}.\end{aligned}$$ By Lemma \[inv-sdp\], the group $G({\operatorname{Rep}}K_5)$ is isomorphic to the dihedral group $D_5$ of order $10$. The unique subgroup $R$ of order $5$ of $G({\operatorname{Rep}}K_5)$ coincides therefore with the stabilizer of $Y$ and $Y'$ under left (or right) multiplication. Since every element $s$ outside of $R$ is of order $2$, then $s\otimes Y \cong Y
\otimes s \cong Y'$. So that we have a decomposition $$\label{a-b}
Y \otimes Y^* \cong Y\otimes Y \cong \bigoplus_{r \in R} r \oplus aY
\oplus bY',$$ where $a, b \geq 0$ and $a+b = 4$. Letting $F: {\operatorname{Rep}}K_5 \to {{\mathcal C}}({\mathbb A}_{4}) = {\operatorname{Rep}}k^{{\mathbb A}_{4}}$ denote the restriction functor, we obtain $$\begin{aligned}
& F(Y) = V_{(123)} \oplus V_{(243)}\oplus V_{(132)}\oplus V_{
(13)(24)}\oplus V_{ (234)},\\
& F(Y') = V_{(124)} \oplus V_{(143)} \oplus V_{(134)} \oplus V_{(12)(34)}
\oplus V_{(142)},\end{aligned}$$ where, for each $s \in {\mathbb A}_4$, $V_s$ denotes the one-dimensional simple $k^{{\mathbb A}_4}$-module corresponding to $s$. Comparing these relations with , we find that $a = b = 2$. Hence the fusion rules of ${\operatorname{Rep}}K_5$ are determined by the condition $G = G({\operatorname{Rep}}K_5) \cong D_5$, $g \otimes Y = Y
\otimes g = Y'$, for every element or order $2$ of $G$, and $$Y\otimes Y \cong \bigoplus_{r \in R} r \oplus 2Y
\oplus 2Y'\cong Y'\otimes Y',$$ where $R$ is the unique subgroup of order $5$ of $G$.
The dual Hopf algebras $J_n^*$, $K_n^*$
---------------------------------------
Let $n \geq 2$ be a natural number and let $H_n = J_n^*$. Recall that there is a split exact sequence of Hopf algebras $$\label{hn-ln} k \longrightarrow k^{C_n} \longrightarrow H_n
\longrightarrow k{\mathbb{S}}_{n-1}
\longrightarrow k.$$ Suppose that $n$ is odd. Let $L_n = K_n^*$, so that there is a split exact sequence of Hopf algebras $$\label{hn-ln-2}k \longrightarrow k^{C_n} \longrightarrow L_n
\longrightarrow k{\mathbb A}_{n-1}
\longrightarrow k.$$ Moreover, by Lemma \[secjnkn\] there is a cocentral exact sequence $$\label{cocentral}
k \longrightarrow L_n \longrightarrow H_n \longrightarrow k{{{\mathbb Z}}_2}
\longrightarrow k.$$
\[dual-next\] Suppose $n \geq 5$. Since $D(H_n) \cong D(J_n)$, then $G({\operatorname{Rep}}D(H_n)) \cong {{\mathbb Z}}_2$. Let $q$ be a prime number. If the category ${\operatorname{Rep}}H_n$ is a ${{\mathbb Z}}_q$-extension or a ${{\mathbb Z}}_q$-equivariantization of a fusion category, then $q = 2$.
Suppose $n$ is odd. In view of [@ext-ty Proposition 3.5], ${\operatorname{Rep}}H_n$ is a ${{\mathbb Z}}_2$-equivariantization of ${\operatorname{Rep}}L_n$. As in the proof of Lemma \[not-ext\], we obtain that if $n \geq 5$, the category ${\operatorname{Rep}}H_n$ is not a ${{\mathbb Z}}_q$-extension of any fusion category. Similarly, ${\operatorname{Rep}}L_n$ is not a ${{\mathbb Z}}_q$-extension or a ${{\mathbb Z}}_q$-equivariantization of any fusion category.
\[grupos-deg2\] Let $n \geq 5$ be a natural number. Then the following hold.
\(i) $G({\operatorname{Rep}}H_n) \cong {{\mathbb Z}}_2$.
\(ii) ${\operatorname{Rep}}(H_5)$ is of type $(1, 2; 2, 1; 3, 2; 4, 2; 8, 1)$ and $H_n$ has no irreducible representation of dimension 2, for all $n > 5$.
Assume in addition that $n$ is odd. Then
\(iii) $G({\operatorname{Rep}}L_n) = {\textbf{1}}$, if $n > 5$.
\(iv) ${\operatorname{Rep}}L_5$ is of type $(1, 3; 3, 1; 4, 3)$ and $L_n$ has no irreducible representation of dimension 2.
Consider the exact sequences , . The respective invariant subgroups $C_n^{{\mathbb{S}}_{n-1}}$ and $C_n^{{\mathbb A}_{n-1}}$ are both trivial. Parts (i) and (iii) follow from Lemma \[inv-abel\].
Since $H_n$ is a split abelian extension of $k^{C_n}$ by $k{\mathbb{S}}_{n-1}$ and the action of ${\mathbb{S}}_{n-1}$ has two orbits $\{e\}$ and $\{z, \dots, z^{n-1}\}$, then the simple $H_n$-modules are classified by pairs $(t, \rho)$, where either $t = e$ and $\rho$ is an irreducible representation of $F_e = {\mathbb{S}}_{n-1}$, or $t = z$ and $\rho$ is an irreducible representation of $F_z = \{a \in {\mathbb{S}}_{n-1}:\, a(n-1) = n-1 \} \cong {\mathbb{S}}_{n-2}$. See [@jm Lemma 3.2]. If $S_{t, \rho }$ is the simple module corresponding to the pair $(t, \rho)$, we have in addition $\dim S_{e, \rho} = \dim \rho$, and $\dim S_{z, \rho} =
[{\mathbb A}_{n-1}: F_z]\dim \rho = (n-1)\dim \rho$. This implies the statement for $H_5$ in part (ii).
Suppose that $n > 5$. Then $\dim S_{z,
\rho} > 2$, for all $\rho$. As observed in Remark \[rmk-ansn-2\], the symmetric group ${\mathbb{S}}_{n-1}$ has no irredubible representation of degree $2$. Therefore we also get that $\dim S_{e, \rho} \neq 2$, for all $\rho$. In conclusion $H_n$ has no irreducible representation of dimension 2, and we obtain (ii).
Suppose that $n$ is odd. Similarly, $L_n$ is a split abelian extension of $k^{C_n}$ by $k{\mathbb A}_{n-1}$ and the action of ${\mathbb A}_{n-1}$ has two orbits $\{e\}$ and $\{z, \dots, z^{n-1}\}$. Hence the simple $L_n$-modules are classified by pairs $(t, \rho)$, where either $t = e$ and $\rho$ is an irreducible representation of $F_e \cap {\mathbb A}_{n-1} = {\mathbb A}_{n-1}$, or $t = z$ and $\rho$ is an irreducible representation of $F_z \cap {\mathbb A}_{n-1} \cong {\mathbb A}_{n-2}$. This implies that ${\operatorname{Rep}}L_5$ is of the prescribed type. As before, $\dim S_{z, \rho} > 2$, for all $\rho$, and since the alternating group ${\mathbb A}_{n-1}$ has no irredubible representation of degree $2$, then also $\dim S_{e, \rho} \neq 2$, for all $\rho$. So that $L_n$ has no irreducible representation of dimension 2. This proves part (iv) and finishes the proof of the lemma.
\[nsol-type\] Let ${{\mathcal C}}$ be a fusion category of type $(1, 3; 3, 1; 4, 3)$. Then ${{\mathcal C}}$ is not solvable.
The assumption on the type of ${{\mathcal C}}$ implies that the simple objects of dimensions $1$ and $3$ generate a fusion subcategory ${{\mathcal B}}$ of ${{\mathcal C}}$ of type $(1, 3; 3, 1)$ and moreover, every fusion subcategory of ${{\mathcal C}}$ is contained in ${{\mathcal B}}$.
Suppose first that ${{\mathcal C}}$ is a ${{\mathbb Z}}_q$-extension of a fusion subcategory ${{\mathcal C}}_0$, for some prime number $q$. Then necessarily ${{\mathcal C}}_0 = {{\mathcal B}}$ and $q = 5$. Hence we have a ${{\mathbb Z}}_5$-faithful grading ${{\mathcal C}}= {{\mathcal C}}_0 \oplus {{\mathcal C}}_1 \oplus \dots \oplus {{\mathcal C}}_4$, with trivial component ${{\mathcal C}}_0 = {{\mathcal B}}$. But ${{\mathcal C}}$ has only 3 classes of simple objects outside of ${{\mathcal B}}$, entailing that such a decomposition is impossible.
Therefore ${{\mathcal C}}$ must be a ${{\mathbb Z}}_q$-equivariantization of a fusion category ${{\mathcal D}}$, for some prime number $q$, where ${{\mathcal D}}$ is also a solvable fusion category. Thus $q = 3$ and ${{\mathcal C}}\cong {{\mathcal D}}^{{{\mathbb Z}}_3}$. The description of simple objects of ${{\mathcal D}}^{{{\mathbb Z}}_3}$ together with the assumption on the type of ${{\mathcal C}}$ imply that ${{\mathcal D}}$ must be of type $(1, 4; 4, 1)$; c.f. Formula . Moreover, the action (by group automorphisms) of ${{\mathbb Z}}_3$ on the set of nontrivial invertible objects of ${{\mathcal D}}$ must be transitive, hence $G({{\mathcal D}}) \cong {{\mathbb Z}}_2 \times {{\mathbb Z}}_2$.
On the other hand, letting $X$ be the unique noninvertible simple object of ${{\mathcal D}}$, we must have $$X^{\otimes 2} \cong \oplus_{Y \in G({{\mathcal D}})}Y \oplus 3X.$$ This means that ${{\mathcal D}}$ is a near-group fusion category of type $(G, 3)$, where $G =G({{\mathcal D}})$. Then it follows from [@siehler Theorem 1.2] that the group $G({{\mathcal D}})$ is cyclic, which leads to a contradiction. Therefore ${{\mathcal C}}$ cannot be solvable. This finishes the proof of the lemma.
\[dual-jn-kn\] Let ${{\mathcal C}}$ be a fusion category and let $n \geq 5$ be a natural number. Then we have:
\(i) If ${{\mathcal C}}$ is Grothendieck equivalent to ${\operatorname{Rep}}H_n$, then ${{\mathcal C}}$ is not solvable.
\(ii) Suppose that $n$ is odd. If ${{\mathcal C}}$ is Grothendieck equivalent to ${\operatorname{Rep}}L_n$, then ${{\mathcal C}}$ is not solvable.
To show part (i), suppose first that $n > 5$. By Lemma \[grupos-deg2\], $G({\operatorname{Rep}}H_n) \cong {{\mathbb Z}}_2$ and ${\operatorname{Rep}}H_n$ has no simple objects of dimension $2$. The claim follows in this case from Proposition \[cor-cyc\].
Consider the case $n = 5$. Suppose on the contrary that ${{\mathcal C}}$ is Grothendieck equivalent to ${\operatorname{Rep}}H_5$ and ${{\mathcal C}}$ is solvable. Then ${{\mathcal C}}$ is of type $(1, 2; 2, 1; 3, 2; 4, 2; 8, 1)$ and, for any prime $q$, ${{\mathcal C}}$ is not a ${{\mathbb Z}}_q$-extension of any fusion subcategory, in view of Remark \[dual-next\]. Therefore ${{\mathcal C}}$ must be a ${{\mathbb Z}}_2$-equivariantization of a solvable fusion category ${{\mathcal D}}$ of dimension $60$. The description of the simple objects of ${{\mathcal D}}^{{{\mathbb Z}}_2}$ together with the assumption on the type of ${{\mathcal C}}$ imply that ${{\mathcal D}}$ must be of type $(1, 3; 3, 1; 4, 3)$. Lemma \[nsol-type\] implies that ${{\mathcal D}}$ is not solvable, which is a contradiction. Thus we get (i).
Let us show (ii). If $n = 5$, the result follows from Lemma \[nsol-type\]. Suppose next that $n > 5$. Since a solvable fusion category contains nontrivial invertible objects, then part (ii) follows from Lemma \[grupos-deg2\] (iii).
Further examples associated to the symmetric group
--------------------------------------------------
Let $n \geq 2$ be a natural number. Consider the matched pair $({\mathbb A}_n, C_2)$, where $C_2 = \langle (12) \rangle \subseteq {\mathbb{S}}_n$, the action $\lhd: {\mathbb A}_n
\times C_2 \to {\mathbb A}_n$ is given by conjugation in ${\mathbb{S}}_n$ and the action $\rhd: {\mathbb A}_n \times C_2 \to C_2$ is trivial. The associated group ${\mathbb A}_n \bowtie
C_2$ is isomorphic to the symmetric group ${\mathbb{S}}_n$.
Let $B_n = k^{{\mathbb A}_n}\# kC_2$ be the split abelian extension associated to this matched pair. We have ${\mathrm{cd}}(B_n) = \{1, 2\}$. The fusion category ${\operatorname{Rep}}B_n$ is Morita equivalent to ${\mathbb{S}}_n$ and therefore it is not solvable if $n \geq 5$.
Since the action $\rhd$ is the trivial one and $C_2 \cong {{\mathbb Z}}_2$, there is a cocentral exact sequence $$\label{cocentral-B_n}
k \longrightarrow k^{{\mathbb A}_n} \longrightarrow B_n \longrightarrow k{{{\mathbb Z}}_2}
\longrightarrow k.$$ In view of [@ext-ty Proposition 3.5] ${\operatorname{Rep}}B_n$ is a ${{\mathbb Z}}_2$-equivariantization of ${\operatorname{Rep}}k^{\mathbb A}_n={{\mathcal C}}({\mathbb A}_n)$.
Moreover, since ${\operatorname{Rep}}B_n$ is Morita equivalent to ${\mathbb{S}}_n$ and the group of invertible objects of ${{\mathcal Z}}({\operatorname{Rep}}{\mathbb{S}}_n)$ is cyclic of order 2, then for all prime number $q$, ${\operatorname{Rep}}B_n$ is not a ${{\mathbb Z}}_q$-extension of any fusion category and if it is a ${{\mathbb Z}}_q$-equivariantization, then $q = 2$ (compare with Proposition \[RepJn\]). In particular, not being a ${{\mathbb Z}}_q$-extention, $B_n$ has no nontrivial central group-like elements; that is, $Z(B_n) \cap G(B_n) = \{1\}$.
Our first statement concerns the dual Hopf algebra $B_n^*$.
\[bn\*\] Let ${{\mathcal C}}$ be a fusion category. Suppose that ${{\mathcal C}}$ is Grothendieck equivalent to ${\operatorname{Rep}}B_n^*$, $n \geq 5$. Then ${{\mathcal C}}$ is not solvable.
The dual Hopf algebra $B_n^*$ fits into a central exact sequence $$\label{central-B_n*}
k \longrightarrow k^{{{\mathbb Z}}_2} \longrightarrow B_n \longrightarrow k{{\mathbb A}_n}
\longrightarrow k.$$ Therefore ${\operatorname{Rep}}B_n^*$ is a ${{\mathbb Z}}_2$-extension of ${\operatorname{Rep}}{\mathbb A}_n$. Hence ${{\mathcal C}}$ is a ${{\mathbb Z}}_2$-extension of a fusion category ${{\mathcal D}}$, which is Grothendieck equivalent to ${\operatorname{Rep}}{\mathbb A}_n$.
Suppose on the contrary that ${{\mathcal C}}$ is solvable. Then so is ${{\mathcal D}}$ and therefore ${{\mathcal D}}_{pt} \neq {\operatorname{Vect}}$. This implies that ${\mathbb A}_n$ has nontrivial one-dimensional representations, which is a contradiction. Then ${{\mathcal C}}$ cannot be solvable, as claimed.
\[g-repbn\] The group $G({\operatorname{Rep}}B_n)$ of invertible objects of the category ${\operatorname{Rep}}B_n$ is isomorphic to the direct product $\widehat{C_2} \times C_{{\mathbb A}_n}(12)$, where $C_{{\mathbb A}_n}(12)$ denotes the centralizer in ${\mathbb A}_n$ of the transposition $(12)$.
There is a group isomorphism $G({\operatorname{Rep}}B_n) \cong G(B_n^*)$. On the other hand, $B_n^*$ is a split abelian extension $B_n^* \cong k^{C_2} \# k{\mathbb A}_n$, associated to the adjoint action $\rhd': C_2 \times {\mathbb A}_n \to {\mathbb A}_n$ and the trivial action $\lhd': C_2 \times {\mathbb A}_n \to C_2$. The result follows from Lemma \[inv-abel\].
\[bn\] Suppose $n \geq 5$. Let $\tilde{{\mathcal C}}$ be a fusion category Grothendieck equivalent to ${\operatorname{Rep}}B_n$. Then $\tilde{{\mathcal C}}$ is not solvable.
Suppose first that $n \geq 7$. By Lemma \[g-repbn\], $G({\operatorname{Rep}}B_n) \cong \widehat{C_2} \times C_{{\mathbb A}_n}(12)$. Note that $C_{{\mathbb A}_n}(12)$ contains the subgroup $\{\sigma \in {\mathbb A}_n:\, \sigma(1) = 1,
\sigma(2) = 2\} \cong {\mathbb A}_{n-2}$. Since $n \geq 7$, the group ${\mathbb A}_{n-2}$ is not solvable. Then $G(\tilde{{\mathcal C}})$ is not solvable neither and then $\tilde{{\mathcal C}}$ is not solvable.
It remains to consider the cases $n = 5$ and $6$. It follows from Lemma \[g-repbn\] that $G({\operatorname{Rep}}B_5) \cong \widehat {C_2} \times {\mathbb{S}}_3$ is non-abelian of order 12. Hence ${\operatorname{Rep}}B_5$ is of type $(1,12;2,27)$. Similarly, $G({\operatorname{Rep}}B_6) \cong \widehat {C_2} \times {\mathbb{S}}_4$ is non-abelian of order 48 and ${\operatorname{Rep}}B_6$ is of type $(1,48; 2,168)$.
Suppose that there exist a solvable fusion category $\tilde {{\mathcal C}}$ which is Grothendieck equivalent to ${{\mathcal C}}$, where ${{\mathcal C}}= {\operatorname{Rep}}B_5$ or ${\operatorname{Rep}}B_6$. By Proposition \[groth-eq\], $G(\tilde{{\mathcal C}}) \cong G({{\mathcal C}})$. Since, for every prime number $q$, ${{\mathcal C}}$ is not a ${{\mathbb Z}}_q$-extension of any fusion category, we have that $\tilde{{\mathcal C}}$ must be a ${{\mathbb Z}}_q$-equivariantization of a fusion subcategory ${{\mathcal D}}$, and ${{\mathcal D}}$ is also a solvable fusion category. Moreover, $q=2$ because ${{\mathbb Z}}_q \subseteq Z(G(\tilde{{\mathcal C}}))$ and ${\operatorname{FPdim}}{{\mathcal D}}= 60$ or ${\operatorname{FPdim}}{{\mathcal D}}= 360$, respectively. Then there is an exact sequence of fusion categories $${\operatorname{Rep}}{{\mathbb Z}}_2 \to \tilde{{\mathcal C}}\to {{\mathcal D}}.$$ Since ${\mathrm{cd}}(\tilde{{\mathcal C}})=\{1,2\}$, it follows that ${\mathrm{cd}}({{\mathcal D}})=\{1,2\}$. The previous exact sequence induces an exact sequence of groups $$1 \to \widehat{{\mathbb Z}}_2 \to G(\tilde{{\mathcal C}}) \to G_0({{\mathcal D}}) \to 1,$$ where $\widehat{{\mathbb Z}}_2$ denotes the group of invertible characters of ${{\mathbb Z}}_2$ and $G_0({{\mathcal D}})$ is the subgroup of $G({{\mathcal D}})$ consisting of isomorphism classes of invertible objects which are ${{\mathbb Z}}_2$-equivariant. See [@fusionrules-equiv Remark 3.1]. As ${{\mathbb Z}}_2$ is a cyclic group, we have that $G_0({{\mathcal D}})$ coincides with the subgroup of fixed points of the induced action of ${{\mathbb Z}}_2$ on the group of invertible objects of ${{\mathcal D}}$.
Observe that, since ${{\mathcal C}}$ is also a ${{\mathbb Z}}_2$-equivariantization of ${{\mathcal C}}_{{{\mathbb Z}}_2} \cong {{\mathcal C}}({\mathbb A}_5)$ or ${{\mathcal C}}({\mathbb A}_6)$, respectively, then the group $G({{\mathcal C}}) \cong G(\tilde {{\mathcal C}})$ also fits into an exact sequence $$1 \to \widehat{{\mathbb Z}}_2 \to G({{\mathcal C}}) \to G_0({{\mathcal C}}_{{{\mathbb Z}}_2}) \to 1.$$ In this case, the subgroup $G_0({{\mathcal C}}_{{{\mathbb Z}}_2})$ is isomorphic to $C_{{\mathbb A}_5}(12)$ or $C_{{\mathbb A}_6}(12)$, respectively. In addition, the group $G({{\mathcal C}})$ contains a unique normal subgroup of order 2. Therefore, $G_0({{\mathcal D}}) \cong G_0({{\mathcal C}}_{{{\mathbb Z}}_2})$ is a non-abelian subgroup of $G({{\mathcal D}})$.
Suppose first that $n = 5$. In this case ${{\mathcal C}}={\operatorname{Rep}}B_5 \cong {{\mathcal C}}({\mathbb A}_5)^{{{\mathbb Z}}_2}$. In particular, $G_0({{\mathcal D}})$ is a subgroup of order 6 of $G({{\mathcal D}})$. A counting argument shows that $G({{\mathcal D}})$ can be of type $(1,12;2,12)$ or else ${{\mathcal D}}$ is pointed. Suppose that ${{\mathcal D}}$ is of type $(1,12;2,12)$. Then ${{\mathcal D}}_{pt}^{{{\mathbb Z}}_2} \subseteq \tilde{{\mathcal C}}$ is a fusion subcategory of dimension $24$ and type $(1,12;2,3)$, containing the central subcategory ${\operatorname{Rep}}{{\mathbb Z}}_2$. Therefore ${{\mathcal C}}$ has a fusion subcategory ${{\mathcal B}}$ of type $(1,12;2,3)$ containing the central subcategory ${\operatorname{Rep}}{{\mathbb Z}}_2$. (In fact, as observed before, $G({{\mathcal C}})$ contains a unique normal subgroup of order $2$; see Lemma \[g-central\].)
Consider the de-equivariantization ${{\mathcal B}}_{{{\mathbb Z}}_2} \subseteq {{\mathcal C}}({\mathbb A}_5)$. We have that $\dim{{\mathcal B}}_{{{\mathbb Z}}_2}=12$ and thus ${{\mathcal B}}_{{{\mathbb Z}}_2}={{\mathcal C}}(H)$, where $H$ is a (${{\mathbb Z}}_2$-stable) subgroup of ${\mathbb A}_5$ of order 12. Since $G({\operatorname{Rep}}B_5) \subseteq {{\mathcal B}}$, then the subgroup $H$ contains the invariant sugroup ${\mathbb A}_5^{{{\mathbb Z}}_2} = C_{{\mathbb A}_5}(12) \cong {\mathbb{S}}_3$. On the other hand, every subgroup of order $12$ of ${\mathbb A}_5$ is isomorphic to ${\mathbb A}_4$, then $H\cong {\mathbb A}_4$. This leads to a contradiction, because ${\mathbb A}_4$ has no subgroup of order $6$. This proves that ${{\mathcal D}}$ cannot be of type $(1,12;2,12)$.
Therefore ${{\mathcal D}}$ must be a pointed fusion category. In this case ${{\mathcal D}}= {{\mathcal C}}(\Gamma,\omega)$, where $\omega:\Gamma \times \Gamma \times \Gamma \to k^*$ is a 3-cocyle and $\Gamma$ is a solvable non-abelian subgroup of order 60. In addition ${{\mathbb Z}}_2$ acts on $\Gamma$ by group automorphisms and $\Gamma^{{{\mathbb Z}}_2} \cong {\mathbb{S}}_3$. Since $\Gamma \neq {\mathbb A}_5$, $\Gamma$ can be isomorphic to ${\mathbb A}_4 \times {{\mathbb Z}}_5$, ${{\mathbb Z}}_{15} \rtimes {{\mathbb Z}}_4$ or ${{\mathbb Z}}_{15} \rtimes ({{\mathbb Z}}_2 \times {{\mathbb Z}}_2)$.
If $\Gamma \cong {\mathbb A}_4 \times {{\mathbb Z}}_5$, the action of ${{\mathbb Z}}_2$ must fix ${\mathbb A}_4$ and ${{\mathbb Z}}_5$. Since $|\Gamma^{{{\mathbb Z}}_2}|=6$ then $\Gamma^{{{\mathbb Z}}_2} \subseteq {\mathbb A}_4$, and we reach a contradiction. Therefore $\Gamma \cong {{\mathbb Z}}_{15} \rtimes {{\mathbb Z}}_4$ or ${{\mathbb Z}}_{15} \rtimes ({{\mathbb Z}}_2 \times {{\mathbb Z}}_2)$. In this case $\Gamma$ has a unique subgroup $L$ of order 15, and then $L$ is ${{\mathbb Z}}_2$-stable and ${{\mathcal C}}(L)^{{{\mathbb Z}}_2}$ is a fusion subcategory of $\tilde{{\mathcal C}}$ of dimension $30$. This implies that ${{\mathcal C}}$ has a fusion subcategory of dimension 30. Such fusion subcategory must correspond to a quotient Hopf algebra of $B_5$ of dimension $30$, which is a contradiction, because $Z(B_5)\cap G(B_5) = \{1\}$. See [@ssld Corollary 1.4.3]. Thus ${{\mathcal D}}$ cannot be pointed. This proves that if $\tilde{{\mathcal C}}$ is Grothendieck equivalent to ${\operatorname{Rep}}B_5$ then $\tilde{{\mathcal C}}$ is not solvable.
Finally, let us consider the case $n = 6$. In this case we have ${{\mathcal C}}:={\operatorname{Rep}}B_6 \cong {{\mathcal C}}({\mathbb A}_6)^{{{\mathbb Z}}_2}$. On the other hand, $G_0({{\mathcal D}})$ is a subgroup of order 24 of $G({{\mathcal D}})$. As before, one can see that ${{\mathcal D}}$ must be of type $(1,24;2,84)$, $(1,72;2,72)$, $(1,120;2,60)$, or else ${{\mathcal D}}$ is pointed.
Suppose that ${{\mathcal D}}$ is of type $(1,72;2,72)$ or $(1,120;2,60)$. In these cases ${{\mathcal D}}_{pt}^{{{\mathbb Z}}_2} \subseteq \tilde{{\mathcal C}}$ is a fusion subcategory of dimension 144 or 240, respectively, containing the central subcategory ${\operatorname{Rep}}{{\mathbb Z}}_2$. Therefore ${{\mathcal C}}$ has a fusion subcategory ${{\mathcal B}}$ of dimension 144 or 240, respectively, containing the central subcategory ${\operatorname{Rep}}{{\mathbb Z}}_2$. The de-equivariantization ${{\mathcal B}}_{{{\mathbb Z}}_2} \subseteq {{\mathcal C}}({\mathbb A}_6)$ is of dimension $\dim{{\mathcal B}}_{{{\mathbb Z}}_2}=72$ or 120, respectively. Then ${{\mathcal B}}_{{{\mathbb Z}}_2}={{\mathcal C}}(H)$, where $H$ is a ${{\mathbb Z}}_2$-stable subgroup of ${\mathbb A}_6$ of order $72$ or 120, respectively. Since ${\mathbb A}_6$ has no subgroups of order 72 or 120, it follows that these types are not possible for ${{\mathcal D}}$.
Suppose next that ${{\mathcal D}}$ is of type $(1,24;2,84)$. It follows from the description of the simple objects of ${{\mathcal D}}^G$ and the fact that ${\mathrm{cd}}({{\mathcal D}}) = {\mathrm{cd}}({{\mathcal C}}) =\{1,2\}$, that ${{\mathbb Z}}_2$ acts trivially on the set ${\operatorname{Irr}}({{\mathcal D}})$; see . In particular, $G({{\mathcal D}}) = G_0({{\mathcal D}}) \cong C_{{\mathbb A}_6}(12) \cong {\mathbb{S}}_4$.
Since ${{\mathcal D}}$ is solvable, then it is a ${{\mathbb Z}}_p$-extension or a ${{\mathbb Z}}_p$-equivariantization, where $p$ is a prime number that divides the dimension of ${{\mathcal D}}$, which is $360$. If ${{\mathcal D}}$ were a ${{\mathbb Z}}_p$-equivariantization then, by Lemma \[g-central\], ${{\mathbb Z}}_p \subseteq Z(G({{\mathcal D}}))$, which is a contradiction because $G({{\mathcal D}}) \cong {\mathbb{S}}_4$. Therefore ${{\mathcal D}}$ must be a ${{\mathbb Z}}_p$-extension of a fusion subcategory ${{\mathcal D}}_e$. The fusion subcategory ${{\mathcal D}}_e$ is of dimension 72, 120 or 180. Furthermore, ${{\mathcal D}}_e$ must be stable under the action of ${{\mathbb Z}}_2$, since this action is trivial on ${\operatorname{Irr}}({{\mathcal D}})$. As before, this implies that ${{\mathcal C}}$ contains a fusion subcategory ${{\mathcal B}}$ of dimension 144, 240 or 360, respectively, containing the central subcategory ${\operatorname{Rep}}{{\mathbb Z}}_2$. Hence ${{\mathcal B}}_{{{\mathbb Z}}_2}={{\mathcal C}}(H)$, where $H$ is a ${{\mathbb Z}}_2$-stable subgroup of ${\mathbb A}_6$ with order 72, 120 or 180, respectively. But ${\mathbb A}_6$ has no subgroups neither of order 72, 120 nor 180, therefore the type $(1,24;2,84)$ is also impossible for ${{\mathcal D}}$.
Suppose finally that ${{\mathcal D}}$ is a solvable pointed fusion category. We have ${{\mathcal D}}= {{\mathcal C}}(\Gamma,\omega)$, where $\omega:\Gamma \times \Gamma \times \Gamma \to k^*$ is a 3-cocyle and $\Gamma$ is a solvable non-abelian subgroup of order $360$. In addition ${{\mathbb Z}}_2$ acts on $\Gamma$ by group automorphisms and the subgroup $\Gamma_0$ of fixed points of $\Gamma$ under this ${{\mathbb Z}}_2$-action is of order $24$. Let $S$ be a Sylow 5-subgroup of $\Gamma$. Since $\Gamma$ is solvable, there exist $H$, a Hall $\{2,5\}$-subgroup of $\Gamma$, and $K$, a Hall $\{3,5\}$-subgroup of $\Gamma$, such that $S \subseteq H$ and $S \subseteq K$. A counting argument shows that $S\unlhd H$ and $S\unlhd K$ and so $S\unlhd \langle H,K \rangle = \Gamma $. Hence $S$ is the unique Sylow 5-subgroup of $\Gamma$ and then $S$ is ${{\mathbb Z}}_2$-stable. In this case ${{\mathcal C}}(S,\omega|_S)^{{{\mathbb Z}}_2}$ is a fusion subcategory of $\tilde{{\mathcal C}}$ of dimension $10$, containing the central subcategory ${\operatorname{Rep}}{{\mathbb Z}}_2$. Therefore ${{\mathcal C}}$ has a fusion subcategory ${{\mathcal B}}$ with dimension 10, containing the central subcategory ${\operatorname{Rep}}{{\mathbb Z}}_2$. The de-equivariantization ${{\mathcal B}}_{{{\mathbb Z}}_2} \subseteq {{\mathcal C}}({\mathbb A}_6)$ is of dimension $5$. Then ${{\mathcal B}}_{{{\mathbb Z}}_2}={{\mathcal C}}(T)$, where $T$ is a ${{\mathbb Z}}_2$-stable subgroup of ${\mathbb A}_6$ of order $5$. We have that $T=\{{\operatorname{id}},(abcde),(acebd),(adbec),(aedcb)\}$, and without loss of generality we may assume $a=1$ and $b=2$. We thus reach a contradiction, since $(12)(12cde)(12)=(21cde) \neq (1ce2d)$. This proves that $\widetilde{{\mathcal C}}$ cannot be solvable and finishes the proof of the theorem.
Solvability and fusion rules of a braided fusion category {#solv-fr-bfc}
=========================================================
Let ${{\mathcal C}}$ be a fusion category. Suppose that ${\operatorname{FPdim}}{{\mathcal C}}$ is an integer (which is always the case if ${{\mathcal C}}$ is solvable). Then the adjoint subcategory ${{\mathcal C}}_{ad}$ is integral [@ENO Proposition 8.27].
Assume that ${{\mathcal C}}$ is braided. Recall that ${{\mathcal C}}$ is solvable and integral, then either it is pointed or it contains a nontrivial Tannakian subcategory [@witt-wgt Proposition 5.2].
\[e-tann\] Let ${{\mathcal C}}$, $\tilde{{\mathcal C}}$ be Grothendieck equivalent braided fusion categories. Suppose that ${{\mathcal C}}$ is solvable. Then the following hold:
\(i) $\tilde {{\mathcal C}}_{pt}$ is a solvable fusion category and it is not trivial if $\tilde {{\mathcal C}}$ is not trivial.
\(ii) If $\tilde {{\mathcal C}}$ is not pointed, then it contains a nontrivial Tannakian subcategory.
Since ${{\mathcal C}}$ is solvable, [@ENO2 Proposition 4.5 (iv)] implies that ${{\mathcal C}}_{pt}
\neq {\operatorname{Vect}}$. In addition ${{\mathcal C}}_{pt}$ is a solvable fusion category. Hence $\tilde
{{\mathcal C}}_{pt} \neq {\operatorname{Vect}}$. We have ${{\mathcal C}}_{pt} \cong {{\mathcal C}}(G({{\mathcal C}}),
\omega)$ for some invertible 3-cocycle $\omega$ on $G({{\mathcal C}})$. By assumption ${{\mathcal C}}$ is solvable, hence the group $G({{\mathcal C}})$ is solvable.
Moreover, $\tilde {{\mathcal C}}_{pt}$ is Grothendieck equivalent to ${{\mathcal C}}_{pt}$ and therefore there exists an isomorphism of groups $G({{\mathcal C}}) \cong
G(\tilde {{\mathcal C}})$. Hence $G(\tilde
{{\mathcal C}})$, and *a fortiori* also $\tilde {{\mathcal C}}_{pt}$, are solvable. This shows part (i).
Suppose that $\tilde {{\mathcal C}}$ is not pointed, so that ${{\mathcal C}}$ is not pointed neither. Note that ${{\mathcal C}}_{ad}$ is Grothendieck equivalent to $\tilde {{\mathcal C}}_{ad}$. If ${{\mathcal C}}_{ad}$ is a proper fusion subcategory, then an inductive argument implies that $\tilde
{{\mathcal C}}_{ad}$ is solvable and therefore so is $\tilde {{\mathcal C}}$, because it is a $U(\tilde
{{\mathcal C}})$-extension of $\tilde {{\mathcal C}}_{ad}$ and the universal grading group $U(\tilde {{\mathcal C}})$ is abelian. Hence we may assume that $\tilde {{\mathcal C}}_{ad} = \tilde {{\mathcal C}}$ (in particular, the same is true for ${{\mathcal C}}$). Since ${{\mathcal C}}$ is solvable, then its Frobenius-Perron dimension is an integer and therefore ${{\mathcal C}}$ is in fact integral. Then $\tilde {{\mathcal C}}$ is also integral. To show part (ii) we may assume that $\tilde {{\mathcal C}}$ is not solvable, in view of [@witt-wgt Proposition 5.2].
By part (i), $\tilde {{\mathcal C}}_{pt}$ is solvable and not trivial. Note that $\tilde {{\mathcal C}}$ cannot contain any nontrivial non-degenerate fusion subcategory. In fact, if ${{\mathcal C}}$ were non-degenerate, then $\tilde{{\mathcal C}}_{ad} = \tilde
{{\mathcal C}}_{pt}' \subsetneq \tilde {{\mathcal C}}$, against the assumption. If, on the other hand, $\tilde {{\mathcal D}}\subseteq \tilde {{\mathcal C}}$ were a proper non-degenerate subcategory, then $\tilde {{\mathcal C}}\cong \tilde {{\mathcal D}}\boxtimes \tilde {{\mathcal D}}'$, and both $\tilde {{\mathcal D}}$ and $\tilde {{\mathcal D}}'$ are Grothendieck equivalent to fusion subcategories of ${{\mathcal C}}$. An inductive argument implies that $\tilde {{\mathcal D}}$ and $\tilde {{\mathcal D}}'$ are solvable and therefore so is $\tilde {{\mathcal C}}$.
Suppose that $\tilde {{\mathcal C}}$ contains no nontrivial Tannakian subcategory. It follows from [@witt-wgt Lemma 7.1] that $\tilde {{\mathcal C}}_{pt} = \tilde {{\mathcal C}}' \cong
{\operatorname{sVect}}$ and $G[\tilde X] = {\textbf{1}}$, for all simple object $\tilde X$ of $\tilde
{{\mathcal C}}$. This implies that ${\operatorname{FPdim}}{{\mathcal C}}_{pt} = 2$ and $G[X] = {\textbf{1}}$, for all simple object $X$ of ${{\mathcal C}}$.
On the other hand, since ${{\mathcal C}}$ is solvable and ${{\mathcal C}}_{ad} = {{\mathcal C}}$, then ${{\mathcal C}}$ must be a ${{\mathbb Z}}_p$-equivariantization of a fusion category ${{\mathcal D}}$ for some prime number $p$. In particular ${{\mathcal C}}$ contains a (pointed) fusion subcategory of dimension $2$, and therefore $p = 2$. It follows from Lemma \[simple-p\] that ${{\mathcal C}}$ has a simple object $X$ of Frobenius-Perron dimension $2$. In addition, for every such simple object $X$, we have $G[X] = {\textbf{1}}$.
The Nichols-Richmond theorem implies that ${{\mathcal C}}$ contains a fusion subcategory $\overline{{\mathcal C}}$ of type $(1, 2; 2, 1; 3, 2)$ or $(1, 3; 3, 1)$ or $(1, 1; 3, 2; 4,
1; 5, 1)$; see [@NR Theorem 11], [@fusion-lowdim Theorem 3.4]. The first possibility cannot hold in this case, because the unique simple object of dimension $2$ of $\overline{{\mathcal C}}$ is necessarily stable under the action of $G(\overline{{\mathcal C}}) \cong
{{\mathbb Z}}_2$. The second possibility contradicts the assumption that ${\operatorname{FPdim}}{{\mathcal C}}_{pt} =
2$. The third possibility is also discarded because $\overline{{\mathcal C}}$ must be a solvable fusion category, whence $\overline{{\mathcal C}}_{pt} \neq {\operatorname{Vect}}$. This contradiction shows that $\tilde{{\mathcal C}}$ must contain a Tannakian subcategory, and hence (ii) holds.
\[e-prime\] Let ${{\mathcal C}}$ be a braided fusion category. Suppose that ${{\mathcal E}}\subseteq {{\mathcal C}}$ is a Tannakian subcategory. Then ${{\mathcal C}}$ is solvable if and only if ${{\mathcal E}}'$ is solvable.
If ${{\mathcal C}}$ is solvable, then every fusion subcategory of ${{\mathcal C}}$ is solvable. In particular, ${{\mathcal E}}'$ is solvable, showing the ’only if’ direction. Conversely, suppose that ${{\mathcal E}}'$ is solvable. Since ${{\mathcal E}}$ is a Tannakian subcategory, it is symmetric, and therefore ${{\mathcal E}}\subseteq {{\mathcal E}}'$. Then ${{\mathcal E}}$ is solvable. Let $G$ be a finite group such that ${{\mathcal E}}\cong {\operatorname{Rep}}G$ as braided fusion categories. Then the group $G$ is solvable, by [@ENO2 Proposition 4.5 (ii)].
Consider the $G$-crossed braided fusion category ${{\mathcal C}}_G$, so that ${{\mathcal C}}\cong
({{\mathcal C}}_G)^G$ is an equivariantization. Furthermore, the category ${{\mathcal C}}_G$ is a $G$-graded fusion category, and the neutral component ${{\mathcal C}}_G^0$ of this grading satisfies $({{\mathcal C}}_G^0)^G \cong {{\mathcal E}}'$ [@DGNOI Proposition 4.56 (i)]. Therefore ${{\mathcal C}}_G^0$ is solvable. Since $G$ is solvable, then so is ${{\mathcal C}}_G$ and also ${{\mathcal C}}\cong ({{\mathcal C}}_G)^G$. This proves the ’if’ direction and finishes the proof of the proposition.
Let $\tilde {{\mathcal C}}$ be a braided fusion category. Suppose that $\tilde {{\mathcal C}}$ is Grothendieck equivalent to a solvable braided fusion category ${{\mathcal C}}$ and $\tilde {{\mathcal C}}$ is not solvable. Assume in addition that ${\operatorname{FPdim}}\tilde {{\mathcal C}}$ is minimum with respect to these properties.
Then $\tilde {{\mathcal C}}$ must satisfy the following conditions:
\(i) $\tilde {{\mathcal C}}_{ad} = \tilde {{\mathcal C}}$.
\(ii) $\tilde {{\mathcal C}}_{pt} \neq {\operatorname{Vect}}$ is a solvable fusion subcategory and $(\tilde
{{\mathcal C}}_{pt})' = \tilde {{\mathcal C}}$.
\(iii) $\tilde {{\mathcal C}}$ contains a nontrivial Tannakian subcategory and for every Tannakian subcategory $\tilde {{\mathcal E}}$, $\tilde {{\mathcal E}}' = \tilde {{\mathcal C}}$.
\(iv) $\tilde {{\mathcal C}}$ contains no proper non-degenerate fusion subcategories.
Indeed, (i) and (iv) can be shown as in the proof Theorem \[e-tann\], (ii) follows from (i) and Lemma \[cent-cpt\], and (iii) follow from Theorem \[e-tann\] and Proposition \[e-prime\].
The character table of a spherical fusion category {#s-char-tbl}
==================================================
Spherical fusion categories
---------------------------
A *spherical structure* on a fusion category ${{\mathcal C}}$ is a natural isomorphism of tensor functors $\psi: {\operatorname{id}}_{{\mathcal C}}\to (\; )^{**}$ such that $$d_+(X) = d_-(X),$$ for all objects $X$ of ${{\mathcal C}}$, where $d_\pm(X) = {\operatorname{Tr}}_\pm({\operatorname{id}}_X)$, and for every endomorphism $f: X \to X$, ${\operatorname{Tr}}_\pm(f) \in k$ are defined as the compositions $${\operatorname{Tr}}_+(f): {\textbf{1}}\longrightarrow X \otimes X^* \overset{\psi_Xf \otimes {\operatorname{id}}}\longrightarrow X^{**} \otimes X \longrightarrow {\textbf{1}},$$ $${\operatorname{Tr}}_-(f): {\textbf{1}}\longrightarrow X^* \otimes X^{**} \overset{{\operatorname{id}}\otimes f\psi_X^{-1} }\longrightarrow X^{*} \otimes X \longrightarrow {\textbf{1}}.$$ Let ${{\mathcal C}}$ be a spherical fusion category, that is, a fusion category endowed with a spherical structure. The quantum dimension of $X \in {{\mathcal C}}$ is denoted by $d_X : = d_+(X) = d_-(X)$, and the quantum dimension of ${{\mathcal C}}$ is defined in the form $\dim {{\mathcal C}}= \sum_{X \in {\operatorname{Irr}}({{\mathcal C}})} d_X^2$. The *quantum trace* of an endomorphism $f: X \to X$ is denoted by ${\operatorname{Tr}}(f) ={\operatorname{Tr}}_+(f) = {\operatorname{Tr}}_-(f)$. See [@DGNOI Subsection 2.4.3], [@ENO Subsection 2.2].
Recall that a fusion category is called *pseudo-unitary* if its global dimension coincides with its Frobenius-Perron dimension. By [@ENO Proposition 8.24], every weakly integral fusion category is pseudo-unitary. It is shown in [@ENO Proposition 8.23] that every pseudo-unitary fusion category admits a canonical spherical structure with respect to which quantum dimensions of objects coincide with their Frobenius-Perron dimensions.
Modular categories and $S$-matrices
-----------------------------------
A *premodular* category is a braided fusion category equipped with a spherical structure. Equivalently, ${{\mathcal C}}$ is a braided fusion category endowed with a *ribbon structure*, that is, a natural automorphism $\theta: {\operatorname{id}}_{{\mathcal C}}\to {\operatorname{id}}_{{\mathcal C}}$ satisfying $$\label{bal}\theta_{X \otimes Y} = (\theta_X \otimes \theta_Y) c_{Y, X}c_{X, Y}, \quad \theta_X^* = \theta_{X^*},$$ for all objects $X, Y$ of ${{\mathcal C}}$ [@bruguieres], [@DGNOI Subsection 2.8.2].
Let ${{\mathcal C}}$ be a premodular category. The *central charge* of ${{\mathcal C}}$ is the ratio $$\xi({{\mathcal C}}) = \frac{\tau^+({{\mathcal C}})}{\sqrt{\dim {{\mathcal C}}}},$$ where $\sqrt{\dim {{\mathcal C}}}$ is the positive square root and $\tau^+({{\mathcal C}}) = \sum_{X \in {\operatorname{Irr}}({{\mathcal C}})}\theta_Xd_X^2$. See [@DGNOI Subsection 6.2].
The $S$-matrix of ${{\mathcal C}}$ is defined in the form $S = (S_{XY})_{X, Y \in {\operatorname{Irr}}({{\mathcal C}})}$, where for all $X, Y \in {\operatorname{Irr}}({{\mathcal C}})$, $$S_{X, Y} = {\operatorname{Tr}}(c_{Y, X}c_{X, Y}) \in k$$ is the quantum trace of the squared braiding $c_{Y, X}c_{X, Y}:X \otimes Y \to Y \otimes X$.
A premodular category ${{\mathcal C}}$ is called *modular* if the $S$-matrix is non-degenerate [@turaev-b] or, equivalently, if it is non-degenerate [@DGNOI Proposition 3.7].
If ${{\mathcal C}}$ is a spherical fusion category, then its Drinfeld center ${{\mathcal Z}}({{\mathcal C}})$ is a modular category of global dimension $\dim {{\mathcal Z}}({{\mathcal C}}) = (\dim {{\mathcal C}})^2$ and central charge $\xi({{\mathcal Z}}({{\mathcal C}})) = 1$ [@mueger-ii], [@DGNOI Example 6.9].
Suppose that ${{\mathcal C}}$ is a modular category. Then for every $X, Y, Z \in {\operatorname{Irr}}({{\mathcal C}})$, the multiplicity $N_{XY}^Z$ of $Z$ in the tensor product $X \otimes Y$ is given by the *Verlinde formula*: $$\label{verlinde}N_{XY}^Z = \frac{1}{\dim {{\mathcal C}}}\sum_{T \in {\operatorname{Irr}}({{\mathcal C}})} \frac{S_{XT} \, S_{YT} \, S_{Z^*T}}{d_T},$$ where $d_T$ denotes the quantum dimension of the object $T$ and $\dim {{\mathcal C}}$ is the quantum dimension of ${{\mathcal C}}$. See [@BK Theorem 3.1.14].
$S$-equivalence of spherical fusion categories {#s-equiv}
----------------------------------------------
Let ${{\mathcal C}}$ and ${{\mathcal D}}$ be spherical fusion categories. We shall say that ${{\mathcal C}}$ and ${{\mathcal D}}$ are *$S$-equivalent* if there exists a bijection $f: {\operatorname{Irr}}({{\mathcal Z}}({{\mathcal C}})) \to {\operatorname{Irr}}({{\mathcal Z}}({{\mathcal D}}))$ such that $f({\textbf{1}}) = {\textbf{1}}$ and $S_{f(X), f(Y)} = S_{X, Y}$, for all $X, Y \in {\operatorname{Irr}}({{\mathcal C}})$.
The following lemma summarizes some of the main properties of $S$-equivalence.
\[s-eq\] Let ${{\mathcal C}}$ and ${{\mathcal D}}$ be spherical fusion categories and suppose that $f:{\operatorname{Irr}}({{\mathcal Z}}({{\mathcal C}}))\to {\operatorname{Irr}}({{\mathcal Z}}({{\mathcal D}}))$ is an $S$-equivalence. Then the following hold:
1. $d_{f(X)} = d_X$, for all $X \in {\operatorname{Irr}}({{\mathcal C}})$. In particular, $\dim {{\mathcal Z}}({{\mathcal C}}) = \dim {{\mathcal Z}}({{\mathcal D}})$.
2. $f:{\operatorname{Irr}}({{\mathcal Z}}({{\mathcal C}}))\to {\operatorname{Irr}}({{\mathcal Z}}({{\mathcal D}}))$ is a Grothendieck equivalence.
3. For every fusion subcategory ${{\mathcal E}}$ of ${{\mathcal Z}}({{\mathcal C}})$ we have $f({{\mathcal E}}') = f({{\mathcal E}})'$. In particular, $f({{\mathcal E}})$ is symmetric (respectively, non-degenerate) if and only if so is ${{\mathcal E}}$.
4. For every fusion subcategory ${{\mathcal E}}$ of ${{\mathcal Z}}({{\mathcal C}})$, $f$ maps the projective centralizer of ${{\mathcal E}}$ to the projective centralizer of $f({{\mathcal E}})$.
For every simple object $X$ of ${{\mathcal Z}}({{\mathcal C}})$ we have $d_X = d_{\textbf{1}}d_X = S_{{\textbf{1}}, X} = S_{{\textbf{1}}, f(X)} = d_{\textbf{1}}d_{f(X)} = d_{f(X)}$, and we get (i). Now part (ii) follows from (i) and the Verlinde formula . Part (iii) follows from the fact that two simple objects $X$ and $Y$ centralize each other if and only if $S_{X, Y} = d_Xd_Y$.
We now show part (iv). Let $X$ and $Y$ be simple objects of ${{\mathcal C}}$. It follows from [@DGNOI Proposition 3.22] that $X$ belongs to the projective centralizer of $Y$ if and only if $X$ belongs to the centralizer of $Y \otimes Y^*$. In view of part (iii) this happens if and only if $f(X)$ centralizes $f(Z)$, for all $Z \in {\operatorname{Irr}}({{\mathcal C}})$ such that $N^Z_{Y\otimes Y^*} \neq 0$. Since, by (ii), $f$ is a Grothendieck equivalence, then $f(Y)^* = f(Y^*)$ (Proposition \[groth-eq\] (iii)), and it follows that the last condition is equivalent to the condition that $f(X)$ centralizes $f(Y)\otimes f(Y)^*$, that is, $f(X)$ belongs to the projective centralizer of $f(Y)$.
\[s-gpttic\] Let ${{\mathcal C}}$ and ${{\mathcal D}}$ be $S$-equivalent spherical fusion categories. Then ${{\mathcal C}}$ is group-theoretical if and only if so is ${{\mathcal D}}$.
We have that ${{\mathcal C}}$ is group-theoretical if and only if ${{\mathcal Z}}({{\mathcal C}})$ is group-theoretical. Suppose that this is the case. In particular ${{\mathcal Z}}({{\mathcal C}})$, and therefore also ${{\mathcal Z}}({{\mathcal D}})$, are integral. Since ${{\mathcal Z}}({{\mathcal C}})$ is a modular category, [@dgno-gpttic Corollary 4.14] implies that it contains a symmetric subcategory ${{\mathcal E}}$ such that ${{\mathcal E}}'_{ad} \subseteq {{\mathcal E}}$. Since every $S$-equivalence preserves centralizers, symmetric subategories and is a Grothendieck equivalence between ${{\mathcal Z}}({{\mathcal C}})$ and ${{\mathcal Z}}({{\mathcal D}})$, this implies that $f({{\mathcal E}})$ is a symmetric subcategory of ${{\mathcal Z}}({{\mathcal D}})$ and $f({{\mathcal E}})'_{ad} = f({{\mathcal E}}'_{ad}) \subseteq f({{\mathcal E}})$ (see Proposition \[groth-eq\] (iv)). Hence ${{\mathcal Z}}({{\mathcal D}})$ and therefore also ${{\mathcal D}}$ are group-theoretical. This implies the theorem.
\[s-eq-groups\] Let $G$ and $\Gamma$ be finite groups and let $\omega: G \times G \times G \to k^*$, $\omega': \Gamma \times \Gamma \times \Gamma \to k^*$ be $3$-cocycles on $G$ and $\Gamma$, respectively. Suppose that the categories ${{\mathcal C}}(G, \omega)$ and ${{\mathcal C}}(\Gamma, \omega')$ are $S$-equivalent. Then $G$ is solvable if and only if so is $\Gamma$.
It is enough to show the ’if’ direction. Thus, let us assume that $G$ is solvable. Let $f: {\operatorname{Irr}}({{\mathcal Z}}({{\mathcal C}}(\Gamma, \omega'))) \to {\operatorname{Irr}}({{\mathcal Z}}({{\mathcal C}}(G, \omega)))$ be an $S$-equivalence. The center of the category ${{\mathcal C}}(\Gamma, \omega')$ contains a Tannakian subcategory ${{\mathcal E}}$ equivalent to ${\operatorname{Rep}}\Gamma$ as braided fusion categories. In view of Lemma \[s-eq\], $f({{\mathcal E}})$ is a symmetric fusion subcategory of ${{\mathcal Z}}({{\mathcal C}}(G, \omega))$ which is Grothendieck equivalent to ${\operatorname{Rep}}\Gamma$. Being symmetric, $f({{\mathcal E}})$ is equivalent as a fusion category to the category ${\operatorname{Rep}}F$ for some finite group $F$. Then $F$ is solvable because ${{\mathcal Z}}({{\mathcal C}}(G, \omega))$ is solvable, by [@ENO2 Proposition 4.5]. Since the categories ${\operatorname{Rep}}\Gamma$ and ${\operatorname{Rep}}F$ are Grothendieck equivalent, then the groups $\Gamma$ and $F$ have the same character table. This implies that $\Gamma$ is solvable. Hence ${{\mathcal C}}(\Gamma, \omega')$ is solvable, as claimed.
\[s-equiv-gt\] Let ${{\mathcal C}}$ and ${{\mathcal D}}$ be $S$-equivalent spherical fusion categories and suppose that ${{\mathcal C}}$ is group-theoretical. Then ${{\mathcal C}}$ is solvable if and only if ${{\mathcal D}}$ is solvable.
Since ${{\mathcal C}}$ is group-theoretical, ${{\mathcal Z}}({{\mathcal C}})$ is equivalent to the center of a pointed fusion category ${{\mathcal Z}}({{\mathcal C}}(G, \omega))$, for some finite group $G$ and $3$-cocycle $\omega$ on $G$. Hence ${{\mathcal Z}}({{\mathcal C}})$ contains a Tannakian subcategory ${{\mathcal E}}$ equivalent to ${\operatorname{Rep}}G$ as braided fusion categories, such that $(\dim {{\mathcal E}})^2 = \dim {{\mathcal Z}}({{\mathcal C}})$.
Being Grothendieck equivalent to ${{\mathcal Z}}({{\mathcal C}})$, ${{\mathcal Z}}({{\mathcal D}})$ is also group-theoretical, in view of Theorem \[s-gpttic\]. Thus ${{\mathcal Z}}({{\mathcal C}})$ is an integral modular category of dimension $(\dim {{\mathcal D}})^2$ and central charge $1$. Note in addition that if $f$ is an $S$-equivalence, then $f({{\mathcal E}})$ is a symmetric subcategory of ${{\mathcal Z}}({{\mathcal D}})$ such that $\dim {{\mathcal Z}}({{\mathcal D}}) = (\dim f({{\mathcal E}}))^2$. Theorem 4.8 of [@dgno-gpttic] implies that ${{\mathcal Z}}({{\mathcal D}})$ is equivalent to the center of a pointed fusion category, that is, ${{\mathcal Z}}({{\mathcal D}}) \cong {{\mathcal Z}}({{\mathcal C}}(\Gamma, \omega'))$, for some finite group $\Gamma$ and $3$-cocycle $\omega'$ on $\Gamma$. Then the theorem follows from Lemma \[s-eq-groups\].
[AAAA]{}
B. Bakalov, A. Kirillov Jr., *Lectures on tensor categories and modular functors*, Amer. Math. Soc., 2001.
A. Brugui\` eres, *Cat' egories pr' emodulaires, modularisations et invariants des vari' et' es de dimension 3*, Math. Ann. **316**, 215–36 (2000).
A. Bruguières and S. Natale, *Exact sequences of tensor categories*, Int. Math. Res. Not. **2011** (24), 5644–5705 (2011).
A. Bruguières and S. Natale, *Central exact sequences of tensor categories, equivariantization and applications*, J. Math. Soc. Japan **66**, 257–287 (2014).
S. Burciu and S. Natale, *Fusion rules of equivariantizations of fusion categories*, J. Math. Phys. **54** `DOI: 10.1063/1.4774293` (2013).
P. Deligne, *Cat' egories tensorielles*, Mosc. Math. J. **2**, 227–248 (2002).
R. Dijkgraaf, V. Pasquier and P. Roche, *Quasi-quantum groups related to orbifold models* In: Proc. Modern Quantum Field Theory, Tata Institute, Bombay, 375–383, 1990.
J. Dong, S. Natale, L. Vendramin, *Frobenius property for fusion categories of small integral dimension*, J. Algebra Appl. **14**, 1550011 (2015).
V. Drinfeld, S. Gelaki, D. Nikshych and V. Ostrik, *On braided fusion categories I*, Sel. Math. New Ser. **16**, 1–119 (2010).
V. Drinfeld, S. Gelaki, D. Nikshych and V. Ostrik, *Group-theoretical properties of nilpotent modular categories*, preprint `arXiv:0704.0195` (2007).
P. Etingof, D. Nikshych and V. Ostrik, *On fusion categories*, Ann. Math. (2) **162**, 581–642 (2005).
P. Etingof, D. Nikshych and V. Ostrik, *Weakly group-theoretical and solvable fusion categories*, Adv. Math. **226**, 176–205 (2011).
A. Ganchev, *Fusion rings and tensor categories*, In: Noncommutative structures in mathematics and physics, Kluwer Academic Publishers, NATO Sci. Ser. II, Math. Phys. Chem. **22**, 295-298 (2001).
S. Gelaki and D. Nikshych, *Nilpotent fusion categories*, Adv. Math. **217**, 1053–1071 (2008).
I. M. Isaacs, *Character theory of finite groups*, Academic Press, 1976.
A. Jedwab and S. Montgomery, *Representations of some Hopf algebras associated to the symmetric group ${\mathbb{S}}_n$*, Algebr. Represent. Theory **12**, 1–17 (2009).
A. Masuoka, *Extensions of Hopf algebras*, Trab. Mat. **41/99**, Fa.M.A.F. (1999).
S. Mattarei, *Character tables and metabelian groups*, J. London Math. Soc. **46**, 92–100, (1992).
S. Mattarei, *An example of $p$-groups with identical character tables and different derived lengths*, Arch. Math. **62**, 12–20 (1994).
S. Montgomery, S. J. Witherspoon, *Irreducible representations of crossed products*, J. Pure Appl. Algebra **129**, 315–326 (1998).
M. M" uger, *From subfactors to categories and topology. II. The quantum double of tensor categories and subfactors*, J. Pure Appl. Algebra **180**, 159–219 (2003).
M. M" uger, *Galois extensions of braided tensor categories and braided crossed G-categories*, J. Algebra **277**, 256–281 (2004).
, [D. Nikshych]{} and [S. Witherspoon]{}, *Fusion subcategories of representation categories of twisted quantum doubles of finite groups*, Int. Math. Res. Not. **2009** (22), 4183–4219 (2009).
D. Naidu and E. Rowell, *A finiteness property for braided fusion categories*, Algebr. Represent. Theory **14**, 837–855 (2011).
S. Natale, *On group-theoretical Hopf algebras and exact factorizations of finite groups*, J. Algebra **270**, 199–211 (2003).
S. Natale, *Semisolvability of semisimple Hopf algebras of low dimension*, Memoirs Amer. Math. Soc. **186** (2007).
S. Natale, *Hopf algebra extensions of group algebras and Tambara-Yamagami categories*, Algebr. Represent. Theory **13**, 673–691 (2010).
S. Natale, *On quasitriangular structures in Hopf algebras arising from exact group factorizations*, Commun. Algebra **39**, 4763–4775 (2011).
S. Natale, *On weakly group-theoretical non-degenerate braided fusion categories*, J. Noncommut. Geom. **8**, 1043–1060 (2014).
W. D. Nichols and M. B. Richmond, *The Grothendieck group of a Hopf algebra*, J. Pure Appl. Algebra **106**, 297–306 (1996)
J. Siehler, *Near-group categories*, Alg. Geom. Topol. **3**, 719–775 (2003).
V. Turaev, *Quantum invariants of knots and 3-manifolds*, de Gruyter Studies in Math. **18**, Berlin, 1994.
[^1]: This work was partially supported by CONICET and SeCYT–UNC
[^2]: This can be seen, for instance, as a consequence of the Nichols-Richmond theorem [@NR Theorem 11]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a multi-wavelength study of the young stellar population in the Cygnus-X DR15 region. We studied young stars forming or recently formed at and around the tip of a prominent molecular pillar and an infrared dark cloud. Using a combination of ground based near-infrared, space based infrared and X-ray data, we constructed a point source catalog from which we identified 226 young stellar sources, which we classified into evolutionary classes. We studied their spatial distribution across the molecular gas structures and identified several groups possibly belonging to distinct young star clusters. We obtained samples of these groups and constructed K-band luminosity functions that we compared with those of artificial clusters, allowing us to make first order estimates of the mean ages and age spreads of the groups. We used a $^{13}$CO(1-0) map to investigate the gas kinematics at the prominent gaseous envelope of the central cluster in DR15, and we infer that the removal of this envelope is relatively slow compared to other cluster regions, in which gas dispersal timescale could be similar or shorter than the circumstellar disk dissipation timescale. The presence of other groups with slightly older ages, associated with much less prominent gaseous structures may imply that the evolution of young clusters in this part of the complex proceeds in periods that last 3 to 5 Myr, perhaps after a slow dissipation of their dense molecular cloud birthplaces.'
author:
- 'S. Rivera-Gálvez, C. G. Román-Zúñiga, E. Jiménez-Bailón, J. E. Ybarra, J. F. Alves and Elizabeth A. Lada'
title: 'The Young Stellar Population of the Cygnus-X DR15 Region'
---
=1
Introduction \[intro\]
======================
The star forming complex of Cygnus-X region is one of the most prominent features in our Galaxy. Originally detected as an extended region with a thermal spectrum, @Piddington:1952aa named the region Cygnus-X, in order to distinguish it from the other nearby known radio source, Cygnus-A. Cygnus-X is composed of several OB associations, dozens of embedded stellar clusters, hundreds of HII regions and over 40 known massive protostars [see @Reipurth:2008aa for an extensive review]. The young star population in Cygnus-X is currently interacting with one of the most massive molecular cloud complexes in the Galaxy, with a total mass of $3\times 10^6\ \mathrm{M}_\odot$ [@Schneider:2006aa], as well as the X-ray emitting Cygnus Superbubble. It has been proposed that Cygnus-X could be the precursor of a globular cluster [@Knodlseder:2000aa].
To study the interaction between recently formed star clusters and their surrounding medium in Cygnus-X could provide very important clues about the present evolution of the complex. It is particularly important to focus on the numerous embedded cluster populations and to compare the properties of clusters across the region. For instance: how embedded clusters proceed from formation to emergence from their parental gas clumps in such a strong ionizing medium? Is cluster formation or evolution in Cygnus-X determined by the local environment? What are the time scales from formation to gas dispersal? Also, what happens after cluster emergence: are subsidiary clusters destined to disperse or could they end up swelling the ranks of the main association? In any case, we should expect that multiwavelength analysis of the young star clusters in Cygnus-X will help to fine tune current ideas about the formation and evolution of embedded stellar clusters or groups.
In this paper we selected to study one of the most prominent embedded cluster populations in Cygnus-X: the region DR15, also listed as cluster 10 in the survey of @Dutra:2001aa. The DR15 region has been related to the HII region G79.306+0.282, the source IRAS 20306+4005/FIR-1 [@Campbell:1982aa], and sources IRS 1, 2 and 3 in the list of @Kleinmann:1979aa [see also @Odenwald:1990aa]. DR15 is located in the Cygnus-X South region, located at an estimated distance of 1.4 kpc [@Rygl:2012aa]. The cluster contains one prominent far-infrared source (FIR-1), which marks the location of a compact HII region formed by a pair of B type embedded sources [@Odenwald:1990aa; @Oka:2001aa; @Kurtz:1994aa], however the cluster hosts various intermediate to massive stars which create a rather complex structure of photodissociation regions, forming a nebulous envelope which glows brightly in infrared images (see Figure \[fig:RGB\]). DR15 sits at the tip of a long (about 10 pc), filamentary pillar that extends into the southern edge of the central OB from DR12, lying in projection about 1 degree south from the Cyg-X3 star. The structure of the pillar appears to be protruding from the DR12 ridge (see Figure \[fig:spitzer\]). The DR12-15 region in Cygnus-X possibly lies in front of the OB2 association, along our line of sight. A large filamentary infrared dark cloud (IRDC) with active star formation lies to the north and west of the cluster. This IRDC appears to be kinematically independent from DR15, as shown by @Schneider:2006aa analysis. Recently, the western segment of the IRDC, has been shown to host a young stellar cluster in interaction with the Luminous Blue Variable (LBV) source G79.29+0.46 [@Rizzo:2008aa; @Jimenez:2010aa]. Distance to G79.29+0.46 has been also estimated to be 1.4 kpc [@Rizzo:2014aa; @Palau:2014aa].
Our main goal in this study is to track the progression of the star formation in the DR15 cluster and its immediate surroundings, by looking at the properties and distributions of the young stellar population. DR15 is especially interesting for being at a very specific stage at which it is already emerging from its parental cloud (some members and parts of the reflection nebulosities are already detectable in visible wavelengths). Moreover, the interesting case it presents at being apparently formed at the tip of a filamentary structure and next to a much younger star forming spot in the IRDC, is worth of detailed attention.
The paper is organized as follows: in Section \[obs\] we describe the datasets used in this work. Section \[analysis\] is dedicated to describe the procedures we followed to acomplish a description of the history of star formation in the region, a result which we discuss in Section \[discussion\]. Finally, a brief summary of our results can be found in Section \[discussion\].
Observations {#obs}
============
For the first part of our analysis, we combined high-quality deep near-IR observations with images and catalogs from the Spitzer Cygnus-X Legacy Survey [@Hora:2009aa hereafter CXLS]. We also made use of archival data from the Chandra X-Ray Observatory. By combining these datasets we prepared a multi-wavelength photometry catalog for the stellar population in DR15, which we use to reconstruct the history of star formation in the region.
Near Infrared Observations {#obs:nir}
--------------------------
Near-infrared images of the Cygnus-X DR15 region were obtained with the OMEGA 2000 camera at the 3.5m telescope of the Calar Alto Observatory, atop Sierra de los Filabres in Almería, Spain, during the nights of February 2nd and March 3rd, 2010, with excellent weather conditions. The dataset consists of 900 second co-added exposures in the $J$, $H$ and $K$ bands (1.209, 1.648 and 2.208 $\mu$m, respectively). The seeing values –measured directly from the average FWHM of stars in the final reduced mosaics– were 1.17, 1.13 and 0.98$\arcsec$ in $J$, $H$ and $K$, respectively.
The reduction of the images and the extraction of Point Spread Function (PSF) photometry lists for all bands were performed with custom `IRAF` pipelines, and the `SExtractor` algorithm [@Bertin:1996aa], following a methodology equivalent to the one described in @Roman:2010aa.
We constructed a near-IR catalog by matching individual filter photometry lists with the aid of [TOPCAT]{} [@Taylor:2005uq]. We replaced saturated sources and a small number of missing detections in the Calar Alto images with values from the 2MASS Point Source Catalog, obtained at the Infrared Processing and Analysis Center (IPAC, Caltech). The 10 percent photometric depth values –indicated by the average brightness at which photometric error reaches a value of 0.1 mag– are 21.5, 20.0 and 18.75 mag in $J$, $H$ and $K$, respectively, and are good indicators of the completeness of the data. These limits are enough to sample the young star population in DR15 down to 0.09 M$_\odot$ in regions of low to moderate extinction ($A_V<25.0$ mag).
{width="6.0in"}
Spitzer Cygnus-X Legacy Survey data {#obs:mir}
------------------------------------
The *Spitzer Space Telescope* has observed the DR15 cluster with the IRAC and MIPS detectors as part of the CXLS. We obtained archival enhanced product mosaics from the Spitzer Heritage Archive as well as a photometric catalog coincident with our region of interest directly from the CXLS Data Release 1 (DR1). The catalog contains calibrated magnitudes for sources detected with IRAC in its four cryogenic mission channels (3.6, 4.5, 5.8 and 8.0 $\mu$m), as well as in the 24 $\mu$m channel of MIPS. Using `TOPCAT` and IDL routines we combined our near-IR and the CXLS DR1 catalogs into a single infrared photometry list with a total of 46983 sources.
{width="6.0in"}
Chandra ACIS Observations {#obs:xray}
-------------------------
The DR15 cluster was observed with the Imaging Array of the Chandra Advanced CCD Imaging Spectrometer (ACIS-I) on 2011 January 25 (ObsID 12390, P.I. Wright) with net exposure time of 39.875 ks.
We processed the archival raw data using the routines from version 4.5 of the Chandra Interactive Analysis of Observations `CIAO` data analysis system [@Fruscione:2006aa]. We used the `chandra\_repro` reprocessing script to recalibrate our event data in order to ensure that consistent calibration updates were applied to the dataset.
The X-ray image processing was performed as follows: first, we created a exposure-corrected image from our data in the broad band (0.5 to 7 keV with an effective energy of 2.3 keV) using `CIAO/fluximage` routine with a bin size of 1. Second, we applied the `wavedetect` tool to our broad band image in order to identify potential X-ray sources. We used wavelet scales from 1 to 16.5 pixels in s.pdf of $\sqrt{2}$ and a source significance threshold of $1\times10^{-6}$.
We performed photometry on the list provided by `wavedetect` using the the `ACIS Extract` (AE) package [@Broos:2010aa]. AE permits an optimal determination of the local background and the best flux extraction apertures based on the PSF of the image. Photometry is then extracted on selected energy bands and a list of source properties, statistics and best fit spectral models is produced. We used three energy bands: Soft, from 0.2 to 2.0 keV, Medium, from 2.0 to 4.0 keV and Hard, from 4.0 to 7.0 keV. From the final list of sources produced by `ACIS Extract` we rejected those with a probability of 1$\%$ or higher of being a background fluctuation $P_B>0.01$). Our final lists contains a total of 131 X-ray sources. From these, a total of 109 (83.2%) sources have a counterpart in our IR catalog. In Figure \[fig:chandra\] we show a false color (RGB) map of the ACIS field for DR15 using images from the three energy bands, overlaid with contours of visual extinction from the NICEST extinction map ww constructed from near-IR data. In there, we see how sources with hard X-ray emission are preferently located in regions of high column density. This is because soft X-ray bands are prone to oscuration by dust. We see how most of the X-ray sources are associated with the molecular cloud and the embedded population, confirming that most of the point source X-ray emission in DR15 comes from young stars.
{width="7.0in"}
FCRAO Observations {#obs:others}
-------------------
We made use of the Five College Radio Observatory (FCRAO) $^{13}$CO(1-0) molecular radio emission map of the South Cygnus-X region from the study of @Schneider:2011aa. The map is a RA-Dec-radial velocity cube, from which we extracted molecular gas properties using standard tools from the `MIRIAD` [@Sault:1995aa] package.
Analysis
=========
We limited our region of study to the area covered with OMEGA 2000, defined as $\mathrm{[RA,\ Dec]}=[307.969040,40.150383]\rightarrow[308.305620,40.405865]$. The analysis described below correspond to sources falling within that area only.
Identification of YSOs {#analysis:ysoid}
-----------------------
We identified Young Stellar Objects (YSOs) in the DR15 region, by applying color and brightness criteria to our multiwavelength catalog. For sources that were detected in the Spitzer IRAC and MIPS 24$\mu$m bands, we classified Class I/0 (embedded protostars) and Class II (Classic T Tauri) sources using the criteria by @Ybarra:2013aa, we also required that these sources had photometric uncertainty values less than 0.25 mag in each band.
Our color criteria @Ybarra:2013aa are essencially based in the color criteria of @Gutermuth:2008aa and @Kryukova:2012aa, but as explained by @Ybarra:2013aa we add an additional \[5.8\]-\[8.0\] criteria for objects that do not have a MIPS 24 $\mu m$ detection. In addition, our use of $JHK$ photometry allows us to identify additional less bright candidates and with the use of X-ray data we are able to identify Class III sources that do not have an infrared excess.
Class III candidate sources were selected from a list of sources in the Chandra catalog that match in position with an infrared point source, after removing those corresponding to Class I/0 and Class II candidates. We added the requirement that $$J-H\ge0$$, because sources with $J-H<0$ are most likely either galaxies or spurious sources associated with diffraction spikes of bright stars. We further depurated this first counterpart list by keeping only those sources that had no evidence of a prominent circumstellar disk (i.e., we only selected stars with mostly photospheric emission). For this last criteria we used the $J-H$ vs $K-[4.5]$ color space:
$$J-H>1.97(K-[4.5]).$$
and for those sources that do not have a detection in \[4.5\] we used:
$$J-H>1.74(H-K).$$
The remaining list of Chandra-NIR counterparts are sources with evidence of a disk that were not previously selected as Class 0/I or II sources. We used the additional near-IR criteria of @Gutermuth:2008aa to select a few more young sources from this group. If a Chandra counterpart had photometry in the first three IRAC bands, and photometric errors less than 0.1 mag in at least \[3.6\] and \[4.5\], then it was classified as a Class 0/I candidate if:
$$[4.5]-[5.8]>0.5 \mathrm{\ and\ } [3.6]-[4.5]>0.7,$$
while for Class II candidate identification, we used:
$$0.2<[3.6]-[4.5]<0.7 \mathrm{\ and\ } 0.5<[4.5]-[5.0]<1.5.$$
A total of 20 sources could not be classified with these criteria if they did lack a detection in one or more bands, but if they fall to the right of the reddening band in the $J-H$ vs $K-[4.5]$ color space, they could be bonafide young sources. For 13 of these sources we determined their class (I or II) by inspecting their spectral energy distrubutions (SED), which we constructed using the SED fitting tool of @Robitaille:2007aa. Four additional sources were identified as possible AGN galaxies. The 3 remaining sources did not have enough IR bands to permit a clear classification and were discarded from the list.
In total, we identified in our selected region a total of 226 YSOs, distributed as follows: 26 Class I/0 candidates, 155 Class II candidates and 45 Class III candidates (11, 69 and 20 percent of the total, respectively). We list all identified YSOs in the tables of the Appendix \[app:ysos\].
We matched our Class I, Class II and Class III catalogs against the catalog of @Kryukova:2014aa, out of 23 sources coinciding with our region of study, 17 were also identified as YSO candidates in this study. The 6 remaining sources lack emission is the \[4.5\] band and could not be confirmed as YSOs using our criteria.
In Figure \[fig:diagrams\] we show the distribution of the classified YSOs in two different color-color diagrams. In the $J-H$ vs. $K-[4.5]$ diagram we can see that the most of Class I sources lie to the right of the reddening band indicating the presence of intense excess emission at IR wavelegths due to the stellar radiation in the dusty material of their envelopes or circumstellar disks. Class II sources also present infrared excess emission, although in a lesser way due the dust clearing within their inner disks. Then, the Class III sources lie within the reddening band or along the dwarf main sequence and lack significant infrared excess. The three groups of sources appear well separated in the \[3.6\]-\[4.5\] vs. \[4.5\]-\[8.0\] two-color diagram.
![Color-color diagrams for YSO sources in the DR15 region, Class I sources are marked with a red dot symbol; Class II sources are marked with a green dot symbol; Class III sources are indicated with a blue dot symbol. Sources marked with gray dot symbols are sources in the field without a YSO designation and photometric errors less than 0.1 mag in all bands. a) $J-H$ vs. $K-[4.5]$ two-color diagram. The solid curve is a model for dwarf main sequence population from the Dartmouth [@Dotter:2008aa] grid; the solid line at its right side is the Classic T-Tauri star locus [@Meyer:1997aa] adapted to this color space as in @Teixeira:2012aa. b) \[3.6\]-\[4.5\] vs. \[4.5\]-\[8.0\] two-color diagram, showing a clear separation between evolutionary classes.[]{data-label="fig:diagrams"}](f4a.pdf "fig:"){width="3.0in"} ![Color-color diagrams for YSO sources in the DR15 region, Class I sources are marked with a red dot symbol; Class II sources are marked with a green dot symbol; Class III sources are indicated with a blue dot symbol. Sources marked with gray dot symbols are sources in the field without a YSO designation and photometric errors less than 0.1 mag in all bands. a) $J-H$ vs. $K-[4.5]$ two-color diagram. The solid curve is a model for dwarf main sequence population from the Dartmouth [@Dotter:2008aa] grid; the solid line at its right side is the Classic T-Tauri star locus [@Meyer:1997aa] adapted to this color space as in @Teixeira:2012aa. b) \[3.6\]-\[4.5\] vs. \[4.5\]-\[8.0\] two-color diagram, showing a clear separation between evolutionary classes.[]{data-label="fig:diagrams"}](f4b.pdf "fig:"){width="3.0in"}
Dust Extinction and the Spatial Distribution of YSOs \[analysis:extmap\]
------------------------------------------------------------------------
In Figure \[fig:ysodist\] we plot the spatial distribution of the YSOs on a dust extinction map. This dust extinction (A$_V$) map for the DR15 region was constructed with the near-IR catalog an optimized version of the near-infrared excess (NICER) algorithm of @Lombardi:2001aa, which estimates extinction using the dust-reddened colors of stars in the background of the cloud. In order to construct the map, we removed all sources with colors indicative of intrinsic infrared excess, which would bias the estimated extinction towards higher values. We used a sigma-clipping scheme to remove outlier values from the final weighted averages at each position. The map was constructed using Nyquist sampling on a equally spaced equatorial grid, smoothing the individual extinction estimates of background stars with a Gaussian filter of 30$\arcsec$ FWHM (this implies a resolution near 0.2 pc at a distance of 1.4 kpc). In Figure \[fig:ysodist\] we see how the map clearly resolves the filamentary structure of the IRDC in the northern part of the field and the morphology of the dust pillar on which DR15 is located. The extinction contours are limited to a maximum level of $A_V$=30 mag and further smoothed on the figure with a factor 3 boxcar, in order to remove some spurious features at the IRDC region. The resultant map shown in Figure \[fig:ysodist\] is in good agreement with a column density map constructed from 250 to 500 micron dust emission images from Herschel/SPIRE[^1], which also allows a much larger dynamic range (up to $A_V\approx$150 mag; Schneider et al., private comm.)
Clearly, the regions with higher column density values are those hosting a majority of the youngest stars. Using our extinction map we determined how the Class I sources are distributed in the highest density regions. Using contours of constant extinction in s.pdf of 1.0 mag, we counted the number of Class I sources above each level, and found that 23 out of 24 sources in our region of study lie above A$_V$=13.0 mag, and 20 out of 24 sources lie above $A_V$=15.0 mag. The fraction of the total of Class I sources, $N(>A_V)/N_{total}$, decreases steeply after that, with only 50 percent of the total number of sources remaining at levels above $A_V$=22.0 mag. It is also important to notice that given the filamentary morphology of the IRDC, the projected area of the map contained above each level decreases very steeply for $A_V>12.0$ mag. We also found that the surface density of Class I sources, $\Sigma_*(>A_V)$, defined as the number of sources divided by the area above a given level, deviates little from a power-law behavior with a slope $\beta \approx 2.9$ in the range $12<A_V<40.0$ mag. All of this is very consistent with a Schmidt type relation like it was found for a set of nearby Giant Molecular Clouds by @Lada:2013aa, except that in those clouds the linear regime appears to be set at a lower extinction interval. Unfortunately, our numbers are too small to attempt a Bayesian analysis like that of @Lada:2013aa, and possibly a comparison with other regions containing IRDCs would be more fair. However, the fact that we find a possible Schmidt-like behavior in a region like Cygnus-X DR15 is very interesting and worth of a dedicated comparative analysis with similar regions, which is beyond the scope of this paper.
Properties of the youngest stars in the DR15 region \[analysis:SED\]
--------------------------------------------------------------------
Using our master infrared catalogs we were able to construct spectral energy distrubutions (SED) for 24 of the Class 0/I sources we identified using the methods previously described. In most cases these sources belong to the cluster population associated with the dark infrared cloud in the northern section of our region of study. In a few cases we were able to complement these SEDs with Herschel Space Telescope PACS mid/far infrared photometry from the catalog of @Ragan:2012aa and 850 $\mu$m photometry from the catalog of @Difrancesco:2008aa. Using the SED fitting tool of @Robitaille:2007aa, we estimated some basic properties for the sources. In Figure \[fig:C1disks\] we show histograms of star masses and the disk accretion rates. The mass distribution suggests that the sources we are able to detect in DR15 are mostly intermediate to massive (mostly solar type and above). The accretion rates are in good agreement with estimates of typical accretion disk for sources with masses above solar [@Fang:2013aa].
![Left: Distribution of Class 0/I candidate sources mass estimates from SED model fits. Right: Distribution of Class 0/I candidate sources disk accretion rate estimates from SED model fits.[]{data-label="fig:C1disks"}](f6a.pdf "fig:"){width="2.5in"} ![Left: Distribution of Class 0/I candidate sources mass estimates from SED model fits. Right: Distribution of Class 0/I candidate sources disk accretion rate estimates from SED model fits.[]{data-label="fig:C1disks"}](f6b.pdf "fig:"){width="2.5in"}
The Star Formation History of DR15 \[analysis:sfh\]
---------------------------------------------------
### Identification of the main stellar groups in DR15 {#analysis:sfh:clusterid}
To further investigate the stellar formation history in the DR15 region we attempted to identify the individual stellar clusters present in our field of study. Two clusters are easy to identify, namely the cluster at the center of the field and the cluster embedded in the IRDC at the north. However, the map of Figure \[fig:ysodist\] shows a good number of sources located around the molecular gas filaments, which may or may not be part of previously formed stellar clusters or groups.
In order to identify significant overdensities of stars in the DR15 region, we constructed surface density maps using the `Gather` algorithm of @Gladwin:1999id. This algorithm is based in turn on the *nearest neighbor* method, which assigns individual surface density values to points on a two-dimensional map based on the equivalent circular area defined by the distance to a $n$th. neighbor point (see @Casertano:1985uq for a description of this method in the particular case of stellar cluster identification; some examples of its use for embedded clusters can be found in @Roman-Zuniga:2008aa [@Gutermuth:2009aa] and @Roman:2015aa). The `gather` algorithm is adequate for identifying individual clusters in a relatively simple layout, optimizing the value of $n$ that defines the surface density measurement and the size of the smoothing kernel used to construct the surface density map.
We made individual `gather` maps for each of the candidate YSO class lists. These maps are shown in Figure \[fig:gather\]. The maps show the concentration of Class I sources at the north IRDC region. Class II sources are distributed over the entire region of study, but still concentrate in a few clearly defined groups. This led us to define some populations of stars, which we use as samples for our analysis. We named the two known clusters as DR15-C and DR15-N because their location center and north of our region of study. We identified and named as well three other groups: DR15-W, DR15-SW and DR15-SE. The purpose of this selection is to compare the age and age spread of these populations with the age of DR15-C.
{width="7.0in"}
### YSO population ages from the K-band Luminosity Function {#analysis:sfh:clusterages}
One of the goals of this study is to reconstruct the history of star formation in the region, by estimating the mean age and age spread of a set of the clusters and groups we identified from the `gather` maps. For this purpose, we constructed the K-band luminosity functions (KLF) of the groups and clusters, and compared them to the KLF of artificial pre-main sequence populations with different age ranges and age spreads. These artificial KLFs were constructed using the pre-main sequence model interpolation code of @Muench:2000aa. Another example of this method applied to a young cluster population can be found in @Roman:2015aa
We selected the samples for DR15-C,DR15-W, DR15-SW and DR15-SE for this analysis. The samples were defined as circular areas that covered the four overdensities. The radii of the circles was chosen as 0.03 deg for DR15-W,SW and SE, and as 0.0225 for DR15-C. The KLF was constructed by limiting the samples by extinction in the following manner: we restricted the sample to those sources that a) fall above an extinction vector corresponding $A_V$=20 mag that reaches the sensitivity limit in a $H$ vs. $J-K$ color-magnitude diagram, and b) fall between an unreddened and unreddened 3 Myr pre-main sequence isochrone, properly shifted to the estimated distance to DR15. The extinction-limited samples assure a minimum contamination from extragalactic sources (which mostly will fall in the area below the extinction vector and the sensitivity limit) and will minimize a bias due to the decrease in the number of detected sources as a function of extinction.
In their study, @Muench:2000aa showed that the observed shape of the KLF (mostly defined by the peak value) of a pre-main sequence population is particularly sensitive to three intrinsic parameters: the underlying Initial Mass Function (IMF), the mean age of the population, and to a lesser extent, the age spread or period of formation. Our analysis consists of constructing a grid of artificial KLF for each of our samples, and compare them with our observed function. We constructed this grid by assuming a fixed IMF and let the time parameters, mean age and age spread, to run free.
For each population (DR15-C,DR15-W, DR15-SW and DR15-SE) we the used code of @Muench:2000aa to simulate artificial KLFs for clusters with the same number of sources and the same distributions for extinction and disk fractions as a function of color ($H-K$). We drew the artificial population from a broken power law IMF using the parameters of the one obtained for the IC 348 cluster (2 Myr old) in the paper of @Muench:2003aa. To draw the artificial populations we used the pre-main sequence stellar evolution model by [@DAntona:1997aa], with $[D/H]=2\times 10^{-5}$ to draw the samples for the artificial clusters . We ran the models using a grid of ages in which we varied the mean age of the cluster between 0.5 and 10 Myr, in s.pdf of 0.5 Myr. For each case, we simulated five age spreads between 1.0 to 5.0 Myr. For each case we simulated 500 clusters.
The method could not be applied succesfully for the DR15-N sample, due to the large extinction variations in that cluster, which does not allow us to obtain a sufficient detection rate in the K band at the highest column density regions. In the remaining clusters, we were able to isolate extinction limited samples satisfactorily.
We determined which age/age spread set fits each of the observed KLFs by averaging the KLF over all simulations for each point in the grid and comparing to the observed KLF, obtaining for each case a reduced $\chi^2$ estimation. In Figures \[fig:nchi1\] and \[fig:nchi2\] we show, for each cluster a contour plot of the reduced $\chi^2$ in a mean age vs age spread plane, indicating the parameter region with the best adjustments.
![The panels show contour maps of the normalized $\chi ^2$ values for the age vs. age spread estimation of clusters DR15-C and DR15-W using pre-main sequence models of the KLF. The most likely values fall within regions with purple and dark blue colors. The white contours indicate 68 and 95 percent confidence limits.The limits of the model grid are indicated by a thin, red line.[]{data-label="fig:nchi1"}](f8a.pdf "fig:"){width="3.0in"} ![The panels show contour maps of the normalized $\chi ^2$ values for the age vs. age spread estimation of clusters DR15-C and DR15-W using pre-main sequence models of the KLF. The most likely values fall within regions with purple and dark blue colors. The white contours indicate 68 and 95 percent confidence limits.The limits of the model grid are indicated by a thin, red line.[]{data-label="fig:nchi1"}](f8b.pdf "fig:"){width="3.0in"}
![Same as in Figure \[fig:nchi1\], but for clusters DR15-SW and DR15-SE[]{data-label="fig:nchi2"}](f9a.pdf "fig:"){width="3.0in"} ![Same as in Figure \[fig:nchi1\], but for clusters DR15-SW and DR15-SE[]{data-label="fig:nchi2"}](f9b.pdf "fig:"){width="3.0in"}
In Table \[tab:results\] we list the results of the artificial KLF modeling analysis for the cluster samples in DR15. For each cluster we also include the number of Class I, Class II and Class III YSOs found in each population. The more embedded clusters DR15-C and DR15-W contain a population that according to our analysis, could have a mean age of 3.0 Myr. The populations flanking the main molecular pillar, DR15-SE and DR15-SW, appear to have mean ages as large as 3.5 and 4.5 Myr, respectively. The contour plots for these two samples, however, appear to have two local minima, suggesting age spreads as short as 3.0-3.5 Myr or as large as 4.5-5.0 Myr. The adjustment to an older age is in agreement with the less embedded status of these two samples. Notice that the confidence ranges in all cases indicate that our method cannot really constrain the age spreads satisfactorily (the 95 percent confidence range in DR15-C is the only one that suggest a constraint towards an age spread of 3.5 to 5 Myr). Still, given the embedded nature of these young star populations, we think it is little plausible that age spreads can be significantly larger than 5 Myr.
[lccccc]{} DR15-C & 3.0\[2.5,3.5\] & 4.5\[3.5,5.0\] & 1 & 8 & 4\
DR15-W & 3.5\[2.5,4.5\] & 4.5\[1.0,5.0\] & 2 & 16 & 5\
DR15-SE & 3.5\[3.0,4.5\] & 3.0\[1.0,5.0\] & 0 & 5 & 0\
DR15-SW & 4.5\[4.0,5.0\] & 3.0\[1.0,5.0\] & 0 & 11 & 2\
DR15-N & – & – & 15 & 21 & 6\
DR15-NE & – & – & 0 & 11 & 3\
\[tab:results\]
The slow removal of the DR15-C cluster molecular envelope {#analysis:sfh:radio_envelope}
---------------------------------------------------------
At 3 Myr of age, we could expect that the DR15-C cluster have removed a significant fraction of its molecular envelope, as it occurs in clusters of a similar age, e.g. IC-348 [@Muench:2003aa], IC 1795 [e.g. @Oey:2005ly Román-Zúñiga et al 2015, in rev.]. In fact our images show how other groups adjacent to DR15-C, like DR15-SE and DR15-SW, which our analysis suggests have similar ages, are associated with much less prominent molecular cloud features. Still, the molecular envelope of DR15-C appears as a well defined structure, both dense and compact, surrounding the cluster atop a dense molecular pillar.
Using the $^{13}$CO(1-0) map of @Schneider:2011aa we made two position-velocity (PV) cuts across the envelope of DR15-C. We used our extinction map and a zero moment (integrated intensity) integration of the $^{13}$CO(1-0) map as a guide. The first cut (L2) runs across the observable structure of the pillar in the region observed. The second cut (L4) runs almost perpendicular to L2. The L2 cut shows a component related to the neck of the pillar, almost 2 pc long with a radial velocity about 4 km/s away from the Cygnus OB2 system velocity (0 km/s). Near the center of the cut, coincidental with the envelope of DR15-C, the PV plot shows a hint of an elliptical shell structure, with a red component moving slightly above 4 km/s and a blue component near 0 km/s. The latter merges into the dark infrared cloud, which shows a very smooth gradient from 2 to 0 km/s. The L4 cut shows for the most part, gas with velocities near 0 km/s but also a much more clear elliptical shell structure at the envelope, which opens from 0 to 5 km/s. The elliptical shell structure corresponds to the expanding envelope of DR15-C.
![Top: Zero moment (integrated intensity) $^{13}$CO(1-0) map of the DR15 region, constructed from the survey of @Schneider:2006aa, the grayscale is indicated in units of km/s. Contours indicate visual extinction, similar to Figure \[fig:ysodist\]. The two red arrow lines labeled as L2 and L4 indicate cuts along which position-velocity (P-V) plots were obtained (see text). Center, Bottom: P-V plots along L2 and L4 cuts, as indicated in top panel; position along the length of each cut is indicated both in arcseconds and in parsecs, asumming a distance of 1.4 kpc.[]{data-label="fig:pvcuts"}](f10a.pdf "fig:"){width="3.0in"}\
![Top: Zero moment (integrated intensity) $^{13}$CO(1-0) map of the DR15 region, constructed from the survey of @Schneider:2006aa, the grayscale is indicated in units of km/s. Contours indicate visual extinction, similar to Figure \[fig:ysodist\]. The two red arrow lines labeled as L2 and L4 indicate cuts along which position-velocity (P-V) plots were obtained (see text). Center, Bottom: P-V plots along L2 and L4 cuts, as indicated in top panel; position along the length of each cut is indicated both in arcseconds and in parsecs, asumming a distance of 1.4 kpc.[]{data-label="fig:pvcuts"}](f10b.pdf "fig:"){width="3.0in"}\
![Top: Zero moment (integrated intensity) $^{13}$CO(1-0) map of the DR15 region, constructed from the survey of @Schneider:2006aa, the grayscale is indicated in units of km/s. Contours indicate visual extinction, similar to Figure \[fig:ysodist\]. The two red arrow lines labeled as L2 and L4 indicate cuts along which position-velocity (P-V) plots were obtained (see text). Center, Bottom: P-V plots along L2 and L4 cuts, as indicated in top panel; position along the length of each cut is indicated both in arcseconds and in parsecs, asumming a distance of 1.4 kpc.[]{data-label="fig:pvcuts"}](f10c.pdf "fig:"){width="3.0in"}
We defined the area of the shell as a rectangle of 5.25$\arcmin \times$3.75$\arcmin$ around the defined center of DR15C. Following the prescription by @Estalella:1999aa, we estimated the column density, N(H$_2$) and the total mass of expanding gas within $3<v_r<6$ km/s, as $M_{out}=$103.5 M$_\odot$. Then we used the method described by [@Qiu:2009aa] to estimate the dynamical time of the component, $t_{dyn}=2.9\times 10^5$ yr. This implies a total mass loss rate of $\dot{M}_{out}=M_{out}/t_{dyn}=360 \mathrm{\ M}_\odot \mathrm{\ Myr}^{-1}$. At this rate, it would take about 3 Myr to remove the total mass of gas in the shell, which we estimate to be 1023 M$_\odot$, using our extinction map, and a distance to the pillar of 1.4 kpc.
This result appears to be consistent with estimates of the mass loss rates in other molecular pillars. For instance, using integral field unit spectroscopy, @Westmoquette:2013aa estimated a mass loss rate of 300 $\mathrm{\ M}_\odot \mathrm{\ Myr}^{-1}$ for the pillars of NGC 3603, which contain a total of about 700 M$_\odot$ of gas. For the “Pillars of creation" in M16, an estimate by @Mcleod:2015aa is of 70 $\mathrm{\ M}_\odot \mathrm{\ Myr}^{-1}$. In both cases, the removal timescales are of the order of 2-3 Myr. However, we need to point out that those estimates are based on models of photoevaporation, while our estimate comes from an estimation of a gas outflow rate from a zero moment map of molecular gas emission. The shell mass could be overestimated, due mainly to a large column density in the line of sight towards Cygnus-X, in which case we would be more consistent with the optical spectroscopy studies. However, what we consider important to note is that the removal of the envelope of DR15-C is relatively slow compared to other embedded cluster regions, where gas dispersal timescales are similar or shorter than the T Tauri timescale, i.e. less than 2 Myr .
Discussion
===========
The main goal of this paper is to reconstruct the history of star formation in the Cygnus-X DR15 region. For this purpose, we made a) an analysis of the spatial distribution of YSO candidate sources classified by evolutionary classes, and b) a comparison of the observed KLF of several young star population samples in our region of study with those of artificial cluster samples drawn from pre-main sequence stellar evolution models.
The classification of YSO candidates and their spatial distribution reveals that several populations of young stars are present in the field of study. At the north/northeast part of our field we identified a very young cluster, DR15-N, hosting 15 Class 0/I sources. The cluster is forming within an infrared dark cloud that runs in the east-west direction at an estimated distance of 1400 pc [@Rizzo:2014aa; @Palau:2014aa]. The cluster also contains a relatively large number (21) of Class II sources and thus it may host the youngest population in the region of study. For this cluster we were not able to construct an unbiased extinction-limited sample to construct a KLF. Instead, we constructed SEDs for all the Class I sources we identified within it, and we compared them to YSO models. We found that in most cases, the models with the best fits correspond to intermediate to high mass YSOs with ages of $\sim 1$ Myr. The mass of the IRDC, estimated from our NICEST extinction map is of $\sim 2400\mathrm{\ M}_\odot$, with an equivalent radii of 5.2 pc. These parameters are in good agreement with massive star forming IRDCs as defined by the analysis of @Kauffmann:2010aa.
The revised distance to Cygnus-X from the study of @Rygl:2012aa is 1.4 kpc, in agreement with the estimations of @Rizzo:2014aa and @Palau:2014aa. From our position-velocity cuts, we note that the emission along the structure of the dark infrared cloud has a null relative velocity compared to the Cygnus-X systemic velocity, same as the DR12-15 structure. It would be difficult to determine if the dark infrared cloud is located at the same distance than the DR15 pillar based only on a radial velocity difference argument. Also, it is important to notice that @Rygl:2012aa could not confirm that Cygnus-X South structures are at the same distance as those in Cygnus-X North where most of their measurements were made. We also lack enough evidence to claim interaction between the DR15 pillar and the IRDC. Using purely morphological arguments we like to comment that the formation of molecular pillars like DR15 is thought to be the result of the interaction between the molecular cloud and the photoionization region formed by the ionizing radiation of the central cluster (Cygnus OB2 in our case), while the IRDC appears to be a highly dense and coherent structure, apparently less affected by the HII regions. Also, our images do not show any obvious pillar structures coincident with the IRDC. These arguments could work in favor of a scenario in which the IRDC could be located at a slightly different distance than the DR15 pillar.
According to our observations and our KLF simulation results, DR15-C is a young cluster with an age of about $~3$ Myrs, but it has a low number of young sources, with only one identified Class I, 8 Class II and 4 Class III sources, all of this despite the presence of the thick, nebulous envelope evident in extinction and $^{13}$CO(1-0) integrated intensity maps. We know the DR15 pillar hosts an embedded cluster, as evidenced by well known far-infrared sources associated with the photodissociation region [@Odenwald:1990aa] and also from our KLF analysis, which indicates that 200 sources could be present within DR15-C, after background and foreground contamination corrections. We do not think the contamination by sources from the DR15-N cluster is too high, as most the density peak of that cluster is located to the northeast and it is more embedded. However, it cannot be fully discarded that some sources in DR15-C are actually sources from DR15-N (and viceversa). Also, there seems to be another YSO group in the filament, south of DR15-C, which reinforces the idea of star formation being active in the pillar.
Our evidence points to a scenario in which a) the star formation episode in DR15-C has probably reached its end or it is near its end and b) the parental gas has dissipated relatively slow, or at least slower that the dispersion time of circumstellar disks of its member stars. Our analysis of the $^{13}$CO(1-0) zero moment map indicates it would take up to 4 Myr to remove the cluster gaseous envelope. However, our observations in other cluster forming regions like the Rosette Molecular Cloud and W3 [e.g. @Roman-Zuniga:2008aa; @Ybarra:2013aa; @Roman:2015aa] suggest that clusters hosting intermediate to massive stars may remove their gas envelopes in periods shorter than the T Tauri timescale. Therefore, our analysis of DR15-C suggests that clusters forming in this kind of pillar structures, could dissipate their envelopes at slower rates.
As shown in recent studies like those of @Westmoquette:2013aa and @Mcleod:2015aa the mass loss rates in pillars due to photoevaporation are of the order of $\sim 10^2\mathrm{\ M}_\odot \mathrm{\ Myr}^{-1}$, not too different from our estimations in the expanding gas shell. Clearly, we are comparing very different methods, that refer to very different processes (shell expansion vs. shell photo-erosion). It is important to notice that the DR15 pillar may not be the same kind of dusty pillar as those in M16, because extinction is too high to allow delineating the structure in DR15 using optical images. The molecular fragment located west of DR15 shows as a shadow with a pillar morphology in optical images, which make us think that DR15 could be a pillar too. Even so, the fact that the mass loss estimates for DR15 coincide in the order of magnitude of the effect with respect to M16 is interesting, and motivates further investigation. A detailed study of the removal of the DR15-C envelope is out of the scope of this paper, but it is the main topic in a close following study (Román-Zúñiga et al., in prep).
DR15-W, DR15-SE and DR15-SW appear to be slightly older populations ($3.5-4.75$ Myrs) with much less gas and dust observed. However, we find evidence of remains of a structure that could have been similar to the pillar associated with DR15-C. It may be possible that the DR15-SE and DR15-SW groups belong to clusters formed before DR15-C, but their ages may not be much older, as evident from the presence of Class II and Class III sources. The estimated ages of the DR15-SW and SE samples suggest that the cluster evolution period, from formation to gas dispersal in the region could be around 5 Myr.
Summary
=======
The Cygnus-X DR15 region presents a prominent gaseous pillar as well as an IRDC, both hosting clusters of young stars. In this investigation, we made a multi-wavelength study of the young stellar population in the region. For this purpose, we processed and analyzed deep, high quality near-IR images of the region, as well as X-ray images from the Chandra Observatory. Using these datasets we obtained photometry catalogs for all point sources we were able to detect. We combined these catalogs with the 3.6 to 24 $\mu$m photometry catalog of the Cygnus-X Spitzer Legacy Survey, resulting in a master catalog containing almost 47 thousand individual sources. From our master catalog, we identified 226 young stellar sources, which we classified according to their evolutionary class related to the prominence of their circumstellar disks. We found that the young sources distribute into 26 Class 0/I, 156 Class II and 45 Class III sources.
From our near-IR we constructed an extinction map, which we used to study the spatial distribution of the young sources in the molecular cloud structure present in the region. We found that the youngest population of this region is currently forming at the IRDC, at the regions of highest column density. Combining this with maps of YSO surface density, we were able to identify several groups, possibly associated to distinct stellar clusters. We obtained extinction-limited samples of these groups in order to construct their K band luminosity functions (KLF). We compared the observed KLFs with those of artificial young cluster populations sampled from interpolation of pre-main sequence models. This allowed us to make first order estimations of the mean ages and age spreads of the cluster population samples.
We constructed SEDs for all Class I sources identified within the IRDC region, which allowed us to estimate their mass and disk accretion rates. These estimations are consistent with the formation of an intermediate mass star cluster, indicating that structures of this kind in Cygnus-X are possibly birth places of massive clusters.
Using the FCRAO $^{13}$CO(1-0) from the study of @Schneider:2011aa we estimated the radial velocity distribution along the IRDC and the DR15 pillar, and we found that the nebulous envelope of the DR15-C cluster at the tip of the pillar is consistent with an expanding shell morphology. We estimated the mass loss rate of gas in this expanding shell and we deduced that the dissipation process of the DR15-C cluster gas envelope is relatively slow, compared to what we found in other studies of cluster forming regions, where the gas dispersal process is shorter than the T Tauri timescale. The mass loss rate we estimated is in the same order of magnitude as mass loss rates by photodissociation found in studies of other gas pillars using optical spectroscopy. This suggests that clusters forming in gas pillars like DR15-C could have a different evolution process than clusters forming at dense clumps in other giant molecular clouds.
The presence of other populations containing Class II and Class III sources at the regions flanking the DR15 pillar, and projected near less dense but still noticeable gaseous structures, with estimated ages of 4 to 5 Myr is suggestive of a process of cluster forming processes that take about that long to form an dissipate in the molecular cloud complex that surrounds the Cygnus-X HII region.
**Acknowledgements:** We thank the referee, Nicola Schneider for providing a comprehensive and constructive review that greatly improved the content of our manuscript. We thank Nicola Schneider for kindly providing us with a copy of the FCRAO $^{13}$CO(1-0) map for our study. CRZ acknowledges support from CONACYT project CB2010-152160, Mexico and programs UNAM-DGAPA-PAPIIT IN103014 and IN116315. EAL acknowledges support from the National Science Foundation through NSF Grant AST-1109679 to the University of Florida.
This study is based on observations collected at the Centro Astronómico Hispano Alemán (CAHA) at Calar Alto, operated jointly by the Max-Planck Institut für Astronomie and the Instituto de Astrofísica de Andalucía (CSIC). We acknowledge the staff at Calar Alto for top of the line queued observations at the 3.5m with OMEGA 2000. We acknowledge use of data products from the 2MASS, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Centre/California Institute of Technology (funded by the USA National Aeronautics and Space Administration and National Science Foundation). This work is partly based 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. The scientific results reported in this article are based to a significant degree on data obtained from the Chandra Data Archive; particularly, we made use of data obtained from the Chandra Source Catalog, provided by the Chandra X-ray Center (CXC) as part of the Chandra Data Archive. We made use of the `pvextractor` tool by Adam Ginsburg, that is part of the Radio Astro Tools repository (http://github.com/radio-astro-tools).
[*Facilities:*]{} , , .
, E. & [Arnouts]{}, S. 1996, , 117, 393
Broos, P. S., Townsley, L. K., Feigelson, E. D., et al. 2010, , 714, 1582
Campbell, M. F., Niles, D., Nawfel, R., et al. 1982, , 261, 550
, S. & [Hut]{}, P. 1985, , 298, 80
, F. & [Mazzitelli]{}, I. 1997, , 68
Di Francesco, J., Johnstone, D., Kirk, H., MacKenzie, T., & Ledwosinska, E. 2008, , 175, 277
Dotter, A., Chaboyer, B., Jevremovi[ć]{}, D., et al. 2008, , 178, 89
Dutra, C. M., & Bica, E. 2001, , 376, 434
Estalella, R. & G. [Anglada]{}. . Number 50 in [Textos Docents]{}. Barcelona: Edicions Universitat de Barcelona, 1999.
Fang, K., et al. 2013, , 207, 5
Fruscione, A., et al. 2006, , 6270
Gladwin, P. P., Kitsionas, S., Boffin, H. M. J., & Whitworth, A. P. 1999, , 302, 305
Gutermuth, R. A., Myers, P. C., Megeath, S. T. et al. 2008, , 674, 336
Gutermuth, R. A., Megeath, S. T., Myers, P. C., et al. 2009, , 184, 18
Hora, J., Gutermuth, R., Carey, S. et al. 2009, in Reionization to Exoplanets: Spitzer’s Growing Legacy, Ed. Ogle, P., ASP Conf. Ser., pp. 26-28
Jim[é]{}nez-Esteban, F. M., Rizzo, J. R., & Palau, A. 2010, , 713, 429
Kauffmann, J., & Pillai, T. 2010, ApJ, 723, L7
Kleinmann, S. G., Sargent, D. G., Telesco, C. M., Joyce, R. R., & Gillett, F. C. 1979, , 227, 126
Kn[ö]{}dlseder, J. 2000, , 360, 539
Kryukova, E., Megeath, S.T., Gutermuth, R.A., et al. 2012, , 144, 31
Kryukova, E., Megeath, S.T., Hora, J.L., et al. 2014, , 148, 11
Kurtz, S., Churchwell, E., & Wood, D. O. S. 1994, , 91, 659
Lada, C. J., Lombardi, M., Roman-Zuniga, C., Forbrich, J., & Alves, J. F. 2013, , 778, 133
, M. & [Alves]{}, J. 2001, , 377, 1023
Mc Leod, A. F., Dale, J. E., Ginsburg, A., et al. 2015, in press, arXiv:1504.03323
Meyer, M. R., Calvet, N., & Hillenbrand, L. A. 1997, , 114, 288
, A. A., [Lada]{}, E. A., & [Lada]{}, C. J. 2000, , 533, 358
, A. A., [Lada]{}, E. A., [Lada]{}, C. J. et al. 2003, , 125, 2029
Odenwald, S. F., Campbell, M. F., Shivanandan, K., et al. 1990, , 99, 288
, M. S., [Watson]{}, A. M., [Kern]{}, K., & [Walth]{}, G. L. 2005, , 129, 393
Oka, T., Yamamoto, S., Iwata, M., et al. 2001, , 558, 176
Palau, A., Rizzo, J. R., Girart, J. M., & Henkel, C. 2014, , 784, LL21
Piddington, J. H., & Minnett, H. C. 1952, Australian Journal of Scientific Research A Physical Sciences, 5, 17
Qiu, K., Zhang, Q., Wu, J. & Chen, H.-R. 2009, , 696, 66
Ragan, S., Henning, T., Krause, O., et al. 2012, , 547, AA49
Reipurth, B., & Schneider, N. 2008, Handbook of Star Forming Regions, Volume I, 36
Rieke, G. H., Lebofsky, M. J., & Low, F. J. 1985, , 90, 900
Rizzo, J. R., Jim[é]{}nez-Esteban, F. M., & Ortiz, E. 2008, , 681, 355
Rizzo, J. R., Palau, A., Jim[é]{}nez-Esteban, F., & Henkel, C. 2014, , 564, AA2
Robitaille, T. P., Whitney, B. A., Indebetouw, R., & Wood, K. 2007, , 169, 328
Rom[á]{}n-Z[ú]{}[ñ]{}iga, C. G., Elston, R., Ferreira, B., & Lada, E. A. 2008, , 672, 861
Rom[á]{}n-Z[ú]{}[ñ]{}iga, C. G., Alves, J. F., Lada, C. J., & Lombardi, M. 2010, , 725, 2232
Rom[á]{}n-Z[ú]{}[ñ]{}iga, C. G., Ybarra, J. E., Meg[í]{}as, G. D., et al. 2015, , 150, 80
Rygl, K. L. J., Brunthaler, A., Sanna, A., et al. 2012, , 539, A79
Sault, R. J., Teuben, P. J., & Wright, M. C. H. 1995, Astronomical Data Analysis Software and Systems IV, 77, 433
Schneider, N., Bontemps, S., Simon, R., et al. 2006, , 458, 855
Schneider, N., Bontemps, S., Simon, R., et al. 2011, , 529, AA1
, M. B. 2005, in Astronomical Society of the Pacific Conference Series, Vol. 347, Astronomical Data Analysis Software and Systems XIV, ed. [P. Shopbell, M. Britton, & R. Ebert]{}, 29
Teixeira, P. S., Lada, C. J., Marengo, M., & Lada, E. A. 2012, , 540, AA83
Westmoquette, M. S., Dale, J. E., Ercolano, B., & Smith, L. J. 2013, , 435, 30
Ybarra, J. E., Lada, E. A., Roman-Z[ú]{}[ñ]{}iga, C. G. et al. 2013, , 769, 140
Young Sources Identified in the DR15 region \[app:ysos\]
========================================================
In Tables \[tab:ClassI\], \[tab:ClassII\] and \[tab:ClassIII\] we list YSO sources identified as Class I, Class II and Class III in our region of study. The tables contain JHK photometry from Calar Alto (or 2MASS when pertinent, see section \[obs:nir\]), 3.6 to 24 $\mu$m photometry from the Spitzer CXLS, and when possible, median energy and total energy flux values for those sources detected with Chandra ACIS (that is the case for all Class III sources). Identifications and positions from our Calar Alto survey are listed for most cases. The remaining sources are listed with 2MASS or CXLS depending if they were detected in those surveys.
[lllllllllllllllllllll]{}
CAHA\_20322208\_402017 & 20:32:22.08 & +40:20:17.2 & 18.827 & 9.999 & 18.101 & 0.02 & 12.026 & 0.02 & 7.411 & 0.015 & 6.021 & 0.015 & 5.066 & 0.015 & 4.136 & 0.015 & 0.395 & 0.018 & — & —\
CAHA\_20322856\_401941 & 20:32:28.56 & +40:19:41.6 & 14.827 & 0.003 & 11.397 & 0.019 & 8.943 & 0.016 & 7.372 & 0.015 & 6.304 & 0.015 & 5.349 & 0.015 & 4.427 & 0.015 & 0.981 & 0.017 & — & —\
CAHA\_20321777\_401408 & 20:32:17.77 & +40:14:08.2 & 15.139 & 0.015 & 12.744 & 0.014 & 10.647 & 0.014 & 8.162 & 0.015 & 7.323 & 0.015 & 6.455 & 0.015 & 4.864 & 0.015 & 1.412 & 0.016 & — & —\
SSTCYGX\_J203222.99\_402021.4 & 20:32:22.99 & +40:20:21.4 & — & — & — & — & — & — & 13.706 & 0.17 & 11.445 & 0.106 & 9.627 & 0.041 & 7.858 & 0.033 & 1.662 & 0.019 & — & —\
CAHA\_20322060\_401950 & 20:32:20.60 & +40:19:50.1 & — & — & — & — & 16.815 & 0.02 & 10.233 & 0.021 & 8.048 & 0.016 & 6.613 & 0.015 & 5.777 & 0.017 & 2.145 & 0.038 & — & —\
CAHA\_20323152\_401352 & 20:32:31.52 & +40:13:53.0 & 20.457 & 0.042 & 14.909 & 0.008 & 14.374 & 0.089 & 8.519 & 0.016 & 7.417 & 0.015 & 6.565 & 0.017 & 5.96 & 0.03 & 2.229 & 0.111 & — & —\
CAHA\_20322113\_402025 & 20:32:21.13 & +40:20:25.6 & 17.814 & 9.999 & 16.994 & 0.013 & 13.068 & 0.009 & 9.281 & 0.016 & 8.148 & 0.016 & 7.201 & 0.016 & 6.35 & 0.017 & 2.316 & 0.021 & 2.694 & 5.880E-15\
CAHA\_20322194\_401937 & 20:32:21.94 & +40:19:37.9 & — & — & 20.053 & 0.075 & 15.564 & 0.017 & 11.613 & 0.021 & 10.495 & 0.02 & 7.246 & 0.016 & 5.487 & 0.021 & 2.668 & 0.056 & — & —\
CAHA\_20322126\_401601 & 20:32:21.26 & +40:16:01.8 & 18.458 & 0.023 & 14.213 & 0.007 & 11.565 & 0.018 & 8.827 & 0.015 & 7.851 & 0.015 & 7.005 & 0.015 & 6.26 & 0.017 & 3.036 & 0.18 & — & —\
CAHA\_20320254\_401838 & 20:32:02.54 & +40:18:39.0 & — & — & — & — & 17.492 & 0.073 & 11.914 & 0.017 & 10.031 & 0.016 & 8.879 & 0.017 & 8.241 & 0.033 & 3.077 & 0.046 & — & —\
CAHA\_20322282\_401940 & 20:32:22.82 & +40:19:40.9 & — & — & — & — & 16.245 & 0.038 & 12.171 & 0.03 & 10.362 & 0.019 & 9.161 & 0.021 & 8.26 & 0.059 & 3.583 & 0.118 & — & —\
CAHA\_20315797\_401835 & 20:31:57.97 & +40:18:35.8 & 21.759 & 0.22 & — & — & — & — & 13.975 & 0.028 & 12.189 & 0.027 & 11.35 & 0.036 & 10.973 & 0.112 & 3.677 & 0.035 & — & —\
SSTCYGX\_J203153.84\_401833.9 & 20:31:53.84 & +40:18:33.9 & — & — & — & — & — & — & 13.039 & 0.018 & 11.039 & 0.016 & 9.802 & 0.019 & 8.87 & 0.028 & 5.383 & 0.092 & — & —\
CAHA\_20322222\_401955 & 20:32:22.22 & +40:19:56.0 & 18.432 & 9.999 & 16.975 & 0.008 & 13.395 & 0.005 & 10.72 & 0.016 & 9.534 & 0.016 & 8.519 & 0.018 & 7.538 & 0.025 & — & — & — & —\
CAHA\_20322014\_401953 & 20:32:20.14 & +40:19:53.6 & 17.771 & 9.999 & 17.345 & 9.999 & 14.847 & 0.006 & 9.945 & 0.016 & 8.612 & 0.016 & 7.726 & 0.017 & 7.577 & 0.026 & — & — & — & —\
CAHA\_20322784\_401942 & 20:32:27.84 & +40:19:42.4 & 19.796 & 0.037 & 15.484 & 0.007 & 12.902 & 0.005 & 10.559 & 0.017 & 9.733 & 0.018 & 8.976 & 0.021 & 8.266 & 0.062 & — & — & — & —\
CAHA\_20322033\_402001 & 20:32:20.33 & +40:20:01.5 & 16.173 & 0.003 & 15.529 & 0.007 & 15.118 & 0.012 & 13.54 & 0.118 & 11.84 & 0.066 & 10.647 & 0.059 & 9.852 & 0.075 & — & — & — & —\
CAHA\_20322899\_401821 & 20:32:28.99 & +40:18:21.2 & — & — & — & — & 16.45 & 0.03 & 14.072 & 0.062 & 13.326 & 0.036 & 12.328 & 0.117 & 11.479 & 0.24 & — & — & — & —\
CAHA\_20322111\_402001 & 20:32:21.11 & +40:20:01.1 & — & — & — & — & 16.686 & 0.047 & 11.988 & 0.033 & 10.318 & 0.023 & 9.153 & 0.019 & 8.38 & 0.027 & — & — & — & —\
CAHA\_20322781\_402032 & 20:32:27.81 & +40:20:32.7 & — & — & — & — & 18.385 & 0.1 & 14.168 & 0.022 & 12.909 & 0.018 & 12.075 & 0.042 & 11.14 & 0.069 & — & — & — & —\
SSTCYGX\_J203226.39\_401847.4 & 20:32:26.39 & +40:18:47.4 & — & — & — & — & — & — & 14.761 & 0.051 & 13.612 & 0.028 & 12.252 & 0.112 & 11.323 & 0.185 & — & — & — & —\
CAHA\_20320404\_401856 & 20:32:04.04 & +40:18:56.6 & 18.559 & 0.009 & 15.162 & 0.01 & 13.261 & 0.01 & 11.352 & 0.017 & 10.546 & 0.016 & 9.895 & 0.03 & 9.09 & 0.064 & 5.954 & 0.42 & 4.636 & 1.717E-14\
CAHA\_20323480\_401629 & 20:32:34.80 & +40:16:29.2 & 17.797 & 0.023 & 14.591 & 0.016 & 12.513 & 0.009 & 10.185 & 0.05 & 9.455 & 0.048 & 8.048 & 0.145 & — & — & — & — & 3.906 & 1.154E-14\
CAHA\_20322210\_401800 & 20:32:22.10 & +40:18:00.3 & 13.681 & 0.003 & 12.608 & 0.006 & 12.447 & 0.023 & 11.337 & 0.021 & 11.004 & 0.022 & 10.653 & 0.153 & 9.602 & 0.323 & — & — & 1.438 & 6.480E-15\
CAHA\_20322358\_401729 & 20:32:23.58 & +40:17:29.9 & 14.661 & 0.003 & 13.556 & 0.005 & 13.146 & 0.005 & 12.243 & 0.08 & 12.081 & 0.112 & 9.757 & 0.146 & — & — & — & — & 1.204 & 4.064E-15\
CAHA\_20325161\_401945 & 20:32:51.61 & +40:19:45.6 & 16.385 & 0.005 & 14.341 & 0.007 & 13.271 & 0.005 & 12.165 & 0.017 & 11.689 & 0.017 & 11.375 & 0.024 & 11.162 & 0.054 & — & — & 3.351 & 1.147E-14\
\[tab:ClassI\]
[lllllllllllllllllllll]{} CAHA\_20324212\_401726 & 20:32:42.12 & +40:17:27.0 & 12.732 & 0.005 & 11.778 & 0.03 & 11.309 & 0.023 & 11.001 & 0.018 & 10.834 & 0.018 & 10.6 & 0.103 & 9.503 & 0.22 & 2.175 & 0.175 & — & —\
TWOM\_20322301\_4019226 & 20:32:23.02 & +40:19:22.7 & 12.204 & 0.025 & 10.916 & 0.021 & 10.071 & 0.014 & 8.789 & 0.015 & 8.154 & 0.015 & 7.561 & 0.015 & 6.622 & 0.016 & 2.94 & 0.037 & — & —\
CAHA\_20323244\_402105 & 20:32:32.44 & +40:21:05.1 & 13.473 & 0.003 & 11.578 & 0.018 & 10.661 & 0.017 & 10.131 & 0.015 & 9.936 & 0.015 & 9.751 & 0.016 & 9.249 & 0.024 & 3.358 & 0.043 & — & —\
CAHA\_20321111\_401916 & 20:32:11.11 & +40:19:16.7 & 19.311 & 0.033 & 15.549 & 0.016 & 13.14 & 0.007 & 11.057 & 0.03 & 9.915 & 0.023 & 9.281 & 0.096 & 8.539 & 0.248 & 3.658 & 0.235 & — & —\
CAHA\_20321736\_401949 & 20:32:17.36 & +40:19:49.5 & 19.194 & 0.02 & 15.036 & 0.006 & 12.555 & 0.004 & 10.287 & 0.017 & 9.536 & 0.016 & 8.877 & 0.063 & 8.238 & 0.205 & 3.748 & 0.159 & — & —\
CAHA\_20320319\_402215 & 20:32:03.19 & +40:22:15.9 & 14.418 & 0.004 & 12.749 & 0.008 & 12.095 & 0.02 & 9.921 & 0.016 & 9.256 & 0.015 & 8.651 & 0.017 & 7.743 & 0.038 & 4.189 & 0.086 & — & —\
CAHA\_20324852\_402104 & 20:32:48.52 & +40:21:04.8 & 13.849 & 0.003 & 12.805 & 0.005 & 12.1 & 0.021 & 10.642 & 0.015 & 9.731 & 0.015 & 9.054 & 0.016 & 8.221 & 0.021 & 4.79 & 0.046 & — & —\
CAHA\_20325574\_402220 & 20:32:55.74 & +40:22:20.6 & — & — & 19.872 & 0.06 & 16.277 & 0.041 & 11.28 & 0.016 & 9.469 & 0.015 & 8.207 & 0.015 & 7.367 & 0.015 & 4.826 & 0.036 & — & —\
CAHA\_20315532\_402216 & 20:31:55.32 & +40:22:16.8 & 14.671 & 0.003 & 13.231 & 0.008 & 12.382 & 0.006 & 11.501 & 0.016 & 11.015 & 0.016 & 10.594 & 0.019 & 9.544 & 0.026 & 4.841 & 0.196 & — & —\
CAHA\_20321148\_401807 & 20:32:11.71 & +40:18:05.1 & 13.322 & 0.039 & 12.351 & 0.042 & 11.517 & 9.999 & 10.704 & 0.019 & 10.358 & 0.018 & 10.261 & 0.051 & 9.379 & 0.117 & 4.949 & 0.169 & — & —\
CAHA\_20320833\_401604 & 20:32:08.33 & +40:16:04.7 & 13.914 & 0.004 & 12.784 & 0.005 & 12.356 & 0.023 & 11.52 & 0.019 & 11.226 & 0.017 & 10.55 & 0.061 & 9.419 & 0.111 & 5.031 & 0.162 & 1.774 & 1.365E-14\
CAHA\_20321289\_401257 & 20:32:12.89 & +40:12:58.0 & 14.167 & 0.011 & 12.895 & 0.021 & 12.215 & 0.02 & 11.591 & 0.016 & 11.367 & 0.016 & 11.211 & 0.04 & 10.694 & 0.089 & 5.4 & 0.076 & 2.069 & 9.625E-15\
CAHA\_20324359\_402121 & 20:32:43.59 & +40:21:21.3 & 14.853 & 0.003 & 12.584 & 0.007 & 11.295 & 0.017 & 9.688 & 0.015 & 9.134 & 0.015 & 8.61 & 0.015 & 7.939 & 0.016 & 5.42 & 0.07 & — & —\
TWOM\_20315270\_4019059 & 20:31:52.71 & +40:19:06.0 & 17.14 & 0.217 & 14.15 & 0.042 & 12.799 & 0.029 & 12.079 & 0.019 & 11.8 & 0.026 & 11.499 & 0.045 & 10.641 & 0.084 & 5.502 & 0.245 & 2.373 & 9.074E-15\
CAHA\_20315661\_401614 & 20:31:56.61 & +40:16:14.0 & 12.979 & 0.005 & 12.093 & 0.019 & 11.357 & 0.016 & 10.733 & 0.016 & 10.412 & 0.016 & 10.187 & 0.034 & 9.427 & 0.073 & 5.586 & 0.221 & 1.511 & 1.315E-14\
CAHA\_20320053\_402026 & 20:32:00.53 & +40:20:27.0 & 15.688 & 0.003 & 13.881 & 0.005 & 12.739 & 0.005 & 11.66 & 0.017 & 10.921 & 0.016 & 10.346 & 0.027 & 9.495 & 0.049 & 5.696 & 0.133 & — & —\
CAHA\_20325540\_402150 & 20:32:55.40 & +40:21:50.7 & 15.923 & 0.005 & 13.857 & 0.007 & 12.355 & 0.007 & 10.89 & 0.016 & 10.376 & 0.016 & 9.785 & 0.016 & 9.012 & 0.018 & 5.954 & 0.056 & — & —\
CAHA\_20330082\_401800 & 20:33:00.82 & +40:18:00.2 & 13.978 & 0.007 & 12.929 & 0.012 & 12.615 & 0.026 & 11.539 & 0.016 & 11.08 & 0.016 & 10.769 & 0.018 & 9.75 & 0.022 & 6.142 & 0.09 & — & —\
CAHA\_20322395\_402218 & 20:32:23.95 & +40:22:18.3 & 15.943 & 0.037 & 14.013 & 0.035 & 12.994 & 0.033 & 11.783 & 0.016 & 11.241 & 0.016 & 10.787 & 0.021 & 10.031 & 0.034 & 6.152 & 0.158 & — & —\
CAHA\_20321650\_402141 & 20:32:16.50 & +40:21:41.1 & 18.196 & 0.012 & 15.382 & 0.016 & 13.564 & 0.021 & 12.04 & 0.017 & 11.356 & 0.016 & 10.867 & 0.029 & 10.155 & 0.057 & 6.186 & 0.177 & — & —\
CAHA\_20320880\_402054 & 20:32:08.80 & +40:20:54.9 & 15.698 & 0.003 & 13.914 & 0.006 & 12.866 & 0.004 & 11.565 & 0.016 & 10.84 & 0.016 & 10.039 & 0.025 & 8.887 & 0.039 & 6.378 & 0.098 & — & —\
CAHA\_20324730\_402217 & 20:32:47.30 & +40:22:17.9 & 14.143 & 0.004 & 13.131 & 0.006 & 12.514 & 0.006 & 11.697 & 0.016 & 11.226 & 0.016 & 10.764 & 0.019 & 10.016 & 0.021 & 6.462 & 0.123 & — & —\
CAHA\_20321184\_401733 & 20:32:11.84 & +40:17:33.8 & 14.27 & 0.002 & 12.991 & 0.005 & 12.4 & 0.009 & 11.543 & 0.017 & 11.161 & 0.017 & 10.727 & 0.055 & 10.004 & 0.086 & 6.559 & 0.163 & — & —\
CAHA\_20325680\_401223 & 20:32:56.80 & +40:12:23.3 & 15.594 & 0.015 & 13.946 & 0.017 & 12.806 & 0.017 & 11.997 & 0.016 & 11.55 & 0.016 & 11.297 & 0.021 & 10.649 & 0.029 & 6.835 & 0.132 & — & —\
CAHA\_20330112\_401449 & 20:33:01.12 & +40:14:49.2 & 14.081 & 0.011 & 13.502 & 0.036 & 12.493 & 0.025 & 11.214 & 0.016 & 10.878 & 0.016 & 10.53 & 0.018 & 9.727 & 0.02 & 6.93 & 0.237 & — & —\
CAHA\_20330982\_401154 & 20:33:09.82 & +40:11:54.6 & 14.388 & 0.01 & 13.128 & 0.012 & 12.538 & 0.024 & 11.265 & 0.016 & 10.82 & 0.016 & 10.411 & 0.017 & 9.723 & 0.017 & 6.99 & 0.199 & — & —\
CAHA\_20325220\_402406 & 20:32:52.20 & +40:24:06.2 & 15.228 & 0.006 & 13.653 & 0.008 & 12.688 & 0.009 & 11.626 & 0.016 & 11.181 & 0.016 & 10.784 & 0.018 & 10.062 & 0.029 & 7.023 & 0.064 & — & —\
CAHA\_20324795\_402214 & 20:32:47.95 & +40:22:14.2 & 18.103 & 0.007 & 15.928 & 0.007 & 14.661 & 0.007 & 13.505 & 0.038 & 12.873 & 0.032 & 12.302 & 0.053 & 11.42 & 0.07 & 7.027 & 0.246 & — & —\
CAHA\_20325444\_401949 & 20:32:54.44 & +40:19:49.5 & 14.826 & 0.004 & 13.712 & 0.007 & 13.046 & 0.005 & 12.011 & 0.016 & 11.721 & 0.016 & 11.454 & 0.025 & 11.034 & 0.037 & 7.044 & 0.183 & — & —\
CAHA\_20320189\_401007 & 20:32:01.89 & +40:10:07.8 & 16.296 & 0.009 & 14.641 & 0.011 & 13.655 & 0.049 & 12.911 & 0.02 & 12.625 & 0.02 & 12.386 & 0.049 & 11.962 & 0.109 & 7.045 & 0.217 & — & —\
CAHA\_20330546\_401215 & 20:33:05.46 & +40:12:15.7 & 14.32 & 0.036 & 13.247 & 0.038 & 12.594 & 0.055 & 11.881 & 0.016 & 11.534 & 0.016 & 11.129 & 0.021 & 10.243 & 0.035 & 7.084 & 0.19 & 1.701 & 1.217E-14\
CAHA\_20330798\_401915 & 20:33:07.98 & +40:19:15.5 & 13.951 & 0.007 & 12.724 & 0.01 & 12.251 & 0.024 & 11.21 & 0.016 & 10.83 & 0.016 & 10.462 & 0.017 & 9.941 & 0.022 & 7.182 & 0.209 & — & —\
CAHA\_20321078\_402356 & 20:32:10.78 & +40:23:56.0 & 16.728 & 0.005 & 14.42 & 0.011 & 12.906 & 0.005 & 11.646 & 0.016 & 10.951 & 0.016 & 10.453 & 0.017 & 9.679 & 0.019 & 7.496 & 0.154 & — & —\
CAHA\_20331260\_402247 & 20:33:12.60 & +40:22:47.2 & 17.066 & 0.011 & 15.378 & 0.096 & 14.455 & 0.096 & 13.466 & 0.024 & 13.22 & 0.021 & 12.702 & 0.059 & 11.862 & 0.087 & 7.898 & 0.246 & — & —\
CAHA\_20321598\_401023 & 20:32:15.98 & +40:10:23.5 & 14.683 & 0.013 & 13.6 & 0.014 & 12.856 & 0.033 & 12.218 & 0.016 & 11.902 & 0.016 & 11.536 & 0.023 & 11.097 & 0.04 & — & — & 1.467 & 1.295E-14\
CAHA\_20323581\_400914 & 20:32:35.81 & +40:09:14.3 & 16.045 & 9.999 & 14.927 & 0.018 & 13.921 & 0.023 & 13.471 & 0.021 & 13.326 & 0.022 & 12.951 & 0.121 & 12.124 & 0.243 & — & — & — & —\
CAHA\_20325027\_401053 & 20:32:50.27 & +40:10:53.5 & 16.775 & 0.023 & 15.134 & 0.023 & 13.962 & 0.025 & 13.741 & 0.034 & 13.334 & 0.032 & 12.821 & 0.084 & 11.766 & 0.129 & — & — & — & —\
CAHA\_20320910\_401441 & 20:32:09.10 & +40:14:41.7 & 15.85 & 0.006 & 13.621 & 0.006 & 12.556 & 0.021 & 11.916 & 0.019 & 11.779 & 0.018 & 11.381 & 0.052 & 10.657 & 0.113 & — & — & — & —\
CAHA\_20321210\_401240 & 20:32:12.10 & +40:12:41.0 & 15.8 & 0.01 & 14.19 & 0.009 & 13.448 & 0.03 & 12.56 & 0.018 & 12.053 & 0.017 & 11.504 & 0.035 & 10.608 & 0.052 & — & — & — & —\
CAHA\_20320975\_401131 & 20:32:09.75 & +40:11:31.3 & 17.666 & 0.01 & 15.457 & 0.011 & 14.374 & 0.035 & 13.816 & 0.027 & 13.654 & 0.033 & 13.254 & 0.131 & 12.599 & 0.17 & — & — & — & —\
CAHA\_20320721\_401330 & 20:32:07.21 & +40:13:30.5 & 15.47 & 0.016 & 14.657 & 0.025 & 14.144 & 0.026 & 13.503 & 0.028 & 13.251 & 0.027 & 12.661 & 0.125 & 11.52 & 0.207 & — & — & — & —\
CAHA\_20320583\_401139 & 20:32:05.83 & +40:11:39.8 & 17.571 & 0.012 & 15.648 & 0.011 & 14.458 & 0.035 & 13.53 & 0.022 & 13.079 & 0.02 & 12.704 & 0.065 & 12.004 & 0.105 & — & — & — & —\
CAHA\_20321307\_401250 & 20:32:13.07 & +40:12:50.0 & 16.731 & 0.009 & 15.006 & 0.01 & 13.919 & 0.027 & 12.918 & 0.024 & 12.355 & 0.021 & 12.049 & 0.051 & 11.509 & 0.141 & — & — & — & —\
CAHA\_20321094\_401157 & 20:32:10.94 & +40:11:57.3 & 14.802 & 0.01 & 13.906 & 0.009 & 13.395 & 0.03 & 13.259 & 0.022 & 13.186 & 0.025 & 12.688 & 0.096 & 11.88 & 0.237 & — & — & — & —\
CAHA\_20321434\_401357 & 20:32:14.34 & +40:13:57.8 & 16.7 & 0.152 & 14.895 & 0.062 & 14.25 & 0.024 & 12.878 & 0.017 & 12.516 & 0.019 & 12.323 & 0.056 & 11.361 & 0.082 & — & — & — & —\
CAHA\_20320296\_401350 & 20:32:02.96 & +40:13:50.9 & 17.226 & 0.009 & 14.281 & 0.007 & 12.922 & 0.024 & 12.062 & 0.018 & 11.914 & 0.019 & 11.574 & 0.078 & 10.965 & 0.228 & — & — & — & —\
CAHA\_20320488\_401316 & 20:32:04.88 & +40:13:16.0 & 16.034 & 0.007 & 14.24 & 0.007 & 13.273 & 0.026 & 12.802 & 0.022 & 12.554 & 0.022 & 11.825 & 0.082 & 10.472 & 0.128 & — & — & — & —\
CAHA\_20320511\_401231 & 20:32:05.11 & +40:12:31.7 & 19.14 & 0.027 & 15.766 & 0.01 & 14.088 & 0.03 & 13.316 & 0.027 & 13.068 & 0.026 & 12.75 & 0.106 & 11.833 & 0.237 & — & — & — & —\
CAHA\_20321626\_401254 & 20:32:16.26 & +40:12:54.2 & 17.864 & 0.009 & 15.707 & 0.01 & 14.282 & 0.023 & 12.711 & 0.017 & 12.039 & 0.017 & 11.581 & 0.049 & 11.111 & 0.103 & — & — & — & —\
CAHA\_20323916\_401223 & 20:32:39.16 & +40:12:23.2 & 14.6 & 0.018 & 12.759 & 0.022 & 12.049 & 0.022 & 11.536 & 0.017 & 11.475 & 0.019 & 11.083 & 0.083 & 10.366 & 0.216 & — & — & — & —\
CAHA\_20323223\_401347 & 20:32:32.23 & +40:13:47.1 & 17.797 & 0.01 & 15.102 & 0.01 & 13.512 & 0.012 & 11.859 & 0.046 & 11.264 & 0.046 & 9.956 & 0.147 & 8.584 & 0.184 & — & — & — & —\
CAHA\_20323296\_401334 & 20:32:32.96 & +40:13:34.5 & 14.659 & 0.008 & 12.915 & 0.021 & 12.118 & 0.022 & 10.675 & 0.024 & 10.098 & 0.021 & 9.251 & 0.097 & 8.062 & 0.161 & — & — & 1.920 & 1.205E-14\
CAHA\_20323055\_401356 & 20:32:30.55 & +40:13:56.2 & 18.796 & 0.015 & 15.275 & 0.01 & 13.144 & 0.017 & 11.272 & 0.046 & 10.48 & 0.033 & 9.671 & 0.127 & 8.665 & 0.228 & — & — & 3.059 & 1.572E-15\
CAHA\_20323042\_401221 & 20:32:30.42 & +40:12:21.1 & 17.321 & 0.015 & 15.434 & 0.015 & 14.292 & 0.022 & 13.077 & 0.045 & 12.658 & 0.028 & 11.364 & 0.121 & 10.082 & 0.182 & — & — & — & —\
CAHA\_20323549\_401308 & 20:32:35.49 & +40:13:08.9 & 17.794 & 0.012 & 14.456 & 0.012 & 12.912 & 0.023 & 11.898 & 0.024 & 11.703 & 0.023 & 10.857 & 0.111 & 9.75 & 0.194 & — & — & — & —\
CAHA\_20325130\_401150 & 20:32:51.30 & +40:11:50.8 & 17.119 & 0.016 & 15.336 & 0.016 & 14.289 & 0.016 & 13.299 & 0.03 & 12.742 & 0.021 & 12.286 & 0.049 & 11.36 & 0.058 & — & — & — & —\
CAHA\_20325783\_401208 & 20:32:57.83 & +40:12:08.4 & 15.826 & 0.015 & 14.792 & 0.017 & 14.297 & 0.021 & 14.107 & 0.027 & 13.942 & 0.029 & 13.696 & 0.095 & 12.871 & 0.135 & — & — & — & —\
CAHA\_20330832\_401317 & 20:33:08.32 & +40:13:17.2 & 15.914 & 0.01 & 15.062 & 0.016 & 14.513 & 0.029 & 14.045 & 0.026 & 13.567 & 0.027 & 13.146 & 0.07 & 12.456 & 0.114 & — & — & — & —\
CAHA\_20330731\_401428 & 20:33:07.31 & +40:14:28.4 & 15.442 & 0.009 & 14.386 & 0.016 & 13.868 & 0.021 & 13.555 & 0.03 & 13.252 & 0.029 & 12.772 & 0.049 & 12.064 & 0.09 & — & — & — & —\
CAHA\_20330608\_401211 & 20:33:06.08 & +40:12:11.6 & 14.922 & 0.01 & 13.97 & 0.015 & 13.365 & 0.036 & 13.081 & 0.023 & 12.807 & 0.023 & 12.479 & 0.048 & 11.86 & 0.058 & — & — & — & —\
CAHA\_20315908\_401647 & 20:31:59.08 & +40:16:47.9 & 14.435 & 0.006 & 13.58 & 0.008 & 13.221 & 0.013 & 12.687 & 0.021 & 12.142 & 0.019 & 11.856 & 0.057 & 10.977 & 0.139 & — & — & 1.570 & 7.604E-15\
CAHA\_20320234\_401726 & 20:32:02.34 & +40:17:26.9 & 18.316 & 0.011 & 15.371 & 0.007 & 13.777 & 0.008 & 12.528 & 0.023 & 11.981 & 0.02 & 11.466 & 0.051 & 10.813 & 0.088 & — & — & — & —\
CAHA\_20320067\_401621 & 20:32:00.67 & +40:16:22.0 & 14.559 & 0.004 & 13.399 & 0.006 & 12.684 & 0.014 & 12.3 & 0.024 & 12.007 & 0.023 & 11.448 & 0.053 & 10.343 & 0.062 & — & — & — & —\
CAHA\_20321121\_401703 & 20:32:11.21 & +40:17:03.5 & 16.9 & 0.004 & 15.195 & 0.006 & 14.244 & 0.011 & 12.913 & 0.022 & 12.358 & 0.02 & 11.972 & 0.076 & 11.223 & 0.171 & — & — & — & —\
CAHA\_20321075\_401624 & 20:32:10.75 & +40:16:24.8 & 13.638 & 0.004 & 12.659 & 0.006 & 11.911 & 0.018 & 11.199 & 0.016 & 10.843 & 0.016 & 10.509 & 0.019 & 10.005 & 0.057 & — & — & — & —\
CAHA\_20320290\_401724 & 20:32:02.90 & +40:17:24.4 & 16.272 & 0.005 & 15.33 & 0.027 & 14.234 & 0.021 & 12.75 & 0.04 & 12.243 & 0.033 & 11.751 & 0.074 & 11.132 & 0.178 & — & — & — & —\
CAHA\_20320046\_401506 & 20:32:00.46 & +40:15:06.1 & 15.626 & 0.006 & 13.109 & 0.009 & 12.109 & 0.018 & 11.339 & 0.023 & 11.244 & 0.023 & 10.875 & 0.043 & 10.259 & 0.107 & — & — & — & —\
CAHA\_20320665\_401519 & 20:32:06.65 & +40:15:20.0 & 14.918 & 0.005 & 12.888 & 0.007 & 11.834 & 0.02 & 10.679 & 0.016 & 10.308 & 0.016 & 9.932 & 0.038 & 9.529 & 0.117 & — & — & — & —\
CAHA\_20320783\_401636 & 20:32:07.83 & +40:16:36.3 & 15.209 & 0.004 & 14.078 & 0.007 & 13.417 & 0.014 & 12.87 & 0.019 & 12.576 & 0.019 & 12.233 & 0.058 & 11.23 & 0.176 & — & — & — & —\
CAHA\_20320403\_401828 & 20:32:04.03 & +40:18:28.1 & 20.924 & 0.058 & 17.029 & 0.012 & 14.745 & 0.008 & 13.107 & 0.033 & 12.324 & 0.023 & 11.886 & 0.098 & 10.963 & 0.229 & — & — & — & —\
CAHA\_20320309\_401738 & 20:32:03.09 & +40:17:38.4 & 19.014 & 0.018 & 15.72 & 0.007 & 13.95 & 0.009 & 12.491 & 0.021 & 11.803 & 0.02 & 11.231 & 0.048 & 10.465 & 0.089 & — & — & — & —\
CAHA\_20320807\_401715 & 20:32:08.07 & +40:17:15.7 & 14.595 & 0.003 & 13.256 & 0.006 & 12.477 & 0.014 & 11.741 & 0.016 & 11.405 & 0.016 & 11.075 & 0.033 & 10.171 & 0.121 & — & — & — & —\
CAHA\_20320848\_401550 & 20:32:08.48 & +40:15:50.7 & 16.506 & 0.003 & 15.376 & 0.006 & 14.944 & 0.015 & 14.359 & 0.061 & 14.224 & 0.063 & 12.389 & 0.102 & 10.463 & 0.09 & — & — & — & —\
CAHA\_20322058\_401825 & 20:32:20.58 & +40:18:25.4 & 18.092 & 0.007 & 15.837 & 0.012 & 14.668 & 0.023 & 12.748 & 0.023 & 12.0 & 0.02 & 11.443 & 0.074 & 10.774 & 0.114 & — & — & — & —\
CAHA\_20322443\_401818 & 20:32:24.43 & +40:18:18.4 & 16.859 & 0.024 & 14.573 & 0.024 & 13.207 & 0.024 & 11.911 & 0.021 & 11.259 & 0.02 & 10.694 & 0.083 & 9.749 & 0.196 & — & — & 2.913 & 1.306E-15\
CAHA\_20322935\_401844 & 20:32:29.35 & +40:18:44.7 & 14.703 & 0.003 & 13.689 & 0.005 & 13.172 & 0.004 & 12.806 & 0.021 & 12.618 & 0.024 & 12.336 & 0.124 & 11.746 & 0.24 & — & — & — & —\
CAHA\_20321655\_401808 & 20:32:16.55 & +40:18:08.8 & 15.139 & 0.003 & 13.688 & 0.007 & 12.914 & 0.005 & 12.245 & 0.022 & 11.862 & 0.019 & 11.021 & 0.072 & 9.779 & 0.069 & — & — & — & —\
CAHA\_20323555\_401840 & 20:32:35.55 & +40:18:40.3 & 18.715 & 0.015 & 15.218 & 0.005 & 13.545 & 0.006 & 12.638 & 0.022 & 12.426 & 0.023 & 12.034 & 0.112 & 11.721 & 0.239 & — & — & — & —\
CAHA\_20323489\_401810 & 20:32:34.89 & +40:18:11.0 & 13.619 & 0.004 & 12.464 & 0.005 & 11.896 & 0.021 & 10.95 & 0.016 & 10.578 & 0.016 & 10.286 & 0.043 & 9.551 & 0.082 & — & — & 1.219 & 1.186E-15\
CAHA\_20323249\_401602 & 20:32:32.49 & +40:16:02.9 & 17.053 & 0.027 & 13.156 & 0.007 & 10.815 & 0.017 & 8.458 & 0.017 & 7.656 & 0.017 & 7.065 & 0.034 & 6.613 & 0.093 & — & — & — & —\
CAHA\_20323176\_401616 & 20:32:31.76 & +40:16:16.5 & 14.715 & 0.005 & 11.98 & 0.018 & 10.456 & 0.017 & 9.25 & 0.031 & 8.842 & 0.054 & 7.789 & 0.142 & 5.961 & 0.131 & — & — & — & —\
CAHA\_20324131\_401807 & 20:32:41.31 & +40:18:07.2 & 17.244 & 0.005 & 15.14 & 0.007 & 14.169 & 0.007 & 13.383 & 0.021 & 12.98 & 0.022 & 12.605 & 0.101 & 11.901 & 0.207 & — & — & — & —\
CAHA\_20323108\_401608 & 20:32:31.08 & +40:16:08.6 & 16.236 & 0.009 & 13.524 & 0.005 & 12.042 & 0.032 & 10.022 & 0.044 & 9.314 & 0.046 & 8.55 & 0.133 & 6.872 & 0.167 & — & — & — & —\
CAHA\_20324014\_401812 & 20:32:40.14 & +40:18:12.7 & 16.204 & 0.004 & 14.726 & 0.007 & 13.956 & 0.005 & 13.485 & 0.029 & 13.178 & 0.035 & 12.759 & 0.13 & 11.857 & 0.213 & — & — & — & —\
CAHA\_20323095\_401649 & 20:32:30.95 & +40:16:49.6 & 13.021 & 0.004 & 10.756 & 0.022 & 9.763 & 0.017 & 9.074 & 0.032 & 8.531 & 0.031 & 7.958 & 0.127 & 6.609 & 0.208 & — & — & 2.708 & 6.967E-15\
CAHA\_20323708\_401737 & 20:32:37.08 & +40:17:37.5 & 13.577 & 0.004 & 11.519 & 0.031 & 10.529 & 0.026 & 9.853 & 0.023 & 9.483 & 0.027 & 8.839 & 0.084 & 7.252 & 0.093 & — & — & 2.110 & 3.729E-15\
CAHA\_20323555\_401605 & 20:32:35.55 & +40:16:05.8 & 13.195 & 0.007 & 10.399 & 0.022 & 8.493 & 0.017 & 6.322 & 0.016 & 5.666 & 0.015 & 5.036 & 0.021 & 4.051 & 0.038 & — & — & — & —\
CAHA\_20330586\_401552 & 20:33:05.86 & +40:15:52.9 & 17.201 & 0.014 & 15.822 & 0.028 & 15.098 & 0.043 & 14.265 & 0.027 & 13.921 & 0.024 & 13.669 & 0.099 & 13.24 & 0.204 & — & — & — & —\
CAHA\_20320906\_402043 & 20:32:09.06 & +40:20:43.0 & 14.099 & 0.002 & 12.956 & 0.006 & 12.338 & 0.004 & 12.017 & 0.017 & 11.927 & 0.018 & 11.576 & 0.078 & 10.78 & 0.15 & — & — & — & —\
CAHA\_20321381\_401908 & 20:32:13.81 & +40:19:08.6 & 20.069 & 0.057 & 16.556 & 0.007 & 14.37 & 0.006 & 12.346 & 0.092 & 11.718 & 0.054 & 10.147 & 0.178 & 8.759 & 0.227 & — & — & — & —\
CAHA\_20322602\_401904 & 20:32:26.02 & +40:19:04.3 & 18.818 & 9.999 & 16.463 & 0.008 & 14.207 & 0.005 & 12.653 & 0.017 & 12.093 & 0.017 & 11.66 & 0.049 & 11.05 & 0.116 & — & — & — & —\
CAHA\_20322626\_402216 & 20:32:26.26 & +40:22:16.8 & 15.245 & 0.003 & 13.693 & 0.007 & 12.927 & 0.004 & 12.298 & 0.017 & 11.832 & 0.016 & 11.483 & 0.035 & 10.776 & 0.072 & — & — & — & —\
CAHA\_20322761\_401914 & 20:32:27.61 & +40:19:14.5 & 20.543 & 0.057 & 16.125 & 0.006 & 13.617 & 0.005 & 12.102 & 0.017 & 11.45 & 0.017 & 11.021 & 0.045 & 10.301 & 0.058 & — & — & — & —\
CAHA\_20322942\_401917 & 20:32:29.42 & +40:19:17.2 & 19.044 & 0.015 & 16.167 & 0.005 & 14.449 & 0.009 & 13.061 & 0.02 & 12.395 & 0.021 & 12.073 & 0.054 & 11.34 & 0.107 & — & — & — & —\
CAHA\_20322974\_401906 & 20:32:29.74 & +40:19:06.3 & 18.843 & 0.012 & 16.098 & 0.007 & 14.651 & 0.007 & 13.529 & 0.025 & 12.964 & 0.023 & 12.437 & 0.067 & 11.733 & 0.081 & — & — & — & —\
CAHA\_20322536\_402054 & 20:32:25.36 & +40:20:54.5 & 14.601 & 0.003 & 12.966 & 0.006 & 12.168 & 0.004 & 11.392 & 0.017 & 11.041 & 0.016 & 10.586 & 0.019 & 9.925 & 0.055 & — & — & — & —\
CAHA\_20322352\_402035 & 20:32:23.52 & +40:20:35.2 & 19.808 & 0.038 & 16.256 & 0.007 & 14.394 & 0.007 & 13.064 & 0.023 & 12.391 & 0.02 & 11.636 & 0.059 & 10.866 & 0.08 & — & — & — & —\
CAHA\_20322337\_402212 & 20:32:23.37 & +40:22:12.2 & 15.341 & 0.003 & 13.874 & 0.008 & 13.182 & 0.004 & 12.554 & 0.017 & 12.186 & 0.017 & 11.763 & 0.028 & 10.759 & 0.04 & — & — & — & —\
CAHA\_20321759\_402226 & 20:32:17.59 & +40:22:26.4 & 17.119 & 0.006 & 15.531 & 0.008 & 14.738 & 0.006 & 13.952 & 0.028 & 13.531 & 0.024 & 13.111 & 0.081 & 12.349 & 0.072 & — & — & — & —\
CAHA\_20321778\_401905 & 20:32:17.78 & +40:19:05.8 & 15.883 & 0.003 & 13.723 & 0.006 & 12.612 & 0.005 & 11.801 & 0.035 & 11.492 & 0.029 & 10.374 & 0.156 & 9.075 & 0.23 & — & — & — & —\
CAHA\_20321874\_402003 & 20:32:18.74 & +40:20:03.2 & 19.097 & 0.014 & 16.012 & 0.007 & 14.24 & 0.007 & 12.801 & 0.018 & 12.146 & 0.023 & 11.508 & 0.063 & 10.417 & 0.126 & — & — & — & —\
CAHA\_20321559\_402131 & 20:32:15.59 & +40:21:31.2 & 17.573 & 0.005 & 15.115 & 0.007 & 13.785 & 0.004 & 13.103 & 0.019 & 12.853 & 0.02 & 12.689 & 0.113 & 11.855 & 0.246 & — & — & — & —\
CAHA\_20323947\_402116 & 20:32:39.47 & +40:21:16.5 & 15.868 & 0.003 & 14.169 & 0.006 & 13.308 & 0.005 & 12.546 & 0.018 & 12.045 & 0.017 & 11.555 & 0.034 & 10.922 & 0.063 & — & — & — & —\
CAHA\_20324423\_401939 & 20:32:44.23 & +40:19:39.4 & 16.834 & 0.004 & 15.494 & 0.007 & 14.667 & 0.009 & 13.761 & 0.023 & 13.168 & 0.022 & 12.481 & 0.054 & 11.586 & 0.071 & — & — & — & —\
CAHA\_20323153\_402048 & 20:32:31.53 & +40:20:49.0 & 18.767 & 0.018 & 15.671 & 0.006 & 14.18 & 0.009 & 13.323 & 0.02 & 13.097 & 0.021 & 12.761 & 0.058 & 12.403 & 0.122 & — & — & — & —\
CAHA\_20323570\_402141 & 20:32:35.70 & +40:21:41.9 & 18.206 & 0.012 & 15.383 & 0.006 & 14.069 & 0.006 & 13.276 & 0.019 & 13.112 & 0.02 & 12.83 & 0.071 & 12.158 & 0.155 & — & — & — & —\
CAHA\_20323809\_401943 & 20:32:38.09 & +40:19:43.7 & 18.638 & 0.015 & 15.449 & 0.006 & 13.945 & 0.007 & 13.14 & 0.02 & 12.916 & 0.02 & 12.593 & 0.078 & 12.069 & 0.157 & — & — & — & —\
CAHA\_20323774\_401916 & 20:32:37.74 & +40:19:16.4 & 20.18 & 0.044 & 16.788 & 0.011 & 14.861 & 0.006 & 13.415 & 0.021 & 12.829 & 0.021 & 12.384 & 0.095 & 11.399 & 0.107 & — & — & — & —\
CAHA\_20325212\_401914 & 20:32:52.12 & +40:19:14.8 & 17.619 & 0.004 & 15.916 & 0.008 & 15.151 & 0.012 & 14.45 & 0.043 & 13.974 & 0.038 & 13.495 & 0.118 & 12.946 & 0.175 & — & — & — & —\
CAHA\_20325635\_402129 & 20:32:56.35 & +40:21:29.0 & 17.837 & 0.008 & 15.611 & 0.007 & 14.432 & 0.009 & 13.374 & 0.019 & 12.82 & 0.017 & 12.432 & 0.035 & 11.756 & 0.056 & — & — & — & —\
CAHA\_20315935\_402411 & 20:31:59.35 & +40:24:11.4 & 15.12 & 0.006 & 13.917 & 0.012 & 13.485 & 0.043 & 12.841 & 0.018 & 12.476 & 0.017 & 12.116 & 0.032 & 11.572 & 0.049 & — & — & — & —\
CAHA\_20315954\_402231 & 20:31:59.54 & +40:22:31.8 & 16.638 & 0.004 & 15.201 & 0.008 & 14.36 & 0.006 & 13.77 & 0.024 & 13.34 & 0.024 & 13.058 & 0.063 & 12.413 & 0.135 & — & — & — & —\
CAHA\_20320915\_402301 & 20:32:09.15 & +40:23:01.9 & 16.366 & 0.004 & 14.708 & 0.009 & 13.754 & 0.004 & 12.628 & 0.022 & 12.111 & 0.019 & 11.678 & 0.047 & 10.876 & 0.095 & — & — & — & —\
CAHA\_20320939\_402250 & 20:32:09.39 & +40:22:50.6 & 13.076 & 0.004 & 11.106 & 0.023 & 10.307 & 0.017 & 9.56 & 0.016 & 9.078 & 0.016 & 8.597 & 0.016 & 7.961 & 0.017 & — & — & 1.920 & 3.140E-14\
CAHA\_20320112\_402313 & 20:32:01.12 & +40:23:13.2 & 17.417 & 0.019 & 15.762 & 0.03 & 14.796 & 0.039 & 13.528 & 0.019 & 12.909 & 0.017 & 12.462 & 0.047 & 11.808 & 0.077 & — & — & — & —\
CAHA\_20322465\_402329 & 20:32:24.65 & +40:23:29.2 & 17.824 & 0.008 & 15.397 & 0.009 & 14.258 & 0.006 & 13.65 & 0.023 & 13.479 & 0.023 & 13.182 & 0.062 & 12.632 & 0.166 & — & — & — & —\
CAHA\_20324555\_402331 & 20:32:45.55 & +40:23:31.7 & 17.385 & 0.006 & 15.647 & 0.007 & 14.723 & 0.005 & 14.097 & 0.024 & 13.687 & 0.023 & 13.609 & 0.115 & 12.844 & 0.202 & — & — & — & —\
CAHA\_20324550\_402231 & 20:32:45.50 & +40:22:31.7 & 16.875 & 0.004 & 15.837 & 0.007 & 15.174 & 0.005 & 14.419 & 0.038 & 13.977 & 0.033 & 13.046 & 0.059 & 12.014 & 0.091 & — & — & — & —\
CAHA\_20330316\_402332 & 20:33:03.16 & +40:23:33.0 & 18.123 & 0.031 & 15.717 & 0.026 & 14.347 & 0.018 & 12.893 & 0.019 & 12.421 & 0.017 & 12.128 & 0.046 & 11.701 & 0.117 & — & — & — & —\
TWOM\_20324634\_4009048 & 20:32:46.35 & +40:09:04.9 & 18.49 & 9.999 & 14.906 & 0.059 & 13.206 & 0.036 & 12.13 & 0.017 & 11.948 & 0.017 & 11.546 & 0.038 & 11.182 & 0.062 & — & — & — & —\
TWOM\_20315282\_4012188 & 20:31:52.82 & +40:12:18.8 & 13.994 & 0.027 & 13.068 & 0.029 & 12.612 & 0.026 & 12.05 & 0.017 & 11.775 & 0.017 & 11.533 & 0.027 & 11.024 & 0.046 & — & — & — & —\
TWOM\_20322814\_4017148 & 20:32:28.14 & +40:17:14.9 & 10.966 & 0.035 & 9.926 & 0.038 & 9.336 & 0.024 & 8.985 & 0.034 & 8.541 & 0.046 & 7.94 & 0.172 & 6.448 & 0.242 & — & — & — & —\
TWOM\_20331278\_4018418 & 20:33:12.79 & +40:18:41.8 & 18.61 & 9.999 & 15.771 & 0.13 & 14.561 & 0.103 & 13.589 & 0.029 & 13.368 & 0.024 & 13.103 & 0.098 & 12.477 & 0.185 & — & — & — & —\
CAHA\_20325255\_401152 & 20:32:52.55 & +40:11:52.0 & 19.187 & 0.034 & 17.342 & 0.018 & 16.09 & 0.017 & 14.485 & 0.027 & 13.904 & 0.024 & 13.684 & 0.111 & 12.616 & 0.157 & — & — & — & —\
CAHA\_20321844\_401724 & 20:32:18.23 & +40:17:27.7 & 18.294 & 0.033 & 16.583 & 0.015 & 15.157 & 0.007 & 13.024 & 0.026 & 12.173 & 0.019 & 11.537 & 0.046 & 10.547 & 0.045 & — & — & — & —\
CAHA\_20322275\_401739 & 20:32:22.75 & +40:17:39.6 & 20.249 & 0.058 & 18.616 & 0.064 & 17.053 & 0.089 & 12.84 & 0.153 & 12.698 & 0.187 & 9.895 & 0.171 & 8.093 & 0.184 & — & — & — & —\
CAHA\_20324785\_401817 & 20:32:47.85 & +40:18:17.7 & 20.327 & 0.059 & 16.924 & 0.008 & 15.097 & 0.008 & 14.101 & 0.04 & 13.779 & 0.046 & 13.599 & 0.16 & 12.532 & 0.219 & — & — & — & —\
CAHA\_20324105\_401849 & 20:32:41.05 & +40:18:49.3 & 18.836 & 0.017 & 16.75 & 0.007 & 15.68 & 0.012 & 14.572 & 0.041 & 14.287 & 0.038 & 13.87 & 0.173 & 12.149 & 0.14 & — & — & — & —\
CAHA\_20325040\_401850 & 20:32:50.40 & +40:18:50.4 & 17.912 & 0.009 & 16.288 & 0.01 & 15.384 & 0.01 & 14.69 & 0.056 & 14.384 & 0.051 & 13.923 & 0.118 & 13.129 & 0.199 & — & — & — & —\
CAHA\_20322563\_401850 & 20:32:25.63 & +40:18:50.8 & 20.371 & 0.064 & 17.687 & 0.01 & 16.093 & 0.02 & 14.339 & 0.037 & 13.47 & 0.029 & 13.026 & 0.19 & 11.419 & 0.218 & — & — & — & —\
CAHA\_20330273\_401903 & 20:33:02.73 & +40:19:03.2 & 18.929 & 0.019 & 16.783 & 0.015 & 15.583 & 0.011 & 14.568 & 0.054 & 14.16 & 0.054 & 13.824 & 0.121 & 12.891 & 0.146 & — & — & — & —\
CAHA\_20323802\_401934 & 20:32:38.02 & +40:19:34.4 & 20.414 & 0.051 & 17.149 & 0.012 & 15.2 & 0.009 & 13.652 & 0.023 & 13.145 & 0.021 & 12.602 & 0.075 & 11.952 & 0.115 & — & — & — & —\
CAHA\_20320973\_402253 & 20:32:09.73 & +40:22:53.4 & 13.802 & 0.004 & 12.505 & 0.009 & — & — & 10.709 & 0.017 & 10.351 & 0.017 & 9.975 & 0.02 & 9.154 & 0.025 & — & — & — & —\
CAHA\_20315905\_401755 & 20:31:59.05 & +40:17:56.0 & — & — & 18.606 & 0.03 & 16.205 & 0.017 & 14.226 & 0.057 & 13.447 & 0.055 & 12.447 & 0.145 & 11.06 & 0.236 & — & — & — & —\
CAHA\_20322723\_401923 & 20:32:27.23 & +40:19:23.1 & — & — & 17.445 & 0.011 & 14.791 & 0.008 & 12.988 & 0.029 & 12.078 & 0.03 & 11.483 & 0.095 & 10.825 & 0.189 & — & — & — & —\
CAHA\_20321904\_401942 & 20:32:19.04 & +40:19:42.4 & — & — & 17.846 & 0.011 & 14.913 & 0.009 & 12.756 & 0.017 & 12.021 & 0.018 & 11.126 & 0.068 & 9.878 & 0.125 & — & — & — & —\
CAHA\_20321823\_402039 & 20:32:18.23 & +40:20:39.4 & — & — & 19.331 & 0.053 & 15.653 & 0.02 & 12.337 & 0.017 & 11.325 & 0.016 & 10.716 & 0.061 & 10.249 & 0.167 & — & — & — & —\
CAHA\_20321538\_402356 & 20:32:15.38 & +40:23:56.9 & — & — & 17.495 & 0.023 & 15.832 & 0.034 & 14.414 & 0.024 & 13.942 & 0.025 & 13.44 & 0.08 & 12.978 & 0.241 & — & — & — & —\
CAHA\_20330081\_401020 & 20:33:00.81 & +40:10:20.1 & — & — & — & — & 18.342 & 0.121 & 15.457 & 0.055 & 14.702 & 0.061 & 13.348 & 0.082 & 12.192 & 0.091 & — & — & — & —\
CAHA\_20322687\_401910 & 20:32:26.87 & +40:19:10.3 & — & — & — & — & 12.997 & 0.006 & 11.497 & 0.016 & 10.941 & 0.016 & 10.605 & 0.031 & 10.164 & 0.069 & — & — & 1.832 & 3.097E-15\
CAHA\_20322007\_401933 & 20:32:20.07 & +40:19:33.6 & — & — & — & — & 16.563 & 0.02 & 13.016 & 0.041 & 11.857 & 0.027 & 11.221 & 0.13 & 9.785 & 0.223 & — & — & — & —\
SSTCYGX\_J203152.69\_401840.5 & 20:31:52.68 & +40:18:40.5 & — & — & — & — & — & — & 13.805 & 0.022 & 12.985 & 0.021 & 12.558 & 0.048 & 12.053 & 0.175 & — & — & — & —\
CAHA\_20321015\_401847 & 20:32:10.15 & +40:18:48.0 & — & — & 17.561 & 0.014 & 14.716 & 0.008 & 13.155 & 0.074 & 12.549 & 0.065 & 11.844 & 0.44 & 10.831 & 0.807 & — & — & 2.431 & 2.387E-15\
CAHA\_20321798\_401922 & 20:32:17.98 & +40:19:22.3 & 19.178 & 0.018 & 15.984 & 0.006 & 14.341 & 0.006 & 12.932 & 0.077 & 12.54 & 0.056 & 11.309 & 0.362 & — & — & — & — & 3.409 & 1.067E-14\
CAHA\_20321061\_401904 & 20:32:10.61 & +40:19:04.8 & — & — & — & — & 17.616 & 0.05 & 14.017 & 0.067 & 13.097 & 0.081 & 11.677 & 0.173 & — & — & — & — & 3.176 & 1.357E-14\
CAHA\_20315867\_401915 & 20:31:58.67 & +40:19:15.1 & 17.58 & 0.006 & 14.666 & 0.005 & 12.948 & 0.008 & 11.306 & 0.016 & 10.714 & 0.016 & 10.313 & 0.019 & 9.896 & 0.045 & 7.262 & 0.41 & 1.862 & 9.616E-15\
CAHA\_20320503\_401715 & 20:32:05.03 & +40:17:16.0 & 12.833 & 0.005 & 11.862 & 0.019 & 11.351 & 0.016 & 10.564 & 0.016 & 10.174 & 0.015 & 9.762 & 0.02 & 8.992 & 0.046 & 6.837 & 0.749 & 1.380 & 6.313E-15\
CAHA\_20320542\_402126 & 20:32:05.42 & +40:21:26.2 & 14.416 & 0.002 & 13.22 & 0.006 & 12.679 & 0.005 & 12.127 & 0.023 & 11.936 & 0.021 & 11.7 & 0.133 & 11.008 & 0.356 & 6.932 & 0.505 & 1.920 & 1.509E-14\
CAHA\_20322109\_401848 & 20:32:21.09 & +40:18:48.1 & 17.17 & 0.005 & 14.563 & 0.005 & 13.189 & 0.004 & 11.819 & 0.031 & 11.28 & 0.023 & 10.886 & 0.187 & 10.03 & 0.32 & — & — & 2.314 & 2.492E-15\
CAHA\_20322641\_401515 & 20:32:26.41 & +40:15:15.6 & 15.269 & 0.007 & 12.963 & 0.007 & 11.691 & 0.018 & 10.435 & 0.025 & 10.028 & 0.025 & 9.808 & 0.219 & 9.412 & 0.829 & — & — & 1.920 & 1.220E-15\
CAHA\_20323239\_401643 & 20:32:32.39 & +40:16:43.3 & 15.99 & 0.005 & 13.31 & 0.005 & 11.504 & 0.026 & 10.049 & 0.071 & 9.374 & 0.043 & 9.051 & 0.477 & — & — & — & — & 3.059 & 7.560E-15\
CAHA\_20323582\_401745 & 20:32:35.82 & +40:17:45.0 & 13.798 & 0.004 & 12.611 & 0.008 & 12.108 & 0.033 & 11.136 & 0.049 & 10.972 & 0.043 & 9.44 & 0.243 & — & — & — & — & 2.781 & 6.038E-15\
CAHA\_20324487\_401834 & 20:32:44.87 & +40:18:34.1 & 15.417 & 0.005 & 13.309 & 0.008 & 12.165 & 0.021 & 11.274 & 0.016 & 10.967 & 0.016 & 10.707 & 0.022 & 10.187 & 0.048 & 7.054 & 0.322 & 2.256 & 8.162E-15\
CAHA\_20325504\_401617 & 20:32:55.04 & +40:16:17.3 & 13.48 & 0.009 & 12.375 & 0.021 & 11.68 & 0.019 & 10.542 & 0.015 & 10.282 & 0.015 & 10.11 & 0.017 & 9.77 & 0.017 & 6.073 & 0.163 & 1.526 & 1.505E-14\
CAHA\_20330681\_401337 & 20:33:06.81 & +40:13:37.9 & 13.464 & 0.017 & 12.831 & 0.024 & 12.525 & 0.027 & 11.991 & 0.016 & 11.836 & 0.017 & 11.689 & 0.027 & 11.2 & 0.044 & 7.413 & 0.258 & 1.613 & 1.026E-14\
\[tab:ClassII\]
[lllllllllllllllllllll]{}
CAHA\_20322152\_401104 & 20:32:21.52 & +40:11:04.6 & 13.842 & 0.016 & 13.002 & 0.015 & 12.596 & 0.028 & 12.559 & 0.019 & 12.548 & 0.021 & 12.395 & 0.078 & 11.608 & 0.169 & — & — & 2.548 & 6.699E-15\
CAHA\_20323285\_401056 & 20:32:32.85 & +40:10:56.5 & 16.605 & 0.017 & 15.081 & 0.012 & 14.321 & 0.02 & 14.065 & 0.232 & 14.028 & 0.171 & — & — & — & — & — & — & 3.438 & 6.162E-15\
CAHA\_20320491\_401359 & 20:32:04.91 & +40:13:59.3 & 13.227 & 0.008 & 12.717 & 0.023 & 12.476 & 0.024 & 12.571 & 0.021 & 12.592 & 0.021 & 12.509 & 0.099 & — & — & — & — & 1.190 & 5.738E-15\
CAHA\_20321916\_401318 & 20:32:19.16 & +40:13:18.2 & 15.919 & 0.011 & 14.424 & 0.01 & 13.758 & 0.025 & 13.357 & 0.055 & 13.247 & 0.067 & 13.21 & 0.55 & — & — & — & — & 2.489 & 3.258E-15\
CAHA\_20323784\_401352 & 20:32:37.84 & +40:13:52.7 & 15.577 & 0.014 & 14.011 & 0.027 & 13.419 & 0.021 & 13.344 & 0.233 & 13.284 & 0.25 & — & — & — & — & — & — & 3.263 & 1.022E-14\
CAHA\_20324208\_401220 & 20:32:42.08 & +40:12:20.3 & 14.24 & 0.014 & 13.071 & 0.017 & 12.52 & 0.022 & 12.381 & 0.019 & 12.255 & 0.02 & 12.494 & 0.123 & — & — & — & — & 1.905 & 7.653E-15\
CAHA\_20323075\_401419 & 20:32:30.75 & +40:14:19.5 & 15.255 & 0.028 & 12.881 & 0.044 & 11.66 & 0.025 & 10.803 & 0.072 & 10.597 & 0.071 & — & — & — & — & — & — & 2.694 & 1.592E-14\
CAHA\_20325387\_401420 & 20:32:53.87 & +40:14:20.4 & 13.2 & 0.015 & 12.069 & 0.019 & 11.614 & 0.018 & 11.383 & 0.016 & 11.305 & 0.016 & 11.215 & 0.024 & 11.149 & 0.149 & — & — & 1.818 & 2.580E-14\
CAHA\_20325533\_401440 & 20:32:55.33 & +40:14:40.3 & 15.64 & 0.014 & 14.432 & 0.015 & 13.914 & 0.014 & 13.712 & 0.023 & 13.654 & 0.027 & 13.552 & 0.093 & — & — & — & — & 2.212 & 8.545E-15\
CAHA\_20320256\_401740 & 20:32:02.56 & +40:17:40.5 & 13.948 & 0.005 & 13.094 & 0.006 & 12.814 & 0.01 & 12.755 & 0.029 & 12.633 & 0.034 & 12.802 & 0.247 & — & — & — & — & 1.467 & 3.822E-15\
CAHA\_20320237\_401701 & 20:32:02.37 & +40:17:01.1 & 13.188 & 0.003 & 12.005 & 0.039 & 11.556 & 0.028 & 11.155 & 0.016 & 11.008 & 0.016 & 10.808 & 0.034 & 10.338 & 0.081 & — & — & 1.511 & 2.660E-14\
CAHA\_20321032\_401752 & 20:32:10.32 & +40:17:53.0 & 14.269 & 0.003 & 13.309 & 0.006 & 12.949 & 0.009 & 12.646 & 0.03 & 12.564 & 0.043 & 12.658 & 0.378 & 12.546 & 1.721 & — & — & 1.073 & 3.459E-15\
CAHA\_20321962\_401812 & 20:32:19.62 & +40:18:12.4 & 13.391 & 0.004 & 12.21 & 0.027 & 11.83 & 0.021 & 11.669 & 0.017 & 11.53 & 0.017 & 11.415 & 0.064 & 11.042 & 0.159 & — & — & 1.292 & 3.995E-15\
CAHA\_20322560\_401805 & 20:32:25.60 & +40:18:05.3 & 19.153 & 0.02 & 15.839 & 0.005 & 14.207 & 0.004 & 13.319 & 0.063 & 13.104 & 0.09 & 13.092 & 0.671 & — & — & — & — & 3.570 & 2.168E-13\
CAHA\_20323482\_401818 & 20:32:34.82 & +40:18:18.4 & 14.571 & 0.021 & 13.428 & 0.028 & 12.995 & 0.029 & 12.655 & 0.027 & 12.56 & 0.037 & 12.213 & 0.136 & — & — & — & — & 1.438 & 4.581E-15\
CAHA\_20324203\_401510 & 20:32:42.03 & +40:15:10.2 & 14.755 & 0.009 & 13.799 & 0.011 & 13.423 & 0.013 & 13.169 & 0.111 & 13.281 & 0.12 & — & — & — & — & — & — & 1.175 & 3.020E-15\
CAHA\_20323083\_401706 & 20:32:30.83 & +40:17:06.1 & 15.738 & 0.006 & 13.446 & 0.007 & 12.33 & 0.009 & 11.896 & 0.078 & 11.932 & 0.113 & — & — & — & — & — & — & 3.380 & 5.798E-14\
CAHA\_20323043\_401643 & 20:32:30.43 & +40:16:43.2 & 15.525 & 0.003 & 13.049 & 0.006 & 11.933 & 0.026 & 11.182 & 0.055 & 11.031 & 0.106 & — & — & — & — & — & — & 2.716 & 7.643E-15\
CAHA\_20325412\_401547 & 20:32:54.12 & +40:15:47.9 & 17.073 & 0.009 & 15.642 & 0.013 & 15.02 & 0.012 & 14.742 & 0.041 & 14.508 & 0.052 & — & — & — & — & — & — & 1.978 & 1.2521E-14\
CAHA\_20325095\_401520 & 20:32:50.95 & +40:15:20.2 & 14.535 & 0.01 & 13.575 & 0.013 & 13.178 & 0.008 & 12.989 & 0.022 & 12.938 & 0.028 & 12.74 & 0.112 & — & — & — & — & 1.219 & 8.092E-15\
CAHA\_20315441\_402153 & 20:31:54.41 & +40:21:53.2 & 13.824 & 0.005 & 12.53 & 0.009 & 12.024 & 0.018 & 11.721 & 0.016 & 11.667 & 0.016 & 11.605 & 0.027 & 11.665 & 0.093 & — & — & 1.745 & 8.670E-15\
CAHA\_20315532\_402127 & 20:31:55.32 & +40:21:27.6 & 14.938 & 0.002 & 13.572 & 0.008 & 12.999 & 0.005 & 12.716 & 0.02 & 12.601 & 0.021 & 12.687 & 0.105 & 12.069 & 0.322 & — & — & 1.657 & 3.869E-15\
CAHA\_20320065\_401931 & 20:32:00.65 & +40:19:31.8 & 17.586 & 0.005 & 15.52 & 0.006 & 14.493 & 0.008 & 13.831 & 0.039 & 13.648 & 0.036 & 13.508 & 0.404 & — & — & — & — & 2.022 & 5.032E-15\
CAHA\_20320051\_402020 & 20:32:00.51 & +40:20:20.2 & 13.538 & 0.002 & 12.756 & 0.006 & 12.468 & 0.005 & 12.438 & 0.033 & 12.417 & 0.044 & 12.317 & 0.12 & 11.485 & 0.299 & — & — & 1.365 & 5.296E-15\
CAHA\_20320956\_401901 & 20:32:09.56 & +40:19:01.3 & 18.248 & 0.009 & 13.633 & 0.006 & 11.067 & 0.016 & 9.439 & 0.017 & 8.995 & 0.017 & 8.7 & 0.039 & 8.721 & 0.137 & — & — & 2.533 & 4.274E-15\
CAHA\_20320899\_401937 & 20:32:08.99 & +40:19:37.3 & 13.226 & 0.003 & 12.039 & 0.021 & 11.586 & 0.016 & 11.278 & 0.021 & 11.204 & 0.029 & 11.309 & 0.203 & 12.854 & 3.984 & — & — & 1.803 & 3.219E-14\
CAHA\_20321374\_401856 & 20:32:13.74 & +40:18:57.0 & 18.195 & 0.007 & 14.942 & 0.007 & 13.283 & 0.005 & 12.483 & 0.081 & 12.185 & 0.058 & 11.035 & 0.227 & — & — & — & — & 2.548 & 1.6081E-14\
CAHA\_20322030\_401901 & 20:32:20.30 & +40:19:01.8 & 15.208 & 0.003 & 13.298 & 0.005 & 12.42 & 0.005 & 11.744 & 0.034 & 11.482 & 0.027 & 11.051 & 0.239 & — & — & — & — & 2.621 & 8.350E-15\
CAHA\_20322943\_401858 & 20:32:29.43 & +40:18:58.9 & 15.257 & 0.011 & 13.795 & 0.007 & 13.134 & 0.006 & 12.803 & 0.023 & 12.699 & 0.025 & 12.67 & 0.107 & 12.967 & 0.732 & — & — & 1.584 & 2.927E-15\
CAHA\_20322803\_401927 & 20:32:28.03 & +40:19:27.6 & 15.742 & 0.012 & 13.726 & 0.017 & 12.758 & 0.024 & 12.166 & 0.021 & 11.895 & 0.042 & 11.74 & 0.115 & — & — & — & — & 1.584 & 4.456E-15\
CAHA\_20321720\_401910 & 20:32:17.20 & +40:19:10.0 & 15.663 & 0.003 & 15.001 & 0.006 & 14.695 & 0.006 & 14.39 & 0.175 & 14.528 & 0.198 & — & — & — & — & — & — & 1.584 & 3.054E-15\
CAHA\_20324072\_402146 & 20:32:40.72 & +40:21:46.2 & 14.345 & 0.003 & 12.705 & 0.006 & 12.007 & 0.022 & 11.619 & 0.016 & 11.456 & 0.016 & 11.386 & 0.027 & 11.445 & 0.059 & — & — & 2.227 & 1.829E-14\
CAHA\_20323427\_401851 & 20:32:34.27 & +40:18:51.1 & 15.769 & 0.005 & 13.771 & 0.005 & 12.887 & 0.006 & 12.478 & 0.017 & 12.353 & 0.019 & 12.112 & 0.08 & 12.208 & 0.269 & — & — & 1.716 & 1.399E-15\
CAHA\_20320116\_402233 & 20:32:01.16 & +40:22:33.9 & 15.79 & 0.002 & 14.135 & 0.008 & 13.341 & 0.006 & 12.809 & 0.02 & 12.754 & 0.021 & 12.698 & 0.059 & 13.085 & 0.224 & — & — & 1.978 & 3.314E-15\
TWOM\_20324184\_4014001 & 20:32:41.84 & +40:14:00.2 & 11.113 & 0.021 & 10.838 & 0.018 & 10.789 & 0.017 & 10.793 & 0.019 & 10.816 & 0.021 & 10.639 & 0.116 & 10.248 & 0.302 & — & — & 0.883 & 4.458E-15\
TWOM\_20330733\_4014348 & 20:33:07.33 & +40:14:34.8 & 11.646 & 0.022 & 11.18 & 0.019 & 11.041 & 0.018 & 10.981 & 0.016 & 10.968 & 0.016 & 10.935 & 0.021 & 10.923 & 0.044 & — & — & 1.131 & 2.840E-14\
TWOM\_20323631\_4020142 & 20:32:36.32 & +40:20:14.3 & 11.918 & 9.999 & 11.508 & 0.027 & 11.443 & 0.026 & 11.432 & 0.016 & 11.48 & 0.017 & 11.409 & 0.073 & 11.268 & 0.231 & — & — & 0.986 & 3.822E-15\
TWOM\_20330655\_4022485 & 20:33:06.55 & +40:22:48.6 & 10.945 & 0.022 & 10.767 & 0.019 & 10.719 & 0.018 & 10.708 & 0.016 & 10.721 & 0.016 & 10.73 & 0.019 & 10.804 & 0.04 & — & — & 0.898 & 3.0744E-15\
CAHA\_20321512\_401714 & 20:32:15.12 & +40:17:14.8 & 14.779 & 0.003 & 13.287 & 0.005 & 12.676 & 0.008 & 12.291 & 0.019 & 12.159 & 0.022 & 12.094 & 0.135 & — & — & — & — & 1.789 & 5.472E-15\
CAHA\_20320777\_401759 & 20:32:07.77 & +40:17:59.4 & 16.206 & 0.092 & 13.979 & 0.038 & 13.316 & 0.009 & — & — & — & — & — & — & — & — & — & — & 4.417 & 2.468E-15\
CAHA\_20323293\_401511 & 20:32:32.93 & +40:15:11.6 & 17.384 & 0.011 & 15.979 & 0.01 & 15.275 & 0.011 & — & — & — & — & — & — & — & — & — & — & 2.650 & 1.288E-13\
CAHA\_20322779\_401512 & 20:32:27.79 & +40:15:12.4 & 18.761 & 0.011 & 16.498 & 0.009 & 15.323 & 0.011 & — & — & — & — & — & — & — & — & — & — & 5.351 & 4.285E-15\
CAHA\_20330360\_401707 & 20:33:03.60 & +40:17:07.4 & 20.784 & 0.065 & 19.02 & 0.044 & 18.445 & 0.094 & — & — & — & — & — & — & — & — & — & — & 2.562 & 7.871E-15\
CAHA\_20315779\_401717 & 20:31:57.79 & +40:17:17.9 & 15.757 & 0.004 & 14.392 & 0.006 & 13.699 & 0.015 & — & — & — & — & — & — & — & — & — & — & 1.920 & 3.300E-15\
CAHA\_20324085\_401304 & 20:32:40.85 & +40:13:04.7 & 20.26 & 0.069 & 17.919 & 0.041 & 17.023 & 0.049 & — & — & — & — & — & — & — & — & — & — & 1.949 & 2.980E-15\
\[tab:ClassIII\]
[^1]: Herschel is an ESA space observatory with science instruments provided by European-led Principal Investigator consortia and with important participation from NASA.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We review the calculation of the spectrum of glueball masses in non-supersymmetric Yang-Mills theory using the conjectured duality between supergravity and large $N$ gauge theories. The glueball masses are obtained by solving the supergravity wave equations in a black hole geometry. The glueball masses found this way are in unexpected agreement with the available lattice data. We also show how to use a modified version of the duality based on rotating branes to calculate the glueball mass spectrum with some of the Kaluza-Klein states of the supergravity theory decoupled from the spectrum.'
address: |
Theoretical Physics Group\
Ernest Orlando Lawrence Berkeley National Laboratory\
University of California, Berkeley, CA 94720\
[and]{}\
Department of Physics\
University of California, Berkeley, CA 94720
author:
- 'Csaba Csáki[^1] and John Terning'
title: 'Glueball Mass Spectrum from Supergravity[^2]'
---
LBNL-42987\
UCB-PTH-99/08\
hep-th/9903142\
Introduction
============
Maldacena’s conjecture [@Malda] relates ${\cal N}=4$ supersymmetric $SU(N)$ gauge theories in the large $N$ limit to Type IIB string theory on an AdS$_5\times {\bf S}^5$ background, where AdS$_5$ is a five dimensional anti-de Sitter space. The metric of this space is given by $${ds^2 \over l_s^2 \sqrt{4 \pi g_s N}} = \rho^{-2} d\rho^2 +
\rho^2
\sum_{i=1}^4 dx_i^2 + d \Omega_5^2
\label{ads5}$$ where $l_s$ is the string length related to the superstring tension, $g_s$ is the string coupling constant and $d\Omega_5$ is the line element on ${\bf S}^5$. The $x_{1,2,3,4}$ directions in AdS$_5$ correspond to ${\bf R}^4$ where the gauge theory lives. The gauge coupling constant $g_4$ of the 4D theory is related to the string coupling constant $g_s$ by $g_4^2 = g_s$. In the ’t Hooft limit ($N \rightarrow \infty$ with $g_4^2N = g_s N$ fixed), the string coupling constant vanishes $g_s \rightarrow 0$. Therefore we can study the 4D theory using the first quantized string theory in the AdS space (\[ads5\]). Moreover if $g_s N \gg 1$, the curvature of the AdS space is small and the string theory is approximated by classical supergravity. Witten extended this proposal to non-supersymmetric theories [@witten]. In his setup supersymmetry is broken by heating up the ${\cal N}=4$ theory, which corresponds to putting the four dimensional theory on a circle and assigning anti-periodic boundary conditions to the fermions. In this case the fermions will get a supersymmetry breaking mass term of the order $T=1/2\pi R$, where $R$ is the radius of the compact coordinate and $T$ is the corresponding temperature, while the scalars (not protected by supersymmetry anymore) will get masses from loop corrections. Thus in the $T\to \infty$ limit this should reproduce a pure (3 dimensional) $SU(N)$ theory in the large $N$ limit, which we will refer to as QCD$_3$. On the string theory side this corresponds to replacing the anti-de Sitter metric by a Schwarzschild metric describing a black hole in the anti-de Sitter space. This metric is given by
$${ds^2 \over l_s^2 \sqrt{4 \pi g_s N}}
=
\left(\rho^2 - {b^4 \over \rho^2}
\right)^{-1} d\rho^2 +
\left(\rho^2 - {b^4 \over \rho^2}
\right) d \tau^2 + \rho^2
\sum_{i=1}^3 dx_i^2 + d\Omega_5^2,
\label{metric}$$
where $\tau$ parameterizes the compactifying circle and the $x_{1,2,3}$ direction corresponding to the ${\bf R}^3$ where QCD$_3$ lives. The horizon of this geometry is located at $\rho=b$ with $$b = {1 \over 2R}=\pi T.
\label{horizonlocation}$$ The supergravity approximation is valid for this theory when the curvature of the space is small, thus when $g_s N\to \infty$. However, in order to obtain the pure gauge theory we have to take the temperature to infinity. In order to keep the intrinsic scale $g_3^2 N =g_4^2 N/R$ of the resulting theory at the scale of QCD, we simultaneously would need to take $g_4^2 N=g_s N \to 0$. Here $g_3$ is the dimensionful gauge coupling of QCD$_3$. This is exactly the opposite limit in which the supergravity approximation is applicable! Thus as expected for any strong-weak duality, the weakly coupled classical supergravity theory and the QCD$_3$ theory are valid in different limits of the ’t Hooft coupling $g_4^2 N$.
From the point of view of QCD$_3$, the radius $R$ of the compactifying circle provides the ultraviolet cutoff scale. Therefore, with the currently available techniques, the Maldacena-Witten conjecture can only be used to study large $N$ QCD with a fixed ultraviolet cutoff $R^{-1}$ in the strong ultraviolet coupling regime, and hope that the results one obtains this way are not very sensitive to removing the cutoff, that is on going from one limit to the other. Since the theory is non-supersymmetric, there is a priori no reason to believe that these two limits have anything to do with each other, since for example there might very well be a phase transition when the ’t Hooft coupling is decreased from the very large values where the supergravity description is valid to the small values where the theory should describe QCD$_3$. Nevertheless, Witten showed that the supergravity theory correctly reproduces several of the qualitative features of a confining 3 dimensional pure gauge theory correctly [@witten]. In particular, he showed that there is an area law in the Wilson loop and that there is a mass gap in the spectrum, both of which are expected features of a confining gauge theory. Here we will address the question of whether any of the quantitative features of the gauge theories are reproduced as well. In particular, we will calculate the glueball mass spectrum of the theory, and find, that it is in reasonable agreement with recent lattice simulations[@COOT].
The Glueball Spectrum in 3 Dimensions
=====================================
In this section we will show how to calculate the glueball spectrum of some of the glueballs in the supergravity approximation in the 3 dimensional case. In the following we will use the notation $J^{PC}$ for the glueballs, where $J$ is the glueball spin, and $P$, $C$ refer to the parity and charge conjugation quantum numbers respectively. In the field theory, one can find operators that have the quantum numbers corresponding to the given glueball states. For example, an operator with quantum numbers $0^{++}$ is given by ${\cal O}_4={\rm Tr} F^2$, or an operator with quantum numbers $0^{--}$ is given by ${\cal O}_6=d^{abc} F_{\mu \alpha}^a F^{b\alpha \beta}
F^c_{\beta \nu}$. According to the refinement of the Maldacena conjecture given in [@ref], one should find a supergravity state corresponding to the chiral primary operators of the original ${\cal N}=4$ conformal theory, which will couple to the supergravity states on the boundary of the AdS space. Assuming this coupling is maintained while heating the system, we can find the supergravity operators coupling to ${\cal O}_4$ and ${\cal O}_6$. The dilaton and the R-R scalar of the supergravity theory combine into a complex massless scalar field. Its real and imaginary parts couple to the dimension 4 scalar operators ${\cal O}_4 = {\rm tr}~ F^2$ and $\tilde {\cal O}_4 = {\rm tr} ~F\wedge F$. The NS-NS and R-R two-forms combine into a complex-valued antisymmetric field $A_{\mu\nu}$, polarized along the ${\bf R}^4$. Its $({\rm AdS~ mass})^2=16$ and thus one can show that it couples to a dimension 6 two-form operator of the ${\cal N}=4$ theory. This operator has been identified as the operator ${\cal O}_6$ [@dastrivedi; @flz]. With this knowledge we would like to calculate the actual glueball mass spectrum corresponding to these operators ${\cal O}_4$ and ${\cal O}_6$. In field theory, in order to calculate the masses of these states one would need to evaluate the correlators $\langle {\cal O}_4 (x) {\cal O}_4 (y) \rangle =
\sum_i c_i e^{-m_i |x-y|}$, where the $m_i$’s are the glueball masses. According to the refinement of the Maldacena conjecture [@ref], this just amounts to solving the supergravity wave equations for the fields that couple to these operators on the boundary. In the case of the $0^{++}$ glueballs, we need to find the solutions of the dilaton equations of motion of the form $\Phi =f(\rho ) e^{ikx}$. This is because in the supergravity theory on AdS$_5\times {\bf S}^5$, the Kaluza-Klein modes on the ${\bf S}^5$ can be classified according to the spherical harmonics of the ${\bf S}^5$, which form representations of the isometry group $SO(6)$ (which is the R-symmetry group of the ${\cal N}=4$ theory). When we put the theory at finite temperature, the states carrying non-trivial $SO(6)$ quantum numbers should eventually decouple from the spectrum, thus the glueballs should be identified with the $SO(6)$ singlet states, which implies a solution of the form $\Phi =f(\rho ) e^{ikx}$ for the dilaton as mentioned above. Thus we will look for normalizable regular solutions to the dilaton equation of motion which will give a discrete spectrum with the glueball masses determined as $k_i^2=-M_i^2$.
In the supergravity description we have to solve the classical equation of motion of the massless dilaton, $$\label{dilaton}
\partial_{\mu} \left[ \sqrt{g} \partial_{\nu}\Phi
g^{\mu \nu} \right] =0 \ ,$$ on the AdS$_5$ black hole background (\[metric\]). Plugging the ansatz $\Phi =f(\rho )e^{ikx}$ into this equation and using the metric of (\[metric\]) one obtains the following differential equation for $f$: $$\rho^{-1} {{d}\over{d \rho}} \left( \left(\rho^4 -
b^4 \right) \rho {{d f}\over{d \rho}} \right) - k^2 f = 0
\label{dilatondiff}$$ Since the glueball mass $M^2$ is equal to $-k^2$, the task is to solve this equation as an eigenvalue problem for $k^2$. In the following we set $b=1$, so the masses are computed in units of $b$. We need to find normalizable solutions to this equations which are also regular at the horizon. For large $\rho$, the black hole metric (\[metric\]) asymptotically approaches the AdS metric, and the behavior of the solution for a $p$-form for large $\rho$ takes the form $\rho^{\lambda}$, where $\lambda$ is determined from the mass $m$ of the supergravity field: $$m^2 = \lambda( \lambda + 4 - 2p)~.
\label{power}$$ Indeed both (\[dilatondiff\]) and (\[power\]) give the asymptotic forms $f \sim 1, \rho^{-4}$, and only the later is a normalizable solution [@witten]. Changing variables to $f=\psi/\rho^4$ we have: $$\left( {\rho^2} - {\rho^6} \right) \,\psi^{\prime\prime} +
\left( 3\,{\rho^5} -7 \rho \right) \,\psi^\prime
+\left( 16 + k^2 \rho^2 \right) \, \psi = 0$$ For large $\rho$ this equation can be solved by series solution with negative even powers: $$\psi = \Sigma_{n=0}^\infty a_{2n} \rho^{-2n}
\label{asymp}$$ Since the normalization is arbitrary we can set $a_0=1$. The first few coefficients are given by: $$a_2= {{k^2}\over{12}}, \ \ a_4 = {{1}\over{2}} + {{k^4}\over{384}}, \ \
a_6 = {{7 k^2}\over{120}} + {{k^6}\over{23040}}.
\label{asymp2}$$ For $n \ge 5$ the coefficients are given by the recursive relation: $$(n^2+4n) a_n = k^2 a_{n-2} + n^2 a_{n-4} ~.$$ Since the black hole geometry is regular at the horizon $\rho=1$, $k^2$ has to be adjusted so that $f$ is also regular at $\rho=1$ [@witten]. This can be done numerically in a simple fashion using a “shooting" technique as follows. For a given value of $k^2$ the equation is numerically integrated from some sufficiently large value of $\rho$ ($\rho \gg k^2$) by matching $f(\rho)$ with the asymptotic solution set by (\[asymp\]) and (\[asymp2\]). The glueball mass $M$ is related to the eigenvalues of $k^2$ by $M^2 = - k^2$ in units of $b^2$. The results obtained this way, together with the results of the lattice simulations [@Teper97] are displayed in Table \[summary\]. Since the lattice results are in units of string tension, we normalize the supergravity results so that the lightest $0^{++}$ state agrees with the lattice result. One should also expect a systematic error in addition to the statistical error denoted in Table \[summary\] for the lattice computations. Similar numerical results have been obtained in [@jev], while a WKB approximation for the eigenvalues of (\[dilatondiff\]) has been obtained in [@Minahan].
state lattice, $N=3$ lattice, $N\rightarrow \infty$ supergravity
--------------- ------------------- -------------------------------- ------------------
$0^{++}$ $4.329 \pm 0.041$ $4.065 \pm 0.055$ 4.07 ([input]{})
$0^{++*}$ $6.52 \pm 0.09$ $6.18 \pm 0.13$ 7.02
$0^{++**}$ $8.23 \pm 0.17$ $7.99 \pm 0.22$ 9.92
$0^{++***}$ - - 12.80
$0^{++****}$ - - 15.67
$0^{++*****}$ - - 18.54
: $0^{++}$ glueball masses in QCD$_3$ coupled to ${\rm tr}~F_{\mu \nu}
F^{\mu\nu}$. The lattice results are in units of the square root of the string tension. The denoted error in the lattice results is only the statistical one.\[summary\]
The $0^{--}$ glueballs can be dealt with similarly by considering the two-form of the supergravity theory, which couples to the operator ${\cal O}_6$. The supergravity equation of motion for the s-wave component of this field is given by $${{3}\over{\sqrt{g}}} \partial_\mu\left[\sqrt{g} \, \partial_{[\mu^\prime}
A_{\mu_1^\prime
\mu_2^\prime]} \, g^{\mu^\prime \mu} g^{\mu_1^\prime \mu_1} g^{\mu_2^\prime
\mu_2}\right]
- 16 g^{\mu_1^\prime \mu_1} g^{\mu_2^\prime \mu_2}
A_{\mu_1^\prime \mu_2^\prime} = 0 ,$$ where $[~~]$ denotes antisymmetrization with strength one. For the pseudoscalar component of $A_{ij}$ the equation reduces to $$\rho\left({\rho^4}\, - 1\right) h''
+ \left( 3 + {\rho^4}\right)h' -\left( {k^2}\,\rho\,
+16\,{\rho^3} \right) h =0 ~,$$ in units where $b=1$. This can be solved similarly as for the case of the $0^{++}$ glueballs, and the results are displayed in Table \[3dtensor\]. Since the supergravity method and the lattice gauge theory compute the glueball masses in different units, one cannot compare the absolute values of the lowest glueball mass obtained using these methods. However it makes sense to compare the lowest glueball masses of different quantum numbers. Using Tables \[summary\] and \[3dtensor\], we find that the supergravity results are in good agreement with the lattice gauge theory computation [@Teper97]: $$\begin{aligned}
&\left(\frac{M_{0^{--}}}{M_{0^{++}}}\right)_{{\rm supergravity}}&= 1.50
\nonumber \\
&\left(\frac{M_{0^{--}}}{M_{0^{++}}}\right)_{{\rm lattice~~~~~}}& =
1.45\pm 0.08\end{aligned}$$
state lattice, $N=3$ lattice, $N\rightarrow \infty$ supergravity
--------------- ----------------- -------------------------------- --------------
$0^{--}$ $6.48 \pm 0.09$ $5.91 \pm 0.25$ 6.10
$0^{--*}$ $8.15 \pm 0.16$ $7.63 \pm 0.37$ 9.34
$0^{--**}$ $9.81 \pm 0.26$ $8.96 \pm 0.65$ 12.37
$0^{--***}$ - - 15.33
$0^{--****}$ - - 18.26
$0^{--*****}$ - - 21.16
: $0^{--}$ glueball masses in QCD$_3$ coupled to ${\cal O}_6$. The lattice results are in units of square root of the string tension. The normalization of the supergravity results is the same as in Table \[summary\].\[3dtensor\]
One can see, that the glueball mass ratios obtained from the supergravity calculation are in reasonable agreement with the lattice results, even though as explained in the introduction these two calculations are in the opposite limits for the ’t Hooft coupling. Therefore, it is important to see, how the ratios are modified once corrections due to string theory are taken into account. The leading string theory corrections can be calculated by using the results of [@GKT], who calculated the first $\alpha'$ corrections to the AdS black-hole metric (\[metric\]). The details of the calculation can be found in [@COOT], here we just give the results for the $0^{++}$ state: $$\begin{aligned}
M_{0^{++}}^2 &=& 11.59\times
(1 -2.78 \zeta(3) \alpha'^3 ) \Lambda_{UV}^2 \nonumber \\
M_{0^{++*}}^2 &=& 34.53\times
(1 -2.43 \zeta(3) \alpha'^3 ) \Lambda_{UV}^2 \nonumber \\
M_{0^{++**}}^2 &=& 68.98\times
(1 - 2.28 \zeta(3) \alpha'^3 ) \Lambda_{UV}^2 \nonumber \\
M_{0^{++***}}^2 &=& 114.9\times
(1 -2.23 \zeta(3) \alpha'^3 ) \Lambda_{UV}^2 \nonumber \\
M_{0^{++****}}^2 &=& 172.3\times
(1 -2.21 \zeta(3) \alpha'^3 ) \Lambda_{UV}^2 \nonumber \\
M_{0^{++*****}}^2 &=& 241.2\times
(1 -2.20 \zeta(3) \alpha'^3 ) \Lambda_{UV}^2 ~,
\label{truecorrection}\end{aligned}$$ where $\Lambda_{UV}=\frac{1}{2R}$ and the correction to the horizon is given by $b=(1-\frac{15}{8} \zeta (3) \alpha'^3)\frac{1}{2R}$. One can see that the string theory corrections are somewhat uniform for the different excited states of the $0^{++}$ glueball, and therefore one could hope that these corrections to the ratios of the glueball masses are small. However, it can be seen that this is probably a too optimistic assumption, by considering the Kaluza-Klein partners of the glueball states. As explained above, the glueball states do not carry quantum numbers under the $SO(6)$ isometry, and are also singlets under the $U(1)$ symmetry corresponding to the compact direction $\tau$. The Kaluza-Klein modes however do carry quantum numbers under $SO(6)\times U(1)$, and they do not correspond to any state in the QCD theory, but rather they should decouple in the $R\to 0, g_4^2N\to 0$ limit from the spectrum. However, in the supergravity limit of finite $R$, $g_4^2N \to \infty$ these states have masses comparable to the light glueballs [@ORT]. This is simply a consequence of the fact, that the masses of the fermions and scalars carrying the $SO(6)
\times U(1)$ quantum numbers is of the order of the temperature $T$, thus their bound states are expected to also have masses of the order of the temperature. However, since the temperature is the only scale in the theory, and so this will also be the cutoff scale of the QCD theory, and thus the mass scale for the glueballs. In particular, the masses of the KK modes of the $0^{++}$ glueballs obtained from the dilaton equation by using the ansatz $\Phi= f(\rho ) e^{ikx} Y_l (\Omega_5)$ are given by [@ORT] $$\begin{tabular}{c|cccc}
$l$ & 0 & 1 & 2 & 3 \\ \hline
$M_l^2$ & 11.59 & 19.43 & 29.26 & 41.10 \\
${M_{l^{*}}}^2$ & 34.53 & 48.07 & 63.60 & 81.11 \\
${M_{l^{**}}}^2$ & 68.98 & 88.24 & 109.5 & 132.7 \end{tabular}
,$$ where we have displayed the unnormalized values of the masses of the different KK modes.
One can explicitly see, that the masses of these KK modes are as expected of the same order as the masses of the glueball states. One might hope that even though the supergravity approximation of these masses is of the same order as for the glueballs, string theory corrections will increase the masses of these states compared to the glueball states. Unfortunately, at least the leading string theory corrections calculated in [@ORT; @COOT] do not support this conclusion. The corrections to the first few KK modes are
$$\begin{aligned}
M_{0}^2 &=& 11.59\times
(1 -2.78 \zeta(3) \alpha'^3) \Lambda_{UV}^2 \nonumber \\
M_{1}^2 &=& 19.43\times
(1 -2.73 \zeta(3) \alpha'^3) \Lambda_{UV}^2 \nonumber \\
M_{2}^2 &=& 29.26\times
(1 - 2.74 \zeta(3) \alpha'^3) \Lambda_{UV}^2\end{aligned}$$
Thus one can see that the masses of these KK modes in fact do need large $\alpha'$ corrections to remove them from the spectrum of states. Then it is not clear why one would get large corrections to the masses of the KK modes but not to the masses of the glueball states. This situation is clearly unsatisfactory, therefore one may try to improve on it by introducing a different supergravity background, where some of these KK modes are automatically decoupled. We will consider this possibility in the next section where we discuss the construction based on rotating branes[@Russo; @CORT; @CRST].
The Glueball Spectrum in 4 Dimensions and the Construction Based on Rotating Branes
===================================================================================
Results similar to the the ones presented in the previous section can be obtained for the glueball mass spectrum in QCD$_4$ by starting from a slightly different construction where the M-theory 5-brane is wrapped on two circles [@witten]. The details of these results can be found in [@COOT; @HO]. Here we will review only the generalized construction based on the rotating M5 brane with one angular momentum, first constructed in [@Russo], and explored in [@CORT]. The metric for this background is given by $$\begin{aligned}
ds^2_{\rm IIA}&=&{2\pi \lambda A \over 3u_0} u \Delta ^{1/2}\bigg[ 4u^2
\big( -dx_0^2+dx_1^2+dx_2^2+dx^2_3\big)
+ { 4A^2\over 9u_0^2} u^2 \ (1-{u_0^6\over u^6 \Delta }) d\theta_2^2
+ {4\ du^2 \over u^2 (1-{a^4\over u^4}-{u_0^6\over u^6 }) }
\nonumber \\
&+& d\theta^2+{\tilde{\Delta}\over \Delta} \sin^2\theta d\varphi^2
+{1\over \Delta } \cos^2\theta d\Omega_2^2
-{4a^2 A u_0^2\over 3u^4\Delta } \sin^2\theta d\theta_2 d\varphi \bigg],
\label{pocho}\end{aligned}$$ where $x_{0,1,2,3}$ are the coordinates along the brane where the gauge theory lives, $u$ is the “radial" coordinate of the AdS space, while the remaining four coordinates parameterize the angular variables of $S^4$, $a$ is the angular momentum parameter, and we have introduced $$\Delta=1-{a^4\cos^2\theta \over u^4}\ ,\ \ \ \ \tilde{\Delta}=1-{a^4\over
u^4} \ ,
\ \ \ \ A\equiv {u_0^4\over u_H^4-\frac{1}{3} a^4}\ , \ \ \ \
u_H^6-a^4 u_H^2-u_0^6=0\ .$$ $u_H$ is the location of the horizon, and the dilaton background and the temperature of the field theory are given by $$\begin{aligned}
e^{2\phi }={8\pi\over 27} {A^3\lambda^3 u^3\Delta^{1/2}\over u_0^3}
{1\over N^2}\ , \ \ \ \ R=(2\pi T_H)^{-1}={A\over 3u_0}\ .
\label{dilz}\end{aligned}$$ Note, that in the limit when $a/u_0\gg 1$, the radius of compactification $R$ shrinks to zero, thus the KK modes on this compact direction are expected to decouple in this theory when we increase the angular momentum $a$. In order to find the mass spectrum of the $0^{++}$ glueballs, we need to again solve the dilaton equations of motion as a function of $a$. This can be done by plugging the background (\[pocho\]) into the dilaton equation of motion $$\partial_{\mu} \left[ \sqrt{g} e^{-2 \Phi}
g^{\mu \nu} \partial_{\nu} \Phi \right]=0.$$ For a dilaton ansatz of the form $\Phi = f(u) e^{ikx}$ we obtain the differential equation $$\partial_u \left[ u(u^6-a^4u^2-u_0^6) f'(u)\right]-k^2 u^3f(u)=0,$$ which can be solved the same way as explained in the previous section, where the eigenvalues are now a function of the angular momentum parameter $a$. The results of this are summarized in Table \[tab:4ddila\]. Note, that while some of the KK modes decouple in the $a\to \infty$ limit, the $0^{++}$ glueball mass ratios change only very slightly, showing that the supergravity predictions are robust for these ratios against the change of the angular momentum parameter.
state lattice, $N=3$ supergravity $a=0$ supergravity $a\to \infty$
------------- ------------------ -------------------- ----------------------------
$0^{++}$ $1.61 \pm 0.15$ 1.61 [(input)]{} 1.61 [(input)]{}
$0^{++*}$ $2.48 \pm 0.18 $ 2.55 2.56
$0^{++**}$ - 3.46 3.48
$0^{++***}$ - 4.36 4.40
: Masses of the first few $0^{++}$ glueballs in QCD$_4$, in GeV, from supergravity compared to the available lattice results. The first column gives the lattice result [@Teper97; @MorningstarPeardon; @Peardon], the second the supergravity result for $a=0$ while the third the supergravity result in the $a\to \infty$ limit. The change from $a=0$ to $a=\infty$ in the supergravity predictions is tiny. Note, that for the excited state the supergravity calculation came before the lattice results.\[tab:4ddila\]
One can similarly calculate the mass ratios for the $0^{-+}$ glueballs, by considering the equations of motion of the RR 1-form in the background (\[pocho\]), since on the D4 brane worldvolume this couples to the operator ${\rm Tr}F\tilde{F}$. To find the glueball spectrum we have to solve the supergravity equation of motion of the RR 1-form $$\partial_{\nu} \left[ \sqrt{g} g^{\mu\rho}g^{\nu\sigma}
(\partial_{\rho}A_{\sigma}-\partial_{\sigma}A_{\rho})\right]=0$$ in the background (\[pocho\]). Using the ansatz $A_{\theta_2} =f(u) e^{ikx}$ leads to the differential equation $$(u^6-a^4u^2-u_0^6)\partial_u\left[ u^3(u^4-a^4)f'(u)\right]
-k^2u^5(u^4-a^4)f(u),$$ which we solve using the same numerical methods as in the previous section. The results are summarized in Table \[tab:0-+\]. Note, that the change in the $0^{-+}$ glueball mass is sizeable when going from $a=0$ to $a\to \infty$, and is in the right direction as suggested by lattice results [@MorningstarPeardon; @Peardon].
state lattice, $N=3$ supergravity $a=0$ supergravity $a\to \infty$
------------- ---------------- -------------------- ----------------------------
$0^{-+}$ 2.59 $\pm$0.13 2.00 2.56
$0^{-+*}$ 3.64 $\pm$0.18 2.98 3.49
$0^{-+**}$ - 3.91 4.40
$0^{-+***}$ - 4.83 5.30
: Masses of the first few $0^{-+}$ glueballs in QCD$_4$, in GeV, from supergravity compared to the available lattice results. The first column gives the lattice result, the second the supergravity result for $a=0$ while the third the supergravity result in the $a\to \infty$ limit. Note that the change from $a=0$ to $a=\infty$ in the supergravity predictions is of the order $\sim 25 \%$.\[tab:0-+\]
One can also calculate the masses of the different Kaluza-Klein modes in the background of (\[pocho\]). One finds, that as expected from the fact that for $a\to \infty$ the compact circle shrinks to zero, the KK modes on this compact circle decouple from the spectrum, leading to a 4 dimensional field theory in this limit. However, the KK modes of the sphere $S^4$ do not decouple from the spectrum even in the $a\to \infty$ limit. These conclusions remain unchanged even in the case when one considers the theory with the maximal number of angular momenta (which is two for the case of QCD$_4$) [@CRST; @JorgeKostas]. In the limit when the angular momentum becomes large, $a/u_0\gg 1$, the theory approaches a supersymmetric limit [@Russo; @CRST] since the supersymmetry breaking fermion masses get smaller with increasing angular momentum [@CG]. Therefore, the limit of increasing angular momentum on one hand does decouple some of the KK modes which makes the theory four dimensional, but at the same time reintroduces the light fermions into the spectrum [@CG].
Conclusions
===========
We have seen how the Witten extension of Maldacena’s conjecture can be used to study pure Yang-Mills theories in the large $N$ limit. These theories reproduce several of the qualitative features of QCD, and one can also study the predictions for the glueball mass spectra. One finds, that the supergravity calculations are in a reasonable agreement with the lattice results, even though they are obtained in the opposite limit of the ’t Hooft coupling. It would be very important to understand, whether this unexpected agreement is purely a numerical coincidence or whether there is any deeper reason behind it.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Hirosi Ooguri, Yaron Oz, Jorge Russo and Konstadinos Sfetsos for several collaborations, based on which this paper has been written. C. C. is a research fellow of the Miller Institute for Basic Research in Science. This work was supported in part the U.S. Department of Energy under Contract DE-AC03-76SF00098, and in part by the National Science Foundation under grant PHY-95-14797.
J. M. Maldacena, [*Adv. Theor. Math. Phys.*]{} [**2**]{} 231 (1998).
E. Witten, [*Adv. Theor. Math. Phys.*]{} [**2**]{} 505 (1998).
C. Csáki, H. Ooguri, Y. Oz and J. Terning, [*JHEP*]{} [**9901**]{} 017 (1999).
S. Gubser, I. Klebanov and A. Polyakov, [*Phys. Lett.*]{} [**B428**]{} 105 (1998); E. Witten, [*Adv. Theor. Math. Phys.*]{} [**2**]{} 253 (1998).
S. Das and S. Trivedi, [*Phys. Lett.*]{} [**B445**]{} 142 (1998).
S. Ferrara, M. A. Lledo, A. Zaffaroni, [*Phys. Rev.*]{} [**D58**]{} 105029 (1998).
M. J. Teper, hep-lat/9711011.
R. de Mello Koch, A. Jevicki, M. Mihailescu and J. Nunes, [*Phys. Rev.*]{} [**D58**]{} 105009 (1998); M. Zyskin, [*Phys. Lett.*]{} [**B439**]{} 373 (1998).
J. Minahan, hep-th/9811156.
S. Gubser, I. Klebanov and A. Tseytlin, [*Nucl. Phys.*]{} [**B534**]{} 202 (1998).
H. Ooguri, H. Robins and J. Tannenhauser, [*Phys. Lett.*]{} [**B437**]{} 77-81 (1998).
J. Russo, hep-th/9808117.
C. Csáki, Y. Oz, J. Russo and J. Terning, [*Phys. Rev.*]{} [**D59**]{} 065008 (1999).
C. Csáki, J. Russo, K. Sfetsos and J. Terning, hep-th/9902067.
A. Hashimoto and Y. Oz, hep-th/9809106.
C. Morningstar and M. Peardon, [*Phys. Rev.*]{} [**D56**]{} 4043 (1997).
M. Peardon, [*Nucl. Phys.*]{} [**B**]{} (Proc. Suppl.) [**63**]{} 22 (1998); C. Morningstar and M. Peardon, hep-lat/9901004.
J. Russo and K. Sfetsos, hep-th/9901056.
M. Cvetič and S. Gubser, hep-th/9903132.
[^1]: Research fellow, Miller Institute for Basic Research in Science.
[^2]: Based on two talks presented at DPF ’99, UCLA, Los Angeles, CA, 5-9 January, 1999.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that off-site processes and multi-orbital physics have a crucial impact on the phase diagram of quantum gas mixtures in optical lattices. In particular, we discuss Bose-Fermi mixtures where the intra- and interspecies interactions induce competing density-induced hopping processes, the so-called bond-charge interactions. Furthermore, higher bands strongly influence tunneling and on-site interactions. We apply a multi-orbital interaction-induced dressing of the lowest band which leads to renormalized hopping processes. These corrections give rise to an extended Hubbard model with intrinsically occupation-dependent parameters. The resulting decrease of the tunneling competes with a decrease of the total on-site interaction energy both affecting the critical lattice depth of the superfluid to Mott insulator transition. In contrast to the standard Bose-Fermi-Hubbard model, we predict a large shift of the transition to shallower lattice depths with increasing Bose-Fermi attraction. The applied theoretical model allows an accurate prediction of the modified tunneling amplitudes and the critical lattice depth both recently observed experimentally.'
author:
- Ole Jürgensen
- Klaus Sengstock
- 'Dirk-Sören Lühmann'
title: 'Density-induced processes in quantum gas mixtures in optical lattices'
---
Quantum gas mixtures in optical lattices are well suited to study in detail interaction-induced effects in condensed matter. They allow for the investigation of systems with spin degree of freedom and with different species of particles that can even obey different quantum statistics. In particular, the experimental realization of atomic mixtures of bosonic and fermionic particles (e.g., $^{87}$Rb-$^{40}$K) in optical lattices [@Gunter2006; @Ospelkaus2006; @Best2009; @Heinze2011] triggered a vivid discussion on the role of inter- and intraspecies interactions. These experiments allow for the observation of the bosonic superfluid to Mott-insulator transition in the presence of fermionic atoms. The prominent feature observed in all experiments is the decay of visibility and condensate fraction of the bosonic subsystem induced by the interaction with the fermionic atoms. Two possible explanations for this drop in bosonic coherence were proposed. First, the process of adiabatic heating while ramping the lattice has been suggested [@Cramer2008; @Cramer2011]. It is caused by different contributions of the atomic species to the total entropy and is therefore specific to the loading procedure of experiments with ultracold gases. Second, an interaction-induced dressing of tunneling and interaction processes has been found that causes a shift of the superfluid to Mott-insulator phase transition[@Luhmann2008b; @Best2009; @Lutchyn2009; @Mering2011]. The latter effect corresponds to a necessary extension of the Hubbard model at zero temperature and is therefore fundamental for various lattice systems. The important role of interaction-induced processes in optical lattices is caused by the specific shape of the Wannier functions and the possibility of high filling factors.
The [*standard*]{} Bose-Fermi-Hubbard model [@Albus2003] is restricted to the lowest single-particle band and on-site interactions. Interestingly, it fails to describe interaction effects in boson-fermion mixtures. Assuming a fermionic band-insulator, which is present in the experimental realizations [@Ospelkaus2006; @Gunter2006; @Best2009; @Heinze2011], the boson-fermion interaction gives rise only to an irrelevant shift of the global chemical potential. Even for realistic assumptions for the confining potential, the interspecies interaction has little influence [@Kollath2004]. Due to an effective screening even a small shift in the opposite direction is predicted. In contrast to the experimental results, the superfluid phase is therefore more stable within this framework, which poses the question of the applicability of the standard Hubbard model in this case. It was pointed out that the Bose-Fermi-Hubbard parameters, i.e., tunneling and on-site interactions can be strongly influenced by the inclusion of higher orbitals and non-local interactions [@Luhmann2008b; @Lutchyn2009; @Mering2011]. Off-site interactions include the so-called bond-charge interactions that have a direct density-dependent influence on the total tunneling [@Hirsch1989; @Strack1993; @Hirsch1994; @Amadon1996; @Mazzarella2006; @Mering2011; @Luhmann2012; @Bissbort2011]. For purely bosonic systems and without taking off-site interactions into account, the inclusion of higher bands has been discussed in, e.g., Ref. [@Li2006; @Larson2009; @Hazzard2010; @Dutta2011; @Sakmann2009; @Sakmann2010; @Busch1998; @Johnson2009; @Will2010; @Buchler2010; @Pilati2011]. Both extensions can be included in extended Hubbard models [@Mering2011; @Luhmann2012; @Bissbort2011] and can be illustrated as effective modifications of the tunneling potentials for both atomic species [@Luhmann2012].
Here, we present quantitative results for the phase diagram of Bose-Fermi mixtures using an extended Hubbard model with multi-orbitally dressed processes [@Luhmann2012; @Bissbort2011]. It includes density-induced tunneling, so-called bond-charge interactions [@Hirsch1989; @Strack1993; @Hirsch1994; @Amadon1996; @Mazzarella2006; @Mering2011; @Luhmann2012; @Bissbort2011], as well as a multi-orbitally renormalized tunneling and interaction processes [@Will2010; @Mering2011; @Luhmann2012; @Bissbort2011]. The physical effects discussed here are in general present for all interacting quantum gas mixtures. The exact results, however, depend on the quantum statistics of the particles and the specific parameters such as the detuning from the light potential and the atomic masses.
As a central result, we present the phase diagram of the superfluid to Mott-insulator transition in a Bose-Fermi mixture in section \[sec:PD\]. Furthermore, we discuss the crucial effect of off-site interactions and the corresponding shortcomings of the standard Bose-Fermi Hubbard model in section \[sec:off-site\_interactions\]. Subsequently, we present the procedure to incorporate higher-band processes in section \[sec:MO\]. Afterwards, the corresponding extended Hubbard models and the implications for the bosonic phase transition are discussed in detail in section \[sec:models\].
[{width="0.8\linewidth"}]{}
Phase diagrams {#sec:PD}
==============
We will now first discuss the resulting phase diagrams of the bosonic superfluid to Mott-insulator transition in the presence of a fermionic band-insulator, where the individual corrections to the standard Hubbard model are discussed in detail below. For concreteness, we choose a mixture of bosonic ${^{87}\mathrm{Rb}}$ and fermionic ${^{40}\mathrm{K}}$ in an optical lattice with a spacing of $a=377 \, \mathrm{nm}$. For the respective wavelength, the Wannier functions of both species are almost identical. The interaction between the bosonic atoms is fixed to a repulsive scattering length of $a_{\mathrm{BB}}=102\, a_0$ [@Will2010], while the attractive interaction between the two species is tunable over a wide range using a Feshbach resonance [@Best2009]. The fermionic nature of the spin-polarized potassium atoms simplifies the system, as we can assume a band-insulating phase and thus a fixed atom-number of one fermion per lattice site. Note that it is possible to directly apply the presented methods and extensions to a variety of atomic quantum gas mixtures in lattices. In particular, both atomic species can be bosonic or fermionic and the generalization to multi-component systems with more than two species is straight forward.
For a fermionic band insulator, the fermionic degrees of freedom are frozen out and the physics can be described by an effective bosonic model that takes into account all effects induced by the interaction with the fermions. In the framework of this paper, we will discuss in detail the derivation of an effective Hamiltonian which reads $$\label{eq:Hext}
\tilde H_\mathrm{ext}= -\sum_{\langle i,j \rangle} \tilde{b}_i^\dagger \tilde{b}_j \tilde J^{{\mathrm{tot}}}_{\hat n_j,\hat n_i}
+ \sum_i \tilde E_{\hat{n}_i} - \mu \sum_i \hat{n}_i.$$ We will see that despite its simplicity it already includes higher-band and bond-charge off-site processes. The latter gives rise to an occupation-dependent tunneling $J^{{\mathrm{tot}}}_{n_j,n_i}= J_{\mathrm{B}}+ (n_i+ n_j-1) X_{\mathrm{BB}}+ 2 X_{\mathrm{BF}}$ even within the lowest single-particle band. Here, $J_{\mathrm{B}}$ is the conventional tunneling; $X_{\mathrm{BB}}$ and $X_{\mathrm{BF}}$ are the bond-charge tunneling elements arising from Bose-Bose and Bose-Fermi interactions, respectively (section \[sec:off-site\_interactions\]). The interaction induced occupation of higher orbitals leads to a further occupation dependency of all parameters, i.e., $\tilde J_{\mathrm{B}}$, $\tilde X_{\mathrm{BB}}$, $\tilde X_{\mathrm{BF}}$ and $\tilde E_n$.
The tilde above the parameters and operators in indicates the multi-orbital dressing as discussed in section \[sec:MO\]. The effective single-band Hamiltonian uses the ground state of the interacting system, called the dressed band, instead of the lowest single-particle band [@Luhmann2012; @Bissbort2011]. The dressed operators $\tilde b_i$ and $\tilde b_i^\dagger$ annihilate and create bosonic particles on site $i$ in this dressed band and $\hat n_i=\tilde b_i^\dagger \tilde b_i$ counts the number of bosons on site $i$. It is important to note that after the transformation to the dressed band, the phase diagrams can be calculated using standard single-band methods. The renormalized on-site energy $\tilde E_n$ is composed of the single particle energies of bosons $\tilde \epsilon_{\mathrm{B},n}$ and fermions $\tilde \epsilon_{\mathrm{F},n}$ as well as the interaction energies for the repulsion between the bosons $\frac12 n(n-1) \tilde U_n$ and the attraction between the species $n \tilde U_{\mathrm{BF},n}$. The chemical potential $\mu$ fixes the total number of bosonic atoms.
After calculating the dressed parameters, we apply Gutzwiller mean-field theory to compute the critical lattice depth of the transition from the superfluid to the Mott-insulator. The phase diagrams of the extended model are shown in [Fig. \[fig:PD\]]{}**a**. The effective chemical potential $\mu - \tilde E_1$ is given in units of the Hubbard on-site interaction $U$, where $\tilde E_1$ is the renormalized on-site energy of one boson and one fermion. For vanishing interaction between the bosons and fermions $a_{{\mathrm{BF}}}=0$ and a repulsive interaction $a_{\mathrm{BB}}=102\, a_0$ among the bosons, the Mott-lobes are contracted compared with the standard Hubbard model. This is a result of a decrease of the on-site energy and an increase of the total tunneling caused by off-site interactions [@Luhmann2012]. For increasing attraction between bosons and fermions the effect is reversed [@Mering2011] and the Mott-lobes are extended exhibiting a critical transition point at much lower lattice depths. This effect can be attributed to a strong reduction of the total tunneling amplitude induced by off-site interactions.
In [Fig. \[fig:PD\]]{}**b** the critical lattice depths for the superfluid to Mott insulator transition for one, two and three bosons per lattice site are shown as a function of the interspecies interaction strength. The solid lines depict the results obtained using the extended model , while the dashed lines correspond to the standard Bose-Fermi Hubbard model, which predicts no dependency on the interspecies interaction $a_{\mathrm{BF}}$ (section \[sec:off-site\_interactions\]). As discussed above, for $a_{\mathrm{BF}}=0$ the Mott-insulator transition is shifted to deeper lattices, where the shift is increased with the bosonic filling. This is mainly caused by the bosonic bond-charge interaction $\tilde X_{\mathrm{BB}}$ enhancing the total tunneling. With increasing attractive interaction $a_{\mathrm{BF}}$ the transition is strongly shifted to shallower lattice depths. Depending on the filling, the shift of the Mott-insulator transition caused by the fermionic atoms is $3$-$4\,{E_\mathrm{R}}$ for $a_{\mathrm{BF}}=-300\, a_0$.
The predicted shift is considerably larger than calculated with the adiabatic band elimination method in Ref. [@Mering2011] which also incorporates the bond-charge interactions. The effective potential approach in Refs. [@Luhmann2008b; @Best2009] does not include the important contributions of fermionic on-site energy and bosonic bond-charge interaction. The extended Bose-Fermi-Hubbard model discussed here contains all relevant energies that can affect the superfluid to Mott-insulator transition at zero temperature. As a result, the Mott-insulator shift in Ref. [@Best2009] can be partly explained by interaction-induced effects. This provides a consistent picture, where the experimental observations [@Ospelkaus2006; @Gunter2006; @Best2009] are a combined effect of the Hubbard extensions *and* the adiabatic heating processes [@Cramer2008; @Cramer2011] which depend on the initial temperature of the quantum gas.
Off-site interactions {#sec:off-site_interactions}
=====================
[{width="0.985\linewidth"}]{}
We now turn back to the full description of the applied extended Hubbard model which features two corrections to the standard Hubbard model. First, off-site interactions lead to a significant contribution to the total tunneling amplitude by changing the effective tunneling potential. Second, the inclusion of multi-band processes causes a modification of all model parameters.
As mentioned above, mixtures of ultracold spin-polarized bosonic and fermionic atoms in optical lattices are usually described by the standard Bose-Fermi Hubbard model. The underlying tight-binding approximation restricts the model to the lowest single-particle orbital and interactions between particles on the same lattice site. The resulting Hubbard Hamiltonian reads $$\begin{split}
\hat H_\text{BFH}=& - \sum_{{\langle i,j \rangle}} (J_{\mathrm{B}}\hat b_i^\dagger \hat b_j
\!+\! J_{\mathrm{F}}\hat f_i^\dagger \hat f_j )
+\frac{U_{\mathrm{BB}}}{2} \sum_i \hat n_i (\hat n_i-1) \\
& +\sum_i U_{\mathrm{BF}}\hat n_i \hat m_i
- \sum_i (\mu_{\mathrm{B}}\hat n_i + \mu_{\mathrm{F}}\hat m_i ),
\end{split}\label{eq:BFHM}$$ Here, $\hat b_i$ ($\hat f_i$) is the bosonic (fermionic) annihilation operator and $\hat n_i$ ($\hat m_i$) the respective particle number operator. In general, the tunneling matrix elements for bosons ($J_{\mathrm{B}}$) and fermions ($J_{\mathrm{F}}$) can have different values. The on-site interaction is fully described by the parameters $U_{\mathrm{BB}}$ and $U_{\mathrm{BF}}$ for intra- and interspecies interaction, respectively. The total number of bosonic and fermionic atoms are fixed by the chemical potentials $\mu_{\mathrm{B}}$ and $\mu_{\mathrm{F}}$. Under common experimental conditions [@Ospelkaus2006; @Gunter2006; @Best2009; @Heinze2011], the fermions are in a band-insulator phase where Pauli-blocking prohibits tunneling. This freezes out the fermionic degrees of freedom and the resulting Hamiltonian captures the behavior of the bosons under the influence of exactly one fermion per lattice site. Consequently, we can set $\hat f_i^\dagger \hat f_j \rightarrow 0,\ \hat m_i \rightarrow 1$ and get $$\begin{split}
\hat H_\text{FBI}=& - \sum_{{\langle i,j \rangle}} J_{\mathrm{B}}\hat b_i^\dagger \hat b_j
+\frac{U_{\mathrm{BB}}}{2} \sum_i \hat n_i (\hat n_i-1) \\
& +\sum_i (U_{\mathrm{BF}}-\mu_{\mathrm{B}}) \hat n_ i .
\end{split}\label{eq:BHM}$$ The interaction energy $U_{\mathrm{BF}}$ between bosons and fermions can be absorbed into an effective chemical potential $\mu_{{\mathrm{eff}}}=\mu_{\mathrm{B}}-U_{\mathrm{BF}}$ and the resulting Hamiltonian does not differ from the standard Bose-Hubbard model. Thus, the behavior of the bosons is not influenced by the homogeneously distributed fermions, which is in contradiction to the experimental observations [@Ospelkaus2006; @Gunter2006; @Best2009].
In the derivation of the standard Hubbard model, it is argued that interaction processes between particles on neighboring lattice sites can be neglected due to their small amplitudes compared with on-site interactions. This argument is however only partly correct, since some of these processes, such as bond-charge interactions, involve the hopping of particles. Compared to the conventional tunneling, these processes can be non-negligible and consequently alter the phase diagrams. In particular, the Wannier functions in optical lattices differ strongly from their counterpart in solid state materials leading to comparably large matrix elements for bond-charge processes. In addition, the possibility of larger fillings in bosonic systems can enlarge these interaction effects. Consequently, off-site interaction processes can be strongly enhanced for optical lattice systems.
Consider the interacting part of the full two-particle Hamiltonian for the lowest band and two neighboring lattice sites $\mathrm{L}$ and $\mathrm{R}$ $$\hat H_\mathrm{int}=
\frac{1}{2} \sum_{ijkl} U_{ijkl}^{\mathrm{BB}}\,
\hat{b}^{\dagger}_i \hat{b}^{\dagger}_j \hat{b}_k \hat{b}_l + \sum_{ijkl} U_{ijkl}^{\mathrm{BF}}\,
\hat{b}^{\dagger}_i \hat{f}^{\dagger}_j \hat{f}_k \hat{b}_l,
\label{eq:Hint}$$ with $i,j,k,l = \mathrm{L, R}$ and $$\begin{split}
\label{eq:Uijkl_BB}
U_{ijkl}^{{\mathrm{BB}}/{\mathrm{BF}}} = {g_{{\mathrm{BB}}/{\mathrm{BF}}}} \int \ \
&w^{{\mathrm{B}}*}_i{(\mathbf{r})}\, w^{{{\mathrm{B}}/{\mathrm{F}}} *}_j(\mathbf{r'})\, V(\mathbf{r},\mathbf{r'}) \\ \times\,
&w_k^{{\mathrm{B}}/{\mathrm{F}}}(\mathbf{r'})\, w_l^{\mathrm{B}}{(\mathbf{r})}\ \ d^3r\, d^3r'.
\end{split}$$ The basis functions $w^{{\mathrm{B}}/{\mathrm{F}}}_i{(\mathbf{r})}$ are the maximally localized Wannier functions that describe a boson/fermion sitting on site $i$ and the interaction strengths are given by $g_{\mathrm{BB}}= \frac{4\pi \hbar^2}{m_{\mathrm{B}}} {a_\mathrm{B}}$ and $g_{\mathrm{BF}}= \frac{2 \pi \hbar^2}{m_\mathrm{r}} a_{\mathrm{BF}}$ with the mass of the bosonic atoms $m_{\mathrm{B}}\ $ and the reduced mass $m_\mathrm{r}$ of boson and fermion. The interaction potential $V(\mathbf{r}, \mathbf{r'})$ we applied describes the scattering properties using a finite-ranged box potential (see [@Luhmann2012]). In a lowest-band treatment, this is usually replaced by contact interactions, i.e., a $\delta$-pseudopotential. In a multi-orbital framework the latter would lead to mathematical subtleties [@Busch1998].
The distinct processes arising from the full two-body Hamiltonian for Bose-Bose and the Bose-Fermi interaction are depicted in [Fig. \[fig:offsite-processes\]]{}. In addition to the on-site interaction, the Bose-Bose interaction leads to the density-density interaction process $V_{\mathrm{BB}}\, \hat n_i \hat n_j$, the correlated tunneling of a particle pair $P_{\mathrm{BB}}\, \hat b_i^{\dagger2} \hat b_j^{2}$, and the bond-charge assisted tunneling $-X_{\mathrm{BB}}\, \hat b_i^{\dagger} (\hat n_i+ \hat n_j) \hat b_j$. The respective matrix elements are $V_{\mathrm{BB}}=U^{\mathrm{BB}}_{ijji}$, $P_{\mathrm{BB}}=U^{\mathrm{BB}}_{iijj}/2$, and $X_{\mathrm{BB}}=-U^{\mathrm{BB}}_{iiij}=-U^{\mathrm{BB}}_{ijjj}$. The amplitudes of these processes are plotted in [Fig. \[fig:ProcessIntegrals\]]{} as solid lines. While the density-density interaction $V_{\mathrm{BB}}$ and the pair tunneling amplitude $P_{\mathrm{BB}}$ are several orders of magnitudes smaller than the the standard Hubbard processes, the bond-charge interaction $X_{\mathrm{BB}}$ is only one order of magnitude smaller than the conventional tunneling $J_{\mathrm{B}}$. As the bond-charge interaction $-X_{\mathrm{BB}}\, \hat b_i^{\dagger} (\hat n_i+ \hat n_j) \hat b_j$ scales with the particle number on both involved sites, it can easily reach non-negligible values and must be accounted for (see also [@Luhmann2012]). In addition, it scales linearly with the interaction strength which can be tuned experimentally. Here, it is essential that the bond-charge interaction contributes to the tunneling of the particles rather than the on-site interaction.
For the Bose-Fermi interaction two more processes exist as the interacting particles are distinguishable ([Fig. \[fig:offsite-processes\]]{}b). First, the cross-tunneling, which is similar to the density-density interaction except that the particles interchange, and second, the bond-charge interaction, where either a bosonic or a fermionic particle tunnels. However, assuming a fermionic band insulator all processes involving the hopping of a fermion are forbidden. The respective amplitudes of the processes induced by Bose-Fermi interaction are plotted as dashed lines in [Fig. \[fig:ProcessIntegrals\]]{}.
In conclusion, only on-site interactions and bond-charge tunneling of bosons have to be taken into account (shaded in [Fig. \[fig:offsite-processes\]]{}), since other processes are prohibited by the fermionic band-insulator or contribute only with extremely small amplitudes. Thus, the necessary extensions of the Hubbard Hamiltonian, i.e. the bond-charge processes in [Fig. \[fig:offsite-processes\]]{}**d** and **i**, read $$\begin{split}
\hat X = \hat X_{\mathrm{BB}}+ \hat X_{\mathrm{BF}}= \hat b_i^\dagger (&U_{iiij}^{\mathrm{BB}}\hat b_i^\dagger \hat b_i + U_{ijjj}^{\mathrm{BB}}\hat b_j^\dagger \hat b_j \\
+ &U_{iiij}^{\mathrm{BF}}\hat f_i^\dagger \hat f_i + U_{ijjj}^{\mathrm{BF}}\hat f_j^\dagger \hat f_j) \hat b_j.
\label{eq:Xsb}
\end{split}$$
[![Matrix elements for the different hopping and interaction processes. The solid lines are the purely bosonic processes, while the dashed lines correspond to Bose-Fermi interaction and fermionic tunneling. Shown are the intra- and interspecies on-site interaction $U_{\mathrm{BB}}$ and $U_{\mathrm{BF}}$ (green), the conventional tunneling amplitudes $J_{\mathrm{B}}$ and $J_{\mathrm{F}}$ (red), the bosonic bond-charge tunneling $X_{\mathrm{BB}}$ and $X_{\mathrm{BF}}$ (blue) as well as density-density interactions $V_{\mathrm{BB}}$ and $V_{\mathrm{BF}}$ (orange) and correlated pair-tunneling $P_{\mathrm{BB}}$ (cyan). \[fig:ProcessIntegrals\] ](Figure3.pdf "fig:"){width="0.8\linewidth"}]{}
An intuitive physical understanding of the bond-charge tunneling induced by Bose-Bose as well as by Bose-Fermi interaction can be obtained by the analogy to an effective tunneling potential [@Luhmann2012]. Assuming contact interactions in the single-band description and using the integral expressions , we can rewrite the expression as $$\label{eq:bond-charge_operator}
\hat{X} = \hat b_i^\dagger \hat b_j {\int d^3{r}\ }w_i^{{\mathrm{B}}*}{(\mathbf{r})}\left(g_{\mathrm{BB}}\hat \rho_{ij}^{\mathrm{BB}}+ g_{\mathrm{BF}}\hat \rho_{ij}^{\mathrm{BF}}\right) w_j^{\mathrm{B}}{(\mathbf{r})},$$ where we introduced the reduced densities $$\begin{aligned}
\hat \rho_{ij}^{\mathrm{BB}}&= \hat n_i |w_i^{\mathrm{B}}{(\mathbf{r})}|^2 + (\hat n_j-1) |w_j^{\mathrm{B}}{(\mathbf{r})}|^2 \\
\hat \rho_{ij}^{\mathrm{BF}}&= \hat m_i |w_i^{\mathrm{F}}{(\mathbf{r})}|^2 + \hat m_j |w_j^{\mathrm{F}}{(\mathbf{r})}|^2.\end{aligned}$$ The $-1$ in the bosonic density corresponds to the exclusion of self-interactions and is directly obtained from the commutation relations. Inside the integral, we can use the locality of the Wannier functions to replace these operators by the density functions $\hat \rho_{ij}^{\mathrm{BB}}\rightarrow \rho_{\mathrm{B}}{(\mathbf{r})}- |w_j^{\mathrm{B}}{(\mathbf{r})}|^2$ and $\hat \rho_{ij}^{\mathrm{BF}}\rightarrow \rho_{\mathrm{F}}{(\mathbf{r})}$. The bond-charge tunneling operator can now easily be unified with the conventional tunneling to find the expression $$\hat J + \hat X = \bra{w_i^{\mathrm{B}}{(\mathbf{r})}} \left(\frac{\mathbf{p}^2}{2m} + V_{{\mathrm{eff}}}{(\mathbf{r})}\right) \ket{w_j^{\mathrm{B}}{(\mathbf{r})}} \hat b_i^\dagger \hat b_j$$ which corresponds to the conventional tunneling in an effective potential $V_{{\mathrm{eff}}}{(\mathbf{r})}= V{(\mathbf{r})}+ g_{\mathrm{BB}}(\rho_{\mathrm{B}}{(\mathbf{r})}- |w_j^{\mathrm{B}}{(\mathbf{r})}|^2) + g_{\mathrm{BF}}\rho_{\mathrm{F}}{(\mathbf{r})}$.
In [Fig. \[fig:offsite-processes\]]{}**k** and **l** the tunneling in effective potentials is sketched. Repulsive interactions, as between the bosons, effectively reduce the lattice depth, while attractive interactions have the opposite effect. Depending on the relative scattering lengths, the total tunneling can be strongly enhanced or reduced. This is consistent with the results for the fermionic tunneling obtained in [@Heinze2011]. The modification of the total tunneling in the lowest band already leads to a considerable deformation of the well-known phase diagram of the superfluid to Mott-insulator transition for bosons as discussed later in detail.
Multi-orbital renormalization {#sec:MO}
=============================
Whereas in the previous section we introduced off-site interactions as an important extension to the standard Hubbard model, we will now discuss another important feature of the Hamiltonian , namely the effective inclusion of higher bands. In the standard Hubbard model approach, only the lowest single-particle band is assumed to be occupied. However, in the strongly correlated system particles are promoted to higher orbitals due to the interaction induced coupling between the bands. By changing their wave functions the particles minimize their on-site interaction energy (see section \[sec:MO\].B). In particular, this also affects the tunneling and other off-site processes as outlined now in section \[sec:MO\].A.
Multi-orbital dressing
----------------------
For the following calculation of the multi-orbital effects, we assume the orbital occupation to be determined purely by on-site interactions, while off-site processes can be neglected due to the much smaller amplitudes. The problem can thus first be reduced to a single lattice site. The single-site ground state is a superposition of many-particle states $\Psi(n)=\sum_{N, M} c_{N,M} \ket{N}\ket{M}$. Here, $\ket{N}\ket{M}=\ket{n_0,n_1,...}\ket{m_0,m_1,...}$ is the product state with $n_\alpha$ bosons and $m_\alpha$ fermions in the Wannier orbital $w_{{\mathrm{B}}/{\mathrm{F}}}^{{(\alpha)}}{(\mathbf{r})}$, where $\alpha$ indicates the orbital. The state $\Psi(n)$ consists of the lowest single-particle band dressed with small contributions of higher bands. At zero temperature the particles will exclusively occupy the dressed band $\Psi(n)$ instead of the lowest single-particle band.
In general, any multi-orbital two-site operator acting on sites ${\mathrm{L}}$ and ${\mathrm{R}}$ can be decomposed in operators of the form $$\hat O = \sum_{\{\alpha\},\{ \beta\}} A^{\{\alpha\},\{ \beta\}} \hat O_{\mathrm{L}}^{\{\alpha\}} \hat O_{\mathrm{R}}^{\{\beta\}},$$ where the summation is over all possible sets of orbitals $\{\alpha\}=\{\alpha_1,\alpha_2,... \}$ and $\{\beta\}=\{\beta_1,\beta_2,... \}$. $A^{\{\alpha\},\{ \beta\}}$ is the amplitude for the corresponding process and $\hat O_i^{\{\alpha\}}$ consists of creation and annihilation operators ${\hat{b}_\text{$i$}^{(\alpha_k)\dagger}}$ and ${\hat{b}_\text{$i$}^{(\alpha_k)}}$ for particles on site $i$ in the orbital $\alpha_k$. In the simplest case – the conventional tunneling from the right to the left site – we have only sets with a single orbital $\{\alpha\}=\alpha$ and $\{\beta\}=\beta$, the operators on the left and the right site $\hat O_{\mathrm{L}}^{\alpha} = {\hat{b}_\text{L}^{(\alpha)\dagger}}$, $\hat O_{\mathrm{R}}^{\beta} = {\hat{b}_\text{R}^{(\beta)}}$, and tunneling amplitudes $A^{\alpha,\beta}$ as defined below. Instead of collapsing this operator to the lowest single-particle band, we reduce it to the dressed band [@Luhmann2012; @Bissbort2011] given by the many-particle ground state $\Psi(n)$. The effective operator $\tilde O$ in the dressed band takes the form $$\tilde O = \tilde A_{{{n_\mathrm{L}}},{{n_\mathrm{R}}}} \tilde O_{\mathrm{L}}\tilde O_{\mathrm{R}},$$ where $\tilde O_i$ is obtained from $\hat{O_i}^{\{\alpha\}}$ by replacing all creation and annihilation operators ${\hat{b}_\text{$i$}^{(\alpha_k)\dagger}}$ and ${\hat{b}_\text{$i$}^{(\alpha_k)}}$ by their counterparts of the dressed band $\tilde b_i^\dagger$ and $\tilde b_i$. The operators of the dressed band fulfill the usual relations $\tilde b_i \ket{\Psi(n)}_i = \sqrt{n} \ket{\Psi(n-1)}_i$ and $\tilde b_i^\dagger \ket{\Psi(n)}_i = \sqrt{n+1} \ket{\Psi(n+1)}_i$.
[![**a** Contributions to the total tunneling in the lowest band (dashed) and the dressed band (solid) for one bosons and one fermion on each site. The total tunneling (black) is composed of conventional tunneling $J$ (green), bosonic bond-charge tunneling $X_{\mathrm{BB}}$ (red), and fermionic bond-charge tunneling $X_{\mathrm{BF}}$ (blue). The absolute value of all matrix elements are enhanced by orbital degrees of freedom but due to their opposite signs the multi-orbital effects partly compensate each other. **b** Deviations of the total tunneling $\tilde{J}^\mathrm{tot}_{n,n} = \tilde J_{\mathrm{B}}+ (n_{\mathrm{L}}+n_{\mathrm{R}}-1)\tilde X_{\mathrm{BB}}+ 2 \tilde X_{\mathrm{BF}}$ from the standard Hubbard tunneling $J_{\mathrm{B}}$ for different bosonic occupations $n_{\mathrm{L}}=n_{\mathrm{R}}$. \[fig:tunneling\_contributions\] ](Figure4.pdf "fig:"){width="0.8\linewidth"}]{}
The effective amplitude $\tilde A_{{{n_\mathrm{L}}},{{n_\mathrm{R}}}}$ is obtained from the matrix element $\bra{\Psi_{\mathrm{F}}} \hat O \ket{\Psi_{\mathrm{I}}}$, where $\Psi_{\mathrm{I}}({{n_\mathrm{L}}},{{n_\mathrm{R}}})$ denotes the initial and $\Psi_{\mathrm{F}}=\Psi({{n_\mathrm{L}}}',{{n_\mathrm{R}}}')$ the final state of the process. It thereby includes the summation over all multi-orbital processes. Since the states are product states of the individual lattice sites $\ket{\Psi({{n_\mathrm{L}}})}\ket{\Psi({{n_\mathrm{R}}})}$, also the effective amplitude $\tilde A$ decomposes into individual site contributions $$\begin{split}
\tilde A_{{{n_\mathrm{L}}},{{n_\mathrm{R}}}} \!= \!\frac{1}{N} \!\sum_{\{\alpha\},\{ \beta\}} \!\! A^{\{\alpha\},\{ \beta\}} &\bra{\Psi({{n_\mathrm{L}}}')} \hat O_{\mathrm{L}}^{\{\alpha\}} \ket{\Psi({{n_\mathrm{L}}})} \\
\times & \bra{\Psi({{n_\mathrm{R}}}')} \hat O_{\mathrm{R}}^{\{\beta\}} \ket{\Psi({{n_\mathrm{R}}})} ,
\end{split}$$ where $N=\bra{\Psi_{\mathrm{F}}} \tilde O_{\mathrm{L}}\tilde O_{\mathrm{R}}\ket{\Psi_{\mathrm{I}}}$ is needed for the correct normalization. Note that the effective amplitude is intrinsically occupation-dependent.
As an example, for the conventional single-particle tunneling of bosons it follows $A^{\alpha,\beta}=-J_{\alpha} \delta_{\alpha,\beta}$ and $N=\sqrt{{{n_\mathrm{L}}}}\sqrt{{{n_\mathrm{R}}}+1}$. Here, $J_{\alpha}=-\bra{w^{(\alpha)}} \frac{\mathbf{p}^2}{2m} + V{(\mathbf{r})}\ket{w^{(\alpha)}}$ is the tunneling amplitude in band $\alpha$. As another example, the multi-orbital bosonic bond-charge operator $$\begin{split}
\hat X_{\mathrm{BB}}= \ \sum_{\{\alpha\},\{\beta\}} &X_{\mathrm{BB}}^{\{\alpha\},\{ \beta\}}\, {\hat{b}_\text{L}^{(\alpha_1)\dagger}}
{\hat{b}_\text{L}^{(\alpha_2)\dagger}}{\hat{b}_\text{L}^{(\alpha_3)}}{\hat{b}_\text{R}^{(\beta_1)}}+\\
\sum_{\{\alpha'\},\{\beta'\}} &X_{\mathrm{BB}}^{\{\alpha'\},\{ \beta'\}}\, {\hat{b}_\text{L}^{(\alpha_1')\dagger}}
{\hat{b}_\text{R}^{(\beta_1')\dagger}}{\hat{b}_\text{R}^{(\beta_2')}}{\hat{b}_\text{R}^{(\beta_3')}}
\end{split}$$ decomposes in left and right part $\hat O_{\mathrm{L}}^{\{\alpha\}}$ and $\hat O_{\mathrm{R}}^{\{\beta\}}$, that consist of either one or three creation/annihilation operators. In contrast to the conventional tunneling, orbital-changing processes are allowed for the multi-orbital bond-charge operator.
The renormalized amplitudes are depicted in [Fig. \[fig:tunneling\_contributions\]]{} for all relevant contributions to the total tunneling in the case of one boson and one fermion per site (solid lines). The amplitudes for conventional tunneling $\tilde J_{n_{\mathrm{L}}, n_{\mathrm{R}}}$ (green line) as well as bond-charge induced tunneling $\tilde X_{{\mathrm{BB}},n_{\mathrm{L}}, n_{\mathrm{R}}}$ (red) and $\tilde X_{{\mathrm{BF}},n_{\mathrm{L}}, n_{\mathrm{R}}}$ (blue) can differ strongly from the lowest band values, which are indicated as dashed lines. The multi-orbital renormalization can enhance the conventional tunneling by up to $30\%$ and the bond-charge induced processes can even be twice as strong as in the lowest single-particle band.
On-site problem and on-site interaction
---------------------------------------
This section is dedicated to the explicit solution of the many-body problem on a single lattice site for a given number of particles. For the multi-orbital dressing procedure (section \[sec:MO\].A) we assume the ground state of the interacting system to be a product state of these solutions. We apply the method of exact diagonalization to compute the ground state $\Psi(n)$ of $n$ bosons and one fermion on a single lattice site and thereby obtain the solution of the full lattice problem. In particular, this leads directly to an occupation-number-dependent on-site energy $E_n$. The respective single-site Hamiltonian reads $$\begin{split}
\label{eq:Hsite}
\hat{H}_\text{site}=&\sum_\alpha \epsilon_{\mathrm{B}}^{{(\alpha)}}\hat{n}^{{(\alpha)}}+ \sum_\alpha \epsilon_{\mathrm{F}}^{{(\alpha)}}\hat{m}^{{(\alpha)}}\\
+ \frac12 &\sum_{{\alpha\beta\gamma\delta}} U_{\mathrm{BB}}^{({\alpha\beta\gamma\delta})}\, {\hat{b}_\text{}^{(\alpha)\dagger}}{\hat{b}_\text{}^{(\beta)\dagger}}{\hat{b}_\text{}^{(\gamma)}}{\hat{b}_\text{}^{(\delta)}} \\
+ &\sum_{{\alpha\beta\gamma\delta}} U_{\mathrm{BF}}^{({\alpha\beta\gamma\delta})}\, {\hat{b}_\text{}^{(\alpha)\dagger}}{\hat{f}_\text{}^{(\beta)\dagger}}{\hat{f}_\text{}^{(\gamma)}}{\hat{b}_\text{}^{(\delta)}},
\end{split}$$ with the bosonic particle number operator $\hat n^{{(\alpha)}}={\hat{b}_\text{}^{(\alpha)\dagger}}{\hat{b}_\text{}^{(\alpha)}}$ and the single-particle energies $\epsilon_{\mathrm{B}}^{{(\alpha)}}$. The operators $\hat m^{{(\alpha)}}$, ${\hat{f}_\text{}^{(\alpha)\dagger}}$, ${\hat{f}_\text{}^{(\alpha)}}$ and the energy $\epsilon_{\mathrm{F}}^{{(\alpha)}}$ are the fermionic analogues. The multi-band interaction amplitudes are defined as $$\begin{split}
U_{{\mathrm{BB}}/{\mathrm{BF}}}^{({\alpha\beta\gamma\delta})} = {g_{{\mathrm{BB}}/{\mathrm{BF}}}} \int \ \
&w^{{{(\alpha)}}*}_{\mathrm{B}}{(\mathbf{r})}\, w^{(\beta)*}_{{\mathrm{B}}/{\mathrm{F}}}(\mathbf{r'})\, V(\mathbf{r},\mathbf{r'}) \\ \times\,
&w^{(\gamma)}_{{\mathrm{B}}/{\mathrm{F}}}(\mathbf{r'})\, w^{(\delta)}_{\mathrm{B}}(\mathbf{r}) \ \ d^3r \, d^3r'.
\end{split}$$ The interaction potential $V(\mathbf{r},\mathbf{r'})$ is a finite-range box potential with a width of $7.5\,\text{nm}$ (see Ref. [@Luhmann2012] for details). We expand this Hamiltonian in the basis of many-particle Fock-states $\ket{N}\ket{M}$. For $n$ bosons and 1 fermion, the ground state is found by applying an exact diagonalization, the so-called configuration interaction method. We restrict the calculation to the lowest 9 bands per spatial direction and use a high energy cutoff [^1]. The on-site energy is directly obtained as the lowest eigenvalue of the matrix and its contributions can be computed as expectation values of the individual operators in using the corresponding eigenvector.
The energy contributions are plotted in [Fig. \[fig:onsite\_contributions\]]{}**a** for $V_0=15\,{E_\mathrm{R}}$ and $n=3$ bosons (solid lines). The values significantly deviate from the lowest-band approximation (dashed lines). The single-particle energies $n \tilde \epsilon_{\mathrm{B}}+ \tilde \epsilon_{\mathrm{F}}$ (green line) are measured relative to the lowest-band values and thus are always positive. The occupation of higher orbitals causes a contraction of the wave functions which leads to an increase of the absolute value of the Bose-Fermi interaction. This results in a large reduction of the total on-site energy (black line) for large scattering lengths ${a_\mathrm{BF}}$. Of course, also the repulsive interaction among the bosonic particles (red line) contributes but is less drastically influenced.
The total on-site energies for various boson numbers are shown in [Fig. \[fig:onsite\_contributions\]]{}**b**. The dashed lines correspond to a calculation with a three times larger interaction range. In general, the contributions of higher bands are reduced with an increasing interaction range. The figure shows that the energy is only weakly affected by a the change of the (finite) interaction range. Additionally, we applied scaling theory to estimate the value of convergence for the on-site energy at the most demanding parameters $V_0=20\,{E_\mathrm{R}}$ and $a_{\mathrm{BF}}=-300\,a_0$ for an interaction range of $7.5\,\text{nm}$. We see that the error (differences in energies) converges exponentially with both the length of the many-particle basis and the number of orbitals. Scaling our results according to the exponential behavior, we are able to determine the converged energy value, where we first perform the scaling in respect to the basis length for a given number of orbitals. The results indicated as crosses in [Fig. \[fig:onsite\_contributions\]]{}**b** show only small deviations and justify the constraints for the basis length and the number of orbitals applied for the solid lines.
The total on-site energy becomes intrinsically occupation-dependent (beyond the dependency in the standard Bose-Fermi Hubbard model ) and can be written in terms of effective n-particle collisions [@Will2010; @Johnson2009; @Bissbort2011] $$\tilde E_n=n \tilde E_1 + \frac12 n(n-1) \bar E_2 + \frac{1}{6} n(n-1)(n-2) \bar E_3\text{...}$$ The first term $\tilde E_1$ describes the interaction energy by two-particle collisions between bosons and fermions. The second term is the interaction energy caused by processes that involve two bosons $\bar{E}_2=\tilde E_2-2 \tilde{E}_1$ and the third term $\bar{E}_3=\tilde E_3 - 3\bar{E}_2 - 3 \tilde{E}_1$ involves three bosons. Although, the restriction to the first three terms is enough to describe well the energies for up to $n=5$ bosons, we use here the exact values for $\tilde E_n$.
[![**a** Total on-site energy $E_n$ and its contributions for $n=3$ bosons and $m=1$ fermion as functions of the interspecies scattering length. The contributions are single-particle energies $\epsilon_{{\mathrm{B}},{\mathrm{F}}}$ (green), Bose-Bose interaction $U_{\mathrm{BB}}$ (red), and Bose-Fermi interaction $U_{\mathrm{BF}}$ (blue). The total on-site energy is visibly lower than the lowest band prediction (dashed lines). **b** The total on-site energies $E_n$ for various numbers of bosons $n$ and one fermion. With increasing interspecies interaction the total energy decreases non-linearly. The dashed lines correspond to a three times larger interaction range. The markers are obtained by applying scaling theory as described in the text. \[fig:onsite\_contributions\] ](Figure5.pdf "fig:"){width="0.8\linewidth"}]{}
Superfluid to Mott-insulator transition {#sec:models}
=======================================
After discussing off-site interactions and multi-orbital renormalizations, we now turn back to the full many-body quantum gas problem. Obviously, from the above results, it is necessary to take both bond-charge interactions and higher bands into account. First, we now define an extended model of the lowest band that includes off-site interactions and discuss its implications. Afterwards, we replace the lowest single-particle band and parameters with the respective dressed analogues and thereby include higher bands in a very efficient way.
The extended Hubbard model of the lowest band reads $$\begin{split}
\hat H_\text{ext}=& - \sum_{{\langle i,j \rangle}} \big[J_{\mathrm{B}}\!+\! X_{\mathrm{BB}}(\hat n_i + \hat n_j - 1) \!+\! 2\, X_{\mathrm{BF}}\big]\hat b_i^\dagger \hat b_j\\
& +\frac{U_{\mathrm{BB}}}{2} \sum_i \hat n_i (\hat n_i-1) - \mu_{{\mathrm{eff}}}\sum_i\hat n_i.
\end{split}\label{eq:extended_BFHM}$$ While the repulsive interaction between the bosons increases the total tunneling, the attractive fermions reduce the bosonic mobility. As one central result and in strong contrast to the predictions of the standard Hubbard model (see section \[sec:off-site\_interactions\]) the superfluid to Mott-insulator transition is shifted. The phase diagrams are shown in [Fig. \[fig:PD\_SB\]]{} for different attractive Bose-Fermi interaction strengths. For strong Bose-Fermi attraction and low bosonic filling, the transition occurs at much shallower lattices, due to the effectively deepened tunneling-potential. The effect is reversed when the repulsion between the bosons becomes stronger than the attraction to the fermions, which is the case for weaker Bose-Fermi interaction and higher bosonic filling. In [Fig. \[fig:PD\]]{}**b** the critical lattice depth for the transition is plotted as a function of the interspecies interaction strength $a_{\mathrm{BF}}$. The dotted lines correspond to the extended Hamiltonian restricted to the lowest band.
When including higher bands we must replace the lowest-band operators with those of the dressed band $\tilde b_i$ and $\tilde b_i^\dagger$. Also, the parameters $J$, $X$ and $U$ must be renormalized as discussed in section \[sec:MO\]. All tunneling contributions, i.e., conventional tunneling, both bond-charge interactions as well as their multi-orbital renormalizations can be combined to one total tunneling parameter $$\tilde J^{{\mathrm{tot}}}_{n_i, n_j}= \tilde J_{n_i,n_j}+\tilde X_{{\mathrm{BB}},n_i,n_j}(n_i + n_j -1)+2 \tilde X_{{\mathrm{BF}},n_i,n_j},$$ which is explicitly occupation-number dependent. The renormalized on-site energy $$\tilde E_n = n \tilde \epsilon_{\mathrm{B},n} + \tilde \epsilon_{\mathrm{F},n} + \frac12 n(n-1) \tilde U_n + n \tilde U_{\mathrm{BF},n}$$ is composed of the renormalized single particle energies of bosons $\tilde \epsilon_{\mathrm{B},n}$ and fermions $\tilde \epsilon_{\mathrm{F},n}$ as well as the interaction energies for the repulsion between the bosons $\frac12 n(n-1) \tilde U_n$ and the attraction between the species $n \tilde U_{\mathrm{BF},n}$. This allows to define the extended Hamiltonian of the dressed band in equation , namely, $$\tilde H_\mathrm{ext}= -\sum_{\langle i,j \rangle} \tilde{b}_i^\dagger \tilde{b}_j \tilde J^{{\mathrm{tot}}}_{\hat n_j,\hat n_i}
+ \sum_i \tilde E_{\hat{n}_i} - \mu \sum_i \hat{n}_i.$$ The dressed-band Hamiltonian now takes into account all higher-band processes and all relevant nearest-neighbor interactions. The multi-orbital corrections of the Bose-Fermi interaction have a strong impact on the chemical potential at which the transition to a certain Mott-lobe occurs. This distorts and shifts the phase diagram along the axis of the chemical potential. Since this effect is not of any physical interest, we plot [Fig. \[fig:PD\]]{}**a** in terms of an effective chemical potential $\mu_{{\mathrm{eff}}}= \mu - \tilde E_{1}$. Concerning the transition point of the bosonic superfluid to Mott-insulator transition, the reduction of the total on-site energy by multi-orbital processes counters the effect of reduced total tunneling. Nonetheless, the total effect on the transition can be a shift of several recoil energies ${E_\mathrm{R}}$ depending on interaction strengths and filling factors ([Fig. \[fig:PD\]]{}**b**). Note that the lowest-band model $\hat H_\mathrm{ext}$ underestimates the impact on the superfluid to Mott-insulator transition by only up to $1 {E_\mathrm{R}}$, which is surprising keeping the strong changes of the individual amplitudes in mind. However, this (coincidental) compensation of the contributing amplitudes depends on the choice of system parameters. Furthermore, it has been demonstrated that the on-site interaction [@Will2010; @Will2011] and the effective tunneling matrix element [@Heinze2011] are experimentally well accessible and can be measured independently.
[![Phase diagrams of the lowest-band model with bond-charge interactions for different scattering lengths. For comparison the results of the standard Hubbard model are shown as a dashed black line. \[fig:PD\_SB\] ](Figure6.pdf "fig:"){width="0.8\linewidth"}]{}
Density-density interactions and pair-tunneling {#sec:other_processes}
===============================================
In the above model, several off-site processes have been neglected due to their small amplitude in the lowest band. In this context the question arises, whether the multi-orbital dressing can enhance them to non-negligible values. The correlated pair tunneling and the density-density interactions both have very small amplitudes because in the integrand the small tail of the Wannier function enters quadratically on both lattice sites. By contrast, in the bond-charge integral the tail is multiplied three times with the center of a Wannier function. However, when taking into account strongly delocalized wave functions of higher bands, this argument no longer applies and all processes need to be reconsidered, since the overlap integrals become comparable for all types of off-site interactions. These contributions are strongly suppressed in the case of density-density interactions, because the initial and final state both depend on higher-band contributions with small coefficients $c_N \ll 1$ to produce a large integral, whereas in the case of pair tunneling one of them can be the ground state $(c_N \approx 1)$. Although, the multi-orbitally renormalized pair tunneling is usually smaller than the conventional tunneling it can, in general, reach the same order of magnitude. Due to this structure it has very bad convergence properties in respect to the total number of contributing bands. As a fourth order contribution in perturbation theory it can, however, be neglected even for rather large amplitudes. Density-density interactions are not as strongly influenced and remain small. This justifies, the restriction to the extensions incorporated in Hamiltonian , which includes all relevant off-site interactions.
Conclusions {#sec:conclusions}
===========
We have discussed the important role of interaction effects in atomic quantum gas mixtures in optical lattices. Off-site interactions as well as higher band processes turn out to have a strong impact on these systems, which we are able to calculate using an extended occupation-dependent Hubbard model. In particular, we have focused on Bose-Fermi mixtures in this paper, where the standard Bose-Fermi-Hubbard model fails to cover all relevant processes. This manifests itself in a strong shift of the superfluid to Mott-insulator transition in the bosonic subsystem, which is not predicted by the standard Hubbard model. The critical lattice depth is shifted towards shallower lattices with increasing Bose-Fermi attraction. Similar corrections are present for all experiments with optical lattices and can be expected to be relevant, e.g., for Bose-Bose mixtures, low-dimensional systems or other lattice geometries. Omitting the condition of having a fermionic band insulator, which we have applied here, the presented extensions of the Bose-Fermi Hubbard model can lead to very rich physics such as the formation of polarons.
We have shown that for optical lattice systems the bond-charge tunneling [@Hirsch1989; @Strack1993; @Hirsch1994; @Amadon1996; @Mazzarella2006; @Mering2011; @Luhmann2012; @Bissbort2011] is the most important contribution of the nearest-neighbor off-site interaction as it can drastically influence the tunneling. Repulsive interactions enhance the total tunneling, whereas the attractive interactions reduce it. In an intuitive picture, this can be described as lowered and increased effective potentials, respectively. Furthermore, higher band processes not only reduce the total on-site energy but also have an impact on the conventional tunneling and the bond-charge interactions. We have treated the problem by dressing [@Luhmann2012; @Bissbort2011] the lowest single-particle band with interaction-induced occupations of higher-orbital states. This leads to a renormalization of interactions and tunneling parameters that become intrinsically occupation-dependent. These parameters have been used in order to define an extended Bose-Fermi-Hubbard model capable of describing effects of higher orbitals and off-site interactions in Bose-Fermi mixtures.
The results show in general that interactions in multi-component systems can have a crucial impact beyond the standard Hubbard treatment. In the presented case, the standard Hubbard model is incapable of describing the interspecies interaction between bosonic and fermionic atoms correctly. While here mainly the effects on the bosonic atoms have been discussed, the mutual interaction affects the fermionic atoms similarly, which has recently been observed in experiment [@Heinze2011].
Acknowledgments
===============
We thank U. Bissbort and W. Hofstetter for stimulating discussions and acknowledge financial support by the German science foundation DFG under grant FOR801.
[30]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, , , , , ****, ().
, , , , , , , ****, ().
, , , , , , , ****, ().
, , , , , , , , ****, ().
, , , , , , ****, ().
, ****, ().
, , , , ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, , , ().
, , , , ****, ().
, , , ****, ().
, ****, ().
, , , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , , , ****, ().
, ****, ().
, ().
, , , , , ****, ().
[^1]: The resulting many-body product basis has a total length of $12000 n^2$, where $n$ is the number of bosons.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The transport coefficients of strongly interacting matter are currently subject of intense theoretical and phenomenological studies due to their relevance for the characterization of the quark-gluon plasma produced in ultra-relativistic heavy-ion collisions (uRHIC). We discuss the connection between the shear viscosity to entropy density ratio, $\eta/s$, and the electric conductivity, $\sigma_{el}$. Once the relaxation time is tuned to have a minimum value of $\eta/s=1/4\pi$ near the critical temperature $T_c$, one simultaneously predicts $\sigma_{el}/T$ very close to recent lQCD data. More generally, we discuss why the ratio of $(\eta/s)/(\sigma_{el}/T)$ supplies a measure of the quark to gluon scattering rates whose knowledge would allow to significantly advance in the understanding of the QGP phase. We also predict that $(\eta/s)/(\sigma_{el}/T)$, independently on the running coupling $\alpha_s(T)$, should increase up to about $\sim 20$ for $T \rightarrow T_c$, while it goes down to a nearly flat behavior around $\simeq 4$ for $T \geq 4\, T_c$. Therefore we in general predict a stronger T dependence of $\sigma_{el}/T$ with respect to $\eta/s$ that in a quasiparticle approach is constrained by lQCD thermodynamics. A conformal theory, instead, predicts a similar T dependence of $\eta/s$ and $\sigma_{el}/T$.'
author:
- 'A. Puglisi'
- 'S. Plumari'
- 'V. Greco'
title: 'Shear viscosity $\eta$ to electric conductivity $\sigma_{el}$ ratio for the Quark-Gluon Plasma'
---
Relativistic Heavy Ion Collider (RHIC) at BNL and Large Hadron Collider (LHC) at CERN have produced a very hot and dense system of strongly interacting particles as in the Early Universe with temperatures largely above $T_c\simeq 160 \rm MeV$ [@Science_Muller; @Shuryak:2003xe; @Aoki:2006we], the transition temperature from nuclear matter to the Quark-Gluon Plasma (QGP) [@lQCD; @Bazavov:2011nk; @Lombardo:2012ix]. The phenomenological studies by viscous hydrodynamics [@Romatschke:2007mq; @Song:2011hk; @Schenke:2010nt; @Niemi:2011ix] and parton transport [@Ferini:2008he; @Plumari_Bari; @Xu:2007jv; @Xu:2008av; @Cassing:2009vt; @Bratkovskaya:2011wp] of the collective behavior have shown that the QGP has a very $\eta/s$, quite close to the conjectured lower-bound limit for a strongly interacting system in the limit of infinite coupling $\eta/s=1/4\pi$ [@Kovtun:2004de]. This suggests that hot QCD matter could be a nearly perfect fluid with the smallest $\eta/s$ ever observed, even less dissipative than the ultra cold matter created by magnetic traps [@O'Hara:2002zz; @Cao:2010wa]. As for atomic and molecular systems a minimum in $\eta/s$ is expected slightly above $T_c$ [@Csernai:2006zz; @Lacey:2006bc].
Another key transport coefficient, yet much less studied, is $\sigma_{el}$. This transport coefficient represents the linear response of the system to an applied external electric field. Several processes occurring in uRHIC as well as in the Early Universe are regulated by the electric conductivity. Indeed HICs are expected to generate very high electric and magnetic fields ($eE\simeq eB \simeq m_{\pi}^2$, with $m_{\pi}$ the pion mass) in the very early stage of the collisions [@Tuchin; @Hirono]. A large value of $\sigma_{el}$ would determines a relaxation time for the electromagnetic field of the order of $\sim 1-2\, fm/c $ [@McLerran:2013hla; @Gursoy:2014aka], which would be of fundamental importance for the strength of the Chiral-Magnetic Effect [@Fukushima:2008xe], a signature of the CP violation of the strong interaction. Also in mass asymmetric collisions, like Cu+Au, the electric field directed from Au to Cu induces a current resulting in charge asymmetric collective flow directly related to $\sigma_{el}$ [@Hirono]. Furthermore the emission rate of soft photons should be directly proportional to $\sigma_{el}$ [@Kapusta_book; @Turbide:2003si; @Linnyk:2013wma]. Despite its relevance there is yet only a poor theoretical and phenomenological knowledge of $\sigma_{el}$ and its temperature dependence. First preliminary studies in lQCD has extracted only few estimates with large uncertainties [@Gupta; @Aarts] and only recently more safe extrapolation has been developed [@Amato; @Ding; @Brandt].
In this Letter, we point out the main elements determining $\sigma_{el}$ for a QGP plasma and in particular its connection with $\eta$. In fact, while one may expect that the QGP is quite a good conductor due to the deconfinement of color charges, on the other hand, the very small $\eta/s$ indicates large scattering rates which can largely damp the conductivity, especially if the plasma is dominated by gluons that do not carry any electric charge.
The electric conductivity can be formally derived from the Green-Kubo formula and it is related to the relaxation of the current-current correlator for a system in thermal equilibrium. It can be written as $\sigma_{el}=V/(3\,T)\, \langle \vec J (t=0)\cdot \vec J (t=0)\rangle \cdot \tau$, where $\tau$ is the relaxation time of the correlator whose initial value can be related to the thermal average $\frac{ \rho\, e^2}{3T}\langle p^2/E^2 \rangle $ [@FernandezFraile:2005ka], with $\rho$ and $E$ the density and energy of the charge carriers. Generalizing to the case of QGP one can write: $$\sigma_{el}=\frac{e^2}{3T} \left \langle \frac{{\vec p}^{\,2}}{E^2}\right\rangle \sum_{j=q,\bar q} f_j^2 \,\tau_j \rho_j=
\frac{e_\star^2}{3T} \left \langle \frac{{\vec p}^{\,2}}{E^2}\right\rangle \tau_q \rho_q
\label{conductivity_qgp}$$ where $e_\star^2=e^2\sum_{j=u,d,s}^{\bar{u},\bar{d},\bar{s}}f_j^2=4e^2/3 $ with $f_j$ the fractional quark charge. Eq. (\[conductivity\_qgp\]) in the non-relativistic limit reduces formally to the Drude formula $\frac{\tau e^2 \rho}{m}$, even if we notice that $\tau$ in Eq.(\[conductivity\_qgp\]) has not to be equal to $1/(\sigma\rho)$ as in the Drude model. The relaxation time of a particle of species $j$ in terms of cross-sections and particle densities can be written in the relaxation time approximation (RTA) as $ \tau_j^{-1}=\sum_{i=q, \bar q , g} \langle \rho_i v_{rel}^{ij} \sigma^{ij}_{tr}\rangle$ where $j=q, \bar q$ while the sum runs over all particle species with $\rho_i$ the density of species $i$, $v_{rel}^{ij}$ is the relative velocity and $\sigma^{ij}_{tr}$ is the transport scattering cross-section. In Ref. [@Cassing_el] it has been shown that RTA is able to describe with quite good approximation $\sigma_{el}$ in agreement with numerically simulation of the Dynamical QP model (DQPM) known as PHSD, see also more generally for a numerical approach Ref.s [@Greif:2014oia; @Puglisi:2014sha].
As done within the Hard-Thermal-Loop (HTL) approach, we will consider the total transport cross section regulated by a screening Debye mass $m_D=g(T)T$, with $g(T)$ being the strong coupling: $$\label{pQCD_cross-section}
\sigma_{tr}^{ij}(s)=\int \frac{d\sigma}{dt} \sin^2 \Theta\,dt=\beta^{ij}\frac{\pi \alpha^2_s}{m_D^2}\frac{s}{s+m_D^2}h(a)$$ where $\alpha_s=g^2/4\pi$, the differential cross section $\frac{d\sigma}{d t}=\frac{d\sigma}{d q^2}\simeq \alpha_s^2/(q^2+m_D^2)^2$ where $q^2=\frac{s}{2}(1-\cos \theta)$. The function $h(a)=4a(1+a)[(2a+1)\ln(1+1/a)-2]$, with $a=m_D^2/s$ accounts for the anisotropy of the scatterings: for $m_D\to \infty$, $h(a)\to 2/3$ and one recovers the isotropic limit The coefficient $\beta^{ij}$ depends on the pair of interacting particles: $\beta^{qq}=16/9$, $\beta^{qq'}=8/9$, $\beta^{qg}=2$, $\beta^{gg}=9$. These factors are directly related to the quark and gluon Casimir factor, for example $\beta^{qq}/\beta^{gg}=(C_F/C_A)^2=(4/9)^2$.
The shear viscosity $\eta$ is known from the Green-Kubo relation to be given by $\eta=V/T\, \langle \Pi_{xy}^2(t=0)\rangle\,\cdot \tau$, where the initial value of the correlator of the transverse components of the energy-momentum tensor can be written as $\frac{\rho}{15 T} \langle p^4/E^2 \rangle$ [@Plumari_visco; @Wesp:2011yy; @Fuini:2010xz]. Hence for a system with different species can be written as [@Sasaki; @Kapusta_qp]: $$\eta=\frac{1}{15 T} \left\langle \frac{p^4}{E^2} \right\rangle \left( \tau_q \rho^{tot}_q + \tau_g \rho_{g}\right)
\label{viscosity}$$ where the relaxation time $\tau_g$ has a similar expression as above with $j=g$ while $\rho^{tot}$ is the sum of all quarks and antiquarks flavour density. The thermodynamical averages entering Eq.s (\[conductivity\_qgp\]) and (\[viscosity\]), will be fixed employing a quasi-particle (QP) model tuned to reproduce the lattice QCD thermodynamics [@Plumari_qpmodel], similarly to [@Levai:1997yx; @Peshier:2002ww; @Bluhm:2010qf; @Bluhm:2004xn]. The quark and gluon masses are given by $m^2_g= 3/4\,g^2 T^2$ and $m^2_q=1/3\,g^2 T^2$ in terms of a running coupling $g(T)$ that is determined by a fit to the lattice energy density, which allows to well describe also the pressure $P$ and entropy density $s$ above $T_c=160 \, \rm MeV$. In Ref. [@Plumari_qpmodel] we have obtained: $$g^2_{QP}(T)={48 \pi^2}/{\left( 11 N_c - 2 N_f\right) \ln\left[ \lambda \left( \frac{T}{T_c}-\frac{T_s}{T_c} \right) \right]^2}$$ with $\lambda = 2.6$, $T_s/T_c=0.57$. We warn that the previous equation is a good parametrization only for $T>1.1\,T_c$. We notice that a self-consistent dynamical model (DQPM), that includes also the pertinent spectral function, has been developed in [@Cassing:2009vt] and leads to nearly the same behavior of the strong coupling $g(T)$. We will consider the DQPM explicitly, showing that the considerations elaborated in this Letter are quite general and can be only marginally affected by particle width. We notice that the only approximation made in deriving Eq.(\[viscosity\]) is to consider $\langle p^4/E^2\rangle$ equal for quarks and gluons. We have verified that $\langle p^4/E^2 \rangle_{g}\simeq \langle p^4/E^2 \rangle_q$ within a 5$\%$ in the QP model but also more generally even when $m_q$ and $m_g$ are largely different but $m_{q,g} \lesssim 3T$, which means that Eq.(3) is valid also for light and strange current quark masses and massless gluons. The $\langle p^4/E^2\rangle$ in a massless approximation is simply $4\epsilon\, T /\rho$, we have checked that the validity of this expression is kept using the QP model (i.e. massive excitation) with a discrepancy of about $2\%$. Hence the first term in Eq. (\[viscosity\]) is determined by the lQCD thermodynamics and does not rely on the detailed $m_{q,g}(T)$ in the QP model. We note that even if the QP model is able to correctly describe the thermodynamics it is not obvious that it correctly describes dynamical quantities like the relaxation times with the same coupling $g(T)$ employed to fit the thermodynamics. However our key point will be to find a quantity independent on $g(T)$, see Eq.(5).
For its general interest and asymptotic validity for $T\rightarrow \infty$, we also consider the behavior of the pQCD running coupling constant for the evaluation of transport relaxation time: $g_{pQCD}(T)=\frac{8\pi^2}{9} \ln^{-1} \left(\frac{2\pi T}{\Lambda_{QCD}} \right)$. On one hand, close to $T_c$, such a case misses the dynamics of the phase transition, on the other hand it allows to see explicitly what is the impact of a different running coupling.
![Shear viscosity to entropy density ratio $\eta/s$: dashed line represents QP model results, dot-dashed line is pQCD, stars is DQPM [@Marty]. Red thick solid line and blue thin solid line are obtained rescaling $g(T)$. Blue dotted line is AdS/CFT result from [@Kovtun:2004de]. Symbols are lattice date: full squares [@Meyer], diamonds and triangles [@Nakamura], open and full circles [@Sakai].[]{data-label="fig:eta_s_Lattice"}](./eta_s_Lattice.eps)
The $\eta/s$ calculated is shown in Fig. \[fig:eta\_s\_Lattice\]: red dashed line is the result for the QP model using $g_{QP}(T)$ for relaxation times and transport coefficient, blue dot-dashed line labeled as $g_{pQCD}$, means that we used the pQCD running coupling for evaluating the relaxation time, green stars are the DQPM [@Marty] and by symbols several lQCD results. We warn that the different lQCD data are obtained with different methods and actions. The main difference between our QP model and DQPM comes from the fact that the latter assumes isotropic scatterings which decrease the relaxation time by about $30-40\%$. Anyway, the $\eta/s$ predicted is toward higher value with respect to the conjectured minimum value of $\eta/s\sim 0.08$, supported also by several phenomenological estimates [@Romatschke:2007mq; @Song:2011hk; @Schenke:2010nt; @Niemi:2011ix; @Ferini:2008he]. However within the QP model it has been discussed in the literature also another approach for $\tau$ where the relaxation times are $\tau_{q,g} = C_{q,g} \,g^4 T \ln (a/g^2)$ [@Khvorostukhin] with $C_{q,g}$ and $a$ fixed to reproduce both the pQCD estimate asymptotically [@Arnold:2003] and a minimum for $\eta/s(T)=1/4\pi$ [@Plumari_qpmodel; @Bluhm:2010qf]. In the T region of interest, the result is quite similar to upscaling the coupling $g(T)$ by a $k$-factor in such a way to have the minimum of $\eta/s(T)=1/4\pi$. Therefore we do not employ the above parametrization but compute the transport coefficients using the definition of $\tau$ of Eq.(\[tau\_transport\]), where enters the cross section in Eq. (\[pQCD\_cross-section\]) with the coupling upscaled. The corresponding curves are shown in Fig. \[fig:eta\_s\_Lattice\] by red thick solid line for the $g_{QP}(T)$ coupling (rescaled by $k=1.59$) and by blue thin solid line for the $g_{pQCD}(T)$ (rescaled by $k=2.08$). One obtains $\tau_g \simeq \tau_q/2 \sim 0.2 \rm\, fm/c$ and also $\eta/s(T)$ roughly linearly rising with $T$ in agreement with quenched lQCD estimates, full circles [@Sakai].
A main point we want to stress is that, once the relaxation time is set to an $\eta/s(T)=0.08$, the $\sigma_{el}/T$ predicted, with the same $\tau_q$ as for $\eta/s$, is in quite good agreement with most of the lQCD data, shown by symbols in Fig. \[fig:K\_el\_cond\] (see caption for details). Therefore a low $\sigma_{el}/T$ is obtained at variance with the early lQCD estimate, Ref. [@Gupta], as a consequence of the small $\tau_{q,g}$ entailed by $\eta/s \simeq 0.08$. In Fig. \[fig:K\_el\_cond\], we show also the predictions of DQPM (green stars) [@Cassing_el; @Marty].
![Electric conductivity $\sigma_{el}/T$ as a function of $T/T_c$: red dashed line represents QP model results, blue dot-dashed line is pQCD, red thick solid line and blue thin solid line are respectively QP and pQCD considering the rescaled $g(T)$ in order to reproduce the minimum of $\eta/s$. Green line are AdS/CFT results from [@Huot]. Green stars represent DQPM [@Cassing_el]. Symbols are Lattice data: grey squares [@Gupta], violet triangles [@Ding], green circle [@Brandt], yellow diamond [@Aarts], orange square[@Buividovich:2010tn] and red diamonds [@Amato].[]{data-label="fig:K_el_cond"}](./el_cond_K_Tc.eps)
In Fig.\[fig:K\_el\_cond\], we also plot by green dotted line the ${\cal N}=4$ Super Yang Mills electric conductivity [@Huot] that predicts a constant behavior for $\sigma_{el}/T=e^2 N_c^2 / (16 \pi)$. We note that in our framework one instead expects that, even if the $\eta/s$ is independent on the temperature, the $\sigma_{el}$ should still have a strong T-dependence. This can be seen noticing that one can write approximately, $\eta/s \simeq T^{-2} \tau \rho$, being $\langle p^4/E^2\rangle \simeq \epsilon T/\rho$, and $\sigma_{el}/T\simeq T/m(T) \eta/s$, being $\langle p^2/E^2 \rangle \simeq T/m(T)$, which means an extra $T$ dependence for $\sigma_{el}$ leading to a steep decrease of $\sigma_{el}/T$ close to $T_c$. $m(T)$ increases as $T\to T_c$ because it is fitted to reproduce the decrease of energy density $\epsilon$ in lQCD. We notice that for a conformal theory $T^{\mu}_{\mu}=\epsilon - 3P=0$, as for massless particles, one has $\sigma_{el}/T\sim \eta/s$ like found in AdS/CFT. It seems that the large interaction measure is the origin of such extra T dependence of $\sigma_{el}/T$ with respect to $\eta/s$. This indication is corroborated also by the recent result in AdS/QCD [@Noronha] that presents a similar strong T dependence for $T<2-3 T_c$ at variance with AdS/CFT.
The $\sigma_{el}$ appears to be self-consistent with a minimal $\eta/s$, but the specific $T$ dependence of both are largely dependent on the modeling of $\tau_{q,g}$, we point out that the ratio $(\eta/s)/(\sigma_{el}/T)$ can be written, from Eq. (\[conductivity\_qgp\]) and Eq. (\[viscosity\]), as: $$\begin{aligned}
\label{ratio}
\frac{\eta/s}{\sigma_{el}/T}=\frac{6}{5}\frac{ T \langle p^2/E^2 \rangle^{-1}}{s\,e_\star^2 } \left\langle \frac{p^4}{E^2} \right\rangle \left( 1+ \frac{\tau_g}{\tau_q} \frac{\rho_g}{\rho^{tot}_{q}} \right).\end{aligned}$$ in terms of generic relaxation times. Eq.(\[ratio\]) is quite general and does not rely on specific features or validity of the quasi-particle model. A main feature of such a ratio is its independence on the $k$-factor introduced above, and, more importantly, even on the $g(T)$ coupling as we can see writing explicitly the transport relaxation time for quarks and gluons: $$\begin{aligned}
\label{tau_transport}
\tau_q^{-1} =\langle \sigma(s)_{tr} v_{rel} \rangle (\rho_q \sum_{i=u,d,s}^{\bar{u},\bar{d},\bar{s}} \beta^{qi} + \rho_g \beta^{qg}) \nonumber\\
\tau_g^{-1}= \langle \sigma(s)_{tr} v_{rel} \rangle \left(\rho_q^{tot} \beta^{qg} + \rho_g\beta^{gg}\right)\end{aligned}$$ where the $\beta^{ij}$ were defined above. Hence the ratio of transport relaxation times appearing in Eq. (\[ratio\]) can be written as: $$\begin{aligned}
\label{tau-ratio}
\frac{\tau_g}{\tau_q} = \frac{{C^q}+ \frac{\rho_g}{\rho_q}}{6 + \frac{\rho_g}{\rho_q} C^g }\end{aligned}$$ where the coefficients $C^{q}=(\beta^{qq} + \beta^{q\bar{q}} + 2 \beta^{q\bar{q}'} + 2 \beta^{qq'})/\beta^{qg}$ and $C^{g}=\beta^{gg}/\beta^{qg}$ are the relative magnitude between quark-(anti-)quark and $gg$ with respect to $q(\bar q)g$ scatterings.
![Shear viscosity $\eta/s$ to $\sigma/T$ ratio as a function of $T/T_c$: red solid line is the QP model, blue dashed line pQCD, green stars DQPM [@Marty]. Orange line is obtained using $C^q=10\,C^q_{pQCD}$, black thin line $C^g=10\,C^g_{pQCD}$. Green dotted line represent AdS/CFT results [@Kovtun:2004de; @Huot]. Symbols are obtained using available lattice data (see text for details).[]{data-label="fig:ratio"}](./ratio_new.eps)
Using the standard pQCD factors for $\beta_{ij}$, $C^{q}|_{pQCD}=\frac{28}{9}\simeq3.1$ and $C^{g}|_{pQCD}=\frac{9}{2}$.
In Fig. \[fig:ratio\] we show $(\eta/s)/(\sigma_{el}/T)$ as a function of $T/T_c$: the red thick solid line is the prediction for the ratio using $g_{QP}(T)$, but it is clear from the Eq. (\[ratio\]) that the ratio is completely independent on the running coupling itself; the result for $g_{pQCD}(T)$ is shown by blue dashed line. The ratio is instead sensitive just to the relative strength of the quark (anti-quark) scatterings with respect to the gluonic ones, hence we suggest that a measurement in lQCD can shed light on the relative scattering rates of quarks and gluons, providing an insight into their relative role. It is not known if such ratios, linked to the Casimir factors of $SU(3)_c$, are kept also in the non-perturbative regime, which may be not so unlikely [@Nakamura:2005hk]. We remark that we have computed the ratio in a very large temperature range $1-10\,T_c$: at large temperatures ($T> 5-10\,T_c$) deviation from the obtained value, $(\eta/s)/(\sigma_{el}/T)\simeq 3$, would be quite surprising, on the other hand for $T<1.2-1.5\, T_c$ one may cast doubts on the validity of the Casimir coefficients. In the following we evaluate also the impact of modified Casimir Coefficients. As $T\rightarrow T_c$ a steep increase is predicted that is essentially regulated by $\langle p^2/E^2 \rangle$. It is interesting to notice that in the massless limit (conformal theory) the factor before the parenthesis in Eq.(\[ratio\]) becomes a temperature independent constant and hence also the ratio. This is in quite close agreement with the AdS/CFT prediction shown by dotted line in Fig. \[fig:ratio\].
We also briefly want to mention that one possible scenario could be that when the QGP approaches the phase transition, the confinement dynamics becomes dominant and the $q\bar q$ scattering, precursors of mesonic states, and di-quark $qq$ states, precursor of baryonic states, are strongly enhanced by a resonant scattering with respect to other channels, as found in a T-matrix approach in the heavy quark sector [@vanHees:2007me]. For this reason, we explore the sensitivity of the ratio $(\eta/s)/(\sigma_{el}/T)$ on the magnitude of $C^{q}$ and $C^{g}$. The orange solid line shows the behavior for an enhancement of the quark scatterings, $C^q=10\,C^q_{pQCD}$. We can see in Fig. \[fig:ratio\] that this would lead to an enhancement of the ratio by about a $40\%$. We also see that instead the ratio is not very sensitive to a possible enhancement of only the $gg$ scattering with respect to the $q\bar q,qq,qg$; in fact even for $C^g=10 C^g_{pQCD}$ one obtains the thin black solid line. This is due to the fact that already in the pQCD case $\tau_g/\tau_q \sim 0.3-0.4$. Furthermore already in the massless limit $\rho_g/\rho_q^{tot} \simeq d_g/d_{q+\bar q}=4/9$ even not dwelling on the details of the QP model where the larger gluon mass further decreases this ratio. Therefore the second term in parenthesis in Eq. (\[ratio\]) is of the order of $10^{-1}$ and further decrease of its value would not be visible because the ratio is anyway dominated by the first term equal to one. We reported in Fig. \[fig:ratio\] also the ratio from the DQPM model, as deduced from [@Marty] and we can see that, even if it is not evaluated through Eq. (\[ratio\]), it is in very good agreement with our general prediction. In Fig. \[fig:ratio\] we also display by symbols the ratio evaluated from the available lQCD data, considering for $4\pi \eta/s \lesssim 4$ while for $\sigma_{el}/T$ we choose red diamonds [@Amato] as a lower limit (filled symbols) and the others in Fig. \[fig:K\_el\_cond\] as an upper limit (open symbols), excluding only the grey squares [@nota-gupta]. To compute $(\eta/s)/(\sigma_{el}/T)$ we do an interpolation between the data point of $\sigma_{el}$. We warn to consider these estimates only as a first rough indications, in fact the lattice data collected are obtained with different actions among them and have quite different $T_c$ with respect to the most realistic one, $T_c\sim 160\,MeV$ [@lQCD; @Bazavov:2011nk], that we employed to tune the QP model [@Plumari_qpmodel]. In this Letter we point out the direct relation between the shear viscosity $\eta$ and the electric conductivity $\sigma_{el}$. In particular, we have discussed why most recent lQCD data [@Amato; @Ding; @Brandt] predicting an electric conductivity $\sigma_{el}\simeq 10^{-2} T$ (for $T < 2\, T_c$) , appears to be consistent with a fluid at the minimal conjectured viscosity $4\pi\eta/s\simeq 1$, while the data of Ref. [@Gupta] appear to be hardly reconcilable with it. Also a steep rise of $\sigma_{el}/T$, in agreement with lQCD data, appears quite naturally in the quasi-particle approach as inverse of the self-energy determining the effective masses needed to correctly reproduce the lQCD thermodynamics. This result is at variance with the AdS/CFT [@Huot], but our analysis suggests that it is due to the conformal thermodynamics that does not reflect the QCD one. It is quite interesting that an AdS/QCD approach [@Noronha], able to correctly describe the interaction measure of lQCD, also modify the AdS/CFT result predicting a strong T dependence of $\sigma_{el}/T$ for $T<2-3\,T_c$. We note that the extra T dependence predicted for $\sigma_{el}/T$ with respect to $\eta/s$ is determined by the $\langle p^2/E^2 \rangle$ constrained to reproduce the lQCD thermodynamics. If instead one imposes conformality with $m=0$, this leads to $\langle p^2/E^2 \rangle=1$ and the T dependence of $\eta/s$ becomes quite similar to the one of $\sigma_{el}/T$ apart from differences that can arise between quark and gluon relaxation times.
We identify the dimensionless ratio $(\eta/s)/(\sigma_{el}/T)$ as not affected by the uncertainties in the running coupling $g(T)$. Moreover due to the fact that gluons do not carry an electric charge, the ratio is regulated by the relative strength and chemical composition of the QGP through the term $(1+ \tau_g \rho_g/\tau_q \rho_q^{tot})$. Our analysis provides the baseline of such a ratio that in this decade will most likely be more safely evaluated thanks to the developments of lQCD techniques. This will provide a first and pivotal insight into the understanding of the relative role of quarks and gluons in the QGP. Deviations from our predictions for $(\eta/s)/(\sigma_{el}/T)$ especially at high temperature $T\gtrsim 2-3\,T_c$, where a quasi-particle picture can be derived from QCD within the HTL scheme [@Andersen:2010wu], would be quite compelling.
V.G. acknowledge the support of the ERC-StG Grant under the QGPDyn project. We thanks M. Ruggieri for carefully reading the manuscript.
[99]{}
B. V. Jacak and B. Muller, *Science* [**337**]{}, 310 (2012).
E. Shuryak, Prog.Part.Nucl.Phys. [**53**]{}, 273 (2004).
Y. Aoki, G. Endrodi, Z. Fodor, S. D. Katz and K. K. Szabo, Nature [**443**]{} (2006) 675
S. Borsanyi et al., JHEP 1009 073 (2010).
A. Bazavov, T. Bhattacharya, M. Cheng, C. DeTar, H. T. Ding, S. Gottlieb, R. Gupta and P. Hegde [*et al.*]{}, Phys. Rev. D [**85**]{} (2012) 054503.
M. P. Lombardo, PoS LATTICE [**2012**]{} (2012) 016
P. Romatschke and U. Romatschke, Phys.Rev.Lett. [**99**]{}, 172301 (2007).
H. Song, S. A. Bass, U. Heinz, T. Hirano and C. Shen, Phys. Rev. C [**83**]{}, 054910 (2011) \[Erratum-ibid. C [**86**]{}, 059903 (2012)\].
B. Schenke, S. Jeon and C. Gale, Phys. Rev. C [**82**]{} (2010) 014903.
H. Niemi, G. S. Denicol, P. Huovinen, E. Molnar and D. H. Rischke, Phys. Rev. Lett. [**106**]{} (2011) 212302. G. Ferini, M. Colonna, M. Di Toro, and V. Greco, Phys.Lett. [**B670**]{}, 325 (2009).
Z. Xu, C. Greiner, and H. Stocker, Phys. Rev. Lett. [**101**]{}, 082302 (2008).
Z. Xu and C. Greiner, Phys. Rev. [**C79**]{}, 014904 (2009).
W. Cassing and E. Bratkovskaya, Nucl. Phys. [**A831**]{}, 215 (2009).
E. Bratkovskaya, W. Cassing, V. Konchakovski, and O. Linnyk, Nucl. Phys. [**A856**]{}, 162 (2011).
S. Plumari and V. Greco, AIP Conf. Proc. [**1422**]{}, 56 (2012).
P. Kovtun, D. Son, and A. Starinets, Phys. Rev. Lett. [**94**]{}, 111601 (2005).
C. Cao, E. Elliott, J. Joseph, H. Wu, J. Petricka, T. Schäfer and J. E. Thomas, Science [**331**]{} (2011) 58.
O’Hara K M, Hemmer S L, Gehm M E, Granade S R and Thomas J E 2002, [*Science*]{} [**298**]{} 2179.
L.P. Csernai, J. Kapusta, and L.D. McLerran, Phys. Rev. Lett. [**97**]{}, 152303 (2006).
R. A. Lacey [*et al.*]{}, Phys. Rev. Lett. [**98**]{}, 092301 (2007).
K. Tuchin, *Adv.High Energy Phys.* 2013 (2013) 490495.
Y. Hirono, M. Hongo, T. Hirano, *arXiv*:1211.1114.
L. McLerran and V. Skokov, arXiv:1305.0774 \[hep-ph\].
U. Gursoy, D. Kharzeev and K. Rajagopal, Phys. Rev. C [**89**]{} (2014) 054905.
K. Fukushima, D. E. Kharzeev and H. J. Warringa, Phys. Rev. D [**78**]{} (2008) 074033.
J. Kapusta, *Finite-temperature Field Theory*, Cambridge monographs on mathematical physics (Cambridge University Press, 1993).
S. Turbide, R. Rapp and C. Gale, Phys. Rev. C [**69**]{} (2004) 014903.
O. Linnyk, W. Cassing and E. Bratkovskaya, Phys. Rev. C [**89**]{} (2014) 034908.
S. Gupta, *Phys. Lett.* B **597**, 57 (2007).
H.-T Ding, A. Francis, O. Kaczmarek, F. Karsch, E. Laermann et all, PoS 185 (2012).
B. B. Brandt, A. Francis, H. B. Meyer, and H. Witting, PoS 186 (2012).
G. Aarts, C. Allton, J. Foley, S. Hands, and S. Kim, *Phys. Rev. Lett.* **99**, 022002 (2007).
P. V. Buividovich, M. N. Chernodub, D. E. Kharzeev, T. Kalaydzhyan, E. V. Luschevskaya and M. I. Polikarpov, *Phys. Rev. Lett.* **105**, 132001 (2010)
A. Amato, G. Aarts, C. Allton, P. Giudice, S. Hands, and J. Skullerud, *Phys. Rev. Lett.* **111**, 172001 (2013).
D. Fernandez-Fraile and A. Gomez Nicola, Phys. Rev. D [**73**]{} (2006) 045025.
W. Cassing, O. Linnyk, T. Steinert and V. Ozvenchuk, *Phys. Rev. Lett.* **110**, 182301 (2013).
S. Plumari, A. Puglisi, F. Scardina and V. Greco, *Phys. Rev. C.* **86**, 054902 (2012).
C. Wesp, A. El, F. Reining, Z. Xu, I. Bouras and C. Greiner, Phys. Rev. C [**84**]{} (2011) 054911.
J. Fuini, III, N. S. Demir, D. K. Srivastava and S. A. Bass, J. Phys. G [**38**]{} (2011) 015004.
M. Greif, I. Bouras, Z. Xu and C. Greiner, arXiv:1408.7049 \[nucl-th\].
A. Puglisi, S. Plumari and V. Greco, arXiv:1408.7043 \[hep-ph\].
C. Sasaki and K. Redlich, *Phys. Rev. C* **79**, 055207 (2009).
P. Chakraborty and J.I. Kapusta, Phys. Rev. [**C83**]{} (2011) 014906.
S. Plumari, W. M. Alberico, V. Greco and C. Ratti, Phys. Rev. [**D84**]{} (2011) 094004.
P. Levai and U. W. Heinz, Phys. Rev. C [**57**]{} (1998) 1879.
A. Peshier, B. Kampfer and G. Soff, Phys. Rev. D [**66**]{} (2002) 094003.
M. Bluhm, B. Kampfer and K. Redlich, Phys. Rev. C [**84**]{} (2011) 025201.
M. Bluhm, B. Kampfer and G. Soff, Phys. Lett. B [**620**]{} (2005) 131.
R. Marty, E. Bratkovskaya, W. Cassing, J. Aichelin, and H. Berrehrah, Phys. Rev. C **88**, 045204 (2013).
A.S. Khvorostukhin, V.D. Toneev and D.N. Voskresensky, *Phys.Rev. C* **83** 035204 (2011).
P. Arnold, G. D. Moore and L. G. Yaffe, JHEP 0305:051 (2003).
S. Sakai and A. Nakamura, Lattice calculation of QGP viscosities: Present results and next project Proc. Sci. LAT2007(2007)221.
S. C. Huot, P. Kovtun, G. D. Moore, A. Starinets and L. G. Yaffe, *JHEP*12 (2006) 015.
A. Nakamura and T. Saito, Phys. Lett. B [**621**]{} (2005) 171.
H. van Hees, M. Mannarelli, V. Greco and R. Rapp, Phys. Rev. Lett. [**100**]{} (2008) 192301.
These data are not considered being the only set more then one order of magnitude above all the other lQCD data.
H. B. Meyer, *Phys. Rev. D* **76** 101701 (2007).
A. Nakamura and S. Sakai, *Phys. Rev. Lett.* **94** 072305 (2005).
After the submission of our paper, we become aware of S.I. Finazzo, Jorge Noronha, Phys.Rev. D89 (2014) 106008, evaluating $\sigma_{el}$ in and AdS/QCD scheme and showing a similar T dependence and by private communication we known that a quite similar behavior is expected.
J. O. Andersen, L. E. Leganger, M. Strickland and N. Su, Phys. Lett. B [**696**]{} (2011) 468.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Abundances of R Coronae Borealis stars (RCBs) and Extreme Helium stars (EHes) are discussed. Recent estimates of the $s$-process elements in hot extreme helium stars show enhancements of lighter $s$-process elements (Y, Zr) relative to heavier $s$-process elements, a characteristic shared by RCB stars. It also suggests that atleast some EHe stars went through an episode of $s$-process element synthesis in their earlier evolution. A majority of RCB stars show a high $^{12}$C/$^{13}$C ratio in their atmospheres. A recent analysis of the spectrum of minority RCB star VCrA, however, shows a lower value between 4 $-$ 10. The implications of these results are discussed.'
author:
- 'N.Kameswara Rao'
title: 'SURFACE COMPOSITIONS OF R CORONAE BOREALIS STARS & EXTREME HELIUM STARS - SOME CONNECTIONS'
---
Introduction
============
It is a great pleasure to be able to participate in this celebration honouring Prof. David Lambert. I would like to convey to David greetings and best wishes from his numerous friends and admirers in India. It is almost a hundred years since Ludendorff discovered the absence of Balmer lines in the spectrum of the F type irregular variable RCrB, the first recognised hydrogen deficient star. About 35 years later, the first helium star was discovered by Popper. The basic problem about the origins of these stars, namely, a single (or double) star of intermediate mass becoming a high luminosity hydrogen deficient star, still remains a mystery. Presently, there are two principal hypotheses to account for their origins. The prime contender is the white dwarf merger scenario (or double degenerate, DD) in which a helium white dwarf secondary is accreted onto a CO (or He) white dwarf primary resulting in the ignition of a helium shell in the accreted envelope. The energy generated forces the star to expand to cool supergiant dimensions. The life time at this stage depends on the supply of helium in the shell and the mass of the envelope. The subsequent evolution would be like a canonical post-AGB contraction to the white dwarf track (Webbink 1984, Iben & Tutukov 1984, Saio & Jeffery 2000). The second scenario proposed is called the late thermal pulse (or final flash, FF), in which an AGB star in its final stages retains a helium shell of enough mass as to get ignited for a last time after it descends on to the white dwarf track. The final thermal pulse forces the star to become a cool supergiant for a second time. The subsequent evolution would again resemble the canonical post-AGB evolution of single stars (Iben et al. 1983). Presently, there are no decisive observational tests predicted to choose either one (or any other alternative) of the scenarios.
Several groups of hydrogen deficient stars have now been identified ranging from high luminosity (low surface gravity) cooler HdC stars, RCBs, EHes to low luminosity (high gravity) hot PG1159 stars (see Jeffery 1996 for a display of various groups in the log $g$, log $T_{\rm eff}$ plane). It is not clear how these various groups of hydrogen deficient stars are related to each other $-$ e.g., EHes and RCBs. Do they represent sucessive stages in evolution? Even some groups are further subdivided and show diversity in properties e.g., majority and minority RCBs which show \[S/Fe\], \[Si/Fe\] values of about 0.5 and much higher, respectively (Lambert & Rao 1994, Asplund et al. 2000). Do they suggest different paths of evolution? In the last few years, stars like Sakurai’s object and FGSge provided us examples of stellar evolution in real time by turning from a normal star to hydrogen deficient born-again giant (Asplund et al. 1997, Gonzalez et al. 1998, Duerbec et al. 2002). How are these born-again giants related to RCBs are some of the intriguing questions that the theory of intermediate mass star evolution is unable to provide. In keeping with the theme of the conference,‘abundances as records of stellar evolution and nucleo- synthesis’, I would like to explore in this presentation the interrelationships of RCBs and EHes and also address some of the issues raised above by seeking clues from surface chemical compositions.
Properties
==========
R Coronae Borealis (RCB) stars are carbon-rich, hydrogen-poor F-G supergiants that exhibit light fades of several magnitudes at irregular intervals due to circumstellar dust formation. They range in $T_{\rm eff}$ from 8000 to 4000 K. After Keenan & Barnbaum’s (1997) spectroscopic investigation even the cool so-called DYPer stars also seem to belong to this class, extending the $T_{\rm eff}$ to 3500 K. The effective gravities range from log $g$ = 0.5 to 1.5. The RCB population in the LMC provides the estimates of absolute magnitudes which range from $M_{V}$ $-$2.5 to $-$5 and corresponding luminosities log (L/$L_{\odot}$) of 4.0 $-$ 3.2 (Alcock et al. 2001). Presently known members in the Galaxy amount to about 35 including the 3 hot RCBs, about 21 are known in the LMC (including DY Per stars) and a lone one in the SMC (Alcock et al. 2001, Morgan et al. 2003). The distribution of the number of RCBs with respect to spectral type suggests a peak around F-G for Galactic ones whereas the LMC population shows a peak at much cooler temperatures. By scaling the LMC population of RCBs to the Galaxy, Alcock et al. estimate more than 3200 to be present in the galaxy most of them being cooler members. Any evolutionary schemes that are proposed to explain the origins should also be able to account for these numbers. The two evolutionary schemes that have been proposed (DD and FF) can account only partially for the estimated number of RCBs present in the Galaxy (Iben, Tutukov, Yungelson 1996). Extreme helium stars are mostly carbon rich, hydrogen poor A-B supergiants. Some of them show short period pulsations with a period of a few days. They range in $T_{\rm eff}$ from 32000 to 9000 K with log $g$ of 0.7 to 4.0 . The luminosities, log L/$L_{\odot}$ ,are estimated to be about 4.4 to 3.0 (Saio & Jeffery 2000 ). Presently there are about 21 known members in the Galaxy and none in either LMC or SMC.
The Galactic distribution of both groups (RCB and EHe) suggests a bulge population (Jeffery et al. 1987) and might belong to the thick disk, although there are suggestions that a few of them might even be part of the halo (eg. U Aqr - Cottrell & Lawson 1998).
The first major study of the surface abundances of a larger number of RCBs by Lambert & Rao (1994) revealed that majority of the stars analysed (14 out of 18) showed similar patterns; particularly \[Si/Fe\] and \[S/Fe\][^1] are around 0.5 and a mild Fe deficiency relative to solar. The minority RCBs are marked by approximately solar Si and S abundances and a severe Fe deficiency (or a very high \[Si/Fe\], \[S/Fe\]). A similar classification by Fe abundance was also suggested for EHe stars (Heber 1986).
Spectral analyses
=================
Before discussing the surface abundances it is appropriate to recall some of the uncertainities involved in arriving at these estimates. Most of the analyses are based on the line-blanketed, LTE atmospheric models computed at Uppsala by Asplund et al. (1997) for RCB stars of $T_{\rm eff}$ 8000 to 6000 K and at Armagh by Jeffery & Heber (1992), Jeffery, Woolf, Pollacco (2001) for EHes with $T_{\rm eff}$ $>$ 9000 K. The continuous opacity in the atmospheres of RCBs is controlled by the photoionization of C[i]{} from excited levels and the gas pressure is provided by helium. In estimating the mass fraction of elements, a crucial parameter needed for RCBs is the C/He ratio (the number density of carbon to helium), which can not be estimated from the spectrum directly (Rao & Lambert 1996, Asplund et al. 2000). This ratio in EHes, which can be directly estimated from the analysis of spectral lines (Jeffrey 1996, Pandey et al. 2001) has a mean value of 1%. It has been assumed in the analyses of RCBs that the same C/He of 1% holds. This value of C/He seems to be consistent with the metallicity expected from galactic distribution (Rao & Lambert 1996, Pandey et al. 2001). One of the surprising outcomes of the atmospheric analysis of RCBs is the so called ‘carbon problem’, as discussed in detail by Asplund et al. (2000)(first noticed by David Lambert). The carbon abundance estimated from the observed C[i]{} lines is four times less than the input carbon abundance for C/He, with less input the C/He ratio is very very small. Recently, based on the analysis of \[C[i]{}\] lines 9850Å and 8727Å, Pandey et al. (2004b) suggest that a chromosphere like temperature rise in the atmosphere might be able to account for the carbon problem. However it was realised by Asplund et al. (2000) that the abundance ratios are largely unaffected by the carbon problem (and the assumed C/He values). An $M_{bol}$ of $-$5 has been assumed for all RCBs in arriving at the appropriate $T_{\rm eff}$ and log $g$, although it is known that LMC RCBs show a range of values. As discussed by Pandey et al. (2001), the element which represents the initial metallicity of the stars is not clear, whether it is Fe or Si, S. A metallicity parameter M has been defined by Pandey et al. based on the Si, S abundances that represents initial Fe abundance. The observed Fe abundance might be affected by things other than the intial metallicity of the star. This property might even apply to hot RCB stars like DY Cen as well. The extreme Fe deficiency is suggested to be a result of the accretion of winnowed gas from dust (Jeffery & Heber 1993).
Abundance Patterns
==================
The surface abundances of RCBs and EHes have been estimated by Asplund et al. (2000), Rao & Lambert (2003, 2004), Jeffery (1996), Pandey et al. (2001) and Pandey et al. (2004a). Both RCBs and EHes show the majority, minority division proposed by Lambert & Rao (1994). Mainly they are differentiated by Fe abundance. The majority cluster around \[Fe\] of $-$1 and the minority show a larger deficiency about $-$1.7 or more. Since a C/He of 1% is assumed for all stars (except the minority RCB star V854Cen for which a value of 10% is suggested -Asplund et al. 1998), the absolute numbers can be compared. 15 stars out of 19 analysed RCBs comprise the majority and 12 out of 14 EHes analysed constitute the majority class. The 4 minority RCBs are VCrA, VZSgr, V3795Sgr and V854Cen and the 2 minority EHes are BD +10 2179 and FQAqr. The mean abundances (normalised to $\log\Sigma\mu_X\epsilon$(X) = 12.15 with $\mu$ as the atomic weight) of each group are shown in Table 1. The dispersion around
[llllll]{} & &\
Element & Z & Maj. RCBs(15) & Min. RCBs(4) & Maj. EHes(12) & Min. EHes(2)\
H & 1 & 6.14$\pm$0.89 & 7.57 & 7.19$\pm$0.95 & 7.35\
He & 2 & 11.54 & 11.54 & 11.54 & 11.54\
C & 6 & 8.91$\pm$0.14 & 9.0$\pm$0.35 & 9.32$\pm$0.22 & 9.05$\pm$0.05\
N & 7 & 8.67$\pm$0.23 & 7.88$\pm$0.20 & 8.36$\pm$0.34 & 7.63$\pm$0.48\
O & 8 & 8.17$\pm$0.41 & 8.22$\pm$0.59 & 8.60$\pm$0.48 & 8.5$\pm$0.4\
Ne & 10 & 8.3(1) & 7.9(1) & 9.11(6)$\pm$0.25 & 7.9(1)\
Na & 11 & 6.13$\pm$0.22 & 5.94 & 6.5(3)$\pm$0.8 &\
Mg & 12 & 6.72(5)$\pm$0.21 & 6.3(3)$\pm$0.22 & 7.40$\pm$0.35 & 6.5$\pm$0.5\
Al & 13 & 5.95$\pm$0.29 & 5.51$\pm$0.14 & 6.07$\pm$0.52 & 5.5$\pm$0.8\
Si & 14 & 7.12$\pm$0.19 & 7.34 & 7.23$\pm$0.60 & 6.5$\pm$0.2\
P & 15 & 5.9(1) & 6.5(1) & 5.62$\pm$0.47 & 4.85$\pm$0.65\
S & 16 & 6.87$\pm$0.33 & 6.93$\pm$0.40 & 7.08$\pm$0.32 & 6.55$\pm$0.55\
Ca & 20 & 5.36$\pm$0.19 & 5.16$\pm$0.07 & 5.84(5)$\pm$0.27 & 4.2(1)\
Sc & 21 & 2.87(6)$\pm$0.19 & 2.89 & 3.3(1) & 2.1(1)\
Ti & 22 & 4.03(8)$\pm$0.14 & 3.6(3)$\pm$0.34 & 4.53(3)$\pm$0.17 & 3.25$\pm$0.05\
Fe & 26 & 6.49$\pm$0.24 & 5.73 & 6.89$\pm$0.32 & 5.7$\pm$0.3\
Ni & 28 & 5.82$\pm$0.24 & 5.44$\pm$0.43 & 5.93(3)$\pm$0.47 & 5.0(1)\
Zn & 30 & 4.34(13)$\pm$0.28 & 4.08$\pm$0.21 & 4.4(2)$\pm$0.2 & 4.14(1)\
Y & 39 & 2.08$\pm$0.50 & 1.94$\pm$0.60 & 2.27$\pm$0.59 & 1.75(1)\
Zr & 40 & 2.09(8)$\pm$0.29 & 2.04$\pm$0.48 & 2.7(4)$\pm$0.60 & 1.83$\pm$0.53\
Ba & 56 & 1.43$\pm$0.56 & 0.97$\pm$0.43 & 1.7 & 0.5\
the mean in each group is surprisingly small particularly for the majority groups of RCBs and EHes $\sim$ 0.27 dex. Only H and the $s$-process elements show a little more dispersion. The similarity in the abundance pattern of the majority RCBs and EHe (Fe of 6.5 and 6.8, respectively) is striking. The mean difference for 15 elements is 0.23 dex. However H, N, Ne and Mg show significant differences. The minority groups RCBs and EHes (Fe of 5.7 for both) also show small differences ($<$ 0.13 dex) for most elements except Si, Ca (may be P and Sc). The fact that the H abundance for the majority of RCBs is lower by 1.0 dex compared to EHes and the N abundance is also higher by 0.3 dex might suggest that RCBs are a later phase in evolution to EHes, However, the larger abundance of Ne and Mg in EHes furthur indicates that $^{14}$N is converted to $^{22}$Ne and $^{25}$Mg by alpha processing, thus EHes might be a later phase to RCBs as is expected from the tracks of post AGB stars in the log $g$, log $T_{\rm eff}$ plane.
CNO abundances
--------------
Figure 1 shows \[N/Fe\] versus Fe and \[N/M\]versus M (the metalicity parameter) for both groups of RCBs and EHes. Clearly, most of the RCBs have N abundances that are predicted from conversion of initial C and O to N or even more. In some cases C, produced in the He burning, might also have been converted to N. The newly discovered RCB star V2552Oph may be one such and illustrates the N enhancement prominently (Rao & Lambert 2003).
Although RCrB and V2552Oph have very similar line spectra and physical parameters, V2552 Oph shows much stronger N[i]{} lines than R CrB. On the other hand, the N abundance in EHes generally lies between the expected N from conversion of initial C to N and the value of N expected from conversion of initial C, O both to N . The minority stars (both groups and DYCen) have N abundances less than that expected from full convertion of C to N. Although the N abundance in many RCBs and in some EHes imply wholesale conversion of O to N via ON cycles, many stars are not O deficient suggesting O is synthesized along with C, i.e., 3$\alpha$ -process was followed by $^{12}$C$(\alpha,\gamma)^{16}$O. Most of the O-rich stars have an observed O/C about 1 implying equal production of C and O.
$s$-Process elements
--------------------
Dramatic enhancement of light $s$-process elements Sr, Y, and Zr was first seen in cool RCB star UAqr (Bond et al. 1979). Asplund et al. (2000) have shown that most of the RCB stars show enhancements \[Y/Fe\] of about 0.8 and \[Ba/Fe\] of about 0.4, i.e., the lighter $s$-process elements are more enhanced than the heavier ones. However, there is a considerable dispersion in the Y and Ba abundances in RCBs.
Both Y and Ba abundances seem to be anticorrelated with H abundance. UAqr shows extraordinary overabundances of $s$-process elements \[Y/Fe\] $\sim$ 3.3 and \[Ba/Fe\] $\sim$ 2.1 (Vanture et al. 1999). Generally these enhancements relative to Fe are consistent with a mild single neutron exposure $\tau_{o}$ $\sim$ 0.1 mb$^{-1}$ These estimates of $s$-process abundances are not available for EHes stars (Asplund et al. 2000). Did EHe stars pass through a phase of $s$-process element production (similar to RCBs ) ? Did they undergo third dredgeup and show $s$-process abundance pattern similar to AGB or post -AGB stars. These are some of the questions that need to be addressed.
In the atmospheres of the hot EHes most of the $s$-process elements exist in doubly ionized state and lack spectral lines in the optical region. \[Let me add a personal note here. In 1996 January, David Lambert and I were attending a conference on spectroscopy in Bombay where we heard a talk by Indrek Martinson discussing the Zr[iii]{} and Y[iii]{} spectra in UV and the availability of fairly decent $gf$-values. This prompted us to apply for $HST$ - $STIS$ spectra in search of Zr and Y abundances in EHe stars.\] Fortunately, strong lines of Y[iii]{}, Zr[iii]{}, Ce[iii]{}, La[iii]{} etc., do occur in the UV where EHe stars have appreciable flux. We could obtain UV spectra of 7 EHe stars with $STIS$ on $HST$. Analyses of the spectra of two EHe stars V1920Cyg and HD124448 demonstrate the similarities in the pattern of $s$-process elements with RCBs. The two stars have the same log $T_{\rm eff}$ and log $g$ but show large differences in \[Y/Fe\] and \[Zr/Fe\] (similar to RCBs) (Figure 3). V1920Cyg has more enhanced abundances of Y, Zr and the range in abundance variations is also very similar to RCBs (Pandey et al. 2004).
Although the abundances of heavy $s$-process elements could not be estimated, the upper limits of the abundances of Ce, Nd do demonstrate the lighter $s$-process elements are more enhanced even in EHe stars. It is generally acknowledged that $^{13}$C$(\alpha,n)^{16}$O is the main source of neutrons to run the $s$-processing in the He-burning shells of intermediate-mass AGB stars. Sufficient amounts of $^{13}$C are to be generated by slow mixing of protons into the $^{12}$C rich intershell regions to generate neutrons. The neutron irradiation occurs in radiative conditions. The heavier the neutron flux the greater is the abundance of heavies relative to light $s$-process elements. Busso et al. (2001) used the distribution of the ratio of heavy $s$-process celements (hs) to the light $s$-process (ls) elements with respect to metallicity to characterize various parameters of neutron exposures during the third dredgeup phase in AGB stars eg. mass of $^{13}$C pocket in the inter shell regions. Reddy et al. (2002) showed that the variation of the \[hs/ls\] with respect to metallicity in post -AGB stars (that went through third dredgeup) is characterized by a model ST/1.5 of Busso et al. (2001).
A plot of \[Y/M\] and \[Zr/M\] versus \[M\] for RCBs and EHes (figure 4 ) shows that the enhancements are positive and both show a similar range in their abundances. We compared the run of the ratio of \[ls/hs\] for RCBs and EHes with respect to the metallicity parameter \[M\] . The estimates for EHes are based on the upper limits for the heavy $s$-process elements and includes data from our ongoing analysis of the $HST$ UV spectra. Estimates of UAqr and the born again giant, Sakurai’s object (during May - Oct 1996) are also included for comparison. Both the groups RCBs and EHes blend together emphasizing the similarity in their ls/hs ratios.
Figure 5 shows a comparison of the trend of \[ls/hs\] in RCB and EHe stars along with that shown by post- AGB stars (Reddy et al. 2002) and the Busso et al’s (2001) model ST/1.5. It is obvious that the trend of \[ls/hs\] with respect to metallicity of RCB and EHes is quite different from that shown by post-AGB stars. It also, probably, suggests the $s$-processing in RCBs and EHes is not a result of the third dredgeup and could have happened when the stars passed through a second AGB phase (presumably).
Minority RCBs and Sakurai’s object (V4334Sgr)
=============================================
Are the minority RCBs born-again giants? The similarity of the abundance pattern of V854Cen with Sakurai’s object, a clear example of born-again giant (final flash object), has been pointed out by Asplund et al. (1998). It is expected that the final flash objects show a C/He ratio much greater than 1%, say about 10 to 30%. The C/He value of 10% has been estimated for Sakurai’s object in 1996 May to October (Asplund et al. 1998). Asplund et al. infer a C/He of 10% for V854Cen. Asplund et al. (1998, 2000) also state that there are indications to suggest that the minority objects V3795Sgr and VZSgr have higher C/He values greater than 1% -may even be 10%.
The abundance patterns of the minority stars suggested more diversity and range, in particular light elements like H, N, O and some $s$-processed elements. Particularly VCrA showed differences relative to other members. However, revised estimates from a recent analysis of higher resolution spectra (Rao & Lambert 2004) suggests similar abundances for N, O, etc., as for the rest of the members. The abundances are more uniform in the minority group as well. The abundance pattern (i.e., log X/Fe) of V CrA is more similar to that of Sakurai’s object during May - October 1996 (Asplund et al. 1997, 1999) for most elements except Mg, Si, S, and Ca. The $s$-process elements in V CrA are also not enhanced as much as in Sakurai’s object.
One of the expectations of the final flash scenario is the presence of $^{13}$C, and a low value of $^{12}$C/$^{13}$C ratio. It is expected that surface protons are mixed into the intershell region trigering CN cycle and converting $^{12}$C to $^{14}$N and $^{13}$C depending on the proton supply (Renzini 1990). In fact, Sakurai’s object showed a $^{12}$C/$^{13}$C value of 4 although the other final flash objects like FGSge (Gonzalez et al. 1998) and V605Aql (which showed a hydrogen deficient carbon star spectrum at maximum light -Lundmark 1919, Clayton & De Marco 1997) did not show the presence of $^{12}$C$^{13}$C bands. Most RCBs have been shown to have a high value of the $^{12}$C/$^{13}$C ratio. Keenan & Barnbaum (1997) detected the $^{12}$C$^{13}$C band at 4744Å in the cool peculiar RCB variable DYPer. It was a surprise to find $^{12}$C$^{13}$C bands in the minority star VCrA (Rao & Lambert 2004). We have synthesized the $^{12}$C$^{12}$C and $^{12}$C$^{13}$C bands of the Swan system (1,0) in the three stars; Sakurai’s object in Oct 1996, V854Cen and VCrA (Figure 6) to match the observations. We used the line list and physical parameters obtained by Asplund et al. (1997) for Sakurai’s object. The estimated ratio of $^{12}$C/$^{13}$C ranged between 3 $-$ 5 for Sakurai’s object (Asplund et al. 1997), has an upper limit of 30 for V854Cen and a value 4 $-$ 10 for VCrA , thus displaying a similarity with Sakurai’s object. The presence of $^{13}$C in the atmosphere of a RCB star does support the suggestion $^{13}$C$(\alpha,n)^{16}$O is the neutron source. It is possible that the evolutionary path for all minority RCBs (may even EHes) is through final flash.
Discovery of a post -AGB hydrogen deficient star in globular cluster M5 (Dixon et al 2004) is an exciting new development which could pin down an age (and possible mass) to the progenitor.
In summary, it now appears that at least some EHe stars show enhanced abundances of light $s$-process elements, e.g., Y, Zr as well as a ls/hs ratio similar to RCBs. The variation of the ls/hs ratio with decreasing metallicity suggests that $s$-processing in RCBs and EHes is not similar to that experienced by post -AGB (and AGB) stars (i.e., ST/1.5 model of Busso et al. 2001). The abundance ratios suggest a single exposure of $\tau_{o}$ of 0.1 to 0.2 mb$^{-1}$. It is likely that this episode of $s$-process element production might have occured when the stars were passing through AGB phase for a second time. The similarity in the abundance patterns of majority RCBs and majority EHes including the $s$-process elements and the presence of enhanced abundances of Ne and Mg in EHes does suggest that EHe phase might be later in evolution to that of RCBs. The minority RCBs seems to be a more coherent group in abundance distribution than earlier estimates indicated. Minority EHes and RCBs show a very similar abundance pattern, except for Si, Sc, and Ca, (elements that could be tied up in dust). RCBs have IR excesses and dust production episodes. The discovery of low $^{12}$C/$^{13}$C ratio (4 $-$ 10) in the minority RCB, VCrA does provide long awaited evidence for the mixing of surface protons to the intershell region and subsequent production of neutrons by $^{13}$C$(\alpha,n)^{16}$O, similar to Sakurai’s object. The similarity of abundance patterns of VCrA and V854Cen to that displayed by Sakurai’s object in 1996 Oct might encourage the suggestion that all minority RCBs are formed through final flash.
Acknowledgements
================
I would like to thank my collaborators David Lambert, Gajendra Pandey, Simon Jeffery for letting me use some results before publication. I would also like to thank Martin Asplund for supplying me atmospheric models and line lists for $^{12}$C$^{13}$C bands. I would like to express my thanks to David Yong, Eswar Reddy and Gajendra Pandey for preparing the figures and other help. I would also like express my appreciation to the organisers of the conference for their generous hospitality in Austin.
References
==========
> `Alcock, C., Allsman, R. A., Alves, D. R., et al. 2001, ApJ,`\
> `554, 298`
>
> `Asplund, M., Gustafsson, B., Kiselman, D., Eriksson, K.`\
> `1997a, A&A, 318, 521`
>
> `Asplund, M., Gustafsson, B., Lambert, D. L., Rao, N. K.`\
> `1997b, A&A, 321, L17`
>
> `Asplund, M., Gustafsson, B., Rao, N. K., Lambert, D. L.`\
> `1998, A&A, 332, 651`
>
> `Asplund, M., Lambert, D. L., Kipper, T., Pollacco, D., `\
> `Shetron, M. D. 1999, A&A, 343, 507`
>
> `Asplund, M., Gustafsson, B., Lambert, D. L., Rao, N. K.,`\
> `2000, A&A, 353, 287`
>
> `Bond, H. E., Luck, R. E., Newman, M. J. 1979, ApJ, 233,`\
> `205`
>
> `Busso, M., Gallino, R., Lambert, D. L., Travaglio, C.,`\
> `Smith, V. V. 2001, ApJ, 557, 802`
>
> `Clayton, G. C., de Marco, O. 1997, AJ, 114, 2679`
>
> `Cottrell, P. L., Lawson, W. A. 1998, PASA, 15, 179`
>
> `Dixon, W. V., Brown, T. M., Landsman, W. B. 2004, ApJ,`\
> `600, L43`
>
> `Duerbeck, H. W., Liller, W., Sterken, C., et al. 2000,`\
> `AJ, 119, 2360`
>
> `Gonzalez, G., Lambert, D. L., Wallerstein, G.,`\
> `Rao, N. K., Smith, V. V., McCarthy, J. K. 1998, ApJS,`\
> `114, 132`
>
> `Heber, U. 1986, in Hydrogen Deficient Stars,`\
> `IAU Coll. 87, ed. K. Hunger, D. Schonberner & `\
> `N. K. Rao, Reidel, Dordrecht, 32`
>
> `Iben, I. Jr., Tutukov, A. V. 1984, ApJS, 55, 335`
>
> `Iben, I, Jr., Tutukov, A. V., Yungelson, L. R. 1996,`\
> `ApJ, 456, 750`
>
> `Iben, I. Jr., Kaler, J. B., Truran, J. W., Renzini, A.`\
> `1983, ApJ, 264, 605`
>
> `Jeffery, C. S. 1996, in ASP Conf. Ser. Vol. 96,`\
> `Hydrogen Deficient Stars, ed. C. S. Jeffery & U. Heber, 152`
>
> `Jeffery, C. S., Heber, U. 1992, A&A, 260, 133`
>
> `Jeffery, C. S., Heber, U. 1993, A&A, 270, 167`
>
> `Jeffery, C. S., Drilling, J. S., Heber, U. 1987, MNRAS,`\
> `226, 317`
>
> `Jeffery, C. S., Woolf, V. M., Pollacco, D. L. 2001, A&A,`\
> `376, 497`
>
> `Keenan, P. C., Barnbaum, C. 1997, PASP, 109, 969`
>
> `Lambert, D. L., Rao, N. K. 1994, J. Astrophys. Astron.,` `15, 47`
>
> `Lundmark, K. 1921, PASP, 33, 814`
>
> `Morgan, D. H., Hatzidimitriou, D., Cannon, R. D., Croke,`\
> `B. F. 2003, MNRAS, 344, 325`
>
> `Pandey, G., Rao, N. K., Lambert, D. L., Jeffery, C. S.,`\
> `Asplund, M. 2001, MNRAS, 324, 937`
>
> `Pandey, G., Lambert, D. L., Rao, N. K., Jeffery, C. S.`\
> `2004a, ApJ, 602, L113`
>
> `Pandey, G., Lambert, D. L., Rao, N. K., Gustafsson, B.,`\
> `Ryde, N., Yong, D. 2004b, MNRAS, 353, 143`
>
> `Rao, N. K., Lambert, D. L. 1996, in ASP Conf. Ser. Vol. 96,`\
> `Hydrogen Deficient Stars, ed. C. S. Jeffery & U. Heber, 43`
>
> `Rao, N. K., Lambert, D. L. 2003, PASP, 115, 1304`
>
> `Rao, N. K., Lambert, D. L. 2004, (in preparation)`
>
> `Renzini,A. 1990, in ASP Conf. Ser, Vol. 11, Confrontation`\
> `between Stellar Pulsation and Evolution, ed. C. Cacciari,`\
> `G. Clementini, 549`
>
> `Reddy, B. E., Lambert, D. L., Gonzalez, G., Yong, D. 2002,`\
> `ApJ, 564, 482`
>
> `Saio, H., Jeffery, C. S. 2000, MNRAS, 313, 671`
>
> `Saio, H., Jeffery, C.S. 2002, MNRAS, 333, 121`
>
> `Vanture, A. D., Zucker, D., Wallerstein, G. 1999, ApJ, 514,`\
> `932`
>
> `Webbink, R. F. 1984, ApJ, 277, 355`
=
J. C. Wheeler : Can you comment on mass measurements for the RCB and EHe stars?
K.Rao: Simon Jeffery and Vincent Woolf have tried to obtain estimates for two pulsating EHe stars V652 Her, LSS 3184. They obtain $\sim$ 0.5 $M_{\odot}$ for the mass. LSS 3184 is carbon rich. If a CO white dwarf has to get into action, this mass estimate is little low.
J. Lattanzio : Can you determine the Mg isotopes for your Mg-rich stars?
K.Rao: No. We have to find some molecule containing Mg to estimate isotope ratios. Since H is weak, MgH is not present, even in cool stars. It is a difficult problem.
J. Cohen : You and some previous speakers have spoken about abundance peculiarities. We need to make sure that the abundance peculiarities are real and not the result of the physical phenomena of gravitational settling and radiative levitation such as seen on the blue HB of globular clusters.
K.Rao: These stars are supergiants with low effective gravities, with rotation, and appreciable microturbulance. As such it is unlikely to expect gravitational settling to operate. Moreover, the abundance pattern is not similar to that seen in blue horizontal branch stars.
[^1]: the \[$i$\] notation refers to log $\epsilon$(i)$^{\star}$ $-$ log $\epsilon$(i)$^{\odot}$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Holomorphic Embedding Load Flow Method (HELM) has been suggested as an alternative approach to solve load flow problems. However, the current literature does not provide any HELM models that can accurately handle general power networks containing PV and PQ buses of realistic sizes. The original HELM paper dealt only with PQ buses, while a second paper showed how to include PV buses but suffered from serious accuracy problems. This paper fills this gap by providing several models capable of solving general networks, with computational results for the standard IEEE test cases provided for comparison. In addition this paper also presents a new derivation of the theory behind the method and investigates some of the claims made in the original HELM paper.'
author:
- 'I. Wallace, D. Roberts, A. Grothey, and K. I. M. McKinnon — [^1]'
bibliography:
- 'helm\_pv\_paper.bib'
title: Alternative PV Bus Modelling with the Holomorphic Embedding Load Flow Method
---
Load flow analysis; Power system modelling; Power system simulation; Power engineering computing; Energy management; Decision support systems.
Introduction
============
flow problems consist of solving a set of equations — the bus-power-equilibrium equations (BPEE) — based on the physical laws surrounding a power network so as to obtain a feasible load flow for the network. The physical laws result in the BPEE being nonlinear, and usually having multiple solutions. In addition, many of these solutions would typically correspond to [*unstable*]{} operating points of a real network, so termed [*low voltage solutions*]{}, and so solving the BPEE for the physical solution is not an easy calculation. Most current methods for solving the BPEE of a network use an iterative approach. These methods all have the same issue — that for specific problems they may not converge to a solution or converge to the “wrong” ([*i.e.,*]{} unstable or low voltage) solution. A 2012 paper [@helm] and a subsequent patent [@helmpatent] by Trias proposed an alternative to these iterative methods, HELM, which claims to address this issue. The idea was to treat the voltages as holomorphic functions of a complex parameter $z$ that scales the demands, and to use the simple-to-calculate solution when $z=0$ to determine the desired solution when $z=1$ making use of the powerful theory of holomorphic functions and analytic continuations. There remained some important gaps in the method as described by Trias however, as it did not provide any detail on how to handle PV buses, and so the method was applicable only on networks with exclusively PQ buses (plus one slack bus). As well, Trias’ papers contain claims about the theoretical properties of the method, namely regarding the method’s ability to always find the “correct” solution, for which only sketchy theoretical justification was given. Subramanian [*et al*]{} [@subramanian] suggest an extension of the HELM approach that can be applied to networks that include PV buses, however this approach suffers from numerical problems.
The contribution of this paper is to provide alternative approaches to model PQ buses within the HELM framework that do not suffer from the above numerical problems. Further, we give an investigation of the theoretical underpinnings of the HELM method.
The organisation of this paper is as follows. In the following section, we recap the HELM method for PQ bus networks as proposed in [@helm; @helmpatent]. In Section \[sec:theory\], we give a derivation of the HELM approach that uses the holomorphic implicit function theorem. This enables us to gain further insight into the theoretical properties of the method. In Section \[sec:subramanian\], we show how the problem changes when considering PV buses, present the current literature and its shortcomings, and then in Section \[sec:PVmodels\] provide our own models which are capable of handling general networks. Section \[sec:results\] presents computational results for the test networks.
Nomenclature {#nomenclature .unnumbered}
============
Sets {#sets .unnumbered}
----
${\mathcal}{B}$
: All Buses.
${\mathcal}{B}_{PQ}$
: PQ Buses.
${\mathcal}{B}_{PV}$
: PV Buses.
Parameters and Variables {#parameters-and-variables .unnumbered}
------------------------
$Y_{ik}$
: ($i,k$) element of bus admittance matrix.
$V_i$
: Complex voltage at bus $i$.
$S_i$
: Complex power injection at bus $i$ ($S_i=P_i+jQ_i$).
$P_i$
: Real power injection at bus $i$.
$Q_i$
: Reactive power injection at bus $i$.
$M_i$
: Prescribed voltage magnitude at PV bus $i$.
$z$
: Complex variable used for holomorphic embedding.
Notation {#notation .unnumbered}
--------
$\delta_{i,j}$
: $1$ if $i\!=\!j$, $0$ otherwise.
HELM PQ Model {#sec:helm}
=============
The HELM method was first introduced by Trias in [@helm]. He begins by considering the BPEE for a PQ-bus network in the general form: $$\begin{aligned}
\sum_{k\in {\mathcal}{B}}Y_{ik}V_k=\frac{S_i^*}{V_i^*},\quad i \in {\mathcal}{B}_{PQ}\label{eq:bpee}
\end{aligned}$$ where without loss of generality we will set bus 0 to be the slack bus, so that $\mathcal{B} = \mathcal{B}_{PQ}\cup\{0\}$. The voltage at the slack bus is known to be $V_0=1$. The remaining $V_i$ are the unknown complex variables. Trias proceeds by setting up a homotopy where the demands are scaled by a [*complex*]{} parameter $z$ and the resulting bus voltages are treated as functions of this complex parameter $$\begin{aligned}
\sum_{k\in {\mathcal}{B}}Y_{ik}V_k(z)=\frac{zS_i^*}{V_i^*(z^*)},\quad i \in {\mathcal}{B}_{PQ}\label{eq:bpeez}
\end{aligned}$$ here $z\!=\!1$ corresponds to the solution of the BPEE while at $z\!=\!0$ a solution can be easily computed. Note that the solution for will in general not be unique for a given $z$, however for $z=0$ a unique solution exists under the condition that $V_i(0)\ne 0\,\forall i$.
The main claim by Trias is that the voltages $V_i(z)$ implicitly defined by are holomorphic functions at $z=0$ and can be analytically continued to obtained the “correct” solution to the BPEE at $z=1$.[^2] This is not obvious, due to the use of the complex conjugate in system which is [*not*]{} a holomorphic function.
Trias circumvents this difficulty by embedding in a larger holomorphic system, namely
$$\begin{aligned}
\sum_{k\in {\mathcal}{B}}Y_{ik}V_k(z)=\frac{zS_i^*}{\overline{V_i}(z)},\quad i \in {\mathcal}{B}_{PQ}\\
\sum_{k\in {\mathcal}{B}}Y^*_{ik}\overline{V_k}(z)=\frac{zS_i}{V_i(z)},\quad i \in {\mathcal}{B}_{PQ}
\end{aligned}\label{eq:helm-first}$$
where $\overline{V_i}(z)$ are additional complex variables formally independent of $V_i(z)$. It is easy to check that these equations are indeed holomorphic as functions of the independent complex variables $z, V_i, \overline{V_i}$ for example by checking the Wirtinger derivatives or the Cauchy-Riemann equations. System is a set of polynomial equations (after multiplying through with the denominator in each case) and Trias uses the theory of resultants and Gröbner bases to deduce that all $V_i$ and $\overline{V_i}$ are holomorphic functions everywhere except for a finite set of singularities - all of them branch points - which will not include 0.
If the additional constraint $$\overline{V_i}(z) = (V_i(z^*))^*, \quad i\in {\mathcal}{B},\label{eq:helm-refcond}$$ which Trias calls the [*reflecting condition*]{}, holds, system reduces to . Trias makes use of the system only to establish that there exist holomorphic solution functions $V_i(z), \overline{V_i}(z)$ and then argues that since we are only interested in those solutions that satisfy the reflecting condition it can be used to eliminate the $\overline{V_i}$. In the remainder of his presentation Trias uses exclusively.
In Section \[sec:theory\] we will show that the reflecting condition is satisfied automatically under the condition that $V_i(0)\ne 0\,\forall i$.
Given that the voltages are holomorphic functions in a neighbourhood of $z\!=\!0$, they, and their reciprocals, can be written as power series expandable about $z\!=\!0$: $$\begin{aligned}
V_i(z)&=\sum_{n=0}^{\infty} V_i[n]z^n, \quad i\in {\mathcal}{B}\\
\frac{1}{V_i(z)}=W_i(z)&=\sum_{n=0}^{\infty} W_i[n]z^n, \quad i\in {\mathcal}{B}. \label{eq:helm-Wdef}\end{aligned}$$ Substituting into one obtains $$\begin{aligned}
\sum_{k\in{\mathcal}{B}}Y_{ik}\sum_{n=0}^{\infty} V_k[n]z^n= zS_i^*\sum_{n=0}^{\infty} W_i^*[n]z^n\label{eq:helm-powseries}
\end{aligned}$$ From , it is now possible to determine the coefficients of the power series up to any desired level. The process is begun by solving for $z\!=\!0$, which (under the condition that $V_i(0)\ne 0$) yields the set of linear equations $$\begin{aligned}
\sum_{k\in {\mathcal}{B}}Y_{ik}V_k[0]=0, \quad i \in {\mathcal}{B}_{PQ}.\label{eq:helm-v0}\end{aligned}$$ Note that the sum on the left includes the slack bus $k=0$ for which $V_0(z)$ is set to $1$ for all $z$. At this point we need to impose
The reduced bus admittance matrix $Y'$ obtained from $Y$ by removing the row and column corresponding to the slack bus is non-singular.
This is a standard assumption and will hold for any sensible power system. In particular, in the absence of shunts and phase shifters the assumption is equivalent to requiring the system to be connected.
Under Assumption 1 system (\[eq:helm-v0\]) has a unique solution. $W_i[0]$ can then be computed using $$W_i[0]=\frac{1}{V_i[0]}$$ Having obtained the initial values for $V$ and $W$, an iterative process can be used to determine the remaining values in the power series up to any desired order of $n$ by equating the coefficients of $z^n$ in , which yields $$\begin{aligned}
\sum_{k\in {\mathcal}{B}}Y_{ik}V_k[n]= S_i^*W^*_i[n\!\!-\!\!1], \quad i \in {\mathcal}{B}_{PQ}\quad n\geq 1\label{eq:helm-vcalc}\end{aligned}$$ where $W_i[n\!-\!1]$ is calculated using the coefficients of lower orders $$\begin{aligned}
W_i[n-1]=-\frac{\sum\limits_{m=0}^{n-2}V_i[n\!-\!m\!-\!1]W_i[m]}{V_i[0]}\label{eq:helm-wcalc}\end{aligned}$$ In and the coefficient matrix of the system of equations is constant, and so factorisation of this matrix needs only to be done once and can be used for all iterations.
Having obtained the power series for the voltages up to some desired level, it is now possible to compute the voltages for each bus. However, a direct summation of the power series for $z\!=\!1$ is insufficient, as the radius of convergence of the power series is typically less than 1. Instead, analytic continuation using Padé approximants [@pade] is used. Padé approximants are a particular type of rational approximations to power series known to have good convergence properties. The $L,M$ Padé approximant is denoted by $[L/M]=P_L(x)/Q_M(x)$, where $P_L(x)$ and $Q_M(x)$ are polynomials of degree less than or equal to $L$ and $M$ respectively.
In [@helm], Trias explains how Stahl’s extremal domain theorem and Stahl’s Padé convergence theorem provide proof that Padé approximants give the maximal analytical continuation. That is, if there is a steady-state solution to the problem, then the diagonal Padé approximants will converge to this answer, while if there is no steady-state solution (voltage collapse), then the Padé approximants will not converge. In fact Stahl’s Theorems [@pade Ch 6] asserts that the diagonal Padé approximants ([*i.e.,*]{} $L=M$) converge to the analytic continuation with the extremal domain of the approximated function. Here extremal domain is understood as the one having a minimal exemption set (in the shape of branch cuts) measured in capacity. The implication is that diagonal Padé approximants (of high enough order) will yield values $V_i(1)$ provided that $z=1$ is not on a branch cut. Such a branch cut is indicated by a line of poles of the approximant.
A slight variant to is given in [@helmpatent], where an additional term is added to the holomorphic embedding $$\begin{aligned}
\sum_{k\in {\mathcal}{B}}Y_{ik}V_k(z)-(1\!\!-\!\!z)y_i=\frac{zS_i^*}{\overline{V_i}(z)},\quad i \in {\mathcal}{B}_{PQ}\\
\sum_{k\in {\mathcal}{B}}Y^*_{ik}\overline{V_k}(z)-(1\!\!-\!\!z)y_i^*=\frac{zS_i}{V_i(z)},\quad i \in {\mathcal}{B}_{PQ}
\end{aligned}\label{eq:helm-alternative}$$ where $$\begin{aligned}
y_i=\sum_{k\in{\mathcal}{B}}Y_{ik}\end{aligned}$$
In this alternative model, at $z\!=\!0$, $V_k(0)=1$ is an obvious solution, as this causes the two terms in the left-hand side of the equations to cancel. This also means that $V_k[0]=1$, eliminating the need for step . At $z\!=\!1$ the $(1\!-\!z)$ term disappears, leaving the equations the equivalent of . This alternative model will be the basis of some of the models which incorporate $PV$ buses shown later on.
Theory {#sec:theory}
======
In this section we will present additional theory for the HELM model. First, we will provide a separate proof that the $V_i$ and $\overline{V_i}$ in are holomorphic using the Complex Implicit Function Theorem (CIFT) [@ift], which we will then extend in Section \[sec:PVmodels\] to include models with PV buses. Next we will show that the reflecting condition is implied by the formulation and thus need not be assumed.
$V$ and $\overline{V}$ Holomorphic {#sec:HELM-holo}
----------------------------------
We begin by defining the functions $$\begin{aligned}
f_i(z,V,\overline{V}):=&\overline{V_i}{\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k - zS_i^*,\quad &i\in {\mathcal}{B}_{PQ}\\
f_{N+i}(z,V,\overline{V}):=&V_i{\sum_{k\in{\mathcal}{B}}}Y^*_{ik}\overline{V_k} - zS_i, \quad &i\in{\mathcal}{B}_{PQ}
\end{aligned}\label{eq:ift}$$ where $N$ is the number of non-slack buses in the network.
The CIFT states that if there exists a seed solution $v, \overline{v}$ with $f(0,v,\overline{v})=0$ and $J$ is non-singular at $(0,v,\overline{v})$, where $$J_{ij}=\frac{\partial f_i}{\partial U_j}, \quad i,j = 1\dots 2N,
\label{eq:CIFTJ}$$ and $U~:=~\{V_1,\dots , V_N,\overline{V}_1,\dots , \overline{V}_N\}$, then there exist holomorphic functions $V_i(z)$ and $\overline{V_i}(z)$ of $z$ that satisfy (\[eq:ift\]) in a neighbourhood of $z\!=\!0$.
In this setup, $f$ is clearly a holomorphic mapping, and the values of $v$ in the seed solution $f(0,v,\overline{v})=0$ are the solution to .
Using , the values of $J$ can be computed as follows $$\begin{aligned}
\frac{\partial f_i}{\partial V_j}(0,v,\overline{v})=\overline{v_i}Y_{ij},\quad i,j=1,\dots,N\\
\frac{\partial f_i}{\partial \overline{V_j}}(0,v,\overline{v})=0,\quad i,j=1,\dots,N\\
\frac{\partial f_{N+i}}{\partial V_j}(0,v,\overline{v})=0,\quad i,j=1,\dots,N\\
\frac{\partial f_{N+i}}{\partial \overline{V_j}}(0,v,\overline{v})=v_iY^*_{ij},\quad i,j=1,\dots,N
\end{aligned}$$ and so $J$ can be rewritten as $$J=\begin{pmatrix}\overline{v'}Y'&0\\0&v'(Y')^*\end{pmatrix}$$ where $v'=\textrm{diag}(v_1,\dots,v_N)$ and $Y'$ represents the admittance matrix without the slack bus row and column. Clearly $J$ is non-singular iff $Y'$ is non-singular which is guaranteed by Assumption 1 and therefore $V(z)$ and $\overline{V}(z)$ are holomorphic functions of $z$.
Reflecting Condition Redundancy {#sec:refcond1}
-------------------------------
Now we will show that, in a neighbourhood of $z=0$, the reflecting condition is redundant. In particular we will show that any solution to (\[eq:helm-first\])—which will automatically satisfy $V_i(0)\ne 0\,\forall i$—must satisfy the reflecting condition. If we do not use the reflecting condition, then instead of we obtain the following set of equations $$\begin{aligned}
\sum_{k\in{\mathcal}{B}}Y_{ik}\sum_{n=0}^{\infty} V_k[n]z^n&= zS_i^*\sum_{n=0}^{\infty} \overline{W_i}[n]z^n\\
\sum_{k\in{\mathcal}{B}}Y_{ik}^*\sum_{n=0}^{\infty} \overline{V_k}[n]z^n&= zS_i\sum_{n=0}^{\infty} W_i[n]z^n
\end{aligned}\label{eq:refcondred}$$ where again $W_i=1/V_i$ and likewise $\overline{W_i}=1/\overline{V_i}$. In comparing coefficients of $z^n$, we get: $$\begin{bmatrix}Y'&0\\0&(Y')^*\end{bmatrix}\begin{bmatrix}V[n]\\ \overline{V}[n]\end{bmatrix} = \begin{bmatrix}r_{1,n-1} \\ r_{2,n-1}\end{bmatrix}
\label{eq:refcondred1}$$ with $$\begin{aligned}
r_{1,n-1}&=\left\{\begin{array}{lr}
-Y_0&n=0\\
S_i^*\overline{W_i}[n-1]&n\geq 1
\end{array}\right.\\
r_{2,n-1}&=\left\{\begin{array}{lr}
-Y_0^*&n=0\\
S_iW_i[n-1]&n\geq 1
\end{array}\right.\\
\end{aligned}\label{eq:refcondrhs1}$$ where $Y_0$ is the slack bus column of the admittance matrix.
Now if we take the complex conjugate of the above system, we obtain: $$\begin{bmatrix}(Y')^*&0\\0&Y'\end{bmatrix}\begin{bmatrix}V^*[n]\\ \overline{V}^*[n]\end{bmatrix} = \begin{bmatrix}r_{1,n-1}^* \\ r_{2,n-1}^*\end{bmatrix}$$ and by rearranging $$\begin{bmatrix}Y'&0\\0&(Y')^*\end{bmatrix}\begin{bmatrix}\overline{V}^*[n]\\ V[n]\end{bmatrix} = \begin{bmatrix}r_{2,n-1}^* \\ r_{1,n-1}^*\end{bmatrix}
\label{eq:refcondred2}$$
From it follows that when $n\!=\!0$, $r_{2,-1}=-Y_0^*=r_{1,-1}^*$. Thus the right-hand side in is the same as in . Under Assumption 1 $Y'$ is non-singular and therefore the solutions of the two systems must be the same: namely $V[0]=\overline{V}[0]^*$.
We now assume that $V[n]=\overline{V}^*[n]$ — and hence $W[n] = \overline{W}[n]^*$ — is true up to $n=k$ and check that for $n=k+1$
$$\begin{aligned}
r_{1,k}&= S^*\overline{W}[k] \\
&=S^*W[k]^*&\textrm{from inductive hypothesis}\\
&=r_{2,k}^*
\end{aligned}$$
Once again we have that the right hand side in is identical to the one in , and so $V[n]=\overline{V}[n]^*$ for $n\geq0$. The reflecting condition holds for all coefficients of the power series of $V_i(z)$ and $\overline{V}_i(z)$ and therefore $\overline{V}_i(z)=V_i(z^*)^*$ holds for the functions themselves. Moreover, if the \[L/M\] Padé Approximant for $V(z)$ is given by $p(z)/q(z)$, then it is straightforward to show that the \[L/M\] Padé Approximant to $\overline{V}(z)$ is given by $(p(z^*))^*/(q(z^*))^*$ and so the reflecting condition holds also for the analytically continued function. From, for example, [@Ahlfors] we know that the analytically continued functions satisfy the polynomial HE equations
Introducing PV Buses {#sec:subramanian}
====================
The HELM method as described in [@helm] had one major deficiency — it did not describe how to deal with networks that include PV buses. When considering PV buses, as seen in Table \[tab:bustypes\], the unknowns in the BPEE are different. At a PQ bus the real and reactive power injections are known and the complex voltage is unknown, whereas at a PV bus the real power injection and the voltage magnitude are known and the reactive power injection and the voltage angle are unknown. As we are solving for different variables in the BPEE, it is necessary to rearrange the equations. Without loss of generality we will consider the systems to be ordered such that the PQ buses are grouped first and then the PV buses come afterwards.
------------------- ------------- ------------------ --------------- ----------------
Type of Voltage Mag Voltage Real Power Reactive Power
Bus (|V|) Angle ($\delta$) Injection (P) Injection (Q)
Slack (V$\delta$) Given Given Unknown Unknown
PQ Bus Unknown Unknown Given Given
PV Bus Given Unknown Given Unknown
------------------- ------------- ------------------ --------------- ----------------
: Types of buses in a power system
\[tab:bustypes\]
In [@subramanian], Subramanian et al present an approach to deal with general networks that may include PV buses. For PQ buses, their holomorphic embedding of the BPEE uses the alternative set of equations .
For PV buses, the voltage magnitude and real power at the bus are known, but not the reactive power. Thus the authors create equations that use only the real power at a bus. Adding a number’s complex conjugate to itself results in eliminating the imaginary part of that number—using this strategy with equation , the authors suggest the following holomorphic embedding to replace (\[eq:helm-alternative\]) for PV buses $$\begin{aligned}
M_i^2{\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k(z)& = z2P_iV_i(z) + (1\!\!-\!\!z)M_i^2y_i - {}\\
&z\!\left(V_i(z)^2\sum_{k=0}^NY_{ik}^*\overline{V_k}(z)\right) , i\in {\mathcal}{B}_{PV}\\
\overline{V_i}(z)V_i(z) &= 1+z(M_i-1),\ \ \ i\in{\mathcal}{B}_{PV}
\end{aligned}\label{eq:mod1-syseqnPV}$$ where $M_i$ is the target voltage magnitude for PV buses. The second equation in the holomorphic embedding, is not explicitly shown in [@subramanian], but is required to provide a path for the voltage magnitudes in the PV buses to start at 1 when $z\!=\!0$ and finish at $M_i$ when $z\!=\!1$.
Now the reflecting condition substituted into and combine to form the holomorphic embedding of the entire network.
When $z\!=\!0$ a solution to the system is simply $V_i(0)=1$, $i\in{\mathcal}{B}$. The power series coefficients are then determined by the same process as in the original HELM method. However, for PV buses, the term $V_i(z)^2\sum_{k=0}^NY_{ik}^*\overline{V_k}(z)$ contains products of three power series. This results in double convolutions, which can have precision limitations and can lead to inaccuracies in the final results. This problem is discussed in [@subramanian].
When applied to the IEEE test cases, the Subramanian model in [@subramanian] had poor convergence in even the simplest 9-Bus case, which is confirmed by our results in Table \[tab:1padecase9\]. Here, $Rs$ refers to the residual when the voltages are substituted into the original BPEE, and $\Delta$ refers to the difference between the model voltage results and the voltage results obtained through MatPower. For more complicated cases the model was unable to provide even approximately correct results, as shown for the 39-Bus case in Table \[tab:1padecase39\].
Padé Order Max $|Rs|$ Max $|\Delta|$
------------ ------------ ----------------
$[5/5]$ 1.4e-01 1.2e-01
$[10/10]$ 3.9e-01 7.4e-02
$[15/15]$ 2.2e-02 4.8e-03
$[20/20]$ 5.4e-03 7.9e-04
$[25/25]$ 4.8e-04 6.6e-05
: Results for IEEE case9 using Subramanian Model[]{data-label="tab:1padecase9"}
Padé Order Max $|Rs|$ Max $|\Delta|$
------------ ------------ ----------------
$[5/5]$ 2.3e+01 4.2e+00
$[10/10]$ 4.8e+00 1.1e+00
$[15/15]$ 1.2e+01 8.5e-01
$[20/25]$ 2.5e+01 2.1e+00
$[25/25]$ 1.1e+01 1.1e+00
: Results for IEEE case39 using Subramanian Model[]{data-label="tab:1padecase39"}
Updating his previous work in [@subramanian], Subramanian provides a revised model in his thesis [@subrthesis]. This revised model no longer has the double convolution issue and is better suited to solving larger models. The model also no longer involves a $(1-z)$ term, instead it splits the admittance matrix into two parts, creating a diagonal matrix for the shunt effects which allows the remaining transmission elements to have zero row sums. The shunt elements are moved to the right-hand side and are multiplied by the complex parameter $z$.[^3] The holomorphic embedding becomes
$$\begin{aligned}
&\!\!\!\!{\sum_{k\in{\mathcal}{B}}}Y_{\text{trans}_{ik}}V_k(z) = \frac{zS_i^*}{V_i^*(z^*)} -zY_{\text{shunt}_{ii}}V_i(z),\ i\in {\mathcal}{B}_{PQ}\\
&\!\!\!\!\left(\!\!V_i(z){\sum_{k\in{\mathcal}{B}}}Y_{\text{trans}_{ik}}^*V_k^*(z^*)\!\!\right)\!\!+\!\!\left(\!\!V_i^*(z^*){\sum_{k\in{\mathcal}{B}}}Y_{\text{trans}_{ik}}V_k(z)\!\!\right) \\
&\ \ = 2zP_i - 2\text{Re}\left\{Y_{\text{shunt}_{ii}}V_i(z)V_i^*(z^*)\right\},\ \ \ i\in{\mathcal}{B}_{PV}\\
&\!\!\!\!\overline{V_i}(z)V_i(z) = 1+z(M_i-1),\ \ \ i\in{\mathcal}{B}_{PV}
\end{aligned}\label{eq:subrthes-syseqnPQ}$$
This has the similar effect of making the seed voltages at every bus equal to $1$. This model is different from our general model, which we provide in the next section. By separating the shunt elements from the admittance matrix, it creates an additional term which is dependent on the square of the voltage. In contrast, our general model is capable of creating either a linearly voltage-dependent term or a voltage-independent term. The results of both of Subramanian’s models are given for the IEEE test cases in Section \[sec:results\].
New PV Models {#sec:PVmodels}
=============
In this section we present a general parametrised model for mixed PQ/PV systems that does not have the drawback of having a double convolution. We will show that for models of this general form the $V$ and $\overline{V}$ are holomorphic functions in $z$ and that the reflecting condition remains redundant. Finally we will present specific choices for the parameters of the general model and investigate the numerical behaviour of the resulting methods.
General Model
-------------
We suggest the following general model: for PQ buses, the following equations are used in the place of (\[eq:helm-alternative\]): $$\begin{aligned}
\overline{V_i}(z) \left( {\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k(z)+ (z\!\!-\!\!1)a_i \right) &+{}\\
(z\!\!-\!\!1)b_i &= zS_i^*\quad i\in {\mathcal}{B}_{PQ}\\
V_i(z)\left( {\sum_{k\in{\mathcal}{B}}}Y^*_{ik}\overline{V_k}(z)+ (z\!\!-\!\!1)a^*_i \right) &+{}\\
(z\!\!-\!\!1)b^*_i &= zS_i\quad i\in {\mathcal}{B}_{PQ}
\end{aligned}\label{eq:genPQ}$$
This is similar to the approach used in both [@helmpatent] and [@subramanian], and is identical if $a_i = y_i$ and $b_i=0,\quad i\in {\mathcal}{B}_{PQ}$. Natural choices for $a_i$ and $b_i$ would be $y_i$ or $0$ and these terms would serve a similar purpose as the additional $(1-z)y_i$ term in (\[eq:helm-alternative\]). $a_i$ is multiplied by the voltage-dependent term, while $b_i$ is voltage-independent. For PV buses we use the following equations in the place of : $$\begin{aligned}
\overline{V_i}(\!z\!)\!\!&\left({\sum_{k\in{\mathcal}{B}}}\!\! Y_{ik}V_k(\!z\!)\!+\!(z\!\!-\!\!1)a_i\!\!\right)\!+\! V_i(\!z\!)\!\!\left({\sum_{k\in{\mathcal}{B}}}\!\!Y^*_{ik}\overline{V_k}(\!z\!)+{}\right.\\
&\left.\vphantom{{\sum_{k\in{\mathcal}{B}}}}(z\!\!-\!\!1)a^*_i\right)+(z\!\!-\!\!1)(b_i+b_i^*) = 2zP_i\quad i\in {\mathcal}{B}_{PV}
\end{aligned}\label{eq:genPV1}$$ $$\begin{aligned}
\overline{V_i}(z)V_i(z)=L_i^2(z)\quad i\in {\mathcal}{B}_{PV}\label{eq:genPV2}\end{aligned}$$ where $L_i(z)$, which is assumed to be real-valued, describes how the target voltage magnitude changes with respect to $z$.\
Equation uses the same approach to eliminate the unknown reactive power as shown in [@subramanian], except there is one less voltage term in both the left and right hand sides, so our model does not give rise to any double convolutions and their possible accuracy problems. Equation gives us the voltage magnitude at each PV bus, which may depend on the value of $z$. The form of $L_i(z)$ is unrestricted except that $L_i(1)$ must equal $M_i$.
Equating coefficients of $z^n,n\geq 1$ in will yield the following equations for $n=1$ $$\begin{aligned}
\overline{V_i}[0]{\sum_{k\in{\mathcal}{B}}}&Y_{ik}V_k[1] + \overline{V_i}[1]\left({\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k[0]-a_i\right)= \\
&S_i^*-\overline{V_i}[0]a_i-b_i, \qquad\qquad\qquad\quad i \in {\mathcal}{B}_{PQ}\\
V_i[0]{\sum_{k\in{\mathcal}{B}}}&Y^*_{ik}\overline{V_k}[1] + V_i[1]\left({\sum_{k\in{\mathcal}{B}}}Y^*_{ik}\overline{V_k}[0]-a^*_i\right)= \\
&S_i-V_i[0]a^*_i-b^*_i,\qquad\qquad \qquad\quad i \in {\mathcal}{B}_{PQ}
\end{aligned}\label{eq:gen-powseries1}$$ and for $n\geq 2$ $$\begin{aligned}
\overline{V_i}[0]&{\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k[n] + \overline{V_i}[k]({\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k[0]-a_i)=\\
&-\!\!\sum_{m=1}^{n-1}\!\overline{V_i}[m]\!\!\left({\sum_{k\in{\mathcal}{B}}}\! Y_{ik}V_k[n-m]+ \delta_{m,n-1}a_i\!\right), i \in {\mathcal}{B}_{PQ}\\
V_i[0]&{\sum_{k\in{\mathcal}{B}}}Y^*_{ik}\overline{V_k}[n] + V_i[k]({\sum_{k\in{\mathcal}{B}}}Y^*_{ik}\overline{V_k}[0]-a^*_i)=\\
&-\!\!\sum_{m=1}^{n-1}\!V_i[m]\!\!\left({\sum_{k\in{\mathcal}{B}}}\! Y_{ik}^*\overline{V_k}[n-m]+\delta_{m,n-1}a_i^*\!\right), i \in {\mathcal}{B}_{PQ}
\end{aligned}\label{eq:gen-powseries2}$$\
Similarly, equating coefficients of $z^n$ in yields for $n=1$ $$\begin{aligned}
\overline{V_i}&[0]{\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k[1] + \overline{V_i}[1]\left({\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k[0]-a_i\right) +{}\\
&V_i[0]{\sum_{k\in{\mathcal}{B}}}Y_{ik}^* \overline{V_k}[1]+V_i[1]\left({\sum_{k\in{\mathcal}{B}}}Y^*_{ik}\overline{V_k}[0]-a^*_i\right) = \\
&\quad 2P_i-\overline{V_i}[0]a_i-V_i[0]a_i^*-b_i-b^*_i,\qquad i\in {\mathcal}{B}_{PV}
\end{aligned}$$ and for $n\geq2$ $$\begin{aligned}
\overline{V_i}&[0]{\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k[n] + \overline{V_i}[n]\left({\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k[0]-a_i\right) + \\
&V_i[0]{\sum_{k\in{\mathcal}{B}}}Y_{ik}^* \overline{V_k}[n]+V_i[n]\left({\sum_{k\in{\mathcal}{B}}}Y^*_{ik}\overline{V_k}[0]-a^*_i\right) = \\
&-\!\!\sum_{m=1}^{n-1}\overline{V_i}[m]\left({\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k[n\!-\!m]+\!\delta_{m,n-1}a_i\!\right)\\
&-\!\!\sum_{m=1}^{n-1}V_i[m]\left({\sum_{k\in{\mathcal}{B}}}Y_{ik}^*\overline{V_k}[n\!-\!m]+\!\delta_{m,n-1}a_i^*\!\right)
\end{aligned}$$\
Finally, equating coefficients of $z^n$ in gives $$\begin{aligned}
\overline{V_i}[0]V_i[n]+\overline{V_i}[n]V_i[0]=&-\sum_{m=1}^{n-1}\overline{V_i}[m]V_i[m-n] + L_i^2[n]\\
\end{aligned}\label{eq:genpowseriesend}$$
Equations – can be written in the following simple form, where the matrix on the left is independent of $n$. $$\begin{bmatrix} A_{PQ_1}&A_{PQ_2}\\ A_{PV_1}&A_{PV_2}\\ A_{PQ_3}&A_{PQ_4}\\A_{PV_3}&A_{PV_4} \end{bmatrix}\begin{bmatrix}V[n]\\ \overline{V}[n]\end{bmatrix}=\begin{bmatrix}r_{PQ_1,n\!-\!1}\\r_{PV_1,n\!-\!1}\\r_{PQ_2,n\!-\!1}\\r_{PV_2,n\!-\!1}\end{bmatrix}\label{eq:genblock}$$
where $$\begin{aligned}
A_{PQ_1,{ij}}&= \overline{V_i}[0]Y_{ij}\\
A_{PV_1,{ij}}&= \overline{V_i}[0]Y_{ij}+\delta_{_{i,j}}\left({\sum_{k\in{\mathcal}{B}}}Y_{ik}^*\overline{V_k}[0]-a_i^*\right)\\
A_{PQ_2,{ij}}&= \delta_{_{i,j}}\left({\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k[0]-a_i\right)\\
A_{PV_2,{ij}}&= V_i[0]Y^*_{ij}+\delta_{_{i,j}}\left({\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k[0]-a_i\right)\\
A_{PQ_3,{ij}}&= \delta_{_{i,j}}\left({\sum_{k\in{\mathcal}{B}}}Y_{ik}^*\overline{V_k}[0]-a_i^*\right)\\
A_{PV_3,{ij}}&= \delta_{_{i,j}}\overline{V_i}[0]\\
A_{PQ_4,{ij}}&= V_i[0]Y_{ij}^*\\
A_{PV_4,{ij}}&= \delta_{_{i,j}}V_i[0]
\end{aligned}$$ and $$\begin{aligned}
r_{PQ_1,n\!-\!1,i}&= \delta_{_{n,1}}\left(S_i^*-\overline{V_i}[0]a_i-b_i\right)-{}\\
& \quad\!\!\!\sum\limits_{m=1}^{n-1}\!\overline{V_i}[m]\!\!\left({\sum_{k\in{\mathcal}{B}}}\limits \!\!Y_{ik}V_k[n\!-\!m]\!+\!\delta_{_{m,n-1}}a_i\!\!\right)\\
r_{PV_1,n\!-\!1,i}&= \delta_{_{n,1}}\left(2P_i-\overline{V_i}[0]a_i-V_i[0]a_i^*-b_i-b_i^*\right)-{}\\
&\quad\!\!\!\sum\limits_{m=1}^{n-1}\!\overline{V_i}[m]\!\!\left({\sum_{k\in{\mathcal}{B}}}\limits \!\!Y_{ik}V_k[n\!-\!m]\!+\!\delta_{_{m,n-1}}a_i\!\!\right)-{}\\
&\quad\!\!\!\sum\limits_{m=1}^{n-1}\!V_i[m]\!\!\left({\sum_{k\in{\mathcal}{B}}}\limits \!\!Y^*_{ik}\overline{V_k}[n\!-\!m]\!+\!\delta_{_{m,n-1}}a_i^*\!\!\right)\\
r_{PQ_2,n\!-\!1,i}&= \delta_{_{n,1}}\left(S_i-V_i[0]a^*_i-b^*_i\right)-{}\\
&\quad\!\!\!\sum\limits_{m=1}^{n-1}\!V_i[m]\!\!\left({\sum_{k\in{\mathcal}{B}}}\limits \!\!Y^*_{ik}\overline{V_k}[n\!-\!m]\!+\!\delta_{_{m,n-1}}a_i^*\!\!\right)\\
r_{PV_2,n\!-\!1,i}&= -\!\!\sum\limits_{m=1}^{n-1}\overline{V_i}[m]V_i[n-m]+L_i^2[n]\end{aligned}$$
Values for $V_i[0]$ and $\overline{V}_i[0]$ can be obtained by solving (\[eq:genPQ\])–(\[eq:genPV2\]) at $z=0$ or equivalently equating the constant terms (coefficient for $z^0$). Unfortunately this leads to a nonlinear system of equations for the $V_i[0],\overline{V}_i[0]$ and a solution for general $a_i, b_i, L_i$ can not be derived easily. However, for each of the specific models that we consider it is possible to state a simple choice for $V_i[0],\overline{V}_i[0]$ that furthermore satisfies $\overline{V}_i[0] = V^*_i[0]$. This choice will be used as the seed, allowing us to calculate all further coefficients of the power series from the recurrence (\[eq:genblock\]).
Holomorphicity and Reflecting Condition
---------------------------------------
We now assume that we have a seed solution $(v, \overline{v})$, with $v_i = V_i[0], \overline{v}_i= \overline{V}_i[0]$, to (\[eq:genPQ\])–(\[eq:genPV2\]) at $z=0$ and that furthermore this seed satisfies $\overline{v} = v^*$.
Proving that this seed can be continued into holomorphic functions $V(z)$ and $\overline{V}(z)$ of $z$ that solve the general model follows a similar approach as in Section \[sec:HELM-holo\]. Converting equations (\[eq:genPQ\])–(\[eq:genPV2\]) into functions to replace $f$ in , we need to prove that $J$, defined in , is non-singular at $(0,v,\overline{v})$. Examination of $J$ and $A$ from at $(0,v,\overline{v})$ shows that the two matrices are equivalent at this point. Therefore $A$ being nonsingular is sufficient for $V(z),\overline{V}(z)$ to be holomorphic using the CIFT. While it is possible to create special networks where this does not hold, especially if no restrictions are placed on the values of $a$ and $b$, for general networks and sensibly chosen values for $a$ and $b$ the condition will hold, as indeed is true for the IEEE models used for the numerical tests in Section \[sec:results\].
The proof that the reflecting condition is implied follows the same pattern as the proof given in Section \[sec:refcond1\]. Note that for the PQ-part of (\[eq:genblock\]) we have $r_{PQ_1,n\!-\!1}=r_{PQ_2,n\!-\!1}^*$. In the PV part, $r_{PV_1,n\!-\!1}$ and $r_{PV_2,n\!-\!1}$ are both real-valued: Assuming that we have already proven that $\overline{V}_i[k] = V^*_i[k]$ for all $k<n$, then $r_{PV_1,n\!-\!1}$, can be written as the sum of a complex number and its complex conjugate, whereas for $r_{PV_2,n\!-\!1}$ we have $\sum_{m=1}^{n-1}\overline{V_i}[m]V_i[n-m] \in {\mathcal}{R}$ since it is equal to its own complex conjugate, and $L_i(z)\in {\mathcal}{R}$ by assumption.
With this in mind, taking the complex conjugates of both sides of (\[eq:genblock\]) we obtain $$\begin{bmatrix} A_{PQ_1}^\ast&A_{PQ_2}^*\\ A_{PV_1}^*&A_{PV_2}^*\\A_{PQ_3}^*&A_{PQ_4}^*\\A_{PV_3}^*&A_{PV_4}^* \end{bmatrix}\begin{bmatrix}V[n]^*\\ \overline{V}[n]^*\end{bmatrix}=\begin{bmatrix}r_{PQ_1,n\!-\!1}^*\\r_{PV_1,n\!-\!1}^*\\r_{PQ_1,n\!-\!1}^*\\r_{PV_2,n\!-\!1}^*\end{bmatrix}$$ which, after noting how $A_{PQ_1}^* = A_{PQ_4}$, $A_{PQ_2}^* = A_{PQ_3}$, $A_{PV_1}=A_{PV_2}$, and $A_{PV_3}=A_{PV_4}$, we can rearrange and rewrite as $$\begin{bmatrix} A_{PQ_1}&A_{PQ_2}\\ A_{PV_1}&A_{PV_2}\\A_{PQ_3}&A_{PQ_4}\\A_{PV_3}&A_{PV_4} \end{bmatrix}\begin{bmatrix}\overline{V}[n]^*\\ V[n]^*\end{bmatrix}=\begin{bmatrix}r_{PQ_1,n\!-\!1}\\r_{PV_1,n\!-\!1}\\r_{PQ_2,n\!-\!1}\\r_{PV_2,n\!-\!1}\end{bmatrix}
\label{eq:refcondpv2}$$
Comparing with the original system and assuming as above that $A$ is non-singular we must have that $V[n]=\overline{V}^*[n]$ and $V^*[n]=\overline{V}[n]$ for $n\geq 0$ as long as the reflecting condition holds for the seed $V_i[0] = V_i(0), \overline{V}_i[0] = \overline{V}_i(0)$. The reflecting condition $\overline{V}(z)=V(z^*)^*$ then follows.
Different choices of $a_i$, $b_i$, and $L_i(z)$ result in different paths for the bus voltages between $z\!=\!0$ and $z\!=\!1$. The obvious choices for $a_i$ and $b_i$ are either to both be 0, eliminating the terms from the model, or for one to be 0 and the other to be $y_i$, which makes the system of equations trivial at $z=0$. There is no single obvious choice for $L_i(z)$, though we have decided to have the voltage magnitudes scale linearly with $z$ in our models. Table \[tab:bustypes1\] provides the values for $a_i, b_i,$ and $L_i(z)$ that we have chosen for the four models we present in this paper.
Model $a_i$ $b_i$ $L_i(z)$ $V_i(0)$
------- --------------------------------------------- -------- ------------------------------ -------------
1 $y_i$ 0 $ 1+z(M_i-1) $ 1
2 ${\sum_{k\in{\mathcal}{B}}}Y_{ik}\lambda_k$ 0 $M_i$ $\lambda_i$
3 0 0 $\|\nu_i\|+z(M_i-\|\nu_i\|)$ $\nu_i$
4 0 $y_i $ $1+z(M_i-1) $ 1
: Outline of PV methods
\[tab:bustypes1\]
Model 1 {#sec:mod4}
-------
The first of our models sets $a_i=y_i$, $b_i=0$, and $L_i=1+z(M_i-1)$. With these values, we obtain a model that is very similar to the Subramanian model given in [@subramanian], differing only in that the double convolutions are removed. Using the reflecting condition, the holomorphic embedding simplifies to $$\begin{aligned}
\textrm{Re} &\left\{\!\!V_i^*(z^*)\left(\!{\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k(z)-(1\!\!-\!\!z)y_i\right)\!\! \right\}=zP_i,&&i \in {\mathcal}{B}\\
\textrm{Im} &\left\{\!\!V_i^*(z^*)\left(\!{\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k(z)-(1\!\!-\!\!z)y_i\right)\!\! \right\}=-zQ_i,\!\!\!\!\!&& i \in {\mathcal}{B}_{PQ}\\
V_i^*&(z^*)V_i(z) = (1+z(M_i-1))^2,&& i\in{\mathcal}{B}_{PV}
\end{aligned}\label{eq:mod4-hol}$$ The last equation gives the PV buses a voltage magnitude of 1 when at $z\!=\!0$, which allows the initial model to have the simple solution of $V_i(0)=1,i\in {\mathcal}{B}$ when $z\!=\!0$, while allowing the voltages in the PV buses to reach their required magnitude, $M_i$, when $z\!=\!1$. All the equations are now real, so we split the voltage $V_i$ into its real and imaginary parts, and so equation becomes $$\begin{bmatrix} A_{_1}&A_{_2}\\ A_{PQ_3}&A_{PQ_4}\\A_{PV_3}&A_{PV_4} \end{bmatrix}\begin{bmatrix}\textrm{Re}\{V[n]\}\\ \textrm{Im}\{V[n]\}\end{bmatrix}=\begin{bmatrix}\textrm{r}_{_{1,n-1}}\\\textrm{r}_{PQ_{2,n-1}}\\\textrm{r}_{PV_{2,n-1}}\end{bmatrix}\label{eq:mod1block}$$
where the A matrix is $$\begin{aligned}
A_{1_{ij}}&= G_{ij}, && i\in {\mathcal}{B}\\
A_{2_{ij}}&= B_{ij}, && i\in {\mathcal}{B}\\
A_{PQ_{3_{ij}}}&= B_{ij}, && i\in {\mathcal}{B}_{PQ}\\
A_{PV_{3_{ij}}}& = 2\delta_{i,j},&& i\in{\mathcal}{B}_{PV}\\
A_{PQ_{4_{ij}}}&= G_{ij},&& i\in {\mathcal}{B}_{PQ}\\
A_{PV_{4_{ij}}}&= 0,&& i\in{\mathcal}{B}_{PV}
\end{aligned}$$
and the right-hand side becomes
$$\begin{aligned}
\!\!\textrm{r}&_{_{1,n-1,i}}= \delta_{_{n,1}}\left(P_i-\text{Re}\left\{y_i\right\}\right)-&{}\\
&\quad\!\!\!\text{Re}\left\{\!\sum\limits_{m=1}^{n-1}\!V_i^*[m]\!\!\left({\sum_{k\in{\mathcal}{B}}}\limits \!\!Y_{ik}V_k[n\!-\!m]\!+\!\delta_{_{m,n-1}}y_i\!\!\right)\!\!\right\},&&i\in{\mathcal}{B}\\
\!\!\textrm{r}&_{PQ_{2,n-1,i}}= \delta_{_{n,1}}\left(-Q_i-\text{Im}\left\{y_i\right\}\right)-&{}\\
&\quad\!\!\!\text{Im}\left\{\!\sum\limits_{m=1}^{n-1}\!V_i^*[m]\!\!\left({\sum_{k\in{\mathcal}{B}}}\limits \!\!Y_{ik}V_k[n\!-\!m]\!+\!\delta_{_{m,n-1}}y_i\!\!\right)\!\!\right\},&&i\in{\mathcal}{B}_{PQ}\\
\textrm{r}&_{PV_{2,n-1,i}}= -\!\!\sum\limits_{m=1}^{n-1}V_i^*[m]V_i[n-m]+L_i^2[n], &&i\in{\mathcal}{B}_{PV}\end{aligned}$$
We may calculate the power series of $V$ (and $\overline{V}$) to any desired degree $n$. This model does not suffer from the double convolution found in [@subramanian], and in Section \[sec:results\] we will show that this model leads to more accurate solutions of the IEEE test cases.
Model 2 {#sec:mod2}
-------
The idea behind Model 2 is to keep the voltage magnitudes of PV buses constant regardless of the value of $z$. This is accomplished by first setting $a_i={\sum_{k\in{\mathcal}{B}}}Y_{ik}\lambda_k$ and $b_i=0$ in and , where all $\lambda_k$ are constant and $\lambda_k=1, k\in{\mathcal}{B}_{PQ}\cup\{0\}$, and $\lambda_k=\|V_k\|, k\in{\mathcal}{B}_{PV}$. As well, in we set $L_i(z)=\|V_i\|$, which is independent of $z$. The holomorphic embedding thus becomes:
$$\begin{aligned}
\textrm{Re}&\left\{ V_i^*(z^*)\left({\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k(z)-{}\right.\right.\\
&\left.\left.(1\!\!-\!\!z){\sum_{k\in{\mathcal}{B}}}Y_{ik}\lambda_k\right)\!\!\right\}=zP_i,\quad i \in {\mathcal}{B}\\
\textrm{Im}&\left\{ V_i^*(z^*)\left({\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k(z)-{}\right.\right.\\
&\left.\left.(1\!\!-\!\!z){\sum_{k\in{\mathcal}{B}}}Y_{ik}\lambda_k\right)\!\!\right\}=-zQ_i,\quad i \in {\mathcal}{B}_{PQ}\\
V_i^*&(z^*)V_i(z) = \lambda_i^2 ,\quad i\in{\mathcal}{B}_{PV}
\end{aligned}\label{eq:mod2-holo}$$
For Model 2, when $z\!=\!0$, $V_i(0)=\lambda_i, \overline{V_i}(0)=\lambda_i$ is a valid solution for $i\in{\mathcal}{B}$.
Here, the $A$ matrix from equation becomes: $$\begin{aligned}
A_{1_{ij}}&= G_{ij}\lambda_i, && i\in {\mathcal}{B}\\
A_{2_{ij}}&= B_{ij}\lambda_i, && i\in {\mathcal}{B}\\
A_{PQ_{3_{ij}}}&= B_{ij}\lambda_i, && i\in {\mathcal}{B}_{PQ}\\
A_{PV_{3_{ij}}}& = 2\delta_{i,j}\lambda_i,&& i\in{\mathcal}{B}_{PV}\\
A_{PQ_{4_{ij}}}&= G_{ij}\lambda_i,&& i\in {\mathcal}{B}_{PQ}\\
A_{PV_{4_{ij}}}&= 0,&& i\in{\mathcal}{B}_{PV}
\end{aligned}$$
while the right-hand side becomes
$$\begin{aligned}
\!\!\textrm{r}&_{_{1,n-1,i}}= \delta_{_{n,1}}\left(P_i-\lambda_i{\sum_{k\in{\mathcal}{B}}}G_{ik}\lambda_k\right)-&{}\\
&\ \!\!\text{Re}\!\left\{\!\sum\limits_{m=1}^{n-1}\!\!V_i^*[m]\!\!\left({\sum_{k\in{\mathcal}{B}}}\limits \!\!Y_{ik}V_k[n\!-\!m]\!+\!\delta_{_{m,n\!-\!1}}\!\!{\sum_{k\in{\mathcal}{B}}}\!Y_{ik}\lambda_k\!\!\right)\!\!\!\right\}\!\!,\ \!i\in{\mathcal}{B}\\
\!\!\textrm{r}&_{PQ_{2,n-1,i}}= \delta_{_{n,1}}\left(-Q_i-\lambda_i{\sum_{k\in{\mathcal}{B}}}B_{ik}\lambda_k\right)-&{}\\
&\ \!\!\text{Im}\!\left\{\!\sum\limits_{m=1}^{n-1}\!\!V_i^*[m]\!\!\left({\sum_{k\in{\mathcal}{B}}}\limits \!\!Y_{ik}V_k[n\!-\!m]\!+\!\delta_{_{m,n\!-\!1}}\!\!{\sum_{k\in{\mathcal}{B}}}\!Y_{ik}\lambda_k\!\!\right)\!\!\!\right\}\!\!,\ \!i\in{\mathcal}{B}_{PQ}\\
\textrm{r}&_{PV_{2,n-1,i}}= -\!\!\sum\limits_{m=1}^{n-1}V_i^*[m]V_i[n-m],\qquad i\in{\mathcal}{B}_{PV}\end{aligned}$$
The $A$ matrix is identical to that in Model 1 save the introduction of $\lambda$ – the main difference is in the right-hand side. We can once again solve for $V$ to any desired value of $n$.
Model 3 {#sec:mod3}
-------
In Model 3 both $a_i$ and $b_i$ are set to 0. This gives a model similar in form to the model used in the HELM method given in [@helmpatent]—in the absence of PV buses they are identical. The solution at $z=0$ requires solving a simple set of equations. We obtain $V(0)=\nu$ and $\overline{V}(0)=\nu^*$, where $\nu$ is the solution to
$${\sum_{k\in{\mathcal}{B}}}Y_{ik} \nu_k = 0\qquad i\in{\mathcal}{B}$$
We also scale the voltage magnitudes linearly with respect to $z$, so $L_i(z)= \|\nu_i\|+z(M_i-\|\nu_i\|)$. The holomorphic embedding thus becomes: $$\begin{aligned}
& \textrm{Re}\left\{V_i^*(z^*){\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k(z)\!\right\}=zP_i , && i \in {\mathcal}{B}\\
& \textrm{Im}\left\{V_i^*(z^*){\sum_{k\in{\mathcal}{B}}}Y_{ik}V_k(z)\!\right\}=-zQ_i , && i \in {\mathcal}{B}_{PQ}\\
& V_i^*(z^*)V_i(z) = \left(\|\nu_i\|+z(M_i-\|\nu_i\|)\right)^2 , && i\in{\mathcal}{B}_{PV}
\end{aligned}\label{eq:mod3-holo}$$
Here, the $A$ matrix from equation becomes: $$\begin{aligned}
A_{1_{ij}}&= \nu_iG_{ij}, && i\in {\mathcal}{B}\\
A_{2_{ij}}&= \nu_iB_{ij}, && i\in {\mathcal}{B}\\
A_{PQ_{3_{ij}}}&= \nu_iB_{ij}, && i\in {\mathcal}{B}_{PQ}\\
A_{PV_{3_{ij}}}& = 2\delta_{i,j}\nu_i,&& i\in{\mathcal}{B}_{PV}\\
A_{PQ_{4_{ij}}}&= \nu_iG_{ij},&& i\in {\mathcal}{B}_{PQ}\\
A_{PV_{4_{ij}}}&= 0,&& i\in{\mathcal}{B}_{PV}
\end{aligned}$$
while the right-hand side becomes $$\begin{aligned}
&\textrm{r}_{_{1,n-1,i}}= \delta_{_{n,1}}P_i-\text{Re}\left\{\sum\limits_{m=1}^{n-1}\!V_i^*[m]\!\!{\sum_{k\in{\mathcal}{B}}}\limits \!\!Y_{ik}V_k[n\!-\!m]\!\right\},\quad\!\! i\in{\mathcal}{B}\\
&\textrm{r}_{PQ_{2,n-1,i}}= -\delta_{_{n,1}}Q_i-{}\\
&\qquad\qquad\qquad \text{Im}\left\{\sum\limits_{m=1}^{n-1}\!V_i^*[m]\!\!{\sum_{k\in{\mathcal}{B}}}\limits \!\!Y_{ik}V_k[n\!-\!m]\!\right\},\quad i\in{\mathcal}{B}_{PQ}\\
&\textrm{r}_{PV_{2,n-1,i}}= -\!\!\sum\limits_{m=1}^{n-1}V_i^*[m]V_i[n-m]+L_i^2[n], \quad i\in{\mathcal}{B}_{PV}\end{aligned}$$
We can once again solve for $V$ to any desired value of $n$.
Model 4
-------
The holomorphic embedding for Model 4 is the same as Model 1 except that the term involving $(1\!-\!z)$ is made voltage independent. This is accomplished by setting $a_i=0$, $b_i=y_i$, and $L_i=1+z(M_i-1)$. The holomorphic embedding becomes $$\begin{aligned}
& \textrm{Re}\left\{V_i^*(z^*){\sum_{k\in{\mathcal}{B}}}\!\!Y_{ik}V_k(z)-(1\!\!-\!\!z)y_i\right\}=zP_i, &&\!\!\!i \in {\mathcal}{B}\\
& \textrm{Im}\left\{V_i^*(z^*){\sum_{k\in{\mathcal}{B}}}\!\!Y_{ik}V_k(z)-(1\!\!-\!\!z)y_i\right\}=-zQ_i,&&\!\!\!i \in {\mathcal}{B}_{PQ}\\
&V_i^*(z^*)V_i(z) = (1+z(M_i-1))^2, &&\!\!\!i\in{\mathcal}{B}_{PV}
\end{aligned}\label{eq:mod5-holo}$$
Here, the $A$ matrix from equation becomes: $$\begin{aligned}
A_{1_{ij}}&= G_{ij}+\delta_{_{i,j}}\textrm{Re}\left\{y_i\right\}, && i\in {\mathcal}{B}\\
A_{2_{ij}}&= B_{ij}-\delta_{_{i,j}}\textrm{Im}\left\{y_i\right\}, && i\in {\mathcal}{B}\\
A_{PQ_{3_{ij}}}&= B_{ij} -\delta_{_{i,j}}\textrm{Im}\left\{y_i\right\},&& i\in {\mathcal}{B}_{PQ}\\
A_{PV_{3_{ij}}}& = 2\delta_{i,j},&& i\in{\mathcal}{B}_{PV}\\
A_{PQ_{4_{ij}}}&= G_{ij}+\delta_{_{i,j}}\textrm{Im}\left\{y_i\right\},&& i\in {\mathcal}{B}_{PQ}\\
A_{PV_{4_{ij}}}&= 0,&& i\in{\mathcal}{B}_{PV}
\end{aligned}$$
while the right-hand side becomes $$\begin{aligned}
\textrm{r}_{_{1,n-1,i}}&= \delta_{_{n,1}}\left(P_i-\text{Re}\left\{y_i\right\}\right)-&{}\\
&\quad\text{Re}\left\{\!\sum\limits_{m=1}^{n-1}\!V_i^*[m]\!{\sum_{k\in{\mathcal}{B}}}\limits \!\!Y_{ik}V_k[n\!-\!m]\!\!\right\},&&i\in{\mathcal}{B}\\
\textrm{r}_{PQ_{2,n-1,i}}&= \delta_{_{n,1}}\left(-Q_i-\text{Im}\left\{y_i\right\}\right)-&{}\\
&\quad\text{Im}\left\{\!\sum\limits_{m=1}^{n-1}\!V_i^*[m]\!{\sum_{k\in{\mathcal}{B}}}\limits \!\!Y_{ik}V_k[n\!-\!m]\!\!\right\},&&i\in{\mathcal}{B}_{PQ}\\
\textrm{r}_{PV_{2,n-1,i}}&= -\!\!\sum\limits_{m=1}^{n-1}V_i^*[m]V_i[n-m]+L_i^2[n], &&i\in{\mathcal}{B}_{PV}\end{aligned}$$
and we can once again solve for $V$ to any desired value of $n$.
Having shown how the new models are able to create the required power series for the bus voltages, the next section will show how successful each model was at solving IEEE test cases.
Computational results {#sec:results}
=====================
The models from Section \[sec:PVmodels\] were created in Matlab and run on a series of seven standard IEEE test cases (9-, 14-, 30-, 39-, 57-, 118-, and 300-Bus) to obtain power series for the voltages. These power series were then run through the Viskovatov Padé approximant algorithm, as this is also the algorithm used in [@helm] and [@subramanian]. The resulting voltage values were then checked to see how well they solved the initial BPEE (\[eq:bpee\]) as well as how closely they resemble the solution obtained through the power flow function in MatPower 4.1, used here as an example of a traditional method of solving a load flow problem.
In Table \[tab:res\], Max $|Rs|$ refers to the maximum absolute residual obtained from the $(N\!-\!1)$ equations when the Padé approximant solutions are substituted back into (\[eq:bpee\]). Max $|\Delta|$ refers to the maximum absolute difference between the voltages obtained from the models and those obtained from MatPower. In each case the \[15/15\] Padé approximant is used to determine the bus voltage values. In all cases double precision was used. The times given in the table are the average of 100 runs. All the models, including MatPower, take roughly the same amount of time for the small systems.
System Model Max |$R_s$| Max $|\Delta|$ Time (s)
-------- ---------- ------------- ---------------- ----------
Subr. 1 2.1932e-02 4.7824e-03 0.0110
Subr. 2 1.2074e-11 1.4392e-12 0.0064
1 1.8938e-11 1.8491e-12 0.0059
2 1.8948e-11 1.8494e-12 0.0059
3 8.5691e-11 1.1914e-11 0.0073
4 4.4744e-12 6.1133e-13 0.0049
MatPower 5.9274e-14 0.0171
Subr. 1 7.4505e-03 1.3738e-03 0.0111
Subr. 2 2.3921e-14 5.8243e-12 0.0052
1 6.9950e-14 5.8213e-12 0.0062
2 1.1199e-12 5.6706e-12 0.0053
3 2.4484e-14 5.8254e-12 0.0049
4 2.4461e-14 5.8235e-12 0.0050
MatPower 1.3180e-10 0.0087
Subr. 1 1.6959e-02 1.5330e-03 0.0147
Subr. 2 2.2901e-14 1.9658e-10 0.0062
1 4.8474e-14 1.9658e-10 0.0068
2 2.3742e-14 1.9658e-10 0.0062
3 4.6547e-14 1.9658e-10 0.0059
4 6.0382e-14 1.9658e-10 0.0060
MatPower 9.7323e-10 0.0091
Subr. 1 1.2063e+01 8.4773e-01 0.0198
Subr. 2 2.2529e-01 2.5868e-02 0.0076
1 7.5746e-07 4.0769e-08 0.0085
2 5.0394e-06 2.4831e-07 0.0075
3 3.0385e-06 9.7468e-08 0.0069
4 1.1003e-09 5.2491e-11 0.0067
MatPower 8.2567e-13 0.0081
Subr. 1 5.8562e+00 1.8105e-01 0.0181
Subr. 2 4.7931e-13 2.2841e-13 0.0084
1 4.0425e-10 8.4430e-11 0.0092
2 2.7227e-09 1.0797e-10 0.0092
3 2.4653e-10 1.2281e-10 0.0088
4 4.8125e-10 2.7309e-10 0.0089
MatPower 3.7036e-12 0.0094
Subr. 1 2.2228e+02 2.4416e+00 0.0713
Subr. 2 8.1703e-11 4.3043e-12 0.0232
1 1.3121e-04 3.1500e-06 0.0192
2 2.0368e-04 1.6476e-05 0.0298
3 2.6099e-02 2.0917e-04 0.0220
4 1.6917e-10 7.6155e-12 0.0174
MatPower 1.4892e-12 0.0111
Subr. 1 1.9832e+02 9.1111e+00 0.1348
Subr. 2 1.4294e-01 4.9324e-02 0.0807
1 7.8944e+01 2.7515e+00 0.0656
2 1.5028e+02 2.4886e+00 0.1225
3 5.7531e+04 2.7865e+01 0.0761
4 2.8486e-04 8.4840e-06 0.0656
MatPower 1.7628e-12 0.0229
: Results for IEEE test cases
\[tab:res\]
From Table \[tab:res\] we can see that the model derived from [@subramanian], Subr. 1, converges much more poorly than the rest. For the 9-Bus system, Subramanian’s first model required a \[60/60\] Padé approximant to achieve a similar level of accuracy to what the other models achieve with the \[15/15\] Padé approximant. The other models perform relatively equally on all systems up to the 300-Bus system except that Subramanian’s second model has difficulty with the 39-Bus system. All of the models have trouble with the 300-Bus system and fail to provide correct voltages except Model 4, which manages to converge slowly to a proper solution.
Below we will further investigate the behaviour for the 300-bus system. Figures \[fig:mod4vc\] and \[fig:mod5vc\] show the power series coefficients obtained in the 300-Bus network using Models 1 and 4 respectively, while Figures \[fig:mod4v1\] and \[fig:mod4v2\] show the corresponding singularities (zeros of the denominator polynomial) of the \[50/50\] Padé approximant. The data for the figures was produced using Maple to 100 significant digits.
We can see from Figure \[fig:mod4v1\] that the Bus 8 Padé approximant derived using Model 1 appears to have a set of poles along the real line between $z=0.1$ and $z=0.2$, and going off the real axis into the upper and lower half-planes, with yet another set around $z=0.8$, indicating cuts in the complex voltage function. This is consistent with a small radius of convergence of the corresponding power series as can be seen from its rapidly increasing coefficients (Figure \[fig:mod4vc\]). From Table \[tab:res\] we see that these singularities adversely affect the rate of convergence of the Padé approximants, and using double precision the model is incapable of providing a sufficiently accurate value at $z=1$. Further investigation was conducted on the 300-Bus network using Maple, where precision can be set to higher levels than double precision. By setting the precision to 200 digits, much higher order of Padé approximants can be computed to greater accuracy, and in this setup convergence to the correct value at $z=1$ is obtained.
On the other hand, Figure \[fig:mod4v2\], shows for Model 4 there are no singularities in the disk centred at $z=0$ of radius 1. The single pole closest to the origin below the real line is a spurious pole and does not appear neighbouring Padé Approximants. Indeed the radius of convergence of the power series is greater than 1 as is indicated by the decreasing nature of its coefficients (Figure \[fig:mod5vc\]).
Case 39PQ with HELM
-------------------
It is not possible to test the original HELM model from [@helm] on examples containing PV buses. Practical examples, including all of the IEEE test cases used by our general models, have PV buses. One workaround to the lack of PQ-only networks is to convert the PV buses in the IEEE test cases into PQ buses. This is accomplished by finding the reactive power output of the PV buses at the power flow solution to the networks at full-load (as obtained by MatPower) and then fixing this value as a parameter at the bus while also now allowing voltage magnitude to vary (as in a PQ bus). The test-case solution should again be a solution to this altered network, and indeed, running these altered IEEE test cases through MatPower results in the same power flows and voltages as the original cases.
Using the HELM model on these modified results in the same solution as MatPower in all cases except the 39-Bus IEEE test case. In this case, the HELM model converges to a solution at $z=1$ which does indeed solve the BPEE correctly, but which however is different from the solution obtained by MatPower. The MatPower solution has voltages magnitudes close to 1 p.u. whereas the voltages magnitudes in the HELM solution are quite high (1.4–1.8 p.u.)
We have attempted to follow both solutions as the load changes between full load ($z=1$) and no load ($z=0$). Figure \[fig:39PQHelm\] shows the profile of the voltages at each bus as HELM moves from the initial no-load case to the full-load case. We have attempted to use HELM to trace the MatPower solution back from $z=1$ to $z=0$ but have been unable to do so: there seems to be a singularity around $z=0.5$ which prevents HELM obtaining solutions for real values of $z$ less than $0.5$ due to precision problems. Instead we have followed the MatPower solution back to the no-load case by using a traditional real homotopy on the non-embedded system (\[eq:bpeez\]). The resulting bus voltage profiles as the load factor changes are shown in Figure \[fig:Matpower\].
By comparing the two figures we see that the HELM solution starts with $V_i(0)\ne 0, \forall i$, whereas for the MatPower solution for one of the buses (Bus 38) $V_i(0)=0$.
The apparent inconsistencies can be reconciled by looking at the stability of each solution. Using the $\frac{d\Delta Q}{dV}$ stability criteria [@powersystems], the MatPower solution—which is stable for the original IEEE test case—becomes unstable in the altered, PQ-only network. Indeed, if more reactive power is demanded at any PQ bus except Bus 32, then the voltage magnitude at that bus increases rather than decreases. The higher-voltage HELM solution, by contrast, is stable, giving credence to the claim that HELM will return a stable solution.
[^1]: I. Wallace, A. Grothey, and K. I. M. McKinnon are with the School of Mathematics, University of Edinburgh, James Clerk Maxwell Building, Edinburgh EH9 3JZ, UK (e-mail: `[email protected]`, `[email protected]`, `[email protected]`).
[^2]: To be precise: The unique seed, or reference, solution at $z=0$ under the condition $V_i(0)\ne 0$ can be continued into a neighbourhood by the implicit function theorem. This continuation is holomorphic at $z=0$.
[^3]: though for simplicity he does not model the shunt conductance term for PV buses in his paper
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We define a tower of affine Temperley-Lieb algebras of type $\tilde{A_{n}}$ and we define Markov elements in those algebras. We prove that any trace over an affine Temperley-Lieb algebras of type $\tilde{A_{2}}$ is uniquely defined by its values on the Markov elements.'
address: 'LAMFA, Université de Picardie - Jules Verne'
author:
- Sadek AL HARBAT
title: 'Markov elements in affine Temperley-Lieb algebras'
---
Introduction
============
In [@Sadek_Thesis] we define a tower $( \widehat{TL}_{n+1}(q))_{n\ge 0} $ of affine Temperley-Lieb algebras of type $\tilde{A_{n}}$ and we prove that there exists a unique Markov trace on this tower. Crucial in the proof is the definition of [*Markov elements*]{} and the following Theorem :
Any trace over $\widehat{TL}_{n+1} (q)$ for $2 \leq n$ is uniquely defined by its values on the Markov elements in $\widehat{TL}_{n+1} (q)$.\
The proof of this Theorem for $3 \leq n$ is given in [@Sadek_2013_1], where we have omitted the case $n=2$, long and technical. We thus present it here for completeness.
Notations
=========
Let $K$ be an integral domain of characteristic $0$. Suppose that $q$ is a square invertible element in $K$ of which we fix a root $\sqrt{q}$. For $x,y$ in a given ring we define $V(x,y):= xyx+xy+yx+x+y+1$. We mean by algebra in what follows $K$-algebra.\
We denote by $B(\tilde{A_{n}})$ (resp. $W(\tilde{A_{n}})$) the affine braid (resp. affine Coxeter) group with $n+1$ generators of type $\tilde{A}$, while we denote by $B(A_{n})$ (resp. $W(A_{n})$) the braid (resp. Coxeter) group with $n$ generators of type $A$, where $n \geq 0 $. Let $W^{c}(\tilde{A_{n}})$ (resp. $W^{c}(A_{n}))$ be the set of fully commutative elements in $W(\tilde{A_{n}})$ (resp. $W(A_{n}))$.\
Let $ n\geq 2$. We define $\widehat{TL}_{n+1} (q)$ to be the algebra with unit given by a set of generators $\left\{g_{\sigma_{1}}, ...,~ g_{\sigma_{n}}, g_{a_{n+1}}\right\}$, with the following relations [@Graham_Lehrer_1998]:\
- $g_{\sigma_{i}} g_{\sigma_{j}} =g_{\sigma_{j}} g_{\sigma_{i}} $, for $1\leq i,j\leq n$ and $ \left| i-j\right| \geq 2$.
- $g_{\sigma_{i}} g_{a_{n+1}} =g_{a_{n+1}} g_{\sigma_{i}} $, for $2\leq i \leq n-1$.
- $g_{\sigma_{i}}g_{\sigma_{i+1}}g_{\sigma_{i}} = g_{\sigma_{i+1}}g_{\sigma_{i}}g_{\sigma_{i+1}}$, for $1\leq i\leq n-1$.
- $g_{\sigma_{i}}g_{a_{n+1}}g_{\sigma_{i}} = g_{a_{n+1}}g_{\sigma_{i}}g_{a_{n+1}}$, for $i= 1, n $.
- $g^{2}_{\sigma_{i}} = (q-1)g_{\sigma_{i}} +q$, for $1\leq i\leq n$.
- $g^{2}_{a_{n+1}} = (q-1)g_{a_{n+1}} +q$,
- $V(g_{\sigma_{i}},g_{\sigma_{i+1}})=V(g_{\sigma_{1}},g_{a_{n+1}}) = V(g_{\sigma_{n}},g_{a_{n+1}})= 0$, for $1\leq i\leq n-1$.\
The set $\left\{ g_{w}: w \in W^{c}(\tilde{A_{n}})\right\}$ is well defined in the usual sense of the theory of Hecke algebra and it is a $K$-basis. We set $T_{a_{n+1}}$ (resp. $T_{\sigma_{i}}$ for $1 \leq i \leq n$) to be $\sqrt{q}g_{a_{n+1}}$ (resp. $\sqrt{q}g_{\sigma_{i}}$ for $1 \leq i \leq n$). Hence, $T_{w}$ is well defined for $w \in W^{c}(\tilde{A_{n}})$, it equals $q^{\frac{l(w)}{2}}g_{w}$. The multiplication associated to the basis $\left\{ T_{w}: w \in W^{c}(\tilde{A_{n}})\right\}$, is given as follows:
$$\begin{aligned}
T_{w} T_{v} &=& T_{wv} ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \text{whenever } l(wv) = l(w) + l (v).\nonumber\\
T_{s} T_{w} &=& \sqrt{q}(q-1) T_{w} +q^{2} T_{sw} ~~~~~~~~ \text{whenever } l(sw) = l(w) - 1, \nonumber
\end{aligned}$$
for $w,v$ in $W^{c}(\tilde{A_{n}})$ and $s$ in $\left\{\sigma_{1},...,~\sigma_{n}, a_{n+1}\right\}$.\
In what follows we suppose that $q+1$ is invertible in $K$, we set $\delta = \frac{1}{2+q+q^{-1}} = \frac{q}{(1+q)^{2}}$ in $K$. In view of [@Graham_Lehrer_2003] , for $1 \leq i \leq n$ we set $ f_{\sigma_{i}}:= \frac{ g_{\sigma_{i}}+1}{q+1}$ and $ f_{a_{n+1}}:= \frac{g_{a_{n+1}}+1}{q+1}$. In other terms $ g_{\sigma_{i}} = (q+1) f_{\sigma_{i}} -1$, and $ g_{a_{n+1}} = (q+1) f_{a_{n+1}} -1$. The set $\left\{ f_{w}: w \in W^{c}(\tilde{A_{n}})\right\}$ is well defined and it is a $K$-basis for $\widehat{TL}_{n+1} (q)$.\
We define the Temperley-Lieb algebra of type $A$ with $n$ generators $TL_{n}(q)$, as the subalgebra of $\widehat{TL}_{n+1} (q)$ generated by $\left\{g_{\sigma_{1}} ,...,~ g_{\sigma_{n}}\right\}$, with $\left\{ g_{w}: w \in W^{c}(A_{n})\right\}$ as $K$-basis.\
Now for $TL_{0}(q) = K$, we consider the following tower:
$$\begin{aligned}
TL_{0}(q)~\subset TL_{1}(q) ~~...\subset TL_{n-1}(q) ~~\subset TL_{n}(q) ~~... \nonumber\\\nonumber
\end{aligned}$$
\[1\_1\][@Jones_1985] There is a unique collection of traces $(\tau_{n+1})_{0 \leq n}$ on $(TL_{n})_{0\leq n}$, such that:\
- $\tau_{1}(1) = 1 $.
- For $ 1 \leq n$, we have $\tau_{n+1}(hT^{\pm1}_{\sigma_{n}}) = \tau_{n}(h) $, for any $h$ in $TL_{n-1}(q)$.
The collection $(\tau_{n+1})_{0 \leq n}$ is called a Markov trace. Moreover, for any $a,b$ and $c$ in $TL_{n}(q)$ and for $ n \geq 1$, every $\tau_{n+1}: TL_{n}(q) \longrightarrow K$ verifies: $$\begin{aligned}
\tau_{n+1}(bT_{\sigma_{n}}c)= \tau_{n}(bc)$ and $\tau_{n+1}(a)=- \frac{1+q}{\sqrt{q}}\tau_{n}(a). \nonumber
\end{aligned}$$
The tower of affine Temperley-Lieb algebras and affine Markov trace
===================================================================
In this section we define a tower of affine Temperley-Lieb algebras, we show that this tower “surjects” onto the tower of Temperley-Lieb algebras mentioned in the introduction, and we define the affine Markov trace.\
We consider the Dynkin diagram of the group $B(\tilde{A_{n}})$. We denote the Dynkin automorphism $(\sigma_{1}\mapsto \sigma_{2} \mapsto ... \sigma_{n} \mapsto a_{n+1} \mapsto \sigma_{1})$ by $\psi_{n+1}$. Notice that $ \sigma_{n} \sigma_{n-1}.. \sigma_{1}a_{n+1} $ acts on $ B(\tilde{A_{n-1}}) $ as $\psi_{n}$ as follows ($\sigma_{1} \mapsto a_{n} \mapsto \sigma_{n-1} \mapsto \sigma_{n-2} \mapsto.. \sigma_{2} \mapsto \sigma_{1}$). We write $(\sigma_{n} .. \sigma_{1} a_{n+1})^{d}h = \psi^{d} \left[ h\right] (\sigma_{n} .. \sigma_{1} a_{n+1})^{d} $, for any $h$ in $ B(\tilde{A_{n-1}}) $, we keep same convention for the affine Temperley-Lieb algebra.\
at (0,0.5) [$\sigma_{1}$]{}; (0,0) circle (2pt);
(0,0) – (1.5, 0);
at (1.5,0.5) [$\sigma_{2}$]{}; (1.5,0) circle (2pt);
(1.5,0) – (5.5, 0);
at (5.5,0.5) [$\sigma_{n-1}$]{}; (5.5,0) circle (2pt);
(5.5,0) – (7, 0);
at (7,0.5) [$\sigma_{n}$]{}; (7,0) circle (2pt);
(7,0) – (3, -3);
(3, -3) circle (2pt); at (3, -3.5) [$a_{n+1}$]{};
(3, -3) – (0, 0);
We have the following injection
$$\begin{aligned}
G_{n}: K[B(\tilde{A_{n-1}})] &\longrightarrow& K[B(\tilde{A_{n}})] \nonumber\\
\sigma_{i} &\longmapsto& \sigma_{i}$ ~~~ \text{for} $1\leq i\leq n-1 \nonumber\\
a_{n} &\longmapsto& \sigma_{n} a_{n+1}\sigma^{-1}_{n} \nonumber
\end{aligned}$$
We prove in [@Sadek_Thesis], to which we refer for details, the following two propositions:\
\[2\_1\] The injection $G_{n}$ induces the following morphism of algebras:
$$\begin{aligned}
F_{n}: \widehat{TL}_{n}(q) &\longrightarrow& \widehat{TL}_{n+1}(q) \nonumber\\
t_{\sigma_{i}} &\longmapsto & g_{\sigma_{i}} \text{ for } 1 \leq i \leq n-1 \nonumber\\
t_{a_{n}} &\longmapsto & g_{\sigma_{n}} g_{a_{n+1}} g^{-1}_{\sigma_{n}}. \nonumber
\end{aligned}$$
The following map is a surjection of algebras
$$\begin{aligned}
E_{n}: \widehat{TL}_{n+1}(q) &\longrightarrow& TL_{n}(q) \nonumber\\
g_{\sigma_{i}} &\longmapsto & g_{\sigma_{i}} \text{ for } 1 \leq i \leq n \nonumber\\
g_{a_{n+1}} &\longmapsto & g_{\sigma_{1}} ... g_{\sigma_{n-1}} g_{\sigma_{n}} g^{-1}_{\sigma_{n-1}} ... g^{-1}_{\sigma_{1}}. \nonumber
\end{aligned}$$
Moreover, the following diagram commutes:
;
\(C) edge\[-myhook,line width=0.42pt\] node\[above, xshift=-5mm, yshift=-2mm, rotate=0\] [$E_{n-1}$]{} (A); (A) edge\[-myonto,line width=0.42pt\] node\[above, xshift=-5mm, yshift=-2mm, rotate=0\] [$ $]{} (C);
\(A) edge\[-myto,line width=0.42pt\] node\[above, xshift=1.5mm, yshift=0mm, rotate=0\] [$F_{n}$]{} (B);
\(D) edge\[-myhook,line width=0.42pt\] node\[below, xshift=1.5mm, yshift=0mm, rotate=0\] [$$]{} (C); (C) edge\[-myto,line width=0.42pt\] (D);
\(B) edge\[-myonto,line width=0.42pt\] node\[above, xshift=5mm, , yshift=-2mm, rotate=0\] [$E_{n} $]{} (D);
Moreover, it is immediate that $E_{n}$ composed with the natural inclusion of $TL_{n}(q)$ into $\widehat{TL}_{n+1}(q)$, gives $Id_{TL_{n}(q)}$.
In view of proposition \[2\_1\] we can consider the tower of affine T-L algebras (it is not known whether it is a tower of faithful arrows or not):\
$$\begin{aligned}
\widehat{TL}_{1}(q) \stackrel{F_{1}}{\longrightarrow} \widehat{TL}_{2}(q) \stackrel{F_{2}} {\longrightarrow}\widehat{TL}_{3}(q) \longrightarrow ~~...~~ \widehat{TL}_{n}(q) \stackrel{F_{n}} {\longrightarrow}\widehat{TL}_{n+1}(q) \longrightarrow ... \nonumber\\\nonumber
\end{aligned}$$
\[5\_2\_1\] We call $(\hat{\tau}_{n})_{1 \leq n}$ an affine Markov trace, if every $\hat{\tau}_{n}$ is a trace function on $\widehat{TL}_{n} (q)$ with the following conditions:\
- $\hat{\tau}_{1}(1) = 1$, (here $\widehat{TL}_{1} (q) = K$).
- $\hat{\tau}_{n+1}(F_{n}(h)T^{\pm1}_{\sigma_{n}}) = \hat{\tau}_{n}(h)$, for all $h \in \widehat{TL}_{n} (q)$ and for $n \geq 1$.
- $\hat{\tau}_{n}$ is invariant under the Dynkin automorphism $\psi_{n}$ for all $n$.\
We notice that the second condition gives us that $\hat{\tau}_{n+1}\big(F_{n}(h)T^{-1}_{\sigma_{n}}\big) = \hat{\tau}_{n}\big(h\big)$, which means that: $$\begin{aligned}
\hat{\tau}_{n+1}\big(F_{n}(h)[\frac{1}{q^{2}}T_{\sigma_{n}}- \frac{q-1}{q\sqrt{q}}] \big) = \hat{\tau}_{n}\big(h). \text{ Thus } \hat{\tau}_{n+1} \big(F_{n}(h)\big) = -\frac{q+1}{\sqrt{q}} \hat{\tau}_{n}\big(h\big). \nonumber\\\nonumber
\end{aligned}$$
The third condition of definition \[5\_2\_1\] is, in fact, not independent, i.e., it results from the first and second conditions (see [@Sadek_2013_1]). Nevertheless, we will keep viewing it as a condition.\
This affine Markov trace does the job topologically, i.e., it gives an invariant for “affine oriented knots” and generalizes, in fact, the Jones invariant, noticing that the set of oriented knots in $S^{3}$ injects naturally into the set of “affine oriented knots”. For further details see [@Sadek_Thesis].
Now, consider the following commutative diagram:\
;
\(A) edge\[-myto,line width=0.42pt\] (B);
\(B) edge\[-myto,line width=0.42pt\] (BB);
(BB) edge\[-myto,line width=0.42pt\] (C);
\(C) edge\[-myto,line width=0.42pt\] (D);
\(F) edge\[-myhook,line width=0.42pt\] (E); (E) edge\[-myto,line width=0.42pt\] (F);
(FF) edge\[-myhook,line width=0.42pt\] (F); (F) edge\[-myto,line width=0.42pt\] (FF);
\(G) edge\[-myhook,line width=0.42pt\] (FF); (FF) edge\[-myto,line width=0.42pt\] (G);
\(H) edge\[-myhook,line width=0.42pt\] (G); (G) edge\[-myto,line width=0.42pt\] (H);
\(A) edge\[-myonto,line width=0.42pt\] (E);
\(B) edge\[-myonto,line width=0.42pt\] (F);
\(C) edge\[-myonto,line width=0.42pt\] (G);
\(D) edge\[-myonto,line width=0.42pt\] (H);
\(E) edge\[-myto,line width=0.42pt\] node\[above, xshift=-5mm, yshift=-2mm, rotate=0\] [$\tau_{1}$]{} (I);
\(F) edge\[-myto,line width=0.42pt\] node\[above, xshift=-3mm, yshift=-2mm, rotate=0\] [$\tau_{2}$]{} (I);
\(G) edge\[-myto,line width=0.42pt\] node\[above, xshift=3mm, yshift=-2mm, rotate=0\] [$\tau_{n}$]{} (I);
\(H) edge\[-myto,line width=0.42pt\] node\[above, xshift=8mm, yshift=-2mm, rotate=0\] [$\tau_{n+1}$]{} (I);
Set $\rho_{n+1}$ to be the trace over $\widehat{TL}_{n+1} (q) $ induced by $\tau_{n+1}$ over $TL_{n} (q) $ for $0 \leq n$. We prove in [@Sadek_2013_1] that $(\rho_{i})_{1 \leq i}$ is an affine Markov trace over $\big( \widehat{TL}_{i} (q)\big)_{1 \leq i}$ and we prove the following Theorem:
\[5\_2\_8\] [@Sadek_2013_1]
There exists a unique affine Markov trace over the tower of $\tilde{A}$-type Temperley-Lieb algebras, namely $(\rho_{i})_{1 \leq i}$.
The proof relies on Theorem \[5\_1\_1\] below, the proof of which separates into two cases: $n=2$ and $n \ge 3$. The latter case is included in [@Sadek_2013_1] while the former appears in the present note.
Markov elements and traces on $\widehat{TL}_{n+1}(q)$
=====================================================
Markov elements
---------------
We consider $F_{n}:\widehat{TL}_{n} (q)\longrightarrow \widehat{TL}_{n+1} (q)$ of proposition \[2\_1\]. In this subsection we set $F:= F_{n}$. We give a definition of Markov elements in $\widehat{TL}_{n+1} (q)$ for $ 2 \leq n $. Then we show that any trace over $\widehat{TL}_{n+1} (q)$ is uniquely determined by its values on those elements.
For $F$ as above, and $n \geq 2$, a Markov element in $\widehat{TL}_{n+1} (q)$ is any element of the form $ A g^{\epsilon}_{\sigma_{n}} B $, where $A$ and $B$ are in $F(\widehat{TL}_{n} (q))$ and $\epsilon \in \left\{ 0,1 \right\}$.
The aim of this subsection is to prove the following theorem for $n=2$.
[@Sadek_2013_1] \[5\_1\_1\] Let $\tau_{n+1}$ be any trace over $\widehat{TL}_{n+1} (q)$ for $2 \leq n$. Then, $\tau_{n+1}$ is uniquely defined by its values on the Markov elements in $\widehat{TL}_{n+1} (q)$.\
The proof of theorem \[5\_1\_1\] for $n=2$ is divided into two parts. In the first we show some general facts, in the second we give the proof for $ n=2$.\
**Part 1**\
In this part, we suppose that $\tau_{n+1}$ is any trace on $\widehat{TL}_{n+1} (q)$. We will apply $\tau_{n+1}$ to $\widehat{TL}_{n+1} (q)$ assuming that $2 \leq n$, and show that $\tau_{n+1}$ is uniquely determined on $\widehat{TL}_{n+1} (q)$ by its values on the positive powers of $g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}$, in addition to its values on Markov elements. From now on we denote by $w$: an arbitrary element in $W^{c}(\tilde{A_{n}})$.
\[5\_1\_3\] In $\widehat{TL}_{n+1} (q) $ we have:\
$$\begin{aligned}
(1)~g_{\sigma_{n}} (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k} &=& (q-1) (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k} + \sum\limits^{i=k-1}_{i=1} f_{i} (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{i} \nonumber \\
& & + A \big(g_{\sigma_{n-1} \sigma_{n-2} ..\sigma_{1}}F(t_{a_{n}})\big)^{k} g_{\sigma_{n}}\prod^{j=k-1}_{j=0} \psi^{j} \big[F((t_{a_{n}})^{-1})\big], \nonumber
\end{aligned}$$
$$\begin{aligned}
(2)~(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k}g_{\sigma_{n}} &=& (q-1) (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k} + \sum^{i=k-1}_{i=1} h_{i} (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{i}\nonumber \\
& & + A \prod^{j=k-1}_{j=0} \phi^{j} \big[(g_{\sigma_{{n-1}}})^{-1}\big] g_{\sigma_{n}}\big(g_{\sigma_{n-1} \sigma_{n-2} ..\sigma_{1}}F(t_{a_{n}})\big)^{k}, \nonumber
\end{aligned}$$
with $ A $ in the ground field, $ f_{i},h_{i} $ in $F (\widehat{TL}_{n} (q)) $ and $ \phi^{-1} = \psi $.
$$\begin{aligned}
g_{\sigma_{n}} \big(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}\big)^{k} &=& \big(q-1\big) \big(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}\big)^{k} \nonumber\\
& & + q g_{\sigma_{n-1} \sigma_{n-2} ..\sigma_{1}}F\big(t_{a_{n}}\big) g_{\sigma_{n}} F\big((t_{a_{n}})^{-1}\big) \big(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}\big)^{k-1} \nonumber\\\nonumber\\
&=& \big(q-1\big) \big(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}\big)^{k} + \nonumber\\
& & q g_{\sigma_{n-1} \sigma_{n-2} ..\sigma_{1}}F\big(t_{a_{n}}\big) g_{\sigma_{n}} \big(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}\big)^{k-1} \psi^{k-1} \big[F((t_{a_{n}})^{-1})\big]. \nonumber\\\nonumber
\end{aligned}$$
So, by induction on $k$, (1) follows. In the very same way we deal with (2), by noticing that: $g_{a_{n+1}}g_{\sigma_{n}} =g^{-1}_{\sigma_{n}} F(t_{a_{n}}) g^{2}_{\sigma_{n}} = (q-1) g_{a_{n+1}} + qg^{-1} _{\sigma_{n}} F(t_{a_{n}}) $.
A main result in [@Sadek_2013_2] is to give a general form for “fully commutative braids”, from which we deduce that any element of the basis of $\widehat{TL}_{n+1} (q)$ (where we have the convention $\sigma_{n+1} = 1$ in $W(\tilde{A_{n}})$ thus $ g_{\sigma_{n} \sigma_{n-1} .. \sigma_{i}} =1$ when $i=n+1$), is either of the form\
$$c (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k} g_{\sigma_{n}\sigma_{n-1} .. \sigma_{i}}$$
or of the form\
$$g_{\sigma_{i_{0}} .. \sigma_{2}\sigma_{1}a_{n+1}} (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k} d g_{\sigma_{n} \sigma_{n-1} .. \sigma_{i}}$$\
where $c$ and $d$ are in $F (\widehat{TL}_{n} (q)) $, $ 1\leq i \leq n+1 $ and $ 0 \leq i_{0} \leq n-1 $ .\
By lemma \[5\_1\_3\] $ c (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k} g_{\sigma_{n}\sigma_{n-1} .. \sigma_{i}} $ is of the form: $$\begin{aligned}
\sum^{j=h}_{j=1}c_{j} (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{j} + M. \nonumber
\end{aligned}$$
Where $h \leq k $, $c_{j}$ is in $F (\widehat{TL}_{n} (q)) $ for any $ j $ and $M$ is a Markov element.\
Now we deal with the second form: $$\begin{aligned}
\tau_{n+1} &\big(&g_{\sigma_{i_{0}} .. \sigma_{2}\sigma_{1}a_{n+1}} c (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k} g_{\sigma_{n} \sigma_{n-1} .. \sigma_{i}}\big) = \tau_{n+1} \big(g_{\sigma_{n} \sigma_{n-1} .. \sigma_{i}} g_{\sigma_{i_{0}} .. \sigma_{2}\sigma_{1}a_{n+1}} c (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k}\big). \nonumber
\end{aligned}$$
For any possible value for $ i_{0} $ or $i$, we see that: $$\begin{aligned}
g_{\sigma_{n} \sigma_{n-1} .. \sigma_{i}} g_{\sigma_{i_{0}} .. \sigma_{2}\sigma_{1}a_{n+1}} c (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k} = c'g_{\sigma_{n}} (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{s} c'' ,\nonumber
\end{aligned}$$
where $ c',c''$ are in $F (\widehat{TL}_{n} (q)) $ and $ s \leq k+1 $. By lemma \[5\_1\_3\] we see that this element is of the form: $$\begin{aligned}
\sum^{j=h}_{j=1}f_{j} (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{j} + M, \nonumber
\end{aligned}$$
where $h \leq k+1 $, $f_{j}$ is in $F (\widehat{TL}_{n} (q)) $ for any $ j $ and $M$ is a Markov element .\
Hence, we see that in order to define $\tau_{n+1}$ uniquely it is enough to have its values on Markov elements and its values on $\Omega (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k}$, where $1 \leq k$ (since if $k$ is equal to 0 then we are again in the case of a Markov element) and $\Omega$ is in $F\big(\widehat{TL}_{n} (q)\big)$.
\[5\_1\_4\] Let $2 \leq n $ then $\tau_{n+1}$ is uniquely defined by its values on Markov elements, in addition to its values on $ (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}} )^{k} $, with $0 \leq k $ .\
In order to determine $\tau_{n+1} \big( h(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k}\big)$, with a positive $k$ and an arbitrary $h$ in $F\big(\widehat{TL}_{n} (q)\big)$, it is enough to treat $\tau_{n+1} \big(F(t_{x})(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k}\big)$, with $x$ in $W^{c}(\tilde{A_{n-1}})$, but the fact that $\tau_{n+1}$ is a trace, in addition to the fact that $g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}$ acts as a Dynkin automorphism on $F \big(\widehat{TL}_{n} (q)\big)$, authorizes us to suppose that $x$ has a reduced expression which ends with $\sigma_{n-1}$.\
Now we show by induction on $l(x)$, that $\tau_{n+1} \big( F(t_{x}) (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k}\big)$ is a sum of values of $\tau_{n+1}$ over $ (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k}$, elements of the form $ h (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{i}$ with $i < k $ and Markov elements, (of course with coefficients in the ground ring which might be zeros).\
For $l(x) = 0$ the property is true. Take $l(x) > 0$, and let $ x = z \sigma_{n-1}$ be a reduced expression, hence: $$\begin{aligned}
\tau_{n+1} \big( F(t_{x}) (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k}\big) &=& \tau_{n+1}\big( F(t_{z}) F(t_{\sigma_{n-1}}) g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k-1}\big) \nonumber\\\nonumber\\
&=& \tau_{n+1} \big(F(t_{z}) \underbrace{g_{\sigma_{n-1}} g_{\sigma_{n}} g_{\sigma_{n-1}}}_{=-V(g_{\sigma_{n-1}}, g_{\sigma_{n}})} g_{ \sigma_{n-2} ..\sigma_{1} a_{n+1}}(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k-1}\big). \nonumber
\end{aligned}$$
Recalling that $V(g_{\sigma_{n-1}}, g_{\sigma_{n}})=0$, this is equal to the following sum: $$\begin{aligned}
& & - \tau_{n+1}\big( F(t_{z}) (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k}\big) \nonumber\\\nonumber\\
& & - \tau_{n+1}\big( F(t_{z}) g_{\sigma_{n-1}} g_{ \sigma_{n-2} ..\sigma_{1}} g_{a_{n+1}}(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k-1}\big)\nonumber\\\nonumber\\
& & - \tau_{n+1}\big( F(t_{z}) g_{ \sigma_{n-2} ..\sigma_{1} a_{n+1}}(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k-1}\big)\nonumber\\\nonumber\\
& & - \tau_{n+1}\big( F(t_{z}) g_{\sigma_{n-1}} g_{ \sigma_{n-2} ..\sigma_{1}}\underbrace{ g_{\sigma_{n}} g_{a_{n+1}}}_{}(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k-1}\big) \nonumber\\
& & - \tau_{n+1}\big( F(t_{z}) g_{ \sigma_{n-2} ..\sigma_{1}} \underbrace{g_{\sigma_{n}} g_{a_{n+1}}}_{}(g_{\sigma_{n} \sigma_{n-1} .. \sigma_{1} a_{n+1}})^{k-1}\big).\nonumber
\end{aligned}$$
Now we apply the induction hypothesis to the first term. The second and the third terms are equal to: $$\begin{aligned}
\tau_{n+1} &\bigg(& F\big(t_{z}\big) g_{\sigma_{n-1}} g_{ \sigma_{n-2} ..\sigma_{1}} F\big( t_{a_{n}}\big) g_{\sigma_{n}} F\big( (t_{a_{n}})^{-1}\big)\big(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}\big)^{k-1}\bigg)\nonumber\\\nonumber\\
&+& \tau_{n+1} \bigg( F\big(t_{z}\big) g_{\sigma_{n-2} ..\sigma_{1}} F\big( t_{a_{n}}\big) g_{\sigma_{n}}F\big( (t_{a_{n}})^{-1}\big)\big(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}\big)^{k-1}\bigg), \nonumber
\end{aligned}$$
which is equal to: $$\begin{aligned}
\tau_{n+1} &\bigg(& \psi^{1-k} \big[ F\big( (t_{a_{n}})^{-1}\big) \big]F\big(t_{z}\big) g_{\sigma_{n-1}} g_{ \sigma_{n-2} ..\sigma_{1}} F\big( t_{a_{n}}\big) \big(g_{\sigma_{n}} (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k-1}\big)\bigg)\nonumber\\\nonumber\\
&+& \tau_{n+1}\bigg(\psi^{1-k} \big[ F\big( (t_{a_{n}})^{-1}\big) \big] F\big(t_{z}\big) g_{ \sigma_{n-2} ..\sigma_{1}} F( t_{a_{n}})\big( g_{\sigma_{n}}(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k-1}\big)\bigg).\nonumber
\end{aligned}$$
The fourth and the fifth terms are equal to: $$\begin{aligned}
\tau_{n+1} &\bigg(& F\big(t_{z}\big) g_{\sigma_{n-1}} g_{ \sigma_{n-2} ..\sigma_{1}} F\big( t_{a_{n}}\big) \big(g_{\sigma_{n}}(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k-1}\big)\bigg) \nonumber\\\nonumber\\
&+& \tau_{n+1}\bigg( F\big(t_{z}\big) g_{\sigma_{n-2} ..\sigma_{1}} F\big(t_{a_{n}}\big) \big(g_{\sigma_{n}}(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k-1}\big)\bigg).\nonumber
\end{aligned}$$
Thus, lemma \[5\_1\_3\] tells us that the property is true for those four terms. This step is to be applied repeatedly, to the powers of $g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}} $ down to an element of the form $ \tau_{n+1}\big(h (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{1}\big)$, arriving to the sum of:
$$\tau_{n+1}( g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})$$ and $$\tau_{n+1} (h' g_{\sigma_{n-1} ..\sigma_{1} a_{n+1}}) ,$$\
which is the sum of values of $ \tau_{n+1}$ on Markov elements, since $h,h'\in F \big(\widehat{TL}_{n} (q)\big)$.
We end this part by the following lemma:
\[5\_1\_5\] Let $ 1\leq k $. Then $ (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k}$ is a sum of two kinds of elements:\
- $g_{\sigma_{n}} \big(g_{\sigma_{n-1} \sigma_{n-2} .. \sigma_{1}} F(t_{a_{n}}) \big)^{j} g_{\sigma_{n}} h $, with $j\leq k $.
- $\big(g_{\sigma_{n-1} \sigma_{n-2} .. \sigma_{1}} F(t_{a_{n}})\big)^{i} g_{\sigma_{n}} f $, with $i < k $,\
with $h,f$ in $F \big(\widehat{TL}_{n} (q)\big)$ and $2\leq n$.\
Moreover, in the first type we have one, and only one element, with $j=k$, in which we have: $$\begin{aligned}
h = \prod^{i=k-1}_{i=0} \phi^{i}\big[F(t^{-1}_{a_{n}})\big]. \nonumber
\end{aligned}$$
Suppose that $k=1$. Then, $$\begin{aligned}
g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}} = g_{\sigma_{n}} \big(g_{\sigma_{n-1} \sigma_{n-2} .. \sigma_{1}} F(t_{a_{n}})\big) g_{\sigma_{n}} F\big(t_{a_{n}}\big)^{-1}, \nonumber
\end{aligned}$$
the property is true.\
Suppose the property is true for $k-1$, then, with $ 2\leq k$, we have: $$\begin{aligned}
(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k} = (g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k-1} g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}. \nonumber
\end{aligned}$$
We apply the property to $(g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}})^{k-1}$, which gives two cases:\
- $g_{\sigma_{n}} \big(g_{\sigma_{n-1} \sigma_{n-2} .. \sigma_{1}} F(t_{a_{n}}) \big)^{j'} g_{\sigma_{n}} h g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}$, with $ j' \leq k-1 $ which is: $g_{\sigma_{n}} \big(g_{\sigma_{n-1} \sigma_{n-2} .. \sigma_{1}} F(t_{a_{n}}) \big)^{j'} g_{\sigma_{n}} g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}} \psi^{-1}\big[h\big]$, which is equal to:\
$$\begin{aligned}
qg_{\sigma_{n}} &\big(& g_{\sigma_{n-1} \sigma_{n-2} .. \sigma_{1}} F(t_{a_{n}}) \big)^{j'+1} g_{\sigma_{n}} F\big( (t_{a_{n}})^{-1}\big) \psi^{-1}\big[h\big] \nonumber\\\nonumber\\
&+& (q-1) g_{\sigma_{n}} g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}\psi^{-1} \bigg[\big(g_{\sigma_{n-1} \sigma_{n-2} .. \sigma_{1}} F(t_{a_{n}}) \big)^{j'} \bigg] \psi^{-1}\left[h\right]. \nonumber
\end{aligned}$$
Since, $ j'+1 \leq k $, the first term is clear to be of the first type, while the second term is equal to: $$\begin{aligned}
& & \big(q-1\big)q g_{\sigma_{n-1} ..\sigma_{1}} F\big(t_{a_{n}}\big) g_{\sigma_{n}} F\big( (t_{a_{n}})^{-1}\big)\psi^{-1} \big[\big(g_{\sigma_{n-1} \sigma_{n-2} .. \sigma_{1}} F(t_{a_{n}}) \big)^{j'} \big] \psi^{-1}\big[h\big] + \nonumber\\\nonumber\\
& & \big(q-1\big)^{2} g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}\psi^{-1} \big[\big(g_{\sigma_{n-1} \sigma_{n-2} .. \sigma_{1}} F(t_{a_{n}}) \big)^{j'} \big] \psi^{-1}\big[h\big] .\nonumber
\end{aligned}$$
Here, the first term is of the second type (with $i=1 < k $), and the second term is of the first type (with $j=1$).\
- $\big(g_{\sigma_{n-1} \sigma_{n-2} .. \sigma_{1}} F(t_{a_{n}})\big)^{i'} g_{\sigma_{n}} f g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}}$, with $ i' < k-1 $, which is: $$\begin{aligned}
\big(&g&_{\sigma_{n-1} \sigma_{n-2} .. \sigma_{1}} F(t_{a_{n}}) \big)^{i'} g_{\sigma_{n}} g_{\sigma_{n} \sigma_{n-1} ..\sigma_{1} a_{n+1}} \psi^{-1} \big[f\big] = \nonumber\\\nonumber\\
& & q \big(g_{\sigma_{n-1} \sigma_{n-2} .. \sigma_{1}} F(t_{a_{n}})\big)^{i'+1} g_{\sigma_{n}} F\big((t_{a_{n}})^{-1}\big) \psi^{-1} \big[f\big] + \nonumber\\\nonumber\\
& & \big(q-1\big) g_{\sigma_{n}} \big(g_{\sigma_{n-1} ..\sigma_{1}} F(t_{a_{n}})\big) g_{\sigma_{n}} F\big((t_{a_{n}})^{-1}\big) \psi^{-1}\big[\big(g_{\sigma_{n-1} \sigma_{n-2} .. \sigma_{1}} F(t_{a_{n}}) \big)^{i'} \big] \psi^{-1} \big[f\big]. \nonumber\\\nonumber
\end{aligned}$$
Since $i'+1 <k $, the first term is of the second type, while the second term is of the first type with $j=1$. The lemma is proven.\
(By induction over $k$ again, the last formula is easy).
**Part 2**\
In this part we will consider a given trace $\tau_{3}$ over $\widehat{TL}_{3} (q)$. The aim is to show that $\tau_{3}$ is uniquely defined by its values on Markov elements. consider $$\begin{aligned}
F_{2}:\widehat{TL}_{2} (q) &\longrightarrow& \widehat{TL}_{3} (q)\nonumber\\
t_{\sigma_{1}} &\longmapsto& g_{\sigma_{1}}\nonumber\\
t_{a_{n}} &\longmapsto& g_{\sigma_{2}}g_{a_{3}} g^{-1}_{\sigma_{2}}. \nonumber
\end{aligned}$$
In this part we will denote $F_{2}$ by $F$.\
Lemma \[5\_1\_4\] tells that we can uniquely determine $\tau_{3}$ by its values over $(g_{\sigma_{2}\sigma_{1}a_{3}})^{k} $ for a positive $k$ beside its values on Markov elements. We know as well by lemma \[5\_1\_5\] that $(g_{\sigma_{2}\sigma_{1}a_{3}})^{k} $ is a sum of two kinds of elements:\
- $g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{j} g_{\sigma_{2}} h $ with $j\leq k $.
- $\big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{i} g_{\sigma_{2}} f $ with $i < k $.\
Here, $h$ and $f$ are in $F\big(\widehat{TL}_{2}(q)\big)$ .\
Moreover, in first type, only when $j=k$, we have: $$\begin{aligned}
h= \prod^{i=k-1}_{i=0} \psi^{i} \bigg[\big(F(t_{a_{2}})^{-1}\big)\bigg]. \nonumber
\end{aligned}$$
In other terms: $$\begin{aligned}
g_{\sigma_{2}} \bigg(g_{\sigma_{1}} F\big(t_{a_{2}}\big) \bigg)^{k} g_{\sigma_{2}} &=& \bigg(g_{\sigma_{2}\sigma_{1}a_{3}}\bigg)^{k}\prod^{i=k-1}_{i=0} \psi^{i} \bigg[\big(F(t_{a_{2}})\big)\bigg] \nonumber\\\nonumber\\
& & - \sum^{r=k-1}_{r=1} \bigg(g_{\sigma_{1}} F\big(t_{a_{2}}\big) \bigg)^{r} g_{\sigma_{2}} f_{r}\prod^{i=k-1}_{i=0} \psi^{i} \bigg[\big(F(t_{a_{2}})\big)\bigg] \nonumber\\\nonumber\\
& & + \sum^{l=k-1}_{l=1} g_{\sigma_{2}}\bigg(g_{\sigma_{1}} F\big(t_{a_{2}}\big) \bigg)^{l} g_{\sigma_{2}} f'_{l}\prod^{i=k-1}_{i=0} \psi^{i} \bigg[\big(F(t_{a_{2}})\big)\bigg]. \nonumber\\\nonumber
\end{aligned}$$
We repeat the same step on $g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{l} g_{\sigma_{2}}$ for every $l$. We deduce the following:\
\[5\_1\_6\]
For every $h > 0$, we have: $g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{h}g_{\sigma_{2}} = \sum\limits^{j=h}_{j=0} c_{j} \big(g_{\sigma_{2}\sigma_{1}a_{3}}\big)^{j} + \sum\limits_{i} M _{i}$.\
Here, $c_{j}$ is in $F\big(\widehat{TL}_{2}(q)\big)$ for every $j$, and $M_{i} $ is a Markov element for every $i$.
Our way to prove Theorem \[5\_1\_1\] for $n=3$, is to show that $\tau_{3} \big( (g_{\sigma_{2}\sigma_{1}a_{3}})^{k}\big)$ is a linear combination of values of $ \tau_{3} $ on Markov elements and values on elements of the form $c (g_{\sigma_{2}\sigma_{1}a_{3}})^{h}$, where $ h < k $ and $c$ in $F\big(\widehat{TL}_{2}(q)\big)$. Then, using the induction in the proof of Lemma \[5\_1\_4\], beside the fact that $\tau_{3} (g_{\sigma_{2}\sigma_{1}a_{3}}) $ is a linear combination of some values of $\tau_{3}$ on Markov elements, we see that the work is done.\
\[5\_1\_7\] Suppose that $r$ and $s$ are positives ,such that $r \leq s $. Then: $$\begin{aligned}
\tau_{3} \bigg(\big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r}g_{\sigma_{1}}g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s}g_{\sigma_{1}}g_{\sigma_{2}}\bigg) = \sum^{j=h}_{j=0} c_{j} \bigg(g_{\sigma_{2}\sigma_{1}a_{3}}\bigg)^{j} + \sum_{i} M _{i} ,\nonumber
\end{aligned}$$
where $h \leq s$, $c_{j}$ is in $F\big(\widehat{TL}_{2}(q)\big)$ for every $j$ and $M_{i} $ is a Markov element for every $i$.
$$\begin{aligned}
\tau_{3} \bigg(\big(g_{\sigma_{1}} &F& (t_{a_{2}})\big)^{r}g_{\sigma_{1}}g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s}g_{\sigma_{1}}g_{\sigma_{2}}\bigg) = \nonumber\\\nonumber\\
&=& \tau_{3} \bigg(\big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r}g_{\sigma_{1}}g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s}g_{\sigma_{1}} \underbrace{ g_{\sigma_{2} \sigma_{1}a_{3}} \big(g_{\sigma_{2}\sigma_{1}a_{3}}\big)^{-1}} g_{\sigma_{2}}\bigg) \nonumber\\\nonumber\\
&=& \tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r}g_{\sigma_{1}}g_{\sigma_{2}} g_{\sigma_{2}\sigma_{1}a_{3}} \psi \big[ \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s}g_{\sigma_{1}} \big]\big(g_{\sigma_{2}\sigma_{1}a_{3}}\big)^{-1} g_{\sigma_{2}}\bigg) \nonumber\\\nonumber\\
&=& \tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r}g_{\sigma_{1}}g_{\sigma_{2}} g_{\sigma_{2}\sigma_{1}a_{3}} \psi \big[\big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s}g_{\sigma_{1}} \big]g^{-1}_{a_{3}}g^{-1}_{\sigma_{1}}\bigg) \nonumber\\\nonumber\\
&=& \tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-1}g_{\sigma_{1}}g^{2}_{\sigma_{2}}g_{\sigma_{1}}g_{a_{3}} \big( F(t_{a_{2}}) g_{\sigma_{1}}\big)^{s}F\big(t_{a_{2}}\big) g^{-1}_{a_{3}} F\big(t_{a_{2}}\big)\bigg) \nonumber\\\nonumber\\
&=& \frac{1-q}{q} \tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-1}g_{\sigma_{1}}g^{2}_{\sigma_{2}}g_{\sigma_{1}}g_{a_{3}} \big( F(t_{a_{2}}) g_{\sigma_{1}}\big)^{s}F\big(t_{a_{2}}\big) F\big(t_{a_{2}}\big)\bigg) \nonumber\\\nonumber\\
&+& \frac{1}{q} \tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-1}g_{\sigma_{1}}g^{2}_{\sigma_{2}}g_{\sigma_{1}}g_{a_{3}} \big( F(t_{a_{2}}) g_{\sigma_{1}}\big)^{s}F\big(t_{a_{2}}\big) g_{a_{3}} F\big(t_{a_{2}}\big)\bigg) . \nonumber
\end{aligned}$$
Now, the term corresponding to $\frac{1-q}{q}$ is $\tau_{3}$ evaluated on the sum of Markov element and an element of style $ c_{j} (g_{\sigma_{2}\sigma_{1}a_{3}})^{1} $. So, We are reduced to the second term, thus, reduced to: $$\begin{aligned}
\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{r-1}g_{\sigma_{1}}g^{2}_{\sigma_{2}}g_{\sigma_{1}}g_{a_{3}} \big( F(t_{a_{2}}) g_{\sigma_{1}}\big)^{s}F \big(t_{a_{2}}\big) g_{a_{3}} F\big(t_{a_{2}}\big)\bigg). \nonumber
\end{aligned}$$
Obviously, we are in the case: $$\begin{aligned}
q \tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-1} g_{\sigma_{1}}g_{\sigma_{1}}&g&_{a_{3}} \big(F(t_{a_{2}}) g_{\sigma_{1}}\big)^{s}F\big(t_{a_{2}}\big) g_{a_{3}} F\big(t_{a_{2}}\big)\bigg) = \nonumber\\\nonumber\\
& & q \tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-1}g^{2}_{\sigma_{1}}g_{a_{3}} \big( F(t_{a_{2}}) g_{\sigma_{1}}\big)^{s} F\big(t^{2}_{a_{2}}\big)g_{\sigma_{2}} \bigg),\nonumber
\end{aligned}$$
since $g_{a_{3}} F\big(t_{a_{2}}\big)= F\big(t_{a_{2}}\big)g_{\sigma_{2}}$.\
Now, $$\begin{aligned}
\tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}})&\big)&^{r-1}g^{2}_{\sigma_{1}}g_{a_{3}} \big( F(t_{a_{2}}) g_{\sigma_{1}}\big)^{s} F\big(t^{2}_{a_{2}}\big)g_{\sigma_{2}} \bigg)= \nonumber\\\nonumber\\
& & \big(q-1\big)\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-1}g^{2}_{\sigma_{1}}g_{a_{3}} F\big(t_{a_{2}}\big) \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s} g_{\sigma_{2}} \bigg)\nonumber\\\nonumber\\
+& & q\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-1}g^{2}_{\sigma_{1}}g_{a_{3}} F(t_{a_{2}}) g_{\sigma_{1}} \big( F(t_{a_{2}}) g_{\sigma_{1}}\big)^{s-1} g_{\sigma_{2}} \bigg).\nonumber\\\nonumber
\end{aligned}$$ Which is equal to the sum:
$$\begin{aligned}
\big(q-1\big)\tau_{3} &\bigg(& \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-1}g^{2}_{\sigma_{1}} F\big(t_{a_{2}}\big) \underbrace{g_{\sigma_{2}} \big( g_{\sigma_{1}} F(t_{a_{2}})\big)^{s} g_{\sigma_{2}}}_{} \bigg) \nonumber\\\nonumber\\
& & + q\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-1}g^{2}_{\sigma_{1}} F\big(t_{a_{2}}\big) g_{\sigma_{2}} g_{\sigma_{1}} \big( F(t_{a_{2}}) g_{\sigma_{1}}\big)^{s-1} g_{\sigma_{2}} \bigg).\nonumber
\end{aligned}$$
Now, the first term is covered by corollary \[5\_1\_6\]. Thus we are interested with the second term: $$\begin{aligned}
\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-1}g^{2}_{\sigma_{1}} &F& \big(t_{a_{2}}\big) g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s-1} g_{\sigma_{1}} g_{\sigma_{2}} \bigg) = \nonumber\\\nonumber\\
& & q\tau_{3} \bigg(\big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s-1} g_{\sigma_{1}} g_{\sigma_{2}} \bigg) \nonumber\\\nonumber\\
&+& \big(q-1\big)\tau_{3} \bigg(\big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-2} g_{\sigma_{1}}F(t^{2}_{a_{2}}) g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s-1} g_{\sigma_{1}} g_{\sigma_{2}}\bigg). \nonumber
\end{aligned}$$
Which is equal to: $$\begin{aligned}
q\tau_{3} \bigg( \underbrace{g_{\sigma_{2}} \big( g_{\sigma_{1}} F(t_{a_{2}})\big)^{r} g_{\sigma_{2}}}_{} &\big(& g_{\sigma_{1}} F(t_{a_{2}})\big)^{s-1} g_{\sigma_{1}} \bigg) + \nonumber\\\nonumber\\
& & \big(q-1\big) \tau_{3} \bigg(\big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-2} g_{\sigma_{1}}F\big(t^{2}_{a_{2}}\big) g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s-1} g_{\sigma_{1}} g_{\sigma_{2}} \bigg) \nonumber
\end{aligned}$$
The first term is covered by corollary \[5\_1\_6\]. We are reduced to $$\begin{aligned}
\bigg(\tau_{3} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-2} g_{\sigma_{1}} F \big(t^{2}_{a_{2}}\big) g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s-1} g_{\sigma_{1}} g_{\sigma_{2}} \bigg), \nonumber
\end{aligned}$$
which is equal to: $$\begin{aligned}
\big(q-1\big)\tau_{3} &\bigg(&\underbrace{g_{\sigma_{2}}\big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-1} g_{\sigma_{2}}}_{} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{s-1} g_{\sigma_{1}}\bigg) + \nonumber\\\nonumber\\
& & q\tau_{3} \bigg(\big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-2} g_{\sigma_{1}} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s-1} g_{\sigma_{1}} g_{\sigma_{2}} \bigg). \nonumber
\end{aligned}$$
The first term is covered by corollary \[5\_1\_6\]. Thus, we see that, in general, the value of $\tau_{3}$ over $\big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r} g_{\sigma_{1}} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{s} g_{\sigma_{1}} g_{\sigma_{2}} $ can be shifted to its value over: $$\begin{aligned}
\big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r-2} g_{\sigma_{1}} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{s-1} g_{\sigma_{1}} g_{\sigma_{2}}. \nonumber
\end{aligned}$$
After a finite number of repetitions of the computation above (with the possibility of exchanging $r$ and $s$), we see that the lemma is proven modulo determining: $$\begin{aligned}
\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{m} \underbrace{g_{\sigma_{1}} g_{\sigma_{2}} g_{\sigma_{1}}}_{} F\big(t_{a_{2}}\big) g_{\sigma_{1}} g_{\sigma_{2}}\bigg). \nonumber
\end{aligned}$$
We see that the terms corresponding to $-g_{\sigma_{1}}$ and -1 correspond to Markov elements. While those who correspond to $-g_{\sigma_{1}}g_{\sigma_{2}}$ and $-g_{\sigma_{2}}$ are covered by corollary \[5\_1\_6\] for $h=1$. Finally the term corresponding to $-g_{\sigma_{2}}g_{\sigma_{1}}$ is covered by corollary \[5\_1\_6\] for $h=m$ .
\[5\_1\_8\] Suppose that $r$ and $s$ are positive such that $r \leq s $. Then: $$\begin{aligned}
\tau_{3} \bigg( g_{a_{3}}F\big(t_{a_{2}}\big)\big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s}g_{a_{3}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r}\bigg) = \sum^{j=h}_{j=0} c_{j} \big(g_{\sigma_{2}\sigma_{1}a_{3}}\big)^{j} + \sum_{i} M _{i}. \nonumber
\end{aligned}$$
Where $h \leq s$ , $c_{j}$ is in $F(\widehat{TL}_{2}(q))$ for every $j$ and $M_{i} $ is a Markov element for every $i$.
$$\begin{aligned}
\tau_{3} \bigg( g_{a_{3}} F \big(t_{a_{2}} &\big)& \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s}g_{a_{3}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r}\bigg) = \nonumber\\\nonumber\\
& & \tau_{3} \bigg( g_{a_{3}} \big(g_{\sigma_{2}\sigma_{1}a_{3}}\big)^{-1} g_{\sigma_{2}\sigma_{1}a_{3}}F\big(t_{a_{2}}\big)\big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s} g_{a_{3}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r}\bigg) = \nonumber\\\nonumber\\
& & \tau_{3} \bigg( g_{a_{3}}\big(g_{\sigma_{2}\sigma_{1}a_{3}}\big)^{-1} \psi \big[F\big(t_{a_{2}}\big)\big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s}\big] g_{\sigma_{2}\sigma_{1}a_{3}}g_{a_{3}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r} \bigg) = \nonumber\\\nonumber\\
& & \tau_{3} \bigg( g^{-1}_{\sigma_{1}} g^{-1}_{\sigma_{2}} g_{\sigma_{1}} \big( F(t_{a_{2}}) g_{\sigma_{1}} \big)^{s} g_{\sigma_{2}}g_{\sigma_{1}}g^{2}_{a_{3}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r} \bigg). \nonumber
\end{aligned}$$
Here, we see that this term is a sum of two terms coming from $g^{2}_{a_{3}} = (q-1) g_{a_{3}} + q $. The term corresponding to $ (q-1) g_{a_{3}} $ is covered the same way as in the last lemma (with $a_{3}$ instead of $\sigma_{2}$ above. Hence we treat the term corresponding to $q$, that is: $$\begin{aligned}
\tau_{3} \bigg( g^{-1}_{\sigma_{1}} \underbrace{g^{-1}_{\sigma_{2}}}_{} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{s} \underbrace{g_{\sigma_{1}} g_{\sigma_{2}}g_{\sigma_{1}}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r}\bigg). \nonumber
\end{aligned}$$
Before applying TL relations, we see in the same way as above, that we are reduced to the next value (otherwise, it is $\tau_{3}$ evaluated on a Markov element): $$\begin{aligned}
\tau_{3} \bigg( g^{-1}_{\sigma_{1}} \underbrace{g_{\sigma_{2}}}_{} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s} \underbrace{g_{\sigma_{1}} g_{\sigma_{2}}g_{\sigma_{1}}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r}\bigg). \nonumber
\end{aligned}$$
We see that the terms corresponding to $-g_{\sigma_{1}}$ and -1 correspond to Markov elements. And those who correspond to $-g_{\sigma_{2}g_{\sigma_{1}}}$ and $-g_{\sigma_{2}}$, are covered by corollary \[5\_1\_6\] for $h=s$.\
The term corresponding to $-g_{\sigma_{1}}g_{\sigma_{2}}$ is: $$\begin{aligned}
\tau_{3} \bigg( g^{-1}_{\sigma_{1}} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s} g_{\sigma_{1}} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r}\bigg), \nonumber
\end{aligned}$$
which is: $$\begin{aligned}
\frac{1-q}{q}\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{s} g_{\sigma_{1}} \underbrace{g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r}g_{\sigma_{2}}}_{}\bigg) + \frac{1}{q}\tau_{3} \bigg( g_{\sigma_{1}} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{s} g_{\sigma_{1}} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{r}\bigg) \nonumber
\end{aligned}$$
The first term is covered by corollary \[5\_1\_6\] for $h=r$. The second follows by lemma \[5\_1\_7\].\
Let us go back to $\tau_{3}( g_{\sigma_{2} \sigma_{1} a_{n+1}})^{k}$. The aim is to show that: $$\begin{aligned}
\tau_{3} \big( g_{\sigma_{2} \sigma_{1} a_{n+1}}\big)^{k} = \tau_{3}\bigg(\sum^{j=h}_{j=0} c_{j} (g_{\sigma_{2}\sigma_{1}a_{3}})^{j} + \Sigma_{i} M _{i} \bigg) ,\nonumber
\end{aligned}$$
where $h < k $. By lemma \[5\_1\_5\], it is sufficient to deal with: $$\begin{aligned}
g_{\sigma_{2}} \bigg( g_{\sigma_{1}} F(t_{a_{2}}) \bigg)^{k}g_{\sigma_{2}}\prod^{i=k-1}_{i=0} \psi^{i} \bigg[\big(F(t_{a_{2}})^{-1}\big)\bigg]. \nonumber
\end{aligned}$$
It is clear that this element is written as a linear combination of four kind of elements:\
- $g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h}$.
- $g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h}g_{\sigma_{1}}$.
- $g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big( F(t_{a_{2}})g_{\sigma_{1}} \big)^{h} F \big(t_{a_{2}}\big)$.
- $g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}}\big(F(t_{a_{2}})g_{\sigma_{1}} \big)^{h} $,\
where $h \leq [\frac{k}{2}] < k $, since $ 1 < k $.\
- We start by $\tau_{3} \bigg( g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}}\big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h} \bigg)$. Which is equal to:
$$\begin{aligned}
\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} \underbrace{g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h}g_{\sigma_{2}}}_{} \bigg), \nonumber
\end{aligned}$$
follows directly, regarding corollary \[5\_1\_6\] .\
- Now we consider $$\tau_{3} \bigg(g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h}g_{\sigma_{1}} \bigg),\label{eq_1}$$
which is equal to: $$\begin{aligned}
& & \tau_{3} \bigg(g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h}g_{\sigma_{1}} g_{\sigma_{2} \sigma_{1} a_{n+1}} \big(g_{\sigma_{2} \sigma_{1} a_{3}} \big)^{-1} \bigg) = \nonumber\\\nonumber\\
& & \tau_{3} \bigg(g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g^{2}_{\sigma_{2}} g_{\sigma_{1}} g_{a_{3}} \psi \big[\big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h}g_{\sigma_{1}}\big] \big(g_{\sigma_{2} \sigma_{1} a_{3}}\big)^{-1} \bigg)= \nonumber\\\nonumber\\
& & \tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g^{2}_{\sigma_{2}} g_{\sigma_{1}} g_{a_{3}} \big(F(t_{a_{2}})g_{\sigma_{1}} \big)^{h}F(t_{a_{2}}) g^{-1}_{a_{3}} F\big(t_{a_{2}}\big)\bigg), \nonumber\\\nonumber
\end{aligned}$$
with the very same steps as used above, we see that we are reduced to : $$\begin{aligned}
& & \tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}}) )^{k-1} g_{\sigma_{1}} g_{a_{3}} \big( F(t_{a_{2}})g_{\sigma_{1}} \big)^{h}F\big(t_{a_{2}}\big) g_{a_{3}} F\big(t_{a_{2}}\big) \bigg)\text{, which is: } \nonumber\\\nonumber\\
& & \tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g_{\sigma_{1}} \underbrace{g_{a_{3}} F\big(t_{a_{2}}\big)}_{}\big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h} \underbrace{g_{a_{3}} F\big(t_{a_{2}}\big)}_{}\bigg)= \nonumber
\end{aligned}$$
$$\begin{aligned}
\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g_{\sigma_{1}} \underbrace{F\big(t_{a_{2}}\big) g_{\sigma_{2}}}_{} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h} \underbrace{ F\big(t_{a_{2}}\big) g_{\sigma_{2}}}_{}\bigg), \text{ which is equal to:} \nonumber
\end{aligned}$$
$$\begin{aligned}
(q-1 &)&\tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h} g_{\sigma_{2}}\bigg) + q\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h-1} g_{\sigma_{1}} g_{\sigma_{2}}\bigg),\nonumber
\end{aligned}$$
we see that corollary \[5\_1\_6\] covers the first term. Thus we see that: $$\begin{aligned}
\tau_{3}\bigg(g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h}g_{\sigma_{1}}\bigg)\text{ in (eq. \ref{eq_1}), is shifted to: }\nonumber
\end{aligned}$$ $$\begin{aligned}
\tau_{3} \bigg(g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h-1}g_{\sigma_{1}} \bigg), \nonumber
\end{aligned}$$
going on in this manner, we arrive to: $$\begin{aligned}
\tau_{3} \bigg(g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}}g_{\sigma_{1}} F\big(t_{a_{2}}\big) g_{\sigma_{1}} \bigg), \nonumber
\end{aligned}$$
with the same steps above, we see that we are reduced to $$\begin{aligned}
\tau_{3} \bigg( \big(g_{\sigma_{1}} F\big(t_{a_{2}}\big) \big)^{k} g_{\sigma_{2}}g_{\sigma_{1}} F(t^{2}_{a_{2}})g_{\sigma_{2}}\bigg), \text{ which is equal to } \nonumber\\\nonumber\\
\big(q-1\big)\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}}g_{\sigma_{1}} F\big(t_{a_{2}}\big)g_{\sigma_{2}}\bigg) + q\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}}g_{\sigma_{1}} g_{\sigma_{2}} \bigg), \nonumber
\end{aligned}$$
corollary \[5\_1\_6\] and TL relations end the job.\
- Here we deal with $\tau_{3} \bigg(g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big( F(t_{a_{2}})g_{\sigma_{1}} \big)^{h}F\big(t_{a_{2}}\big) \bigg)$, which is:
$$\begin{aligned}
\tau_{3} \bigg( F\big(t_{a_{2}} &\big)& g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big( F(t_{a_{2}})g_{\sigma_{1}} \big)^{h} \bigg) \nonumber\\\nonumber\\
&=& \tau_{3} \bigg( g_{a_{3}} F\big(t_{a_{2}}\big) \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big( F(t_{a_{2}})g_{\sigma_{1}} \big)^{h} \bigg)\nonumber\\\nonumber\\
&=& \tau_{3} \bigg( g_{a_{3}} \big(F(t_{a_{2}})g_{\sigma_{1}} \big)^{k} F\big(t_{a_{2}}\big) g_{\sigma_{2}} \big(F(t_{a_{2}})g_{\sigma_{1}} \big)^{h} \bigg), \nonumber
\end{aligned}$$
but, $g_{\sigma_{2}} = F\big(t^{-1}_{a_{2}}\big) g_{a_{3}} F\big(t_{a_{2}}\big)$, thus:
$$\begin{aligned}
& & \tau_{3} \bigg( g_{a_{3}} \big(F(t_{a_{2}})g_{\sigma_{1}} \big)^{k} g_{a_{3}} F\big(t_{a_{2}}\big) \big(F(t_{a_{2}})g_{\sigma_{1}} \big)^{h} \bigg) \nonumber\\\nonumber\\
=& & \tau_{3} \bigg( g_{a_{3}}g^{-1}_{\sigma_{2} \sigma_{1} a_{3}}g_{\sigma_{2} \sigma_{1} a_{3}} \big(F(t_{a_{2}})g_{\sigma_{1}} \big)^{k} g_{a_{3}} F\big(t_{a_{2}}\big) \big(F(t_{a_{2}})g_{\sigma_{1}} \big)^{h} \bigg) \nonumber\\\nonumber\\
=& & \tau_{3} \bigg( g^{-1}_{\sigma_{2}} \psi \big[ \big(F(t_{a_{2}})g_{\sigma_{1}} \big)^{k}\big] g_{\sigma_{2} \sigma_{1} a_{3}} g_{a_{3}} F\big(t_{a_{2}}\big) \big( F(t_{a_{2}})g_{\sigma_{1}} \big)^{h} g^{-1}_{\sigma_{1}} \bigg), \nonumber
\end{aligned}$$
as we have done above, using the quadratic relations, we see that we are reduced to:
$$\begin{aligned}
& & \tau_{3}\bigg(g_{\sigma_{2}} \psi \big[ \big(F(t_{a_{2}})g_{\sigma_{1}} \big)^{k} \big] g_{\sigma_{2}} g_{\sigma_{1}} F\big(t_{a_{2}}\big) \big( F(t_{a_{2}})g_{\sigma_{1}} \big)^{h} g^{-1}_{\sigma_{1}}\bigg) \nonumber\\\nonumber\\
=& & \tau_{3} \bigg(g_{\sigma_{2}} \psi \big[ \big(F(t_{a_{2}})g_{\sigma_{1}} \big)^{k} \big] g_{\sigma_{2}} g_{\sigma_{1}} F\big(t^{2}_{a_{2}}\big) \big( g_{\sigma_{1}} F(t_{a_{2}})\big)^{h-1} \bigg) \nonumber\\\nonumber\\
=& & \big(q-1\big)\tau_{3} \bigg(g_{\sigma_{2}} \psi \big[ \big(F(t_{a_{2}})g_{\sigma_{1}} \big)^{k} \big] g_{\sigma_{2}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h} \bigg) + \nonumber\\\nonumber\\
& & q\tau_{3} \bigg( g_{\sigma_{2}} \psi \big[ \big(F(t_{a_{2}})g_{\sigma_{1}} \big)^{k} \big] g_{\sigma_{2}} g_{\sigma_{1}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h-1} \bigg), \nonumber
\end{aligned}$$
the first term is covered by corollary \[5\_1\_6\]. For the second we see that it is equal to: $$\begin{aligned}
q\tau_{3} \bigg(g_{\sigma_{2}} \big(g_{\sigma_{1}} &F& (t_{a_{2}})\big)^{k} g_{\sigma_{2}} g_{\sigma_{1}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h-1} \bigg), \text{ which is equal to} \nonumber\\\nonumber\\
& & q\big(q-1\big)\tau_{3}\bigg(g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big( g_{\sigma_{1}} F(t_{a_{2}})\big)^{h-1} \bigg) + \nonumber\\\nonumber\\
& & q^{2}\tau_{3} \bigg(g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} F\big(t_{a_{2}}\big) \big(g_{\sigma_{1}} F(t_{a_{2}})\big)^{h-2} \bigg),\nonumber\\\nonumber
\end{aligned}$$
the first term is obviously, covered by corollary \[5\_1\_6\], for the second one we see that it is case 3 itself, but with $ h-2 $ instead of $h$. Thus, we get two elements for $\tau_{3}$ to be evaluated on:\
- $g_{\sigma_{2}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}}F\big(t_{a_{2}}\big) g_{\sigma_{1}} F\big(t_{a_{2}}\big)$,
- $g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} F\big(t_{a_{2}}\big) \big( g_{\sigma_{1}} F(t_{a_{2}})\big)^{2}$.\
For \[$~b~$\] we can repeat what we have done until arriving to: $$\begin{aligned}
\tau_{3} \bigg(g_{\sigma_{2}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k}g_{\sigma_{2}} g_{\sigma_{1}} g_{\sigma_{1}} F\big(t_{a_{2}}\big) \bigg), \text{ which is the following sum:} \nonumber
\end{aligned}$$ $$\begin{aligned}
\big(q-1\big)\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} g_{\sigma_{1}} F\big(t_{a_{2}}\big) g_{\sigma_{2}} \bigg) + q\tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k}\underbrace{g_{\sigma_{2}} F\big(t_{a_{2}} \big) g_{\sigma_{2}}}_{F(t_{a_{2}}) g_{\sigma_{2}}F(t_{a_{2}})} \bigg), \nonumber
\end{aligned}$$
obviously, the first term is covered by corollary \[5\_1\_6\], the second term is a Markov element.\
For \[$~a~$\] we see that: $$\begin{aligned}
\tau_{3} \bigg( g_{\sigma_{2}} &\big(& g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} F\big(t_{a_{2}}\big) g_{\sigma_{1}} F\big(t_{a_{2}}\big) \bigg) \nonumber\\\nonumber\\
=\tau_{3} \bigg( g_{\sigma_{2}} &\big(& g_{\sigma_{1}} F(t_{a_{2}})\big)^{k-1} g_{\sigma_{1}} \underbrace{ F\big(t_{a_{2}}\big) g_{\sigma_{2}} F\big(t_{a_{2}}\big)}_{g^{2}_{\sigma_{2}} g_{a_{3}}} g_{\sigma_{1}} F\big(t_{a_{2}}\big) \bigg) \nonumber\\\nonumber\\
=& & \big(q-1\big) \tau_{3} \bigg( g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g_{\sigma_{1}} g_{\sigma_{2}} g_{a_{3}} g_{\sigma_{1}} F\big(t_{a_{2}}\big) \bigg) + \nonumber\\\nonumber\\
& & q\tau_{3} \bigg( g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g_{\sigma_{1}}g_{a_{3}} g_{\sigma_{1}} F\big(t_{a_{2}}\big) \bigg),\nonumber
\end{aligned}$$
the first term is covered by corollary \[5\_1\_6\], since it is equal to: $$\begin{aligned}
\big(q-1\big) \tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g_{\sigma_{1}} F\big(t_{a_{2}}\big) g_{\sigma_{2}} g_{\sigma_{1}} F\big(t_{a_{2}}\big) g_{\sigma_{2}} \bigg).\nonumber
\end{aligned}$$
For the second term, we see that: $$\begin{aligned}
& & q\tau_{3} \bigg( g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g_{\sigma_{1}} g_{a_{3}} g_{\sigma_{1}} F\big(t_{a_{2}}\big) \bigg) \nonumber\\\nonumber\\
& & =q\tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g_{\sigma_{1}} g_{a_{3}} g_{\sigma_{1}} F\big(t_{a_{2}}\big) g_{\sigma_{2}} \bigg) \nonumber
\end{aligned}$$
$$\begin{aligned}
=q\tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g_{\sigma_{1}} g_{a_{3}} g_{\sigma_{1}} g_{a_{3}} F\big(t_{a_{2}}\big) \bigg) \nonumber\\\nonumber\\
=q\tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g^{2}_{\sigma_{1}} g_{a_{3}} g_{\sigma_{1}} F\big(t_{a_{2}} \big) \bigg), \nonumber
\end{aligned}$$
which is a Markov element, since $ g_{a_{3}} = F(t_{a_{2}})g_{\sigma_{2}} F(t^{-1}_{a_{2}})$ .\
- We deal with $\tau_{3} \bigg( g_{\sigma_{2}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big( F(t_{a_{2}}) g_{\sigma_{1}} \big)^{h} \bigg) $, using the same techniques:
$$\begin{aligned}
& & \tau_{3} \bigg( g_{\sigma_{2}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big( F(t_{a_{2}})g_{\sigma_{1}} \big)^{h} \bigg) \nonumber\\\nonumber\\
=& & \tau_{3} \bigg( g_{\sigma_{2}} \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big( F(t_{a_{2}})g_{\sigma_{1}} \big)^{h} g_{\sigma_{2} \sigma_{1} a_{3}}g^{-1}_{\sigma_{2} \sigma_{1} a_{3}} \bigg) \nonumber\\\nonumber\\
=& & \tau_{3} \bigg( g_{\sigma_{2}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g^{2}_{\sigma_{2}} g_{\sigma_{1}} g_{a_{3}} \big(g_{\sigma_{1}}F(t_{a_{2}}) \big)^{h} g^{-1}_{\sigma_{2} \sigma_{1} a_{3}} \bigg), \nonumber
\end{aligned}$$
so, we are reduced to: $$\begin{aligned}
& & \tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g_{\sigma_{1}} g_{a_{3}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h} g_{a_{3}} F\big(t_{a_{2}}\big) \bigg). \text{ Which is equal to: } \nonumber\\\nonumber\\
& & \tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} \underbrace{ g_{\sigma_{1}} g_{a_{3}}g_{\sigma_{1}}}_{V(g_{\sigma_{1}}, g_{a_{3}})} F\big(t_{a_{2}}\big) \big(g_{\sigma_{1}}F(t_{a_{2}}) \big)^{h-1} F\big(t_{a_{2}}\big) g_{\sigma_{2}} \bigg), \nonumber
\end{aligned}$$
for -1 and $-g_{\sigma_{1}}$ it is a Markov element. For $- g_{a_{3}}g_{\sigma_{1}}$ we see that:
$$\begin{aligned}
& & \tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g_{a_{3}} g_{\sigma_{1}} F\big(t_{a_{2}}\big) \big(g_{\sigma_{1}}F(t_{a_{2}}) \big)^{h-1} F\big(t_{a_{2}}\big) g_{\sigma_{2}}\bigg) \nonumber\\\nonumber\\
=& & \tau_{3} \bigg( g_{a_{3}} F\big(t_{a_{2}}\big) \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g_{a_{3}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h} \bigg), \nonumber
\end{aligned}$$
which is covered by lemma \[5\_1\_8\].\
For $- g_{a_{3}}$, we see that: $$\begin{aligned}
\tau_{3} &\bigg(& \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g_{a_{3}} F\big(t_{a_{2}} \big) \big( g_{\sigma_{1}}F(t_{a_{2}}) \big)^{h-1} F\big(t_{a_{2}}\big)g_{\sigma_{2}} \bigg) \nonumber\\\nonumber\\
=\tau_{3} &\bigg(& \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-2} g_{\sigma_{1}} F\big(t^{2}_{a_{2}}\big) g_{\sigma_{2}} \big(g_{\sigma_{1}}F(t_{a_{2}}) \big)^{h-2} g_{\sigma_{1}} F\big(t^{2}_{a_{2}}\big) g_{\sigma_{2}} \bigg) \nonumber\\\nonumber\\
=&\big(&q-1\big) \tau_{3} \bigg( \big(g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-1} g_{\sigma_{2}} \big(g_{\sigma_{1}}F(t_{a_{2}}) \big)^{h-2} g_{\sigma_{1}} F\big(t^{2}_{a_{2}} \big) g_{\sigma_{2}} \bigg) + \nonumber\\\nonumber\\
q\tau_{3} &\bigg(& \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-2}g_{\sigma_{1}} g_{\sigma_{2}} \big(g_{\sigma_{1}}F(t_{a_{2}}) \big)^{h-2} g_{\sigma_{1}} F\big(t^{2}_{a_{2}}\big) g_{\sigma_{2}} \bigg), \nonumber
\end{aligned}$$
the first term is covered by corollary \[5\_1\_6\]. We do the same thing with $F(t^{2}_{a_{2}})$ in the second term, we arrive to: $$\begin{aligned}
q^{2} \tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k-2}g_{\sigma_{1}} g_{\sigma_{2}} \big( g_{\sigma_{1}}F(t_{a_{2}}) \big)^{h-2} g_{\sigma_{1}} g_{\sigma_{2}} \big), \nonumber
\end{aligned}$$
which is the case of lemma \[5\_1\_7\].\
For $-g_{\sigma_{1}} g_{a_{3}}$ we see that: $$\begin{aligned}
\tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}}) &\big)&^{k-1} g_{\sigma_{1}} g_{a_{3}} F\big(t_{a_{2}}\big) \big(g_{\sigma_{1}}F(t_{a_{2}}) \big)^{h-1} F\big(t_{a_{2}}\big) g_{\sigma_{2}} \bigg) \nonumber\\\nonumber\\
=\tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}}) &\big)&^{k} g_{\sigma_{2}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h-1} F\big(t_{a_{2}}\big) g_{\sigma_{2}} \bigg) \nonumber\\\nonumber\\
=&\big(& q-1\big) \tau_{3} \bigg( \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{h-1} g_{\sigma_{2}} \bigg) + \nonumber\\\nonumber\\
q\tau_{3} &\bigg(& \big( g_{\sigma_{1}} F(t_{a_{2}}) \big)^{k} g_{\sigma_{2}} \big(g_{\sigma_{1}}F(t_{a_{2}}) \big)^{h-2} g_{\sigma_{1}} g_{\sigma_{2}} \bigg), \nonumber
\end{aligned}$$
corollary \[5\_1\_6\] covers the first term, while the second term is covered by (1) from our four cases.\
J. J. Graham and G. I. Lehrer. The representation theory of affine Temperley-Lieb algebras. L’Enseignement Mathematique, 44, 173-218, 1998.\[Graham\_Lehrer\_1998\]
J. J. Graham and G. I. Lehrer. Diagram algebras, Hecke algebras and decomposition numbers at roots of unity. Annales Scientifiques de lÉcole Normale Supérieure, 36, Issue 4:479-524, 2003. \[Graham\_Lehrer\_2003\]
S. Al Harbat.Markov trace on a tower of affine Temperley-Lieb algebras of type $\tilde{A_{n}}$. 2013. [ *arXiv:1311.7092*](http://arxiv.org/abs/1311.7092) \[Sadek\_2013\_1\]
S. Al Harbat.A classification of affine fully commutative elements. 2013. [*arXiv:1311.7089v1*](http://xxx.tau.ac.il/abs/1311.7089) \[Sadek\_2013\_2\]
S. Al Harbat. On the affine braid group, affine Temperley-Lieb algebra and Markov trace. PH.D Thesis, 2013. \[Sadek\_Thesis\]
V. F. R. Jones. A polynomial invariant for knots via Von Neumann algebras. Bulletin, American Mathematical Society, 12, No. 1:103-111, 1985. \[Jones\_1985\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Numerical simulations are used in this work to investigate aspects of microstructure and microsegregation during rapid solidification of a -based superalloy in a laser powder bed fusion additive manufacturing process. Thermal modeling by finite element analysis simulates the laser melt pool, with surface temperatures in agreement with *in situ* thermographic measurements on Inconel 625. Geometric and thermal features of the simulated melt pools are extracted and used in subsequent mesoscale simulations. Solidification in the melt pool is simulated on two length scales. For the multicomponent alloy Inconel 625, microsegregation between dendrite arms is calculated using the Scheil-Gulliver solidification model and DICTRA software.[^1] Phase-field simulations, using – as a binary analogue to Inconel 625, produced microstructures with primary cellular/dendritic arm spacings in agreement with measured spacings in experimentally observed microstructures and a lesser extent of microsegregation than predicted by DICTRA simulations. The composition profiles are used to compare thermodynamic driving forces for nucleation against experimentally observed precipitates identified by electron and X-ray diffraction analyses. Our analysis lists the precipitates that may form from FCC phase of enriched interdendritic compositions and compares these against experimentally observed phases from heat treatments at two temperatures: stress relief at () or homogenization at ().'
address:
- 'Materials Science and Engineering Division, Material Measurement Laboratory, National Institute of Standards and Technology, 100 Bureau Drive, Gaithersburg, Maryland 20899, U. S. A.'
- 'Theiss Research, 7411 Eads Avenue, La Jolla, CA 92037'
- |
Intelligent Systems Division, Engineering Laboratory,\
National Institute of Standards and Technology, 100 Bureau Drive, Gaithersburg, Maryland 20899, U. S. A.
- 'Materials Measurement Science Division, Material Measurement Laboratory, National Institute of Standards and Technology, 100 Bureau Drive, Gaithersburg, Maryland 20899, U. S. A.'
author:
- Trevor Keller
- Greta Lindwall
- Supriyo Ghosh
- Li Ma
- 'Brandon M. Lane'
- Fan Zhang
- 'Ursula R. Kattner'
- 'Eric A. Lass'
- 'Jarred C. Heigel'
- Yaakov Idell
- 'Maureen E. Williams'
- 'Andrew J. Allen'
- 'Jonathan E. Guyer'
- 'Lyle E. Levine'
bibliography:
- 'Acta\_Additive625.bib'
date: 'May 1, 2017'
title: 'Application of Finite Element, Phase-field, and CALPHAD-based Methods to Additive Manufacturing of -based Superalloys'
---
Additive manufacturing; Finite element analysis (FEA); CALPHAD; Phase-field simulations; Microsegregation
Introduction
============
Inconel 625 (IN625) is a -based superalloy used for turbine parts that is strengthened by substitutional alloying elements such as , , and . Laser powder bed fusion (L-PBF), an additive manufacturing technique, presents opportunities to reduce the cost of making IN625 parts with appropriate geometries and internal cooling channels for high-temperature applications. Heat treatment is often necessary following additive manufacturing to relieve residual stress [@Denlinger2015; @Tremsin2016], and to homogenize the microstructure [@Xu2013; @Mostafaei2016]. Recent work shows that common heat treatments promote precipitation of secondary phases [@Zhang2017], which degrade mechanical properties (such as indentation hardness) in IN625 [@Suave2014]. Heat treatment schedules for wrought IN625 were designed to avoid these same precipitates [@Floreen1994]; however, there are substantial microstructural differences between wrought and L-PBF material [@Idell2016], with significant microsegregation of as-solidified material of particular interest here. IN625 processed by welding, casting, or directional solidification exhibits less homogeneity than wrought material, and typically contains and Laves phase precipitates [@DuPont1996; @Formenti2005]. Finding suitable stress-relieving and homogenizing heat treatments without sacrificing strength is an iterative process, but numerical modeling can help narrow the search.
3D finite element modeling has been crucial to understanding the L-PBF process. Models describing single laser tracks across the powder bed surface are routinely used to simulate heat dissipation through the solid substrate [@Song2012; @Yin2012]. Multi-track models allow for coupling residual stress evolution to these thermal profiles [@Li2010; @Hussein2013; @Hodge2014]. Recent models improve the thermal modeling using coupled Calculation of Phase Diagrams (CALPHAD) methods for accurate prediction of the melt pool boundary and solidification microstructure [@Smith2016]. Such advances in modeling the L-PBF system configuration and relevant materials properties are critical to accurate finite element analysis (FEA) predictions of the real process under conditions of interest.
Phase-field methods are commonly used to simulate microstructural features, between atomistic and continuum length scales. A scalar-valued order parameter defined throughout a spatial domain is used in these methods to label the presence or absence of some phase, hence “phase field.” The best available phase-field models for solidification were performed in 2D for binary alloys [@Karma1998; @Echebarria2004]. The simplified geometry and approximation of multicomponent alloys as binaries allows for efficient computation, and the models produce quantitatively correct mass redistribution across the solidifying interface. The model has been applied successfully to tungsten arc welding of – [@Farzadi2008], laser powder forming of – [@Fallah2012] and – [@Nie2014], and electron beam additive manufacturing of – – by combining with as a virtual element with mass fraction [@Gong2015].
A third approach to studying microsegregation during solidification uses DICTRA software [@Andersson2002], which implements a one-dimensional model for diffusion-controlled phase transformation. Despite its simplicity, this approach has the advantage of allowing for simulation of microsegregation and back-diffusion in multicomponent materials by combining CALPHAD thermodynamic and kinetic materials descriptions. DICTRA is routinely applied to multicomponent alloys, including -based superalloys [@Jablonski2009].
These modeling techniques—FEA, phase-field, and CALPHAD-based—are already in use studying various aspects of L-PBF, separately and increasingly in cooperation, for a variety of alloy systems. It is our goal to integrate all three models together to achieve high fidelity simulations of dendritic solidification in L-PBF IN625, with direct comparison against experimental results. This effort will improve the fundamental understanding of solidification in this system and produce input data for modeling solid state transformations in the future.
Numerical methods
=================
Finite element thermal model\[sec:meth\_thermal\]
-------------------------------------------------
Using the commercial FEA code ABAQUS [@Abaqus2013], a non-linear, transient, thermal model was designed and executed to obtain the global temperature history generated during laser irradiation of one layer of powder covering a solid substrate. The simulated powder layer thickness is , combining the nominal layer height with of underlayer densification. This matches the steady-state powder layer thickness observed in corresponding experimental builds. Both single-track and multiple-track laser scans across the metal powder layer were modeled, in which one “track” is a linear path to be followed by the laser separated by from adjacent laser pathways traversed in anti-parallel directions. The width of the molten pool in the transverse direction is , so material at the midpoint melts on adjacent scans. To reduce computation time, the elements that interact with the laser beam are finely meshed within the diameter of the laser, and a coarse mesh was used for the surrounding loose powder and substrate, visible in Figure \[fig:fea\_meltpool\]. Ma *et al.* [@Ma2015] described this model previously, with detailed discussion of appropriate parameter values.
{width="\textwidth"}
Heat transfer in the L-PBF process was modeled using the energy balance equation with Fourier’s law of heat conduction and internal sources of heat [@Bird2007], $$\label{eqn:fourier}
\frac{\partial\left(\rho c_p T\right)}{\partial t} = \nabla\cdot\left(\kappa\nabla T\right) + Q,$$ in which the evolution of temperature $T$ with time $t$ depends on material density $\rho$, thermal conductivity $\kappa$, specific heat capacity $c_p$ which depends on latent heat, and internal heat $Q$ which depends on radiative loss. The temperature-dependent bulk material density and specific heat were calculated from a Scheil-Gulliver simulation for the nominal IN625 composition and using the TCNI8 thermodynamic database [@TCNI8] within the Thermo-Calc software [@Andersson2002]. The initial condition assumed a uniform temperature of throughout the specimen at time $t = 0$. Adiabatic conditions were applied to all boundaries except the top surface, on which the boundary condition is $$\label{eqn:thermalbc}
\left(-\kappa\nabla T\right)\cdot\hat{n} = q_s + h\left(T-T_e\right) + \varepsilon\sigma\left(T^4 - T_e^4\right).$$ The three terms on the right-hand side represent heat input from the laser, heat convection due to flowing process gas, and radiation. Equation \[eqn:thermalbc\] depends on the surface normal $\hat{n}$, laser input heat $q_s$, convective heat transfer coefficient $h$, thermal radiation coefficient $\varepsilon$, the Stefan-Boltzmann constant $\sigma$, and ambient temperature $T_e$. The laser input was modeled after the single-mode continuous wave fiber laser ($\lambda = \SI{1070}{\nano\meter}$) used in our experimental L-PBF system (described in Sec. \[sec:expt\]). Interaction between the laser and material is modeled using a Gaussian expression for surface heat flux [@Roberts2009], $$\label{eqn:gauss}
q_s = \frac{2AP}{\pi r_b^2}\exp\left(\frac{-2r^2}{r_b^2}\right),$$ with power $P = \SI{195}{\watt}$, powder bed absorption coefficient $A = \num{0.50}$, laser beam radius $r_b = \SI{50}{\micro\meter}$, and radial distance to the beam centerline $r$ measured in . The simulated laser scanning speed was .
Each element stores its temperature and a Boolean variable indicating whether it has ever exceeded the liquidus temperature, $T_\ell$. Elements in the powder layer are initialized with this variable set to “false,” indicating a powder state. Upon melting, the variable switches to “true,” indicating a bulk state. Substrate elements are initialized with this variable set to “true.” There is no mechanism for switching this melt-state variable from “true” to “false:” the fused material can never revert to powder. The materials parameters $\rho$ and $\kappa$ for each element depend on both variables, $T$ and melt-state. In the powder state, $\kappa$ depends on the packing fraction, particle size distribution, particle morphology, and thermal conductivity of the bulk material and process gas [@Rombouts2005; @Tanaka2012]. In the simulations described in this work, $\kappa$ was specified in the range from to after [@Childs2005]. As $T$ rises during the first melting event, $\rho$ and $\kappa$ linearly increase from their powder to bulk values when $T$ is above the solidus temperature $T_s$ and below the liquidus temperature $T_\ell$ given by the IN625 phase diagram [@TCNI8; @Andersson2002]. If $T$ exceeds $T_s$, then the melt-state variable switches to “true;” $\rho$ and $\kappa$ are thereafter functions of $T$, only. Note that the bulk state variable does not differentiate between solid and liquid phases: the FEA model uses $T$ to choose phase-dependent materials properties appropriately. Latent heat effects are captured in the evaluation of $c_p$ [@Abaqus2013].
CALPHAD-based solidification models\[sec:meth\_thermo\]
-------------------------------------------------------
To estimate the extent of microsegregation during solidification of a material, the Scheil-Gulliver model [@Gulliver1909; @Scheil1942] is often applied. The model can be used for multicomponent materials, provided that a thermodynamic description for the multicomponent alloy is available. The model assumes perfect mixing in the liquid and no diffusion in the solid phase. These conditions are not found in nature, so the result is a theoretical limit: less segregation is expected during real solidification processes than the Scheil-Gulliver model predicts, since finite diffusion in both the liquid and solid phases will contribute to mass redistribution.
Microsegregation predictions may also be obtained using DICTRA software [@Andersson2002] that include the effect of diffusion in liquid and solid during solidification. Local equilibrium is assumed at the interface between phases, in this case liquid and $\gamma$, and flux-balance is maintained for each element. Solutions to the diffusion equation and the assumption of local equilibrium at the phase interface are used to determine the tie-line that satisfies flux-balance. Because the simulations are 1D, the effects of dendrite tip diffusion are not included.
For the IN625 segregation simulation, we used the commercial thermodynamic database TCNI8 [@TCNI8] and the NIST Superalloy mobility database [@Campbell2002]. The DICTRA simulation domain was , which is half of the secondary dendrite arm spacing measured from experimental microstructures (e.g., Figure \[fig:experiment\]). Temperature was specified as spatially uniform, but time-dependent, with values taken directly from the FEA thermal model. To reduce the computational complexity, metals contributing mass fractions below were excluded, producing a simplified system composition – – – – – .
Phase-field solidification model
--------------------------------
We used a quantitative phase-field model in 2D to study the cellular/dendritic nature of the solidification of the melt pool [@Echebarria2004], with a simplified representation of IN625 as a – binary system with only FCC $\gamma$ phase and liquid phase. The bulk composition has a mass fraction of . The scalar phase parameter $\phi$ indicates whether a point in the 2D field is liquid $(\phi=-1)$, solid $(\phi=1)$, or within the solid-liquid interface $(-1<\phi<1)$. The phase-field is not conserved, and evolves in time $t$ and space $(x,y)$ as $$\begin{aligned}
\label{eqn:phi}
\frac{\partial \phi}{\partial t} &= \frac{1}{\tau_0{a(\hat{n})}^2}\bigg(W_{0}^2\nabla \left[{a(\hat{n})}^2 \nabla\phi\right] + \phi - \phi^3\\\nonumber
&- \frac{\partial}{\partial x} \left[ a(\hat{n}) \frac{\partial a(\hat{n})}{\partial\hat{n}} \frac{\partial\phi}{\partial y}\right]
+ \frac{\partial}{\partial y} \left[ a(\hat{n}) \frac{\partial a(\hat{n})}{\partial\hat{n}} \frac{\partial\phi}{\partial x}\right]\\\nonumber
&- \frac{\lambda}{1-k} (1-\phi^2)^2 \left[\exp\left(u\right) - 1 + \frac{T- T_0}{m_{\ell} c_0/k}\right]\bigg).
\end{aligned}$$ The dimensionless interfacial energy $a(\hat{n}) = 1 + \epsilon_4 \cos (4\theta)$, with four-fold anisotropy of magnitude $\epsilon_4$, interface normal vector $\hat{n} = -\frac{\nabla\phi}{|\nabla \phi|}$, and orientation angle $\theta = \arctan\left(\frac{\partial\phi}{\partial y}/\frac{\partial\phi}{\partial x}\right)$. The non-dimensional deviation of chemical potential, $u = \ln\left(\frac{2ck/c_0}{1+k-(1-k)\phi}\right)$, is defined with respect to the equilibrium chemical potential at a reference temperature $T_0$ and system composition $c_0$. The frozen temperature approximation is applied such that a linear temperature profile with constant gradient $G$ translates along the growth axis $(x)$ with constant velocity $V_s$ [@Smith1955; @Mullins1964; @Langer1980; @Merchant1990; @Huntley1993]: $T(x,t) = T_0 + G(x - V_s t)$. Interface thickness $W_0$ and relaxation time constant $\tau_0$ are related through the capillary length, $d_0 = a_1 W_0/\lambda$. Asymptotic analysis, performed by enforcing local equilibrium at the interface as its width vanishes, also links these quantities through a dimensionless coupling parameter $\lambda$ and diffusion constant in the liquid $D_{\ell}$ [@Karma1996; @Echebarria2004]: $\tau_0 = a_2\lambda W_{0}^{2}/D_{\ell}$. The fitting parameters $a_1 = 0.8839$ and $a_2 = 0.6267$ depend on the forms of the free energy functional and free energy density, respectively [@Karma1998]. $W_0$ is therefore the only free parameter, chosen to be .
Composition is modeled with a conserved field $c$, and evolves as $$\begin{aligned}
\label{eqn:conc}
\frac{\partial c}{\partial t} &= \nabla \cdot \left( \frac{1-\phi}{2} D_\ell \left[ 1+k-(1-k)\phi\right] \frac{c_0}{k} \nabla \exp\left(u\right)\right.\\\nonumber
&+ \left.\frac{W_0}{2\sqrt{2}} (1-k) \frac{c_0}{k} \exp\left(u\right) \frac{\partial \phi}{\partial t} \hat{n}\right).
\end{aligned}$$ This expression neglects the effect of thermal gradients on diffusion, or the Soret effect [@Groot1984a], which contributes to macrosegregation under low solidification velocities [@Zheng1998], but not microsegregation during rapid solidification. The phase diagram of the – system exhibits a practically linear liquidus with constant slope $m_{\ell}=\SI{-10.5}{\kelvin\per\percent}$, measured with respect to mass percentage , and constant partition coefficient $k=0.48$ in this dilute region. Equations \[eqn:phi\] and \[eqn:conc\] were solved on a uniform rectilinear grid using a finite volume method and an explicit time marching scheme with zero-flux boundary conditions. Model parameters for a dilute solution of in were used directly from Nie *et al.* [@Nie2014], summarized in their Table 1. Further analysis of solidification microstructures in dilute – alloys using this model are reported elsewhere [@Ghosh2017].
Experimental methods and results\[sec:expt\]
============================================
Test cubes of IN625 were additively manufactured by the NIST Engineering Laboratory using an EOSINT M270 (EOS GmbH, Krailling, Germany); for L-PBF system details, the interested reader may refer to [@Lane2016]. The EOS NickelAlloy IN625 powder (EOS GmbH, Krailling, Germany) was supplied with compositions listed in Table \[tab:AM625comp\], as measured by inductively coupled plasma atomic emission spectroscopy (ICP) and flame atomic absorption spectroscopy (FAAS), or by X-ray fluorescence spectroscopy (XRF, conforming to [@ASTMxrf]). All measured values are within the standard ranges for IN625 [@ASTM625]. For calibration of the FEA thermal model, multiple-track laser scans were made on () thick solutionized IN625 plate (High Performance Alloys, Inc. Windfall, IN, USA). For both IN625 media, the fiber laser power was and scan speed was .
Element Standard Range Supplied (ICP) Supplied (XRF)
--------- ---------------- ---------------- ----------------
minimum balance balance
to
to
to
maximum maximum
maximum maximum
maximum
maximum maximum
maximum
maximum
$^*$ maximum
$^*$ maximum maximum maximum
$^*$ maximum
: Allowable and measured mass fractions of constituent elements for IN625 L-PBF feedstock used in this work [@ASTM625], as determined by ICP, FAAS (indicated by an asterisk, $^*$), or XRF.[]{data-label="tab:AM625comp"}
Scanning electron microscopy (SEM) specimens were polished to a surface finish using standard metallographic preparation techniques, then etched for in aqua regia.
Specimens for transmission electron microscopy (TEM) were prepared using electrical discharge machining to cut disks in diameter, which were then electropolished until electron-transparent with a perchloric acid, methanol, and butanol solution. Figure \[fig:experiment\] provides a representative secondary electron micrograph from the as-built L-PBF IN625 specimen: the primary dendrite arms are spaced and the secondary arms are spaced . Regions without secondary arms are also present: the wide range of localized solidification conditions produced by L-PBF can produce regions of cellular growth, so we describe the experimental microstructures as “cellular/dendritic.” Figure \[fig:temcarbides\] provides representative bright field scanning transmission electron micrographs (STEM) showing precipitates as light and dark circular spots, to in diameter. Precipitates are found near cell/dendrite boundaries in as-built material, but are not apparently localized to microstructural features after stress relief at for . Electron diffraction patterns confirm the presence of , , and carbides in stress-relieved material, whereas diffraction patterns of precipitates in as-built material could not be indexed due to residual stress and high dislocation densities.
![Representative cross section through as-built L-PBF IN625 specimen, etched with aqua regia, showing cellular/dendritic microstructure. Hitachi S-4700 secondary electron image with accelerating potential, working distance. This image has been manipulated to increase its contrast.[]{data-label="fig:experiment"}](Figure2){width="\textwidth"}
{width="\textwidth"}
X-ray diffraction experiments were performed at the ultra-small angle X-ray scattering (USAXS) facility at the Advanced Photon Source at Argonne National Laboratory [@Ilavsky2009; @Ilavsky2013], with specimen preparation and measurement conditions reported previously [@Zhang2017]. The relative distance and tilt between the specimen and detector are calibrated using powder. The sample was heated to and held at for , then cooled to , corresponding to the manufacturer’s recommended stress-relieving heat treatment for IN625 [@EOS625]. The sample was then heated to and held at for , then cooled back to , a treatment suggested by previous work [@Zhang2017]. The heating rates were , the cooling rates were , and the temperature stability was .
![Synchrotron XRD observations of phase evolution during heat treatments of AM IN625 with scattering vector magnitude $q = 4\pi\sin(\theta)/\lambda$ expressed in terms of scattering angle $\theta$ and X-ray wavelength $\lambda$. Top panel: experimental XRD patterns of the AM IN625 sample acquired at following the specified heat treatments. Bottom panels: simulated powder XRD patterns for dominant phases ($\gamma$, , and Laves) based on the lattice symmetry and parameters described in the text.[]{data-label="fig:XRD625"}](Figure4){width="\textwidth"}
From the room temperature XRD patterns, Figure \[fig:XRD625\], only the FCC cubic lattice ($\gamma$ phase) with a lattice parameter of $(0.358 \pm 0.001)$ could be identified. The first annealing step to led to the formation of carbides, most of which are with a cubic lattice of $Fd\bar{3}m$ group and a lattice parameter of $(1.011 \pm 0.003)$ . The second heating step to preserved the carbides and promoted the formation of an intermetallic Laves phase. The Laves phase has a hexagonal lattice of $P6_3/mmc$ group, with lattice parameters $a = (0.481 \pm 0.001)$ and $c = (1.565 \pm 0.004)$ . The uncertainties in these values are reported with confidence. Hence, we conclude that from a statistical point of view, the dominant precipitates after this two-step heat treatment are carbides and the Laves phase.
Simulation results
==================
Finite element thermal simulations\[sec:fea\_results\]
------------------------------------------------------
To validate our FEA thermal model, we compared its surface temperature prediction against *in situ* thermographic measurements. Details of the thermographic measurement setup were published in [@Lane2016], and data used for the single-track comparison given here are described in [@Heigel2017]. The same thermal camera settings used in [@Lane2016; @Heigel2017] were used here: integration time, and to spectral range.
As described in [@Lane2016], thermographic imaging of laser scans on metal powder produces highly stochastic temperature fields with localized gradients due to the varying surface structure and emissivity, which inhibit true temperature measurement. In contrast, scans on flat plates of bulk metal result in smooth temperature gradients, and single-line scans create steady-state melt pools that simplify comparisons to FEA simulations. Therefore, FEA simulation results are compared against experimental laser scans on bare plates, without powder. A representative temperature profile on IN625 plate was taken from one frame of the thermal video after the melt pool attained nominal steady state. Figure \[fig:heatprofile\] compares the measured and predicted thermal profiles for a single track scan without powder using a scan speed of and laser power .
The emissivity of the surface of a real specimen—also known as emittivity—is highly dependent on processing, and values for the IN625 melt pool, rolled plate, and powder surfaces are unknown. Therefore, the thermal camera signal cannot be directly converted to temperature for comparisons to FEA model predictions. Instead, we scale thermographic data based on an observed solidification zone following a technique developed for analysis of – – powder [@Yadroitsev2014]. A shoulder in the thermographic profile behind the melt pool is attributed to the solidification zone, apparent between to in Figure \[fig:heatprofile\]. The profile is scaled by an assumed emittivity and converted to temperature units such that the temperature in this zone matches the solidus temperature, $T_s = \SI{1587}{\kelvin}$, which we calculated for IN625 using the TCNI8 thermodynamic database [@TCNI8]. This technique uses $T_s$ as a reference temperature, and results in emittivity of $\epsilon = 0.13$, which is a reasonable value for a semi-specular metallic surface at the same spectral range and viewing angle $(43.7^\circ)$ as the thermal imaging setup [@Touloukian1970]. The calibrated range of the thermal camera was to . Temperatures above this range saturate the camera, and temperatures below approach the noise floor of the camera detector. Applying this emittivity maps the reportable temperature range to to , as shown by the solid blue line in Figure \[fig:heatprofile\].
![Results of FEA thermal modeling (red dashed line) and *in situ* thermographic measurement (solid and dotted blue lines) for single-line scan on bare IN625 plate. Shoulder in measurement data corresponds to melt pool boundary. Dotted blue lines indicate thermographic data outside the calibrated range but still shown for clarity.[]{data-label="fig:heatprofile"}](Figure5){width="\textwidth"}
Figure \[fig:heatschedule\] shows the predicted temperature as a function of time at a point on the surface located midway between the centers of two of three anti-parallel laser scan tracks. Full melting and re-melting occurs as the laser beam traverses the two nearest tracks, passing the same distance away from the measurement point both times. Heating without melting is observed as the laser scans material along the third track. This three-track temperature profile was used directly in DICTRA simulations of IN625 solidification.
![Temperature on the surface of a single powder layer as a function of time at a position midway between two melt pool centers from a three track scanning FEA thermal model. The surface material fully melts during two near passes, with the molten pool overlapping this surface midpoint by ; on the third pass, removed from the melt pool boundary, this material is reheated without melting.[]{data-label="fig:heatschedule"}](Figure6){width="\textwidth"}
Figure \[fig:fea\_meltpool\] shows the thermal profile of the melt pool in cross-section. The rectangular meshing elements near the surface measure $\SI{10}{\micro\meter} \times \SI{10}{\micro\meter} \times \SI{6}{\micro\meter}$. Contours illustrate the temperature field, with the solid-liquid interface coinciding with the border between red and orange bands. The transformation angle $\alpha$ between the melt pool boundary and laser scanning direction indicates the local solidification speed, $V_s = V_b \cos\alpha$ [@Farzadi2008].
FEA thermal model results were also used to determine solidification parameters to be employed for the phase-field simulations described in Section \[sec:pf\_results\]. Average cooling rate $\dot{T}=\SI{e5}{\kelvin\per\second}$ was calculated from Figure \[fig:heatschedule\], measuring the slope of the line connecting the maximum and minimum temperatures of the second peak to its subsequent valley. A constant value for thermal gradient $G=\SI{e7}{\kelvin\per\meter}$ was chosen for the simulations. It is representative of the values along the solidus contour in Figure \[fig:fea\_meltpool\], $T_s=\SI{1587}{\kelvin}$: we computed $G=|\nabla T|$ in the range from $\SI{0.57e7}{\kelvin\per\meter}$ to $\SI{2.2e7}{\kelvin\per\meter}$, for mesh points to the left of the melt pool minimum. Values taken along the solidus contour behind the melt pool minimum yielded a range of angles $\alpha$ from to for $V_b=\SI{0.8}{\meter\per\second}$, and $V_s$ ranged between and Note that the solidifying interface experiences localized thermal gradients and solidification speeds: the cooling rate $\dot{T} = GV_s$ in L-PBF processes is not constant.
CALPHAD-based solidification simulations
----------------------------------------
Figure \[fig:segphase\]a represents solidification of the interdendritic region as predicted by the Scheil-Gulliver model and by DICTRA for the thermal profile in Figure \[fig:heatprofile\]. Based on the low fraction of secondary phases observed in as-solidified experimental specimens (*e.g.*, Figures \[fig:temcarbides\] and \[fig:XRD625\]), we made the simplifying assumption that liquid solidifies into $\gamma$ phase, only. Due to the high cooling rate, the DICTRA results show solidification behavior very similar to the Scheil-Gulliver model (Figure \[fig:segphase\]a). Figure \[fig:segphase\]b shows the Scheil-Gulliver model prediction of liquid composition. Microsegregation occurs for all elements, with particularly high segregation of and into the liquid. In the last solidified liquid, $T=\SI{1250}{\kelvin}$, the and mass fractions are as high as and , respectively. The last solidified liquid is also enriched in whereas and are depleted.
{width="\textwidth"}
{width="\textwidth"}
Detailed DICTRA simulation results are shown in Figure \[fig:segcomp\], with an inset showing results near the centerline of interdendritic liquid. The composition profiles show microsegregation from the secondary dendrite core ($x = \SI{0}{\nano\meter}$) to the interdendritic region ($x = \SI{150}{\nano\meter}$). Since applying the FEA thermal profile (Figure \[fig:heatschedule\]) for the DICTRA simulation leads to complete melting at the second scan temperature peak, the resulting segregation in Figure \[fig:segcomp\] occurs during this cooling. The third laser scan, centered from the FEA measurement point, leads to minor homogenization of the segregated profiles. This is, however, only notable over a distance less than (inset, Figure \[fig:segcomp\]) and can have no significant effect on the microsegregation profile. Figure \[fig:segcomp\] shows most of the microsegregation occurs within the region near the last solidified liquid, in good agreement with a recent synchrotron SAXS study of the homogenization kinetics of L-PBF IN625 [@Zhang2017], where a novel analysis directly linked the length of X-ray streaks to the length scale of the segregation. It can further be concluded that the segregation is in the same range as in the Scheil-Gulliver simulation. The main difference is that the DICTRA simulation results in a greater amount of and somewhat less in the interdendritic region compared to the last solidified liquid composition obtained from the Scheil-Gulliver model, as expected.
Phase-field solidification simulations {#sec:pf_results}
--------------------------------------
Several phase-field simulations were performed with different values for $V_s$, but constant $G$, in order to study microstructural evolution under various cooling conditions of interest. Simulations were initialized with a planar solid-liquid interface with random perturbations up to one grid spacing in location, and using a hyperbolic tangent profile along the growth axis to smooth the step change over several grid spacings for numerical stability. The microstructures evolved according to Equations \[eqn:phi\] and \[eqn:conc\], and simulations ran until steady state was achieved. Depending on solidification conditions, interfacial instabilities can grow into dendrites (with secondary arms) or cells (without). For the parameters chosen in this work, only cellular domains resulted, as represented in Figure \[fig:pfresult\] with $V_s = \SI{25}{\milli\meter\per\second}$. In Figure \[fig:pfresult\], the steady-state mass fraction of in the cell center line was . At the midpoint of the intercellular liquid, the mass fraction of was . Intercellular liquid is shown pinching off at the root of intercellular grooves with mass fraction of . This composition is below the eutectic composition of . Thus secondary solid phases, which are not available to these simulations, are not expected to form. The main point is that the microsegregation found by this model is considerably less than is predicted by a Scheil-Gulliver analysis of binary –: such an analysis for $k=0.48$ predicts a mass fraction of in the cell centerline, in solid formed when only liquid remains, and likely formation of secondary phases from the last liquid to solidify.
![Cellular microstructure in – predicted by phase-field simulation after of growth with $V_s = \SI{25}{\milli\meter\per\second}$, as illustrated by the scalar composition field $c$. Image represents $\SI{6}{\micro\meter}\times \SI{12}{\micro\meter}$ window near the growth front. Note steady-state enrichment of undercooled liquid near cell tips and pinch-off of liquid droplets at the root of intercellular grooves. diffuses out of these droplets into the surrounding solid, resulting in the linear pattern of disks with fading intensity. Arrow indicates growth direction.[]{data-label="fig:pfresult"}](Figure9){width="\textwidth"}
The primary dendrite arm spacing $\lambda_1$, a commonly reported metric for cellular/dendritic microstructures, was averaged from multiple simulations with the same $V_s$ by dividing the sum of simulation domain widths by the sum of the numbers of cells advancing at steady state. We measured $\lambda_1$ in the range from to from our simulations, depending on the cooling rate $\dot{T}$ which ranged from to , as shown in Figure \[fig:pdas\_hunt\]. For reference, we compared our results with the analytical models of Hunt [@Hunt1979], $$\lambda_1 = A(k \Gamma \Delta T_0 D_\ell)^{0.25} G^{-0.5}V_s^{-0.25}
\label{eqn:hunt}$$ with $A=2.83$, and Kurz and Fisher [@Kurz1981], $$\lambda_1 = A(\Gamma \Delta T_0 D_\ell / k)^{0.25} G^{-0.5}V_s^{-0.25}
\label{eqn:kurzapprox}$$ with $A=4.3$, under the simplifying assumption that undercooling $\Delta T\approx\Delta T_0 = T_{\ell}-T_s$, measured from the equilibrium phase diagram at $c_0$. $\Gamma=\SI{3.65e-7}{\kelvin\meter}$ is the Gibbs-Thomson coefficient. The proportionality constants depend on the 3D geometry assumed for the dendrite arrays: Hunt assumed spherical dendrite tips and derived $A(k \Gamma \Delta T_0 D_\ell)^{0.25}$, while Kurz and Fisher assumed ellipsoidal tips and derived $A(\Gamma \Delta T_0 D_\ell / k)^{0.25}$. From the line of best fit through our simulation data points, we calculate $A = 6.8$ from Equation \[eqn:hunt\], and $A = 4.7$ from Equation \[eqn:kurzapprox\]. Figure \[fig:pdas\_hunt\] shows that neither model provides an objectively good fit. The deviation is likely due to the combination of our reduced geometry (2D rather than 3D), rapid cooling $(\Delta T \neq \Delta T_0)$, and more complex tip geometry than either simple model. This result is supported by a body of experimental evidence demonstrating large effects of $c_0$, $G$, and $V_s$ on dendrite tip geometries and $\lambda_1$ [@Ungar1985; @Liu1995; @Kirkaldy1995].
![Comparison of primary dendrite arm spacing measured from simulated microstructures with two analytical models for simple cellular/dendritic tip geometries: spherical (Hunt [@Hunt1979]) and ellipsoidal (Kurz and Fisher [@Kurz1981]).[]{data-label="fig:pdas_hunt"}](Figure10){width="\textwidth"}
Discussion
==========
Experimental observations via TEM reveal significant precipitation near cell/dendrite boundaries in as-built L-PBF material (Figure \[fig:temcarbides\]). Stress relief at the manufacturer’s recommended temperature () was found to promote further growth of carbides, with precipitation and growth from the $\gamma$ matrix observed by *in situ* synchrotron XRD (Figure \[fig:XRD625\]). Further treatment at promoted growth and Laves phase precipitation: after at temperature, persisted and Laves phase grew to measurable levels. This growth is despite the fact that homogeneous IN625 of nominal composition is single-phase $\gamma$, as established experimentally and from thermodynamics [@Floreen1994; @TCNI8].
Investigation into the cause of the observed precipitation during heat treatments requires knowledge of the solidification rate, which can not be directly determined from the laser scan speed. *In situ* thermographic measurements on bare IN625 plate provided the melt pool width and thermal gradients on the surface. FEA simulations were calibrated against these surface data, then used to extract details of the full 3D melt pool shape during three-track L-PBF simulations (Figure \[fig:fea\_meltpool\] and Section \[sec:fea\_results\]). For a fiber laser operating at and scanning a single powder layer at , the solidification rate is between at the bottom of the melt pool and at its trailing edge. The cooling rate along this interface was on the order of , a rapid solidification condition.
Experimentally, these conditions produce fine cellular dendrites. To assess the level of microsegregation, we performed modeling on three levels: Scheil-Gulliver, with no diffusion in $\gamma$ and perfectly mixed liquid; DICTRA, with multicomponent diffusion in 1D for both $\gamma$ and liquid; and phase-field, with 2D cells/dendrites of $\gamma$ growing into liquid.
To assess how microsegregation correlates to the experimentally identified precipitates, we look at the thermodynamic contribution to the driving force for nucleation of various phases from the $\gamma$ matrix [@Hillert2007], ignoring interfacial energy contributions and kinetic obstacles. The phase with largest driving force would be expected to nucleate first. Comparison of these driving force values may provide additional insight regarding the nucleation of precipitates in dendritically segregated material. In Table \[tab:drivingforce\], the phases most likely to nucleate are listed in decreasing order of driving force for the composition of solid when liquid remains (mass fraction – C– Cr– – – ). On this basis, we expect -rich carbide to nucleate first; if it cannot, we expect . If cannot form, a -rich phase is expected, or BCC, followed by -rich $\sigma$, $P$, Laves, or . As Table \[tab:drivingforce\] demonstrates, the specific nucleation order depends on temperature. Note that this is a sequence of energetic favorability, not a predicted time-evolution: phases toward the bottom of the table can only precipitate if local conditions preclude any of the more-favorable phases from doing so. The favored precipitate is also not the same everywhere: local variations in composition and other materials properties will change the tabulated sequence. Competitive growth and coarsening may occur when time and diffusion are factored in, but these effects are beyond the scope of this paper.
------------------------------------- -- ------------------------------------- --
Phase Phase
${{\ensuremath\underline{MC}}}$ ${{\ensuremath\underline{MC}}}$
${{\ensuremath\underline{\mu}}}$
BCC
BCC ${{\ensuremath\underline{\mu}}}$
${{\ensuremath\underline{\sigma}}}$ ${{\ensuremath\underline{Laves}}}$
Laves liquid
${{\ensuremath\underline{\delta}}}$ ${{\ensuremath\underline{\sigma}}}$
$\gamma''$ $\gamma''$
${{\ensuremath\underline{\delta}}}$
------------------------------------- -- ------------------------------------- --
: Thermodynamic driving force for nucleation of secondary phases from $\gamma$ for the enriched (interdendritic) composition, – C– Cr– – – , at the stress relief and homogenization treatment temperatures. Greater values indicate larger driving forces for nucleation. Equilibrium phases for each temperature are underlined.[]{data-label="tab:drivingforce"}
The nucleation sequence suggested by the driving force calculations can be compared to the calculated equilibrium phases for these compositions at the stress relieving and homogenizing temperatures. Thermodynamic calculations show , BCC, $\delta$ and $\sigma$ phases to be stable at both temperatures, while $\mu$ phase is only stable at the lower temperature. Therefore, as the highly segregated profiles diffuse during long heat treatments, solid state transformations are to be expected—an active topic of research beyond the scope of this paper.
It is important to note that the nucleation sequence and the ability for a phase to form are dependent on several mechanisms and not only the nucleation driving force. A complete description of nucleation must include the interfacial energy between the matrix and secondary phase, which depends on the interfacial area and crystallographic details. For large or complex unit cells in particular, kinetic effects play an important role: a thermodynamically stable phase may be unable to nucleate or grow if its constituent elements are not sufficiently mobile in the matrix at the specified temperature and composition. This may explain, for example, the fact that BCC phase has not been observed in our experiments despite its apparently high driving force for nucleation. Furthermore, the thermodynamic database used for the driving force calculation may contain uncertainties and is also limited to a certain energy resolution. This could, in particular, be the case when different phases have similar driving forces, e.g. and BCC in our case.
Even if and Laves are not equilibrium phases at these conditions, their driving forces for nucleation are high enough to put them early in the nucleation sequence: earlier than the equilibrium phases $\delta$ and $\sigma$. This could explain the experimental observation of these phases. It is also interesting to note that the -rich carbide is an equilibrium phase at temperatures as high as , which could complicate the homogenization treatment.
The binary phase-field simulations clearly show microsegregation in the cellular growth pattern. The spherical droplets in Figure \[fig:pfresult\] form at high solidification speeds as the advancing cells pinch off liquid pockets to maintain the steady-state intercellular groove depth. These liquid droplets are enriched to more than twice the initial composition, depending on cooling rate, with some loss to diffusion during solidification. Experiments involving rapid solidification of electron-beam melted –– alloys produced morphologically similar droplets, which precipitated intermetallic compounds during solidification [@Boettinger1988]. While our models did not consider secondary solid phases, and therefore could not simulate precipitation in the evolving microstructure, the microsegregation patterns are real and significant. It is interesting to note that even the highly concentrated droplets are enriched less than half as much as the microsegregation predicted by the DICTRA simulations. To some extent, this difference reflects model geometry: the phase-field simulations were performed in 2D with realistically curved dendrite tips, while DICTRA simulations represent 1D planar solidification. Curvature effects significantly affect diffusion. In addition, under rapid solidification conditions, local chemical equilibrium across the solid-liquid interface is not expected to hold, and results in the natural phenomenon known as solute trapping [@Baker1969]. DICTRA enforces local equilibrium without accounting for solute trapping, and therefore over-estimates the extent of microsegregation in these simulations. While the phase-field model does not enforce local equilibrium, the diffuse interface artificially magnifies the effect. Our phase-field model includes an anti-trapping flux in Equation \[eqn:conc\], to correct the spurious contribution [@Karma1998]. The correction is accurate to second-order in $\frac{W_0}{d_0}$, the dimensionless ratio of interface thickness to capillary length, and is exact for low solidification velocities. Under the rapid solidification conditions investigated here, this second-order expression does not cancel the spurious effect, which artificially decreases the simulated microsegregation. Under L-PBF conditions, which produce both curved interfaces and solute trapping in real material, the phase-field model produces a more accurate estimate of microsegregation than DICTRA for the the cellular/dendritic microstructure.
There are also confounding factors that affect the exact compositions predicted by our phase-field and DICTRA models. The real temperature-time cycle during L-PBF is more complex than the FEA model employed in this study: for example, the effects of building additional layers on the temperature profile are neglected here. This simplified thermal history may discount the influence of solid state diffusion, producing artificially high compositions in enriched regions of the microstructure. This work also models the laser input as a surface heat source, only, which neglects the effects of particle ejection on heat and mass transfer in the melt pool [@Khairallah2016; @Teng2016]. Furthermore, thermophysical properties such as $k$ and $m_{\ell}$ depend on the solidification speed, and may deviate significantly from equilibrium values in the rapid solidification regime, a topic explored in greater detail elsewhere [@Ghosh2017]. Finally, powder grains of somewhat different sizes and composition lead to compositional variations throughout the part, which may produce highly localized segregation patterns not captured in our models.
Conclusion
==========
From this work, we conclude:
- For additively manufactured Inconel 625, heat treatments of at and promote precipitation of secondary phases from the $\gamma$ matrix.
- FEA simulations of the moving 3D melt pool shape predict maximum solidification rates only times the laser scan speed, accompanied by a cooling rate of . This L-PBF process occurs under rapid solidification conditions.
- The Scheil-Gulliver model was used to compute the maximum possible extent of microsegregation for the multicomponent alloy IN625. DICTRA simulations that include the role of diffusion in $\gamma$ reduce the degree of microsegregation near the end of solidification. Phase-field simulations of cellular/dendritic microstructures further reduce microsegregation, finding compositions times the Scheil-Gulliver prediction.
- Carbides (, , and ), topologically close-packed ($\delta$, $\mu$, $\sigma$), and Laves phases have negative thermodynamic driving forces for nucleation from the FCC phase under the conditions of microsegregation and temperature studied: these may spontaneously precipitate under the conditions of temperature and microsegregation investigated.
This effort demonstrates the viability of cooperatively modeling L-PBF processes using several techniques to assess microstructural phenomena in IN625. Work in progress will extend this tool chain to investigate the effects of interfacial energy and diffusion in microsegregated regions on competitive precipitation and solid state transformations. With a complete microstructure model, we plan to evaluate the whole process in order to find a better way to prepare this material for service.
Acknowledgment {#acknowledgment .unnumbered}
==============
Use of the Advanced Photon Source, an Office of Science User Facility operated for the U.S. Department of Energy (DOE) Office of Science by Argonne National Laboratory, was supported by the U.S. DOE under Contract No. DE-AC02-06CH11357.
[^1]: Certain commercial entities, equipment, or materials may be identified in this document in order to describe an experimental procedure or concept adequately. Such identification is not intended to imply recommendation or endorsement by the National Institute of Standards and Technology (NIST), nor is it intended to imply that the entities, materials, or equipment are necessarily the best available for the purpose.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Вивчаються відображення з розгалуженням, які задовольняють деяку умову спотворення модуля сімей кривих. У ситуації, коли область визначення відображень є локально зв’язною на своїй межі, відображена область є регулярною, а мажоранта, яка відповідає за спотворення модуля сімей кривих, є інтегровною, доведено, що сім’ї вказаних відображень з одною умовою нормування є одностайно неперервними в замиканні вихідної області.'
author:
- 'Євген О.Севостьянов, Наталія С. Ількевич'
title: ОДНОСТАЙНА НЕПЕРЕРВНІСТЬ ПО ПРОСТИХ КІНЦЯХ ВІДОБРАЖЕНЬ З УМОВОЮ НОРМУВАННЯ
---
\[section\] \[section\] \[section\] \[section\]
\[section\] \[section\]
\[section\]
We study branching mappings that satisfy some condition of distortion of the modulus of families of paths. In a situation where the definition domain of mappings is locally connected on its boundary, the mapped domain is regular, and the majorant responsible for distortion of the modulus of families of paths is integrable, it is proved that the families of all specified mappings with one normalization condition are equicontinuous in the closure of the given domain.
Вступ
=====
Дану роботу присвячено відображенням з обмеженим і скінченним спотворенням, які активно вивчаються останнім часом (див., напр., [@GRY]–). Нещодавно в наших спільних роботах ми дослідили ситуації, в яких відображення з так званою оберненою нерівністю Полецького мають неперервне межове продовження, а їх сім’ї є одностайно неперервними як у внутрішніх, так і межових точках області, див., напр., – і [@SSD]. В даній статті ми розглянемо ще один важливий випадок, коли відображення можуть мати розгалуження, а області мають складну структуру, при цьому, відображення фіксують принаймні одну точку області. Зауважимо, що клас відображень з оберненою нерівністю Полецького включає до себе відображення з обмеженим спотворенням і відображення зі скінченним спотворенням довжини (див., напр., , [@Ri теорема 6.7.II] і [@MRSY теорема 8.5]).
Звернемося до означень. Нехай $y_0\in {\Bbb R}^n,$ $0<r_1<r_2<\infty$ і $$\label{eq1**}
A(y_0, r_1,r_2)=\left\{ y\,\in\,{\Bbb R}^n:
r_1<|y-y_0|<r_2\right\}\,.$$ Для заданих множин $E,$ $F\subset\overline{{\Bbb R}^n}$ і області $D\subset {\Bbb R}^n$ позначимо через $\Gamma(E,F,D)$ сім’ю всіх кривих $\gamma:[a,b]\rightarrow \overline{{\Bbb R}^n}$ таких, що $\gamma(a)\in E,\gamma(b)\in\,F$ і $\gamma(t)\in D$ при $t \in [a,
b].$ Якщо $f:D\rightarrow {\Bbb R}^n$ – задане відображення, $y_0\in f(D)$ і $0<r_1<r_2<d_0=\sup\limits_{y\in f(D)}|y-y_0|,$ то через $\Gamma_f(y_0, r_1, r_2)$ ми позначимо сім’ю всіх кривих $\gamma$ в області $D$ таких, що $f(\gamma)\in \Gamma(S(y_0, r_1),
S(y_0, r_2), A(y_0,r_1,r_2)).$ Нехай $Q:{\Bbb R}^n\rightarrow [0,
\infty]$ – вимірна за Лебегом функція. Будемо говорити, що [*$f$ задовольняє обернену нерівність Полецького*]{} в точці $y_0\in
f(D),$ якщо співвідношення $$\label{eq2*A}
M(\Gamma_f(y_0, r_1, r_2))\leqslant \int\limits_{f(D)} Q(y)\cdot
\eta^n (|y-y_0|)\, dm(y)
$$ виконується для довільної вимірної за Лебегом функції $\eta:
(r_1,r_2)\rightarrow [0,\infty ]$ такій, що $$\label{eqA2}
\int\limits_{r_1}^{r_2}\eta(r)\, dr\geqslant 1\,.$$ Зауважимо, що нерівності (\[eq2\*A\]) добре відомі в теорії квазірегулярних відображень і виконуються для них при $Q=N(f,
D)\cdot K, $ де $N(f, D)$ – максимальна кратність відображення в $D,$ а $K\geqslant 1$ – деяка стала, яка може бути обчислена як $K={\rm ess \sup}\, K_O(x, f),$ $K_O(x, f)=\Vert
f^{\,\prime}(x)\Vert^n/|J(x, f)|$ при $J(x, f)\ne 0;$ $K_O(x, f)=1$ при $f^{\,\prime}(x)=0,$ і $K_O(x, f)=\infty$ при $f^{\,\prime}(x)\ne 0,$ але $J(x, f)=0$ (див., напр., або [@Ri теорема 6.7.II]). Відображення $f:D\rightarrow {\Bbb R}^n$ називається [*дискретним*]{}, якщо прообраз $\{f^{-1}\left(y\right)\}$ кожної точки $y\,\in\,{\Bbb R}^n$ складається з ізольованих точок, і [*відкритим*]{}, якщо образ будь-якої відкритої множини $U\subset D$ є відкритою множиною в ${\Bbb R}^n.$ Відображення $f$ області $D$ на $D^{\,\prime}$ називається [*замкненим*]{}, якщо $f(E)$ є замкненим в $D^{\,\prime}$ для будь-якої замкненої множини $E\subset D$ (див., напр., ). Означення простого кінця, яке використовується нижче, може бути знайдено в роботі [@IS], див. також –. Тут і далі $\overline{D}_P$ позначає поповнення області $D$ її простими кінцями, а $E_D=\overline{D}_P\setminus D$ – множина всіх простих кінців в $D.$ Говоримо, що обмежена область $D$ в ${\Bbb R}^n$ [*регулярна*]{}, якщо $D$ може бути квазіконформно відображена на область з локально квазіконформною межею, замикання якої є компактом в ${\Bbb R}^n,$ крім того, кожен простий кінець $P\subset E_D$ є регулярним. Зауважимо, що у просторі ${\Bbb R}^n$ кожний простий кінець регулярної області містить ланцюг розрізів з властивістю $d(\sigma_{m})\rightarrow 0$ при $m\rightarrow\infty,$ і навпаки, якщо у кінця є вказана властивість, то він – простий (див. [@Na теорема 5.1]). Крім того, замикання $\overline{D}_P$ регулярної області $D$ є [*метризовним*]{}, при цьому, якщо $g:D_0\rightarrow D$ – квазіконформне відображення області $D_0$ з локально квазіконформною межею на область $D,$ то для $x, y\in
\overline{D}_P$ покладемо: $$\label{eq1A}
\rho(x, y):=|g^{\,-1}(x)-g^{\,-1}(y)|\,,$$ де для $x\in E_D$ елемент $g^{\,-1}(x)$ розуміється як деяка (єдина) точка межі $D_0,$ коректно визначена з огляду на [@Na теорема 4.1].
Сформулюємо тепер основний результат даної статті. Для цього, для областей $D, D^{\,\prime}\subset {\Bbb R}^n,$ $n\geqslant 2,$ точок $a\in D,$ $b\in D^{\,\prime}$ і вимірної за Лебегом функції $Q:D^{\,\prime}\rightarrow [0, \infty]$ позначимо через ${\frak
S}_{a, b, Q }(D, D^{\,\prime})$ сім’ю всіх відкритих дискретних і замкнених відображень $f$ області $D$ на $D^{\,\prime},$ що задовольняють умову (\[eq2\*A\]) для кожного $y_0\in D^{\,\prime},$ причому $f(a)=b.$ Виконується наступне твердження.
\[th2\] [ Припустимо, що область $D$ має слабо плоску межу, жодна із зв’язних компонент якої не вироджена. Якщо $Q\in L^1(D^{\,\prime}),$ і область $D^{\,\prime}$ є регулярною, то будь-яке $f\in{\frak
S}_{a, b, Q }(D, D^{\,\prime})$ неперервно продовжується до відображення $\overline{f}:\overline{D}\rightarrow
\overline{D^{\,\prime}}_P,$ причому, $\overline{f}(\overline{D})=\overline{D^{\,\prime}}_P$ і сім’я ${\frak S}_{a, b, Q }(\overline{D}, \overline{D^{\,\prime}}),$ яка складається з усіх продовжених відображень $\overline{f}:\overline{D}\rightarrow \overline{D^{\,\prime}}_P,$ одностайно неперервна в $\overline{D}.$ ]{}
Зауважимо, що теорему \[th2\] можна застосувати для достатньо широкого спектру областей $D^{\,\prime}.$ Зокрема, за теоремою Рімана регулярною областю є будь-яка однозв’язна область, межа якої містить більше одної точки. Більше того, кожна обмежена скінченно зв’язна область є конформним образом області, межа якої складається зі скінченної кількості кіл і ізольованих точок (див., напр., [@Gol теорема V.6.2]). Оскільки для конформних відображень ізольовані точки є усувними, то вихідна область може вважатися регулярною і без вироджених межових компонент.
Доведення основного результату ґрунтується на певних властивостях відображень зі збереженням діаметру прообразу деякого континууму. Наступна лема за деяких інших припущень на відображення і області, що розглядаються, була доведена в , і . Нехай $h$ – хордальна відстань в $\overline{{\Bbb R}^n}$ (див., напр., означення 12.1 в [@Va]).
\[lem3\] [ Нехай $n\geqslant 2,$ $D$ і $D^{\,\prime}$ – області в ${\Bbb
R}^n,$ причому, $D$ має слабо плоску межу, жодна компонента зв’язності якої не вироджується в точку, а область $D^{\,\prime}$ є регулярною. Нехай також $A$ – невироджений континуум в $D^{\,\prime}$ і $\delta>0.$ Припустимо, $f_m$ – послідовність відкритих, дискретних і замкнених відображень області $D$ на $D^{\,\prime}$ з наступною властивістю: для кожного $m=1,2,\ldots$ знайдеться континуум $A_m\subset D,$ $m=1,2,\ldots ,$ такий, що $f_m(A_m)=A$ і $h(A_m)\geqslant \delta>0.$ Якщо кожне $f_m$ задовольняє співвідношення (\[eq2\*A\]) для кожного $y_0\in
D^{\,\prime},$ причому, $Q\in L^1(D^{\,\prime}),$ то знайдеться $\delta_1>0$ таке, що $$h(A_m,
\partial D)>\delta_1>0\quad \forall\,\, m\in {\Bbb
N}\,.$$ ]{}
Через компактність простору $\overline{{\Bbb R}^n}$ межа області $D$ не порожня і є компактом, так що відстань $h(A_m,
\partial D)$ коректно визначена.
Проведемо доведення від супротивного. Припустимо, що висновок леми не є вірним. Тоді для кожного $k\in {\Bbb N}$ знайдеться номер $m=m_k$ такий, що $h(A_{m_k},
\partial D)<1/k.$ Можна вважати, що послідовність $m_k$ зростає по $k.$ Оскільки $A_{m_k}$ – компакт, то знайдуться $x_k\in A_{m_k}$ і $y_k\in
\partial D$ такі, що $h(A_{m_k},
\partial D)=h(x_k, y_k)<1/k$ (див. малюнок \[fig3\]).
![До доведення леми \[lem3\][]{data-label="fig3"}](Graphic3A.eps)
Оскільки $\partial D$ – компактна множина, ми можемо вважати, що $y_k\rightarrow y_0\in
\partial D$ при $k\rightarrow \infty;$ тоді також $x_k\rightarrow y_0\in \partial D$ при $k\rightarrow \infty.$ Нехай $K_0$ – компонента зв’язності $\partial D,$ яка містить точку $y_0.$ Очевидно, $K_0$ – континуум в $\overline{{\Bbb R}^n}.$ Оскільки $\partial D$ – слабо плоска, за теоремою 1 в [@Sev] відображення $f_{m_k}$ має неперервне продовження $\overline{f}_{m_k}\colon\overline{D}\rightarrow
\overline{D^{\,\prime}}_P.$ Нехай $\rho$ – одна з метрик у (\[eq1A\]), і нехай $g:D_0\rightarrow D$ – квазіконформне відображення деякої області $D_0$ з локально квазіконформною межею на $D,$ яке відповідає метриці $\rho$ у (\[eq1A\]). Оскільки $\overline{f}_{m_k}$ неперервне на компакті $\overline{D},$ відображення $\overline{f}_{m_k}$ є рівномірно неперервним у $\overline{D}$ відносно метрики $\rho$ при кожному фіксованому $k.$ Іншими словами, для кожного $\varepsilon>0$ знайдеться $\delta_k=\delta_k(\varepsilon)<1/k$ таке, що $$\label{eq3B}
\rho(\overline{f}_{m_k}(x),\overline{f}_{m_k}(x_0))<\varepsilon
$$ $$\forall\,\, x,x_0\in \overline{D},\quad h(x, x_0)<\delta_k\,, \quad
\delta_k<1/k\,,$$ де, як звично, $h$ – хордальна метрика в $\overline{{\Bbb R}^n}.$ Оберемо $\varepsilon>0$ таким, щоб $$\label{eq5D}
\varepsilon<(1/2)\cdot {\rm dist}\,(\partial D_0, g^{\,-1}(A))\,.$$ Позначимо $B_h(x_0, r)=\{x\in \overline{{\Bbb R}^n}: h(x, x_0)<r\}.$ Для фіксованого $k\in {\Bbb N},$ покладемо $$B_k:=\bigcup\limits_{x_0\in K_0}B_h(x_0, \delta_k)\,,\quad k\in {\Bbb
N}\,.$$ Оскільки $B_k$ – окіл континуума $K_0,$ за [@HK лема 2.2] знайдеться окіл $U_k$ множини $K_0,$ такий, що $U_k\subset B_k$ і $U_k\cap D$ зв’язна. Можна вважати, що $U_k$ – відкрита, так що $U_k\cap D$ є лінійно зв’язною (див. [@MRSY пропозиція 13.1]). Нехай $h(K_0)=m_0.$ Тоді знайдуться $z_0, w_0\in K_0$ такі, що $h(K_0)=h(z_0, w_0)=m_0.$ Отже, знайдуться послідовності $\overline{y_k}\in U_k\cap D,$ $z_k\in U_k\cap D$ і $w_k\in U_k\cap
D$ такі, що $z_k\rightarrow z_0,$ $\overline{y_k}\rightarrow y_0$ і $w_k\rightarrow w_0$ при $k\rightarrow\infty.$ Можна вважати, що $$\label{eq2B}
h(z_k, w_k)>m_0/2\quad \forall\,\, k\in {\Bbb N}\,.$$ Оскільки множина $U_k\cap D$ лінійно зв’язна, ми можемо з’єднати точки $z_k,$ $\overline{y_k}$ і $w_k,$ використовуючи деяку криву $\gamma_k\in U_k\cap D.$ Як завжди, ми позначаємо через $|\gamma_k|$ носій (образ) кривої $\gamma_k$ в області $D.$ Тоді $f_{m_k}(|\gamma_k|)$ – компактна множина в $D^{\,\prime}.$ Якщо $x\in|\gamma_k|,$ то знайдеться $x_0\in K_0$ таке, що $x\in B(x_0,
\delta_k).$ Зафіксуємо довільне $\omega\in A\subset D.$ Оскільки $x\in|\gamma_k|$ і, більше того, $x$ – внутрішня точка $D,$ ми можемо використовувати запис $f_{m_k}(x)$ замість $\overline{f}_{m_k}(x).$ Зі співвідношень (\[eq3B\]) і (\[eq5D\]), а також за нерівністю трикутника, ми отримаємо, що $$\rho(f_{m_k}(x),\omega)\geqslant
\rho(\omega,
\overline{f}_{m_k}(x_0))-\rho(\overline{f}_{m_k}(x_0),f_{m_k}(x))\geqslant$$ $$\label{eq4C}\geqslant {\rm dist}\,(\partial D_0, g^{\,-1}(A))-(1/2)\cdot {\rm
dist}\,(\partial D_0, g^{\,-1}(A))=$$ $$=(1/2)\cdot {\rm dist}\,(\partial
D_0, g^{\,-1}(A))>\varepsilon$$ для достатньо великих $k\in {\Bbb N},$ де $${\rm dist}(\partial D_0, g^{\,-1}(A)):=\inf\limits_{x\in
\partial D_0, y\in g^{\,-1}(A)} |x-y|\,.$$ Переходячи до $\inf$ в (\[eq4C\]) по всіх $x\in |\gamma_k|$ і $\omega\in A,$ ми отримаємо, що $$\label{eq5B}
\rho(f_{m_k}(|\gamma_k|), A)>\varepsilon,\qquad k=1,2,\ldots \,.$$ Тепер покажемо, що знайдеться $\varepsilon_1>0$ таке, що $$\label{eq6B}
{\rm dist}\,(f_{m_k}(|\gamma_k|), A)>\varepsilon_1, \quad\forall\,\,
k=1,2,\ldots \,,$$ де ${\rm dist},$ як звично, позначає евклідову відстань між множинами $A, B\subset{\Bbb R}^n.$ Дійсно, нехай співвідношення (\[eq6B\]) не має місця, тоді для числа $\varepsilon_l=1/l,$ $l=1,2,\ldots$ знайдуться елементи $\xi_l\in
|\gamma_{k_l}|$ і $\zeta_l\in A$ такі, що $$\label{eq7A}
|f_{m_{k_l}}(\xi_l)-\zeta_l|<1/l\,,\quad l=1,2,\ldots \,.$$ Можна вважати, що послідовність $k_l,$ $l=1,2,\ldots,$ є зростаючою. Оскільки $A$ є компактом, ми також можемо вважати, що послідовність $\zeta_l$ збігається до $\zeta_0\in A$ при $l\rightarrow\infty.$ За нерівністю трикутника і по (\[eq7A\]) будемо мати, що $$\label{eq8A}
|f_{m_{k_l}}(\xi_l)-\zeta_0|\rightarrow 0\,,\quad
l\rightarrow\infty\,.$$ З іншого боку, нагадаємо, що $\rho(f_{m_k}(x),
\omega)=|g^{\,-1}(f_{m_k}(x))-g^{\,-1}(\omega)|,$ де $g:D_0\rightarrow D$ – деяке квазіконформне відображення області $D_0$ на $D,$ див. (\[eq1A\]). Зокрема, відображення $g^{\,-1}$ є неперервним у $D,$ отже, з огляду на нерівність трикутника і (\[eq8A\]), ми отримаємо, що $$|g^{\,-1}(f_{m_{k_l}}(\xi_l))-g^{\,-1}(\zeta_l)|\leqslant$$ $$\label{eq9A}
\leqslant
|g^{\,-1}(f_{m_{k_l}}(\xi_l))-g^{\,-1}(\zeta_0)|+|g^{\,-1}(\zeta_0)-g^{\,-1}(\zeta_l)|\rightarrow
0,\quad l\rightarrow\infty\,.$$ Проте, за означенням метрики $\rho$ і з (\[eq9A\]) випливає, що $$\rho(f_{m_{k_l}}(|\gamma_{k_l}|), A)\leqslant$$ $$\leqslant \rho(f_{m_{k_l}}(\xi_l), \zeta_l)=
|g^{\,-1}(f_{m_{k_l}}(\xi_l))-g^{\,-1}(\zeta_l)|\rightarrow 0, \quad
l\rightarrow\infty\,,$$ що суперечить (\[eq5B\]). Отримана суперечність вказує на справедливість співвідношення (\[eq6B\]).
Покриємо множину $A$ кулями $B(x, \varepsilon_1/4),$ $x\in A.$ Оскільки $A$ компакт, ми можемо вважати, що $A\subset
\bigcup\limits_{i=1}^{M_0}B(x_i, \varepsilon_1/4),$ $x_i\in A,$ $i=1,2,\ldots, M_0,$ $1\leqslant M_0<\infty.$ За означенням, $M_0$ залежить тільки від $A,$ зокрема, $M_0$ не залежить від $k.$ Покладемо $$\label{eq5C}
\Gamma_k:=\Gamma(A_{m_k}, |\gamma_k|, D)\,.$$ Нехай $\Gamma_{ki}:=\Gamma_{f_{m_k}}(x_i, \varepsilon_1/4,
\varepsilon_1/2),$ іншими словами, $\Gamma_{ki}$ складається з усіх кривих $\gamma:[0, 1]\rightarrow D,$ таких що $f_{m_k}(\gamma(0))\in
S(x_i, \varepsilon_1/4),$ $f_{m_k}(\gamma(1))\in S(x_i,
\varepsilon_1/2)$ і $\gamma(t)\in A(x_i, \varepsilon_1/4,
\varepsilon_1/2)$ при $0<t<1.$ Покажемо, що $$\label{eq6C}
\Gamma_k>\bigcup\limits_{i=1}^{M_0}\Gamma_{ki}\,.$$ Справді, нехай $\widetilde{\gamma}\in \Gamma_k,$ іншими словами, $\widetilde{\gamma}:[0, 1]\rightarrow D,$ $\widetilde{\gamma}(0)\in
A_{m_k},$ $\widetilde{\gamma}(1)\in |\gamma_k|$ і $\widetilde{\gamma}(t)\in D$ при $0\leqslant t\leqslant 1.$ Тоді $\gamma^{\,*}:=f_{m_k}(\widetilde{\gamma})\in \Gamma(A,
f_{m_k}(|\gamma_k|), D^{\,\prime}).$ Оскільки кулі $B(x_i,
\varepsilon_1/4),$ $1\leqslant i\leqslant M_0,$ утворюють покриття компакта $A,$ знайдеться $i\in {\Bbb N}$ таке, що $\gamma^{\,*}(0)\in B(x_i, \varepsilon_1/4)$ і $\gamma^{\,*}(1)\in
f_{m_k}(|\gamma_k|).$ За співвідношенням (\[eq6B\]), $|\gamma^{\,*}|\cap B(x_i, \varepsilon_1/4)\ne\varnothing\ne
|\gamma^{\,*}|\cap (D^{\,\prime}\setminus B(x_i, \varepsilon_1/4)).$ Отже, за [@Ku теорема 1.I.5.46] знайдеться $0<t_1<1$ таке, що $\gamma^{\,*}(t_1)\in S(x_i, \varepsilon_1/4).$ Можна вважати, що $\gamma^{\,*}(t)\not\in B(x_i, \varepsilon_1/4)$ при $t>t_1.$ Покладемо $\gamma_1:=\gamma^{\,*}|_{[t_1, 1]}.$ З (\[eq6B\]) випливає, що $|\gamma_1|\cap B(x_i,
\varepsilon_1/2)\ne\varnothing\ne |\gamma_1|\cap (D\setminus B(x_i,
\varepsilon_1/2)).$ Отже, за [@Ku теорема 1.I.5.46] знайдеться $t_1<t_2<1$ таке, що $\gamma^{\,*}(t_2)\in S(x_i, \varepsilon_1/2).$ Можна вважати, що $\gamma^{\,*}(t)\in B(x_i, \varepsilon_1/2)$ при всіх $t<t_2.$ Вважаючи $\gamma_2:=\gamma^{\,*}|_{[t_1, t_2]},$ зауважимо, що крива $\gamma_2$ є підкривою $\gamma^{\,*},$ яка належить $\Gamma(S(x_i, \varepsilon_1/4), S(x_i, \varepsilon_1/2),
A(x_i, \varepsilon_1/4, \varepsilon_1/2)).$
Остаточно, $\widetilde{\gamma}$ має підкриву $\widetilde{\gamma_2}:=\widetilde{\gamma}|_{[t_1, t_2]},$ таку, що $f_{m_k}\circ\widetilde{\gamma_2}=\gamma_2,$ причому, $\gamma_2\in
\Gamma(S(x_i, \varepsilon_1/4), S(x_i, \varepsilon_1/2), A(x_i,
\varepsilon_1/4, \varepsilon_1/2)).$ Отже, співвідношення (\[eq6C\]) встановлене. Покладемо $$\eta(t)= \left\{
\begin{array}{rr}
4/\varepsilon_1, & t\in [\varepsilon_1/4, \varepsilon_1/2],\\
0, & t\not\in [\varepsilon_1/4, \varepsilon_1/2]\,.
\end{array}
\right.$$ Зауважимо, що $\eta$ задовольняє співвідношення (\[eqA2\]) при $r_1=\varepsilon_1/4$ і $r_2=\varepsilon_1/2.$ Оскільки відображення $f_{m_k}$ задовольняє співвідношення (\[eq2\*A\]), то припускаючи тут $y_0=x_i,$ отримаємо: $$\label{eq8C}
M(\Gamma_{f_{m_k}}(x_i, \varepsilon_1/4, \varepsilon_1/2))\leqslant
(4/\varepsilon_1)^n\cdot\Vert Q\Vert_1<c<\infty\,,$$ де $c$ – деяка додатна стала і $\Vert Q\Vert_1$ – $L_1$-норма функції $Q$ в $D^{\,\prime}.$ З (\[eq6C\]) і (\[eq8C\]), враховуючи напівадитивність модуля сімей кривих, отримаємо: $$\label{eq4B}
M(\Gamma_k)\leqslant
\frac{4^nM_0}{\varepsilon_1^n}\int\limits_{D^{\,\prime}}Q(y)\,dm(y)\leqslant
c\cdot M_0<\infty\,.$$ З іншого боку, оскільки за умовою область $D$ має слабо плоску межу, з огляду на умову (\[eq2B\]), ми отримаємо, що $M(\Gamma_k)\rightarrow\infty$ при $k\rightarrow\infty,$ що суперечить (\[eq4B\]). Отримане протиріччя доводить лему.
Доведення теореми \[th2\]
=========================
Доведемо теорему \[th2\] від супротивного. Припустимо, що ${\frak
S}_{a, b, Q}(\overline{D}, \overline{D^{\,\prime}})$ не є одностайно неперервною в деякій точці $x_0\in\partial D.$ Тоді знайдуться точки $x_m\in D$ і відображення $f_m\in {\frak S}_{a, b, Q}(\overline{D},
\overline{D^{\,\prime}}),$ $m=1,2,\ldots ,$ такі що $x_m\rightarrow
x_0$ при $m\rightarrow\infty$ і, причому, при деякому $\varepsilon_0>0$ $$\label{eq15}
h(f_m(x_m), f_m(x_0))\geqslant\varepsilon_0\,,\quad m=1,2,\ldots\,.$$ Оберемо довільним чином точку $y_0\in D^{\,\prime},$ $y_0\ne b,$ і з’єднаємо її з точкою $b$ деякої кривою в $D^{\,\prime},$ яку ми позначимо $\alpha.$ Покладемо $A:=|\alpha|.$ Нехай $A_m$ – повне підняття кривої $\alpha$ при відображенні $f_m$ з початком в точці $a$ (воно існує за ). Зауважимо, що $h(A_m,
\partial D)>0$ за замкненістю відображення $f_m$ (бо, зокрема, відкриті дискретні і замкнені відображення є таким, прообраз компакту при яких є компактом, див. ). Тепер можливі наступні випадки: або $h(A_m)\rightarrow 0$ при $m\rightarrow\infty,$ або $h(A_{m_k})\geqslant\delta_0>0$ при $k\rightarrow\infty$ для деякої зростаючої послідовності номерів $m_k$ і деякого $\delta_0>0.$
У першому з цих випадків, очевидно, $h(A_m, \partial D)\geqslant
\delta>0$ при деякому $\delta>0.$ Тоді сім’я відображень $\{f_m\}_{m=1}^{\infty}$ одностайно неперервна в точці $x_0$ за теоремою 1 в [@Sev], що суперечить умові (\[eq15\]).
У другому випадку, якщо $h(A_{m_k})\geqslant\delta_0>0$ при $k\rightarrow\infty,$ ми також маємо, що $h(A_{m_k}, \partial
D)\geqslant \delta_1>0$ при деякому $\delta_1>0$ за лемою \[lem3\]. Знову ж таки, теоремою 1 в [@Sev] сім’я $\{f_{m_k}\}_{k=1}^{\infty}$ є одностайно неперервною в точці $x_0,$ і це суперечить умові (\[eq15\]).
Отже, в обох з двох можливих випадків ми прийшли до протиріччя з (\[eq15\]), і це вказує на невірність припущення про відсутність одностайної неперервності сім’ї ${\frak S}_{a, b, Q}(D,
D^{\,\prime})$ в $\overline{D}.$ Теорема доведена. $\Box$
\[ex1\] Розглянемо сім’ю відображень $f_n(z)=z^n,$ $n=1,2,\ldots \,,$ $z\in
{\Bbb B}^2=\{z\in {\Bbb C}: |z|<1\}.$ Зауважимо, що $f_n$ є відображеннями з обмеженим спотворенням як гладкі відображення, дилатація котрих дорівнює одиниці. Отже, $f_n$ задовольняє нерівність (\[eq2\*A\]) при $Q(z)=N(f_n, {\Bbb B}^2)=n,$ де, як звично, $N$ – функція кратності, визначена співвідношеннями $$N(y, f, {\Bbb B}^2)\,=\,{\rm
card}\,\left\{z\in {\Bbb B}^2: f(z)=y\right\}\,, \quad N(f, {\Bbb
B}^2)\,=\,\sup\limits_{y\in{\Bbb C}}\,N(y, f, {\Bbb B}^2)$$ (див., напр., або [@Ri теорема 6.7.II]). Всі відображення $f_n$ є дискретними, що перевіряється безпосередньо, крім того, зберігають межу одиничного круга і тому є замкненими (див., напр., ). Відображення $f_n$ також фіксують точку $0,$ тому вони задовольняють всі умови теореми \[th2\], але в той самий час у нерівності (\[eq2\*A\]) нема спільної інтегровної функції $Q,$ яка б забезпечувала всю сім’ю відображень $f_n,$ $n=1,2,\ldots .$ Внаслідок цього, сім’я відображень $f_n$ не є одностайно неперервною на межі одиничного круга, що перевіряється шляхом безпосередніх обчислень.
Аналогічний приклад можна побудувати у просторі. Нехай $x\in {\Bbb
B}^n,$ $x=(r\cos\varphi, r\sin\varphi, x_3,x_4,\ldots\,x_n),$ де, як звично, $x_1=r\cos\varphi,$ $x_2=r\sin\varphi,$ $0\leqslant\varphi<
2\pi,$ $0\leqslant r<\infty.$ Для натурального $m\in {\Bbb N}$ покладемо $f_m(x)=(r\cos m\varphi, r\sin m\varphi,
x_3,x_4,\ldots\,x_n).$ Шляхом безпосередніх обчислень можна переконатися, що $K_O(x, f_m)=m^{n-1}$ (див., напр., [@Re приклад 3 пункту 4.3.I]). Зауважимо, що $f_m$ є відображеннями з обмеженим спотворенням як гладкі відображення в ${\Bbb R}^n\setminus {\Bbb
R}^{n-2},$ де ${\Bbb R}^{n-2}=\{x\in {\Bbb R}^n: x_{n-1}=x_n=0\},$ дилатація котрих дорівнює $m^{n-1}.$ Отже, $f_m$ задовольняє нерівність (\[eq2\*A\]) при $Q(x)=N(f_m, {\Bbb B}^n)\cdot
m^{n-1}=m^n$ в області ${\Bbb B}^n$ (див., напр., або [@Ri теорема 6.7.II]). Всі відображення $f_m$ є дискретними, що перевіряється безпосередньо, крім того, зберігають межу одиничної кулі і тому замкнені (див., напр., ). Відображення $f_m$ також фіксують точку $0,$ але не мають спільної мажоранти $Q$ в (\[eq2\*A\]). Неважко переконатися, що сім’я відображень $\{f_m\}_{m=1}^{\infty}$ не є одностайно неперервною на одиничній сфері.
[99]{}
Gutlyanskii V. Ya., Ryazanov V. I., Yakubov E., *The Beltrami equations and prime ends*, Український математичний вiсник, **12**, № 1, 2015, p. 27–-66; translation *The Beltrami equations and prime ends*, J. Math. Sci. (N.Y.), **210**, no. 1, 2015, p. 22–-51.
Ковтонюк Д.А., Рязанов В.И., *К теории простых концов для пространственных областей*, Укр. мат. журнал, **67**, № 4, 2015, с. 467–479; translation ”On the theory of prime ends for space mappings”, Ukrainian Math. J., **67**, no. 4, 2015, p. 528–-541.
Kovtonyuk D.A., Ryazanov V.I., *Prime ends and Orlicz-Sobolev classes*, St. Petersburg Math. J., **27**, no. 5, 2016, p. 765–788.
Martio O., Rickman S., and Väisälä J., *Definitions for quasiregular mappings*, Ann. Acad. Sci. Fenn. Ser. A1., **448**, 1969, p. 1–40.
Martio O., Ryazanov V., Srebro U. and Yakubov E., *On $Q$-homeomorphisms*, Ann. Acad. Sci. Fenn. Ser. A1, **30**, no. 1, 2005, p. 49–69.
Martio O., Ryazanov V., Srebro U. and Yakubov E., *Moduli in Modern Mapping Theory.* – New York: Springer Science + Business Media, LLC, 2009.
Näkki R., *Prime ends and quasiconformal mappings*, J. Anal. Math., **35**, 1979, p. 13–40.
Reshetnyak Yu.G., *Space mappings with bounded distortion.* – Transl. Math. Monographs 73, AMS, 1989.
Rickman S., *Quasiregular mappings.* – Berlin: Springer-Verlag, 1993.
Väisälä J., *Lectures on $n$-Dimensional Quasiconformal Mappings*, Lecture Notes in Math. **229**, Berlin etc.: Springer–Verlag, 1971.
Vuorinen M., *Exceptional sets and boundary behavior of quasiregular mappings in $n$-space*, Ann. Acad. Sci. Fenn. Ser. A 1. Math. Dissertationes, **11**, 1976, p. 1–44.
Севостьянов Е.А., Скворцов С.А., *О сходимости отображений в метрических пространствах с прямыми и обратными модульными условиями*, Укр. мат. журнал, **70**, № 7, 2018, с. 952–967; transl. Sevost’yanov E.A., Skvortsov S.A., *On the Convergence of Mappings in Metric Spaces with Direct and Inverse Modulus Conditions*, Ukr. Math. J., **70,** no. 7, 2018, p. 1097–1114.
Севостьянов Е.А., Скворцов С.А., *О локальном поведении одного класса обратных отображений,* Укр. мат. вестник, **15**, № 3, 2018, с. 399–417; translation Sevost’yanov E.A., Skvortsov S.A., *On the local behavior of a class of inverse mappings*, J. Math. Sci., **241**, no. 1, 2019, p. 77–89.
Sevost’yanov, E.A. and Skvortsov S.A., *On mappings whose inverse satisfy the Poletsky inequality*, Ann. Acad. Scie. Fenn. Math., **45**, 2020, p. 259–277.
Sevost’yanov E.A., Skvortsov S.O., Dovhopiatyi O.P., *On mappings satisfying the inverse Poletsky inequality* // www. arxiv. org, arXiv:1904.01513.
Ильютко Д.П., Севостьянов Е.А., *О простых концах на римановых многообразиях*, Укр. мат. вестник, **15**, № 3, 2018, c. 358–392; translation *On prime ends on Riemannian manifolds* in J. Math. Sci., **241**, no. 1, 2019, p. 47–63.
Голузин Г.М., *Геометрическая теория функций комплексного переменного.* – М.: Наука, ФИЗМАТГИЗ, 1966.
Севостьянов Е.А., *Межове продовження відображень з оберненою нерівністю Полецького по простих кінцях*, Укр. мат. журнал (подана до друку).
Herron J. and Koskela P., *Quasiextremal distance domains and conformal mappings onto circle domains*, Compl. Var. Theor. Appl., **15**, 1990, p. 167–179.
Куратовский К., *Топология*, т. 2. – М.: Мир, 1969.
КОНТАКТНА ІНФОРМАЦІЯ
[[**Євген Олександрович Севостьянов**]{}\
[**1.**]{} Житомирський державний університет ім. І. Франко\
кафедра математичного аналізу, вул. Велика Бердичівська, 40\
м. Житомир, Україна, 10 008\
[**2.**]{} Інститут прикладної математики і механіки НАН України,\
вул. Добровольського, 1\
м. Слов’янськ, Україна, 84 100\
e-mail: [email protected]]{}
[[**Наталія Сергіївна Ількевич**]{}\
Житомирський державний університет ім. І. Франко\
кафедра фізики і охорони праці, вул. Велика Бердичівська, 40\
м. Житомир, Україна, 10 008\
e-mail: [email protected] ]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper considers multiaccess multiple-input multiple-output (MIMO) systems with finite rate feedback. The goal is to understand how to efficiently employ the given finite feedback resource to maximize the sum rate by characterizing the performance analytically. Towards this, we propose a joint quantization and feedback strategy: the base station selects the strongest users, jointly quantizes their strongest eigen-channel vectors and broadcasts a common feedback to all the users. This joint strategy is different from an individual strategy, in which quantization and feedback are performed across users independently, and it improves upon the individual strategy in the same way that vector quantization improves upon scalar quantization. In our proposed strategy, the effect of user selection is analyzed by extreme order statistics, while the effect of joint quantization is quantified by what we term “the composite Grassmann manifold”. The achievable sum rate is then estimated by random matrix theory. Due to its simple implementation and solid performance analysis, the proposed scheme provides a benchmark for multiaccess MIMO systems with finite rate feedback.'
author:
- |
Wei Dai$^{\dagger}$, Brian C. Rider$^{\dagger\dagger}$ and Youjian(Eugene) Liu$^{\dagger}$\
$^{\dagger}$Department of Electrical and Computer Engineering, University of Colorado at Boulder\
$^{\dagger\dagger}$Department of Mathematics, University of Colorado at Boulder
bibliography:
- 'Bib/\_Heath.bib'
- 'Bib/\_love.bib'
- 'Bib/\_Rao.bib'
- 'Bib/\_Tse.bib'
- 'Bib/FeedbackMIMO\_append.bib'
- 'Bib/MIMO\_basic.bib'
- 'Bib/\_Dai.bib'
- 'Bib/RandomMatrix.bib'
- 'Bib/\_Verdu.bib'
- 'Bib/Books.bib'
- 'Bib/Multi\_Access.bib'
- 'Bib/Math.bib'
- 'Bib/\_Jindal.bib'
- 'Bib/\_BC\_Feedback.bib'
title: 'Joint Beamforming for Multiaccess MIMO Systems with Finite Rate Feedback$^{*}$'
---
\[sec:Introduction\]Introduction
================================
This paper considers multiaccess systems, corresponding to the uplink of cellular systems, where both the base station and the multiple users are equipped with multiple antennas. Multiple antenna systems, also known as multiple-input multiple-output (MIMO) systems, provide significant benefit over single antenna systems in terms of increased spectral efficiency and/or reliability. The full potential of MIMO though requires perfect channel state information (CSI) at both the transmitter and the receiver. While it is often reasonable to assume that the receiver has perfect CSI through a pilot signal, assuming perfect CSI at the transmitter (CSIT) is typically unrealistic. In many practical systems, the transmitter obtains CSI through a finite rate feedback from the receiver. Note that a wireless fading channel may have infinitely many channel states, and a finite rate feedback implies that CSIT is imperfect. One expects a performance degradation, and here we focus on the quantitative effect of finite rate feedback and the corresponding design.
Insight from single user MIMO systems with finite rate feedback proves beneficial. Single user systems are similar to multiaccess systems in the sense that there is only one receiver in both systems. The receiver knows the channel states perfectly and helps transmitters adapt their signals to maximize throughput. The essential difference between these two types of systems lies in the modes of antenna cooperation. In single user MIMO systems, all the transmit antennas are able to cooperate in sending a given message. In multiaccess systems, different users have independent messages, and transmit antennas belonging to one user cannot aid the transmission of another user’s message. Due to this additional constraint, the analysis and design of multiaccess systems becomes more complicated. Still, we will borrow insight from single user systems to simplify the design of multiaccess systems. For single user MIMO systems, strategies to maximize throughput with perfect CSIT and without CSIT are derived and analyzed in [@Telatar_EuroTele99_Capacity_MIMO]. When only finite rate feedback is available, the focus has moved toward the development of suboptimal strategies as a simplification. The dominant approach is based on power on/off strategy, in which a data stream is either turned on with a pre-determined constant power or turned off (zero power). Systems with only one stream are considered in [@Sabharwal_IT03_Beamforming_MIMO; @Love_IT03_Grassman_Beamforming_MIMO; @Rao_SP07_Feedback_High_Resolution]. Systems with multiple independent streams are investigated in [@Honig_Allerton03_Benefits_Limited_Feedback_Wireless_Channels; @Rao_icc05_MIMO_spatial_multiplexing_limit_feedback; @Love_IT2005_limited_feedback_unitary_precoding; @Heath_ICASSP05_Quantization_Grassmann_Manifold; @Dai_ISIT05_Power_onoff_strategy_design; @Dai_05_Power_onoff_strategy_design_finite_rate_feedback; @Dai_IT2008_Quantization_Grassmannian_manifold]. It appears that power on/off strategy is near optimal compared to the optimal power water-filling allocation [@Dai_05_Power_onoff_strategy_design_finite_rate_feedback].
We aim to understand how to efficiently employ the given finite feedback resource to maximize the sum rate by characterizing performance analytically. The full multiaccess MIMO problem still appears behind reach mathematically and is left for the future. In this paper, we propose a *suboptimal* strategy by borrowing insight and methods from single user systems. Specifically, the base station selects the strongest users, jointly quantizes their strongest eigen-channel vectors and broadcasts a common feedback to all the users. Instead of designing a specific quantization code book, we show that the performance of a random code book is optimal in probability. After receiving feedback information, a selected on-user employ power on/off strategy and transmit along the beamforming vector selected by the feedback. Here, joint quantization and feedback are employed based on the plain fact that vector quantization is better than scalar quantization [@Cover_Elements_Information_Theory Ch. 13]. (The precise gain will be verified empirically.) It is also worth noting that, as we shall discuss in Section \[sec:Suboptimal-Feedback-Strategies\] and \[sec:Simulations-and-Discussion\], antenna selection can be viewed as a simplified version of the proposed scheme.
This approach differs from the ongoing research for broadcast channels (BC) with finite rate feedback. While there is a well known duality between broadcast and multiaccess systems [@Vishwanath_IT2003_Duality_BC_MAC], this duality requires full CSI at both the transmitters and the receivers and is not available when only partial CSIT is provided. When CSIT is available only through finite rate feedback, broadcast systems suffer from the so called interference domination phenomenon [@Sharif_IT05_MIMO_BC_Feedback; @Jindal_IT06_BC_Feedback]. The major effort in research is to limit the interference among users. Sharif and Hassibi select the near orthogonal channels when the number of users is sufficiently large [@Sharif_IT05_MIMO_BC_Feedback; @Jindal_IT06_BC_Feedback]. As the number of users is comparable to the number of antennas at the base station, Jindal shows that the feedback rate should be proportional to signal-to-noise ratio (SNR) if the number of users turned on is fixed [@Jindal_IT06_BC_Feedback], while we show that the number of users should be adapt to the SNR if the feedback rate is given [@Dai_CISS2007_broadcast_multiaccess_channels_single_antenna_users]. However, the interference domination phenomenon does not appear in multiaccess systems. Note that the search of near orthogonal channels suffers from exponential increasing complexity. Neither the results nor the methods for broadcast systems can be directly applied to the problem discussed in this paper.
Though the strategy in this paper is relatively simple, the corresponding performance analysis is nontrivial. Our main analytical result is an upper bound on the sum rate, which to our knowledge is the best to date. The effect of user/antenna selection is analyzed by extreme order statistics, and the effect of eigen-channel vectors joint quantization is quantified via *the composite Grassmann manifold*. Interestingly, the complicated effect of imperfect CSIT and feedback is eventually described by a single constant, which we term *the power efficiency factor*. Successful evaluation of the power efficiency factor enables us characterize the upper bound on the sum rate. The anticipated goodness of the upper bound is supported by simulation of several systems with a large range of SNRs.
The rest of this paper is organized as follows. The general model for multiaccess systems with finite rate feedback is described in Section \[sec:System-Model\]. The mathematical results developed for performance analysis are assembled in Section \[sec:Mathematical-Results\]. The antenna selection strategy is analyzed in Section \[sub:Antenna-Selection\]. Then a suboptimal strategy is proposed and analyzed in Section \[sub:General-Beamforming-Strategy\]. In Section \[sec:Simulations-and-Discussion\], simulation results are presented and discussed. Finally, Section \[sec:Conclusion\] summarizes the paper.
\[sec:System-Model\]System Model
================================
Assume that there are $L_{R}$ antennas at the base station and $N$ users communicating with the base station. Assume that the user $i$[^1] has $L_{T,i}$ transmit antennas $1\leq i\leq N$. Throughout we will set $L_{T,1}=\cdots=L_{T,N}=L_{T}$. The signal transmission model is $$\mathbf{Y}=\sum_{i=1}^{N}\mathbf{H}_{i}\mathbf{T}_{i}+\mathbf{W},$$ where $\mathbf{Y}\in\mathbb{C}^{L_{R}\times1}$ is the received signal at the base station, $\mathbf{H}_{i}\in\mathbb{C}^{L_{R}\times L_{T}}$ is the channel state matrix for user $i$, $\mathbf{T}_{i}\in\mathbb{C}^{L_{T}\times1}$ is the transmitted Gaussian signal vector for user $i$ and $\mathbf{W}\in\mathbb{C}^{L_{R}\times1}$ is the additive Gaussian noise vector with zero mean and covariance matrix $\mathbf{I}_{L_{R}}$. We assume the Rayleigh fading channel model: the entries of $\mathbf{H}_{i}$’s are independent and identically distributed (i.i.d.) circularly symmetric complex Gaussian variables with zero mean and unit variance ($\mathcal{CN}\left(0,1\right)$), and $\mathbf{H}_{i}$’s are independent across $i$.
We further assume that there exists a feedback link from the base station to the users. At the beginning of each channel use, the channel states $\mathbf{H}_{i}$’s are perfectly estimated at the receiver (the base station). This assumption is valid in practice since most communication standards allow the receiver to learn the channel states from pilot signals. A common message, which is a function of the channel states, is sent back to all users through the feedback link. We assume that the feedback link is rate limited and error-free. The feedback directs the users to choose their Gaussian signal covariance matrices. In a multiaccess communication system, different users cannot cooperate in terms of information message, leading to $\mathrm{E}\left[\mathbf{T}_{i}\mathbf{T}_{j}^{\dagger}\right]=\mathbf{0}$ for $i\ne j$. Let $\mathbf{T}=\left[\mathbf{T}_{1}^{\dagger}\cdots\mathbf{T}_{N}^{\dagger}\right]^{\dagger}$ be the overall transmitted Gaussian signal for all users and $\mathbf{\Sigma}\triangleq\mathrm{E}\left[\mathbf{T}\mathbf{T}^{\dagger}\right]$ be the overall signal covariance matrix. Then $\mathbf{\Sigma}$ is an $NL_{T}\times NL_{T}$ block diagonal matrix whose $i^{\mathrm{th}}$ diagonal block is the $L_{T}\times L_{T}$ covariance matrix $\mathrm{E}\left[\mathbf{T}_{i}\mathbf{T}_{i}^{\dagger}\right]$. Let $\mathbf{H}=\left[\mathbf{H}_{1}\mathbf{H}_{2}\cdots\mathbf{H}_{N}\right]$ be the overall channel state matrix. An extension of [@Lau_IT04_Capacity_Memoryless_Block_Fading] shows that the optimal feedback strategy is to feedback the index of an appropriate covariance matrix, which is a function of current channel state $\mathbf{H}$. Last, assume that there is a covariance matrix codebook $\mathcal{B}_{\mathbf{\Sigma}}=\left\{ \mathbf{\Sigma}_{1},\cdots,\mathbf{\Sigma}_{K_{\mathcal{B}}}\right\} $ (with finite size) declared to both the base station and the users, where each $\mathbf{\Sigma}_{k}\in\mathcal{B}_{\mathbf{\Sigma}}$ is the overall signal covariance matrix with block diagonal structure just described, and $K_{\mathcal{B}}$ is the size of the codebook. The feedback function $\varphi$ is a map from $\left\{ \mathbf{H}\in\mathbb{C}^{L_{R}\times NL_{T}}\right\} $ onto the index set $\left\{ 1,\cdots,K_{\mathcal{B}}\right\} $. Subjected to this finite rate feedback constraint$$\left|\mathcal{B}_{\mathbf{\Sigma}}\right|=K_{\mathcal{B}}$$ and the average total transmission power constraint$$\mathrm{E}_{\mathbf{H}}\left[\mathrm{tr}\left(\mathbf{\Sigma}_{\varphi\left(\mathbf{H}\right)}\right)\right]\leq\rho,$$ the sum rate of the optimal feedback strategy is given by $$\underset{\mathcal{B}_{\mathbf{\Sigma}}}{\sup}\;\underset{\varphi\left(\cdot\right)}{\sup}\;\mathrm{E}_{\mathbf{H}}\left[\log\left|\mathbf{I}_{L_{R}}+\mathbf{H}\mathbf{\Sigma}_{\varphi\left(\mathbf{H}\right)}\mathbf{H}^{\dagger}\right|\right].\label{eq:sum_rate_optimal}$$ Here, since only symmetric systems are concerned, the total power constraint $\rho$ is equivalent to individual power constraint $\rho/N$. Note that the optimal strategy involves two coupled optimization problems. It is difficult, if not impossible, to find its explicit form and performance. Instead, we shall study two suboptimal strategies and characterize their sum rates in Section \[sec:Suboptimal-Feedback-Strategies\].
\[sec:Mathematical-Results\]Preliminaries
=========================================
This section assembles mathematical results required for later analysis. The reader may proceed directly to Section \[sec:Suboptimal-Feedback-Strategies\] for the main engineering results.
\[sub:Extreme-Chi2\]Order Statistics for Chi-Square Random Variables
--------------------------------------------------------------------
Define $X_{i}=\sum_{j=1}^{L}\left|h_{i,j}\right|^{2}$ where $h_{i,j}\;1\leq j\leq L,\;1\leq i\leq n$ are i.i.d. $\mathcal{CN}\left(0,1\right)$. Then each $X_{i}$ has a Chi-square distribution with probability density functions (PDF) $$f_{X}\left(x\right)=\frac{1}{\left(L-1\right)!}x^{L-1}e^{-x}.$$ Denote the corresponding cumulative distribution function (CDF) by $F_{X}\left(x\right)$. Next introduce the order statistics for these variables: that is the non-decreasing list $X_{\left(1:n\right)}\leq X_{\left(2:n\right)}\leq\cdots\leq X_{\left(n:n\right)}$ connected with each realization. Here, the subscript $\left(k:n\right)$ indicates that $X_{\left(k:n\right)}$ is the $k^{\mathrm{th}}$ minima. (We follow the convention of [@JanosGalambos1987_extreme_order_statistics].) Note of course that ties occur with probability zero and can be broken arbitrarily.
We will need the following, which is proved in Appendix \[sub:Proof-of-Extreme-Order-Statistics\].
\[lem:Expectation-extreme-chi2\]With the notation set out above, for any fixed positive integer $s$ it holds $$\underset{n\rightarrow+\infty}{\lim}\;\mathrm{E}\left[\frac{\sum_{k=1}^{s}X_{\left(n-k+1:n\right)}-sa_{n}}{b_{n}}\right]=s\left(\mu_{1}+1-\sum_{k=1}^{s}\frac{1}{k}\right),\label{eq:Expectation-extreme-chi2}$$ where $$a_{n}=\inf\left\{ x:\;1-F_{X}\left(x\right)\leq\frac{1}{n}\right\} ,$$ $$b_{n}=\frac{\sum_{i=0}^{L-1}\frac{L-i}{i!}a_{n}^{i}}{\sum_{i=0}^{L-1}\frac{1}{i!}a_{n}^{i}},$$ and $\mu_{1}=\int_{-\infty}^{+\infty}xde^{-e^{-x}}$ may be computed numerically.
The limiting result in expectation immediately provides the following approximation for a fixed $s$: $$\mathrm{E}\left[\sum_{k=1}^{s}X_{\left(n-k+1:n\right)}-sa_{n}\right]=sb_{n}\left(\mu_{1}+1-\sum_{i=1}^{s}\frac{1}{i}\right)\left(1+o\left(1\right)\right).\label{eq:approx-order-statistics}$$ The shape of $F_{X}$ guarantees that $a_{n}$ and so $b_{n}$ are finite for any fixed $n$ but tend to infinity and one respectively with this parameter.
\[sub:Conditioned-Eigen\]Conditioned Eigenvalues of the Wishart Matrix
----------------------------------------------------------------------
Let $\mathbf{H}\in\mathbb{L}^{n\times m}$ be a random $n\times m$ matrix whose entries are i.i.d. Gaussian random variables with zero mean and unit variance, where $\mathbb{L}$ is either $\mathbb{R}$ or $\mathbb{C}$. Throughout, we refer to $\mathbf{H}$ as the standard Gaussian random matrix. Let $\lambda_{1}\ge\lambda_{2}\ge\cdots\ge\lambda_{n}$ be the ordered eigenvalues of $\mathbf{W}=\mathbf{H}\mathbf{H}^{\dagger}$ ($\mathbf{W}$ is Wishart distributed [@Muirhead_book82_multivariate_statistics]).
This subsection takes up an estimate of $\mathrm{E}\left[\left.\lambda_{1}\right|\mathrm{tr}\left(\mathbf{W}\right)\right]$, where $\mathrm{tr}\left(\cdot\right)$ is the usual matrix trace. In particular, while a closed formula for this object would be rather involved, we may use random matrix theory to obtain an approximation. The first ingredient is the following.
\[lem:conditional-expectation-Wishart\]Let $\mathbf{H}\in\mathbb{L}^{n\times m}$ (w.l.o.g. $n\le m$)[^2] be a standard random Gaussian matrix. Let $\lambda_{1}\ge\lambda_{2}\ge\cdots\ge\lambda_{n}$ be the ordered eigenvalues of $\mathbf{W}=\mathbf{H}\mathbf{H}^{\dagger}$. Then$$\mathrm{E}\left[\lambda_{i}|\mathrm{tr}\left(\mathbf{W}\right)=c\right]=\zeta_{i}c,$$ where$$\zeta_{i}=\mathrm{E}\left[\lambda_{i}|\mathrm{tr}\left(\mathbf{W}\right)=1\right].\label{eq:zeta-definition}$$ $$,\label{eq:zeta-definition}$$ $\beta=1$ if $\mathbb{L}=\mathbb{R}$ or $\beta=2$ if $\mathbb{L}=\mathbb{C}$, and $\left|\Delta_{n}\left(\bm{\lambda}\right)\right|=\prod_{i<j}^{n}\left(\lambda_{i}-\lambda_{j}\right)$.
The joint density of the ordered eigenvalues of $\mathbf{W}$ is known to be $$K_{m,n,\beta}^{-1}e^{-\frac{\beta}{2}\sum_{i}\lambda_{i}}\prod_{i=1}^{m}\lambda_{i}^{\frac{\beta}{2}\left(n-m+1\right)-1}\left|\Delta_{n}\left(\bm{\lambda}\right)\right|^{\beta},$$ where $\lambda_{1}\geq\cdots\geq\lambda_{m}\geq0$, $\left|\Delta_{n}\left(\bm{\lambda}\right)\right|=\prod_{i<j}^{n}\left(\lambda_{i}-\lambda_{j}\right)$, $$\beta=\begin{cases}
1 & \mathrm{if}\;\mathbb{L}=\mathbb{R}\\
2 & \mathrm{if}\;\mathbb{L}=\mathbb{C}\end{cases},$$ and $K_{m,n,\beta}$ is a normalizing factor ([@Muirhead_book82_multivariate_statistics pg. 107] for the real case and [@Telatar_EuroTele99_Capacity_MIMO] for the complex case). Write out the formula for $\mathrm{E}\left[\lambda_{i}|\sum_{i=1}^{m}\lambda_{i}=c\right]$ and use the variable change $\lambda_{i}'=\frac{\lambda_{i}}{c}$. After some elementary calculations, it can be shown that $$\zeta_{i}=\frac{\int_{\underset{\lambda_{1}\geq\cdots\geq\lambda_{n}}{\sum\lambda_{j}=1}}\lambda_{i}\prod_{j=1}^{n}\lambda_{j}^{\frac{\beta}{2}\left(m-n+1\right)-1}\left|\Delta_{n}\left(\bm{\lambda}\right)\right|^{\beta}\prod_{j=1}^{n}d\lambda_{j}}{\int_{\underset{\lambda_{1}\geq\cdots\geq\lambda_{n}}{\sum\lambda_{j}=1}}\prod_{j=1}^{n}\lambda_{j}^{\frac{\beta}{2}\left(m-n+1\right)-1}\left|\Delta_{n}\left(\bm{\lambda}\right)\right|^{\beta}\prod_{j=1}^{n}d\lambda_{j}}=\mathrm{E}\left[\lambda_{i}|\mathrm{tr}\left(\mathbf{W}\right)=1\right].$$
Given the preceding observation, we require an estimate for $\zeta_{1}$ in (\[eq:zeta-definition\]). For this we turn to the asymptotic behavior of the ordered eigenvalues.
\[lem:condition-expectation-asymptotics\]Let $\lambda_{1}\ge\lambda_{2}\ge\cdots\ge\lambda_{n}$ be the ordered eigenvalues of $\frac{1}{m}\mathbf{H}\mathbf{H}^{\dagger}$, where $\mathbf{H}\in\mathbb{L}^{n\times m}$ $\left(\mathbb{L}=\mathbb{R}\;\mathrm{or}\;\mathbb{C}\right)$ is a standard random Gaussian matrix. As $n,m\rightarrow\infty$ with $\frac{m}{n}\rightarrow\bar{m}\in\mathbb{R}^{+}$, for a given $\tau\in\left(0,\min\left(1,\bar{m}\right)\right)$, $$\begin{aligned}
\bar{\zeta}_{\tau} & :=\underset{\left(n,m\right)\rightarrow\infty}{\lim}\mathrm{E}\left[\frac{1}{n}\left(\sum_{1\le i\le n\tau}\lambda_{i}\right)\right]\\
& =\frac{\bar{m}}{2\pi}\left[\frac{1+\frac{1}{\bar{m}}-a}{2}\sqrt{\left(\lambda^{+}-a\right)\left(a-\lambda^{-}\right)}+\frac{2}{\bar{m}}\left(\frac{\pi}{2}+\sin^{-1}\frac{\sqrt{\bar{m}}\left(1+\frac{1}{\bar{m}}-a\right)}{2}\right)\right],\end{aligned}$$ where $\lambda^{+}=\left(1+\sqrt{\frac{1}{\bar{m}}}\right)^{2}$, $\lambda^{-}=\left(1-\sqrt{\frac{1}{\bar{m}}}\right)^{2}$, and $a\in\left(\lambda^{-},\lambda^{+}\right)$ satisfies $$\begin{aligned}
\tau & =\frac{\bar{m}}{2\pi}\left[-\sqrt{\left(\lambda^{+}-a\right)\left(a-\lambda^{-}\right)}+\frac{1+\bar{m}}{\bar{m}}\left(\frac{\pi}{2}+\sin^{-1}\frac{\sqrt{\bar{m}}\left(1+\frac{1}{\bar{m}}-a\right)}{2}\right)\right.\\
& \quad\quad\quad\left.-\frac{\left|\bar{m}-1\right|}{\bar{m}}\left(\frac{\pi}{2}-\sin^{-1}\frac{\sqrt{\bar{m}}}{2}\frac{\left(1+\frac{1}{\bar{m}}\right)a-\left(1-\frac{1}{\bar{m}}\right)^{2}}{a}\right)\right].\end{aligned}$$
This lemma is an extension of Theorem \[thm:truncate-function-RMT\] in Appendix \[sub:Random-Matrix-Theory\] with explicit evaluation of the integrals appearing in that statement.
Motivated by the observation that the expectation of a fixed fraction of the ordered eigenvalues converges to its limit quickly[@Dai_05_Power_onoff_strategy_design_finite_rate_feedback], we approximate $\zeta_{1}$ by $\bar{\zeta}_{\frac{1}{n}}$ for fixed finite $n$ and $m$.
\[sub:CGManifold\]The Grassmann Manifold and the Composite Grassmann Manifold
-----------------------------------------------------------------------------
As demonstrated in [@Dai_ISIT05_Power_onoff_strategy_design; @Dai_05_Power_onoff_strategy_design_finite_rate_feedback], the Grassmann manifold is closely related to eigen-channel vector quantization, and here we introduce the composite Grassmann manifold. The results developed here help quantify the effect of eigen-channel vector quantization in multiaccess systems (see Section \[sub:General-Beamforming-Strategy\] for details).
The *Grassmann manifold* $\mathcal{G}_{n,p}\left(\mathbb{L}\right)$ is the set of all $p$-dimensional planes (through the origin) in the $n$-dimensional Euclidean space $\mathbb{L}^{n}$, where $\mathbb{L}$ is either $\mathbb{R}$ or $\mathbb{C}$. A generator matrix $\mathbf{P}\in\mathbb{L}^{n\times p}$ for a plane $P\in\mathcal{G}_{n,p}\left(\mathbb{L}\right)$ is a matrix whose columns are orthonormal and span $P$. For a given $P\in\mathcal{G}_{n,p}\left(\mathbb{L}\right)$, its generator matrix is not unique: if $\mathbf{P}$ generates $P$ then $\mathbf{PU}$ also generates $P$ for any $p\times p$ orthogonal/unitary matrix $\mathbf{U}$ (with respect to $\mathbb{L}=\mathbb{R}/\mathbb{C}$ respectively) [@Conway_96_PackingLinesPlanes]. The chordal distance between two planes $P_{1},P_{2}\in\mathcal{G}_{n,p}\left(\mathbb{L}\right)$ can be defined by their generator matrices $\mathbf{P}_{1}$ and $\mathbf{P}_{2}$ via $$d_{c}\left(P_{1},P_{2}\right)=\sqrt{p-\mathrm{tr}\left(\mathbf{P}_{1}^{\dagger}\mathbf{P}_{2}\mathbf{P}_{2}^{\dagger}\mathbf{P}_{1}\right)}.$$ The isotropic measure $\mu$ on $\mathcal{G}_{n,p}\left(\mathbb{L}\right)$ is the Haar measure on $\mathcal{G}_{n,p}\left(\mathbb{L}\right)$[^3]. Let $O\left(n\right)$ and $U\left(n\right)$ be the sets of $n\times n$ orthogonal and unitary matrices respectively. Let $\mathbf{A}\in O\left(n\right)$ when $\mathbb{L}=\mathbb{R}$, or $\mathbf{A}\in U\left(n\right)$ when $\mathbb{L}=\mathbb{C}$. For any measurable set $\mathcal{M}\subset\mathcal{G}_{n,p}\left(\mathbb{L}\right)$ and arbitrary $\mathbf{A}$, $\mu$ satisfies $$\mu\left(\mathbf{A}\mathcal{M}\right)=\mu\left(\mathcal{M}\right).$$
Given above definitions, the distortion rate tradeoff on the Grassmann manifold is quantified in [@Dai_Globecom05_Quantization_bounds_Grassmann_manifold; @Dai_IT2008_Quantization_Grassmannian_manifold]. A quantization $\mathfrak{q}$ is a mapping from $\mathcal{G}_{n,p}\left(\mathbb{L}\right)$ to a discrete subset of $\mathcal{G}_{n,p}\left(\mathbb{L}\right)$, which is typically called a code $\mathcal{C}$. For the sake of application, the quantization $$\begin{aligned}
\mathfrak{q}:\;\mathcal{G}_{n,p}\left(\mathbb{L}\right) & \rightarrow\mathcal{C}\\
Q & \mapsto\mathfrak{q}\left(Q\right)=\arg\;\underset{P\in\mathcal{C}}{\min}\; d_{c}\left(P,Q\right)\end{aligned}$$ is of particular interest. Define the distortion metric on $\mathcal{G}_{n,p}\left(\mathbb{L}\right)$ as the squared chordal distance. Let $Q\in\mathcal{G}_{n,p}\left(\mathbb{L}\right)$ be isotropically distributed (the probability measure is the isotropic measure). For a given code $\mathcal{C}$, the distortion associated with this codebook is defined as $$D\left(\mathcal{C}\right)=\mathrm{E}_{Q}\left[\underset{P\in\mathcal{C}}{\min}\; d_{c}^{2}\left(P,Q\right)\right].$$ For a given code size $K$ where $K$ is a positive integer, the distortion rate function is $$D^{*}\left(K\right)=\underset{\mathcal{C}:\left|\mathcal{C}\right|=K}{\inf}\; D\left(\mathcal{C}\right).$$ In [@Dai_Globecom05_Quantization_bounds_Grassmann_manifold; @Dai_IT2008_Quantization_Grassmannian_manifold], we quantify the distortion rate function by constructing tight lower and upper bounds. The results are summarized as follows.
\[lem:DRF-GM\]Consider the distortion rate function on $\mathcal{G}_{n,p}\left(\mathbb{L}\right)$. Let $t=\beta p\left(n-p\right)$,$$\beta=\left\{ \begin{array}{ll}
1 & \mathrm{if}\;\mathbb{L}=\mathbb{R}\\
2 & \mathrm{if}\;\mathbb{L}=\mathbb{C}\end{array}\right.,$$ $$c_{n,p,p,\beta}=\begin{cases}
\frac{1}{\Gamma\left(\frac{t}{2}+1\right)}\prod_{i=1}^{p}\frac{\Gamma\left(\frac{\beta}{2}\left(n-i+1\right)\right)}{\Gamma\left(\frac{\beta}{2}\left(p-i+1\right)\right)} & \mathrm{if}\; p\le\frac{n}{2}\\
\frac{1}{\Gamma\left(\frac{t}{2}+1\right)}\prod_{i=1}^{n-p}\frac{\Gamma\left(\frac{\beta}{2}\left(n-i+1\right)\right)}{\Gamma\left(\frac{\beta}{2}\left(n-p-i+1\right)\right)} & \mathrm{otherwise}\end{cases}.$$ When $K$ is sufficiently large ($c_{n,p,p,\beta}^{-\frac{2}{t}}2^{-\frac{2\log_{2}K}{t}}\le1$ necessarily), $$\begin{aligned}
& \frac{t}{t+2}c_{n,p,p,\beta}^{-\frac{2}{t}}2^{-\frac{2\log_{2}K}{t}}\left(1+o\left(1\right)\right)\leq D^{*}\left(K\right)\nonumber \\
& \quad\quad\quad\leq\frac{2}{t}\Gamma\left(\frac{2}{t}\right)c_{n,p,p,\beta}^{-\frac{2}{t}}2^{-\frac{2\log_{2}K}{t}}\left(1+o\left(1\right)\right).\label{eq:DRF_bounds_GM}\end{aligned}$$
To analyze the joint quantization problem arising in multiaccess MIMO systems (see Section \[sub:General-Beamforming-Strategy\] for details), we introduce the *composite Grassmann manifold*. The $m$-composite Grassmann manifold $\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ is a Cartesian product of $m$ many $\mathcal{G}_{n,p}\left(\mathbb{L}\right)$’s. Denote $P^{\left(m\right)}$ an element in $\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$. Then $P^{\left(m\right)}$ can be written as $\left(P_{1},\cdots,P_{m}\right)$ where $P_{i}\in\mathcal{G}_{n,p}\left(\mathbb{L}\right)$ $1\leq i\leq m$. For any $P_{1}^{\left(m\right)},P_{2}^{\left(m\right)}\in\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$, the chordal distance between them is well defined by $$d_{c}\left(P_{1}^{\left(m\right)},P_{2}^{\left(m\right)}\right):=\sqrt{\sum_{j=1}^{m}d_{c}^{2}\left(P_{1,j},P_{2,j}\right)},$$ where $P_{1}^{\left(m\right)}=\left(P_{1,1},\cdots,P_{1,m}\right)$, $P_{2}^{\left(m\right)}=\left(P_{2,1},\cdots,P_{2,m}\right)$ and $P_{i,j}\in\mathcal{G}_{n,p}\left(\mathbb{L}\right)$ ($i=1,2$ and $j=1,2,\cdots,m$). The isotropic measure on $\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ can be induced from the isotropic measure on $\mathcal{G}_{n,p}\left(\mathbb{L}\right)$: it is just the product of the isotropic measures on the composed copies of $\mathcal{G}_{n,p}\left(\mathbb{L}\right)$.
One goal will be to characterize the distortion rate function on $\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$. By analogy with the above discussion let a code $\mathcal{C}$ be any discrete subset of $\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$, and consider the quantization function $$\mathfrak{q}\left(Q^{\left(m\right)}\right)=\underset{P^{\left(m\right)}\in\mathcal{C}}{\arg\;\min}\; d_{c}\left(P^{\left(m\right)},Q^{\left(m\right)}\right).\label{eq:quantization-fn-CGM}$$ Let the distortion metric on $\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ be the squared chordal distance. The distortion associated with $\mathcal{C}$ is given by $$D\left(\mathcal{C}\right)=\mathrm{E}_{Q^{\left(m\right)}}\left[\underset{P^{\left(m\right)}\in\mathcal{C}}{\min}\; d_{c}^{2}\left(P^{\left(m\right)},Q^{\left(m\right)}\right)\right],$$ where $Q^{\left(m\right)}\in\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ is isotropically distributed. For all $K\in\mathbb{Z}^{+}$, the distortion rate function is defined as $$D^{*}\left(K\right)=\underset{\mathcal{C}:\left|\mathcal{C}\right|=K}{\inf}\; D\left(\mathcal{C}\right).$$ The following theorem characterizes $D^{*}\left(K\right)$ on $\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$.
\[thm:DRF-CGM\]Consider the distortion rate function on $\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$. Let $t$, $c_{n,p,p,\beta}$ and $\beta$ be defined as in Lemma \[lem:DRF-GM\]. When $K$ is sufficiently large ($\frac{\Gamma^{\frac{2}{mt}}\left(mt+1\right)}{\Gamma^{\frac{2}{t}}\left(t+1\right)}c_{n,p,p,\beta}^{-\frac{2}{t}}2^{-\frac{2\log_{2}K}{mt}}\le1$ necessarily), $$\begin{aligned}
& \frac{mt}{mt+2}\left(\frac{\Gamma^{\frac{2}{mt}}\left(m\frac{t}{2}+1\right)}{\Gamma^{\frac{2}{t}}\left(\frac{t}{2}+1\right)}c_{n,p,p,\beta}^{-\frac{2}{t}}2^{-\frac{2\log_{2}K}{mt}}\right)\left(1+o\left(1\right)\right)\le D^{*}\left(K\right)\nonumber \\
& \quad\quad\quad\le\frac{2}{mt}\Gamma\left(\frac{2}{mt}\right)\left(\frac{\Gamma^{\frac{2}{mt}}\left(m\frac{t}{2}+1\right)}{\Gamma^{\frac{2}{t}}\left(\frac{t}{2}+1\right)}c_{n,p,p,\beta}^{-\frac{2}{t}}2^{-\frac{2\log_{2}K}{mt}}\right)\left(1+o\left(1\right)\right).\label{eq:DRF-CGM}\end{aligned}$$
The detailed proof is given in Appendix \[sub:proof\_DRF\_composite\_GM\], but we mention here that the upper bound is derived by calculating the average distortion of random codes, which turn out to be asymptotically optimal in probability. Further, the lower and upper bounds differ only in the constant factors: $\frac{mt}{mt+2}$ for the lower bound and $\frac{2}{mt}\Gamma\left(\frac{2}{mt}\right)$ for the upper bound. As $n,K\rightarrow\infty$ with $\frac{\log_{2}K}{n}\rightarrow r$, this discrepency vanishes and we precisely characterize the asymptotic distortion rate function.
\[thm:DRF-CGM-Asymptotics\]Fix $p$ and $m$. Let $n,K\rightarrow\infty$ with $\frac{\log_{2}K}{n}\rightarrow r$. If $r$ is sufficiently large ($mp2^{-\frac{2}{\beta mp}r}\le1$ necessarily), then $$\underset{\left(n,K\right)\rightarrow\infty}{\lim}D^{*}\left(K\right)=mp2^{-\frac{2}{\beta mp}r},$$ where $\beta=1$ if $\mathbb{L}=\mathbb{R}$, and $\beta=2$ if $\mathbb{L}=\mathbb{C}$. Furthermore, let $\mathcal{C}_{\mathrm{rand}}\subset\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ be a code random generated from the isotropic distribution and with size $K$. Then for $\forall\epsilon>0$, $$\underset{\left(n,K\right)\rightarrow\infty}{\lim}\Pr\left(D\left(\mathcal{C}_{\mathrm{rand}}\right)>mp2^{-\frac{2}{\beta mp}r}+\epsilon\right)=0.$$
The proof of this theorem follows from those in [@Dai_IT2008_Quantization_Grassmannian_manifold Theorem 3] and is omitted here.
This theorem provides a formula for the distortion rate function at finite $n$ and $K$: $$D^{*}\left(K\right)=\frac{2}{mt}\Gamma\left(\frac{2}{mt}\right)\left(\frac{\Gamma^{\frac{2}{mt}}\left(m\frac{t}{2}+1\right)}{\Gamma^{\frac{2}{t}}\left(\frac{t}{2}+1\right)}c_{n,p,p,\beta}^{-\frac{2}{t}}2^{-\frac{2\log_{2}K}{mt}}\right)\left(1+o\left(1\right)\right).\label{eq:DRF-CGM-approx}$$ By the asymptotic optimality of random codes, we have employed random codes for our analysis, and approximate the corresponding distortion rate function by ignoring the higher order terms behind (\[eq:DRF-CGM-approx\]).
\[sub:CGMatrix\]Calculations Related to Composite Grassmann Matrices
--------------------------------------------------------------------
A *composite Grassmann matrix* $\mathbf{P}^{\left(m\right)}$ is a generator matrix generating $P^{\left(m\right)}\in\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$, and we denote the set of composite Grassmann matrices by $\mathcal{M}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$. A composite Grassmann matrix $\mathbf{P}^{\left(m\right)}=\left[\mathbf{P}_{1}\cdots\mathbf{P}_{m}\right]\in\mathcal{M}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ generates a plane $P^{\left(m\right)}=\left(P_{1},\cdots,P_{m}\right)\in\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$, where $\mathbf{P}_{1},\cdots,\mathbf{P}_{m}$ are the generator matrices for $P_{1},\cdots,P_{m}$ respectively. Note that the generator matrix $\mathbf{P}_{i}$ for a plane $P_{i}\in\mathcal{G}_{n,p}\left(\mathbb{L}\right)$ is not unique. The composite Grassmann matrix $\mathbf{P}^{\left(m\right)}\in\mathcal{M}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ generating $P^{\left(m\right)}\in\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ is not unique either: let $\mathbf{U}^{\left(m\right)}$ is an arbitrary $pm\times pm$ block diagonal matrix whose $i^{\mathrm{th}}$ ($1\le i\le m$) diagonal block is a $p\times p$ orthogonal/unitary matrix (w.r.t. $\mathbb{L}=\mathbb{R}/\mathbb{C}$ respectively); if $\mathbf{P}^{\left(m\right)}$ generates $P^{\left(m\right)}$, then $\mathbf{P}^{\left(m\right)}\mathbf{U}^{\left(m\right)}$ generates $P^{\left(m\right)}$ as well. View $\mathcal{M}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ as a Cartesian product of $m$ many $\mathcal{M}_{n,p}^{\left(1\right)}\left(\mathbb{L}\right)$. Then the isotropic measure $\mu$ on $\mathcal{M}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ is simply the product of Haar measure on each composed $\mathcal{M}_{n,p}^{\left(1\right)}\left(\mathbb{L}\right)$’s. We say a $\mathbf{P}^{\left(m\right)}\in\mathcal{M}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ is isotropically distributed if the corresponding probability measure is the isotropic measure $\mu$.
Note now that we are interested in quantifying $\mathrm{E}\left[\log\left|\mathbf{I}+c\mathbf{P}^{\left(m\right)}\mathbf{P}^{\left(m\right)\dagger}\right|\right]$, for a constant $c\in\mathbb{R}^{+}$ and isotropically distributed $\mathbf{P}^{\left(m\right)}\in\mathcal{M}_{n,1}^{\left(m\right)}\left(\mathbb{C}\right)$. The asymptotic behavior of this quantify is derived by random matrix theory techniques.
\[thm:bds\_CGMatrix\]Let $\mathbf{P}^{\left(m\right)}\in\mathcal{M}_{n,1}^{\left(m\right)}\left(\mathbb{C}\right)$ be isotropically distributed. For all positive real numbers $c$, as $n,m\rightarrow\infty$ with $\frac{m}{n}\rightarrow\bar{m}\in\mathbb{R}^{+}$,$$\begin{aligned}
& \underset{\left(n,m\right)\rightarrow\infty}{\lim}\frac{1}{n}\mathrm{E}\left[\log\left|\mathbf{I}+c\mathbf{P}^{\left(m\right)}\mathbf{P}^{\left(m\right)\dagger}\right|\right]\nonumber \\
& =\log\left(1+c\bar{m}-\frac{1}{4}\mathcal{F}\left(c,\bar{m}\right)\right)+\bar{m}\log\left(1+c-\frac{1}{4}\mathcal{F}\left(c,\bar{m}\right)\right)-\frac{\mathcal{F}\left(c,\bar{m}\right)}{4c},\label{eq:CGM-Shannon-Transform}\end{aligned}$$ where $$\mathcal{F}\left(z,\bar{m}\right)=\left(\left(1+\lambda^{-}z\right)^{1/2}-\left(1+\lambda^{+}z\right)^{1/2}\right)^{2},$$ $\lambda^{+}=\left(1+\sqrt{1/\bar{m}}\right)^{2}$ and $\lambda^{-}=\left(1-\sqrt{1/\bar{m}}\right)^{2}$.
Let $\mathbf{H}\in\mathbb{C}^{n\times m}$ be a standard Gaussian matrix. Let $\mathbf{P}^{\left(m\right)}\in\mathcal{G}_{n,1}^{\left(m\right)}\left(\mathbb{C}\right)$ be isotropically distributed. As $n,m\rightarrow\infty$ with a positive ratio, the eigenvalue statistics of $\mathbf{P}^{\left(m\right)}\mathbf{P}^{\left(m\right)\dagger}$ and $\frac{1}{m}\mathbf{H}\mathbf{H}^{\dagger}$ are asymptotically the same. Indeed, the Raleigh-Ritz criteria shows that the discrepancy between corresponding eigenvalues of these two matrices is bounded (multiplicatively) above and below by the minimum and maximum column norms of $\frac{1}{m}\mathbf{H}\mathbf{H}^{\dagger}$, both of which converge to one almost surely. Thus,$$\underset{\left(n,m\right)\rightarrow\infty}{\lim}\frac{1}{n}\mathrm{E}\left[\log\left|\mathbf{I}+c\mathbf{P}^{\left(m\right)}\mathbf{P}^{\left(m\right)\dagger}\right|\right]=\underset{\left(n,m\right)\rightarrow\infty}{\lim}\frac{1}{n}\mathrm{E}\left[\log\left|\mathbf{I}+c\frac{m}{n}\frac{1}{m}\mathbf{H}\mathbf{H}^{\dagger}\right|\right].$$ Now, it is a basic result in random matrix theory [@Verdu_random_matrix_theory_wireless_communications Eq. (1.10)] (also see Appendix \[sub:Random-Matrix-Theory\]) that the empirical distribution of the eigenvalues of $\frac{1}{m}\mathbf{H}\mathbf{H}^{\dagger}$ converges to the Marcenko-Pastur law given by $$d\mu_{\lambda}=\left(1-\bar{m}\right)^{\dagger}\delta\left(\lambda\right)+\bar{m}\frac{\sqrt{\left(\lambda-\lambda^{-}\right)^{+}\left(\lambda^{+}-\lambda\right)^{+}}}{2\pi\lambda}\cdot d\lambda$$ almost surely, where $\left(z\right)^{+}=\max\left(0,z\right)$. Thus, $$\underset{\left(n,m\right)\rightarrow\infty}{\lim}\frac{1}{n}\mathrm{E}\left[\log\left|\mathbf{I}+c\mathbf{P}^{\left(m\right)}\mathbf{P}^{\left(m\right)\dagger}\right|\right]\rightarrow\int\log\left(1+c\bar{m}\lambda\right)\cdot d\mu_{\lambda}$$ since $\log\left(1+c\bar{m}\lambda\right)$ is a bounded continuous function on the spectral support. The resulting integral is evaluated in [@Verdu_IT1999_Spectral_efficiency_CDMA], and the proof is finished.
For finite $n$ and $m$, we substitute $\bar{m}=\frac{m}{n}$ into (\[eq:CGM-Shannon-Transform\]) to approximate $\frac{1}{n}\mathrm{E}\left[\log\left|\mathbf{I}+c\mathbf{P}^{\left(m\right)}\mathbf{P}^{\left(m\right)\dagger}\right|\right]$.
\[sec:Suboptimal-Feedback-Strategies\]Suboptimal Strategies and the Sum Rate
============================================================================
Given finite rate feedback, the optimal strategy (\[eq:sum\_rate\_optimal\]) involves two coupled optimization problems: one is with respect to the feedback function $\varphi$ and the other optimization is over all possible covariance matrix codebooks. The corresponding design and analysis are extremely complicated, and instead we study suboptimal power on/off strategies. Motivated by the near optimal power on/off strategy for single user MIMO systems [@Dai_ISIT05_Power_onoff_strategy_design; @Dai_05_Power_onoff_strategy_design_finite_rate_feedback], we assume:
[T1)]{}
: Power on/off strategy: The $i$th user covariance matrix is of the form $\mathbf{\Sigma}_{i}=P_{\mathrm{on}}\mathbf{Q}_{i}\mathbf{Q}_{i}^{\dagger}$, where $P_{\mathrm{on}}$ is a fixed positive constant to denote on-power and $\mathbf{Q}_{i}$ is the beamforming matrix for user $i$. Denote each column of $\mathbf{Q}_{i}$ an *on-beam* and the number of the columns of $\mathbf{Q}_{i}$ by $s_{i}$ ($0\leq s_{i}\leq L_{T}$), then $\mathbf{Q}_{i}^{\dagger}\mathbf{Q}_{i}=\mathbf{I}_{s_{i}}$. Here, $s_{i}$ is the number of data streams (or on-beams) for user $i$ ($s_{i}=0$ implies that the user $i$ is off). The user $i$ with $s_{i}>0$ is referred to as an on-user.
[T2)]{}
: Constant number of on-beams: Let $s=\sum_{i=1}^{N}s_{i}$, the total number of on-beams, be constant independent of the specific channel realization for a given SNR. With this assumption, $P_{\mathrm{on}}=\rho/s$.
Using a constant number of on-beams is motivated by the fact that it is asymptotically optimal to turn on a constant fraction of all eigen-channels as $L_{T},L_{R}\rightarrow\infty$ with a positive ratio, see [@Dai_05_Power_onoff_strategy_design_finite_rate_feedback] which also demonstrates the good performance of this strategy. While the number of on-beams is independent of channel realizations, it is a function of SNR. Realize though that typically SNR changes on a much larger time scale than block fading. Keeping the number of on-beams constant enables the base station to keep the feedback and decoding processing from one fading block to another, and therefore reduces complexity of real-world systems.
These two assumptions essentially add extra structure to the input covariance matrix $\mathbf{\Sigma}$. Given this structure, we propose a joint quantization and feedback strategy in Section \[sub:General-Beamforming-Strategy\], which we term “general beamforming strategy‘". As we shall see in Section \[sub:Comments\], antenna selection can be viewed as a special case of general beamforming. Due to the simplicity of antenna selection, we next discuss its main features.
\[sub:Antenna-Selection\]Antenna Selection
------------------------------------------
The antenna selection strategy is described as follows. Index all $NL_{T}$ antennas by $i$ ($i=1,\cdots,NL_{T}$). Then $$\mathbf{Y}=\sum_{i=1}^{NL_{T}}\mathbf{h}_{i}X_{i}+\mathbf{W},$$ where $\mathbf{h}_{i}$ is the $i^{\mathrm{th}}$ column of the overall channel state matrix $\mathbf{H}$ (defined in Section \[sec:System-Model\]), and $X_{i}$ is the Gaussian data source corresponding to the antenna $i$. Power on/off assumptions (T1) and (T2) imply that either $\mathrm{E}\left[X_{i}^{2}\right]=\frac{\rho}{s}$ or $\mathrm{E}\left[X_{i}^{2}\right]=0$. Indeed, for a specific user, its input covariance matrix can be written as $\frac{\rho}{s}\mathbf{Q}\mathbf{Q}^{\dagger}$ where $\mathbf{Q}$ is obtained from intercepting some columns from the identity matrix. Given a channel realization $\mathbf{H}$, the base station selects $s$ many antennas according to
[F1)]{}
: Antenna selection criterion. Sort the channel state vectors $\mathbf{h}_{i}$’s increasingly according to their Frobenius norms such that $\left\Vert \mathbf{h}_{\left(1:NL_{T}\right)}\right\Vert \le\left\Vert \mathbf{h}_{\left(2:NL_{T}\right)}\right\Vert \le\cdots\le\left\Vert \mathbf{h}_{\left(NL_{T}:NL_{T}\right)}\right\Vert $, where $\left\Vert \cdot\right\Vert $ denotes the Frobenius norm. Then the antennas corresponding to $\mathbf{h}_{\left(NL_{T}-s+1:NL_{T}\right)},\cdots,\mathbf{h}_{\left(NL_{T}:NL_{T}\right)}$ are selected to be turned on.
To feedback the antenna selection information, totally $\log_{2}{NL_{T} \choose s}$ many bits are needed. The corresponding signal model then reduces to $$\mathbf{Y}=\sum_{k=1}^{s}\mathbf{h}_{\left(NL_{T}-k+1:NL_{T}\right)}X_{k}+\mathbf{W}.$$
Let $\mathbf{h}_{\left(NL_{T}-k+1:NL_{T}\right)}=n_{k}\bm{\xi}_{k}$ where $n_{k}=\left\Vert \mathbf{h}_{\left(NL_{T}-k+1:NL_{T}\right)}\right\Vert $ and $\bm{\xi}_{k}=\mathbf{h}_{\left(NL_{T}-k+1:NL_{T}\right)}/n_{k}$. Define $\mathbf{\Xi}:=\left[\bm{\xi}_{1}\cdots\bm{\xi}_{s}\right]$. Then the sum rate $\mathcal{I}$ is upper bounded by $$\begin{aligned}
\mathcal{I} & :=\mathrm{E}_{\mathbf{H}}\left[\log\left|\mathbf{I}_{L_{R}}+\frac{\rho}{s}\mathbf{\Xi}\mathrm{diag}\left[n_{1}^{2},\cdots,n_{s}^{2}\right]\mathbf{\Xi}^{\dagger}\right|\right]\nonumber \\
& \le\mathrm{E}_{\mathbf{\Xi}}\left[\log\left|\mathbf{I}_{L_{R}}+\frac{\rho}{s}\eta L_{R}\mathbf{\Xi}\mathbf{\Xi}^{\dagger}\right|\right],\label{eq:sum-rate-ub}\end{aligned}$$ where $$\eta:=\frac{1}{sL_{R}}\mathrm{E}_{\mathbf{n}^{2}}\left[\sum_{k=1}^{s}n_{k}^{2}\right],\label{eq:power-efficiency-factor}$$ and the inequality comes from the concavity of $\log\left|\cdot\right|$ function [@Cover_book91_information_theory] and the fact that $\mathbf{\Xi}$ and $\mathbf{n}^{2}:=\left[n_{1}^{2}\cdots n_{s}^{2}\right]^{\dagger}$ are independent [@Edelman_Rao_2005_Random_Matrix_Theory Eq. (3.9)]. We refer to $\eta$ as *the power efficiency factor* as it describes the power gain of choosing the strongest antennas against random antenna selection: if antennas are selected randomly with the total power constraint increased to $\rho\eta$, the average received signal power is the same as that of our antenna selection strategy.
Based on the upper bound (\[eq:sum-rate-ub\]), the sum rate can be approximately quantified. Note that $\left\Vert \mathbf{h}_{i}\right\Vert $’s are i.i.d. r.v. with PDF $f\left(x\right)=\frac{1}{\left(L_{R}-1\right)!}x^{L_{R}-1}e^{-x}$. An application of (\[eq:approx-order-statistics\]) provides an accurate approximation of $\eta$. Furthermore, note that $\mathbf{\Xi}\in\mathcal{M}_{L_{R},1}^{\left(s\right)}\left(\mathbb{C}\right)$ is isotropically distributed. Substituting $c=\frac{\rho}{s}\eta L_{R}$ and $\bar{m}=\frac{s}{L_{R}}$ into (\[eq:CGM-Shannon-Transform\]) estimates the upper bound (\[eq:sum-rate-ub\]). Simulations in Section \[sec:Simulations-and-Discussion\] show that this theoretical calculation gives a good approximation to the true sum rate.
\[sub:General-Beamforming-Strategy\]General Beamforming Strategy
----------------------------------------------------------------
In this subsection, we propose a power on/off strategy with general beamforming: the base station selects the strongest users, jointly quantizes their strongest eigen-channel vectors and broadcasts a common feedback to all the users; then the on-users transmit along the fedback beamforming vectors.
We consider this suboptimal strategy for its implementational simplicity and tractable performance analysis. The user selection is only based on the Frobenius norm of the channel realization, which does not require complicated matrix computations. Note that only a few users are chosen among a large number of total users available and that singular value decomposition is performed only after user selection in our strategy. The computation complexity is much lower than a user selection strategy depending on eigenvalues of the channel matrices. For each selected user, only the strongest eigen-channel is used. This assumption imposes a nice symmetric structure and makes analysis tractable.
In particular, for transmission, along with assumptions T1) and T2), we add one more constraint:
[T3)]{}
: There is at most one on-beam per user, that is, $s_{i}=0$ or $s_{i}=1$. Note that this also implies that the total number of on-streams $s$ is the same as the number of on-users.
For a given channel realization $\mathbf{H}$, we select the on-users according to
[F2)]{}
: User selection criterion. Sort the channel state matrices $\mathbf{H}_{i}$’s such that $\left\Vert \mathbf{H}_{\left(1:N\right)}\right\Vert \le\left\Vert \mathbf{H}_{\left(2:N\right)}\right\Vert \le\cdots\le\left\Vert \mathbf{H}_{\left(N:N\right)}\right\Vert $, where $\left\Vert \cdot\right\Vert $ is the Frobenius norm. Then the users corresponding to $\mathbf{H}_{\left(N-k+1:N\right)},$ $\cdots,\mathbf{H}_{\left(N:N\right)}$ are selected to be turned on.
After selecting the on-users, the base stations also quantizes their strongest eigen-channel vectors. Consider the singular value decomposition $\mathbf{H}_{\left(N-k+1:N\right)}=\mathbf{U}_{k}\mathbf{\Lambda}_{k}\mathbf{V}_{k}^{\dagger}$ where the diagonal elements of $\mathbf{\Lambda}_{k}$ are decreasingly ordered. Let $\mathbf{v}_{k}$ be the column of $\mathbf{V}_{k}$ corresponding to the largest singular value of $\mathbf{\Lambda}_{k}$. Then the matrix $$\mathbf{V}:=\left[\mathbf{v}_{1}\cdots\mathbf{v}_{s}\right]\in\mathcal{M}_{L_{T},1}^{\left(s\right)}\left(\mathbb{C}\right),$$ where $\mathcal{M}_{L_{T},1}^{\left(s\right)}\left(\mathbb{C}\right)$ is the set of composite Grassmann matrices (defined in Section \[sub:CGMatrix\]). In order to quantize $\mathbf{V}$, the base station constructs a codebook $\mathcal{B}\subset\mathcal{M}_{L_{T},1}^{\left(s\right)}\left(\mathbb{C}\right)$ with $\left|\mathcal{B}\right|=2^{R_{\mathrm{q}}}$ where $R_{\mathrm{q}}$ is the feedback bits available for eigen-channel vector quantization. *Note that random codebooks are asymptotically optimal in probability (Theorem \[thm:DRF-CGM-Asymptotics\]), we assume that $\mathcal{B}$ is randomly generated from the isotropic distribution.* For a given eigen-channel vector matrix $\mathbf{V}$, the base station quantizes $\mathbf{V}$ via the
[F3)]{}
: Eigen-channel vector quantization function $$\varphi\left(\mathbf{V}\right)=\underset{\mathbf{B}\in\mathcal{B}}{\arg\;\max}\;\sum_{k=1}^{s}\left|\mathbf{v}_{k}^{\dagger}\mathbf{b}_{k}\right|^{2},\label{eq:quantization-fn-eigenchannels}$$ where $\mathbf{b}_{k}$ is the $k^{\mathrm{th}}$ column of $\mathbf{B}\in\mathcal{B}$. Indeed, let $P^{\left(m\right)},Q^{\left(m\right)}\in\mathcal{G}_{L_{T},1}^{\left(s\right)}\left(\mathbb{C}\right)$ be the composite planes generated by $\mathbf{V}$ and $\mathbf{B}$ respectively. Then (\[eq:quantization-fn-eigenchannels\]) is equivalent to the quantization function on the composite Grassmann manifold defined in (\[eq:quantization-fn-CGM\]).
After quantization, the base station broadcasts the user selection information (requiring $\log_{2}{N \choose s}$ many feedback bits) and the index of eigen-channel vector quantization to the users. The corresponding signal model is now reduced to $$\begin{aligned}
\mathbf{Y} & =\sum_{k=1}^{s}\mathbf{H}_{\left(N-k+1:N\right)}\mathbf{b}_{k}X_{k}+\mathbf{W}\\
& =\sum_{k=1}^{s}\tilde{\mathbf{h}}_{k}X_{k}+\mathbf{W},\end{aligned}$$ where $\tilde{\mathbf{h}}_{k}:=\mathbf{H}_{\left(N-k+1:N\right)}\mathbf{b}_{k}$ is the equivalent channel for the on-user $k$.
The point is that the joint quantization (\[eq:quantization-fn-eigenchannels\]) efficiently employs the feedback resource. It differs from an individual quantization where each $\mathbf{v}_{k}$ is quantized independently: separate codebooks $\mathcal{B}_{1},\cdots,\mathcal{B}_{s}$ are constructed for quantization of $\mathbf{v}_{1},\cdots,\mathbf{v}_{s}$ respectively, and the quantization function is $$\varphi^{\prime}\left(\mathbf{V}\right)=\prod_{k=1}^{s}\underset{\mathbf{b}\in\mathcal{B}_{k}}{\arg\;\max}\;\left|\mathbf{v}_{k}^{\dagger}\mathbf{b}\right|$$ where $\prod$ is the Cartesian product. Indeed, individual quantization is a special case of joint quantization obtained by restricting the codebook to be a Cartesian product of several individual codebooks. It is thus obvious that joint quantization achieves a gain tied to that of vector over scalar quantization.
Certainly the sum rate depends on the codebook. Still, when random codebooks are considered, it is reasonable to focus upon the ensemble average sum rate. Let $\tilde{\mathbf{h}}_{k}=n_{k}\bm{\xi}_{k}$ and $\mathbf{\Xi}=\left[\bm{\xi}_{1}\cdots\bm{\xi}_{s}\right]$, where $n_{k}=\left\Vert \tilde{\mathbf{h}}_{k}\right\Vert $ and $\bm{\xi}_{k}=\tilde{\mathbf{h}}_{k}/n_{k}$. Then the average sum rate satisfies $$\begin{aligned}
\bar{\mathcal{I}}_{\mathrm{rand}} & =\mathrm{E}_{\mathcal{B}}\left[\log\left|\mathbf{I}_{L_{R}}+\frac{\rho}{s}\mathbf{\Xi}\mathrm{diag}\left[n_{1}^{2},\cdots,n_{s}^{2}\right]\mathbf{\Xi}^{\dagger}\right|\right]\nonumber \\
& \le\mathrm{E}_{\mathbf{\Xi}}\left[\log\left|\mathbf{I}_{L_{R}}+\frac{\rho}{s}\mathrm{E}_{\mathcal{B}}\left[\eta\right]L_{R}\mathbf{\Xi\Xi}^{\dagger}\right|\right],\label{eq:average-sum-rate-ub}\end{aligned}$$ where $\eta$ is defined in (\[eq:power-efficiency-factor\]). The inequality in the second line follows from Jensen’s inequality and the next fact.
\[thm:prob-properties\]$\bm{\xi}_{k}$’s $1\le k\le s$ are independent and isotropically distributed. Furthermore, $\bm{\xi}_{k}$’s are independent of $n_{k}$’s.
Consider the singular value decomposition of a standard Gaussian matrix $\mathbf{H}=\mathbf{U}\mathbf{\Lambda}\mathbf{V}^{\dagger}$. It is well known that $\mathbf{U}$ and $\mathbf{V}$ are independent and isotropically distributed, and both of them are independent of $\mathbf{\Lambda}$ [@Edelman_Rao_2005_Random_Matrix_Theory Eq. (3.9)]. Now let $\mathbf{U}_{k}\mathbf{\Lambda}_{k}\mathbf{V}_{k}^{\dagger}$ be the singular value decomposition of $\mathbf{H}_{\left(N-k+1:N\right)}$ $1\le k\le s$. Since we choose users only according to their Frobenius norms, the choice of $\mathbf{H}_{\left(N-k+1:N\right)}$ only depends on $\mathbf{\Lambda}$ but is independent of $\mathbf{U}_{k}$ and $\mathbf{V}_{k}$. The independence among $\mathbf{U}_{k}$, $\mathbf{V}_{k}$ and $\mathbf{\Lambda}_{k}$ still holds. Note that the equivalent channel vector $\tilde{\mathbf{h}}_{k}=\mathbf{U}_{k}\mathbf{\Lambda}_{k}\mathbf{V}_{k}^{\dagger}\mathbf{b}_{k}=\mathbf{U}_{k}\hat{\bm{\xi}}_{k}n_{k}$ where $\mathbf{\Lambda}_{k}\mathbf{V}_{k}^{\dagger}\mathbf{b}_{k}=\hat{\bm{\xi}}_{k}n_{k}$. Since $\mathbf{b}_{k}$ depends only on $\mathbf{V}_{k}$, $\mathbf{U}_{k}$ is independent of $\hat{\bm{\xi}}_{k}$. Thus $\bm{\xi}_{k}=\mathbf{U}_{k}\hat{\bm{\xi}}_{k}$ is isotropically distributed [@James_54_Normal_Multivariate_Analysis_Orthogonal_Group]. Now the fact that $\mathbf{U}_{k}$’s are independent across $k$ implies that $\bm{\xi}_{k}$’s are independent across $k$ [@James_54_Normal_Multivariate_Analysis_Orthogonal_Group].
Next realize that $n_{k}$ is only a function of $\mathbf{\Lambda}_{k}$ and $\mathbf{V}_{k}$, both of which are independent of $\mathbf{U}_{k}$. $\mathbf{U}_{k}$’s are independent of $n_{k}$’s (and isotropically distributed). It follows that $\bm{\xi}_{k}$’s and $n_{k}$’s are independent [@James_54_Normal_Multivariate_Analysis_Orthogonal_Group].
The calculation of $\mathrm{E}_{\mathcal{B}}\left[\eta\right]$ proceeds as follows. To simplify notation, let $\mathbf{H}_{\left(k\right)}=\mathbf{H}_{\left(N-k+1:N\right)}$ and $n_{\left(k\right)}^{2}=\left\Vert \mathbf{H}_{\left(k\right)}\right\Vert ^{2}$. Let $\bar{n}_{\left(\cdot\right)}^{2}=\frac{1}{s}\sum_{k=1}^{s}\mathrm{E}\left[n_{\left(k\right)}^{2}\right]$. Let $\lambda_{k,j}$ ($1\le k\le s$ and $1\le j\le L_{R}$) be the decreasingly ordered eigenvalues of $\mathbf{H}_{\left(k\right)}\mathbf{H}_{\left(k\right)}^{\dagger}$, and $\zeta_{j}=\mathrm{E}\left[\left.\lambda_{k,j}\right|n_{\left(k\right)}^{2}=1\right]$, defined in Lemma \[lem:conditional-expectation-Wishart\]. For a quantization codebook $\mathcal{B}$, let $\mathbf{v}_{k}$ be the $k^{\mathrm{th}}$ column of $\mathbf{V}$, and $\mathbf{b}_{k}$ be the $k^{\mathrm{th}}$ column of $\mathbf{B}=\varphi\left(\mathbf{V}\right)\in\mathcal{B}$. Define $$\begin{aligned}
\gamma & :=\mathrm{E}_{\mathcal{B},\mathbf{V}}\left[\sum_{k=1}^{s}\left|\mathbf{v}_{k}^{\dagger}\mathbf{b}_{k}\right|^{2}\right].\label{eq:def-gamma}\end{aligned}$$ $\mathrm{E}_{\mathcal{B}}\left[\eta\right]$ is a function of $\gamma$.
\[thm:n\_bar\]Let the random codebook $\mathcal{B}$ follows the isotropic distribution. Then$$\begin{aligned}
\mathrm{E}_{\mathcal{B}}\left[\eta\right] & =\mathrm{E}_{\mathcal{B},\mathbf{V}}\left[\sum_{k=1}^{s}\left|\mathbf{v}_{k}^{\dagger}\mathbf{b}_{k}\right|^{2}\right]\nonumber \\
& =L_{R}\left(\frac{\gamma}{s}\zeta_{1}+\frac{s-\gamma}{s}\frac{1-\zeta_{1}}{L_{T}-1}\right)\bar{n}_{\left(\cdot\right)}^{2}.\label{eq:power-efficiency-calculation}\end{aligned}$$
The proof is contained in Appendix \[sub:Proof-of-Theorem-n\_bar\].
To make use of this formula, the constant $\zeta_{1}$ can be well approximated by $\bar{\zeta}_{1/L_{R}}$ using our results in Section \[sub:Conditioned-Eigen\], and $\bar{n}_{\left(\cdot\right)}^{2}$ can be estimated by (\[eq:approx-order-statistics\]). Let $R_{q}$ be the quantization rate on eigen-channel vector quantization. As a function of $R_{q}$, an approximation of $\gamma$ is provided at the end of Section \[sub:CGManifold\]. Put together we have our estimate of $\mathrm{E}_{\mathcal{B}}\left[\eta\right]$. And to estimate the average sum rate, we only need to substitute the value of $\mathrm{E}_{\mathcal{B}}\left[\eta\right]$ into the bound (\[eq:average-sum-rate-ub\]) and then evaluate it via (\[eq:CGM-Shannon-Transform\]).
\[sub:Comments\]Comments
------------------------
### Choice of $s$
The number of on-beams $s$ should be chosen to maximize the sum rate keeping in mind that it is a function of SNR $\rho$. Given that our proved bound accurately approximates the sum rate (when $s\ll N$ and $R_{\mathrm{q}}$ are large enough), the optimal number of on-beams $s^{*}$ can be found by a simple search.
### Antenna Selection and General Beamforming
The antenna selection can be viewed as a special case of general beamforming where a beamforming vector has a particular structure - it must be a column of the identity matrix. Note that general beamforming requires total feedback rate $\log_{2}{N \choose s}+R_{\mathrm{q}}$ bits while antenna selection needs $\log_{2}{NL_{T} \choose s}=\log_{2}{N \choose s}+s\log_{2}L_{T}+O\left(\frac{1}{N}\right)$ bits for feedback. Antenna selection can be viewed as general beamforming with $R_{\mathrm{q}}=s\log_{2}L_{T}$. One difference between antenna selection and general beamforming is that antenna selection does not assume one on-beam per on-user (Assumption T3)). In antenna selection, multiple antennas corresponding to the same user can be turned on simultaneously. As a result, the sum rate achieved by antenna selection is expected to be better than that of general beamforming with $R_{\mathrm{q}}=s\log_{2}L_{T}$. This is supported in our simulations.
\[sec:Simulations-and-Discussion\]Simulations and Discussion
============================================================
Simulations for antenna selection and general beamforming strategies are presented in Fig. \[cap:Antenna-Selection\] and \[cap:Beamforming\] respectively. Fig. \[cap:Antenna-Selection\] shows the sum rate of antenna selection versus SNR. The circles are simulated sum rates, the solid lines are simulated upper bounds (\[eq:sum-rate-ub\]), the plus markers are the sum rates calculated by theoretical approximation, and the dotted lines are the sum rates corresponding to the case where there is no CSIT at all. In the simulations, the value of $s$ is chosen to maximize the sum rate according to our theoretical analysis. Fig. \[cap:Beamforming\] illustrates how the sum rate increases as the eigen-channel vectors quantization rate $R_{\mathrm{q}}$ increases. Here, the $s$ is fixed to be $4$. The dash-dot lines denote perfect beamforming, which corresponds to $R_{\mathrm{q}}=+\infty$. The circles are for our proposed joint strategy, the solid lines are simulated upper bounds (\[eq:sum-rate-ub\]), the up-triangles are for antenna selection and the down-triangles are for individual eigen-channel vectors quantization (recall the detailed discussion in Section \[sub:General-Beamforming-Strategy\]). We observe the following.
- The upper bounds (\[eq:sum-rate-ub\]) and (\[eq:average-sum-rate-ub\]) appear to be good approximations to the sum rate.
- The sum rate increases as the number of users $N$ increases. Fig. \[cap:Antenna-Selection\] compares the $N=32$ and $N=256$ cases. Our analysis bears out that increasing $N$ results in an increase in the equivalent channel norms according to extreme order statistics. The power efficiency factor increases and therefore the sum rate performance improves.
- The loss due to eigen-channel vector quantization decreases exponentially as $R_{\mathrm{q}}$ increases. According to Theorem \[thm:DRF-CGM\], the decay rate is $\frac{1}{s\left(L_{T}-1\right)}R_{\mathrm{q}}$. When $L_{T}$ is not large (which is often true in practice), a relatively small $R_{\mathrm{q}}$ may be good enough. In Fig. \[cap:Beamforming\], as $L_{T}=2$ and $s=4$, $R_{\mathrm{q}}=12$ bits is almost as good as perfect beamforming.
- Our proposed joint strategy achieves better performance than individual quantization. Note that the effect of eigen-channel vectors quantization is characterized by a single parameter $\gamma$. Joint quantization yields larger $\gamma$, larger power efficiency factor, and therefore better performance.
- Antenna selection is only slightly better than general beamforming with $R_{\mathrm{q}}=s\log_{2}L_{T}$. As has been discussed in Section \[sub:Comments\], the performance improvement is due to excluding the assumption T3).
![\[cap:Beamforming\]General Beamforming: Sum Rate versus $R_{\mathrm{q}}$.](Graph/AntennaSel)
![\[cap:Beamforming\]General Beamforming: Sum Rate versus $R_{\mathrm{q}}$.](Graph/GeneralBeamforming)
\[sec:Conclusion\]Conclusion
============================
This paper proposes a joint quantization and feedback strategy for multiaccess MIMO systems with finite rate feedback. The effect of user choice is analyzed by extreme order statistics and the effect of eigen-channel vector quantization is quantified by analysis on the composite Grassmann manifold. By asymptotic random matrix theory, the sum rate is well approximated. Due to its simple implementation and solid performance analysis, the proposed scheme provides a benchmark for multiaccess MIMO systems with finite rate feedback.
\[sub:Random-Matrix-Theory\]Random Matrix Theory
------------------------------------------------
Let $\mathbf{H}\in\mathbb{L}^{n\times m}$ be a standard Gaussian random matrix, where $\mathbb{L}$ is either $\mathbb{R}$ or $\mathbb{C}$. Let $\lambda_{1},\cdots,\lambda_{n}$ be the $n$ singular values of $\frac{1}{m}\mathbf{H}\mathbf{H}^{\dagger}$. Define the empirical distribution of the singular values $$\mu_{n,\bm{\lambda}}\left(\lambda\right)\triangleq\frac{1}{n}\left|\left\{ j:\;\lambda_{j}\le\lambda\right\} \right|.$$ As $n,m\rightarrow\infty$ with $\frac{m}{n}\rightarrow\bar{m}\in\mathbb{R}^{+}$, the empirical measure converges to the Marcenko-Pastur law $$d\mu_{\lambda}=\left(\left(1-\bar{m}\right)^{+}\delta\left(\lambda\right)+\frac{\bar{m}\sqrt{\left(\lambda-\lambda^{-}\right)^{+}\left(\lambda^{+}-\lambda\right)^{+}}}{2\pi\lambda}\right)\, d\lambda\label{eq:spectrum-pdf}$$ almost surely, where $\lambda^{\pm}=\left(1\pm\sqrt{\frac{1}{\bar{m}}}\right)^{2}$ and $\left(x\right)^{+}=\max\left(x,0\right)$ (A good reference for this type of result is [@Verdu_random_matrix_theory_wireless_communications Eq. (1.10)]). Define$$\lambda_{t}^{-}\triangleq\begin{cases}
0 & \mathrm{if}\;\beta\ge1\\
\lambda^{-} & \mathrm{if}\;\beta<1\end{cases}.$$ Consider as well a linear spectral statistic $$g\left(\frac{1}{m}\mathbf{H}\mathbf{H}^{\dagger}\right)=\frac{1}{n}\sum_{i=1}^{n}g\left(\lambda_{i}\right).$$ If $g$ is Lipschitz on $\left[\lambda_{t}^{-},\lambda^{+}\right]$, then we also have that $$\underset{\left(n,m\right)\rightarrow\infty}{\lim}g\left(\frac{1}{m}\mathbf{H}\mathbf{H}^{\dagger}\right)=\int g\left(\lambda\right)d\mu_{\lambda}$$ almost surely, see for example [@Guionnet_Zeitouni_2000] for a modern approach.
The asymptotic properties of the maximum eigenvalue will figure into our analysis. Denote the largest eigenvalue by $\lambda_{1}$.
\[prop:Asymptotic-Lambda-Max\]Let $n,m\rightarrow\infty$ linearly with $\frac{m}{n}\rightarrow\bar{m}\in\mathbb{R}^{+}$.
1. $\lambda_{1}\rightarrow\lambda^{+}$ almost surely.
2. All moments of $\lambda_{1}$ also converge.
The almost sure convergence goes back to [@Bai_Yin_Krishnaih_1988; @Bai_Yin_1993]. The convergence of moments is implied by the tail estimates in [@Ledoux_2005_Orthogonal_Polynomials]. A direct application of this proposition is that for $\forall A_{n}\subset\mathbb{R}^{n}$ such that $\mu_{n,\bm{\lambda}}\left(A_{n}\right)\rightarrow0$, $\mathrm{E}_{\bm{\lambda}}\left[\lambda_{1},\; A_{n}\right]\rightarrow0$.
\[thm:truncate-function-RMT\]Let $\mathbf{H}\in\mathbb{L}^{n\times m}$ ($\mathbb{L}=\mathbb{R}/\mathbb{C}$) be standard Gaussian matrix and $\lambda_{i}$ be the $i^{\mathrm{th}}$ largest eigenvalue of $\frac{1}{m}\mathbf{H}\mathbf{H}^{\dagger}$.
1. Let $g\left(\lambda\right)=f\left(\lambda\right)\cdot\chi_{\left[a,\lambda^{+}\right]}\left(\lambda\right)$ for some $a<\lambda^{+}$ where $f\left(\lambda\right)$ is Lipschitz continuous on $\left[\lambda^{-},\lambda^{+}\right]$ and $\chi_{\left[a,\lambda^{+}\right]}\left(\lambda\right)$ is the indicator function on the set $\left[a,\lambda^{+}\right]$, then as $n,m\rightarrow\infty$ with $\frac{m}{n}\rightarrow\bar{m}\in\mathbb{R}^{+}$, $$\underset{\left(n,m\right)\rightarrow\infty}{\lim}\int g\left(\lambda\right)\cdot d\mu_{n,\bm{\lambda}}\left(\lambda\right)=\int g\left(\lambda\right)\cdot d\mu_{\lambda}$$ almost surely and $$\underset{\left(n,m\right)\rightarrow\infty}{\lim}\mathrm{E}_{\bm{\lambda}}\left[\frac{1}{n}\sum_{i=1}^{n}g\left(\lambda_{i}\right)\right]=\int g\left(\lambda\right)\cdot d\mu_{\lambda}.$$
2. For $\forall a\in\left(\lambda_{t}^{-},\lambda^{+}\right)$, $$\mathrm{E}_{\bm{\lambda}}\left[\frac{1}{n}\left|\left\{ \lambda_{i}:\;\lambda_{i}\ge a\right\} \right|\right]=\int_{a}^{\lambda^{+}}d\mu_{\lambda}.$$
3. For $\forall\tau\in\left(0,\min\left(1,\bar{m}\right)\right)$, $$\underset{\left(n,m\right)\rightarrow\infty}{\lim}\mathrm{E}\left[\frac{1}{n}\left(\sum_{1\le i\le n\tau}\lambda_{i}\right)\right]=\int_{a}^{\lambda^{+}}\lambda\cdot d\mu_{\lambda},$$ where $a\in\left(\lambda^{-},\lambda^{+}\right)$ satisfies $$\tau=\int_{a}^{\lambda^{+}}d\mu_{\lambda}.$$
<!-- -->
1. Though $g\left(\lambda\right)$ is not Lipschitz continuous on $\left[\lambda_{t}^{-},\lambda^{+}\right]$, we are able to construct sequences of Lipschitz functions $g_{k}^{+}\left(\lambda\right)$ and $g_{k}^{-}\left(\lambda\right)$ such that $g_{k}^{\pm}\left(\lambda\right)$’s are Lipschitz continuous on $\left[\lambda_{t}^{-},\lambda^{+}\right]$ for all $k$, $g_{k}^{+}\left(\lambda\right)\ge g\left(\lambda\right)$ and $g_{k}^{-}\left(\lambda\right)\le g\left(\lambda\right)$ for $\lambda\in\left[\lambda_{t}^{-},\lambda^{+}\right]$, and $g_{k}^{\pm}\left(\lambda\right)\rightarrow g\left(\lambda\right)$ pointwisely as $k\rightarrow\infty$. Due to their Lipschitz continuity, $g_{k}^{\pm}\left(\lambda\right)$’s are integrable with respect to $\mu_{\lambda}$. Then we have $$\begin{aligned}
\underset{k\rightarrow\infty}{\lim}\underset{\left(n,m\right)\rightarrow\infty}{\lim}\int g_{k}^{-}\left(\lambda\right)\cdot d\mu_{n,\bm{\lambda}}\left(\lambda\right) & \le\underset{\left(n,m\right)\rightarrow\infty}{\lim}\int g\left(\lambda\right)\cdot d\mu_{n,\bm{\lambda}}\left(\lambda\right)\\
& \le\underset{k\rightarrow\infty}{\lim}\underset{\left(n,m\right)\rightarrow\infty}{\lim}\int g_{k}^{+}\left(\lambda\right)\cdot d\mu_{n,\bm{\lambda}}\left(\lambda\right),\end{aligned}$$ while $$\underset{k\rightarrow\infty}{\lim}\underset{\left(n,m\right)\rightarrow\infty}{\lim}\int g_{k}^{-}\left(\lambda\right)\cdot d\mu_{n,\bm{\lambda}}\left(\lambda\right)=\underset{k\rightarrow\infty}{\lim}\int g_{k}^{-}\left(\lambda\right)\cdot d\mu\left(\lambda\right)=\int g\left(\lambda\right)\cdot d\mu\left(\lambda\right)$$ and $$\underset{k\rightarrow\infty}{\lim}\underset{\left(n,m\right)\rightarrow\infty}{\lim}\int g_{k}^{+}\left(\lambda\right)\cdot d\mu_{n,\bm{\lambda}}\left(\lambda\right)=\int g\left(\lambda\right)\cdot d\mu\left(\lambda\right)$$ almost surely. This proves the almost sure statement, and the convergence of the expectation follows from dominated convergence.
2. follows from the first part upon setting $g\left(\lambda\right)=\chi_{\left[a,\lambda^{+}\right]}\left(\lambda\right)$.
3. Since $a\in\left(\lambda^{-},\lambda^{+}\right)$, there exists an $\epsilon>0$ such that $\left(a-\epsilon,a+\epsilon\right)\subset\left(\lambda^{-},\lambda^{+}\right)$. For any $\delta>0$, define the events$$A_{n,a+\epsilon}=\left\{ \bm{\lambda}:\;\frac{\left|\left\{ \lambda_{i}:\;\lambda_{i}\ge a+\epsilon\right\} \right|}{n}<\tau\right\} ,$$ $$A_{n,a-\epsilon}=\left\{ \bm{\lambda}:\;\frac{\left|\left\{ \lambda_{i}:\;\lambda_{i}\ge a-\epsilon\right\} \right|}{n}>\tau\right\} ,$$ $$B_{n,a+\epsilon,\delta}=\left\{ \bm{\lambda}:\;\left|\frac{1}{n}\sum_{\lambda_{i}\ge a+\epsilon}\lambda_{i}-\int_{a+\epsilon}^{\lambda^{+}}\lambda\cdot d\mu_{\lambda}\right|<\delta\right\} ,$$ and $$B_{n,a-\epsilon,\delta}=\left\{ \bm{\lambda}:\;\left|\frac{1}{n}\sum_{\lambda_{i}\ge a-\epsilon}\lambda_{i}-\int_{a-\epsilon}^{\lambda^{+}}\lambda\cdot d\mu_{\lambda}\right|<\delta\right\} .$$ According to the first part of this theorem, it can be verified that $\forall\epsilon>0$, as $\left(n,m\right)\rightarrow\infty$, $\mu_{n,\bm{\lambda}}\left(A_{n,a+\epsilon}\right)\rightarrow1$, $\mu_{n,\bm{\lambda}}\left(A_{n,a-\epsilon}\right)\rightarrow1$, $\mu_{n,\bm{\lambda}}\left(B_{n,a+\epsilon,\delta}\right)\rightarrow1$, and $\mu_{n,\bm{\lambda}}\left(B_{n,a-\epsilon,\delta}\right)\rightarrow1$. Then for sufficiently large $n$, $$\begin{aligned}
& \mathrm{E}_{\bm{\lambda}}\left[\frac{1}{n}\sum_{i\le n\tau}\lambda_{i}\right]\nonumber \\
& \ge\mathrm{E}_{\bm{\lambda}}\left[\frac{1}{n}\sum_{i\le n\tau}\lambda_{i},\; A_{n,a+\epsilon}\cap B_{n,a+\epsilon,\delta}\right]\nonumber \\
& \overset{\left(a\right)}{\ge}\mathrm{E}_{\bm{\lambda}}\left[\frac{1}{n}\sum_{\lambda_{i}\ge a+\epsilon}\lambda_{i},\; A_{n,a+\epsilon}\cap B_{n,a+\epsilon,\delta}\right]\nonumber \\
& \overset{\left(b\right)}{\ge}\mathrm{E}_{\bm{\lambda}}\left[\int_{a+\epsilon}^{\lambda^{+}}\lambda\cdot d\mu_{\lambda}-\delta,\; A_{n,a+\epsilon}\cap B_{n,a+\epsilon,\delta}\right]\nonumber \\
& =\left(\int_{a+\epsilon}^{\lambda^{+}}\lambda\cdot d\mu_{\lambda}-\delta\right)\mu_{n,\bm{\lambda}}\left(A_{n,a+\epsilon}\cap B_{n,a+\epsilon,\delta}\right)\nonumber \\
& \ge\left(\int_{a+\epsilon}^{\lambda^{+}}\lambda\cdot d\mu_{\lambda}-\delta\right)\left(1-\delta\right),\label{eq:E-eigenvalue-lb}\end{aligned}$$ where $\mathrm{E}\left[\cdot,A\right]$ denotes the expectation operation on the measurable set $A$, $\left(a\right)$ and $\left(b\right)$ follow from the definition of $A_{n,a+\epsilon}$ and $B_{n,a+\epsilon,\delta}$ respectively. Similarly, when $n$ is large enough, $$\begin{aligned}
& \mathrm{E}_{\bm{\lambda}}\left[\frac{1}{n}\sum_{i\le n\tau}\lambda_{i}\right]\nonumber \\
& \le\mathrm{E}_{\bm{\lambda}}\left[\frac{1}{n}\sum_{i\le n\tau}\lambda_{i},\; A_{n,a-\epsilon}\right]+\mathrm{E}_{\bm{\lambda}}\left[\lambda_{1},\; A_{n,a-\epsilon}^{c}\right]\nonumber \\
& \overset{\left(c\right)}{\le}\mathrm{E}_{\bm{\lambda}}\left[\frac{1}{n}\sum_{\lambda_{i}\ge a-\epsilon}\lambda_{i},\; A_{n,a-\epsilon}\right]+\delta\nonumber \\
& \le\mathrm{E}_{\bm{\lambda}}\left[\int_{a-\epsilon}^{\lambda^{+}}\lambda\cdot d\mu_{\lambda}+\delta,\; A_{n,a-\epsilon}\cap B_{n,a-\epsilon,\delta}\right]\nonumber \\
& \quad\quad+\mathrm{E}_{\bm{\lambda}}\left[\lambda_{1},\; A_{n,a-\epsilon}\cap B_{n,a-\epsilon,\delta}^{c}\right]+\delta\nonumber \\
& \overset{\left(d\right)}{\le}\left(\int_{a-\epsilon}^{\lambda^{+}}\lambda\cdot d\mu_{\lambda}+\delta\right)\mu_{n,\bm{\lambda}}\left(A_{n,a-\epsilon}\cap B_{n,a-\epsilon,\delta}\right)+2\delta\nonumber \\
& \le\int_{a-\epsilon}^{\lambda^{+}}\lambda\cdot d\mu_{\lambda}+3\delta,\label{eq:E-eigenvalue-ub}\end{aligned}$$ where $\left(c\right)$ and $\left(d\right)$ are an application of Proposition \[prop:Asymptotic-Lambda-Max\]. Now let $\delta\downarrow0$ and then $\epsilon\downarrow0$. Then we have proved that $$\underset{\left(n,m\right)\rightarrow\infty}{\lim}\mathrm{E}\left[\frac{1}{n}\left(\sum_{1\le i\le n\tau}\lambda_{i}\right)\right]=\int_{a}^{\lambda^{+}}\lambda\cdot d\mu_{\lambda}.$$
\[sub:Proof-of-Extreme-Order-Statistics\]Proof of Lemma \[lem:Expectation-extreme-chi2\]
----------------------------------------------------------------------------------------
As the first step, we compute the asymptotic distribution and expectation of $X_{\left(n:n\right)}$. It can be verified that $$1-F_{X}\left(y\right)=\int_{y}^{+\infty}f_{X}\left(x\right)dx=e^{-y}\left(\sum_{i=0}^{L-1}\frac{1}{i!}y^{i}\right),$$ and for $\forall a>0$,$$\begin{aligned}
& \int_{a}^{+\infty}1-F_{X}\left(y\right)dy\\
& =e^{-a}\left(\sum_{i=0}^{L-1}\frac{1}{i!}a^{i}+\sum_{i=0}^{L-2}\frac{1}{i!}a^{i}+\cdots+\sum_{i=0}^{0}\frac{1}{i!}a^{i}\right)\\
& =e^{-a}\left(\sum_{i=0}^{L-1}\frac{L-i}{i!}a^{i}\right).\end{aligned}$$ For $0<t<+\infty$, define $$R\left(t\right)=\frac{\int_{t}^{+\infty}\left(1-F_{X}\left(y\right)\right)dy}{1-F_{X}\left(t\right)}.$$ Then $$\underset{t\rightarrow+\infty}{\lim}R\left(t\right)=\underset{t\rightarrow+\infty}{\lim}\frac{\sum_{i=0}^{L-1}\frac{L-i}{i!}t^{i}}{\sum_{i=0}^{L-1}\frac{1}{i!}t^{i}}=1.\label{eq:_R(t)_limit}$$ Now let $$a_{n}=\inf\left\{ x:1-F_{X}\left(x\right)\leq\frac{1}{n}\right\} ,$$ and $$b_{n}=R\left(a_{n}\right)=\frac{\sum_{i=0}^{L-1}\frac{L-i}{i!}a_{n}^{i}}{\sum_{i=0}^{L-1}\frac{1}{i!}a_{n}^{i}}.$$ It can be verified that $a_{n}\overset{n\rightarrow\infty}{\longrightarrow}+\infty$, and that $b_{n}\overset{n\rightarrow\infty}{\longrightarrow}1$ by (\[eq:\_R(t)\_limit\]). Furthermore,$$\begin{aligned}
& \underset{n\rightarrow\infty}{\lim}n\left[1-F_{X}\left(a_{n}+xb_{n}\right)\right]\nonumber \\
& =\underset{n\rightarrow\infty}{\lim}\frac{1-F_{X}\left(a_{n}+xb_{n}\right)}{1-F_{X}\left(a_{n}\right)}\nonumber \\
& =\underset{n\rightarrow\infty}{\lim}e^{-xb_{n}}\frac{\sum_{i=0}^{L-1}\frac{1}{i!}\left(a_{n}+xb_{n}\right)^{i}}{\sum_{i=0}^{L-1}\frac{1}{i!}\left(a_{n}\right)^{i}}\nonumber \\
& =e^{-x}.\label{eq:_exponential_element}\end{aligned}$$ Therefore, for all $x\in\mathbb{R}$ and sufficiently large $n$, $$\begin{aligned}
& P\left(X_{\left(n:n\right)}<a_{n}+b_{n}x\right)\\
& =\left[1-\frac{1}{n}n\left(1-F_{X}\left(a_{n}+b_{n}x\right)\right)\right]^{n}\\
& =\exp\left(n\cdot\log\left(1-\frac{1}{n}e^{-x}\left(1+o\left(1\right)\right)\right)\right)\\
& =\exp\left(-e^{-x}\left(1+o\left(1\right)\right)\right)\\
& \overset{n\rightarrow\infty}{\longrightarrow}\exp\left(-e^{-x}\right).\end{aligned}$$ This identifies the limiting distribution, and the tail is of sufficient decay to conclude that $$\underset{n\rightarrow+\infty}{\lim}\mathrm{E}\left[\frac{X_{\left(n:n\right)}-a_{n}}{b_{n}}\right]=\int_{-\infty}^{+\infty}xde^{-e^{-x}}:=\mu_{1}.$$
Given the law of the first maxima $X_{\left(n:n\right)}$, the distribution and the expectation of the $k^{\mathrm{th}}$ maxima follow easily. With $z_{n}=a_{n}+b_{n}x$, $$P\left(X_{\left(n-k+1:n\right)}\le z_{n}\right)=\sum_{t=0}^{k-1}{n \choose t}\left(1-F_{X}\left(z_{n}\right)\right)^{t}F_{X}^{n-t}\left(z_{n}\right).$$ According to (\[eq:\_exponential\_element\]), ${n \choose t}\left(1-F_{X}\left(z_{n}\right)\right)^{t}\overset{n\rightarrow\infty}{\longrightarrow}\frac{1}{t!}e^{-tx}$ and $F_{X}^{n-t}\left(z_{n}\right)\overset{n\rightarrow\infty}{\longrightarrow}e^{-e^{-x}}$. Thus$$P\left(\frac{X_{\left(n-k+1:n\right)}-a_{n}}{b_{n}}\le x\right)\overset{n\rightarrow\infty}{\longrightarrow}\exp\left(-e^{-x}\right)\sum_{t=0}^{k-1}\frac{1}{t!}e^{-tx}.$$ Denote it by $H_{k}\left(x\right)$. The corresponding PDF is given by $$h_{k}\left(x\right)=H_{k}^{'}\left(x\right)=e^{-e^{-x}}\frac{1}{\left(k-1\right)!}e^{-kx}.\label{eq:h_k_(x)}$$ Define $\mu_{k}=\int_{-\infty}^{+\infty}xh_{k}\left(x\right)dx$. It can be verified that Evaluating $\mu_{k}^{x}$ gives an iterative formula$$\begin{aligned}
\mu_{k} & =\frac{1}{\left(k-1\right)!}\int_{-\infty}^{+\infty}xe^{-kx}e^{-e^{-x}}dx\\
& =0-\frac{1}{\left(k-1\right)!}\int_{-\infty}^{+\infty}e^{-e^{-x}}d\left(xe^{-(k-1)x}\right)\\
& =\frac{1}{\left(k-2\right)!}\int_{-\infty}^{+\infty}xe^{-\left(k-1\right)x}e^{-e^{-x}}dx\\
& \quad\quad-\frac{1}{\left(k-1\right)!}\int_{-\infty}^{+\infty}e^{-\left(k-1\right)x}e^{-e^{-x}}dx\\
& =\mu_{k-1}^{x}-\frac{1}{k-1},\end{aligned}$$ where the last step follows the fact that $\frac{1}{\left(k-2\right)!}e^{-\left(k-1\right)}\exp\left(-e^{-x}\right)$ is the asymptotic pdf of $\left(k-1\right)^{\mathrm{th}}$ maxima. Therefore,$$\underset{n\rightarrow+\infty}{\lim}\mathrm{E}\left[\frac{X_{\left(n-k+1:n\right)}-a_{n}}{b_{n}}\right]=\mu_{k}=\mu_{1}-\sum_{i=1}^{k-1}\frac{1}{i}.$$ and so also,$$\underset{n\rightarrow+\infty}{\lim}\mathrm{E}\left[\frac{\sum_{k=1}^{s}X_{\left(n-k+1:n\right)}-sa_{n}}{b_{n}}\right]=\sum_{k=1}^{s}\mu_{k}=s\mu_{1}-\sum_{i=1}^{s}\frac{s-i}{i}.$$
\[sub:proof\_DRF\_composite\_GM\]Proof of Theorem \[thm:DRF-CGM\]
-----------------------------------------------------------------
The proof of Theorem \[thm:DRF-CGM\] is similar to that of Theorem 2 in our earlier paper [@Dai_IT2008_Quantization_Grassmannian_manifold]; the difference being that the composite Grassmann manifold is of interest here while the “single” Grassmann manifold is the focus in that work. The key step of this proof is the volume calculation of a small ball in the composite Grassmann manifold. Given the volume formula, the upper and lower bounds follow from the exact arguments in [@Dai_IT2008_Quantization_Grassmannian_manifold].
A metric ball in $\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ centered at $P^{\left(m\right)}\in\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ with radius $\delta\ge0$ is defined as $$B_{P^{\left(m\right)}}\left(\delta\right):=\left\{ Q^{\left(m\right)}\in\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right):\; d_{c}\left(P^{\left(m\right)},Q^{\left(m\right)}\right)\le\delta\right\} .$$ The volume of $B_{P^{\left(m\right)}}\left(\delta\right)$ as the probability of an isotropically distributed $Q^{\left(m\right)}\in\mathcal{G}_{n,p}^{\left(m\right)}\left(\mathbb{L}\right)$ in this ball: $$\mu\left(B_{P^{\left(m\right)}}\left(\delta\right)\right):=\Pr\left(Q^{\left(m\right)}\in B_{P^{\left(m\right)}}\left(\delta\right)\right).$$ Since $\mu\left(B_{P^{\left(m\right)}}\left(\delta\right)\right)$ is independent of the choice of the center $P^{\left(m\right)}$, we simply denote it by $\mu^{\left(m\right)}\left(\delta\right)$. We have:
\[thm:volume-CGM\]When $\delta\le1$, $$\mu^{\left(m\right)}\left(\delta\right)=\frac{\Gamma^{m}\left(\frac{t}{2}+1\right)}{\Gamma\left(m\frac{t}{2}+1\right)}c_{n,p,p,\beta}^{m}\delta^{mt},\label{eq:volume-CGM}$$ where $c_{n,p,p,\beta}$ and $t$ are defined in Lemma \[lem:DRF-GM\].
Let us drop the subscript of $c_{n,p,p,\beta}$ during the proof. In [@Dai_IT2008_Quantization_Grassmannian_manifold], we proved that for a single Grassmann manifold, $\mu^{\left(1\right)}\left(d_{c}^{2}\le x\right)=\mu^{\left(1\right)}\left(\sqrt{x}\right)=cx^{\frac{t}{2}}\left(1+O\left(x\right)\right)$ when $x\le1$, and it can be verified that $$d\mu\left(d_{c}^{2}\le x\right)=\frac{t}{2}cx^{\frac{t}{2}-1}\left(1+O\left(x\right)\right)\cdot dx.$$ By the definition of the volume, $d\mu^{\left(2\right)}\left(x\right)/dx$ is a convolution of $d\mu\left(x\right)/dx$ and $d\mu\left(x\right)/dx$. So, $$\begin{aligned}
\frac{d\mu^{\left(2\right)}\left(x\right)}{dx} & =\int_{0}^{x}\frac{t^{2}}{4}c^{2}\tau^{\frac{t}{2}-1}\left(x-\tau\right)^{\frac{t}{2}-1}\left(1+O\left(\tau\right)\right)\left(1+O\left(x-\tau\right)\right)d\tau\\
& \overset{\left(a\right)}{=}\frac{t^{2}}{4}c^{2}x^{t-1}\int_{0}^{1}y^{\frac{t}{2}-1}\left(1-y\right)^{\frac{t}{2}-1}\left(1+O\left(xy\right)+O\left(x\left(1-y\right)\right)\right)dy\\
& =\frac{t^{2}}{4}c^{2}x^{t-1}\frac{\Gamma\left(\frac{t}{2}\right)\Gamma\left(\frac{t}{2}\right)}{\Gamma\left(t\right)}\left(1+O\left(x\right)\right),\end{aligned}$$ where $\left(a\right)$ follows from the variable change $\tau=xy$. A calculation produces $$\mu^{\left(2\right)}\left(d_{c}^{2}\le x\right)=\frac{\Gamma\left(\frac{t}{2}+1\right)\Gamma\left(\frac{t}{2}+1\right)}{\Gamma\left(t+1\right)}c^{2}x^{t}\left(1+O\left(x\right)\right).$$ By mathematical induction, we reach (\[eq:volume-CGM\]). Note that $\delta\le1$ is required in every step.
Based on the volume formula, an upper bound on the distortion rate function $D^{*}\left(K\right)$ on the composite Grassmann manifold $$D^{*}\left(K\right)\le\frac{2}{mt}\Gamma\left(\frac{2}{mt}\right)\frac{\Gamma^{\frac{2}{mt}}\left(m\frac{t}{2}+1\right)}{\Gamma^{\frac{2}{t}}\left(\frac{t}{2}+1\right)}c_{n,p,p,\beta}^{-\frac{2}{t}}2^{-\frac{2\log_{2}K}{mt}}\left(1+o\left(1\right)\right)$$ is derived by calculating the average distortion of random codes (see [@Dai_IT2008_Quantization_Grassmannian_manifold] for details). Furthermore, by the sphere packing/covering argument (again see [@Dai_IT2008_Quantization_Grassmannian_manifold] for details), the lower bound $$\frac{mt}{mt+2}\frac{\Gamma^{\frac{2}{mt}}\left(mt+1\right)}{\Gamma^{\frac{2}{t}}\left(t+1\right)}c_{n,p,p,\beta}^{-\frac{2}{t}}2^{-\frac{2\log_{2}K}{mt}}\left(1+o\left(1\right)\right)\le D^{*}\left(K\right)$$ is arrived at. Theorem \[thm:DRF-CGM\] is proved.
\[sub:Proof-of-Theorem-n\_bar\]Proof of Theorem \[thm:n\_bar\]
--------------------------------------------------------------
The key step is to prove that $\mathrm{E}_{\mathcal{B},\mathbf{H}}\left[\mathbf{V}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\mathbf{V}_{k}\right]=\mathrm{diag}\left[\frac{\gamma}{s},\frac{s-\gamma}{s\left(L_{T}-1\right)},\cdots,\frac{s-\gamma}{s\left(L_{T}-1\right)}\right]$, where $\mathbf{V}_{k}$ is from the singular value decomposition $\mathbf{H}_{\left(k\right)}=\mathbf{U}_{k}\mathbf{\Lambda}_{k}\bar{\mathbf{V}}_{k}^{\dagger}$. Let $\mathbf{V}_{k}=\left[\mathbf{v}_{k}\bar{\mathbf{V}}_{k}\right]$ where $\bar{\mathbf{V}}_{k}\in\mathbb{C}^{L_{T}\times\left(L_{T}-1\right)}$ is composed of all the columns of $\mathbf{V}_{k}$ except $\mathbf{v}_{k}$. Let $\mathbf{V}=\left[\mathbf{v}_{1}\cdots\mathbf{v}_{s}\right]$. Recall our feedback function $\varphi\left(\mathbf{V}\right)$ in (\[eq:quantization-fn-eigenchannels\]) and definition of $\gamma$ in (\[eq:def-gamma\]). Then the fact that $$\mathrm{E}_{\mathcal{B},\mathbf{H}}\left[\mathbf{v}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\mathbf{v}_{k}\right]=\frac{\gamma}{s}$$ is implied by the following lemma.
\[lem:correlation\_v\_k\_b\_k\]Let $\mathbf{V}\in\mathcal{M}_{L_{T},1}^{\left(s\right)}$ be isotropically distributed and $\mathcal{B}\subset\mathcal{M}_{L_{T},1}^{\left(s\right)}$ be randomly generated from the isotropic distribution. Let $\mathbf{B}=\varphi\left(\mathbf{V}\right)$ where $\varphi\left(\cdot\right)$ is given in (\[eq:quantization-fn-eigenchannels\]) and $\gamma$ is given by (\[eq:def-gamma\]). Then $$\mathrm{E}_{\mathcal{B},\mathbf{V}}\left[\mathbf{V}^{\dagger}\mathbf{B}\mathbf{B}^{\dagger}\mathbf{V}\right]=\frac{\gamma}{s}\mathbf{I}_{s}.$$
Let $\mathbf{Z}=\mathrm{E}_{\mathcal{B},\mathbf{V}}\left[\mathbf{V}^{\dagger}\mathbf{B}\mathbf{B}^{\dagger}\mathbf{V}\right]$. For any $\theta\in\left[0,2\pi\right)$, let $\mathbf{A}_{k}=\mathrm{diag}\left[1,\cdots,1,e^{j\theta},1,\cdots,1\right]$ be obtained by replacing the $k^{\mathrm{th}}$ diagonal element of $\mathbf{I}$ with $e^{j\theta}$. It can be verified that $\mathbf{V}\mathbf{A}_{k}\in\mathcal{M}_{L_{T},1}^{\left(s\right)}$ is isotropically distributed, and $\varphi\left(\mathbf{V}\mathbf{A}_{k}\right)=\varphi\left(\mathbf{V}\right)=\mathbf{B}$. We have $$\begin{aligned}
\mathbf{Z} & =\mathrm{E}_{\mathcal{B},\mathbf{V}\mathbf{A}_{k}}\left[\mathbf{A}_{k}^{\dagger}\mathbf{V}^{\dagger}\mathbf{B}\mathbf{B}^{\dagger}\mathbf{V}\mathbf{A}_{k}\right]\\
& =\mathbf{A}_{k}^{\dagger}\mathrm{E}_{\mathcal{B},\mathbf{V}}\left[\mathbf{V}^{\dagger}\mathbf{B}\mathbf{B}^{\dagger}\mathbf{V}\right]\mathbf{A}_{k}\\
& =\mathbf{A}_{k}^{\dagger}\mathbf{Z}\mathbf{A}_{k},\end{aligned}$$ where the first equality is obtained by changing the variable from $\mathbf{V}$ to $\mathbf{VA}_{k}$, and the second equality is obtained by replacing the measure of $\mathbf{VA}_{k}$ with the measure of $\mathbf{V}$. Then $\left(\mathbf{Z}\right)_{k,j}=e^{-j\theta}\left(\mathbf{Z}\right)_{k,j}$ for $j\ne k$, which is only possible if $\left(\mathbf{Z}\right)_{k,j}=0$. Therefore, $\mathbf{Z}$ is a diagonal matrix.
Now let $\mathbf{P}\in\mathbb{R}^{s\times s}$ be a permutation matrix generated by permutating rows/columns of the identity matrix. Let $\mathcal{B}\mathbf{P}=\left\{ \mathbf{BP}:\;\mathbf{B}\in\mathcal{B}\right\} $. Then $\mathbf{VP}\in\mathcal{M}_{L_{T},1}^{\left(s\right)}$ and $\mathbf{BP}\in\mathcal{M}_{L_{T},1}^{\left(s\right)}$ are isotropically distributed. It can be verified that $\varphi_{\mathcal{B}\mathbf{P}}\left(\mathbf{VP}\right)=\mathbf{BP}=\varphi_{\mathcal{B}}\left(\mathbf{V}\right)\mathbf{P}$, where the subscript $\varphi$ emphasizes the choice of codebook. Then,$$\begin{aligned}
\mathbf{Z} & =\mathrm{E}_{\mathcal{B}\mathbf{P},\mathbf{VP}}\left[\left(\mathbf{VP}\right)^{\dagger}\varphi_{\mathcal{B}\mathbf{P}}\left(\mathbf{VP}\right)\varphi_{\mathcal{B}\mathbf{P}}\left(\mathbf{VP}\right)^{\dagger}\left(\mathbf{VP}\right)\right]\\
& =\mathbf{P}^{\dagger}\mathrm{E}_{\mathcal{B}\mathbf{P},\mathbf{V}}\left[\mathbf{V}^{\dagger}\varphi_{\mathcal{B}\mathbf{P}}\left(\mathbf{VP}\right)\varphi_{\mathcal{B}\mathbf{P}}\left(\mathbf{VP}\right)^{\dagger}\mathbf{V}\right]\mathbf{P}\\
& =\mathbf{P}^{\dagger}\mathrm{E}_{\mathcal{B}\mathbf{P},\mathbf{V}}\left[\mathbf{V}^{\dagger}\varphi_{\mathcal{B}}\left(\mathbf{V}\right)\mathbf{P}\mathbf{P}^{\dagger}\varphi_{\mathcal{B}}\left(\mathbf{V}\right)^{\dagger}\mathbf{V}\right]\mathbf{P}\\
& =\mathbf{P}^{\dagger}\mathrm{E}_{\mathcal{B},\mathbf{V}}\left[\mathbf{V}^{\dagger}\varphi_{\mathcal{B}}\left(\mathbf{V}\right)\varphi_{\mathcal{B}}\left(\mathbf{V}\right)^{\dagger}\mathbf{V}\right]\mathbf{P}\\
& =\mathbf{P}^{\dagger}\mathbf{Z}\mathbf{P},\end{aligned}$$ where the first equality is obtained by variables change, and the second and fourth equality follows from measure replacement. It follows that $\left(\mathbf{Z}\right)_{i,i}=\left(\mathbf{Z}\right)_{j,j}$ for $1\le i,j\le s$.
Finally, $\mathbf{Z}=\frac{\gamma}{s}\mathbf{I}$ follows from the fact that $\mathrm{tr}\left(\mathbf{Z}\right)=\mathrm{E}\left[\mathrm{tr}\left(\mathbf{V}^{\dagger}\mathbf{B}\mathbf{B}^{\dagger}\mathbf{V}\right)\right]=\gamma$.
We evaluate $$\mathrm{E}\left[\mathbf{V}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\mathbf{V}_{k}\right]=\left[\begin{array}{cc}
\mathrm{E}\left[\mathbf{v}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\mathbf{v}_{k}\right] & \mathrm{E}\left[\mathbf{v}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\bar{\mathbf{V}}_{k}\right]\\
\mathrm{E}\left[\bar{\mathbf{V}}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\mathbf{v}_{k}\right] & \mathrm{E}\left[\bar{\mathbf{V}}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\bar{\mathbf{V}}_{k}\right]\end{array}\right].$$ For any unitary matrix $\mathbf{U}_{r}\in\mathbb{C}^{\left(L_{T}-1\right)\times\left(L_{T}-1\right)}$, $\left[\mathbf{v}_{k},\bar{\mathbf{V}}_{k}\mathbf{U}\right]$ is also isotropically distributed. Employ the method in the proof of Lemma \[lem:correlation\_v\_k\_b\_k\] to find that$$\mathrm{E}\left[\mathbf{v}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\bar{\mathbf{V}}_{k}\right]=\mathrm{E}\left[\mathbf{v}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\bar{\mathbf{V}}_{k}\right]\mathbf{U},$$ $$\mathrm{and}\quad\mathrm{E}\left[\bar{\mathbf{V}}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\bar{\mathbf{V}}_{k}\right]=\mathbf{U}^{\dagger}\mathrm{E}\left[\bar{\mathbf{V}}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\bar{\mathbf{V}}_{k}\right]\mathbf{U}.$$ Therefore, $\mathrm{E}\left[\mathbf{v}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\bar{\mathbf{V}}_{k}\right]=\mathbf{o}^{\dagger}$ and $\mathrm{E}\left[\bar{\mathbf{V}}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\bar{\mathbf{V}}_{k}\right]=c\mathbf{I}_{L_{T}-1}$ for some constant $c$. Note that $\mathrm{E}\left[\mathbf{v}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\mathbf{v}_{k}\right]=\frac{\gamma}{s}$ and $\mathrm{E}\left[\mathrm{tr}\left(\mathbf{V}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\mathbf{V}_{k}\right)\right]=1$. Hence, $c=\frac{s-\gamma}{s\left(L_{T}-1\right)}$ and $\mathrm{E}\left[\mathbf{V}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\mathbf{V}_{k}\right]=\mathrm{diag}\left[\frac{\gamma}{s},\frac{s-\gamma}{s\left(L_{T}-1\right)},\cdots,\frac{s-\gamma}{s\left(L_{T}-1\right)}\right]$.
Finally, $$\begin{aligned}
\mathrm{E}_{\mathcal{B}}\left[\eta\right] & =\frac{1}{sL_{R}}\mathrm{E}_{\mathcal{B},\mathbf{H}}\left[\sum_{k=1}^{s}n_{k}^{2}\right]\\
& =\frac{1}{sL_{R}}\sum_{k=1}^{s}\mathrm{E}_{\mathcal{B},\mathbf{H}}\left[\mathrm{tr}\left(\mathbf{H}_{\left(k\right)}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\mathbf{H}_{\left(k\right)}^{\dagger}\right)\right]\\
& =\frac{1}{sL_{R}}\sum_{k=1}^{s}\mathrm{tr}\left(\mathrm{E}_{\mathcal{B},\mathbf{H}}\left[\mathbf{V}_{k}^{\dagger}\mathbf{b}_{k}\mathbf{b}_{k}^{\dagger}\mathbf{V}_{k}\right]\mathrm{E}_{\mathbf{H}}\left[\mathbf{\Lambda}_{k}^{\dagger}\mathbf{\Lambda}_{k}\right]\right)\\
& =\frac{1}{sL_{R}}\sum_{k=1}^{s}\left(\frac{\gamma}{s}\zeta_{1}\mathrm{E}_{\mathbf{H}}\left[n_{\left(k\right)}^{2}\right]+\frac{s-\gamma}{s}\frac{\left(1-\zeta_{1}\right)\mathrm{E}_{\mathbf{H}}\left[n_{\left(k\right)}^{2}\right]}{L_{T}-1}\right)\\
& =\frac{1}{L_{R}}\left(\frac{\gamma}{s}\zeta_{1}+\frac{s-\gamma}{s}\frac{1-\zeta_{1}}{L_{T}-1}\right)\bar{n}_{\left(\cdot\right)}^{2},\end{aligned}$$ where the third line follows from the fact that $\mathbf{\Lambda}_{k}$ is independent of $\mathbf{V}_{k}$ and $\mathbf{b}_{k}$.
[^1]: When a user joins the multiaccess system, a unique index is assigned and keeps constant. A user in a multiaccess system is aware of the corresponding index.
[^2]: If $n>m$, $\mathrm{E}\left[\lambda_{i}|\mathrm{tr}\left(\mathbf{HH}^{\dagger}\right)=c\right]=0$ for $i>m$ and $\mathrm{E}\left[\lambda_{i}|\mathrm{tr}\left(\mathbf{HH}^{\dagger}\right)=c\right]=\zeta_{i}^{\prime}c$ for $i\le m$, where $\zeta_{i}^{\prime}:=\frac{1}{c}\mathrm{E}\left[\lambda_{i}|\mathrm{tr}\left(\mathbf{H}^{\dagger}\mathbf{H}\right)=c\right]$. The calculation of $\zeta_{i}^{\prime}$ for $i\le m$ is included in Theorem \[lem:conditional-expectation-Wishart\] as well.
[^3]: The Haar measure is well defined for locally compact topological groups [@Haar_1933_Haar_Measure; @Muirhead_book82_multivariate_statistics], and therefore for the Grassmann manifold, the composite Grassmann manifold and the composite Grassmann matrices. Here, the group right and left operations are clear from context.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In distributed evolutionary algorithms, migration interval is used to decide migration moments. Nevertheless, migration moments predetermined by intervals cannot match the dynamic situation of evolution. In this paper, a scheme of setting the success rate of migration based on subpopulation diversity at each interval is proposed. With the scheme, migration still occurs at intervals, but the probability of immigrants entering the target subpopulation will be determined by the diversity of this subpopulation according to a proposed formula. An analysis shows that the time consumption of our scheme is acceptable. In our experiments, the basement of parallelism is an evolutionary algorithm for the travelling salesman problem. Under different value combinations of parameters for the formula, outcomes for eight benchmark instances of the distributed evolutionary algorithm with the proposed scheme are compared with those of a traditional one, respectively. Results show that the distributed evolutionary algorithm based on our scheme has a significant advantage on solutions especially for high difficulty instances. Moreover, it can be seen that the algorithm with the scheme has the most outstanding performance under three value combinations of above-mentioned parameters for the formula.'
author:
- 'Chengjun Li and Jia Wu, [^1] [^2]'
bibliography:
- 'IEEEabrv.bib'
- 'calling.bib'
title: Subpopulation Diversity Based Selecting Migration Moment in Distributed Evolutionary Algorithms
---
[Shell : Bare Demo of IEEEtran.cls for Journals]{}
Distributed evolutionary algorithm, subpopulation diversity, migration moment, success rate of migration.
Introduction
============
stochastic search methods based on population, evolutionary algorithms (EAs) have been applied successfully in many fields [@alba2002parallelism; @wu2013artificial]. Nevertheless, for nontrivial problems, EAs tend to premature convergence. Fortunately, using a parallel EA (PEA) often leads to a superior numerical performance because EAs are naturally prone to parallelism [@alba2002parallelism]. Among types of PEAs, distributed EAs (DEAs), which can be readily implemented in distributed memory MIMD computers [@alba2002parallelism], are most popular [@alba2002parallelism; @cantu1998survey].
In a DEA, a large population is divided into a number of subpopulations. The parallel operator, migration, exchanges individuals between subpopulations at intervals [@herrera1999hierarchical]. As a result, global convergence behavior is improved. The first idea of DEA can be traced back to [@bossert1967mathematical]. So far, many papers focus on DEAs. In recent years, the application range of DEA is wider and wider for more and more suitable parallel computing environments being available. For instance, in [@tosun2013robust; @zidi2013distributed; @bulnes2013parallel; @rocha2014hybrid; @devos2014simultaneous], there are DEAs for problems of different fields. On such a background, it is meaningful to improve migration operator for enhancing DEAs. As one way for this purpose, diversity based migration has been studied in some papers (e.g. [@denzinger2003improving; @power2005promoting; @wei2009maintain; @araujo2011diversity]).
The main motivation of this paper is as follows. According to literature, diversity based migration do help DEAs to get better solutions. However, existing related studies only focus on selecting fit individuals for migration. In fact, diversity based migration may also be realized by selecting migration moments, which is determined only by the parameter, migration interval in traditional DEAs. Nevertheless, such a topic is rarely seen in literature.
A scheme of setting the success rate of migration based on subpopulation diversity at each interval is proposed in this paper for DEAs. The idea behind it is selecting migration moments based on not only interval but also subpopulation diversity. More precisely, migration still occurs at intervals, but the probability of immigrants entering the target subpopulation, $p$, will be decided by the diversity of this subpopulation, $d$, according to a formula building the relationship of them. In detail, at migration intervals, $d$ is computed first. Then, $p$ is calculated according to the formula. At the probability, $p$, immigrants enter the subpopulation .
An experiment is conducted on eight instances of the Travelling Salesman Problem (TSP) from [@reinelt1991tsplib] to compare the outcomes of the DEA with the proposed scheme and those of a traditional DEA. The only difference between them is that the proposed scheme is used only in the former. For the formula in the scheme, two parameters, $\alpha$ and $\beta$, are required. In our experiment, the common parameters in both algorithms are set the same value. Moreover, the algorithm based on the scheme is tested under nine value combinations of $\alpha$ and $\beta$. Results show that the DEA based on our scheme can get better and more stable solutions than the traditional one. Further, under some value combinations of $\alpha$ and $\beta$, the advantage of the algorithm with our scheme is much more remarkable.
The rest of this paper is organized as follows. Related researches are described in Section 2. In Section 3, the proposed scheme is presented. Then, experimental results are shown and analyzed in Section 4. Finally, a conclusion and a prospect are dealt with in Section 5.
Related Researches
==================
Distributed Evolutionary Algorithms
-----------------------------------
A DEA can be considered as the upgrade of a EA to enhance solving ability. In [@herrera1999hierarchical], it is expressed as the Algorithm \[alg1\] whose *italic part* is the migration operation.
Generate at random a population, $P$, of individuals Divide $P$ into $SP_1$, …, $SP_n$ subpopulations Define a neighborhood structure for $SP_x$, $x=1$, …, $n$ Apply, during $i$ generations, stochastic operators of EA *Send $s$ individuals to neighboring subpopulations* *Receive $s$ individuals from neighboring subpopulations*
Main parallel parameters in DEAs are introduced as below:
- Migration topology (neighborhood structurein Algorithm \[alg1\]), it describes which subpopulations send individuals to which subpopulations [@skolicki2005influence].
- Migration strategy, it defines how to select emigrants and choice replaced individuals in each subpopulation.
- Migration rate (rin Algorithm \[alg1\]), it specifies the quantity of individuals which emigrate from each subpopulation in one migration round.
- Migration interval (iin Algorithm \[alg1\]), it is the every certain number of generations between a migration round and the next one [@skolicki2005influence].
- Subpopulation size, it is the quantity of individuals in each subpopulation.
- Quantity of subpopulation (nin Algorithm \[alg1\]), it is the number of subpopulations.
On one hand, semi-isolation can maintain difference among subpopulations. On the other hand, for subpopulations, at intervals, migration is expected to afford individuals whose fitness are close to that of local ones but have some different building blocks [@cohoon1987punctuated]. After migration, immigrants participating in variation operations can help to resist premature convergence if they do have some new building blocks. In this way, stagnation is postponed.
Diversity Based Migration in DEAs
---------------------------------
Diversity refers to differences among individuals, which can be measured at the genotype or phenotype levels [@vcrepinvsek2013exploration]. Thus, diversity of a colony shows the its convergence degree. Though diversity measures at the latter level often demands much less time consumption, they are not as accurate as ones at the former one because the difference on fitness is not equivalent to that on chromosome structure. Some instances on diversity based migration from literature are listed as below. Their diversity measures are all at the genotype level. In [@denzinger2003improving], a weighted sum of fitness and diversity measure is used in a DEA for selecting emigrants and replaced individuals. Results showed a reduction in the time taken to find the optimal solution across a range of benchmark problems. In the DEA presented by [@power2005promoting], the subpopulation representative what is an individual having the lowest average of distance to other all individuals in this subpopulation is found by calculation. Then, it is selected to be the emigrant. Furthermore, if migration rate is set more than one, some other individuals are chosen as emigrants based on the distance from it. This strategy has been tested on a number of problems and has consistently outperformed the standard migration strategy. The scheme in [@wei2009maintain] only allows different enough individuals migrate to a subpopulation in order to keeping diversity of this subpopulation. To realize this scheme, a method for evaluating the similarity of two subpopulations is presented. More recently, multikulti methods, are proposed in [@araujo2011diversity] for selecting individuals in source subpopulation different enough to ones in target subpopulation . In order to do this, information on the composition of source subpopulation is required when emigrants are selected. Different ways of providing this information in a concise manner are considered. The results of experiments prove the usefulness of the multikulti strategies. Besides, The success of this kind of strategies is explained via the measurement of entropy as a representation of population diversity. In conclusion, existing schemes for diversity based migration concentrate on selecting individuals for migration with the guide of diversity.
Proposed Scheme Description
===========================
Preliminary
-----------
Migration may bring new building blocks for target subpopulation, while increases the similarity between source subpopulation and target one inevitably. Hence, there are many occasions not fit for migration during a run. For instance, in a subpopulation, migration may not be required when getting better individuals by variation operators is easy. In such a case, migration have little positive influence but still decreases the difference between subpopulations. Anyhow, migration moments predetermined by intervals cannot match the dynamic situation of evolution. Consequently, inappropriate migration cannot be avoided effectively.
Scheme to Solve Existing Problem
--------------------------------
Our thinking is that more migration chances should be provided when subpopulation diversity is low. In general, the computation for subpopulation diversity is time-consuming in most EAs. Consequently, it is not fit to decide migration moments only based on subpopulation diversity since, in such a plan, the computation for subpopulation diversity has to be executed frequently. Also, setting a value of subpopulation diversity as the threshold for migration is not adopted by us. Two of reasons are listed below. Firstly, the fitting value is difficult to find and may vary much from one occasion to another. Besides, migration may not occur at all in some periods, such as the initial stage. In our scheme, the attempt of migration still occurs at intervals, but the probability of immigrants entering the target subpopulation, $p$, will be computed according to the formula that builds the relationship between subpopulation diversity, $d$, and the probability, $p$. At first, the simplest formula to describe the relationship between $p$ and $d$, Formula \[eq:01\], is designed by us. $$p =1 - d
\label{eq:01}$$ In the initial stage of a run, the value of $d$ is close to the maximum, 1. According to Formula \[eq:01\], $p$ is near to 0 at this time. Then, $d$ decreasing monotonically makes $p$ increasing monotonically. In the end, $p$ approaches 1 since $d$ has come to the minimum, 0. Then, Formula \[eq:02\], where $\alpha\in[0,+\infty]$, and Formula \[eq:03\], $\beta\in[0,+\infty]$ are considered. $$p =1 - d^\alpha
\label{eq:02}$$ $$p =(1 - d)^\beta
\label{eq:03}$$ Finally, Formula \[eq:04\] is adopted since the three former ones can be regarded as special cases of it. $$p =(1 - d^\alpha)^\beta
\label{eq:04}$$ Essentially, $\alpha$ and $\beta$ are just used to adjust the relationship between $p$ and $d$.
The flowchart of a DEA with our scheme is shown as Algorithm \[alg2\]. The scheme is in text of Algorithm \[alg2\].
Generate at random a population, $P$, of individuals Divide $P$ into $SP_1$, …, $SP_n$ subpopulations Define a neighborhood structure for $SP_x$, $x=1$, …, $n$ Apply, during $i$ generations, stochastic operators of EA *Send $s$ individuals to neighboring subpopulations* *Receive $s$ individuals from neighboring subpopulations*
Essentially, our scheme devotes to control the probability of completing migration according to subpopulation diversity at intervals rather than to improve subpopulation diversity.
Analysis of Our Scheme on Running Time
--------------------------------------
Provided that all supopulations in DEAs remains in exact synchronization, the consuming time for the two algorithms can be expressed as below, respectively. Let $\Delta t_e$ be the consuming time of all evolutionary operations in one generation, $\Delta t_d$ be that of once diversity computation, $\Delta t_m$ be that of a migration round, $i$ be interval and $g$ be total generations. In a DEA based on the proposed scheme, the total consuming time in a run, $t_w$, can be computed as follow: $$t_w = (\Delta t_e + \frac{\Delta t_d + \Delta t_m} {i} ) \times g$$ In a DEA without the proposed scheme, $\Delta t'_m$ be the consuming time of a migration round. Then, the total consuming time in a run, $t_n$ can be computed as follow: $$t_n = (\Delta t_e + \frac{\Delta t'_m} {i} ) \times g$$ If above two types of DEA are based on the same EA and same in settings of common parameters, comparisons can be done as follow: Firstly, $$\Delta t_m \le \Delta t'_m$$ mainly because, in the DEA with the scheme, a migration round requires the maximum time, $\Delta t'_m$ , only when $p$ is satisfied. Moreover, the total consuming time of the former algorithm have a peculiar part, $$\frac {\Delta t_d} {i} \times g$$ Therefore, the difference of these two algorithms on total consuming time in a run can be measured as below: $$\frac {\Delta t_d + \Delta t_m - \Delta t'_m} {i} \times g$$ The total consuming time of the former is larger than that of the latter because $$\Delta t_d \gg \Delta t'_m - \Delta t_m$$ That is, consuming time of once diversity computation is much more than the difference in that of a migration round for different algorithms.
In the scheme, subpopulation diversity is computed only at intervals. Although the diversity measure may vary according to chromosome coding, the extra consuming time in a DEA for the proposed scheme will always be acceptable if migration interval, $i$, is large enough. In fact, to maintain the difference between subpopulations, $i$ is always large in the majority of DEAs. Therefore, this scheme can be widely used in DEAs.
Experiment Studies
==================
Selected Problem and EA for It
------------------------------
The famous NP-hard problem, TSP, is selected for our experiment. [@reinelt1991tsplib] has its benchmark instances. Some EA for the TSP have remarkable performance. For instance, the one presented in [@nagata2013powerful] outperforms state-of-the-art heuristic algorithms in finding very high-quality solutions on instances with up to 200,000 cities.
It should be stressed that the purpose of our experiments is not to find better solutions of any problem than ever but to test our scheme by comparing the performance of a DEA with it and that of a traditional one. For this purpose, the more powerful a EA is, the larger instances, which demand much more on resource, should be used. In consideration of this, the EA in [@cai2005improved] proposed for years is used in our experiment. It is based on the inver-over operator [@tao1998inver], the selection method that each individual competes with its offspring only and so-called mapping operator [@cai2005improved]. The flow of mapping is as follow. Firstly, two individuals are selected at random. Then, a segment of chromosome in the individual having worse fitness is selected at random. After that, a segment which has the same number of cities and the same first city is searched in the other individual. As soon as it is found, the former segment in the worse individual is replaced by it. Finally, in the worse individual, the other part of the chromosome is adjusted according to the latter steps of partially mapped crossover [@goldberg1985alleles] which is a traditional operator used in EAs for the TSP. It can be seen from the flow that mapping belongs to crossover. In this EA, evolutionary velocity, $v$, which is calculated as the Formula \[eq:05\], decides whether the mapping operator should be executed. $${v} = \frac{\left|{{f_b} - {f_b}'} \right|}{\Delta g}
\label{eq:05}$$ In this formula, $f_b$ is the fitness of the current best individual, $f'_{b}$ is the that of the previous one and $\Delta g$ denotes the generations between the appearance of the previous best individual and that of the current one. Mapping is carried out only when $v$ is lower than a threshold value.
Diversity Measure in Our DEA
----------------------------
Matrix M in Formula \[eq:mat\] is connection matrix of TSP tour [@chang2010dynamic]. $$\mathbf{M} =
\left[{\begin{array}{cccc}
a_{00} \quad & \quad a_{01} & \quad \cdots & \quad a_{0(k-1)} \\
a_{10} \quad & \quad a_{11} & \quad \cdots & \quad a_{1(k-1)} \\
\cdots \quad & \quad \cdots & \quad \cdots & \quad \cdots \\
a_{(k-1)0} \quad & \quad \cdots & \quad \cdots & \quad a_{(k-1)(k-1)}
\end{array}} \right]
\label{eq:mat}$$ In the matrix, $k$ is the number of cities and $a_{lm} \in \{0,1\}, (0\le l\le k-1, 0\le m\le k-1)$. $a_{lm} = 1$ represents that there is a connection from city$(l+1)$ to city$(m+1)$ in tour, while $a_{lm} = 0$ denotes that such a connection does not exist. Let $I_x$ and $I_y$ be two individuals. Then, each of them has a $k\times k$ connection matrix. Let $k'$ be the number of rows which are same in the two matrixes. Then, the difference between $I_x$ and $I_y$ can be defined as Formula \[eq:diff\] [@chang2010dynamic]. $$D(I_x,I_y) = 1 - \frac{k'}{k}
\label{eq:diff}$$ Then, in [@chang2010dynamic], subpopulation diversity is defined as Formula \[form:five\], where $ni$ is subpopulation size and $C$ means combination. $$d = \frac{\sum{D(I_x,I_y)}}{C_{ni}^2}
\label{form:five}$$ In this paper, Formula \[form:five\] is replaced by Formula \[form:six\] in order to decrease computation complexity. $$d = \frac{\sum{D(I_b,I_n)}}{ni-1}
\label{form:six}$$ In Formula \[form:six\], $I_b$ is the best individual in subpopulation. $I_n$ represents each individual in it other than $I_b$. For once computation of subpopulation diversity, this change makes that total times of the calculation for the difference between two individuals greatly reduce from $C_{ni}^2$ to $ni-1$ in each subpopulation. Although the average of distance between pairwise individuals is replaced by that from the best individual to another one may lead to some error in the resulting value of subpopulation diversity, convergence degree can still be reflected with much less computation. As the steps shown in Algorithm \[alg2\], $d$ of each subpopulation is computed at intervals to obtain each $p$.
Experiment to Compare Result
----------------------------
Our experiment is carried on a Drawing TC5000A computing platform. It has 1264 2.6GHz cores. Its memory capacity is 1.5TB. In our experiment, $\alpha$ and $\beta$ are set different value combinations to find the fit relationship between $p$ and $d$. In detail, their values are both get from the set, $\{0.5,1.0,2.0\}$. In total, there are nine value combinations. Fig. \[figure:01\]-\[figure:09\] are graphs of the function expressed by Formula \[eq:04\] under different value combination of $\alpha$ and $\beta$, respectively. Then, the outcomes under each value combination are compared with those of a traditional DEA, respectively. The algorithms in the experiment are both based on the EA introduced in Subsection 4.1 and are the same in setting of common parameters. The thinking of setting for these common parameters comes from [@li2014global]. In detail, common parameters except migration interval are set the same value in both algorithms. Then, for each instance, the traditional algorithm runs thirty times independently under five equal difference intervals, respectively. Under each of the nine value combinations of $\alpha$ and $\beta$, so does the DEA with our scheme. The value of common parameters except migration interval is listed in Table \[table:01\].
\[table:01\]
[|m[1.4cm]{}<|m[2.5cm]{} <|m[3.6cm]{}|]{} \[20\][2.0cm]{}[Evolutionary parameters]{} & Mutation rate ($p_{mu}$) & Changing during a run according to Formula \[form:seven\] based on the initial value, 0.02.\
& Crossover rate & 1-$p_{mu}$\
& Mapping rate ($p_{ma}$) & Changing during a run according to Formula \[form:eight\] based on the initial value, 0.05.\
& Threshold value of evolutionary velocity & 5000\
\[7\][2.0cm]{}[Parallel parameters]{} & Migration topology & Ring\
& Migration strategy & Random-random\
& Migration size & 1\
& Subpopulation size & 100\
& Quantity of subpopulation & 16\
& Terminal criterion & 2000 migration rounds having been done\
This table shows that mutation rate, $p_{mu}$, and mapping rate, $p_{ma}$, change during a run according to Formula \[form:seven\] and Formula \[form:eight\], respectively. $$p_{mu} = {p_{mu0}} \times (1-\frac{g_n}{g}\times 0.5)
\label{form:seven}$$ $$p_{ma} = {p_{ma0}} \times (\frac{{{g_n} \times 2}}{g} + 1)
\label{form:eight}$$ In Formula \[form:seven\], $p_{mu0}$ is the initial $p_{mu}$. Similarly, $p_{ma0}$ is the initial $p_{ma}$ in Formula \[form:eight\]. In both formulas, $g_n$ denotes current generations and $g$ represents the maximal ones.
Since the EA used in our experiment can find the optimum solution of many TSP instances smaller than a280, eight much larger instances in [@reinelt1991tsplib] from pcb442 to vm1084 are chosen in our experiment. The results of the traditional algorithm are listed in Table \[table:02\]. Those of the DEA based on our scheme under different value combinations of $\alpha$ and $\beta$ are listed in Table \[table:03\]-\[table:11\], respectively. In these tables, each instance corresponds to five migration intervals. Under each interval, the average of outcomes and the standard deviation of them is given. Besides, the optimal solution of each instance provided by [@reinelt1991tsplib] is listed in each table. In Table \[table:03\]-\[table:11\], results having significant difference with those in Table \[table:02\] in terms of t-test with 95% confidence are highlighted by ***italics and bold***.
\[table:02\]
[|c|c|c|m[1.5cm]{}<|c|]{} Instance & Interval &
----------
Outcomes
average
----------
: Results of traditional DEA
&
-----------
Standard
deviation
-----------
: Results of traditional DEA
&
----------
Optimal
solution
----------
: Results of traditional DEA
\
\*[pcb442]{} &150000 &50939.3 &7.01 &\
&200000 &50937.7 &7.84 &\
&250000 &50935.0 &24.85 & 50778\
&300000 &50937.5 &7.54 &\
&350000 &50934.6 &10.94 &\
\*[p654]{} &50000 &34643.4 &1.22 &\
&60000 &34643.4 &1.22 &\
&70000 &34643.5 &1.57 & 34643\
&80000 &34643.0 &0.00 &\
&90000 &34643.1 &0.37 &\
\*[d657]{} &200000 &49092.7 &36.28 &\
&250000 &49084.5 &27.21 &\
&300000 &49063.9 &39.07 &48912\
&350000 &49083.2 &34.98 &\
&400000 &49065.3 &35.79 &\
\*[u724]{} &150000 &42143.3 &43.23 &\
&200000 &42105.0 &37.78 &\
&250000 &42086.1 &43.66 &41910\
&300000 &42085.6 &32.77 &\
&350000 &42068.8 &33.05 &\
\*[rat783]{} &150000 &8829.1 &11.71 &\
&200000 &8821.8 &9.11 &\
&250000 &8818.9 &7.82 &8806\
&300000 &8815.6 &5.97 &\
&350000 &8815.6 &7.50 &\
\*[dsj1000]{} &800000 &18786697.6 &27651.24 &\
&1000000 &18756550.5 &20746.29 &\
&1200000 &18761931.7 &19044.51 &18659688\
&1400000 &18762151.6 &24886.79 &\
&1600000 &18757890.0 &21787.65 &\
\*[pr1002]{} &500000 &259680.9 &253.01 &\
&600000 &259551.0 &233.13 &\
&700000 &259553.3 &249.21 & 259045\
&800000 &259399.3 &159.45 &\
&900000 &259459.9 &208.12 &\
\*[vm1084]{} &600000 &239785.1 &204.25 &\
&800000 &239773.2 &177.70 &\
&1000000 &239779.5 &187.50 & 239297\
&1200000 &239738.4 &178.93 &\
&1400000 &239739.5 &188.81 &\
\[table:03\]
[|c|c|c|m[1.5cm]{}<|c|]{} Instance & Interval &
----------
Outcomes
average
----------
: Results of DEA with our scheme when $\alpha=0.5$ and $\beta=0.5$
&
-----------
Standard
deviation
-----------
: Results of DEA with our scheme when $\alpha=0.5$ and $\beta=0.5$
&
----------
Optimal
solution
----------
: Results of DEA with our scheme when $\alpha=0.5$ and $\beta=0.5$
\
\*[pcb442]{} &150000 &50936.8 &8.37 &\
&200000 &50935.6 &8.11 &\
&250000 &50927.8 &24.62 & 50778\
&***300000*** &***50924.8*** &***33.10*** &\
&350000 &50934.5 &12.86 &\
\*[p654]{} &50000 &34643.1 &0.51&\
&60000 &34643.1 &0.37 &\
&70000 &34643.0 &0.00 & 34643\
&80000 &34643.1 &0.37 &\
&90000 &34643.1 &0.37 &\
\*[d657]{} &200000 &49077.9 &32.29 &\
&***250000*** &***49057.9*** &***32.71*** &\
&300000 &49059.6 &31.23 &48912\
&***350000*** &***49057.0*** &***30.24*** &\
&400000 &49059.1 &45.58 &\
\*[u724]{} &***150000*** &***42099.2*** &***31.61*** &\
&***200000*** &***42074.6*** &***33.18*** &\
&250000 &42073.3 &40.11 &41910\
&***300000*** &***42064.9*** &***23.90*** &\
&350000 &42060.3 &36.93 &\
\*[rat783]{} &150000 &8825.0 &7.97 &\
&***200000*** &***8817.3*** &***7.62*** &\
&250000 &8815.9 &8.28 &8806\
&300000 &8814.5 &6.23 &\
&350000 &8814.8 &5.52 &\
\*[dsj1000]{} &***800000*** &***18761927.0*** &***25664.25*** &\
&1000000 &18762398.9 &23733.48 &\
&***1200000*** &***18748623.4*** &***24506.86*** &18659688\
&*1400000* &*18745199.9* &*27584.08* &\
&*1600000* &*18744510.3* &*23523.72* &\
\*[pr1002]{} &500000 &259551.6 &268.12 &\
&600000 &259504.9 &223.44 &\
&***700000*** &***259383.7*** &***172.11*** &259045\
&800000 &259375.9 &233.52 &\
&900000 &259387.2 &239.40 &\
\*[vm1084]{} &600000 &239736.0 &164.15 &\
&***800000*** &***239676.7*** &***141.30*** &\
&1000000 &239702.1 &163.49 & 239297\
&1200000 &239682.9 &179.91 &\
&***1400000*** &***239637.6*** &***122.91*** &\
![Graph of function for our scheme when $\alpha=0.5$ and $\beta=0.5$ []{data-label="figure:01"}](function1){width="3.0in"}
[|c|c|c|m[1.5cm]{}<|c|]{} Instance & Interval &
----------
Outcomes
average
----------
: Results of DEA with our scheme when $\alpha=0.5$ and $\beta=1.0$[]{data-label="table:04"}
&
-----------
Standard
deviation
-----------
: Results of DEA with our scheme when $\alpha=0.5$ and $\beta=1.0$[]{data-label="table:04"}
&
----------
Optimal
solution
----------
: Results of DEA with our scheme when $\alpha=0.5$ and $\beta=1.0$[]{data-label="table:04"}
\
\*[pcb442]{} &***150000*** &***50931.7*** &***12.66*** &\
&200000 &50934.3 &10.46 &\
&250000 &50929.5 &29.45 & 50778\
&***300000*** &***50932.7*** &***10.44*** &\
&350000 &50927.1 &22.43 &\
\*[p654]{} &50000 &34643.0 &0.00 &\
&60000 &34643.0 &0.00 &\
&70000 &34643.0 &0.00 & 34643\
&80000 &34643.0 &0.00 &\
&90000 &34643.0 &0.00 &\
\*[d657]{} &***200000*** &***49055.4*** &***35.44*** &\
&***250000*** &***49057.3*** &***42.35*** &\
&300000 &49048.0 &38.85 &48912\
&***350000*** &***49045.9*** &***37.06*** &\
&***400000*** &***49036.6*** &***32.54*** &\
\*[u724]{} &***150000*** &***42070.7*** &***33.66*** &\
&***200000*** &***42067.9*** &***29.89*** &\
&***250000*** &***42054.6*** &***31.99*** &41910\
&***300000*** &***42037.9*** &***39.68*** &\
&***350000*** &***42036.7*** &***35.86*** &\
\*[rat783]{} &*150000* &*8822.8* &*9.91* &\
&*200000* &*8815.8* &*7.32* &\
&250000 &8816.2 &7.01 &8806\
&300000 &8812.9 &5.72 &\
&350000 &8812.9 &8.13 &\
\*[dsj1000]{} &***800000*** &***18756031.3*** &***24623.43*** &\
&1000000 &18748806.2 &26546.89 &\
&***1200000*** &***18748989.6*** &***22840.60*** &18659688\
&***1400000*** &***18735264.4*** &***20742.59*** &\
&***1600000*** &***18738518.6*** &***26541.26*** &\
\*[pr1002]{} &***500000*** &***259433.6*** &***190.88*** &\
&***600000*** &***259335.0*** &***165.78*** &\
&***700000*** &***259303.5*** &***140.85*** & 259045\
&***800000*** &***259280.7*** &***138.09*** &\
&***900000*** &***259354.0*** &***199.96*** &\
\*[vm1084]{} &***600000*** &***239683.8*** &***162.64*** &\
&***800000*** &***239665.7*** &***136.93*** &\
&***1000000*** &***239671.2*** &***135.21*** & 239297\
&***1200000*** &***239638.1*** &***145.54*** &\
&***1400000*** &***239623.6*** &***126.22*** &\
![Graph of function for our scheme when $\alpha=0.5$ and $\beta=1.0$ []{data-label="figure:02"}](function2){width="3.0in"}
[|c|c|c|m[1.5cm]{}<|c|]{} Instance & Interval &
----------
Outcomes
average
----------
: Results of DEA with our scheme when $\alpha=0.5$ and $\beta=2.0$[]{data-label="table:05"}
&
-----------
Standard
deviation
-----------
: Results of DEA with our scheme when $\alpha=0.5$ and $\beta=2.0$[]{data-label="table:05"}
&
----------
Optimal
solution
----------
: Results of DEA with our scheme when $\alpha=0.5$ and $\beta=2.0$[]{data-label="table:05"}
\
\*[pcb442]{} &***150000*** &***50932.6*** &***11.09*** &\
&***200000*** &***50930.4*** &***10.13*** &\
&250000 &50926.6 &27.04 & 50778\
&***300000*** &***50919.6*** &***32.26*** &\
&***350000*** &***50922.3*** &***23.90*** &\
\*[p654]{} &50000 &34643.0 &0.00 &\
&60000 &34643.0 &0.00 &\
&70000 &34643.0 &0.00 & 34643\
&80000 &34643.0 &0.00 &\
&90000 &34643.0 &0.00 &\
\*[d657]{} &***200000*** &***49050.0*** &***31.05*** &\
&***250000*** &***49054.0*** &***41.66*** &\
&300000 &49045.9 &34.37 &48912\
&***350000*** &***49028.7*** &***31.89*** &\
&***400000*** &***49042.3*** &***36.07*** &\
\*[u724]{} &***150000*** &***42072.2*** &***31.06*** &\
&***200000*** &***42045.2*** &***27.53*** &\
&***250000*** &***42045.9*** &***30.77*** &41910\
&***300000*** &***42032.3*** &***29.95*** &\
&***350000*** &***42026.7*** &***26.53*** &\
\*[rat783]{} &150000 &8828.1 &8.83 &\
&200000 &8817.9 &6.69 &\
&***250000*** &***8813.9*** &***6.77*** &8806\
&300000 &8815.0 &6.24 &\
&350000 &8813.4 &6.44 &\
\*[dsj1000]{} &***800000*** &***18741740.2*** &***22220.14*** &\
&***1000000*** &***18739983.0*** &***24819.99*** &\
&***1200000*** &***18738972.7*** &***18942.56*** &18659688\
&***1400000*** &***18724773.5*** &***27336.00*** &\
&***1600000*** &***18732647.5*** &***20957.86*** &\
\*[pr1002]{} &***500000*** &***259277.2*** &***136.14*** &\
&***600000*** &***259402.0*** &***201.64*** &\
&***700000*** &***259342.0*** &***189.56*** & 259045\
&***800000*** &***259296.1*** &***180.98*** &\
&***900000*** &***259288.9*** &***132.95*** &\
\*[vm1084]{} &600000 &239691.4 &169.62 &\
&800000 &239690.1 &167.67 &\
&***1000000*** &***239637.8*** &***128.26*** & 239297\
&***1200000*** &***239638.7*** &***171.76*** &\
&***1400000*** &***239585.6*** &***133.70*** &\
![Graph of function for our scheme when $\alpha=0.5$ and $\beta=2.0$ []{data-label="figure:03"}](function3){width="3.0in"}
[|c|c|c|m[1.5cm]{}<|c|]{} Instance & Interval &
----------
Outcomes
average
----------
: Results of DEA with our scheme when $\alpha=1.0$ and $\beta=0.5$[]{data-label="table:06"}
&
-----------
Standard
deviation
-----------
: Results of DEA with our scheme when $\alpha=1.0$ and $\beta=0.5$[]{data-label="table:06"}
&
----------
Optimal
solution
----------
: Results of DEA with our scheme when $\alpha=1.0$ and $\beta=0.5$[]{data-label="table:06"}
\
\*[pcb442]{} &150000 &50938.3 &6.86 &\
&200000 &50937.4 &8.53 &\
&250000 &50934.7 &10.22 & 50778\
&300000 &50933.5 &27.43 &\
&350000 &50921.9 &35.34 &\
\*[p654]{} &50000 &34643.3 &1.18 &\
&60000 &34643.6 &2.30 &\
&70000 &34643.1 &0.37 & 34643\
&80000 &34643.0 &0.00 &\
&90000 &34643.0 &0.00 &\
\*[d657]{} &200000 &49076.7 &42.28 &\
&250000 &49083.5 &33.96 &\
&300000 &49070.3 &32.63 &48912\
&***350000*** &***49060.2*** &***29.56*** &\
&400000 &49058.7 &42.51 &\
\*[u724]{} &***150000*** &***42093.8*** &***39.47*** &\
&***200000*** &***42081.5*** &***31.40*** &\
&250000 &42081.8 &30.52 &41910\
&***300000*** &***42064.9*** &***28.86*** &\
&350000 &42071.3 &30.47 &\
\*[rat783]{} &150000 &8825.9 &10.50 &\
&200000 &8818.5 &8.52 &\
&250000 &8816.4 &6.64 &8806\
&300000 &8814.4 &6.72 &\
&350000 &8814.0 &7.31 &\
\*[dsj1000]{} &***800000*** &***18773122.2*** &***23162.37*** &\
&1000000 &18761402.1 &26171.98 &\
&1200000 &18755819.7 &15957.70 &18659688\
&***1400000*** &***18746441.0*** &***23247.20*** &\
&1600000 &18752288.1 &19447.99 &\
\*[pr1002]{} &***500000*** &***259517.0*** &***273.21*** &\
&600000 &259468.7 &236.99 &\
&700000 &259449.1 &199.82 & 259045\
&***800000*** &***259515.5*** &***272.15*** &\
&900000 &259499.3 &226.40 &\
\*[vm1084]{} &600000 &239770.3 &162.83 &\
&800000 &239730.6 &137.43 &\
&1000000 &239693.5 &151.11& 239297\
&1200000 &239694.1 &158.66 &\
&1400000 &239670.1 &132.08 &\
![Graph of function for our scheme when $\alpha=1.0$ and $\beta=0.5$ []{data-label="figure:04"}](function4){width="3.0in"}
[|c|c|c|m[1.5cm]{}<|c|]{} Instance & Interval &
----------
Outcomes
average
----------
: Results of DEA with our scheme when $\alpha=1.0$ and $\beta=1.0$[]{data-label="table:07"}
&
-----------
Standard
deviation
-----------
: Results of DEA with our scheme when $\alpha=1.0$ and $\beta=1.0$[]{data-label="table:07"}
&
----------
Optimal
solution
----------
: Results of DEA with our scheme when $\alpha=1.0$ and $\beta=1.0$[]{data-label="table:07"}
\
\*[pcb442]{} &150000 &50931.5 &20.76 &\
&200000 &50929.5 &26.70 &\
&250000 &50933.9 &12.78 & 50778\
&***300000*** &***50931.7*** &***10.29*** &\
&350000 &50931.2 &22.20 &\
\*[p654]{} &50000 &34643.1 &0.37 &\
&60000 &34643.1 &0.37 &\
&70000 &34643.4 &1.52 & 34643\
&80000 &34643.0 &0.00 &\
&90000 &34643.0 &0.00 &\
\*[d657]{} &***200000*** &***49067.4*** &***43.30*** &\
&***250000*** &***49063.7*** &***37.62*** &\
&300000 &49055.5 &36.67 &48912\
&***350000*** &***49050.3*** &***32.32*** &\
&400000 &49046.9 &43.36 &\
\*[u724]{} &***150000*** &***42094.9*** &***29.82*** &\
&***200000*** &***42064.9*** &***42.94*** &\
&250000 &42086.1 &42067.8 &34.59\
&***300000*** &***42066.3*** &***25.19*** &\
&350000 &42054.6 &30.69 &\
\*[rat783]{} &150000 &8824.0 &8.83 &\
&200000 &8819.7 &7.47 &\
&***250000*** &***8815.1*** &***6.77*** &8806\
&300000 &8814.0 &7.05 &\
&350000 &8812.5 &6.69 &\
\*[dsj1000]{} &***800000*** &***18763625.5*** &***21831.34*** &\
&1000000 &18762785.8 &24970.92 &\
&***1200000*** &***18746381.2*** &***20736.39*** &18659688\
&***1400000*** &***18745791.7*** &***17386.04*** &\
&***1600000*** &***18743302.0*** &***18939.43*** &\
\*[pr1002]{} &***500000*** &***259435.1*** &***224.69*** &\
&***600000*** &***259417.3*** &***222.38*** &\
&***700000*** &***259359.6*** &***169.53*** & 259045\
&800000 &259341.7 &198.52 &\
&***900000*** &***259317.7*** &***187.90*** &\
\*[vm1084]{} &600000 &239739.4 &183.33 &\
&800000 &239694.4 &158.19 &\
&1000000 &239691.5 &168.34 & 239297\
&1200000 &239699.0 &143.17 &\
&1400000 &239668.0 &129.28 &\
![Graph of function for our scheme when $\alpha=1.0$ and $\beta=1.0$ []{data-label="figure:05"}](function5){width="3.0in"}
[|c|c|c|m[1.5cm]{}<|c|]{} Instance & Interval &
----------
Outcomes
average
----------
: Results of DEA with our scheme when $\alpha=1.0$ and $\beta=2.0$[]{data-label="table:08"}
&
-----------
Standard
deviation
-----------
: Results of DEA with our scheme when $\alpha=1.0$ and $\beta=2.0$[]{data-label="table:08"}
&
----------
Optimal
solution
----------
: Results of DEA with our scheme when $\alpha=1.0$ and $\beta=2.0$[]{data-label="table:08"}
\
\*[pcb442]{} &***150000*** &***50934.0*** &***10.64*** &\
&***200000*** &***50931.3*** &***11.15*** &\
&250000 &50932.7 &12.33 & 50778\
&***300000*** &***50931.5*** &***10.96*** &\
&***350000*** &***50927.6*** &***14.41*** &\
\*[p654]{} &50000 &34643.0 &0.00 &\
&60000 &34643.0 &0.00 &\
&70000 &34643.0 &0.00 & 34643\
&80000 &34643.1 &0.37 &\
&90000 &34643.0 &0.00 &\
\*[d657]{} &***200000*** &***49052.7*** &***35.84*** &\
&***250000*** &***49053.1*** &***36.05*** &\
&300000 &49050.2 &34.89 &48912\
&***350000*** &***49031.9*** &***34.44*** &\
&***400000*** &***49041.6*** &***47.01*** &\
\*[u724]{} &***150000*** &***42070.6*** &***37.78*** &\
&***200000*** &***42056.0*** &***37.16*** &\
&***250000*** &***42035.9*** &***27.50*** &41910\
&***300000*** &***42031.8*** &***30.09*** &\
&***350000*** &***42033.0*** &***30.42*** &\
\*[rat783]{} &150000 &8824.6 &9.78 &\
&200000 &8817.7 &8.45 &\
&***250000*** &***8814.6*** &***6.62*** &8806\
&300000 &8812.7 &6.50 &\
&350000 &8814.3 &6.14 &\
\*[dsj1000]{} &***800000*** &***18740584.7*** &***20775.90*** &\
&***1000000*** &***18743355.3*** &***20831.26*** &\
&***1200000*** &***18741427.6*** &***24245.91*** &18659688\
&***1400000*** &***18729556.2*** &***19250.64*** &\
&***1600000*** &***18733965.3*** &***25656.26*** &\
\*[pr1002]{} &***500000*** &***259394.7*** &***204.43*** &\
&***600000*** &***259367.8*** &***198.61*** &\
&***700000*** &***259331.4*** &***180.84*** & 259045\
&800000 &259410.4 &179.61 &\
&***900000*** &***259307.4*** &***181.98*** &\
\*[vm1084]{} &600000 &239724.6 &145.35 &\
&800000 &239717.7 &149.77 &\
&***1000000*** &***239638.4*** &***148.45*** & 239297\
&***1200000*** &***239642.8*** &***116.59*** &\
&***1400000*** &***239644.8*** &***142.91*** &\
![Graph of function for our scheme when $\alpha=1.0$ and $\beta=2.0$ []{data-label="figure:06"}](function6){width="3.0in"}
[|c|c|c|m[1.5cm]{}<|c|]{} Instance & Interval &
----------
Outcomes
average
----------
: Results of DEA with our scheme when $\alpha=2.0$ and $\beta=0.5$[]{data-label="table:09"}
&
-----------
Standard
deviation
-----------
: Results of DEA with our scheme when $\alpha=2.0$ and $\beta=0.5$[]{data-label="table:09"}
&
----------
Optimal
solution
----------
: Results of DEA with our scheme when $\alpha=2.0$ and $\beta=0.5$[]{data-label="table:09"}
\
\*[pcb442]{} &***150000*** &***50933.8*** &***9.27*** &\
&200000 &50936.4 &8.19 &\
&250000 &50934.8 &10.53 & 50778\
&300000 &50936.1 &9.49 &\
&350000 &50937.4 &8.42 &\
\*[p654]{} &50000 &34643.4 &1.52 &\
&60000 &34643.1 &0.37 &\
&70000 &34643.1 &0.51 & 34643\
&80000 &34643.3 &1.14 &\
&90000 &34643.0 &0.00 &\
\*[d657]{} &200000 &49085.9 &38.80 &\
&250000 &49095.7 &30.77 &\
&300000 &49075.7 &29.14 &48912\
&350000 &49065.2 &40.71 &\
&400000 &49066.2 &32.46 &\
\*[u724]{} &***150000*** &***42113.1*** &***41.00*** &\
&200000 &42097.8 &41.37 &\
&250000 &42084.7 &43.12 &41910\
&300000 &42076.8 &30.95 &\
&350000 &42077.5 &37.93 &\
\*[rat783]{} &150000 &8828.0 &8.28 &\
&200000 &8818.5 &6.06 &\
&***250000*** &***8814.8*** &***6.23*** &8806\
&300000 &8816.0 &5.43 &\
&350000 &8814.6 &5.70 &\
\*[dsj1000]{} &800000 &18776618.6 &27998.28 &\
&1000000 &18761359.1 &26059.44 &\
&1200000 &18753840.4 &22750.80 &18659688\
&1400000 &18753787.1 &26156.22 &\
&1600000 &18752323.7 &21359.45 &\
\*[pr1002]{} &500000 &259622.7 &228.65 &\
&600000 &259523.0 &314.84 &\
&700000 &259479.1 &211.93 & 259045\
&***800000*** &***259500.5*** &***208.99*** &\
&900000 &259519.0 &196.58 &\
\*[vm1084]{} &600000 &239765.1 &122.05 &\
&800000 &239744.2 &134.48 &\
&***1000000*** &***239774.2*** &***153.76*** & 239297\
&1200000 &239675.7 &164.24 &\
&1400000 &239661.6 &125.02 &\
![Graph of function for our scheme when $\alpha=2.0$ and $\beta=0.5$ []{data-label="figure:07"}](function7){width="3.0in"}
[|c|c|c|m[1.5cm]{}<|c|]{} Instance & Interval &
----------
Outcomes
average
----------
: Results of DEA with our scheme when $\alpha=2.0$ and $\beta=1.0$[]{data-label="table:10"}
&
-----------
Standard
deviation
-----------
: Results of DEA with our scheme when $\alpha=2.0$ and $\beta=1.0$[]{data-label="table:10"}
&
----------
Optimal
solution
----------
: Results of DEA with our scheme when $\alpha=2.0$ and $\beta=1.0$[]{data-label="table:10"}
\
\*[pcb442]{} &***150000*** &***50934.5*** &***9.73*** &\
&200000 &50935.6 &25.76 &\
&250000 &50937.1 &8.39 & 50778\
&300000 &50935.8 &8.25 &\
&350000 &50935.2 &11.55 &\
\*[p654]{} &50000 &34643.3 &0.69 &\
&60000 &34643.1 &0.37 &\
&70000 &34643.0 &0.00 & 34643\
&80000 &34643.0 &0.00 &\
&90000 &34643.1 &0.37 &\
\*[d657]{} &200000 &49078.2 &42.55 &\
&250000 &49078.7 &47.63 &\
&300000 &49063.7 &42.80 &48912\
&350000 &49068.5 &41.41 &\
&400000 &49065.7 &40.75 &\
\*[u724]{} &***150000*** &***42112.0*** &***30.64*** &\
&***200000*** &***42081.4*** &***26.93*** &\
&250000 &42080.1 &38.89 &41910\
&300000 &42069.0 &38.67 &\
&350000 &42064.8 &27.71 &\
\*[rat783]{} &150000 &8827.5 &8.08 &\
&***200000*** &***8817.2*** &***4.93*** &\
&250000 &8816.7 &8.42 &8806\
&300000 &8816.5 &6.18 &\
&350000 &8815.4 &5.77 &\
\*[dsj1000]{} &***800000*** &***18762595.7*** &***30773.90*** &\
&1000000 &18765975.2 &33663.58 &\
&1200000 &18759551.7 &22376.08 &18659688\
&1400000 &18755129.5 &24348.98 &\
&1600000 &18753479.1 &21177.11 &\
\*[pr1002]{} &500000 &259586.6 &235.77 &\
&600000 &259520.7 &222.06 &\
&700000 &259454.7 &244.08 & 259045\
&800000 &259351.2 &196.07 &\
&900000 &259401.3 &166.26 &\
\*[vm1084]{} &600000 &239746.5 &145.48 &\
&800000 &239690.4 &160.28 &\
&1000000 &239742.5 &204.31 & 239297\
&1200000 &239709.6 &160.58 &\
&1400000 &239715.3 &155.95 &\
![Graph of function for our scheme when $\alpha=2.0$ and $\beta=1.0$[]{data-label="figure:08"}](function8){width="3.0in"}
[|c|c|c|m[1.5cm]{}<|c|]{} Instance & Interval &
----------
Outcomes
average
----------
: Results of DEA with our scheme when $\alpha=2.0$ and $\beta=2.0$[]{data-label="table:11"}
&
-----------
Standard
deviation
-----------
: Results of DEA with our scheme when $\alpha=2.0$ and $\beta=2.0$[]{data-label="table:11"}
&
----------
Optimal
solution
----------
: Results of DEA with our scheme when $\alpha=2.0$ and $\beta=2.0$[]{data-label="table:11"}
\
\*[pcb442]{} &150000 &50937.6 &8.36 &\
&200000 &50938.5 &8.21 &\
&250000 &50935.4 &10.23 & 50778\
&300000 &50934.7 &10.39 &\
&350000 &50924.3 &31.84 &\
\*[p654]{} &50000 &34643.1 &0.51 &\
&60000 &34643.0 &0.00 &\
&70000 &34643.0 &0.00 & 34643\
&80000 &34643.0 &0.00 &\
&90000 &34643.0 &0.00 &\
\*[d657]{} &***200000*** &***49060.4*** &***42.18*** &\
&***250000*** &***49054.0*** &***35.61*** &\
&300000 &49046.5 &39.59 &48912\
&***350000*** &***49050.9*** &***30.57*** &\
&400000 &49052.1 &35.16 &\
\*[u724]{} &***150000*** &***42080.9*** &***27.08*** &\
&200000 &42089.8 &35.98 &\
&250000 &42078.1 &35.95 &41910\
&300000 &42071.6 &42.63 &\
&350000 &42052.2 &39.33 &\
\*[rat783]{} &***150000*** &***8821.6*** &***9.10*** &\
&200000 &8819.5 &8.56 &\
&250000 &8816.2 &6.56 &8806\
&300000 &8815.4 &6.30 &\
&350000 &8814.8 &5.12 &\
\*[dsj1000]{} &***800000*** &***18761976.8*** &***25253.82*** &\
&1000000 &18760042.3 &14768.00 &\
&1200000 &18755479.9 &22505.88 &18659688\
&***1400000*** &***18748354.6*** &***22603.43*** &\
&1600000 &18746204.2 &23893.81 &\
\*[pr1002]{} &500000 &259537.5 &306.89 &\
&***600000*** &***259412.2*** &***244.67*** &\
&700000 &259464.5 &237.79 & 259045\
&800000 &259410.7 &225.29 &\
&900000 &259372.2 &182.52 &\
\*[vm1084]{} &600000 &239789.1 &252.07 &\
&800000 &239706.1 &172.80 &\
&1000000 &239746.1 &168.84 & 239297\
&***1200000*** &***239653.7*** &***146.65*** &\
&1400000 &239660.9 &140.98 &\
![Graph of function for our scheme when $\alpha=2.0$ and $\beta=2.0$ []{data-label="figure:09"}](function9){width="3.0in"}
Based on the results in Table \[table:02\]-\[table:11\], the difficulty of a instance to the two algorithms, DF, can be expressed according to Formula \[form:nine\], where $f_{ave}$ denotes the average of all solutions for one instance obtained in our experiment and $f_o$ represents the optimal solution provided by [@reinelt1991tsplib]. $$DF = \frac{f_{ave}-f_o}{f_o}
\label{form:nine}$$
The value of difficulty is listed in Table \[table:12\].
Instance pcb442 p654 d657 u724
------------ ---------- ---------- ---------- ----------
Difficulty 0.003048 0.000003 0.003040 0.003842
Instance rat783 dsj1000 pr1002 vm1084
Difficulty 0.001337 0.004947 0.001482 0.001682
: Difficulty of each instance for DEAs[]{data-label="table:12"}
According to Table \[table:12\], the difficulty of each instance can be classified into three levels. That of p654 belongs to the lowest level. Rat783, pr1002 and vm1084 have an intermediate level difficulty. The rest four instances, pcb442, d657, u724, dsj1000, are ones with high difficulty.
It can be seen in Table \[table:02\]-\[table:11\] that, for the lowest difficulty instance, p654, significant difference between solutions of the DEA based on our scheme and those of the traditional one can be found in no case. For the intermediate difficulty instances, the DEA with the proposed scheme significantly wins in forty-five cases out of one hundred and thirty-five ones (45/135) and statistically loses in two cases (2/135). It should be noted that, for high difficulty instances, the DEA with our scheme yields significantly better outcomes than its peer in ninety cases out of one hundred and eighty ones (90/180). Meanwhile, there are no significant differences in all the rest cases. Also, the tables show that the solutions’ standard deviation of the DEA with our scheme is less than that of the traditional one in two hundred and twenty-six cases out of the all ones (226/360).
Moreover, it can been seen that, under different value combinations of $\alpha$ and $\beta$, the performance of the algorithm with our scheme is significant different. When $\alpha=0.5$, $\beta=1.0$ or $\alpha=0.5$, $\beta=2.0$, for the all eight instances, the algorithm significantly wins in the most cases (27/40) and never statistically loses. When $\alpha=1.0$, $\beta=2.0$, the DEA with our scheme also has a good behavior. In detail, it significantly wins in twenty-six cases (26/40) and never statistically loses. On the whole, the winning rate is 135/360, while the losing rate is 2/360. Besides, under the three outstanding value combinations, solutions of the DEA with our scheme have a better standard deviation in eighty-two cases out of one hundred and twenty ones (82/120). This rate is better than that on the whole (226/360). Fig. \[figure:01\]-\[figure:09\] show that, under the three outstanding value combinations, the graph of the function in Formula \[eq:04\] have the same characteristic which is distinguished from that under the other six ones.
In conclusion, results shows that the DEA based on our scheme has an advantage on solutions. Moreover, a value combination of $\alpha$ and $\beta$ which makes the graphing slope of the function in Formula \[eq:04\] increase monotonically from $-\infty$ to zero when $d\in [0,1]$ is fit for our scheme.
Conclusion
==========
In this paper, we have presented the scheme of setting the success rate of migration based on subpopulation diversity at each interval for DEAs. Under the control of the scheme, immigrants enter the target subpopulation at a probability, which is the function of subpopulation diversity according to Formula \[eq:04\], at intervals. In our experiment, eight instances of the TSP are used to test algorithms. In detail, under nine different value combinations of parameters required for Formula \[eq:04\], outcomes of the algorithm with the scheme are compared with those of the traditional DEA, respectively. The experimental results show that, especially for high difficulty instances, the DEA based on our scheme has a significant advantage on solutions. Moreover, under three value combinations of the parameters making graphing slope of the function in Formula \[eq:04\] increase monotonically from $-\infty$ to zero when $d\in [0,1]$, the algorithm with the scheme has most outstanding performance.
Based on the key factor in the behavior of evolutionary computation, diversity, we propose this scheme and discuss its parameters setting. This scheme can be used to improve solutions of DEAs for diversified problems. To apply the scheme, first of all, finding the diversity measure of a certain chromosome coding is necessary. In future, how to use this scheme together with some other diversity based methods should be studied to further improve DEAs. Also, the simplified calculation method for subpopulation diversity to reduce the time complexity used in our scheme should be discussed in theory.
Acknowledgment {#acknowledgment .unnumbered}
==============
The authors would like to thank Dr. Dunhui Xiao and assistant researcher Jian Wang for their valuable suggestions.
[^1]: C. Li is with the School of Computer Science, China University of Geosciences, Wuhan 430074, China, and the Centre for Quantum Computation and Intelligent Systems, Faculty of Engineering & Information Technology, University of Technology Sydney, Australia (e-mail: [email protected].)
[^2]: J. Wu is with Centre for Quantum Computation and Intelligent Systems, Faculty of Engineering & Information Technology, University of Technology Sydney, Australia (e-mail: [email protected])
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A BPS limit is systematically derived for straight multi- D- and DF-strings from the D3$\bar{\mathrm{D}}3$ system in the context of boundary superstring field theory. The BPS limit is obtained in the limit of thin D(F)-strings, where the Bogomolnyi equation supports singular static multi-D(F)-string solutions. For the BPS multi-string configurations with arbitrary separations, BPS sum rule is fulfilled under a Gaussian type tachyon potential and reproduces exactly the descent relation. For the DF-strings ($(p,q)$-strings), the distribution of fundamental string charge density coincides with its energy density and the Hamiltonian density takes the BPS formula of square-root form.'
---
=.22in
[hep-th/0703144]{}
[[**BPS Limit of Multi- D- and DF-strings\
in Boundary String Field Theory** ]{}\
Gyungchoon Go, Akira Ishida, Yoonbai Kim\
[*Department of Physics, Sungkyunkwan University, Suwon 440-746, Korea*]{}\
[gcgo, ishida, [email protected]]{} ]{}
Introduction
============
When the system of D-brane and ${\bar {\rm D}}$-brane decays, lower-dimensional D-branes of codimension-two are produced as the representative nonperturbative degrees [@Sen:2004nf]. When the D3${\bar {\rm D}}$3 is considered, the D-strings or DF-strings ($(p,q)$-strings) are particularly intriguing as cosmic string candidates [@Copeland:2003bj; @Dvali:2003zj]. As has been done for the cosmic strings from the Nielsen-Olesen vortices in Abelian-Higgs model, the straight strings saturating the Bogomolnyi bound [@Bogomolny:1975de] enable us to study various dynamical issues analytically [@VS].
The tachyon dynamics for D3${\bar {\rm D}}$3 is described in the several contexts, and the boundary string field theory (BSFT or background-independent string field theory) [@Witten:1992qy] should be an appropriate language with taking into account string off-shell contributions [@Gerasimov:2000zp]. In BSFT of D$p{\bar
{\rm D}}p$ for superstring theory, the effective BSFT action for a complex tachyon field was derived and the descent relation for single codimension-two brane was obtained in an exact form from the energy density difference between the false and true vacua [@Kraus:2000nj].
Since the kinetic term of the BSFT action is very complicated, the static multi-D-string configurations and the related issues have been dealt in the limited references [@Jones:2002si; @Jones:2003ae; @Jones:2003ew]. For BPS static kinks and rolling tachyons in the BSFT of an unstable D-brane, the equations of motion from the BSFT action were analyzed [@Hashimoto:2001rk; @Sugimoto:2002fp; @Kim:2006mg] and even an exact topological BPS kink solution was obtained [@Kim:2006mg]. Additionally, in the Dirac-Born-Infeld (DBI) type effective field theory (EFT) of a complex tachyon field and U(1)$\times$U(1) gauge fields [@Sen:2003tm; @Garousi:2004rd], some studies have been made recently. The single thin BPS vortex satisfying the descent relation was reproduced [@Sen:2003tm], the solutions corresponding thick D- and DF-strings were found in the presence of radial electric field [@Kim:2005tw], and the gravitating solutions including black brane structure were obtained [@Kim:2006xi]. In relation with cosmic strings, the BPS limit for static straight multi-D(F)-strings was established [@Kim:2006xi].
In this Letter, we will consider the D$p{\bar {\rm D}}p$ action in super-BSFT and derive rigorously a BPS limit for static straight multi- D- and DF-strings. To be specific, the BPS limit is achieved in the limit of zero thickness, the pressure components and off-diagonal stress component vanish in the plane orthogonal to string direction, a BPS sum rule based on the descent relation of codimension-two branes is satisfied under a Gaussian-type tachyon potential. The form of first-order Bogomolnyi equation is the same as that in DBI-type EFT, and the multi-BPS-D(F)-string solutions also satisfy the Euler-Lagrange equation. The obtained BPS properties may open new windows to tackle dynamical and cosmological issues with the D(F)-strings [@Jackson:2004zg; @Tye:2005fn; @Firouzjahi:2006vp] in BSFT.
In section \[sec2\], we derive the BPS limit for multi-D(F)-strings in the context of BSFT. In section \[sec3\], we show that the Euler-Lagrange equation for the tachyon field does not support static regular topological vortex solution, which may imply uniqueness of singular BPS solutions as static D-vortex solutions. We conclude in section \[sec4\] with brief discussions on further studies.
BPS Multi- D- and DF-strings {#sec2}
============================
In BSFT for superstrings, off-shell BSFT action $S$ is obtained through an identification with worldsheet partition function $Z$, $S=Z$ [@Marino:2001qc]. For the system of D$p{\bar {\rm D}}p$ in their coincidence limit, the BSFT action of the tachyon field $T$ and its complex conjugate ${\bar T}$, coupled to an Abelian gauge field $A_{\mu}$ with $F_{\mu\nu}=\partial_{\mu}A_{\nu}-\partial_{\nu}A_{\mu}$, is given by [@Kraus:2000nj] $$\label{ac9}
S=-2{\cal T}_p \int d^{p+1}x \,V(T,{\bar T})
\sqrt{-\det (\eta_{\mu\nu}+F_{\mu\nu})}\,{\mathcal F}(y_+){\mathcal F}(y_-),$$ where ${\mathcal T}_{p}$ is tension of the D$p$-brane. The runaway tachyon potential is Gaussian type, $$\label{tpo}
V(T,{\bar T})=e^{-T{\bar T}},$$ and functional form of the derivative term is $$\label{ktr}
{\mathcal F}(y_{\pm})=\frac{y_{\pm}\,4^{y_{\pm}}\,
\Gamma(y_{\pm})^2}{2\Gamma(2y_{\pm})},$$ where the variables, $$\label{ypm}
y_{\pm}=2(G^{\mu\nu}\partial_\mu T \partial_\nu {\bar T})
\pm 2\sqrt{(G^{\mu\nu}\partial_\mu T \partial_\nu T)
(G^{\rho\sigma}\partial_\rho {\bar T} \partial_\sigma {\bar T})
+(\theta^{\mu\nu}\partial_\mu T \partial_\nu {\bar T})^2} \, ,$$ are expressed in terms of open string metric $G^{\mu\nu}$ and noncommutativity parameter $\theta^{\mu\nu}$ as $$\label{gth}
G^{\mu\nu}=\left(\frac{1}{\eta+F}\right)^{(\mu\nu)},\qquad
\theta^{\mu\nu}=\left(\frac{1}{\eta+F}\right)^{[\mu\nu]} \, .$$
Let us consider static multi-D(F)-strings from D3${\bar {\rm D}}$3 ($p=3$), which are stretched parallel to $z$-direction. An appropriate ansatz for the D-string and fundamental string is $$\label{ans}
T=T(x^{i}),\quad -F_{0z}=E_{z}(x^{i}), \quad (i=1,2),$$ where all the other components of the field strength are assumed to be vanishing. Substitution of the electric field $E_{z}$ (\[ans\]) into the Bianchi identity, $\partial_{\mu}F_{\nu\rho}+\partial_{\nu}F_{\rho\mu}
+\partial_{\rho}F_{\mu\nu}=0$, forces $E_{z}$ to be constant. The static tachyon field (\[ans\]) with constant electric field $E_{z}$ leads to tachyon equation, $$\begin{aligned}
\lefteqn{ 2\partial_i \left\{ V \sqrt{1-E_z^2}
\left[ \mathcal{F}'(y_+) {\mathcal F}(y_-) \left(
\eta^{ij} \partial_j T +\frac{
\eta^{ij}\partial_j {\bar T} (\eta^{kl} \partial_k T \partial_l T)}
{\sqrt{|\eta^{ij}\partial_i T \partial_j T|^2}}
\right) \right. \right. } \nonumber\\
&& \left. \left. +{\mathcal F}(y_+) \mathcal{F}'(y_-) \left(
\eta^{ij} \partial_j T -\frac{
\eta^{ij}\partial_j {\bar T} (\eta^{kl} \partial_k T \partial_l T)}
{\sqrt{|\eta^{ij}\partial_i T \partial_j T|^2}} \right) \right] \right\}
= \sqrt{1-E_z^2}\,{\mathcal F}(y_+)
{\mathcal F}(y_-) \frac{\partial V}{\partial {\bar T}},
\label{stq}\end{aligned}$$ where $y_{\pm}$ in (\[ypm\]) reduce to $$\label{spm}
y_{\pm}=2(\eta^{ij}\partial_i T \partial_j {\bar T})
\pm 2\sqrt{|\eta^{ij}\partial_i T \partial_j T|^2} \;.$$ These field configurations automatically satisfy the equation of the gauge field $A_{\mu}$, $\partial_\mu \Pi^{\mu\nu}=0$, where $\Pi^{\mu \nu}\equiv \partial{\mathcal L}/\partial(\partial_\mu A_\nu)$. Since the momentum densities, $T^{0i}$ and $T^{0z}$, and some off-diagonal stress components, $T^{iz}=T^{zi}$, are vanishing under the ansatz (\[ans\]), the conservation of energy-momentum tensor becomes $$\label{cos}
\partial_{j}T^{ji}=0,$$ and it is equivalent to the tachyon equation (\[stq\]) for nontrivial configurations.
To investigate the BPS limit of the D(F)-strings, we examine the pressure components perpendicular to the D(F)-strings $$\begin{aligned}
T^{x}_{\; x}&=&-2{\cal T}_{3} V(T)\sqrt{1-E_z^2}\left\{
{\mathcal F}(y_+){\mathcal F}(y_-)+2[\mathcal{F}'(y_+)\mathcal{F}(y_-)
y_+^{xx}+\mathcal{F}(y_+)\mathcal{F}'(y_-) y_-^{xx}]
\right\} ,
\label{txx}\\
T^{y}_{\; y}&=&-2{\cal T}_{3} V(T)\sqrt{1-E_z^2}\left\{
{\mathcal F}(y_+){\mathcal F}(y_-)+2\left[\mathcal{F}'(y_+)\mathcal{F}(y_-)
y_+^{yy}+\mathcal{F}(y_+)\mathcal{F}'(y_-) y_-^{yy}\right]
\right\},
\label{tyy}\end{aligned}$$ where $y_{\pm}^{ij}$ are defined by $$\label{yij}
y_{\pm}^{ij}=-2\partial_{(i}T \partial_{j)}\bar{T} \mp
\frac{(\partial_i T \partial_j T)(\partial_k \bar{T} \partial_k \bar{T})
+(\partial_i \bar{T} \partial_j \bar{T})(\partial_k T \partial_k T)}
{\sqrt{|\partial_k T \partial_k T|^2}}.$$ As a necessary condition, pressure difference is required to vanish; $$\begin{aligned}
T^{x}_{\; x}-T^{y}_{\; y}&=&8{\cal T}_{3}V(T)\sqrt{1-E_z^2}\,
(|\partial_x T|^2-|\partial_y T|^2) \Bigg\{
\left[\mathcal{F}'(y_+)\mathcal{F}(y_-)+\mathcal{F}(y_+)\mathcal{F}'(y_-)
\right]\nonumber\\
&&\qquad +\frac{\left[\mathcal{F}'(y_+)\mathcal{F}(y_-)
-\mathcal{F}(y_+)\mathcal{F}'(y_-)
\right](|\partial_x T|^2+|\partial_y T|^2)}
{\sqrt{|(\partial_x T)^2+(\partial_y T)^2|^2}}\Bigg\}\nonumber\\
&=&4{\cal T}_{3}V(T)\sqrt{1-E_z^2}\,
\left[(\overline{\partial_xT+i\partial_yT})(\partial_xT-i\partial_yT)
+(\overline{\partial_xT-i\partial_yT})(\partial_xT+i\partial_yT)\right]
\nonumber\\
&&\times\Bigg\{
\left[\mathcal{F}'(y_+)\mathcal{F}(y_-)+\mathcal{F}(y_+)\mathcal{F}'(y_-)
\right]\nonumber\\
&&\hspace{7mm}+\frac{\left[\mathcal{F}'(y_+)\mathcal{F}(y_-)
-\mathcal{F}(y_+)\mathcal{F}'(y_-) \right](|\partial_x
T|^2+|\partial_y T|^2)} {2\sqrt{|(\partial_x T)^2+(\partial_y T)^2|^2}}\Bigg\}
\label{x-y}\\
&=&0.\end{aligned}$$
We read first-order Cauchy-Riemann equation as Bogomolnyi equation from vanishing pressure difference (\[x-y\]) $$\label{Beq}
(\partial_x\pm i\partial_y)T=0,\qquad (\partial_x
\ln\tau=\pm\partial_y\chi~\textrm{and}~
\partial_y \ln\tau=\mp\partial_x\chi),$$ where $T=\tau e^{i\chi}$.[^1] By using (\[Beq\]), we easily check that the remaining off-diagonal stress component becomes automatically zero; $$\begin{aligned}
T^{x}_{\; y}&=&-4{\cal T}_{3} V(T)\sqrt{1-E_z^2}\,[
\mathcal{F}'(y_+)\mathcal{F}(y_-)
y_+^{xy}+\mathcal{F}(y_+)\mathcal{F}'(y_-) y_-^{xy}]
\label{xy0}\\
&=&4{\cal T}_{3} V(T)\sqrt{1-E_z^2}\,
(\partial_xT\partial_y\bar{T}+\partial_yT\partial_x\bar{T})\nonumber\\
&&\hspace{-2mm}\times\left\{[\mathcal{F}'(y_+)\mathcal{F}(y_-)
+\mathcal{F}(y_+)\mathcal{F}'(y_-)]
+\frac{[\mathcal{F}'(y_+)\mathcal{F}(y_-)
-\mathcal{F}(y_+)\mathcal{F}'(y_-)](|\partial_xT|^2+|\partial_yT|^2)}
{\sqrt{|(\partial_x T)^2+(\partial_y T)^2|^2}}
\right\}\nonumber\\
&=&2{\cal T}_{3} V(T)\sqrt{1-E_z^2}\,\left[
(\partial_xT\pm i\partial_yT)(\overline{\partial_yT\pm i\partial_xT})
+(\overline{\partial_yT\mp i\partial_xT})
(\partial_xT\mp i\partial_yT)\right]\nonumber\\
&&\hspace{-2mm}\times\left\{[\mathcal{F}'(y_+)\mathcal{F}(y_-)
+\mathcal{F}(y_+)\mathcal{F}'(y_-)]
+\frac{[\mathcal{F}'(y_+)\mathcal{F}(y_-)
-\mathcal{F}(y_+)\mathcal{F}'(y_-)](|\partial_xT|^2+|\partial_yT|^2)}
{\sqrt{|(\partial_x T)^2+(\partial_y T)^2|^2}}\right\}\nonumber\\
&\stackrel{(\ref{Beq})}{=}&0.
\label{txy0}\end{aligned}$$
For the $n$ straight strings (anti-strings) spread arbitrarily on the $(x,y)$-plane, the ansatz on the tachyon phase $\chi$ is $$\label{Bph}
\chi=\pm\sum_{p=1}^n\theta_p=\pm\sum_{p=1}^n \tan^{-1}\frac{y-y_p}{x-x_p}.$$ Then the tachyon amplitude $\tau$ is obtained as an exact solution of the Bogomolnyi equation (\[Beq\]), $$\label{Bam}
\tau=\prod_{p=1}^n \tau_{\rm BPS} |{\bf x}-{\bf x}_p|.$$ Inserting the BPS solutions (\[Bph\])–(\[Bam\]) into the formula (\[spm\]), we obtain $$y\equiv y_{\pm}=4\partial_{x}{\bar T}\partial_{x}T=
4 \tau_{{\rm
BPS}}^{2}\prod_{p=1}^{n}(\tau_{{\rm BPS}}|{\bf x}-{\bf x}_p|)^2
\sum_{q,r=1}^n\frac{\cos \theta_{qr}} {(\tau_{{\rm BPS}}|{\bf
x}-{\bf x}_q|)\; (\tau_{{\rm BPS}}|{\bf x}-{\bf x}_r|)},\label{y}$$ where $\theta_{qr}$ is the angle between two vectors, $({\bf x}-{\bf
x}_q)$ and $({\bf x}-{\bf x}_r)$. Substituting (\[Bph\])–(\[y\]) into the pressure components (\[txx\])–(\[tyy\]), we have $-T^{x}_{\; x}=-T^{y}_{\;
y}=2{\cal T}_{3}V\sqrt{1-E_{z}^{2}} \,{\mathcal F}({\mathcal
F}-2y{\mathcal F}')$. Therefore, the pressure components (\[txx\])–(\[tyy\]) vanish only in the limit of zero thickness of each vortex, $\tau_{{\rm BPS}}\rightarrow \infty$, due to the rapidly-decaying tachyon potential $V(\tau)$ (\[tpo\]) except for the site of each vortex ${\bf x}={\bf x}_{p}$, i.e., $\displaystyle{\lim_{\tau_{{\rm BPS}\rightarrow\infty}}( -T^{x}_{\;
x})|_{{\bf x}={\bf x}_{p}} =\lim_{\tau_{{\rm
BPS}\rightarrow\infty}}( -T^{y}_{\; y})|_{{\bf x}={\bf x}_{p}}
=\frac{\pi}{2}{\cal T}_{3}\sqrt{1-E_{z}^{2}}}\,$. This nonvanishing pressure at each D(F)-string location is different from the character of BPS vortices in Abelian gauge theories with Higgs mechanism where the pressure components vanish everywhere including vortex points [@Bogomolny:1975de; @Hong:1990yh]. The stress component $T^{x}_{\; y}$ also vanishes for the BPS configuration as shown in (\[xy0\])–(\[txy0\]), and then the conservation of energy-momentum tensor (\[cos\]) reduces to $\partial_{x}T^{x}_{\;
x}=0$ and $\partial_{y}T^{y}_{\; y}=0$. For the aforementioned pressure components of the BPS D(F)-strings in the infinite $\tau_{{\rm BPS}}$ limit, the equations hold when the derivatives are considered as weak derivatives [@Eva]. As $\tau_{{\rm
BPS}}\rightarrow \infty$, the static singular solution (\[Bph\])–(\[Bam\]) of BPS equation satisfies the conservation of energy-momentum tensor (\[cos\]), which is equivalent to the tachyon equation (\[stq\]) for nontrivial tachyon configurations. In the section \[sec3\], we also show that the tachyon equation (\[stq\]) does not support regular static straight D(F)-string solution.
For the static configurations of $\dot{T}=\dot{\bar{T}}=0$ with constant $E_{z}$, the conjugate momenta of the tachyon field and its complex conjugate vanish, $\Pi_{T}\equiv \partial {\mathcal L}/\partial {\dot T}=0$ and $\Pi_{{\bar T}}\equiv \partial {\mathcal L}/\partial {\dot {\bar T}}=0$, and the conjugate momentum of the gauge field $\Pi_{z}$ is $$\label{piz}
\Pi_{z}=-\frac{E_z}{\sqrt{1-E_z^2}} \;2{\cal T}_{3}V
{\mathcal F}(y_+){\mathcal F}(y_-).$$ The Hamiltonian density obtained by a Legendre transform leads to the BPS formula for DF-strings (($p,q$)-strings) [@Witten:1995im], $$\label{bpsH}
{\mathcal H}=\sqrt{\Pi_{z}^{2}+\left[2{\cal T}_{3}V
{\mathcal F}(y_{+}){\mathcal F}(y_{-})\right]^{2}}\, ,$$ where the limit of D-strings, ${\mathcal H}|_{\Pi_{z}=0}=2{\cal T}_{3}
V{\mathcal F}(y_{+}){\mathcal F}(y_{-})$, is trivially involved in the absence of fundamental string charge density $\Pi_{z} = 0$. Plugging the conjugate momentum (\[piz\]), the Hamiltonian density (\[bpsH\]) coincides exactly with the energy density $-T^{t}_{\; t}$, and, due to the boost symmetry along the $z$-direction, the multi-D(F)-string configuration satisfies $T^{t}_{\; t}=T^{z}_{\; z}$; $$\begin{aligned}
\label{edt}
-T^{t}_{\; t}=-T^{z}_{\; z}=\frac{2{\cal T}_{3} V(T)}{\sqrt{1-E_z^2}}
{\cal F}(y_+){\cal F}(y_-).\end{aligned}$$ Noticing easily that the energy density is proportional to the electric flux density as $${\cal H}=-\frac{\Pi_{z}}{E_{z}},$$ we read for the DF-strings that the charge distribution of fundamental string part is exactly proportional to the energy density of D-string part, which is confined at each string site in $(x,y)$-plane.
If we require a BPS sum rule to the energy per unit D(F)-string length for the BPS configuration with $y_{+}=y_{-}$, $$\begin{aligned}
\label{suru}
{\cal T}_{1}|n|=\frac{H}{\int_{-\infty}^{\infty} dz}
=\frac{2{\cal T}_{3}}{\sqrt{1-E_z^2}}\int d^{2}x\lim_{\tau_{{\rm BPS}}
\rightarrow \infty}V{\mathcal F}^{2}
=(2\pi\sqrt{\alpha'}\,)^{2}\frac{{\cal T}_{3}}{\sqrt{1-E_z^2}}|n|,
\quad (\alpha'=2),\end{aligned}$$ the descent relation of a D(F)-string, $$\begin{aligned}
\label{dsc1}
{\cal T}_{1}=(2\pi\sqrt{\alpha'}\,)^{2}{\cal
T}_{3}/\sqrt{1-E_z^2}\,,\end{aligned}$$ is correctly reproduced, and a constraint condition for a BPS sum rule is achieved for the tachyon potential, $$\begin{aligned}
\label{Vco}
4\pi^{2}|n|=\int d^{2}x\lim_{\tau_{{\rm BPS}}\rightarrow \infty}V(\tau)
{\mathcal F}^{2}.\end{aligned}$$ Since the integrand, $\displaystyle{\lim_{\tau_{{\rm
BPS}}\rightarrow \infty} V(\tau){\mathcal F}^{2}}$, has infinity at each string site ${\bf x}={\bf x}_{p}$ and vanishes at ${\bf
x}\ne{\bf x}_{p}$ in the BPS limit of infinite $\tau_{{\rm BPS}}$, the condition (\[Vco\]) is reexpressed by a local form, $$\label{sde}
\lim_{\tau_{{\rm BPS}}\rightarrow\infty}
V(\tau){\mathcal F}^{2}=4\pi^{2}\sum_{p=1}^{n}\delta^{(2)}
({\bf x}-{\bf x}_{p}).$$ In summary, the energy-momentum tensor of $n$ D(F)-strings is in the BPS limit, $$\begin{aligned}
-T^{\mu}_{\;\nu}={\cal T}_{1}\sum_{p=1}^{n}\,{\rm
diag}\,\left(\delta^{(2)} ({\bf x}-{\bf x}_{p}),
\frac{1-{E_{z}}^2}{16\pi}{\mathbb I} ({\bf x}-{\bf x}_{p}),
\frac{1-{E_{z}}^2}{16\pi}{\mathbb I} ({\bf x}-{\bf x}_{p}),
\delta^{(2)}({\bf x}-{\bf x}_{p})\right),\end{aligned}$$ where ${\mathbb I} ({\bf x}-{\bf x}_{p})$ has unity at ${\bf x}={\bf x}_{p}$ and zero at ${\bf x}\ne {\bf x}_{p}$. Note that the pressure components orthogonal to the string direction vanish in the limit of critical electric field, $|E_{z}|\rightarrow 1$.
From now on, let us perform the integration (\[Vco\]) with the tachyon potential (\[tpo\]) and show that it reproduces the required value for saturating the BPS sum rule. First, we consider a single D(F)-string of $n=1$ at an arbitrary position. In this case, $y$ is independent of $x$, $y=4\tau_{\rm BPS}^2$, so is ${\mathcal
F}$. Then, a rescaling ${\tilde {\bf x}}=\tau_{{\rm BPS}}{\bf x}$ with a translation in (\[Vco\]) provides a definite integral without explicit dependence of $\tau_{{\rm BPS}}$; $$\begin{aligned}
\label{1i1}
\int d^2x \lim_{\tau_{{\rm BPS}}\rightarrow \infty}
e^{-(\tau_{\rm BPS}|{\bf x}|)^2} \mathcal{F}(4\tau_{\rm BPS}^2)^2
&=&\lim_{\tau_{{\rm BPS}}\rightarrow \infty}
\frac{\mathcal{F}(4\tau_{\rm BPS}^2)^2}{\tau_{\rm BPS}^2}
\int d^2{\tilde x}\, e^{-|{\tilde {\bf x}}|^2}.
\label{1i2}\end{aligned}$$ If we perform the Gaussian integral for arbitrary $\tau_{{\rm BPS}}$ and take the limit of infinite $\tau_{{\rm BPS}}$ by using the asymptotic form of $\mathcal{F}(y)^{2}$, $\mathcal{F}(y)^2=\pi
y+\pi/8+{\cal O}(y^{-1})$, in (\[1i1\]), then value of the integral is $4\pi^{2}$, which satisfies the descent relation. Second, we consider the superimposed D(F)-strings of arbitrary $|n|$. Now $y$ of ${\mathcal F}(y)$ has $x$-dependence as $y=4n^2
\tau_{\rm BPS}^2(\tau_{\rm BPS}|{\bf x}|)^{2n-2}$, and then we use the same rescaling of ${\bf x}$ as $$\begin{aligned}
\label{aaa}
\lefteqn{ \int d^2x\lim_{\tau_{{\rm BPS}}\rightarrow \infty}
e^{-(\tau_{\rm BPS}|{\bf x}|)^{2n}}
\mathcal{F}(4n^2 \tau_{\rm BPS}^2(\tau_{\rm BPS}|{\bf x}|)^{2n-2})^2}\\
&=&4\pi n^2 \lim_{\tau_{{\rm BPS}}\rightarrow \infty}
\int d^2{\tilde x}\,e^{-|{\tilde {\bf x}}|^{2n}}|{\tilde {\bf x}}|^{2n-2}
\frac{\mathcal{F}(4n^2 \tau_{\rm BPS}^2|{\tilde {\bf x}}|^{2n-2})^2}{
4n^{2}\pi\tau_{\rm BPS}^{2}|{\tilde {\bf x}}|^{2n-2}}.
\label{ni1}\end{aligned}$$ As $\tau_{{\rm BPS}}$ increases, the integrand with explicit $\tau_{{\rm BPS}}$ dependence becomes $$\begin{aligned}
\lefteqn{
\lim_{\tau_{{\rm BPS}}\rightarrow \infty}
\frac{\mathcal{F}(4n^2 \tau_{\rm BPS}^2|{\tilde {\bf x}}|^{2n-2})^2}{
4n^{2}\pi\tau_{\rm BPS}^{2}|{\tilde {\bf x}}|^{2n-2}}}
\nonumber\\
&=&\lim_{\tau_{{\rm BPS}}\rightarrow \infty}\left[
\frac{(\sqrt{\pi\times 4n^2 \tau_{\rm BPS}^2|{\tilde {\bf x}}|^{2n-2}}
\,)^2}{4n^{2}\pi\tau_{\rm BPS}^2|{\tilde {\bf x}}|^{2n-2}}
+{\mathcal O}(1/\tau_{{\rm BPS}}^{2},1/|{\tilde {\bf x}}|^{2n-2})\right]
\label{Fap}\\
&=&1
\label{ni9}\end{aligned}$$ with keeping $|{\tilde {\bf x}}|$ finite. For infinite $|{\tilde
{\bf x}}|$, the integrand vanishes due to the exponential term. Since ${\mathcal F}(y)^{2}$ is analytic for every non-negative $y$, the integrand is finite at ${\tilde {\bf x}}=0$, and the asymptotic form of $\mathcal{F}(y)^{2}$ guarantees finiteness of the integral (\[ni1\]) for finite $\tau_{{\rm BPS}}$, we can take infinite $\tau_{{\rm BPS}}$ limit to ${\mathcal F}(y)^{2}/\pi y$ part in (\[ni1\]). Therefore, value of the integral (\[ni1\]) is $4\pi^{2}n$ which fits (\[Vco\]). Third, we consider the case of $n$ separated D(F)-strings where the distance between any pair of D(F)-strings is much larger than $1/\tau_{{\rm BPS}}$. When ${\bf
x}\ne{\bf x}_s$ $(s=1,2,\ldots,n)$, it is obvious that $y$ in (\[y\]) diverges in the $\tau_{\rm BPS}\to\infty$ limit for any tachyon field. When ${\bf x}={\bf x}_s$, the term with $p=q=r$ in (\[y\]) survives and hence $y\to\infty$ in this BPS limit. Thus we see that $y$ always becomes infinite in the $\tau_{\rm
BPS}\to\infty$ limit. Accordingly, ${{\cal{F}}(y)}^2$ in the integral (\[Vco\]) diverges everywhere. Let us examine the tachyon potential part in (\[Vco\]). When ${\bf x}={\bf x}_p$ $(p=1,2,\ldots,n)$, $\tau$ in (\[Bam\]) vanishes and then the tachyon potential has unity, $V(\tau=0)=1$. When ${\bf x}\ne{\bf
x}_s$, it vanishes in the infinite $\tau_{\rm BPS}$ limit and the integrand in (\[Vco\]) also vanishes due to the exponential damping of the tachyon potential despite of the leading divergent term of $\mathcal{F}$, ${\mathcal F}(y)\rightarrow \sqrt{\pi y}\, $. Therefore, among $n^2$-terms in $y$ (\[y\]) specified by the $(q,
r)$-indices, the $n$-terms with $q=r$ contribute to the integral (\[Vco\]). In addition, functional shape of the integrand diverges at each string site but vanishes away from the location of each D(F)-string. In what follows, we will show that the contribution of each term at ${\bf x}={\bf x}_p$ to the integration is exactly the same as that of delta function given in single D(F)-string (\[1i2\]) as far as the distance $|{\bf x}_p-{\bf x}_q|$ for any $p$ and $q$ ($p\ne q$) is sufficiently larger than $1/\tau_{{\rm
BPS}}$. Since only the neighborhoods of D(F)-string sites, ${\bf
x}={\bf x}_p$, contribute to (\[Vco\]) in performing the $(x,y)$-integration and become sufficiently small for infinite $\tau_{\rm BPS}$, only the leading terms of $V$ and ${\cal{F}}^2$ can contribute nonvanishing value to the integral (\[Vco\]). To be specific, we can replace the integrand $V{\cal{F}}^2$ and then perform the integration as follows, $$\begin{aligned}
\lim_{\tau_{\rm BPS}\to\infty}\int d^2x V\mathcal{F}^2&=& 4\pi\int
d^2x \lim_{\tau_{\rm BPS}\to\infty} \sum_{s=1}^n
\exp\left[-\left(\prod\limits_{\scriptstyle p=1\atop\scriptstyle
(p\ne s)}^n \tau_{\rm BPS} |{\bf x}_s-{\bf x}_p|\right)^2
\left(\tau_{\rm BPS} |{\bf x}-{\bf x}_s|\right)^2
\right]\nonumber\\
&&\times\tau_{\rm BPS}^2 \left(\prod\limits_{\scriptstyle
q=1\atop\scriptstyle (q\ne s)}^n \tau_{\rm BPS} |{\bf x}_s-{\bf
x}_q|
\right)^2\nonumber\\
&=&4\pi^2n,\end{aligned}$$ which is exactly the value in (\[Vco\]). Fourth, we consider the case of arbitrary BPS configuration where $n_{p}$ D(F)-strings among the $n$ D(F)-strings are superimposed at an ${\bf x}_{p}$ with $\displaystyle{n=\sum_{p}n_{p}}$. If we replace the integration (\[1i2\]) by (\[aaa\])–(\[ni9\]), the integration reproduces the value in (\[Vco\]) by applying repeatedly the above third argument. In synthesis, the aforementioned four arguments lead to a conclusion that the Gaussian type tachyon potential (\[tpo\]) fulfills the integration (\[Vco\]) in the thin BPS limit.
Nonexistence of Nonsingular D- and DF-string Solutions {#sec3}
======================================================
In this section, we deal with the tachyon equation (\[stq\]) and discuss nonexistence of the monotonically-increasing nonsingular D-vortex solution connecting the boundary conditions at the origin, $\tau(|{\bf x}|=0)=0$, and infinity, $\tau(|{\bf x}|=\infty)=\infty$. This perhaps supports uniqueness of the singular BPS multi-D(F)-string solutions obtained in the previous section.
Suppose that we have $n$ superimposed straight D(F)-strings stretched along the $z$-axis. Since the electric field $E_{z}$ is actually canceled in both sides of the tachyon equation (\[stq\]), we have $$\begin{aligned}
\label{Teq2}
\frac{1}{r} \frac{d}{dr}\left[r e^{-\tau^2}\tau'
{\mathcal F}'(y_+){\mathcal F}(y_-) \right] = \tau e^{-\tau^2}
{\mathcal F}(y_+)\left[ \frac{n^2}{r^2}{\mathcal F}'(y_-)
-\frac{1}{4}{\mathcal F}(y_-)\right] \,,\end{aligned}$$ where $y_\pm$ in (\[ypm\]) become $$\begin{aligned}
y_+=4 \tau'^2\,,\qquad y_-=\frac{4n^2}{r^2}\tau^2 \,.\end{aligned}$$ The D(F)-string solutions of our interest are given by monotonically-increasing tachyon configurations connecting the boundary conditions, $\tau(r=0)=0$ and $\tau(r=\infty)=\infty$.
Expansion of the tachyon amplitude $\tau$ near the origin is $$\begin{aligned}
\label{rze}
\tau (r) \approx \tau_0 r^n (1-\tau_{1}r^{2}+...\, ),\end{aligned}$$ where $\tau_0$ is an undetermined constant determined by the behavior at asymptotic region. Since the coefficient of subleading term $\tau_{1}$ is always positive irrespective $n$, $$\tau_{1}=\left\{
\begin{array}{lc}
\displaystyle{\frac{{\mathcal F}(4\tau_{0}^{2})
-8\tau_{0}^{2}{\mathcal F}'(4\tau_{0}^{2})}{
32[{\mathcal F}'(4\tau_{0}^{2})+8\tau_{0}^{2}{\mathcal F}''(4\tau_{0}^{2})]}}
\, ,
& (n=1) \\
\displaystyle{ \frac{1}{24\ln 2}\left\{64\tau_{0}^{2}\left[8(\ln
2)^{2}-\frac{\pi^{2}}{3} \right]+\frac{1}{4}\right\}}\, ,
&(n=2)\\
\displaystyle{\frac{1}{32(\ln 2)(n+1)}}\, ,
&(n\ge 3)
\end{array}
\right. ,$$ increasing tendency of the tachyon field $\tau(r)$ decreases as $r$ increases. If we try expansion at asymptotic region by using a power law, $\tau\sim \tau_\infty r^k,~(k>0)$, or a logarithmic increase, $\tau\sim \tau_\infty \ln r$, many possibilities are ruled out by the tachyon equation (\[Teq2\]) and survived cases are $$\label{rinf}
\tau(r)\approx \tau_{\infty 0}r^{1+k}+\tau_{\infty 1}r^{1+k-l}+...,
\qquad (k>0,~0<l<2k),$$ where both $\tau_{\infty 0}$ and $\tau_{\infty 1}$ are not determined by the tachyon equation (\[Teq2\]). The leading term is rapidly increasing since $\displaystyle{\lim_{r\rightarrow
\infty}\tau'\rightarrow \infty}$. Comparison of the power series solutions near the origin (\[rze\]) and at the asymptotic region (\[rinf\]) suggests that smooth connection of both increasing tachyon profiles seems unlikely.
Another possibility is the solution with maximum value, i.e., the tachyon amplitude increases near the origin, reaches a maximum value $\tau_{{\rm m}}$ at a finite coordinate $r=r_{{\rm m}}$, and then starts to decrease with $d^{2}\tau/dr^{2}|_{r=r_{{\rm m}}}<0$. Expansion near $r=r_{{\rm m}}$ gives $$\label{rme}
\tau(r)\approx \tau_{{\rm m}}\left[1-\frac{1}{2}\tau_{{\rm
m}2}(r-r_{{\rm m}})^{2} +...\, \right],$$ where the coefficient $\tau_{{\rm m}2}$ is $$\begin{aligned}
\tau_{{\rm m}2}=\frac{{\mathcal F}(y_{-}^{{\rm m}})-\frac{4n^{2}}{
r_{{\rm m}}^{2}} {\mathcal F}'(y_{-}^{{\rm m}})}{8(\ln 2)
{\mathcal F}(y_{-}^{{\rm m}})},\qquad y_{-}^{{\rm
m}}=\frac{4n^{2}\tau_{{\rm m}}^{2}}{r_{{\rm m}}^{2}}.\end{aligned}$$ In order to have the maximum $\tau_{\rm m}$=$\tau(r_{\rm m})$, $\tau_{\rm m}$ and $r_{\rm m}$ should satisfy the following inequality, $$\label{maxcon}
\tau_{\rm m}^2 > {\frac{d}{d({\rm ln}{y_-^{\rm m}})}\rm
ln{\mathcal F}}(y_-^{\rm m}).$$ Numerical works support that every regular solution with finite $\tau_{0}$ has the maximum value $\tau_{{\rm m}}$ at a finite $r_{{\rm m}}$ irrespective of $n$ as shown in Fig. \[fig1\].
\[1.3\][![$\tau_{0}$ vs. $\tau_{{\rm m}}$. $n=1$ for dotted line, $n=2$ for dashed line, and $n=3$ for solid line.[]{data-label="fig1"}](vorBfig01.eps "fig:")]{}
-2.0cm
Probably, there does not exist any static nonsingular monotonically-increasing D(F)-string solution of the tachyon equation (\[Teq2\]) with $\tau(0)=0$ and $\tau(\infty)=\infty$. Since the aforementioned discussion does not rule out the singular solution $\tau$ with infinite slope $d\tau/dr\sim\infty$, the BPS solutions (\[Bph\])–(\[Bam\]) are free from the argument of nonexistence. This conclusion, the nonexistence of regular static topological non-BPS D(F)-string solutions, is consistent with the same result of nonexistence in DBI EFT [@Kim:2005tw].
Conclusion {#sec4}
==========
The system of D3${\bar {\rm D}}$3 has been considered in the scheme of super-BSFT EFT (\[ac9\]) including a complex tachyon and U(1)$\times$U(1) gauge fields. From the vanishing pressure difference, the first-order Bogomolnyi equation (\[Beq\]) was derived and straight topological BPS multi-D(F)-string configurations were given as exact static solutions (\[Bph\])–(\[Bam\]) which also satisfy the conservation of energy-momentum tensor (\[cos\]). Since the forms of derived Bogomolnyi equation and singular BPS solutions coincide exactly with those in DBI type EFT, this BPS structure seems universal and is consistent with type II superstring theories. The expression of energy was rewritten by the BPS sum rule for the BPS multi-D(F)-string solutions (\[suru\]), and reproduced the descent relation for codimension-two objects (\[dsc1\]), which allowed to interpret the obtained vortex-strings as BPS D1-branes in IIB string theory. This results in a constraint condition for the BPS tachyon potential (\[Vco\]), and the Gaussian type potential of BSFT (\[tpo\]) fulfills the condition. Since it is nothing but making a sum of delta functions in the thin BPS limit (\[sde\]), the uniqueness of BPS tachyon potential seems unlikely. When the $z$-component of constant electric field (\[ans\]) is turned on, the conjugate momentum of the gauge field, the charge density of fundamental strings (\[piz\]), is confined along the D-strings. In addition, the corresponding Hamiltonian density takes a BPS formula (\[bpsH\]), $\sqrt{p^{2}+q^{2}}$ form for the D1-charge density $q$ and the fundamental string charge density $p$, so that the configuration with constant electric field along the string direction is the DF-string (or $(p,q)$-string) from D3$\bar{\mathrm{D}}3$. Though we checked the conditions for BPS vortex configurations explicitly, the form of obtained BPS limit is different from the usual BPS bound for vortices, of which energy minimum is saturated only when the Bogomolnyi equations are satisfied. In this sense, the BPS bound for codimension-two branes from D${\bar {\rm D}}$ system needs further study. We also checked the possibility that the tachyon equation (\[Teq2\]) could possess a nonsingular D(F)-string solutions and the analysis supported negative answer.
Since we achieved a BPS limit of multi-vortex-strings, it may open systematic study of classical dynamics of BPS multi-D(F)-strings, particularly moduli space dynamics in the context of BSFT. Studies of the D(F)-strings in curved spacetime naturally have cosmological implication as candidates of cosmic superstrings.
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank Dongho Chae and Taekyung Kim for helpful discussion. This work is the result of research activities (Astrophysical Research Center for the Structure and Evolution of the Cosmos (ARCSEC)) (A.I.) and was supported by the Korea Research Foundation Grant funded by the Korean Government (MOEHRD, Basic Research Promotion Fund) (KRF-2006-311-C00022) (Y.K.).
[100]{}
For a review, see A. Sen, “Tachyon dynamics in open string theory,” Int. J. Mod. Phys. A [**20**]{}, 5513 (2005) \[arXiv:hep-th/0410103\], and references therein. E. J. Copeland, R. C. Myers and J. Polchinski, “Cosmic F- and D-strings,” JHEP [**0406**]{}, 013 (2004) \[arXiv:hep-th/0312067\]. G. Dvali and A. Vilenkin, “Formation and evolution of cosmic D-strings,” JCAP [**0403**]{}, 010 (2004) \[arXiv:hep-th/0312007\]. E. B. Bogomolny, “Stability of classical solutions,” Sov. J. Nucl. Phys. [**24**]{}, 449 (1976) \[Yad. Fiz. [**24**]{}, 861 (1976)\]. For a review, see A. Vilenkin and E. P. S. Shellard, [*Cosmic strings and other topological defects*]{}, (Cambridge University Press, 1984) or T. W. B. Kibble, “Cosmic strings reborn?,” arXiv:astro-ph/0410073. E. Witten, “On background independent open string field theory,” Phys. Rev. D [**46**]{}, 5467 (1992) \[arXiv:hep-th/9208027\]; “Some computations in background independent off-shell string theory,” Phys. Rev. D [**47**]{}, 3405 (1993) \[arXiv:hep-th/9210065\]; K. Li and E. Witten, “Role of short distance behavior in off-shell open string field theory,” Phys. Rev. D [**48**]{}, 853 (1993) \[arXiv:hep-th/9303067\]; S. L. Shatashvili, “Comment on the background independent open string theory,” Phys. Lett. B [**311**]{}, 83 (1993) \[arXiv:hep-th/9303143\], “On the problems with background independence in string theory,” Alg. Anal. [**6**]{}, 215 (1994) \[arXiv:hep-th/9311177\]. A. A. Gerasimov and S. L. Shatashvili, “On exact tachyon potential in open string field theory,” JHEP [**0010**]{}, 034 (2000) \[arXiv:hep-th/0009103\]; D. Kutasov, M. Marino and G. W. Moore, “Some exact results on tachyon condensation in string field theory,” JHEP [**0010**]{}, 045 (2000) \[arXiv:hep-th/0009148\]; “Remarks on tachyon condensation in superstring field theory,” arXiv:hep-th/0010108. P. Kraus and F. Larsen, “Boundary string field theory of the DD-bar system,” Phys. Rev. D [**63**]{}, 106004 (2001) \[arXiv:hep-th/0012198\]; T. Takayanagi, S. Terashima and T. Uesugi, “Brane-antibrane action from boundary string field theory,” JHEP [**0103**]{}, 019 (2001) \[arXiv:hep-th/0012210\]. N. T. Jones and S. H. H. Tye, “An improved brane anti-brane action from boundary superstring field theory and multi-vortex solutions,” JHEP [**0301**]{}, 012 (2003) \[arXiv:hep-th/0211180\]. N. T. Jones, L. Leblond and S. H. H. Tye, “Adding a brane to the brane anti-brane action in BSFT,” JHEP [**0310**]{}, 002 (2003) \[arXiv:hep-th/0307086\]. N. T. Jones and S. H. H. Tye, “Spectral flow and boundary string field theory for angled D-branes,” JHEP [**0308**]{}, 037 (2003) \[arXiv:hep-th/0307092\]. K. Hashimoto and S. Hirano, “Metamorphosis of tachyon profile in unstable D9-branes,” Phys. Rev. D [**65**]{}, 026006 (2002) \[arXiv:hep-th/0102174\]. S. Sugimoto and S. Terashima, “Tachyon matter in boundary string field theory,” JHEP [**0207**]{}, 025 (2002) \[arXiv:hep-th/0205085\]; J. A. Minahan, “Rolling the tachyon in super BSFT,” JHEP [**0207**]{}, 030 (2002) \[arXiv:hep-th/0205098\]; A. Ishida, Y. Kim and S. Kouwn, “Homogeneous rolling tachyons in boundary string field theory,” Phys. Lett. B [**638**]{}, 265 (2006) \[arXiv:hep-th/0601208\]. C. Kim, Y. Kim, O. K. Kwon and H. U. Yee, “Tachyon kinks in boundary string field theory,” JHEP [**0603**]{}, 086 (2006) \[arXiv:hep-th/0601206\]. A. Sen, “Dirac-Born-Infeld action on the tachyon kink and vortex,” Phys. Rev. D [**68**]{}, 066008 (2003) \[arXiv:hep-th/0303057\]. M. R. Garousi, “D-brane anti-D-brane effective action and brane interaction in open string channel,” JHEP [**0501**]{}, 029 (2005) \[arXiv:hep-th/0411222\]. Y. Kim, B. Kyae and J. Lee, “Global and local D-vortices,” JHEP [**0510**]{}, 002 (2005) \[arXiv:hep-th/0508027\]; I. Cho, Y. Kim and B. Kyae, “DF-strings from D3 D3-bar as cosmic strings,” JHEP [**0604**]{}, 012 (2006) \[arXiv:hep-th/0510218\]. T. Kim, Y. Kim, B. Kyae and J. Lee, “Cosmic D- and DF-strings from D3D-bar3: Black strings and BPS bound,” arXiv:hep-th/0612285. M. G. Jackson, N. T. Jones and J. Polchinski, “Collisions of cosmic F- and D-strings,” JHEP [**0510**]{}, 013 (2005) \[arXiv:hep-th/0405229\]; A. Hanany and K. Hashimoto, “Reconnection of colliding cosmic strings,” JHEP [**0506**]{}, 021 (2005) \[arXiv:hep-th/0501031\]; E. J. Copeland, T. W. B. Kibble and D. A. Steer, “Collisions of strings with Y junctions,” Phys. Rev. Lett. [**97**]{}, 021602 (2006) \[arXiv:hep-th/0601153\]; “Constraints on string networks with junctions,” Phys. Rev. D [**75**]{}, 065024 (2007) \[arXiv:hep-th/0611243\]. S. H. Tye, I. Wasserman and M. Wyman, “Scaling of multi-tension cosmic superstring networks,” Phys. Rev. D [**71**]{}, 103508 (2005) \[Erratum-ibid. D [**71**]{}, 129906 (2005)\] \[arXiv:astro-ph/0503506\]; E. J. Copeland and P. M. Saffin, “On the evolution of cosmic-superstring networks,” JHEP [**0511**]{}, 023 (2005) \[arXiv:hep-th/0505110\]; M. Hindmarsh and P. M. Saffin, “Scaling in a SU(2)/Z(3) model of cosmic superstring networks,” JHEP [**0608**]{}, 066 (2006) \[arXiv:hep-th/0605014\]. H. Firouzjahi, L. Leblond and S. H. Henry Tye, “The (p,q) string tension in a warped deformed conifold,” JHEP [**0605**]{}, 047 (2006) \[arXiv:hep-th/0603161\]; S. Thomas and J. Ward, “Non-Abelian (p,q) strings in the warped deformed conifold,” JHEP [**0612**]{}, 057 (2006) \[arXiv:hep-th/0605099\]. M. Marino, “On the BV formulation of boundary superstring field theory,” JHEP [**0106**]{}, 059 (2001) \[arXiv:hep-th/0103089\]; V. Niarchos and N. Prezas, “Boundary superstring field theory,” Nucl. Phys. B [**619**]{}, 51 (2001) \[arXiv:hep-th/0103102\]. J. Hong, Y. Kim and P. Y. Pac, “On the multivortex solutions of the Abelian Chern-Simons-Higgs theory,” Phys. Rev. Lett. [**64**]{}, 2230 (1990); R. Jackiw and E. J. Weinberg, “Selfdual Chern-Simons vortices,” Phys. Rev. Lett. [**64**]{}, 2234 (1990). L. C. Evans, Partial differential equations, (American Mathematical Society, Providence, 1998).
E. Witten, “Bound states of strings and p-branes,” Nucl. Phys. B [**460**]{}, 335 (1996) \[arXiv:hep-th/9510135\].
[^1]: We call the first-order Cauchy-Riemann equation the Bogomolnyi equation since every BPS D(F)-string configuration is a solution of this equation and the gauged version of this equation was one of the Bogomolnyi equations in $(2+1)$-dimensional Abelian-Higgs model and its analogues.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The present paper is dedicated to illustrating an extension of polar duality between Fano toric varieties to a more general duality, called *framed* duality, so giving rise to a powerful method of producing mirror partners of hypersurfaces and complete intersections in toric varieties, of any Kodaira dimension. In particular, the class of projective hypersurfaces and their mirror partners are studied in detail. Moreover, many connections with known Landau-Ginzburg mirror models, Homological Mirror Symmetry and Intrinsic Mirror Symmetry, are discussed.'
address: 'Dipartimento di Matematica, Università di Torino, via Carlo Alberto 10, 10123 Torino'
author:
- Michele Rossi
bibliography:
- 'MILEA.bib'
title: An extension of polar duality of toric varieties and its consequences in Mirror Symmetry
---
Introduction {#introduction .unnumbered}
============
Polar duality between reflexive polytopes gives the well known Batyrev duality between Fano toric varieties, inducing a mirror symmetry between generic sections of their anti-canonical divisors [@Batyrev94]. Borisov and Batyrev extended this duality to complete intersections described by a nef partition of the anti-canonical divisor of a Fano toric variety [@Borisov], [@BB96]. By thinking of Batyrev duality as a duality between toric varieties *framed* by their anti-canonical divisor, the present paper is devoted to show how *deforming* the Batyrev-Borisov duality by allowing a more general framing, in principle just given by an effective torus invariant Weil divisor (see Definitions \[def:ftv\] and \[def:wftv\] of framed (ftv) and weak framed (wftv) toric varieties, respectively). In general, such a deformed correspondence, here called *framed* duality ($f$-duality), between framed and weak framed toric varieties, is not involutive, but imposing some further conditions on the framing gives back an involutive duality, here called a *calibrated $f$-process*, incorporating the classical duality between Fano toric varieties as a very particular case.
Consequently, $f$-duality is able to describing an enormous number of $f$-mirror symmetric pairs of not necessarily Calabi-Yau hypersurfaces and complete intersections in toric varieties, sensibly improving the current knowledge of mirror partners of non [Calabi-Yau ]{}varieties (see e.g. [@Givental96], [@Hori-Vafa], [@Seidel], [@Efimov], [@KKOY], [@Krawitz], [@HSSW], [@GKR])
For instance, as a very particular but interesting case, a generic projective hypersurface in $\P^n$ of degree $d\geq n+1$, can be thought of a framing of $\P^n$, so admitting (at least) one $f$-mirror dual partner given by an hypersurface in a suitable finite quotient of a weighted projective $n$-space, whose weights are essentially assigned by the framing itself (see §\[sez:ipersuperfici\]). This construction turns out to nicely extending, the pivotal Greene-Plesser description of a mirror partner of the quintic threefold [@GP], to higher degrees. Moreover, for lower degrees, a generic projective hypersurface in $\P^n$ of degree $d\leq n$ can be thought of a weak framing of $\P^n$, whose associated $f$-mirror dual partner can no more be a complete variety (see §\[ssez:NegKod\]), sharing strict relations with Landau-Ginzburg (LG) mirror models proposed by Givental [@Givental96], [@Givental-ICM].
Furthermore, $f$-duality turns out to extending many known dualities between Calabi-Yau hypersurfaces and complete intersections in toric varieties: this is the case of the Berglund-Hübsch duality [@Berglund-Hubsch] and, more generally, of the recent Artebani-Comparin-Guilbot duality [@ACG] (see §\[ssez:BHK\] and \[ssez:ACG\])).
More in general, $f$-duality opens up to a lot of stimulating connections with many expects of the current status of art of research in mirror symmetry. For instance, it suggests a suitable re-parameterizations of the Landau-Ginzburg mirror model of a complete intersection in a toric variety, proposed by Hori and Vafa in 2000 [@Hori-Vafa], in such a way that, at least for projective hypersurfaces, $f$-duality turns out to exhibit a suitable compactification of this re-parameterized LG mirror model, extending to higher degrees what observed by Hori and Vafa for [Calabi-Yau ]{}projective hypersurfaces (see §\[ssez:HoriVafa\]). Since mostly of the currently proposed LG mirror models are modelled on the so called *Hori-Vafa recipe*, the previous observation introduces possible re-parameterizations of all these LG models, which should be compared with the known ones from the point of view of Homological Mirror Symmetry (HMS) (see considerations given in §\[ssez:iperellittica\]).
More deeply, by observing a natural Landau-Ginzburg/Hypersurface correspondence, conjecturally extending the LG/CY correspondence studied by [@Chiodo-Ruan] (see §\[sssez:LG/Hyp\]), a framing turns out to be the hypersurface counterpart of the “anchoring at infinity” of the compactified LG models proposed by Katzarkov, Kontsevich and Pantev [@KKP] (see \[ssez:KKP-compct\]), so opening the door to conceivable connections with the log-geometry of the Gross-Siebert Intrinsic Mirror Symmetry [@GS-IMS].
Moreover, $f$-duality explains quite well why, passing from a framing to a weak framing, that is, loosing positivity properties, in terms of Kodaira dimension, of the hypersurface (complete intersection) we are considering, translates in loosing completeness properties of the associated mirror partner, so well justifying a description of mirror symmetry in terms of a duality between associated LG models (see Remark \[rem:negKod\]).
A further important remark, is that, since $f$-duality is a duality between framed toric varieties, that is, between pairs given by a complete toric variety and a sufficiently positive torus invariant Weil divisor, multiple mirror partners can, in principle, be assigned by a changing of framing in the same linear equivalence class (see §\[ssez:MWeb\]). This means that, one should think of mirror duality more in terms of a connection between nodes in a web (the Mirror Web) rather than a phenomenon connecting pairs of mirror partners, that is, a symmetry, as done for [Calabi-Yau ]{}varieties. Notice that the multiple mirror phenomenon is a well known one, also for [Calabi-Yau ]{}varieties, after e.g. the R[ø]{}dland example [@Rodland] (see Remark \[rem:mult.mirr\] and references therein).
Finally, following the lines given by Batyrev for [Calabi-Yau ]{}varieties in [@Batyrev02], a conjectural approach, to extending $f$-mirror symmetry beyond a toric embedding, is sketched in §\[ssez:Tdegenarazione\], by means of toric degeneration and geometric transitions.
As Batyrev-Borisov duality, $f$-duality is just a construction to propose candidate mirror partners. After that, one has to prove they are effectively mirror partners, by checking various instances of mirror symmetry. Beyond the [Calabi-Yau ]{}setup, understanding which are those mirror symmetric instances is a bit more involved (see §\[ssez:mirrortest\]). Probably, the deepest way of checking mirror symmetry is the one proposed by Kontsevich’s HMS. But this seems to be a very difficult approach and we refer it to future works. In this paper, a large section is dedicated to check several matching of (stringy) Hodge numbers in the case of projective hypersurfaces of non-negative Kodaira dimension (see §\[sez:ipersuperfici\]). One side of this check (we call $A$-side) turns out to be easily computable (Theorem \[thm:m\*=k\]). The other side of this check (so called $B$-side) is sensibly more intricate, requiring the introduction of suitable partial Gorenstein resolutions, on which studying and computing the involved stringy Hodge numbers (Theorems \[thm:B-mirror0\], \[thm:B-mirror1\] and \[thm:B-mirror2\]). Here, combinatorial computation are rather tricky (see §\[ssez:combinatorica\]). Notice that, the resolution process is quite more natural in the [Calabi-Yau ]{}case, as the [Calabi-Yau ]{}constraint imposes a sort of *canonical* resolutions, so called *maximal projective crepant partial (MPCP)* resolutions [@Batyrev94 Def. 2.2.13]. For Kodaira positive hypersurfaces, the choice of needed resolution is absolutely free, after one construct a suitable partial resolution of $\Q$-Gorenstien singularities to Gorenstein ones (Proposition \[prop:risoluzione\]). This fact makes such a $B$-side check a bit unsatisfactory, as one can go on by blowing up the mirror model, until $h^{1,1}$ reaches the desired value for obtaining a matching with the complex moduli number of the corresponding projective hypersurface.
Anyway, this is a case by case checking, quickly becoming essentially impossible for more general hypersurface and complete intersections in toric varieties, due to the wild singularities $f$-duality produces. According with Chiodo and Ruan [@Chiodo-Ruan], it is generally believed that considering suitably associated LG models may sensibly simplify singularities and give rise to alternative way of checking mirror symmetry. The already mentioned LG/Hypersurface correspondence, presented in §\[ssez:K-dualità\], allows one to drawing an alternative conjectural approach to checking mirror symmetry (see Remark \[rem:K-dualità\]). A more detailed and rigorous exposition of these aspects is postponed to next papers.
This paper is organized as follows. §\[sez:preliminari\] is devoted to introduce the needed notation on toric varieties, their divisors, hypersurfaces and associated stratifications. §\[sez:dualita-ftv\] is dedicated to the definition of framed toric varieties and framed duality. In §\[sez:framingPn\] an important class of framed toric varieties admitting a calibrated $f$-process is presented, namely projective spaces endowed with suitable framings. Then §\[sez:dualita-hyp\] is devoted to present mirror symmetric consequences of $f$-duality for hypersurfaces in complete toric varieties. In the following §\[sez:ipersuperfici\], all these considerations are applied to the important class of examples given by hypersurfaces in $\P^n$ of degree $d\ge n+1$. Then in §\[sez:CI\], $f$-duality is extended to complete intersections subvarieties in complete toric varieties. Finally in §\[sez:open\] many further considerations and open problems are collected: in particular in §\[ssez:NegKod\] weak framed toric varieties are defined, with an application to hypersurfaces in $\P^n$ of degree $d\le n$ and connections with Givental’s LG mirror models.
It is a pleasure to thank M. Artebani for several clarifications about many aspects treated in [@ACG] and T. Hübsch for his interested comments and interesting suggestions, giving rise to perspectives in §\[ssez:Hubsch\]. Many thanks also to S. Filippini, for useful conversation during her last visit in Turin, and to G. Bini for his considerations. I am also indebt with D. Grinberg and M. Riedel, via the `Math Stack Exchange` platform, for useful hints in proving combinatorial Lemmas in §\[ssez:combinatorica\]. Many computations and proofs’ prototypes have been partially performed by means of several Maple routines, mostly of them jointly written with L. Terracini, and some of them based on the Maple package `Convex` [@Convex].
Preliminaries and notation on toric varieties {#sez:preliminari}
=============================================
A *$n$–dimensional toric variety* is an algebraic normal variety $X$ containing the *torus* $T:=(\C^*)^n$ as a Zariski open subset such that the natural multiplicative self–action of the torus can be extended to an action $T\times X\rightarrow X$.
Let us quickly recall the classical approach to toric varieties by means of *cones* and *fans*. For proofs and details the interested reader is referred to the extensive treatments [@Danilov], [@Fulton], [@Oda] and the recent and quite comprehensive [@CLS].
As usual $M$ denotes the *group of characters* $\chi : T \to \C^*$ of $T$ and $N$ the *group of 1–parameter subgroups* $\lambda : \C^* \to T$. It follows that $M$ and $N$ are $n$–dimensional dual lattices via the pairing $$\begin{array}{ccc}
M\times N & \longrightarrow & \operatorname{Hom}(\C^*,\C^*)\cong\C^*\\
\left( \chi,\lambda \right) & \longmapsto
& \chi\circ\lambda
\end{array}$$ which translates into the standard paring $\langle u,v\rangle=\sum u_i v_i$ under the identifications $M\cong\Z^n\cong N$ obtained by setting $\chi(\tt)=\tt^{\uu}:=\prod t_i^{u_i}$ and $\lambda(t)=t^{\v}:=(t^{v_1},\ldots,t^{v_n})$.
Cones and affine toric varieties {#ssez:TV}
--------------------------------
Define $N_{\R}:=N\otimes \R$ and $M_{\R}:=M\otimes\R\cong \operatorname{Hom}(N,\Z)\otimes\R \cong \operatorname{Hom}(N_{\R},\R)$.
A *convex polyhedral cone* (or simply a *cone*) $\sigma$ is the subset of $N_{\R}$ defined by $$\sigma = \langle \v_1,\ldots,\v_s\rangle:=\{ r_1 \v_1 + \dots + r_s \v_s \in N_{\R} \mid r_i\in\R_{\geq 0} \}$$ Vectors $\v_1,\ldots,\v_s\in N_{\R}$ are said *to generate* $\sigma$; $\v_i$ is called a *primitive* generator if it generates the semigroup $\langle\v_i\rangle\cap N$. A cone $\s=\langle \v_1,\ldots,\v_s\rangle$ is called *rational* if $\v_1,\ldots,\v_s\in N$, *simplicial* if $\v_1,\ldots,\v_s$ are $\R$–linear independent and *non-singular* if primitive generators $\v_1,\ldots,\v_s$ can be extended to giving a basis of the lattice $N$.
A cone $\s$ is called *strongly convex* or *pointed* if it does not contain a linear subspace of positive dimension of $N_{\R}$.
The *dual cone $\s^{\vee}$ of $\s$* is the subset of $M_{\R}$ defined by $$\sigma^{\vee} = \{ \uu \in M_{\R} \mid \forall\ \v \in \sigma \quad \langle \uu, \v \rangle \ge 0 \}$$ A *face $\tau$ of $\s$* (denoted by $\tau <\s$) is the subset defined by $$\tau = \sigma \cap \uu^{\bot} = \{\v \in \sigma \mid \langle \uu, \v \rangle = 0 \}$$ for some $\uu\in \sigma ^{\vee}$. Observe that also $\tau$ is a cone.
A *facet* $\tau$ of a cone $\s$ is a codimension 1 face, denoted by $\tau<^1\s$.
Gordon’s Lemma ensures that the semigroup $S_{\s}:=\s^{\vee}\cap M$ is *finitely generated*. Then also the associated $\C$–algebra $A_{\s}:=\C[S_{\s}]$ is finitely generated. A choice of $m$ generators gives a presentation of $A_{\s}$ $$A_{\s}\cong \C[X_1,\dots,X_m]/I_{\s}$$ Then $U_{\s}:=\operatorname{Spec}(A_{\s})\subset\C^m$ is an *affine toric variety*. Since a closed point $x\in U_{\s}$ is an evaluation of elements in $\C[S_{\s}]$ satisfying the relations generating $I_{\s}$, then it can be identified with a semigroup morphism $x:S_{\s}\rightarrow\C$ assigned by thinking of $\C$ as a multiplicative semigroup. In particular the *characteristic morphism* $$\label{caratteristico}
\begin{array}{cccc}
x_{\s}&:\s^{\vee}\cap M & \longrightarrow & \C\\
& \uu & \longmapsto & \left\{\begin{array}{cc}
1 & \text{if $\uu\in\s^{\bot}$} \\
0 & \text{otherwise}
\end{array}
\right.
\end{array}$$ which is well defined since $\s^{\bot}<\s^{\vee}$, defines a *characteristic point* $x_{\s}\in U_{\s}$ whose torus orbit $O_{\s}$ turns out to be a $(n-\dim(\s))$–dimensional torus embedded in $U_{\s}$.
Fans and toric varieties {#ssez:fans&tv}
------------------------
A *fan* $\Si$ is a finite set of cones $\s\subset N_{\R}$ such that
1. for any cone $\s\in\Si$ and for any face $\tau<\s$ then $\tau\in\Si$,
2. for any $\s,\tau\in\Si$ then $\s\cap\tau<\s$ and $\s\cap\tau<\tau$.
For every $i$ with $0\leq i\leq n$ denote by $\Si(i)\subset \Si$ the subset of $i$–dimensional cones, called the *$i$–skeleton of $\Si$*.
A fan $\Si$ is called *simplicial* if every cone $\s\in\Si$ is rational and simplicial, and is called *non-singular* if every such cone is non-singular. The *support* of a fan $\Si$ is the subset $|\Si|\subset N_{\R}$ obtained as the union of all of its cones i.e. $$|\Si|:= \bigcup_{\s\in\Si} \s \subset N_{\R}\ .$$ If $|\Si|=N_{\R}$ then $\Si$ will be called *complete*.
Since for any face $\tau <\s$ the semigroup $S_{\s}$ turns out to be a sub-semigroup of $S_{\tau}$, there is an induced immersion $U_{\tau}\hookrightarrow U_{\s}$ between the associated affine toric varieties which embeds $U_{\tau}$ as a principal open subset of $U_{\s}$. Given a fan $\Si$ one can construct *an associated toric variety $X(\Si)$* by patching all the affine toric varieties $\{U_{\s}\ |\ \s\in\Si \}$ along the principal open subsets associated with any common face. Moreover *for every toric variety $X$ there exists a fan $\Si$ such that $X\cong X(\Si)$*. It turns out that:
- *$X(\Si)$ is non-singular if and only if the fan $\Si$ is non-singular,*
- *$X(\Si)$ is complete if and only if the fan $\Si$ is complete.*
Let $\v_\rho$ be a primitive generator of the ray $\rho\in \Si(1)$. Up to an identification $N\cong\Z^n$, where $n:=\dim X$, and setting $m:=|\Si(1)|$ $$V=\left(\v_\rho\,|\,\rho\in\Si(1)\right)=(\v_1\,\cdots\,\v_m)$$ gives a $n\times m$ integer matrix called *a fan matrix of $\Si$*. Notice that $\Si$ determines $V$ up to the choice of a basis of $N$ and of a permutation of columns (i.e. generators $\v_\rho$), that is, $V$ and $V'$ are *equivalent* fan matrices if $$\label{M-equivalenza}
\exists\,A\in\GL(n,\Z)\,,\ \exists\,B\in\mathfrak{S}_m\leq \GL(m,\Z)\quad V'=A\cdot V\cdot B$$
Divisors on Toric varieties
---------------------------
Let $\mathcal{W}(X)$ denote the group of Weil divisors of a toric variety $X=X(\Si)$. Then its subgroup of *torus–invariant Weil divisors* is given by $$\mathcal{W}_T(X)=\left\langle D_{\rho } \mid \rho \in \Sigma (1)\right\rangle_{\Z} =
\bigoplus_{\rho \in \Sigma (1)}\Z\cdot D_{\rho }$$ where $D_{\rho}=\overline{\T\cdot x_\rho}$, being $\T\cong\operatorname{Hom}(N,\C^*)$ the acting torus and $x_\rho$ the distinguished point of $\rho$, as defined in (\[caratteristico\]). Let $\mathcal{P}(X)\subset\mathcal{W}(X)$ be the subgroup of *principal divisors* and $\v_{\rho}$ be the generator of the monoid $\rho\cap N$. Then the morphism $$\label{div}
\begin{array}{llll}
div : & M & \longrightarrow & \mathcal{P}(X)\cap \mathcal{W}_{T}(X)=:
\mathcal{P}_{T}(X) \\
& \uu & \longmapsto & div(\uu):=\sum_{\rho \in \Sigma (1)}\langle \uu,\v_{\rho }\rangle
D_{\rho }
\end{array}$$ is surjective. Let $V=(\v_1,\ldots,\v_{n+r})$ be a fan matrix of $\Si$, with respect to a chosen identification $N\cong\Z^n$. Then the transposed matrix $V^T$ is a representative matrix of the $\Z$-linear morphism $div$ defined in (\[div\]), with respect to the basis $\{D_1,\ldots,D_{n+r}\}$ of $\mathcal{W}_T(X)$.
Let $\operatorname{Pic}(X)$ be the group of line bundles modulo isomorphism. It is well known that for an *irreducible* variety $X$ the map $D\mapsto\mathcal{O}_X(D)$ induces an isomorphism $\mathcal{C}(X)/\mathcal{P}(X)\cong\operatorname{Pic}(X)$, where $\mathcal{C}(X)\subset\mathcal{W}(X)$ denotes the subgroup of Cartier divisors. The divisor class group is defined as the group of Weil divisors modulo rational (hence linear) equivalence, i.e. $\operatorname{Cl}(X):=\mathcal{W}(X)/\mathcal{P}(X)$. Then the inclusion $\mathcal{C}(X)\subset\mathcal{W}(X)$ passes through the quotient giving an immersion $\operatorname{Pic}(X)\hookrightarrow \operatorname{Cl}(X)$.
A toric variety $X=X(\Si)$ is called *non-degenerate* if the support $|\Si|$ spans $N_{\R}$: in particular this means that it cannot admit torus factors, or, equivalently, that $H^0(X,\cO_X^*)\cong\C^*$. Then, the cardinality of the 1-skeleton is given by $$|\Si(1)|=n+r$$ where $r:=\operatorname{rk}\operatorname{Pic}(X)\geq 1$ is the *Picard number of* $X$, also called *the rank of $X$*, in the following.
\[def:Fmatrice\] An *$F$–matrix* is a $n\times (n+r)$ matrix $V$ with integer entries, satisfying the conditions:
- $\operatorname{rk}(V)=n$;
- $V$ is *$F$–complete* i.e. $\langle V\rangle=N_{\R}\cong\R^n$ ;
- all the columns of $V$ are non zero;
- if ${\bf v}$ is a column of $V$, then $V$ does not contain another column of the form $\lambda {\bf v}$ where $\lambda>0$ is a real number.
A $F$–matrix $V$ is called *reduced* if every column of $V$ is composed by coprime entries .
For instance, a fan matrix of a complete toric variety $X(\Si)$ is always a reduced $F$–matrix.
### Notation {#sssez:SF}
Given a reduced $F$-matrix $V$, in the following $\SF(V)$ will denote the set of all complete and simplicial fans whose 1-skeleton is given by all the rays generated by the columns of $V$. Moreover, $$\P\SF(V)\subset\SF(V)$$ will denote the subset of those fans whose associated toric variety $X(\Si)$ is projective.
Polytopes of divisors and associated fans and varieties {#ssez:politopi}
-------------------------------------------------------
A *polytope* $\D\subset M_{\R}$ is the convex hull of a finite set $S$ of points, that is $\D=\operatorname{Conv}(S)$.
If $S\subseteq M$ then $\D$ is called a *lattice polytope*. When $\D$ is a full dimensional polytope its presentation as an intersection of closed half-spaces has an especially nice form, because each facet $\Phi<^1\D$ has a unique supporting affine hyperplane. We denote such an hyperplane and the corresponding closed half-space as $$H_\Phi=\{\m\in\M_\R\,|\,\langle\m,\n_\phi\rangle=-a_\Phi\}\quad,\quad H^+_\Phi=\{\m\in\M_\R\,|\,\langle\m,\n_\phi\rangle\geq-a_\Phi\}$$ where $(\n_\Phi,a_\Phi)\in N_\R\times \R$ is unique up to multiplication by a positive real number. We call $\n_\Phi$ an *inward pointing normal* vector of the facet $\Phi$. It follows that $$\label{ss-intersezione}
\D=\bigcap_{\Phi<^1\D} H^+_\Phi= \{\m\in M_\R\,|\,\forall\,\Phi<^1\D\quad\langle\m,\n_\Phi\rangle\geq-a_\Phi\}$$ The relative interior of $\D$ will be denoted by $\operatorname{Relint}{\D}$, or simply $\operatorname{Int}\D$ when $\D$ is full dimensional.
In the following *we will consider full dimensional polytopes only, unless otherwise advised*.
The *polar polytope* $\D^*$ of a polytope $\D\subseteq M_\R$ containing the origin $\0\in M$ as an interior point, that is $\0\in\operatorname{Int}{\D}$, is defined as follows $$\label{polare}
\D^*:=\{\n\in N_\R\,|\,\forall\,\m\in\D\quad\langle\n,\m\rangle\geq -1\}\subseteq N_\R$$ It is a full dimensional polytope in $N_\R$ with $\0\in\operatorname{Int}{\D^*}$ and $(\D^*)^*=\D$. In particular, if $\D$ admits the presentation given in (\[ss-intersezione\]) then $$\label{polare2}
\D^*=\operatorname{Conv}(\{a_\Phi^{-1}\n_\Phi\,|\,\forall\,\Phi<^1\D\})\subseteq N_\R$$ (see [@CLS Exer. 2.2.1]). Clearly, in general, $\D^*$ is not a lattice polytope in $N$, even if $\D$ is a lattice polytope in $M$. A lattice polytope $\D$ is called *reflexive* if $\0\in\operatorname{Int}\D$ and $\D^*$ is still a lattice polytope. By [@Batyrev94 Thm. 4.1.6] $$\D\ \text{is reflexive}\ \Longleftrightarrow\ \operatorname{Int}\D\cap M=\{\0\}$$ Given a divisor $D=\sum_{\rho\in\Si(1)}a_\rho D_\rho\in\Weil(X(\Si))$, the following polyhedron $$\label{div-politopo}
\D_D:=\{\m\in M_\R\,|\,\forall\,\rho\in\Si(1)\quad\langle\m,\v_\rho\rangle\geq -a_\rho\}=\{\m\in M_\R\,|\,V^T\cdot\m \geq -\aa\}$$ is called the *polyhedron associated to $D$*, where $V=\left(\v_\rho\right)_{\rho\in\Si(1)}$ is a fan matrix of $X$ and $\aa=\left(a_\rho\right)_{\rho\in\Si(1)}$ is the column vector of coefficients of $D$. In general it is not a polytope, but just a polyhedron as intersection of a finitely many closed half spaces.
\[prop:gg\] If $X(\Si)$ is complete then, for any $D$ in $\Weil(X)$, the associated polyhedron $\D_D$ is a polytope. Moreover:
1. $D$ is *basepoint free*, that is $\cO_X(D)$ is generated by global section, if and only if $\D_D=\operatorname{Conv}(\{\m_\s\in M\,|\,\s\in\Si(n)\})$,
2. $D$ is ample if and only if $\D_D=\operatorname{Conv}(\{\m_\s\in M\,|\,\s\in\Si(n)\})$ and $\s\neq\s'$ implies $\m_\s\neq\m_{\s'}$.
Recall that a Weil divisor $D$ is *semi-ample* if a positive multiple $kD$, $k\in\N$, is basepoint free (hence Cartier). In particular, if $X(\Si)$ is complete and $D$ semi-ample, then $$\D_{kD}=k\D_D=\operatorname{Conv}(\{\m_\s\in M\,|\,\s\in\Si(n)\})$$ is a lattice polytope.
\[prop:semiampio\] Let $|\Si|$ be convex of full dimension. Then $D$ is semi-ample if and only if $kD$ is numerically effective (nef), for some $k\in\N$, that is $kD$ is Cartier and $kD\cdot C\geq 0$, for any complete curve $C\subset X$.
Starting from a lattice polytope $\D$ one can construct a *projective* toric variety $\P_\D$ as follows. For any nonempty face $\phi< \D$ consider the dual cone $\s_\phi^\vee\subseteq N_\R$ of the cone $$\s_\phi:=\{r(\m-\m')\ |\ \m\in\D\ ,\ \m'\in \phi\ ,\ r\in\R_{\geq 0}\}\subseteq M_{\R}$$ Then $\Si_{\D}^\perp:=\{\s_\phi^\vee\ |\ \phi< \D\}$ turns out to be a fan, called the *normal fan* of the polytope $\D$, and $\P_\D$ is the associated toric variety. It is projective as there exists an ample divisor $H$ of $\P_{\D}$ whose associated polytope is precisely $\D$.
A further *complete* toric variety $\XX_\D$ can be associated with a lattice polytope $\D$ such that $\0\in\operatorname{Int}\D$. Namely, for every facet $\Phi<\D$ consider the cone *projecting $\Phi$ from the origin*, that is $$\label{cono su facciata}
\s_\Phi:=\{r\m\,|\,\m\in\Phi\ ,\ r\in\R_{\geq 0}\}\subseteq M_\R$$ Then $\Si_\D:=\{\tau\,|\,\exists\,\Phi<^1\D:\ \tau<\s_\Phi\}$ turns out to be a fan, called the *fan over the polytope $\D$*, and $\XX_\D$ is the associated toric variety. It is complete as the support $|\Si_\D|$ is the whole $M_\R$ (clearly for $\XX_\D$, the role of the dual lattices $M,N$ is reversed with respect to $\P_\D$). This is a direct consequence of the following
\[prop:Fmatricedipolitopo\] Given an identification $M\cong\Z^n$ and a lattice polytope $$\D:=\operatorname{Conv}(\{\m_i\in M\,|\,i=1,\ldots,m\})$$ let $V_\D=\left(
\begin{array}{ccc}
\v_1 & \cdots & \v_m \\
\end{array}
\right)$ be the $n\times m$ integer matrix defined by the generators $\v_i$ of the semi-groups $\langle\m_i\rangle\cap M$, for any $1\leq i\leq m$. Then $V_\D$ is a reduced $F$-matrix if and only if $\0\in\operatorname{Int}\D$. In particular, if $\0\in\operatorname{Int}\D$ then $V_\D$ is a fan matrix of $\XX_\D$.
Assume $V_\D$ is an $F$-matrix. Then $V_\D$ is clearly reduced as all the $\v_i$’s are primitive. Moreover, choosing the first column $\v_1$ of $V_\D$, the opposite vector $-\v_1$ belongs to the cone $\langle\v_2,\ldots,\v_n\rangle$, by . Then $$\0=\v_1-\v_1=\v_1 +\sum_{j=2}^m \lambda_j\v_j=\sum_{i=1}^m \mu_i\m_i$$ where $$\mu_1={\|\v_1\|\over\|\m_1\|}>0\ ,\quad \forall\,j\geq 2\quad \mu_j=\lambda_j{\|\v_j\|\over\|\m_j\|}>0$$ One can then conclude that $\0\in\operatorname{Int}\operatorname{Conv}(\{\m_k\}_{k=1}^m)=\operatorname{Int}\D$ by setting $\mu:=\sum_k\mu_k$ and writing $\0=\sum_k (\mu_k/\mu)\,\m_k$.
Viceversa, assume $\0\in\operatorname{Int}\D=\operatorname{Int}\operatorname{Conv}(\{\m_k\}_{k=1}^m)$. Conditions (a), (c) and (d) in Definition \[def:Fmatrice\] are clearly satisfied. To show that $V_\D$ is $F$-complete, for any vector $\v\in M_\R$ consider the polytope $$\D':=\operatorname{Conv}(\m_1\,\ldots,\m_m,-\v)\subseteq M_\R$$ Since $\D\subseteq\D'$, one has that $\0\in\operatorname{Int}\D'$, meaning that $$\exists\,\mu_1>0,\ldots,\mu_m>0,\mu>0\,:\quad \sum_k\mu_k +\mu=1\ ,\quad\0=\sum_k\mu_k\m_k-\mu\v$$ meaning that $$\v=\sum_k\l_k\v_k\ ,\quad\text{with}\quad \forall\,k\ \l_k={\mu_k\,\|\m_k\|\over\mu\,\|\v_k\|}>0\ \Longrightarrow\ \v\in\langle V_\D\rangle$$
\[rem:reflexive\] If $\D\subseteq M_\R$ is a reflexive polytope, then $$\P_\D\cong\XX_{\D^*}\quad\text{and}\quad\XX_\D\cong\P_{\D^*}$$ These isomorphisms are induced by identity morphisms of lattices $N$ and $M$, respectively.
\[cor:smallres\] Let $\D$ be a lattice polytope such that $\0\in\operatorname{Int}(\D)$ and $V_\D$ be the fan matrix of $\XX_\D$, as constructed in the previous Proposition \[prop:Fmatricedipolitopo\]. Then, for any $\Si\in\SF(V_\D)$ which is a refinement of $\Si_\D$, the associated toric variety $X(\Si)$ is a $\Q$-factorial small resolution of $\XX_\D$.
In fact, every refinement $\Si\in\SF(V)$ of $\Si_\D$ is obtained by a simplicial subdivision of cones in $\Si_\D$. In particular, the induced birational resolution $X(\Si)\longrightarrow\XX_\D$ is small, as $\Si(1)=\Si_\D(1)=\{\langle\v_1\rangle,\ldots,\langle\v_m\rangle\}$.
Cones of divisors {#ssez:coni&div}
-----------------
Let $X(\Si)$ be a complete toric variety. Then there is a short exact sequence $$\label{complete deg sequence}
\xymatrix@1{0\ar[r]& M \ar[r]^-{div}_-{V^T} & *!U(.45){\bigoplus\limits_{\rho \in \Sigma (1)} \Z \cdot D_{\rho}}
\ar[r]^-d_-Q & \operatorname{Cl}(X) \ar[r] & 0 }$$ (see e.g. [@Fulton §3.4], [@CLS Prop. 4.2.5]). The representative matrices, $V$ and $Q$, of the $\Z$-linear morphisms $div$ and $d$, respectively, gives a fan matrix and *weight matrix*, respectively, of $X$. Since $X$ is complete, $V$ is a reduced $F$-matrix and $Q$ is a *Gale dual matrix* of $V$, which *can be assumed to be positive* . This means that:
- *the image $\operatorname{Im}(d)=d(\Weil(X))$ of the degree morphism in (\[complete deg sequence\]), can be assumed contained in the positive orthant $\R^r_+$ of $\R^r\cong\operatorname{Cl}(X)\otimes\R$, being $r$ the Picard number of $X$*.
Recall that every divisor of $X$ is linearly equivalent to a torus invariant divisor. This means that, in the isomorphism $\operatorname{Cl}(X)\otimes\R\cong\R^r$ the cone $\overline{\operatorname{Eff}}(X)\subseteq\operatorname{Cl}(X)\otimes\R$, which is the closure of the cone generated by classes of effective divisors, is identified with cone $\langle Q\rangle$ generated by the columns of $Q$, that is $$\label{iso-coni}
\overline{\operatorname{Eff}}(X)\cong\langle Q\rangle$$ Let us now introduce the following:
### Notation {#sssez:notazione}
Let $A$ be a $d\times m$ matrix. For any subset $I\subseteq\{1,\ldots,m\}$ we will denote by $A_I$ the sub-matrix of $A$ obtained by considering the columns indexed by $I$, only, and by $A^I$ the complementary sub-matrix of $A_I$ in $A$, that is, the one obtained by considering only the columns not indexed by $I$.
Coming back to the situation of a complete toric variety $X(\Si)$, of dimension $n$ and Picard number $r$, to every cone $\s\in\Si$ one can associate a subset $I\subseteq\{1,\ldots,m=n+r\}$ such that $$\s=\langle V_I\rangle\subseteq N_\R$$ Define $\I_\Si:=\{I\in\mathfrak{P}(\{1,\ldots,m\})\,|\,\langle V_I\rangle\in\Si\}$, that is, $\Si=\{\langle V_I\rangle\,|\,I\in\I_\Si\}$. Then set $$\operatorname{Mov}(Q):=\bigcap_{i=1}^m \langle Q^{\{i\}}\rangle\quad,\quad \operatorname{Nef}(\I):=\bigcap_{I\in\I} \langle Q^I\rangle\quad \text{(for any}\ \I\subseteq\mathfrak{P}(\{1,\ldots,m\})\,)$$ Recall that the cone $\overline{\operatorname{Mov}}(X)$, which is the closure of the one generated by classes of *movable* divisors, and the cone $\operatorname{Nef}(X)$, generated by classes of nef divisors, are both sub-cones of the effective cone $\overline{\operatorname{Eff}}(X)$. Then, the isomorphism (\[iso-coni\]) descends to give isomorphisms $$\overline{\operatorname{Mov}}(X)\cong\operatorname{Mov}(Q)\quad,\quad \operatorname{Nef}(X)\cong\operatorname{Nef}(\I_\Si)$$ More precisely, recalling notation \[sssez:SF\], we get the following
*[@CLS Thm. 15.1.10(c)]*\[prop:nef\] If $V=\left(
\begin{array}{ccc}
\v_1 & \ldots & \v_{n+r} \\
\end{array}
\right)$ is an $F$–matrix then, for every fan $\Si\in\P\SF(V)$ there is a natural isomorphism $$\operatorname{Pic}(X(\Si))\otimes\R\cong\operatorname{Cl}(X)\otimes\R\cong \R^r$$ taking the cones $$\operatorname{Nef}(X(\Si))\subseteq\overline{\operatorname{Mov}}(X(\Si))\subseteq\overline{\operatorname{Eff}}(X(\Si))\subseteq\R^r$$ to the sub-cones of the positive orthant $$\operatorname{Nef}(\I_\Si)\subseteq\operatorname{Mov}(Q)\subseteq\langle Q\rangle\subseteq\R^r_+$$ In particular, if $d:\mathcal{W}_T(X(\Si))\to\operatorname{Cl}(X(\Si))$ is the degree morphism, then a Weil divisor $D$ on $X(\Si)$ admits a nef (ample) positive multiple if and only if its class $[D]=d(D)\in\operatorname{Nef}(\I_\Si)$ ($d(D)\in\operatorname{Relint}\operatorname{Nef}(\I_\Si)$, resp.).
The following is a useful application to $\Q$-factorial small resolutions of a complete toric variety $X(\Si)$.
\[prop:risoluzioni\]
Let $X(\Si)$ be a complete toric variety and $V$ be a fan matrix of $X$. Then, for any $\Xi\in\SF(V)$ giving a refinement of $\Si$, the identity morphism of the common lattice $N$, namely $\operatorname{id}_N:N\longrightarrow N$, induces a fans’ morphism from $\Xi$ to $\Si$. The induced morphism of toric varieties $$\vf:Y(\Xi)\longrightarrow X(\Si)$$ is a $\Q$-factorial small resolution. Then, there is the following isomorphism of divisorial exact sequences $$\xymatrix{0\ar[r]& M \ar[d]^-{\operatorname{id}_M}_-{I_n}\ar[r]^-{div}_-{V^T} & \Weil(X)\ar[d]^-{\vf^*}_-{I_m}\ar[r]^-d_-Q & \operatorname{Cl}(X)\ar[d]^-{\overline{\vf}^*}_-{I_r} \ar[r] & 0\\
0\ar[r]& M \ar[r]^-{div}_-{V^T} & \Weil(Y)\ar[r]^-d_-Q & \operatorname{Cl}(Y) \ar[r] & 0}$$ where $\vf^*$ is the pull-back of Weil divisors and $\overline{\vf}^*$ is the induced morphism on divisor classes. By composing $\overline{\vf}^*$ with isomorphisms $\operatorname{Pic}(X)_\R\cong\R^r$ and $\operatorname{Pic}(Y)_\R\cong\R^r$, defined in the previous Proposition \[prop:nef\], there follows the following identification of nested divisorial sub-cones of the positive orthant $\R^r_+$ $$\xymatrix{\operatorname{Nef}(\I_{\Si})\cong\operatorname{Nef}(X)\ar@{^(->}[dd]\ar@<-1ex>@{^(->}[rd]&\\
&\left(\operatorname{Mov}(Q)\cong\overline{\operatorname{Mov}}(X)\stackrel{\overline{\vf}^*_\R}{\cong}\overline{\operatorname{Mov}}(Y)\right)
\,\subseteq\, \left(\langle Q\rangle\cong\overline{\operatorname{Eff}}(X)\stackrel{\overline{\vf}^*_\R}{\cong}\overline{\operatorname{Eff}}(Y)\right)\\
\operatorname{Nef}(\I_{\Xi})\cong\operatorname{Nef}(Y)\ar@{^(->}[ur]&
}$$ Moreover, for every Weil divisor $D=\sum_{\rho\in\Si(1)}a_\rho D_\rho\in\Weil(X)$, there follows the identification of associated polytopes $$\{\m\in M_\R\,|\,V^T\cdot\m \geq -\aa\}=\D_D=\operatorname{id}_M(\D_D)=\D_{\vf^*(D)}$$ where $\aa=(a_\rho)_{\rho\in\Si(1)}$. In particular $D$ is semi-ample if and only if its class $[D]$ belongs to $\operatorname{Nef}(X)$. Then there exists a positive integer $k\in\N$ such that $$\operatorname{Conv}(\{\m_\s\in M\,|\,\s\in\Si(n)\})=k\D_D\stackrel{\operatorname{id}_M}{=}k\D_{\vf^*(D)}
=\operatorname{Conv}(\{\m_I\in M\,|\,I\in\I_{\Xi(n)}\})$$ where $$\label{mI}
\m_I:=-k(V_I^T)^{-1}\cdot \aa_I$$ being $\aa_I$ the subvector of $\aa$ whose entries are indexed by $I$. In particular, this means that, for every $\s\in\Si(n)$, $\m_\s=\m_I$ for any $I\in\I_{\Xi(n)}$ such that $\langle V_I\rangle\subseteq \s$.
Results on divisorial cones are direct consequences of the previous Proposition \[prop:nef\]. In particular[^1] $$\operatorname{Nef}(\I_\Si)=\bigcap_{I\in\I_\Si} \langle Q^I\rangle\subseteq\bigcap_{J\in\I_\Xi} \langle Q^J\rangle = \operatorname{Nef}(\I_\Xi)$$ because $\Xi$ is a refinement of $\Si$, so implying that $$\forall\,I\in\I_\Si\ \exists\,J\in\I_\Xi\,: \langle V_J\rangle \subseteq \langle V_I\rangle\ \Longleftrightarrow\ \langle Q^I\rangle \subseteq \langle Q^J\rangle$$ Results on divisorial polytopes come, on the one hand, directly from the definition given in (\[div-politopo\]) and, on the other hand, from Proposition \[prop:semiampio\], giving that $[D]\in\operatorname{Nef}(X)$, and Proposition \[prop:gg\]. In fact, a positive multiple $kD$ is besepoint free, so giving that $k\D_D=\operatorname{Conv}(\{\m_\s\in M\,|\,\s\in\Si(n)\})$. Since $\operatorname{Nef}(X)\subseteq\operatorname{Nef}(Y)$, $[\vf^*(D)]=\overline{\vf}^*([D])\in\operatorname{Nef}(Y)$ so giving that $\vf^*(D)$ is semi-ample, too, still by Proposition \[prop:nef\]. Therefore $$k\D_D=k\D_{\vf^*(D)}=\operatorname{Conv}(\{\m_\tau\in M\,|\,\tau\in\Xi(n)\})=\operatorname{Conv}(\{\m_I\in M\,|\,I\in\I_{\Xi(n)}\})$$ where $\m_\tau=\m_I$ whenever $\tau=\langle V_I\rangle$ and $\m_I$ is defined as in (\[mI\]).
Non-degenerate hypersurfaces and their stratification {#ssez:regular}
-----------------------------------------------------
Given a toric variety $X=X(\Si)$, let $\T\subseteq X$ be the maximal acting torus on $X$. Consider a Laurent polynomial $$f=\sum_{\substack{m\in M\\\text{finite}}} c_m\chi^m\ ,\quad c_m\in\C^*$$ Denote by $Z_f\subseteq\T$ the zero-locus of $f$ in $\T$ and let $Y_f$ be its closure in $X$.
\[def:Si-regolare\] A Laurent polynomial $f$, and the associated hypersurfaces $Z_f\subseteq\T$ and $Y_f\subseteq X(\Si)$, are called *non-degenerate* $\Si$ (or, equivalently, *$\Si$-regular*) if, for every $\s\in\Si$, the associated $\s$-stratum $Y_{f,\s}:=Y_f\cap \T\cdot x_\s$ is empty or a smooth subvariety of codimension 1 in the torus orbit $\T\cdot x_\s$. In other words, non-degenerate means that $Y_f$ admits only *transversal intersections* with all the torus orbits $\T\cdot x_\s$, $\s\in\Si$.
A duality between framed toric varieties {#sez:dualita-ftv}
========================================
Let us start the present section by introducing the main character of this paper.
\[def:ftv\] A *framed toric variety* is a couple $(X,D)$ where:
- $X$ is a complete toric variety, with $\dim(X)=n$ and $\operatorname{rk}(\operatorname{Pic}(X))=r$,
- $D=\sum_{\rho\in\Si(1)}a_\rho D_\rho=\sum_{i=1}^m a_i D_i\in\Weil(X)$, with $m=n+r$, is a *strictly effective* (that is $a_i>0$, for every $i$), torus invariant Weil divisor, called a *framing* of $X$.
A *morphism of framed toric varieties* $f:(X,D)\longrightarrow(X',D')$ is a morphism of underlying toric varieties $f:X\longrightarrow X'$ inducing a well defined pull-back morphism on torus invariant Weil divisors $f^*:\Weil(X')\longrightarrow\Weil(X)$ such that $f^*D'=D$. If $f$ is an isomorphism of toric varieties, then it gives an *isomorphism of framed toric varieties* $f:(X,D)\cong(X',D')$. It is well defined the category **ftv** of framed toric varieties.
Framed duality
--------------
Given a ftv $$\left(X(\Si),D_\aa=\sum_{\rho\in\Si(1)} a_\rho D_\rho\right)=:(X,\aa)$$ consider the polytope associated with $D_\aa$ $$\label{Delta_a}
\D_{\aa}:=\D_{D_\aa}=\{\m\in M_\R\,|\,V^T\cdot\m\geq -\aa\}$$ being $V$ a fan matrix of $X$. In general, $\D_\aa$ is the convex hull of a finite subset $\{\m_1,\ldots,\m_s\}\subset M_\Q:=M\otimes\Q$. Let $k\in\N$ be the minimum positive integer such that $k\m_i\in M$, for every $i=1,\ldots,s$. Then $$k\D_\aa=\D_{{kD_\aa}}=\operatorname{Conv}(k\m_1,\ldots,k\m_s)$$ is a lattice polytope. Since ${D_\aa}$ is strictly effective, that is $-\aa<\0$, certainly $\0\in\operatorname{Int}\D_\aa$, meaning that $\0\in\operatorname{Int}(k\D_\aa)$, for any positive integer $k\in\N$.
On the other hand, define the *integer part* of a polytope $\D\subseteq M_\R$ as $$[\D]:=\operatorname{Conv}(\{\m\in M\cap\D\})$$ Clearly, if $\D$ is a lattice polytope then $[\D]=\D$.
\[def:Deltapolytope\] The *framing polytope* (*$f$-polytope*) of a ftv $(X,\aa)$ is the lattice polytope $\D(X,\aa)\subseteq M_\R$ so defined: $$\label{k0}
\D(X,\aa):=[k_0\D_\aa]\quad,\quad k_0:=\min\{k\in\N\,|\,\0\in\operatorname{Int}[k\D_\aa]\}$$
For what observed above, $k_0$ is well defined. Notice that $k_0$ may be bigger than 1: in fact it may happen that $\0$ is not an interior point of the integer part $[\D_\aa]$, as the following Example \[ex:k0>1\] shows. On the other hand $k_0=1$ when $\D_\aa$ is a lattice polytope: in this case $\D(X,\aa)=\D_\aa$.
\[ex:k0>1\] Consider the ftv given by $(X,\aa)=(\P(1,2,5),(2,1,1))$. Then $$\D_\aa=\operatorname{Conv}\left(
\begin{array}{ccc}
3 & {-3/ 5} & {-3/ 2} \\
1 & {-4/ 5} & 1 \\
\end{array}
\right)\ \Longrightarrow\ [\D_\aa]=\operatorname{Conv}\left(
\begin{array}{cccc}
3 & -1 & 1 & -1 \\
1 & 1 & 0 & 0 \\
\end{array}
\right)$$ Then $\0\not\in\operatorname{Int}[\D_\aa]$. But $\0\in\operatorname{Int}[2\D_\aa]$, so giving $k_0=2$ in the previous Definition \[def:Deltapolytope\].
\[rem:anticanonico\] Assume $\aa=\1$, that is $D_\aa=-K_X$ is the anti-canonical divisor of the complete toric variety $X$. Then, if $\D_\1$ is a reflexive polytope, the above construction gives $$\D(X,\aa)=\D_\aa=\D_\1=\D_{-K_X}$$
### The $f$-dual ftv and its small $\Q$-factorial resolutions {#sssez:deltadual}
Associated with the construction of the lattice polytope $\D(X,\aa)$ there is the complete toric variety $$\XX_\aa:=\XX_{\D(X,\aa)}$$ given by the fan $\Si_\aa:=\Si_{\D(X,\aa)}$ over the lattice polytope $\D(X,\aa)$. Let $\Lambda_\aa$ be the fan matrix of $\XX_\aa$ constructed in Proposition \[prop:Fmatricedipolitopo\]: it is a $n\times m'$ integer matrix. Given a fan matrix $V$ of $X$, which is a $n\times m$ integer matrix, define $$M_\aa:=V^T\cdot\L_\aa\in\mathbf{M}(m\times m';\Z)$$ Let $\bb=(b_j)_{j=1}^{m'}$ be *the minimum strictly positive column vector* such that $$\label{b}
M_\aa^T+B\geq \0\quad\text{where}\quad B:=\underbrace{\left(\,\bb\ \cdots\ \bb\,\right)}_{m\ \text{times}}\,\in \mathbf{M}(m'\times m;\N)$$
\[def:dual-ftv\] Calling $D'_\bb:=\sum_{j=1}^{m'}b_jD'_j$, where $D'_1,\ldots,D'_{m'}$ are the torus invariant prime divisors generating $\Weil(\XX_\aa)$, then $(\XX_\aa,\bb):=(\XX_\aa,D'_\bb)$ is a ftv, called *the framed dual ($f$-dual) of $(X,\aa)$*.
The following statement is a direct application of the Propostion \[prop:risoluzioni\].
\[cor:risoluzioni\] For every fan $\Xi\in\SF(\L_\aa)$ such that $\Xi$ refines $\Si_\aa$ there is a well defined birational morphism $\vf:Y(\Xi)\longrightarrow \XX_\aa$ which is a $\Q$-factorial small resolution. In particular, for any such $\Xi$, $(Y(\Xi),\vf^*D'_\bb)$ is a $\Q$-factorial ftv.
\[rem:natural\] In a sense, the choice of the framing $\bb$ given in (\[b\]) is *natural*. In fact, let us assume that the polytope $\D_\aa$ is a lattice polytope with *primitive vertices*, that is $\D_\aa=\operatorname{Conv}(\L_\aa)$. Then, definition (\[Delta\_a\]) of $\D_\aa$ ensures that the framing $\aa$ of $X$ is the minimum strictly positive column vector such that $$\label{Ma}
M_\aa+ A=V^T\cdot\L_\aa + A\geq 0\quad \text{where}\quad A:=\underbrace{\left(\,\aa\ \cdots\ \aa\,\right)}_{m'\ \text{times}}\,\in \mathbf{M}(m\times m';\N)$$
\[rem:batyrev\] Recalling previous Remarks \[rem:reflexive\] and \[rem:anticanonico\], if $\aa=\1$ and $\D_\1$ is a reflexive polytope, then *$f$-duality is Batyrev’s duality* between Fano toric varieties, as defined in [@Batyrev94].
\[rem:negKod\] Hypothesis of strictly effectiveness, given in Definition \[def:ftv\] for a framing $\aa$, is needed to getting good properties of the associated polytope $f$-polytope $\D(X;\aa)$ and, consequently, of the $f$-dual toric variety $\XX_\aa$, as its completeness. Dropping that hypothesis leads to an asymmetric duality, as $\XX_\aa$ can no more be complete: this fact is quite reminiscent of the Givental’s LG mirror model construction [@Givental96]. We will briefly discuss this aspect in §\[ssez:NegKod\], introducing the concept of a *weak framing*, to which the interested reader is referred. In a sense, dropping strictly effectiveness is the key to understanding when looking for a LG mirror model rather than for a complete mirror partner.
### $f$-process as double $f$-duality {#ssez:Deltaprocesso}
By definition, we call *$f$-process the double application of $f$-duality*. This gives rise to a ftv $(\XX_\bb, \cc):=(\XX_\bb, D''_\cc)$ where:
- calling $$\label{rel1}
\D_{\bb}=\{\n\in N_\R\,|\,\L_\aa^T\cdot\n \geq -\bb\}=\{\n\in N_\R\,|\,\forall\,1\leq j\leq m'\quad\langle\n,\ll_j\rangle \geq -b_j\}$$ $\XX_\bb$ is the complete toric variety associated with the fan $\Si_\bb:=\Si_{\D(\XX_\aa,\bb)}$ over the lattice polytope $$\label{k1}
\D(\XX_\aa,\bb):=[k_1\D_\bb]\subseteq N_\R\quad,\quad k_1:=\min\{k\in\N\,|\,\0\in\operatorname{Int}[k\D_\bb]\}$$
- $D''_\cc=\sum_{l=1}^{m''} c_l D''_l$, where $D''_1,\ldots,D''_{m''}$ are the torus invariant prime divisors generating $\Weil(\XX_\bb)$ and $\cc=(c_l)_{l=1}^{m''}$ is the minimum strictly positive column vector such that $$\label{c}
M_{\aa,\bb}^T+C\geq 0\,, \ \text{where}\ M_{\aa,\bb}:=\L^T_\aa\cdot\L_\bb\,,\ C:=\underbrace{\left(\,\cc\ \cdots\ \cc\,\right)}_{m'\ \text{times}}\,\in \mathbf{M}(m''\times m';\N)$$ being $\L_\bb$ the fan matrix of $\XX_\bb$, defined by the primitive generators associated with the vertices of $\D_\bb$, as in Proposition \[prop:Fmatricedipolitopo\].
Let $\NN:=\operatorname{Conv}(V)=\operatorname{Conv}(\{\v_1,\ldots,\v_m\})$ be the lattice polytope associated with the fan matrix $V=(\v_1\ \cdots\ \v_m)$ of $X$. Then $$\label{inclusione}
\NN\subseteq [\D_\bb]\subseteq\D(\XX_\aa,\bb)$$ as, for any column $\v_i$ of $V$, relation (\[b\]) gives that $$\L_\aa^T\cdot\v_i\geq -\bb\ \stackrel{(\ref{rel1})}{\Longrightarrow}\ \v_i\in\D_\bb\cap N\subseteq [\D_\bb]$$
\[def:Deltaproc banale\] A $f$-process $$\label{Deltaproc}
(X,\aa)\stackrel{f-dual}{\rightsquigarrow}(\XX_\aa,\bb)\stackrel{f-dual}{\rightsquigarrow}(\XX_\bb,\cc)$$ is called *calibrated* if there exist $\Xi\in\SF(V)$ and $\Xi'\in\SF(\L_\bb)$, refining $\Si$ and $\Si_\bb$, respectively, and an isomorphism of toric varieties $f:Y(\Xi)\stackrel{\cong}{\longrightarrow}Y'(\Xi')$ such that, calling $\vf:Y(\Xi)\longrightarrow X(\Si)$ and $\vf':Y'(\Xi')\longrightarrow \XX_\bb(\Si_\bb)$ the $\Q$-factorial resolutions associated with the choice of $\Xi$ and $\Xi'$, respectively, one has $$\vf^*D_\aa=(\vf'\circ f)^*D''_\cc$$ In particular, there is an induced birational isomorphism in codimension 1, say $\check{f}:X\dashrightarrow \XX_\bb$, fitting in the following commutative diagram $$\xymatrix{Y\ar[d]^-\vf\ar[r]^-f_-\cong&Y'\ar[d]^-{\vf'}\\
X\ar@{-->}[r]^-{\check{f}}&\XX_\bb}$$
\[rem:ftviso\] Notice that, in the notation of the previous Definition \[def:Deltaproc banale\], *both $(Y,\vf^*D_\aa)$ and $(Y',(\vf')^*D''_\cc)$ are still framed toric varieties*. In fact, the birational transform $\vf^*D_\aa=\sum_i a_i\vf^*D_i$ is still a strictly effective divisor, as $$\Weil(Y)=\bigoplus_{i=1}^{n+r}\Z\cdot\vf^*(D_i)$$ being $\vf$ a *small* birational contraction. Analogously for $(Y',(\vf')^*D''_\cc)$.
Consequently, the condition of being calibrated can be restated by asking that *$f:(Y,\vf^*D_\aa)\stackrel{\cong}{\longrightarrow}(Y',(\vf')^*D''_\cc)$ is a ftv isomorphism*.
\[thm:Deltatriviale\] Let $V=(\v_1\ \cdots\ \v_m)$, $\L_\aa=(\ll_1\ \cdots\ \ll_{m'})$ and $\L_\bb$ be the fan matrices of $X$, $\XX_\aa$ and $\XX_\bb$, respectively, constructed above. Then, up to identifying lattices $M$ (hence $N$) of $X$ and $\XX_\bb$, the $f$-process *(\[Deltaproc\])* is calibrated if and only if $$\begin{aligned}
\label{min}
V&=&\L_\bb\quad\text{(up to a permutation of columns)}\\
\nonumber
\min_{1\leq j\leq m'}\langle\v_i,\ll_j\rangle&=&-a_i\quad\text{(for all $i$ with $1\leq i\leq m$\,)}
\end{aligned}$$ In particular, recalling (\[k0\]) and (\[k1\]), $k_0=1=k_1$, that is, $$\D(X,\aa)=[\D_\aa]\quad\text{and}\quad \D(\XX_\aa,\bb)=[\D_\bb]$$
If (\[Deltaproc\]) is a calibrated $f$-process then there exist $\Xi\in\SF(V)$ and $\Xi'\in\SF(\L_\bb)$, refining $\Si$ and $\Si_\bb$, respectively, and a ftv isomorphism $$f:(Y(\Xi),\vf^*D_\aa)\stackrel{\cong}{\longrightarrow}(Y'(\Xi'),(\vf')^*D''_\cc)$$ as described in Definition \[def:Deltaproc banale\]. In particular, this means that $Y$ and $Y'$ admit equivalent fan matrices, as defined in relation (\[M-equivalenza\]), that is $$\label{A,B}
\exists\,A\in\GL(n,\Z)\,,\ \exists\,B\in\mathfrak{S}_m\leq \GL(m,\Z)\quad \L_\bb=A\cdot V\cdot B$$ Therefore $m''=m$ and inclusion (\[inclusione\]) implies that $\0\in [\D_\bb]$. Hence, $k_1=1$ in (\[k1\]) and $\D(\XX_\aa,\bb)=[\D_\bb]$. Calling $M$ and $M'$ the characters’ lattices of acting tori on $Y$ and $Y'$, respectively, and recalling Proposition \[prop:risoluzioni\], condition (\[A,B\]) comes from the following commutative diagram between associated divisorial short exact sequences $$\label{div-diagram-MM'}
\xymatrix{0\ar[r]& M \ar[r]^-{div}_-{V^T} & \Weil(Y)\ar[d]^-{(f^*)^{-1}}_-{B^T}\ar[r]& \operatorname{Cl}(Y)\ar[d]^-{(\overline{f}^*)^{-1}} \ar[r] & 0\\
0\ar[r]& M'\ar[u]^-{A^T} \ar[r]^-{div}_-{\L_\bb^T} & \Weil(Y')\ar[r] & \operatorname{Cl}(Y') \ar[r] & 0}$$ This actually means that, up to a change of bases in $M$ and $M'$, matrices $A,B$ in (\[A,B\]) and (\[div-diagram-MM’\]) can be chosen as $A=I_n$ and $B=I_m$, so that $\L_\bb=V$. Therefore, via $f,\vf,\vf'$ lattices $N$ and $M$ of $X,Y,Y'$ and $\XX_\bb$ can be identified as above, so giving an identification $$\label{Weil=}
\Weil(X)\cong\Weil(Y)\cong\Weil(Y')\cong\Weil(\XX_\bb)$$ under which, generators $D_i$ are identified with generators $D''_i$, and the ftv isomorphism $f$ gives $\aa=\cc$. Definition (\[c\]) of $\cc$ with $V=\L_\bb$ imply that $$\label{a=c}
\forall\,1\leq i\leq m\quad a_i=c_i=\max\left(\{1\}\cup\{-\langle\v_i,\ll_j\rangle\,|\,1\leq j\leq m'\}\right)$$ Since $\L_\aa$ is a reduced $F$-matrix, Proposition \[prop:Fmatricedipolitopo\] gives that $$\begin{aligned}
\0\in\operatorname{Conv}(\L_\aa)&\Longrightarrow&\0=\sum_{j=1}^{m'} x_j\ll_j\ \text{with $x_j\geq 0$ and $\sum_jx_j=1$} \\
&\Longrightarrow&\forall\,i\quad 0=\sum_jx_j\langle\v_i,\ll_j\rangle\\
&\Longrightarrow&\forall\,i\ \exists\,j\,:\quad \langle\v_i,\ll_j\rangle<0 \\
&\Longrightarrow&\forall\,i\quad \max\left(\{-\langle\v_i,\ll_j\rangle\,|\,1\leq j\leq m'\}\right)\geq 1\\
&\Longrightarrow&\forall\,i\quad -a_i=\min\left(\{\langle\v_i,\ll_j\rangle\,|\,1\leq j\leq m'\}\right)
\end{aligned}$$
For the converse, assume $\L_\bb=V$, up to a permutation of columns. Then $\SF(V)=\SF(\L_\bb)$ and, for any choice $\Xi\in\SF(V)$ there exists $\Xi'\in\SF(\L_\bb)$ and an isomorphism of toric varieties $f:Y(\Xi)\cong Y'(\Xi')$. We can then identify lattices $N$ and $M$ of $Y$ and $Y'$. Moreover, via the $\Q$-factorial small resolutions $\vf:Y\longrightarrow X$, $\vf':Y'\longrightarrow \XX_\bb$, we can also identify lattices $N$ and $M$ of $X,Y,Y'$ and $\XX_\bb$, as above, so getting identifications (\[Weil=\]) for torus invariant Weil divisors. Then, the $f$-process (\[Deltaproc\]) is calibrated if $\aa=\cc$. The latter is guaranteed by the second condition in (\[min\]), as $$\label{c=a}
\forall\,1\leq i\leq m\quad -c_i:=\min_{j}\langle\v_i,\ll_j\rangle=-a_i\ \Longrightarrow\ \cc=\aa$$ Reasoning as above, one then has $k_1=1$.
Moreover, still recalling Proposition \[prop:Fmatricedipolitopo\], relations (\[a=c\]),(\[c=a\]) give $$\0\in\operatorname{Conv}(\L_\aa)\subseteq\{\m\in M_\R\,|\,V^T\cdot \m\geq -\aa\}=\D_\aa\ \Longrightarrow\ k_0=1$$
Assume $\D_\aa$ be a lattice polytope with primitive vertices, that is $$\D_\aa=\operatorname{Conv}(\L_\aa)$$ Then the $f$-process (\[Deltaproc\]) is calibrated if and only if $V=\L_\bb$, up to an identification of lattices $M$ (hence $N$) of $X$ and $\XX_\bb$ and a permutation of columns.
In fact, the second condition in (\[min\]), in Theorem \[thm:Deltatriviale\], is immediately attained, as observed in Remark \[rem:natural\].
By the previous results, a calibrated $f$-process is the key ingredient to introduce an involutive duality between framed toric varieties, largely extending the classical Batyrev duality between Fano toric varieties: in fact the latter can be thought of the particular case of an $f$-duality associated with an ample anti-canonical framing (see the following \[ssez:Bat-dualita\]).
Framing $\P^n$ and associated dual partners {#sez:framingPn}
===========================================
A projective space $\P^n$ is a smooth and complete toric variety associated with the fan matrix $$\label{V}
V=\left(
\begin{array}{ccc}
I_n & | &-\mathbf{1} \\
\end{array}
\right)=\left(
\begin{array}{cccc}
\e_1 & \cdots & \e_n & -\mathbf{1}\\
\end{array}
\right)
\in\mathbf{M}(n,n+1;\Z)$$ and the unique fan $\Si\in\SF(V)$, given by all the faces of the $n+1$, maximal, $n$-dimensional cones, generated by every choice of $n$ columns of $V$. For this complete toric variety, it turns out that *condition (\[min\]) in Theorem \[thm:Deltatriviale\] is satisfied for a sufficiently large number of framing*.
\[thm:dualita\] Let $D_\aa=\sum_{i=0}^{n+1}a_iD_i$ be a strictly effective divisor of $\P^n$. Then $(\P^n,D_\aa)$ is a ftv.
For every $i=1,\ldots,n+1$, define $d_i:=\gcd(\{a_j\,|\,j\neq i\})$ and assume that $$\label{convenzione}
a_1\leq a_2\leq \cdots \leq a_{n+1}\quad\text{and}\quad \gcd(a_1,\ldots,a_{n+1})=1$$ Let $\NN'\subset N_\R\cong\R^n$ be the polytope given by the convex hull of suitable multiples of the standard basis, as follows $$\NN'=\operatorname{Conv}\left(\0,\e_1,{a_n\over a_{n-1}}\e_2,\ldots,{a_n\over a_1}\e_n\right)\quad(\text{notation as in \emph{(\ref{V})}}\,)$$ Then, the $f$-process associated with the ftv $(\P^n,D_\aa)$ is calibrated if and only if the following conditions hold:
- $\operatorname{Conv}(\NN'\cap N)=\operatorname{Conv}\left(\{\0\}\cup\left\{\left[{a_n\over a_{n-i+1}}\right]\e_i\,|\,\forall\,1\leq i\leq n \right\}\right)$,
- $\exists\,i,j\in\{1,\ldots,n+1\}:\ i\neq j\,,\ d_i=d_j=1$.
In this case, the associated $f$-dual ftv is given by $(\XX_\aa,D'_\bb)$ with $$\begin{aligned}
\XX_\aa &\cong& \P(\q)/G_\aa\quad\text{where $\q$ is the reduced weight vector of}\ \aa
\\
D'_\bb &=& \sum_{i=1}^{n+1}b_iD'_i\quad\text{where}\ b_i=\left\{\begin{array}{cc}
a_{n+1}/d_i & \text{for}\ i\leq n \\
a_{n}/ d_{n+1} & \text{for}\ i= n+1
\end{array}\right.
\end{aligned}$$ being $D'_1,\ldots,D'_{n+1}$ the torus invariant prime divisors generating $\Weil(\XX_\aa)$ and $G_\aa$ a finite abelian group of order $$\label{Ga-order}
|G_\aa|=\left(\sum_{i=1}^{n+1}a_i\right)^{n-1}$$ whose action on the weighted projective space $\P(\q)$ is represented by a *torsion matrix* $\Ga$ as follows: by setting $$G_\aa\cong\Z/\tau_1\Z\oplus\cdots\oplus\Z/\tau_s\Z$$ with $\tau_1|\tau_2|\cdots|\tau_s$, the action is given by $$\begin{aligned}
&\xymatrix{\left(\bigoplus_{k=1}^s\Z/\tau_k\Z\right)\times\P(\q)\ar[rrr]^-{\Ga=([\g_{k,j}]_{\tau_k})}&&&\P(\q)}\hskip1.5truecm& \\
&\xymatrix{\left(([\ve_1]_{\tau_1},\ldots,[\ve_s]_{\tau_s}),[x_1:\ldots :x_{n+1}]\right)\ar@{|->}[r]&\left[\left(\prod_{k=1}^s\exp\left({2\pi i \g_{k,j}\ve_k\over\tau_k}\right)\right)x_j\right]_{j=1}^{n+1} }&
\end{aligned}$$ where $$\Ga=\left(
\begin{array}{ccc}
[\g_{1,1}]_{\tau_1} & \cdots & [\g_{1,n+1}]_{\tau_1} \\
\vdots & & \vdots \\
{[}\gamma_{s,1}]_{\tau_s} & \cdots & [\g_{s,n+1}]_{\tau_s} \\
\end{array}
\right)$$ is represented by $(\g_{k,j})\in \mathbf{M}(s,n+1;\Z)$ constructed by means of the next Algorithm \[algoritmoG\].\
In particular, if $\aa$ is a reduced weight vector then $\XX_\aa\cong\P(\aa)/G_\aa$ and $$\bb=\left(
\begin{array}{cccc}
a_{n+1} & \cdots & a_{n+1} & a_n \\
\end{array}
\right)$$
\[algoritmoG\] The torsion matrix $\Ga$, representing the $G_\aa$-action giving $\XX_\aa=\P(\q)/G_\aa$ in the previous Theorem \[thm:dualita\], is defined in display (3) of [@RT-Erratum Thm. 3.2]. Namely:
1. consider a fan matrix $\widetilde{\L}$ of $\P(\q)$ such that $A\cdot\L_\aa=\b\cdot\widetilde{\L}$, with $$A\in\GL_n(\Z)\quad\text{and}\quad\b=\operatorname{diag}\left(\underbrace{1,\ldots,1}_{n-s},\tau_1,\ldots\tau_s\right)$$
2. consider the following matrix $U_\q\in\GL_{n+1}(\Z)$ sending the transposed weight vector $\q^T$ in Hermite normal form (HNF): $$U_\q=\left(
\begin{array}{c}
\uu \\
\widetilde{\L} \\
\end{array}
\right)\ \Longrightarrow\ U_\q\cdot\q^T=\left(
\begin{array}{c}
1 \\
0 \\
\vdots \\
0 \\
\end{array}
\right)$$
3. let $^{n+1-s}U_\q$ be the submatrix of $U_\q$ given by the upper $n+1-s$ rows and consider the matrix $W\in\GL_{n+1}(\Z)$ sending the transposed matrix $^{n+1-s}U_\q^T$ in HNF, that is $$W\cdot\ ^{n+1-s}U_\q^T=\operatorname{HNF}\left(\,^{n+1-s}U_\q^T\right)$$
4. consider the submatrices $_s\widetilde{\L}$ and $_sW$ of $\widetilde{\L}$ and $W$, respectively, assigned by the lower $s$ rows and define the following $s\times s$ integer matrix $$G:=_s\widetilde{\L}\cdot\, _sW^T\in\mathbf{M}(s,s;\Z)$$
5. finally, consider $U_G\in\GL_s(\Z)$ sending the transposed matrix $G^T$ in HNF, that is $U_G\cdot G^T=\operatorname{HNF}(G^T)$, and define $$(\g_{k,i}):=U_G\cdot\,_sW\in \mathbf{M}(s,n+1;\Z)\ \Longrightarrow\ \Ga:=(\g_{k,i})\mod \boldsymbol{\tau}$$
The first part of this proof will describe the $f$-dual ftv $(\XX_\aa,\bb)$ under condition (\[convenzione\]). Then the $f$-process $(\P^n,\aa)\leftrightsquigarrow(\XX_\aa,\bb)$ will be shown to be calibrated if and only if conditions (a) and (b) are satisfied: that is, assuming (\[convenzione\], conditions (a) and (b) are equivalent to conditions (\[min\]) in Theorem \[thm:Deltatriviale\].
$\P^n$ is a smooth and complete toric variety whose Picard group $$\operatorname{Pic}(\P^n)\cong\operatorname{Cl}(\P^n)\cong\Z\cdot h$$ is generated by the hyperplane class $h=[D_1]=\cdots =[D_{n+1}]$, associated with the torus invariant prime divisors generating $\Weil(\P^n)\cong\bigoplus_{i=1}^{n+1}\Z\cdot D_i$. In particular $h$ is a very ample class, so giving that every strictly effective divisor is necessarily very ample, that is, for every ftv $(\P^n,D_{\aa})$, $D_\aa$ ia very ample divisor. Recalling Proposition \[prop:gg\] (2) and relation (\[mI\]) in Proposition \[prop:risoluzioni\], the associated lattice polytope $\D_\aa=\D_{D_\aa}$ is given by $$\begin{aligned}
\label{DeltaaConv}
\nonumber
\D_\aa&=&\operatorname{Conv}\left(\left\{-((V^{\{i\}})^T)^{-1}\cdot\aa^{\{i\}}\,|\,i=1,\ldots,n+1\right\}
\right)\\
&=&\operatorname{Conv}\left(
\begin{array}{c}
|\aa|- a_1 \\
-a_2 \\
-a_3\\
\vdots \\
-a_n \\
\end{array}
\begin{array}{c}
-a_1 \\
|\aa|- a_2\\
-a_3\\
\vdots \\
-a_n \\
\end{array}
\cdots
\begin{array}{c}
-a_1 \\
-a_2\\
\vdots \\
-a_{n-1} \\
|\aa|- a_n\\
\end{array}
\begin{array}{c}
-a_1 \\
-a_2\\
\vdots \\
-a_{n-1} \\
-a_n\\
\end{array}
\right)
\end{aligned}$$ where we set $|\aa|:=\sum_{i=1}^{n+1}a_i$. Then the associated reduced $F$-matrix $\L_\aa$ is $$\label{Lambda_a}
\L_\aa=\left(
\begin{array}{c}
{(|\aa|- a_1)/ d_1} \\
-{a_2/ d_1} \\
\vdots \\
-{a_n/ d_1} \\
\end{array}
\begin{array}{c}
\\
\cdots\\
\cdots\\
\\
\end{array}
\begin{array}{c}
-a_1/ d_n \\
\vdots \\
-a_{n-1}/ d_n \\
(|\aa|- a_n)/ d_n\\
\end{array}
\begin{array}{c}
-a_1/ d_{n+1} \\
-a_2/ d_{n+1}\\
\vdots \\
-a_n/ d_{n+1}\\
\end{array}
\right)$$ so giving $$\label{LamdaV}
\L_\aa^T\cdot V=\left(
\begin{array}{c}
(|\aa|- a_1)/ d_1 \\
-{a_1/ d_2} \\
\vdots \\
-a_{1}/ d_{n}\\
-a_{1}/ d_{n+1} \\
\end{array}
\begin{array}{c}
-a_2/ d_{1} \\
(|\aa|- a_2)/ d_{2}\\
-a_2/d_3\\
\vdots \\
-a_{2}/ d_{n+1} \\
\end{array}
\begin{array}{c}
\cdots\\
\cdots\\
\cdots\\
\cdots \\
\cdots \\
\end{array}
\begin{array}{c}
-a_{n}/ d_{1} \\
\vdots \\
-a_{n}/d_{n-1}\\
(|\aa|- a_n)/ d_{n}\\
-a_{n}/d_{n+1}\\
\end{array}
\begin{array}{c}
-a_{n+1}/ d_{1} \\
-a_{n+1}/d_2\\
\vdots \\
-a_{n+1}/d_n\\
(|\aa|- a_{n+1})/ d_{n+1}\\
\end{array}\right)$$ Recalling that $a_1\leq \cdots \leq a_{n+1}$, there follows $$\bb=\left(
\begin{array}{c}
a_{n+1}/d_1 \\
\vdots\ \\
a_{n+1}/d_n \\
a_n/d_{n+1} \\
\end{array}
\right)$$ Moreover $$\left(
\begin{array}{ccc}
d_1a_1 & \cdots & d_{n+1}a_{n+1} \\
\end{array}
\right)\cdot \L_\aa^T= \0$$ meaning that the reduced weight vector $\q$ of $(d_1a_1,\,\cdots\,,d_{n+1}a_{n+1})$ is a *weight vector* of $\XX_\aa$, in the sense explained in §\[ssez:coni&div\], that is a representative matrix of the class morphism $d$ in the short exact sequence (\[complete deg sequence\]). Hence $\XX_\aa$ is a suitable quotient of the weighted projective space (WPS) $\P(\q)$ by the action of a finite abelian group $G_\aa$. The action of $G_\aa$ on $\P(\q)$ is described by item 6 in [@RT-Erratum Thm. 3.2], so giving items from (2) to (5) in Algorithm \[algoritmoG\]. The isomorphism type of $G_\aa$ can be determined by in item (1) of Algorithm \[algoritmoG\], that is, by looking for a fan matrix $\widetilde{\L}$ of $\P(\q)$ and a switching matrix $\b=\operatorname{diag}\left(\1_{n-s},\tau_1,\ldots\tau_s\right)$, such that $A\cdot\L_\aa=\b\cdot\widetilde{\L}$, for some $A\in\GL_n(\Z)$. Then $$G_\aa\cong\bigoplus_{i=1}^s\Z/\tau_i\Z$$ with $\tau_1|\tau_2|\cdots|\tau_s$. In particular $|G_\aa|=\prod_{i=1}^s\tau_i=\det\b$. Then, to prove (\[Ga-order\]), notice that, on the one hand Binnet theorem gives $$\forall\,i=1,\ldots,n+1\quad \left|\det\left(\L_\aa^{\{i\}}\right)\right|=\left|\det\widetilde{\L}^{\{i\}}\right|\cdot\det\b=q_i\det\b$$ On the other hand, we claim that, under condition (\[convenzione\]), $$\label{pesi}
\forall\,i=1,\ldots,n+1\quad \left|\det\left(\L_\aa^{\{i\}}\right)\right|=q_i\,|\aa|^{n-1}$$ so giving $\det\b=|\aa|^{n-1}$ and then (\[Ga-order\]). In fact, the reduction $\q=(q_1\,\cdots\,q_{n+1})$ is obtained by setting $$q_i:={d_ia_i\over\operatorname{lcm}(\{\d_j\,|\,j\neq i\})}\quad \text{where}\quad \d_j:=\gcd(\{d_ka_k\,|\,k\neq j\})$$ Moreover, condition (\[convenzione\]) implies that $$\label{lcm=prod}
\forall\,i=1,\ldots,n+1\quad \operatorname{lcm}(\{\d_j\,|\,j\neq i\})=\prod_{j=1}^{n+1}d_j$$ Notice that (\[lcm=prod\]) implies (\[pesi\]), as $$\left|\det\left(\L_\aa^{\{i\}}\right)\right|= {a_i|\aa|^{n-1}\over\prod_{j\neq i}d_j}={d_ia_i|\aa|^{n-1}\over\prod_{j=1}^{n+1}d_j}=q_i\,|\aa|^{n-1}$$ To show (\[lcm=prod\]), notice that, for any $i=1,\ldots,n+1$, $$\begin{aligned}
d_i=\gcd(\{a_k\,|\,k\neq i\})\ \Longrightarrow\ \forall\,k\neq i\quad d_i|d_ka_k&\Longrightarrow& d_i|\d_i\\
\forall\,j,k\neq i\quad d_j|d_ka_k\ \text{as}\ \left\{\begin{array}{cc}
d_j|d_j & \text{for $k=j$} \\
d_j|a_k & \text{for $k\neq j$}
\end{array}\right.&\Longrightarrow& d_j|\d_i\\
&\Longrightarrow&\operatorname{lcm}(d_1,\ldots,d_{n+1})| \d_i\end{aligned}$$ Recall that $\gcd(a_1,\ldots,a_{n+1})=1$ implies that $\gcd(d_j,d_k)=1$, for any $j\neq k$ [@RT-wps Prop. 3]. Therefore $\operatorname{lcm}(d_1,\ldots,d_{n+1})=\prod_{j=1}^{n+1}d_j$, so giving that $$\forall\,i=1,\ldots,n+1\quad \prod_{j=1}^{n+1}d_j\,|\,\d_i$$ On the other hand, $\d_i=\gcd(\{d_ka_k\,|\,k\neq i\})$. Then $\d_i|d_ka_k$, for any $k\neq i$. Recall that $\gcd(d_k,a_k)=1$ [@RT-wps Prop. 3]. Hence, for any $i=1,\ldots,n+1$, $$\forall\,k\neq i\quad \d_i|d_ka_k\ \Longrightarrow\ \left\{\begin{array}{cccc}
\exists\,k:& \d_i|d_k&\Longrightarrow&\d_i|\prod_{j=1}^{n+1}d_j\\
\forall\,k\neq i& \d_i|a_k&\Longrightarrow&\d_i|\gcd(\{a_k\,|\,k\neq i\})=d_i\\
&&\Longrightarrow& \d_i |\prod_{j=1}^{n+1}d_j
\end{array}\right.$$ $$\Longrightarrow\hskip1.3truecm \d_i|\prod_{j=1}^{n+1}d_j\hskip2.5truecm$$ In conclusion, $\d_i=\prod_{j=1}^{n+1}d_j$ for any $i=1,\ldots,n+1$. Then (\[lcm=prod\]) immediately follows.
Moreover, notice that $\q$ is also the reduced vector of $\aa$. In fact $$\forall\,i\quad q_i={d_ia_i\over\operatorname{lcm}(\{\d_j\,|\,j\neq i\})}={d_ia_i\over\prod_{j=1}^{n+1}d_j}={a_i\over \prod_{j\neq i}d_j}={a_i\over \operatorname{lcm}(\{d_j\,|\,j\neq i\})}$$ Notice that if $\aa$ is already a reduced weight vector, then $$d_1=\cdots=d_{n+1}=1\ \Longrightarrow\ \q=\aa\ ,\quad \bb= \left(
\begin{array}{c}
a_{n+1} \\
\vdots\ \\
a_{n+1} \\
a_n \\
\end{array}
\right)$$ Therefore: $(\XX_\aa:=\P(\q)/G_\aa,D'_\bb$) is the $f$-dual ftv of $(\P^n,D_\aa)$.
We are now going to considering the $f$-process associated with $(\P^n,D_\aa)$. By the definition of $\D_\bb$ given in (\[rel1\]), and noticing that $\XX_\aa$ is the toric variety associated with the fan $\Si_{\D_\aa}$, which turns out to be the unique one in $\SF(\L_\aa)$, we get $$\label{Delta_b}
\D_\bb=\{\n\in N_\R\,|\,\L_\aa^T\cdot\n \geq -\bb\}=\operatorname{Conv}(\{\n_i\in N_\R\,|\,1\leq i\leq n+1\}$$ where $\n_i=-\left(\left(\L_\aa^{\{i\}}\right)^T\right)^{-1}\cdot\bb^{\{i\}}$, so giving $$\begin{aligned}
\forall\,i=1,\ldots,n\quad\n_i&=&-{a_{n+1}-a_n\over |\aa|}\,\mathbf{1}+{a_{n+1}\over a_i}\left(1-{a_{n+1}-a_n\over |\aa|}\right)\,\e_i \\
\n_{n+1} &=& -\mathbf{1}\\\end{aligned}$$ Notice that, in this expression of $\D_\bb$
- the dependence on $d_1,\ldots,d_{n+1}$ completely disappeared,
- since $a_n\leq a_{n+1}$, it follows that $$\label{disuguaglianze}
0\leq{a_{n+1}-a_n\over|\aa|}< 1\quad\text{and}\quad 0<{a_{n+1}\over a_i}\left(1-{a_{n+1}-a_n\over |\aa|}\right)\leq {a_{n+1}\over a_i}$$
Inequalities in (\[disuguaglianze\]) imply that, for every $i=1,\ldots,n$, the $j$-th entry of $\n_i$ has to satisfy the relations $$\label{componenti}
\forall\,j\neq i\quad -1<n_{j,i}\leq 0\quad\text{and}\quad \left\{\begin{array}{cc}
n_{i,i} =1 & \text{if $a_i=a_{n+1}$} \\
n_{i,i}>1 & \text{if $a_i<a_{n+1}$}
\end{array}\right.$$ where the inequality $n_{i,i}>1$ is obtained as follows: $$n_{i,i}={a_{n+1}\over a_i}-\left(1+{a_{n+1}\over a_i}\right){a_{n+1}-a_n\over |\aa|}={a_{n+1}|\aa|-(a_i+a_{n+1})(a_{n+1}-a_n)\over a_i|\aa|}$$ Therefore, $$n_{i,i}>1\ \Longleftrightarrow\ (a_{n+1}-a_i)|\aa|>(a_i+a_{n+1})(a_{n+1}-a_n)$$ and the latter follows immediately by hypothesis on $\aa$ and the first condition in (\[convenzione\]).
Calling $\NN':=\D_\bb\cap N_{\R}^\geq$, where $N_{\R}^\geq$ represents the positive orthant in the chosen identification $N_{\R}\cong \R^n$, (\[componenti\]) give that $$\NN'=\operatorname{Conv}\left(\0,\e_1,{a_n\over a_{n-1}}\e_2,\ldots,{a_n\over a_1}\e_n\right)$$ as one can check by intersecting the hyperplane passing through $\n_1,\ldots,\n_n$ with coordinate axes. Moreover, $$\D_\bb\cap N=\{-\mathbf{1}\}\cup(\NN'\cap N)$$ and $$\label{[Deltab]}
\D(\XX_\aa,\bb)=[\D_\bb]=\operatorname{Conv}\left(\left\{\e_1,\left[{a_n\over a_{n-1}}\right]\e_2,\ldots,\left[{a_n\over a_{1}}\right]\e_n,-\mathbf{1}\right\}\right)\ \Longleftrightarrow\ \L_\bb=V$$ where the last equality has to be understood up to a possible permutation of columns. This means that the first condition (\[min\]) in Theorem \[thm:Deltatriviale\] is equivalent to condition (a) in the statement. Moreover, recalling expression (\[LamdaV\]) of the transposed matrix of $\L_\bb^T\cdot\L_\aa=V^T\cdot\L_\aa$, the second condition in (\[min\]) can be attained if and only if condition (b) in the statement is assumed, that is, if and only if at least two of $d_i$’s equal 1. Then Theorem \[thm:Deltatriviale\] ensures that the $f$-process associated with $(\P^n,D_\aa)$ is calibrated if and only if conditions (a) and (b) hold.
\[cor:ipersuperfici\] Let $Y_d\subseteq\P^n$ be a projective hypersurface of degree $d\geq n+1$. Then there always exists a framing $D_{\aa_0}$ of $\P^n$ such that $Y_d\sim D_{\aa_0}$ and the $f$-process associated with $(\P^n,D_{\aa_0})$ is calibrated.
It suffices choosing ${\aa_0}=(\underbrace{1,\ldots,1}_{n\ \text{times}},\d:=d-n)=(\1,\d)$. It clearly satisfies conditions (\[convenzione\]) and (b) of Theorem \[thm:dualita\]. Moreover $$\NN'=\operatorname{Conv}\left(\0,\e_1,\ldots,\e_n\right)=\operatorname{Conv}(\NN'\cap N)$$ so giving condition (a), too. Then, thesis follows by theorems \[thm:Deltatriviale\] and \[thm:dualita\].
\[rem:s-ample\] Consider the framing $\aa_0=(\1_n,\d)$ of $\P^n$ introduced in the previous Corollary \[cor:ipersuperfici\]. The dual ftv is then given by $$\left(\XX_{\aa_0},\bb_0\right)=\left(\P(\1_n,\d)/\left(\Z/d\Z\right)^{n-1}\ ,\ (\underbrace{\d,\ldots,\d}_{\text{$n$ times}},1)\right)$$ (notice that $G_{(\1,\d)}\cong\left(\Z/d\Z\right)^{n-1}$, by the following Lemma \[lem:Ga\]). Recalling (\[Delta\_b\]), the polytope $d\D_{\bb_0}$ is a lattice polytope, convex hull of $n+1$ lattice points associated to the maximal cones of the fan $\Si_{\aa_0}$ of $\XX_{\aa_0}$, that is, $\cO_{\XX_{\aa_0}}(d D'_{\bb_0})$ is a globally generated line bundle and $D'_{\bb_0}$ is semi-ample, by Proposition \[prop:gg\].\
Since $\XX_{\aa_0}$ has Picard number 1, this is enough to ensure that $dD'_{\bb_0}$ is an ample divisor of $\XX_{\aa_0}$.\
Moreover, calling $\pi:\P(\1_n,\d)\twoheadrightarrow\XX_{\aa_0}$ the canonical quotient associated with the $\left(\Z/d\Z\right)^{n-1}$-action, the pull-back $\pi^*(D'_{\bb_0})=\sum_{j}b_j\pi^*(D'_j)$ turns out to be the generator of $\operatorname{Pic}(\P(\1_n,\d))\cong\Z$ and a very ample divisor of the universal 1-covering $\P(\1_n,\d)$ of $\XX_{\aa_0}$, as guaranteed by [@RT-wps Prop. 8].
A duality between hypersurfaces in toric varieties {#sez:dualita-hyp}
==================================================
Let us come back to the general setting presented in § \[sez:dualita-ftv\] and consider an hypersurface $Y$ in a complete toric variety $X$. Assume that:
1. there exists a divisor $D_\aa\in\Weil(X)$ such that $Y$ is a generic element in the linear system $|D_\aa|:=d^{-1}\left([D_\aa]\right)$, where $d$ is the class morphism in (\[complete deg sequence\]),
2. $(X,D_\aa)$ is a ftv satisfying conditions (\[min\]) in Theorem \[thm:Deltatriviale\], that is the $f$-process $$(X,\aa)\stackrel{f-dual}{\rightsquigarrow}(\XX_\aa,\bb)\stackrel{f-dual}{\rightsquigarrow}(\XX_\bb,\cc)$$ is calibrated.
\[def:mirror\] A generic element $Y^\vee\in|D'_\bb|:=d^{-1}\left([D'_\bb]\right)$ is called *a $f$-mirror partner of $Y\in|D_\aa|$*.
\[rem:famiglie\] One can explicitly describe the defining polynomials of both $Y$ and $Y^\vee$ in the Cox rings of $X$ and $\XX_\aa$, respectively. Namely:
- the polytope $\D(X,\aa)$ is the Newton polytope of $Y\in|D_\aa|$; call $\overline{\L}_\aa$ a matrix whose columns are given by all the lattice points in $\D(X,\aa)$: it is well defined up to a permutation of columns; setting $l:=|\D(X,\aa)\cap M|$, then $\overline{\L}_\aa$ is a $n\times l$ integer matrix; recalling (\[Ma\]), define $$\overline{M}_\aa:= V^T\cdot \overline{\L}_\aa\quad\text{and}\quad \overline{A}:=\underbrace{\left(\,\aa\ \cdots\ \aa\,\right)}_{l\ \text{times}}\,\in \mathbf{M}(m\times l;\N)\,;$$ then the polynomial of $Y$ is given by $$\label{f-WT}
f=\sum_{j=1}^l c_j\x^{\m_j} \in \operatorname{Cox}(X)\cong\C[x_1,\ldots,x_m]$$ where $\m_j=(m_{i,j})$ is the $j$-th column of $\overline{M}_\aa+\overline{A}$ and $\x^{\m_j}:=\prod_{i=1}^m x_i^{m_{i,j}}$;
- the polytope $\D(\XX_\aa,\bb)$ is the Newton polytope of $Y^\vee\in|D'_\bb|$; call $\overline{\L}_\bb$ a matrix whose columns are given by all the lattice points in $\D(\XX_\aa,\bb)$; setting $l':=|\D(\XX_\aa,\bb)\cap N|$, then $\overline{\L}_\bb$ is a $n\times l'$ integer matrix; define $$\overline{M}_{\aa,\bb}:= \L_\aa^T\cdot \overline{\L}_\bb\quad\text{and}\quad \overline{B}:=\underbrace{\left(\,\bb\ \cdots\ \bb\,\right)}_{l'\ \text{times}}\,\in \mathbf{M}(m'\times l';\N)\,;$$ then the polynomial of $Y^\vee$ is given by $$\label{fdual}
f^\vee=\sum_{j=1}^l c_j\x^{\n_j} \in \operatorname{Cox}(\XX_\aa)\cong\C[x_1,\ldots,x_{m'}]$$ where $\n_j=(n_{i,j})$ is the $j$-th column of $\overline{M}_{\aa,\bb}+\overline{B}$ and $\x^{\n_j}:=\prod_{i=1}^{m'} x_i^{n_{i,j}}$.
Notice that both $f$ and $f^\vee$ are homogeneous polynomials, with respect to degrees induced by class groups. In fact, columns of both $\overline{M}_\aa$ and $\overline{M}_{\aa,\bb}$ determine trivial divisors, up to linear equivalence. Then $$\deg(f)=[D_\aa]\in\operatorname{Cl}(X)\quad\text{and}\quad\deg(f^\vee)=[D'_\bb]\in\operatorname{Cl}(\XX_\aa)$$
\[ex:D\] To fixing ideas, consider the following example, that is a running example throughout the present section. Actually, it is the easiest case of the big class of examples given by projective hypersurfaces of general type, extensively studied in the next §\[sez:ipersuperfici\].
Consider the ftv $(X,\aa)=(\P^2,(1,1,2))$. A fan matrix of $X$ is given by $$V=\left(
\begin{array}{ccc}
1 & 0 & -1 \\
0 & 1 & -1 \\
\end{array}
\right)$$ Consequently, the polytope $\D_\aa =\D_{D_\aa}$ is given by $$\D_\aa = \operatorname{Conv}(\L_\aa)\ ,\quad\text{with}\ \L_\aa=\left(
\begin{array}{ccc}
3 & -1 & -1 \\
-1 & 3 & -1 \\
\end{array}
\right)$$ (see Fig. \[Fig1\]). $\XX_\aa$ is the unique complete and $\Q$-factorial toric variety whose fan matrix is given by $\L_\aa$. It is a quotient of the weighted projective space $\P(\aa)=\P(1,1,2)$ by the action of $\Z/4\Z$ given by sending $$\label{azione2}
\Z/4\Z\times\P(1,1,2)\ni(\overline{\ve},[x_1:x_2:x_3])\mapsto[\mu x_1:x_2:\mu^{-1}x_3]\in\P(1,1,2)$$ being $\mu=\exp(\ve\pi i/2)$ (see also the next Lemma \[lem:Ga\]). As explained in Remark \[rem:famiglie\] $$V^T\cdot \L_\aa+\overline{A} = \left( \begin {array}{ccccccccccccccc} 0&1&0&2&1&0&3&2&1&0&4&3&2&1&0\\
4&3&3&2&2&2&1&1&1&1&0&0&0&0&0\\
0&0&1&0&1&2&0&1&2&3&0&1&2&3&4\end {array} \right)$$ and the family $\mathcal{Y}_\aa$ of plane quartics has general element given by the zero-locus of the polynomial $$\begin{aligned}
f_\aa&=&c_1x_2^4+c_2x_1x_3+c_3x_2^3x_3+ c_4x_1^2x_2^2+c_5x_1x_2^2x_3\\
&&c_6x_2^2x_3^2+c_7x_1^3x_2+c_8x_1^2x_2x_3
+c_9x_1x_2x_3^2+c_{10}x_2x_3^3\\
&&+c_{11}x_1^4+c_{12}x_1^3x_3+c_{13}x_1^2x_3^2
+c_{14}x_1x_3^3+c_{15}x_3^4
\end{aligned}$$ Dually, observing that $$\L_\aa^T\cdot V = \left( \begin {array}{ccc} 3&-1&-2\\
-1&3&-2\\
-1&-1&2\end {array} \right)$$ the framing of $\XX_\aa$ is given by the minimum positive vector $\bb$ such that $$\L_\aa^T\cdot V + \left(
\begin{array}{ccc}
\bb & \bb & \bb \\
\end{array}
\right)\geq \0\ \Longrightarrow\ \bb=\left(
\begin{array}{c}
2 \\
2 \\
1 \\
\end{array}
\right)$$
![\[Fig1\] Example \[ex:D\]: polytopes $\D_\aa\subset M_\R$ and $[\D_\bb]\subseteq\D_\bb\subset N_\R$. ](D.jpg){width="11truecm"}
Then $$\D_\bb=\operatorname{Conv}\left( \begin {array}{ccc} 5/4&-1/4&-1\\
-1/4&5/4&-1\end {array} \right)\ \Longrightarrow\ [\D_\bb]=\operatorname{Conv}(V)$$ so giving that the $f$-process $(X,\aa)\leftrightsquigarrow (\XX_\aa,\bb)$ is calibrated: in fact, hypothesis of Theorem \[thm:dualita\] and Corollary \[cor:ipersuperfici\] are satisfied. Now, observing that $$\L_\aa^T\cdot \overline{V}+\overline{B} = \left( \begin {array}{cccc} 1&5&2&0\\
5&1&2&0\\
0&0&1&3\end {array} \right)$$ there follows that the general element of the dual family $\mathcal{Y}_\bb$ of $\mathcal{Y}_\aa$ is a quotient, by the $\Z/4\Z$-action described in (\[azione2\]), of the zero-locus in $\P(1,1,2)$ of the weighted homogeneous polynomial $$f_\bb=c_1x_1x_2^5+c_2x_1^5x_2+c_3x_1^2x_2^2x_3+c_4x_3^3$$
Generalizing Batyrev’s duality {#ssez:Bat-dualita}
------------------------------
Definition \[def:mirror\] is clearly motivated by the case when $X$ is a Fano toric variety and $\aa=\1$, that is $D_\aa=-K_X$. In fact, recalling Remark \[rem:batyrev\], in this case $f$-duality gives precisely the Batyrev’s polar duality, inducing the well known *mirror symmetry* $Y\leftrightsquigarrow Y^\vee$, being $Y$ and $Y^\vee$ both [Calabi-Yau ]{}varieties, up to suitable *crepant* resolutions of singularities (see the following Definition \[def:crepant\]) .
Topological mirror test and Hodge diamond symmetry {#ssez:mirrortest}
--------------------------------------------------
Let $Y$ be a generic hypersurface in a toric variety $X$ of degree $[D_\aa]\in\operatorname{Cl}(X)$. If $Y$ is quasi-smooth and $X$ is $\Q$-factorial and complete, then there is a well defined concept of (coarse) moduli space $\mathcal{M}_Y$ (see e.g. [@BC §13] and the recent [@Bunnet]). In this case, define $m_Y$ to be the dimension of the tangent space to $\mathcal{M}_{Y}$ at $[Y]$. By [@BC Prop. 13.7] one has $$m_Y=\dim\P\left(H^0(X,\cO_X(D_\aa)\right) -\dim\left(\operatorname{Aut}(X)\right)$$ For a $\Q$-factorial and complete toric variety $X$, $\operatorname{Aut}(X)$ is an affine algebraic group of dimension $$\label{aut}
\dim(\operatorname{Aut}(X))= \dim(X)+\sum_{\Theta} l^*(\Theta)$$ where $\Theta$ ranges on the facets of the anti-canonical polytope $\D_{-K_X}=\D_\1$ [@CoxKatz Prop. 3.6.1 and 3.6.2][^2], [@Cox §4] and $l^*(\Theta)$ denotes *the number of lattice points in the relative interior of the polytope $\Theta$*. Moreover, $$\label{h0}
h^0(X,\cO_X(D_{\aa}))=l(\D_{\aa})$$ so giving $$\label{mY}
m_Y= l(\D_\aa)-1-n-\sum_{\Theta<^1\D_{-K_X}} l^*(\Theta)$$ Unfortunately, conditions (1) and (2) opening the present §\[sez:dualita-hyp\] are not sufficient to guaranteeing quasi-smoothness neither of $Y$ nor of a $f$-mirror $Y^\vee$ of $Y$. Then, in any case, in the following, the *numbers $m_Y$ and $m_{Y^\vee}$ of complex moduli of $Y$ and $Y^\vee$*, respectively, will be combinatorially defined as the right term in (\[mY\]). Namely $$\begin{aligned}
\label{mYY}
\nonumber
m_Y &:=& l(\D_\aa)-1-n-\sum_{\Theta<^1\D_{-K_X}} l^*(\Theta) \\
m_{Y^\vee} &:=& l(\D_\bb) - 1- n- \sum_{\Theta<^1\D_{-K_{\XX_\aa}}} l^*(\Theta)\end{aligned}$$
On the other hand, given a suitable resolution $\widehat{Y}\longrightarrow Y$, it is well defined the *[Kähler ]{}moduli space* of $\widehat{Y}$ as the quotient of the *complexified [Kähler ]{}cone* under the action of the automorphism group $\operatorname{Aut}(\widehat{Y})$ [@CoxKatz §6.2]. Define $$k_{\widehat{Y}}\quad\text{be the dimension of the {K\"{a}hler }moduli space of $\widehat{Y}$}$$ also called *the number of [Kähler ]{}moduli of $\widehat{Y}$*. If $\widehat{Y}$ is a smooth projective hypersurface in a complete toric variety then, bythe weak Lefschetz Theorem, its [Kähler ]{}cone, that is $\operatorname{Nef}(\widehat{Y})$, has dimension given by $h^{1,1}(\widehat{Y})$, and $\operatorname{Aut}(\widehat{Y})$ turns out to acting as a finite group (apply an argument similar to that given in [@CoxKatz §6.2.3], there proposed for a [Calabi-Yau ]{}toric hypersurface), so that $k_{\widehat{Y}}=h^{1,1}(\widehat{Y})$.
\[def:A,B-mirror\] Assume $n=\dim X\geq 4$. Then we will say that:
- the ordered couple $(Y,Y^\vee)$ *satisfies the A-side topological mirror test* if there exists a (partial) resolution of singularities $\widehat{Y}\longrightarrow Y$ such that $\widehat{Y}$ is (quasi-)smooth and $$k_{\widehat{Y}}= m_{Y^\vee}$$ In this case, we will also say that *$Y^\vee$ is an $A$-mirror of $Y$*;
- the ordered couple $(Y,Y^\vee)$ *satisfies the B-side topological mirror test* if there exists a (partial) resolutions of singularities $\widehat{Y}^\vee\longrightarrow Y^\vee$ such that $\widehat{Y}^\vee$ is (quasi-)smooth and $$k_{\widehat{Y}^\vee}=m_Y$$ Then, we will also say that *$Y^\vee$ is a $B$-mirror of $Y$*;
- the ordered couple $(Y,Y^\vee)$ *satisfies the Hodge diamond $A$-symmetry* if $$h^1(\widehat{\Omega}_{\widehat{Y}})=:h^{1,1}(\widehat{Y})=h^{n-2,1}(\widehat{Y}^\vee):=
h^{n-2}(\widehat{\Omega}_{\widehat{Y}})$$ where $\widehat{\Omega}:=i_*(\Omega)$ is the sheaf of Zariski differentials and $i$ is the inclusion of the smooth locus;
- the ordered couple $(Y,Y^\vee)$ *satisfies the Hodge diamond $B$-symmetry* if $$h^{n-2,1}(\widehat{Y})=h^{1,1}(\widehat{Y}^\vee)$$
Moreover, if both $(i)$ and $(ii)$ are satisfied we will say that the $f$-mirror partner $Y^\vee$ of $Y$ is actually a *topological mirror* partner of $Y$ (and viceversa), and if both $(iii)$ and $(iv)$ are satisfied we will say that the $f$-mirror partner $Y^\vee$ of $Y$ is an *Hodge mirror* partner of $Y$ (and viceversa).
Again, the above nomenclature is clearly inspired by the Calabi-Yau/Fano toric case for the toric hypersurface $Y\subset X$. The interested reader is referred to [@CoxKatz §6.1.2] and therein references, for a definition of $m_Y$: see in particular [@CoxKatz Prop. 6.1.3]. Due to the well known Bogomolov-Tian-Todorov-Ran Theorem and the [Calabi-Yau ]{}condition, if $Y$ is a [Calabi-Yau ]{}hypersurface in a Fano toric variety $X$, then $$m_Y=h^1(\widehat{Y},\mathcal{T}_{\widehat{Y}})=h^{2,1}(\widehat{Y})$$ meaning that, in the [Calabi-Yau ]{}case, $(i)\Leftrightarrow (iii)$ and $(ii)\Leftrightarrow (iv)$ and being a topological mirror partner is equivalent to being a Hodge mirror partner.
Mirror Web vs Mirror Symmetry {#ssez:MWeb}
-----------------------------
Let $Y$ be a hypersurface in a toric variety $X$, both satisfying above conditions (1) and (2) opening the present §\[sez:dualita-hyp\]. Notice that the divisor $D_\aa\in\Weil(X)$ satisfying condition (1), may not be unique. Assume there exist two distinct divisors $D_{\aa_1}\sim D_{\aa_2}$ such that $Y\in|D_{\aa_1}|=|D_{\aa_2}|$ and $(X,D_{\aa_i})$ is a ftv, for both $i=1,2$. Then, $f$-duality may assign two distinct mirror partners $Y^\vee_i\in|D'_{\bb_i}|$, $i=1,2$, which, a priori, may be even non-isomorphic: observe that, in general, $D'_{\bb_1},D'_{\bb_2}$ are divisors living in distinct toric varieties $\XX_{\aa_1}$ and $\XX_{\aa_2}$, respectively.
Such a phenomenon does not occur in the Calabi-Yau/Fano toric case, as there is a unique strictly effective divisor in the anti-canonical class of $X$, given by $D_\1\in[-K_X]$. In general, it makes then more sense to speak about a concept of *mirror web $\mathcal{MW}$ of toric hypersurfaces* rather then about *mirror symmetry*. More precisely:
- hypersurfaces connected by means of a calibrated $f$-process give rise to what will be called the *$f$-mirror web* $f\mathcal{MW}$,
- hypersurfaces of dimension $\geq 3$ and connected by means of a calibrated $f$-process originating a topological mirror pair, give rise to the sub-web $T\mathcal{MW}\subset f\mathcal{MW}$,
- hypersurfaces of dimension $\geq 3$ and connected by means of a calibrated $f$-process originating a Hodge mirror pair, give rise to the sub-web $H\mathcal{MW}\subset f\mathcal{MW}$.
For [Calabi-Yau ]{}hypersurfaces $H\mathcal{MW}=T\mathcal{MW}= f\mathcal{MW}$.
\[rem:mult.mirr\] A similar phenomenon of multiple mirror partners is not a new one. As observed by Chiodo and Ruan [@Chiodo-Ruan Rem. 1], examples of multiple mirrors can be easily obtained in the context of Berglund-Hübsch-Kravitz (BHK) duality between [Calabi-Yau ]{}hypersurfaces of (suitable quotients of) weighted projective spaces. But probably, the deepest known example of multiple mirrors is the R[ø]{}dland one [@Rodland], then further studied and generalized by Borisov, Caldărăru and Libgober [@Borisov-C],[@Borisov-L] and Kuznetsov [@Kuznetsov]. Moreover, this fact is well known from the point of view of Homological Mirror Symmetry, where the construction of Landau-Ginzburg (LG) mirror models is not in general expected to producing unique mirror partners (see e.g. considerations following Def. 2.2 in [@KKOY] and the next Example \[ssez:iperellittica\]).
Generalizing Artebani-Comparin-Guilbot (ACG) duality {#ssez:ACG}
----------------------------------------------------
In [@ACG] M. Artebani, P. Comparin and R. Guilbot presented a way of extending Batyrev’s duality of families of anti-canonical hypersurfaces in Fano toric varieties, to suitable sub-families whose associated Newton polytope is *canonical*, that is, a lattice sub-polytope of the anti-canonical polytope admitting the origin as a unique interior point [@ACG §2]. As observed in §\[ssez:Bat-dualita\], $f$-duality is an extension of Batyrev’s duality. Then $f$-duality applies to give an extension of ACG-duality beyond the realm of [Calabi-Yau ]{}hypersurfaces.
Namely, set the following assumptions:
1. let $(X,D_\aa)$ be a ftv admitting a calibrated $f$-process,
2. let $\D$ be a lattice sub-polytope of $\D(X,\aa)$ containing the origin as an interior point: thinking of $\D$ as a Newton polytope, it describes a sub-family $\mathcal{Y}_\D\subseteq\mathcal{Y}_\aa$ of the family of hypersurfaces in $X$ of degree $[D_\aa]\in \operatorname{Cl}(X)$;
3. consider the toric variety $\XX_\D$, which is complete by Proposition \[prop:Fmatricedipolitopo\], and the framing $D_\v$, assigned by the minimum strictly positive column vector $\v$ such that $$V_\D^T\cdot V + \underbrace{\left(\,\v\ \cdots\ \v\,\right)}_{m_\D\ \text{times}}\geq \0$$ where $V$ and $V_\D$ are fan matrices of $X$ and $\XX_\D$, respectively, and $m_\D=\operatorname{rk}\Weil(\XX_\D)$; assume the ftv $(\XX_\D,D_\v)$ admitting a calibrated $f$-process;
4. finally assume that $\D(\XX_\aa,\bb)=[\D_\bb]$ (recall the last assertion in Theorem \[thm:Deltatriviale\]) is a lattice sub-polytope of $\D(\XX_\D,\v)=[\D_\v]$, where $$\D_\v:=\{\n\in N_\R\,|\,V_\D^T\cdot\n\geq -\v\}$$
Then, thinking of $[\D_\bb]$ as a Newton polytope, it describes a sub-family $\mathcal{Y}_\bb\subseteq\mathcal{Y}_\v$ of the family of hypersurfaces in $\XX_\D$ of degree $[D_\v]\in\operatorname{Cl}(\XX_\D)$.
The family $\mathcal{Y}_\bb$ is called a *$f$-ACG dual family* of the family $\mathcal{Y}_\D$.
1. By construction, $\mathcal{Y}_\D$ is a $f$-ACG dual family of $\mathcal{Y}_\bb$.
2. If $\D=\D_\aa$, then $f$-ACG duality reduces to $f$-duality exhibiting $\mathcal{Y}_\bb$ as a $f$-dual family of $\mathcal{Y}_\aa$.
3. If $\aa=\1$, that is $D_\aa=-K_X$, than $\bb=\1$, too, and $(\D,\D_\aa)$ turns out to be a *good pair* in the sense of [@ACG Def.1.4]. In particular, assumptions (1), (3) and (4) follow immediately, and $f$-ACG duality reduces to giving just ACG duality between families $\mathcal{Y}_\D$ and $\mathcal{Y}_\bb$ of [Calabi-Yau ]{}varieties [@ACG Thm. 1].
\[ex:D-ACG\] To better understand the level of generalization introduced by $f$-duality, the present example should be compared with [@ACG Ex. 3.3].
Consider the ftv $(X,\aa)=(\P^2,(1,1,2))$ given in Example \[ex:D\] and notation there introduced. We are then looking for a suitable sub-family of plane quartics admitting a $f$-ACG dual family. Consider the sub-polytope $\D\subseteq\D_\aa$ given by $$\D=\operatorname{Conv}(V_\D)\ ,\quad\text{with}\ V_\D=\left(
\begin{array}{cccc}
1 & -1 & 1 & -1 \\
-1 & 1 & 1 & -1 \\
\end{array}
\right)$$
![\[Fig2\] Example \[ex:D-ACG\]: Newton polytopes $\D\subseteq\D_\aa\subset M_\R$ and $[\D_\bb]~\subseteq~[\D_\v]~\subset~N_\R$. Notice that $\D_\bb\nsubseteq\D_\v$.](D-ACG.jpg){width="12truecm"}
One can then easily check that (see Fig. \[Fig2\]):
- as observed in Example \[ex:D\], a $f$-dual ftv $(\XX_\aa,\bb)$ of $(X,\aa)$ is given by choosing $\bb:=(2,2,1)$, where $\XX_\aa$ is a quotient of the weighted projective space $\P(\aa)=\P(1,1,2)$ by the action of $\Z/4\Z$ described in (\[azione2\]); in particular, the $f$-process $(X,\aa)\leftrightsquigarrow (\XX_\aa,\bb)$ is calibrated, satisfying assumption (1) above;
- clearly the origin of $M$ is an interior point of $\D$, so giving assumption (2);
- recalling Remark \[rem:famiglie\], and observing that $$V^T\cdot \overline{V}_\D+\overline{A} = \left( \begin {array}{ccccccccc} 2&1&0&2&1&0&2&1&0\\
2&2&2&1&1&1&0&0&0\\
0&1&2&1&2&3&2&3&4\end {array} \right)$$ the sub-family $\mathcal{Y}_\D\subseteq\mathcal{Y}_\aa$ of plane quartics has general element given by the zero-locus of the polynomial $$\begin{aligned}
f_\D&=& c_4x_1^2x_2^2+c_5x_1x_2^2x_3+c_6x_2^2x_3^2+c_8x_1^2x_2x_3\\
&&+c_9x_1x_2x_3^2+c_{10}x_2x_3^3+c_{13}x_1^2x_3^2+c_{14}x_1x_3^3+c_{15}x_3^4
\end{aligned}$$
- the toric variety $\XX_\D$, which is the unique complete and $\Q$-factorial one, whose fan matrix is given by $V_\D$, is a quotient of $\P^1\times\P^1$ by the action of $\Z/2Z$ defined by sending $$(\eta,([x_1:x_2],[y_1:y_2])\mapsto ([\nu x_1:\nu^{-1}x_2],[y_1:y_2])\ ,\quad\text{where}\ \nu=\exp(\eta\pi i)$$
- observing that $$V_\D^T\cdot V = \left( \begin {array}{ccc} 1&-1&0\\
-1&1&0\\
1&1&-2\\
-1&-1&2\end {array} \right)$$ the framing $\v$ of $\XX_\D$ is given by the minimum positive vector such that $$V_\D^T\cdot V + \left(
\begin{array}{ccc}
\v & \v & \v \\
\end{array}
\right)\geq \0\ \Longrightarrow\ \v=\left(
\begin{array}{c}
1 \\
1 \\
2 \\
1 \\
\end{array}
\right)$$
- lattice polytopes $[\D_\bb]$ and $[\D_\v]$ are given by $$[\D_\bb]=\operatorname{Conv}(V)\subseteq[\D_\v]=\operatorname{Conv}(\L_\v)\,,\ \text{with}\ \L_\v=\left(
\begin{array}{ccccc}
1 & -1 & 0 & 0 & -1 \\
0 & 0 & 1 & -1 & -1 \\
\end{array}
\right)$$ so guaranteeing assumption (4);
- assumption (3), that is, $(\XX_\D,\v)$ is admitting a calibrated $f$-process, is checked by observing that $$\L_\v^T\cdot V_\D= \left( \begin {array}{cccc} 1&-1&1&-1\\
-1&1&-1&1\\
-1&1&1&-1\\
1&-1&-1&1\\
0&0&-2&2\end {array} \right)\ \Longrightarrow\ \w=\left(
\begin{array}{c}
1 \\
1 \\
1 \\
1 \\
2 \\
\end{array}
\right)$$ and noticing that $$[\D_\w]=[\{\n\in N_\R\,|\,\L_\v^T\cdot \n\geq -\w\}]=\D$$
Therefore, there is a well defined $f$-ACG dual family $\mathcal{Y}_\D^\vee=\mathcal{Y}_{[\D_\bb]}$ of the family $\mathcal{Y}_\D$, described by $[\D_\bb]$ as a Newton polytope of hypersurfaces inside the family $\mathcal{Y}_\v$ of hypersurfaces of degree $[D_\v]\in\operatorname{Cl}(\XX_\D)$. Namely, $$V_\D^T\cdot \overline{V}+\left(
\begin{array}{cccc}
\v & \v & \v & \v \\
\end{array}
\right)
=\left( \begin {array}{cccc} 0&2&1&1\\
2&0&1&1\\
3&3&2&0\\
0&0&1&3\end {array} \right)$$ so giving that the general element of $\mathcal{Y}_{[\D_\bb]}$ is a quotient, by the $\Z/2\Z$-action described above, of the zero-locus of the polynomial $$f_{[\D_\bb]}=c_1x_2^2y_1^3+c_2x_1^2y_1^3+c_3x_1x_2y_1^2y_2+c_4x_1x_2y_2^3$$
Notice that, given assumptions from (1) to (4) above, it is not true, in general, that $\D_\bb$ is a sub-polytope of $\D_\v$. In fact, in the previous Example \[ex:D-ACG\] $$\D_\bb=\operatorname{Conv}\left( \begin {array}{ccc} 5/4&-1/4&-1\\
-1/4&5/4&-1\end {array} \right)\ \nsubseteq\ \D_\v=\left( \begin {array}{cccc} 1&0&-1/2&-3/2\\
0&1&-3/2&-1/2\end {array} \right)$$ (see the right part of Fig. \[Fig2\]).
Generalizing Berglund-Hübsch-Krawitz (BHK) duality {#ssez:BHK}
--------------------------------------------------
In 1993, physicist Berglund and Hübsch [@Berglund-Hubsch] presented a first generalization of the mirror symmetric Greene-Plesser construction [@GP]. Their construction appeared just before the Batyrev’s one [@Batyrev94] and it is, in a sense, “orthogonal” to the latter. The intersection between the two is just the Greene-Plesser example of the quintic threefold. The Berglund-Hübsch construction was later refined by Krawitz [@Krawitz]. For this reason, Artebani, Comparin and Guilbot called this construction the *Berglund-Hübsch-Krawitz (BHK) duality*. In [@ACG §4] they showed how their new ACG-duality generalizes BHK-duality from [Calabi-Yau ]{}hypersurfaces of (a quotient of) a weighted projective space, whose polynomial is of Delsarte type, that is, same number of monomials and variables, to [Calabi-Yau ]{}hypersurfaces of (a quotient of) a $\Q$-Fano toric variety [@ACG §4.2] [^3].
In the previous §\[ssez:ACG\], we introduced the $f$-ACG duality, which is a generalization of ACG-duality to suitable subfamilies of hypersurfaces in toric varieties. Clearly the same approach gives a generalization of BHK-duality, which will be called $f$-BHK duality. Namely,
- if a lattice sub-polytope $\D\subseteq [\D_\aa]$ satisfies assumption from (1) to (4) in §\[ssez:ACG\], then sub-families $\mathcal{Y}_{\D(0)}$ and $\mathcal{Y}_{\D(0)}^\vee=\mathcal{Y}_{[\D_\bb](0)}$, generated by vertices of the lattice polytopes $\D$ and $[\D_\bb]$, respectively, will be called *$f$-BHK dual families*.
\[ex:D-BHK\] Consider the previous Example \[ex:D-ACG\] and the sub-family $\mathcal{Y}_{\D(0)}\subset \mathcal{Y}_{\D}$, whose general element is the zero-locus of the polynomial $$f_{\D(0)}=c_4x_1^2x_2^2+c_6x_2^2x_3^2+c_{13}x_1^2x_3^2+c_{15}x_3^4$$ Then its $f$-BHK dual family is given by the sub-family $\mathcal{Y}_{[\D_\bb](0)}\subset\mathcal{Y}_{[\D_\bb]}$, whose general element is a quotient, by the $\Z/2\Z$-action described above, of the zero-locus of the polynomial $$f_{[\D_\bb(0)]}=c_1x_2^2y_1^3+c_2x_1^2y_1^3+c_4x_1x_2y_2^3$$
In the previous Example \[ex:D-BHK\], both $f_{[\D_\bb](0)}$ and $f_{\D(0)}$ are no more polynomials of Delsarte type. Notice that this fact may also occur in the ACG generalization of BHK-duality (see [@ACG Ex. 4.12]).
Framed duality as a generalized Krawitz duality {#ssez:K-dualità}
-----------------------------------------------
Given a pair of framed toric varieties linked by a calibrated $f$-process $$(X,\aa)\stackrel{\text{$f$-process}}{\leftrightsquigarrow}(\XX_\aa,\bb)$$ generic hypersurfaces $Y\in|D_\aa|$ and $Y^\vee\in|D_\bb|$ may be very singular, making quite difficult finding suitable resolutions $\widehat{Y}$ and $\widehat{Y}^\vee$ and compute all the needed Hodge numbers to check the various instances of mirror symmetry as explained in §\[ssez:mirrortest\]. According with Chiodo and Ruan [@Chiodo-Ruan], it is generally believed that considering suitably associated Landau-Ginzburg (LG) models may sensibly simplify singularities and giving rise to alternative way of checking mirror symmetry.
In the present section, a sort of a *LG/Hypersurface correspondence* is presented, as an extension of the LG/CY correspondence, studied by Chiodo and Ruan [@Chiodo-Ruan], in the case of Delsarte [Calabi-Yau ]{}hypersurfaces, and also by Chiodo, Kalashnikov and Veniani in the recent [@CKV], beyond the [Calabi-Yau ]{}setting. As it will be observed in the next §\[sez:ipersuperfici\], in the case of projective hypersurfaces, the associated LG models turn out to be even smooth. A similar LG/Hypersurface correspondence, translates the mirror duality at a level of LG models. The latter has been described, for hypersurfaces of Delsarte type in weighted projective spaces, by Krawitz [@Krawitz] by means of an extension of Berglund-Hübsch duality without any [Calabi-Yau ]{}condition. ACG extension of BHK-duality and, furthermore, considerations given in the previous §\[ssez:ACG\] and §\[ssez:BHK\], allows us *to think of f-duality in terms of a generalized Krawitz duality*, as steted in the following Proposition \[prop:Krawitz\]. Compare also with the more recent [@HSSW], where He, Si, Shen and Webb give an interesting improvement of Krawitz duality.
### A LG/Hypersurface correspondence {#sssez:LG/Hyp}
Given a ftv $(X,\aa)$ and a generic hypersurface $Y\in|D_\aa|$, let $\T\cong(\C^*)^n$ be the acting torus on $X$. Consider the torus hypersurface $Z:=\T\cap Y$. Recalling Remark \[rem:famiglie\](a), $Y$ is the zero locus of the polynomial $f$ in (\[f-WT\]), generated by the columns of the matrix $\overline{M}_\aa+\overline{A}$. Consider the Laurent polynomial $$f_\aa:={f\over \x^\aa}\in\C[\x,\x^{-1}]$$ generated by the columns of the matrix $\overline{M}_\aa$. Notice that, in $\T$ both $f$ and $f_\aa$ admit the same zero-locus $Z$, that is, $$Z=\T\cap f^{-1}(0)=\T\cap f_\aa^{-1}(0)$$ In particular, $f_\aa$ defines a function $f_\aa:\T\longrightarrow\C$, so giving a LG model $(\T,f_\aa)$.
On the other hand, following Remark \[rem:famiglie\](b), let $\T_\aa\cong(\C^*)^n$ be the acting torus on $\XX_\aa$ and $Z^\vee:=\T_\aa\cap Y^\vee$. Consider the Laurent polynomial $$f^\vee_\bb:={f^\vee\over \x^\bb}\in\C[\x,\x^{-1}]$$ where $f^\vee$ is the polynomial given in (\[fdual\]), generated by the columns of the matrix $\overline{M}_{\aa,\bb}+\overline{B}$. In particular, $f^\vee_\bb$ defines a function $f^\vee_\bb:\T_\aa\longrightarrow\C$, so giving a LG model $(\T_\aa,f^\vee_\bb)$.
\[prop:Krawitz\] If the $f$-process $(X,\aa)\leftrightsquigarrow(\XX_\aa,\bb)$ is calibrated then the Landau-Ginzburg models $(\T,f_\aa)$ and $(\T_\aa,f^\vee_\bb)$ are related by a generalized Krawitz duality (K-duality), that is, $$M_{\aa,\bb}=M_\aa^T$$ This gives rise to the following commutative diagram of LG/Hypersurfaces correspondences and mirror dualities $$\xymatrix{Y\ar@{<~>}[d]_-{\text{LG/Hyp}}\ar@{<~>}[rr]^-{\text{$f$-duality}}&&Y^\vee
\ar@{<~>}[d]^-{\text{LG/Hyp}}\\
(\T,f_\aa)\ar@{<~>}[rr]^-{\text{K-duality}}&&(\T_\aa,f^\vee_\bb)}$$
In fact, by Theorem \[thm:Deltatriviale\], one can assume $\L_\bb=V$, up to a change of generators in lattices $M$ and a permutation of columns. Then $$M_{\aa,\bb}= \L_\aa^T\cdot\L_\bb= \L_\aa^T\cdot V = M_{\aa}^T$$
\[rem:K-dualità\] The previous Proposition \[prop:Krawitz\] leads to an alternative conjectural approach, of checking mirror symmetry for an $f$-mirror pair $(Y,Y^\vee)$, following the lines described in [@Chiodo-Ruan]. Namely, Krawitz established a Mirror Theorem for LG models whose superpotentials are given by *quasi-homogeneous and non-degenerate* Delsarte polynomials, linked by Berglund-Hübsch duality [@Krawitz Thm. 1.1]: the Krawitz mirror map is constructed by means of a bi-graded isomorphism between suitable graded vector spaces associated with the involved superpotentials. See also [@HSSW] for an interesting improvement of this Mirror Theorem for LG models. Then, Chiodo and Ruan proved, under the further [Calabi-Yau ]{}condition, that those graded vector spaces are related with the cohomology of suitable resolutions $\widehat{Y}$ and $\widehat{Y}^\vee$ of $Y$ and $Y^\vee$, respectively [@Chiodo-Ruan Thm. 16, Cor. 17]. Superpotentials involved in the statement of Proposition \[prop:Krawitz\], can be assumed quasi-homogeneous, by considering $f$ and $f^\vee$ rather than $f_\aa$ and $f^\vee_\bb$, respectively. But in general they cannot be assumed neither non-degenerate nor Delsarte, so imposing a deep revision of the Krawitz construction. Moreover, the lack of any CY condition imposes a deeper understanding of relations between Chen-Ruan cohomology and the usual cohomology of $Y$ and $Y^\vee$ (in this sense, consider also [@CKV], for a slight relaxation of the CY condition).
KKP-compactification of associated LG models and log geometry {#ssez:KKP-compct}
-------------------------------------------------------------
Landau-Ginzburg models associated with an $f$-mirror pair $(Y,Y^\vee)$ as in §\[sssez:LG/Hyp\], admit a *compactification* in the sense of Katzarkov-Kontsevich-Pantev [@KKP Def. 2.4], exhibiting a log geometry which is that of a log [Calabi-Yau ]{}defined by Gross and Siebert [@GS-IMS Def. 1.10], where the simple normal crossings divisor $D$ is replaced by the framing of the considered ftv.
Namely, under notation introduced in the previous §\[sssez:LG/Hyp\], $K$-dual superpotential functions $f_\aa:\T\longrightarrow\C$ and $f^\vee_\bb:\T_\aa\longrightarrow\C$ admit the following properifications $$\xymatrix{\T\ar[d]^-{f_\aa}\ar@{^(->}[r]&X\ar[d]^-{\overline{f}_\aa:=[f:\x^\aa]}\\
\C\ar@{^(->}[r]&\P^1}\quad \xymatrix{\ar@{<~>}[rr]<-18pt>^-{\text{$K$-duality}}&&}\quad
\xymatrix{\T_\aa\ar[d]^-{f^\vee_\bb}\ar@{^(->}[r]&\XX_\aa\ar[d]^-{\overline{f}^\vee_\bb:=[f^\vee:\x^\bb]}\\
\C\ar@{^(->}[r]&\P^1}$$ Notice that:
1. $\overline{f}_\aa^{-1}([0:1])=Y\subset X$ and $\overline{f}_\aa^{-1}([1:0])=D_\aa\subset X$
2. $(\overline{f}^\vee_\bb)^{-1}([0:1])=Y^\vee\subset \XX_\aa$ and $(\overline{f}^\vee_\bb)^{-1}([1:0])=D'_\bb\subset \XX_\aa$
3. families $\mathcal{Y}_\aa=\{Y\in|D_\aa|\}$ and $\mathcal{Y^\vee}_\bb=\{Y^\vee\in|D'_\bb|\}$ give rise to corresponding families of LG models $\{(\T,f_\aa)\}$ and $\{(\T_\aa,f^\vee_\bb)\}$, respectively, whose variation turns out to be “anchored at infinity” by their framing; when (suitable resolutions of) $Y$ and $Y^\vee$ are [Calabi-Yau ]{}varieties, then these families of LG models are precisely those considered in [@KKP], meaning that, in this case, their spaces of “anchored” versal deformations are smooth [@KKP Thm. A];
4. recalling the Gross-Siebert definition of a log [Calabi-Yau ]{}pair [@GS-IMS Def. 1.10], one has $$K_X+D_\aa\sim D_{\aa-\1}\quad\text{and}\quad K_{\XX_\aa}+D'_\bb\sim D_{\bb-\1}$$ so giving effective divisors supported on $\bigcup_iD_i$ and $\bigcup_jD'_j$, respectively; by this point of view, framed toric varieties $(X,\aa)$ and $(\XX_\aa,\bb)$ may be understood as *log pairs*, no more [Calabi-Yau ]{}as $D_\aa$ and $D'_{\bb}$ have only normal crossings; this gives an hint about how thinking of the $f$-duality in the context of *Intrinsic Mirror Symmetry* [@GS-IMS].
Mirror partners of hypersurfaces of degree $d\geq n+1$ in $\P^n$ {#sez:ipersuperfici}
================================================================
Degree $d$ hypersurfaces in $\P^n$ are parameterized by the projective space $$\P\left(H^0(\P^n,\cO_{\P^n}(d)\right)$$ The action of $\P\GL(n+1)$ on $\P^n$ extends naturally to an action on the parameter space $\P\left(H^0(\cO_{\P^n}(d)\right)$. Recalling (\[mYY\]), define the *number of complex moduli* of a generic (smooth) hypersurface $Y=Y_d\subset\P^n$, of degree $d$, to be the following one $$\label{m_d^n}
m_d^n:=\dim \P\left(H^0(\cO_{\P^n}(d)\right) - \dim \P\GL(n+1) ={n+d\choose d}-(n+1)^2$$ which is actually the dimension of the moduli space $\mathcal{M}^n_d$ of degree $d$ hypersurfaces in $\P^n$, well defined after Mumford’s GIT [@Mumford], as $\P\GL(n+1)$ is a reductive group.\
On the other hand, if $n\geq 4$, Weak Lefschetz Theorem implies that the *Picard number* of $Y$ is given by $$k_d^n:=h^{1,1}(Y)=b_2(Y)=b_2(\P^n)=1$$ which is also called the *number of [Kähler ]{}moduli* of $Y$, being $k_d^n$ the dimension of the complexified [Kähler ]{}cone of $Y$ [@CoxKatz §6.2].
\[rem:combinatorica\] From the combinatorial point of view, consider the framing of $\P^n$ given in Corollary \[cor:ipersuperfici\], that is $D_{\aa_0}\sim Y_d$ with $${\aa_0}=(\underbrace{1,\ldots,1}_{n\ \text{times}},d-n)$$ Then $$\label{Deltaa}
\D_{\aa_0}=\operatorname{Conv}(\L_{\aa_0})\quad\text{with}\quad \L_{\aa_0}=\left(
\begin{array}{c}
d-1 \\
-1 \\
\vdots \\
-1 \\
\end{array}
\begin{array}{c}
\\
\cdots\\
\cdots\\
\\
\end{array}
\begin{array}{c}
-1 \\
\vdots \\
-1 \\
d-1\\
\end{array}
\begin{array}{c}
-1 \\
-1\\
\vdots \\
-1\\
\end{array}\right)$$ is the Newton polytope associated with the generic polynomial in $H^0(\P^n,\cO_{\P^n}(d))$. Recalling (\[aut\]) and (\[h0\]), one has $$m_d^n= l(\D_{\aa_0})-1-n-\sum_{\Theta<^1\D_\1} l^*(\Theta)$$ as the anti-canonical polytope $\D_{-K_{\P^n}}$ is given by the following sub-polytope of $\D_{\aa_0}$ $$\D_{-K_{\P^n}}=\D_\1=\operatorname{Conv}\left(\begin{array}{c}
d-2 \\
-1 \\
\vdots \\
-1 \\
\end{array}
\begin{array}{c}
\\
\cdots\\
\cdots\\
\\
\end{array}
\begin{array}{c}
-1 \\
\vdots \\
-1 \\
d-2\\
\end{array}
\begin{array}{c}
-1 \\
-1\\
\vdots \\
-1\\
\end{array}\right)$$ On the other hand, by Remark \[rem:reflexive\], $\P^n=\P_{\D_\1}\cong\XX_\NN$ with $\NN=\D_\1^*=\operatorname{Conv}(V)$ and $V$ is the fan matrix given in (\[V\]). Then [@Batyrev94 Prop. 4.4.1] gives $$\label{k}
k_d^n=h^{1,1}(\P^n)=l(\NN)-l^*(\NN)-n = l(\operatorname{Conv}(V))-1-n =1$$
A-side mirroring
----------------
The $f$-dual ftv of $(\P^n,D_{\aa_0})$, as given by Theorem \[thm:dualita\], is $(\XX_{\aa_0},D'_{\bb_0})$ with $$\begin{aligned}
\XX_{\aa_0} &\cong& \P(1,\ldots,1,d-n)/G_{\aa_0}\\
D'_{\bb_0} &=& \sum_{i=1}^{n+1}b_iD'_i\quad\text{where}\ b_i=\left\{\begin{array}{cc}
d-n & \text{for}\ i\leq n \\
1 & \text{for}\ i= n+1
\end{array}\right.
\end{aligned}$$
\[lem:Ga\] $G_{\aa_0}\cong\left(\Z/d\Z\right)^{n-1}$ and its action on $\P({\aa_0})$ can be written as follows $$\begin{aligned}
&\xymatrix{\left(\Z/d\Z\right)^{n-1}\times\P(1,\ldots,1,d-n)\ar[rr]^-{\Ga}&&\P(1,\ldots,1,d-n)}\hskip3.8truecm& \\
&\xymatrix{\left((\overline{\ve}_1,\ldots,\overline{\ve}_{n-1}),[x_1:\ldots :x_{n+1}]\right)\ar@{|->}[r]&\left[\mu_1x_1:\cdots:\mu_{n-1}x_{n-1}:x_{n}:\left(\prod_{j=i}^{n-1}\mu_j\right)^{-1}x_{n+1} \right] }&
\end{aligned}$$ where $\mu_j:=\exp\left({2\pi i\over d}\ve_j\right)$. It can then be represented by the following torsion matrix $$\label{azione}
\Ga=\left(
\begin{array}{ccc}
\overline{I}_{n-1} & \overline{\0}_{n-1} & (d-1)\cdot\overline{\mathbf{1}}_{n-1} \\
\end{array}
\right)\in\M(n-1,n+1;\Z/d\Z)$$
First of all, we need to compute the torsion coefficients $\tau_1|\cdots|\tau_s$. At this purpose we determine a fan matrix $\widetilde{\L}_{\aa_0}$ of the covering wps $\P({\aa_0})$.
Since ${\aa_0}=(1,\ldots,1,d-n)$, we can choose $$\label{L-hat}
\widetilde{\L}_{\aa_0}=\left(
\begin{array}{ccc}
I_{n-1} & -\1_{n-1} & \0_{n-1} \\
\0_{n-1}^T & d-n & -1 \\
\end{array}
\right)\in \M(n, n+1, \Z)$$ as ${\aa_0}\cdot \widetilde{\L}_{\aa_0}^T=\0_n^T$. As a second step we have to determine a matrix $$B\in\GL(n,\Q)\cap\M(n,\Z):\quad B\cdot \widetilde{\L}_{\aa_0}=\L_{\aa_0}$$ where $\L_{\aa_0}$ is the fan matrix of $\XX_{\aa_0}$ presented in (\[Deltaa\]). Such an integer matrix $B$ exists by (see also [@RT-QUOT Rem. 2.4]) and is given by $$\label{betamat}
B=\left(
\begin{array}{ccccc}
d-1&-1 & \cdots & -1 & 1 \\
-1 & d-1 &\cdots & -1 & 1 \\
\vdots & & \ddots & \vdots & \vdots \\
-1 & \ldots & -1 & d-1 & 1 \\
-1 & \ldots &-1 & -1 & 1 \\
\end{array}
\right)$$ Then, torsion coefficients are given by entries different than 1 in the diagonal Smith Form $\b$ of $B$, namely given by $$\label{smith}
\b=A\cdot B\cdot C =\operatorname{diag}(1,\underbrace{d,\ldots,d}_{\text{$n-1$ times}})$$ for suitable matrices $A,C\in\GL_n(\Z)$. That is, $s=n-1$ and $\tau_1=\cdots=\tau_{n-1}=d$. In particular, $G_{\aa_0}\cong (\Z/d\Z)^{n-1}$.\
A torsion matrix $\Ga$ is a representative matrix of the torsion part of the class morphism from $\Weil(\XX_{\aa_0})$ to $\operatorname{Cl}(\XX_{\aa_0})$ and it is characterized by properties from (i) to (iv) in the proof of item (6) in [@RT-Erratum Thm. 3.2], that is:
- $\Ga=(\g_{kj})$ with $\g_{kj}\in \Z/d\Z$,
- $\Ga\cdot(\,^1U_{\aa_0})^T\equiv\0_{n-1} \mod d$, being $U_{\aa_0}\in\GL_{n+1}(\Z)$ a matrix switching the transposed weight vector $\aa_0^T$ in Hermite normal form,
- $\Ga\cdot\L_{\aa_0}^T \equiv\0_{n-1,n} \mod d$,
- $\Ga\cdot(\, _{n-1}(C^{-1}\cdot\widetilde{\L}_{\aa_0}))^T\equiv I_{n-1} \mod d$, where $C$ is given in (\[smith\]), since the bottom $n-1$ rows of $C^{-1}\cdot\widetilde{\L}_{\aa_0}$ are sent by $\Ga$ to a set of generators of $$\operatorname{Tors}(\operatorname{Cl}(\XX_{\aa_0}))\cong(\Z/dZ)^{n-1}$$
Assume $\Ga$ is given as in (\[azione\]). Then (i) and (iii) are clear and (ii) follows by choosing $$\,^1U_{\aa_0}=\left(
\begin{array}{ccccc}
0 & \cdots & 0 & 1 & 0 \\
\end{array}
\right)$$ Finally, condition (iv) is verified up to a basis change in $\Weil(\P(\aa_0))$. In fact, suppressing the $n$-th column from $\widetilde{\L}_{\aa_0}$, by it follows that $$1=\det\left(\L_{\aa_0}^{\{n\}}\right)\ \Longrightarrow\ \L_{\aa_0}^{\{n\}}\in\GL_n(\Z)$$ being 1 the $n$-th entry in $\aa_0$. Then (iv) is satisfied by setting $$C=\L_{\aa_0}^{\{n\}}\cdot\left(
\begin{array}{cc}
\0_{n-1} & 1 \\
I_{n-1} & 0 \\
\end{array}
\right)^{-1}$$
Let $\D_{\bb_0}:=\D_{D'_{\bb_0}}$ be the polytope associated with $D'_{\bb_0}$. Then $\D(\XX_{\aa_0},{\bb_0})=[\D_{\bb_0}]$ is the Newton polytope of the generic section in $H^0(\XX_{\aa_0},\cO_{\XX_{\aa_0}}(D'_{\bb_0}))$, meaning that $$h^0(\XX_{\aa_0},\cO_{\XX_{\aa_0}}(D'_{\bb_0}))=l(\D_{\bb_0})=l(\D(\XX_{\aa_0},{\bb_0}))$$ Moreover, $D'_{\bb_0}$ turns out to be a semi-ample divisor of $\XX_{\aa_0}$, as observed in Remark \[rem:s-ample\].
\[thm:m\*=k\] The family of hypersurfaces $Y^\vee\subseteq\XX_{\aa_0}$, obtained as zero-locus of sections in $H^0(\XX_{\aa_0},\cO_{\XX_{\aa_0}}(D'_{\bb_0}))$, depends on a unique complex modulus, that is, $m_{Y^\vee}=1$. If $n\geq 4$ then $m_{Y^\vee}$ equals the number $k_d^n$ of [Kähler ]{}moduli of projective hypersurfaces $Y_d\subseteq\P^n$ of degree $d$, that is $$m_{Y^\vee}=k_d^n=1$$ By Definition \[def:A,B-mirror\], this means that $Y^\vee$ is an $A$-mirror of $Y$.
Since ${\bb_0}=(d-n,\ldots,d-n,1)$, relation (\[Delta\_b\]) gives that $$\D_{\bb_0}=\operatorname{Conv}\left(\begin{array}{c}
{nd+1-n^2\over d} \\
-{d-n-1\over d} \\
\vdots \\
-{d-n-1\over d} \\
\end{array}
\begin{array}{c}
\\
\cdots\\
\cdots\\
\\
\end{array}
\begin{array}{c}
-{d-n-1\over d} \\
\vdots \\
-{d-n-1\over d} \\
{nd+1-n^2\over d}\\
\end{array}
\begin{array}{c}
-1 \\
-1\\
\vdots \\
-1\\
\end{array}\right)$$ One can then directly check that $$\D(\XX_{\aa_0},{\bb_0})=[\D_{\bb_0}]=\operatorname{Conv}(V)=:\NN$$ as already shown by relation (\[\[Deltab\]\]). Therefore $\0\in\operatorname{Int}([\D_{\bb_0}])$, meaning that $k_1=1$ in the first item of §\[ssez:Deltaprocesso\]. In particular, one gets $$\label{h0}
h^0(\XX_{\aa_0},\cO_{\XX_{\aa_0}}(D'_{\bb_0}))=l(\D_{\bb_0})=l(\NN)=n+2$$ and a generic section $f\in H^0(\XX_{\aa_0},\cO_{\XX_{\aa_0}}(D'_{\bb_0}))$ can be written as follows $$\label{f}
f=\left(\sum_{i=1}^n c_i\, x_i^d\cdot \prod_{j=1}^n x_j^{d-n-1}\right)+c_{n+1}\,x_{n+1}^{n+1}+c_{n+2}\,\left(\prod_{j=1}^n x_j^{d-n}\right)\cdot x_{n+1}$$ in the Cox ring $\C[x_1,\ldots,x_{n+1}]$ of $\XX_{\aa_0}$. Recall now that $\XX_{\aa_0}\cong\P({\aa_0})/G_{\aa_0}$, with ${\aa_0}=(1,\ldots,1,d-n)$, and consider the automorphism of $\P({\aa_0})$ represented by the diagonal matrix $\d=\operatorname{diag}(\g_1,\ldots,\g_n,\g_{n+1})$, where $\g_1,\ldots,\g_{n+1}$ are solutions of the following equations $$\begin{aligned}
\label{automorfismo}
\nonumber
\g_k\hskip1.8truecm=&1&\text{if $c_k=0$, for $1\leq k\leq n+1$} \\
\g_i^d\left(\prod_{j=1}^n\g_j\right)^{d-n-1} =& c_i &\text{for $1\leq i\leq n$ with $c_i\neq 0$}\\
\nonumber
\g_{n+1}^{n+1}\hskip1.6truecm =& c_{n+1} &\text{if $c_{n+1}\neq 0$}\end{aligned}$$ By the previous Lemma \[lem:Ga\], the action of $G_{\aa_0}$ can be assumed diagonal, meaning that $\d$ commutes with such an action, giving rise to an automorphism of $\XX_{\aa_0}$ making $f$ equivalent to the section $$\label{mirror}
f'= \left(\sum_{i=1}^n \epsilon_i\, x_i^d\cdot \prod_{j=1}^n x_j^{d-n-1}\right)+\epsilon_{n+1}\,x_{n+1}^{n+1}+\psi\,\left(\prod_{j=1}^n x_j^{d-n}\right)\cdot x_{n+1}$$ where $\epsilon_k=\left\{\begin{array}{cc}
0 & \text{if $c_k=0$} \\
1 & \text{otherwise}
\end{array}
\right.$.
Then $\psi\in\C$ turns out to be the unique complex modulus of the family of hypersurfaces $Y^\vee\subseteq\XX_{\aa_0}$ of degree $[D'_{\bb_0}]\in\operatorname{Cl}(\XX_{\aa_0})$.
\[rem:moduli\] Notice that the general hypersurface $Y^\vee\in|D'_{\bb_0}|$ is not quasi-smooth. Then results by Batyrev and Cox [@BC] and Bunnet [@Bunnet] cannot be applied to guarantee a good definition of a moduli space $\mathcal{M}_{Y^\vee}$. By the way, observe that the computation performed in Theorem \[thm:m\*=k\] is consistent with the definition of $m_{Y^\vee}$ given in $(\ref{mYY})$. In fact $$\begin{aligned}
\dim\P\left(H^0(\XX_{\aa_0},\cO_{\XX_{\aa_0}}(D'_{\bb_0}))\right)- \dim\operatorname{Aut}(\XX_{\aa_0})&=&l(\D_{\bb_0})-1-n\\
&=& l(\NN)-1-n\,=\,1
\end{aligned}$$ The last equality is obtained by recalling (\[h0\]). The former follows by (\[aut\]), just observing that the anti-canonical polytope $\D_{-K_{\XX_{\aa_0}}}$ is given by the following sub-polytope of $\D_{\bb_0}$ $$\D_{-K_{\XX_{\aa_0}}}=\operatorname{Conv}\left(
\begin{array}{cccc}
\e_1 & \cdots & \e_n & -1/(d-n)\1 \\
\end{array}
\right)$$ whose facets do not contain any lattice point in their relative interior.
According with the Hori-Vafa LG mirror model {#ssez:HoriVafa}
--------------------------------------------
In their pivotal, and unpublished, paper [@Hori-Vafa], Hori and Vafa proposed, from a physical point of view, Landau-Ginzburg (LG) mirror models of hypersurfaces and complete intersections in a complete toric variety. Their construction is consistent with the interpretation of Mirror Symmetry as T-duality. In particular, for the projective hypersurface of degree $d=n+1$ in $\P^{n+1}$, a suitable quotient of their LG mirror model still proposes the mirror construction previously given by Greene and Plesser [@GP], for $n=3$, and Batyrev [@Batyrev94], in the general case.
Namely, for the projective, degree $d$, hypersurface $Y=Y_d\subset\P^n$, the Hori-Vafa recipe proposes the LG mirror model $(\L_{d,\psi},w)$, where (see [@Hori-Vafa 5.4] and notation introduced in [@KKOY]):
- $\L_{d,\psi}\cong(\C^*)^{n+1}$ is the choice of an irreducible component of the reducible torus hypersurface $$\L_{d}:=\left\{\prod_{i=1}^{n+1}x_i^d= \tau y^d\right\}\subset(\C^*)^{n+1}\times\C^*=(\C^*)^{n+2}$$ being $\psi^{-d}=\tau=e^{t}\in\C^*$, with $t$ the [Kähler ]{}volume of $Y_d$,
- $w_{d,\psi}:\L_{d,\psi}\longrightarrow\C$ is the holomorphic function defined by setting $$w_{d,\psi}=w_{d|\L_{d,\psi}}$$ being $w_d:\C^{n+2}\longrightarrow\C$ defined by $$w_d(x_1,\ldots,x_{n+1},y)=\sum_{i=1}^{n+1}x_i^{d}+y\,\Longrightarrow\,
w_{d,\psi}(\x)=\sum_{i=1}^{n+1}x_i^{d}+\psi \prod_{i=1}^{n+1}x_i$$ and called the *superpotential* of the LG model.
When $d=n+1$, the superpotential $w_{n+1,\psi}$ turns out to be equivariant the $\C^*$-action defining $\P^n$ and invariant the action of $G_\1\cong(\Z/(n+1)\Z)^{n-1}$ described in Lemma \[lem:Ga\] and defining $\XX_\1=\P^n/G_\1$, so getting the following picture $$\label{LGquoziente1}
\xymatrix{ \{\0\}\ar@{^(->}[r]&\C&\L_{n+1,\psi}\cong(\C^*)^{n+1}\ar@{^(->}[r]\ar[l]_-{w_{n+1,\psi}}
\ar@{->>}[d]_-{/(\C^*\times G_\1)}&\C^{n+1}\setminus\{\0\}\ar@{->>}[d]_-{/(\C^*\times G_{\1})}\\
w_{n+1,\psi}^{-1}(0)/(\C^*\times G_\1)\ar[u]\ar@{^(->}[rr]&&\T\ar@{^(->}[r]&\XX_\1}$$ where $\T$ is the acting torus on $\XX_\1$. Then the Batyrev’s mirror $Y^\vee$ of $Y_{n+1}$ is precisely the closure $$Y^\vee=\overline{w_{n+1,\psi}^{-1}(0)/(\C^*\times G_\1)}\subset\overline{\T}=\XX_\1$$ induced by the open embedding $\T\hookrightarrow \XX_\1$.
\[rem:LG-d\] Unfortunately, for $d\geq n+2$ the Hori-Vafa LG mirror model does no more admit a similar compactification process, as the superpotential $w_{d,\psi}$ is no more quasi-homogeneous, although we know that *a compact mirror model $Y^\vee_d$ of $Y_d$ should exist*, as defined in Definition \[def:mirror\].
### LG mirror model of the projective hypersurface of degree $d$ {#sssez:HVinvariante}
To bypassing troubles observed in Remark \[rem:LG-d\], replace the Hori-Vafa LG mirror model with the LG model $(\widetilde{\L}_{d,\psi},\widetilde{w}_{d,\psi})$ where
- $\widetilde{\L}_{d,\psi}\cong(\C^*)^{n+1}$ is the choice of an irreducible component of the reducible torus hypersurface $$\widetilde{\L}_{d}:=\left\{x_{n+1}^{n+1}\cdot\prod_{i=1}^{n}x_i^{(n+1)(d-n)}= \tau y^{n+1}\right\}\subset(\C^*)^{n+1}\times\C^*=(\C^*)^{n+2}$$ being $\psi^{-(n+1)}=\tau=e^{t}\in\C^*$, with $t$ the [Kähler ]{}volume of $Y_d$,
- $\widetilde{w}_{d,\psi}:\widetilde{\L}_{d,\psi}\longrightarrow\C$ is the holomorphic function defined by setting $$\widetilde{w}_{d,\psi}=\widetilde{w}_{d|\L_{\psi}}$$ being $\widetilde{w}_d:\C^{n+2}\longrightarrow\C$ defined by $$\begin{aligned}
\widetilde{w}_d(x_1,\ldots,x_{n+1},y)&=&\left(\sum_{i=1}^n x_i^d\cdot \prod_{j=1}^n x_j^{d-n-1}\right)+x_{n+1}^{n+1}+y\\
\Longrightarrow\quad
\widetilde{w}_{d,\psi}(\x)&=&\left(\sum_{i=1}^n x_i^d\cdot \prod_{j=1}^n x_j^{d-n-1}\right)+x_{n+1}^{n+1}+\psi \,x_{n+1}\,\prod_{j=1}^n x_j^{d-n}
\end{aligned}$$
\[conj:LGmirror\] For $d\geq n+1$, a LG mirror model of the projective hypersurface $Y_d\subset\P^n$, of [Kähler ]{}modulus $t=-(n+1)\ln(\psi)$, is given by $((\C^*)^{n+1},\widetilde{w}_{d,\psi})$.
Following Kontsevich [@Kontsevich], proving this Conjecture means showing that the derived categories of coherent sheaves, from the complex point of view, and the Fukaya category of lagrangian structures, from the symplectic point of view, are each other equivalent on the two mirror partners involved (Homological Mirror Symmetry, HMS). This is a quite difficult topic. Here, just some evidences will be provided.
First of all, notice that, when $d=n+1$, the LG model $((\C^*)^{n+1},\widetilde{w}_{n+1,\psi})$ is precisely the Hori-Vafa LG model $((\C^*)^{n+1},w_{n+1,\psi})$.
As a second evidence, consider the fact that, under the weighted $\C^*$-action on $(\C^*)^{n+1}$, given by $$\label{azione4}
\xymatrix{(\l,\x)\ar@{|->}[r]&(\l x_1,\ldots,\l x_n,\l^{d-n}x_{n+1})}$$ the superpotential $\widetilde{w}_{d,\psi}$ is equivariant. In Hori-Vafa notation, this means that there is a *gauged linear sigma model* associated with the LG model $((\C^*)^{n+1},\widetilde{w}_{d,\psi})$, whose gauge action is the weighted one presented in (\[azione4\]). Moreover, $\widetilde{w}_{d,\psi}$ is also equivariant with respect to the action of $G_{\aa_0}\cong (\Z/d\Z)^{n-1}$ described in Lemma \[lem:Ga\] and defining $\XX_{\aa_0}=\P(\aa_0)/G_{\aa_0}$, recalling that $\aa_0=(1,\ldots,1,d-n)$. There is then an analogous picture generalizing (\[LGquoziente1\]) as follows $$\label{LGquoziente}
\xymatrix{ \{\0\}\ar@{^(->}[r]&\C&\widetilde{\L}_{d,\psi}\cong(\C^*)^{n+1}\ar@{^(->}[r]\ar[l]_-{\widetilde{w}_{d,\psi}}
\ar@{->>}[d]_-{/(\C^*\times G_{\aa_0})}&\C^{n+1}\setminus\{\0\}\ar@{->>}[d]_-{/(\C^*\times G_{{\aa_0}})}\\
\widetilde{w}_{d,\psi}^{-1}(0)/(\C^*\times G_{\aa_0})\ar[u]\ar@{^(->}[rr]&&\T\ar@{^(->}[r]&\XX_{\aa_0}}$$ where $\T$ is the acting torus on $\XX_{\aa_0}$. Then the $f$-mirror $Y^\vee$ of $Y_{d}$, as proposed in Definition \[def:mirror\], is precisely the closure $$Y^\vee=\overline{\widetilde{w}_{d,\psi}^{-1}(0)/(\C^*\times G_{\aa_0})}\subset\overline{\T}=\XX_{\aa_0}$$ induced by the open embedding $\T\hookrightarrow \XX_{\aa_0}$.
As a final evidence, notice that the picture described by diagram (\[LGquoziente\]) is strongly related with general LG/Hypersurface correspondence sketched in §\[sssez:LG/Hyp\] and its compactification given in §\[ssez:KKP-compct\]. In a sense, the latter turns out to be the quotient of $((\C^*)^{n+1},\widetilde{w}_{d,\psi})$ by the action of $\C^*\times G_{\aa_0}$ defining $\XX_{\aa_0}$ as a Cox quotient.
\[rem:LGmirrors\] Taking into account what just observed, relating the LG model here presented with those described in §\[sssez:LG/Hyp\] and §\[ssez:KKP-compct\], one could argue that the LG model $(\T_{\aa_0},f^\vee_{\bb_0})$ would be a more appropriated LG mirror model for $Y_d\subset\P^n$ than the one proposed in Conjecture \[conj:LGmirror\]. On the other hand, one may expect that these two LG mirror models turn out to be equivalent by the HMS point of view. These are all completely open tasks, at least as far as the author’s knowledge allows!
A-side of the topological mirror web
------------------------------------
The following result characterizes which framing $D_{\aa}$ of $\P^n$, among those satisfying conditions (\[convenzione\]), (a), (b) in Theorem \[thm:dualita\], give rise to $f$-mirror partners, of the generic $Y_d\subset\P^n$, sharing an A-side mirror behaviour.
\[prop:m\*=k in Pn\] Let $D_{\aa}$ be a framing of $\P^n$ satisfying conditions (\[convenzione\]), (a) and (b) in Theorem \[thm:dualita\] and assume $n\geq 4$. Then, the following facts are equivalent:
1. Theorem \[thm:m\*=k\] holds for the ftv $(\P^n,D_{\aa})$, that is, $$m_{Y^\vee}=k_d^n=1$$ and $Y^\vee$ is an $A$-mirror of $Y$,
2. the number of lattice points in $\D_{\bb}$ equals the number of lattice points in $\NN=\operatorname{Conv}(V)$, i.e. $$l(\D_{\bb})=l(\NN)=n+2$$
3. $\D(\XX_{\aa},\bb):=[\D_{\bb}]=\NN=:\operatorname{Conv}(V)$,
4. $\left[a_n /a_1\right]=1$.
Recall that a fan matrix $\L_{\aa}$ of $\XX_{\aa}$ is given by (\[Lambda\_a\]). Then the anti-canonical polytope $\D_{-K_{\XX_{\aa}}}$ is given by $$\begin{aligned}
\label{-K_a pol}
\nonumber
\D_{-K_{\XX_{\aa}}}&=&\operatorname{Conv}\left(\left\{-\left(\left(\L_{\aa}^{\{i\}}\right)^T\right)^{-1}\cdot\1\,|\,1\leq i\leq n+1\right\}\right)\\
&=&\left(
\begin{array}{ccccc}
1/a_1 & 0 & \ldots & 0&-1/a_{n+1} \\
0&1/a_2&&\vdots&\vdots\\
\vdots & & \ddots & 0&\vdots \\
0 & \cdots & 0 & 1/a_n&-1/a_{n+1} \\
\end{array}
\right)
\end{aligned}$$ In particular, every facet of $\D_{-K_{\XX_{\aa}}}$ does not contain any interior point. Then, recalling (\[mYY\]), one find that $$m_{Y^\vee}= \dim\P\left(H^0(\XX_{\aa},\cO_{\XX_{\aa}}(D'_{\bb}))\right)- \dim\operatorname{Aut}(\XX_{\aa})=l(\D_{\bb})-1-n$$ Then clearly $m_{Y^\vee}=1$ if and only if $l(\D_{\bb})=n+2=l(\NN)$, so giving the equivalence between items (1) and (2) in the statement. Moreover, notice that relation (\[\[Deltab\]\]) gives $$[\D_{\bb}]=\operatorname{Conv}\left(\left\{\e_1,\left[{a_n\over a_{n-1}}\right]\e_2,\ldots,\left[{a_n\over a_{1}}\right]\e_n,-\mathbf{1}\right\}\right)\supseteq \NN$$ Therefore $$l(\D_{\bb})= l(\NN)\ \Longleftrightarrow\ [\D_{\bb}]=\NN$$ so proving the equivalence between items (2) and (3). Finally, notice that, relation (\[\[Deltab\]\]), again, recalling convention (\[convenzione\]), ensures the equivalence between items (3) and (4).
Condition (4) in the statement of Proposition \[prop:m\*=k in Pn\], together with convention (\[convenzione\]), implies that a framing $\aa$ of $\P^n$, satisfying one of the equivalent conditions in Proposition \[prop:m\*=k in Pn\], presents necessarily in the following shape $$\aa=(\underbrace{a,\ldots,a}_{\text{$n$ times}},\d:=d-na)=(a\1_n,\d)\quad\text{with}\quad 1\leq a\leq \d$$ Notice that condition (b) in Theorem \[thm:dualita\] gives $(a,\d)=1$, then $\aa$ is reduced if and only if $\aa=\aa_0=(\1_n,\d)$ and $a=1$. Assume $a\geq 2$: then the reduced weight vector $\q$ of $\aa$ is just given by $\q=(\1_n,\d)$. Therefore $$(\XX_\aa,\bb)=\left(\P(\1_n,\d)/G_\aa\ ,\ (\underbrace{\d,\ldots,\d}_{\text{$n$ times}},1)\right)$$ with $|G_\aa|=d^{n-1}$. Then, the same argument used in Remark \[rem:s-ample\] shows that $dD'_\bb$ is base point free, $D'_\bb$ is semi-ample and $\pi^*(D'_\bb)$ is a very ample divisor generating $\operatorname{Pic}(\P(\1_n,\d))\cong\Z$.
The last condition (4) in Proposition \[prop:m\*=k in Pn\] implies that a framing $D_{\aa}\neq D_{\aa_0}$ can give rise to an $A$-mirror partner of $Y_d\subset\P^n$ only if $d\geq 2n+3$. In particular, recalling considerations given in §\[ssez:MWeb\], for $n=4$, the minimum value of the degree $d$ realizing an effective $A$-mirror web, that is, giving rise to multiple $A$-mirrors, is $d=11$, with the two framing $\aa_0=(1,1,1,1,7)$ and $\aa_1=(2,2,2,2,3)$. Here, the two mirror partners of the generic $Y^4_{11}\subset\P^4$ are given by (a suitable desingularization of) the generic hypersurfaces $$\begin{aligned}
Y^\vee_{35}&=&\left\{\left(\sum_{i=1}^4 x_i^{11}\cdot \prod_{j=1}^4 x_j^{6}\right)+\,x_{5}^{5}+\psi\,\left(\prod_{j=1}^4 x_j^{7}\right)\cdot x_{5}=0\right\}\subset\P(\1_4,7)/(\Z/11\Z)^3 \\
Y^\vee_{15}&=&\left\{\left(\sum_{i=1}^4 x_i^{11}\cdot \prod_{j=1}^4 x_j\right)+x_{5}^{5}+\psi\,\left(\prod_{j=1}^4 x_j^{3}\right)\cdot x_{5}=0\right\}\subset\P(\1_4,3)/(\Z/11\Z)^3
\end{aligned}$$ These two mirror models are not isomorphic as they have singular loci of different dimension: $$\begin{aligned}
\operatorname{Sing}\left(Y^\vee_{35}\right)&=&\bigcup_{i=1}^4\{x_i=x_5=0\}\ \Longrightarrow\ \dim\left(\operatorname{Sing}\left({Y^\vee_{35}}\right)\right)=2 \\
\operatorname{Sing}\left(Y^\vee_{15}\right)&=&\left(\bigcup_{1\leq i < j\leq 4}\{x_i=x_j=x_5=0\}\right)\\
&&\cup\left(\bigcup_{i=1}^4\left\{x_i=x_5=\sum_{\substack{1\leq j\leq 4\\j\neq i}}x_j^{11}=0\right\}\right)\\
&\Longrightarrow&\dim\left(\operatorname{Sing}\left({Y^\vee_{15}}\right)\right)=1
\end{aligned}$$
\[rem:Hodge-A\] By means of methods like those employed by Batyrev and Borisov [@BB96], one can check that, calling $Y^\vee$ the $f$-mirror partner assigned to $Y_d\subset\P^n$ by the choice of the framing $\aa_0=(\1_n,\d)$, and assuming $n\geq 4$ and $d=n+\d\geq n+2$, then $$\label{h21dual}
h^{n-2,1}(Y^\vee)= l^*(2\D_{\bb_0})-l^*(\D_{\bb_0})-n > 1= m_{Y^\vee}$$ where $\bb_0=(\d\cdot\1_n,1)$. For instance, if $n=4$ and $\d=2$ then $h^{n-2,1}(Y^\vee)=5$. Consequently, with that framing, there is no hope of getting any Hodge diamond $A$-symmetry, beyond the [Calabi-Yau ]{}setup.
More in detail, the first equality in (\[h21dual\]) can be obtained by the interplay of the following exact sequences $$\xymatrix{0\ar[r]&\cO_{Y^\vee}(-D'_{\bb_0})\ar[r]&\widehat{\Omega}_{\XX_{\aa_0}|Y^\vee}\ar[r]&
\widehat{\Omega}_{Y^\vee}\ar[r]&0}$$ $$\xymatrix{0\ar[r]&\widehat{\Omega}_{\XX_{\aa_0}}\ar[r]&\bigoplus_{i=1}^{n+1}\cO_{\XX_{\aa_0}}(-D'_i)\ar[r]&
\operatorname{Cl}(\XX_{\aa_0})\otimes\cO_{\XX_{\aa_0}}\ar[r]&0}$$ where $\widehat{\Omega}:=i_*\Omega$ is the sheaf of Zariski differentials, being $i$ the inclusion of the smooth locus in the involved varieties, and recalling that $$h^p(\cO_{Y^\vee}(-D'_{\bb_0}))=\left\{\begin{array}{cc}
l^*(2\D_{\bb_0})-l^*(\D_{\bb_0}) & \text{if}\ p=n-1 \\
0 & \text{otherwise}
\end{array}
\right.$$ (see also [@CLS Thm. 9.2.7] for a detailed proof of the latter). For more details, the interested reader is referred to the incoming paper [@R-fCI], where all these expects will be discussed for projective complete intersections and their $f$-mirror partners, following the lines described in the next §\[sez:CI\].
B-side mirroring
----------------
Assuming $n\geq 4$, the other side of the mirroring process, so called *B-side*, is that of comparing the [Kähler ]{}moduli $k_{\widehat{Y}^\vee}$ with either the complex moduli $m^n_d$, as computed in (\[m\_d\^n\]) (see also Remark \[rem:combinatorica\]) or the Hodge number $h^{n-2,1}(Y_d)$, for a generic hypersurface $Y_d\subset\P^n$ and a generic hypersurface $\widehat{Y}^\vee\in|\widehat{D}'_\bb|$ in $\widehat{\XX}_\aa$, where $$(\P^n,D_\aa)\leftrightsquigarrow(\XX_\aa,D'_\bb)$$ is a calibrated $f$-process, with $Y_d\sim D_\aa$, and $(\widehat{\XX}_\aa,\widehat{D}'_\bb)\longrightarrow(\XX_\aa,D'_\bb)$ is a sufficiently good resolution. The Hodge number $h^{n-2,1}(Y_d)$ can be computed, e.g., by means of the Griffths’ theory on Poincaré residues [@Griffiths]. A comparison with (\[m\_d\^n\]) immediately shows that, for $d\geq n+2$, $$m_d^n={n+d\choose d}-(n+1)^2\neq {2d-1\choose n}-(n+1){d\choose n}=h^{n-2,1}(Y_d)$$ meaning that, also in this case, we cannot hope in a symmetry into the Hodge diamond, as in the A-side for the framing $\aa_0=(\1,d-n)$ (recall Remark \[rem:Hodge-A\]).
### Stringy Hodge numbers of $\Q$-factorial complete toric varieties
Let us now recall the definition of *stringy $E$-function* and *stringy Hodge numbers* in the particular case of a $\Q$-factorial and complete toric variety $X(\Si)$.
Being $X$ $\Q$-factorial, for every maximal cone $\s\in \Si(n)$ there exists a unique subset $I\subset\{1,\ldots,|\Si(1)|\}$ such that $|I|=n$ and $\s=\langle V_I\rangle$, where $V$ is a fan matrix of $X$. Then, since $X$ is complete, there exists a well defined continuous function $\vf_K:N_\R\longrightarrow\R$, called the *canonical support function*, constructed by setting $$\label{KsuppFunct}
\forall\,\n\in N_\R ,\ \forall\,I:\n\in\langle V_I\rangle\quad\vf_K(\n):=-\langle\m_I,\n\rangle\quad\text{where}\quad \m_I:=-(V_I^T)^{-1}\cdot\1_n$$ In particular, $\vf_K$ satisfies the following properties:
1. $\vf_K(\v_i)=1$ for every column $\v_i$ of the fan matrix $V$ of $X$,
2. $\vf_K$ is linear on each cone $\s\in\Si$.
\[def:stringy\] The *stringy $E$-function* of a complete and $\Q$-factorial toric variety $X(\Si)$ is the following $$\label{E-funzione}
\Est(X;u,v):=(uv-1)^{\dim X}\sum_{\s\in\Si}\left(\sum_{\n\in\operatorname{Int}(\s)\cap N}(uv)^{-\vf_K(\n)}\right)$$ where $\operatorname{Int}(\s)$ denotes the relative interior of $\s$. The *stringy Euler number of $X$* is then defined by $$e_{\text{st}}(X):=\lim_{u,v\to 1} \Est(X;u,v)$$
\[rem:gorenstein\] Batyrev introduced the stringy $E$-function and Euler number for a normal, irreducible, algebraic variety $X$ with at worst log-terminal singularities [@Batyrev98 Def. 3.1, Def. 3.3] and proved that $\Est(X;u,v)$ is given by (\[E-funzione\]) when $X$ is a $\Q$-Gorenstein toric variety [@Batyrev98 Thm. 4.3]. Moreover,
- *if $X$ has Gorenstein singularities then $\Est(X;u,v)$ turns out to be a polynomial* [@Batyrev98 Prop. 4.4].
Assume that $\Est(X;u,v)$ is a polynomial with $$\Est(X;u,v)=\sum_{p,q}a_{p,q}u^pv^q$$ Up to a sign, its coefficients are defined to be the *stringy Hodge numbers* of $X$, namely $$\hst^{p,q}(X):=(-1)^{p+q}a_{p,q}$$
\[rem:Poincare\] By Poincaré duality [@Batyrev98 Thm. 3.7, Rem. 3.9], if $\Est(X;u,v)$ is a polynomial then $$\deg \Est(X;u,v)=2\dim X$$ In particular, it turns out that $$\forall\,p,q\quad\hst^{p,q}(X)=\hst^{q,p}(X)\ ,\quad\hst^{0,0}(X)=1=\hst^{\dim X,\dim X}(X)$$
Consider the case of a smooth and complete toric variety $X$. The *$E$-polynomial* of $X$ is defined as: $$E(X;u,v):=\sum_{p,q}(-1)^{p+q}h^{p,q}(X)u^pv^q$$
\[prop:smooth-E\] Let $X$ be a smooth and complete toric variety. Then $\Est(X;u,v)=E(X;u,v)$. In particular, $$\forall\,p,q\quad \hst^{p,q}(X)=h^{p,q}(X)$$
\[def:crepant\] Let $X$ be a projective algebraic variety with at worst canonical Gorenstein singularities. A birational morphism $\phi:Y \longrightarrow X$ is called a *crepant partial resolution* of $X$ if $\phi^*K_X=K_Y$. If $Y$ is smooth than $\phi$ is called a *crepant resolution* of $X$.
\[prop:E-crepant\] Let $\phi:Y\longrightarrow X$ be a crepant partial resolution. Then $\Est(Y;u,v)=\Est(X;u,v)$. In particular, Proposition \[prop:smooth-E\] ensures that, if $\phi$ is a crepant resolution then $$E(Y;u,v)=\Est(X;u,v)\ \Longrightarrow\ \forall\,p,q\quad h^{p,q}(Y)=\hst^{p,q}(X)$$
### Stringy Hodge numbers of $\XX_\aa$
Consider, now, the $\Q$-factorial and complete toric variety $\XX_{\aa}$, associated with the fan $\Si_\aa=\Si_{\D(\P^n,\aa})$ over the polytope $\D(\P^n,\aa)$. Assume that conditions (\[convenzione\]), (a) and (b) in Theorem \[thm:dualita\] are satisfied.
\[prop:n+1\] Recall the fan matrix $\L_{\aa}=\left(
\begin{array}{ccc}
\ll_1 & \cdots & \ll_{n+1} \\
\end{array}
\right)
$ of $\XX_{\aa}$ given in (\[Lambda\_a\]) and the associated polytope $\D_{\aa}=\operatorname{Conv}(\ll_1,\ldots,\ll_{n+1})=\D(\P^n,\aa)$. Then the following are equivalent:
1. $|\aa|=n+1$,
2. $\D_\aa$ is a reflexive polytope,
3. $\XX_\aa$ has Gorenstein singularities.
In particular, by setting $$\begin{aligned}
\psi_\aa(0)&:=&1\\
\forall\,h\in\N\setminus\{0\}\quad\psi_\aa(h)&:=&\left|(h\D_{\aa}\setminus(h-1)\D_\aa)\cap M\right|=l(h\D_\aa)-l((h-1)\D_\aa)
\end{aligned}$$ if one of facts from (1) to (3) holds then $$\label{E-funzione a}
\Est(\XX_{\aa};u,v)= (uv-1)^{n}\sum_{h\geq 0}\psi_\aa(h)(uv)^{-h}$$ is a polynomial of degree $2n$.
(1) $\Rightarrow$ (2): clearly $|\aa|=n+1$ if and only if $\aa=\1_{n+1}$. Then (\[DeltaaConv\]) shows that $\D_\aa$ is a lattice polytope admitting $\0$ as the unique interior lattice point.
(2) $\Rightarrow$ (3): To checking that $\XX_{\aa}$ has only Gorenstein singularities is the same as checking that $$\label{gorenstein}
\forall\,\m\in M\quad \vf_K(\m)\in\Z$$ Given $\m\in M$ let $$\label{h}
h(\m):=\min\{l\in\N\,|\,\m\in l\D_\aa\}$$ Being $\D_\aa$ a reflexive polytope, $\m$ has to belong to the boundary of $h\D_\aa$, otherwise $\m\in(h-1)\D_\aa$, against the definition of $h$. Therefore, $\vf_K(\m)=-h\in\Z$, so giving (\[gorenstein\]).
(3) $\Rightarrow$ (1): by contradiction, assume $|\aa|>n+1$; then $$\aa=\left(
\begin{array}{c}
a_1 \\
\vdots \\
a_n \\
a_{n+1} \\
\end{array}
\right)\geq \left(
\begin{array}{c}
1 \\
\vdots \\
1 \\
2 \\
\end{array}
\right)=:\aa'$$ By (\[DeltaaConv\]), this means that $$\D_\aa\supset \D_{\aa'}=\operatorname{Conv}\left(
\begin{array}{c}
n+1 \\
-1 \\
-1\\
\vdots \\
-1 \\
\end{array} \,
\begin{array}{c}
-1 \\
n+1\\
-1\\
\vdots \\
-1 \\
\end{array}
\,\cdots\,
\begin{array}{c}
-1 \\
-1\\
\vdots \\
-1 \\
n+1\\
\end{array}
\,
\begin{array}{c}
-1 \\
-1\\
\vdots \\
-1 \\
-1\\
\end{array}
\right)$$ Notice that $\e_1,\dots,\e_n$ are all interior lattice points of $\D_{\aa'}$. Then they are also interior lattice point of $\D_\aa$, which cannot be a reflexive polytope.
Finally, (\[E-funzione a\]) follows by (\[E-funzione\]), recalling that $\Si_{\aa}$ is a complete fan in $M_\R$ and noticing that, being $\D_\aa$ a reflexive polytope, $\vf_K(\m)=-h$, where $h$ is defined in (\[h\]). The fact that $\Est(\XX_\aa;u,v)$ is a polynomial of degree $2n$ follows by statement $(*)$ in Remark \[rem:gorenstein\], recalling that $\XX_\aa$ has Gorenstein singularities by item (3), and by Remark \[rem:Poincare\].
Putting together Proposition \[prop:E-crepant\] and Proposition \[prop:n+1\], one gets immediately the following
\[cor:hdgnbr\_n+1\] If $|\aa|=n+1$ then $\Est(\XX_\aa;u,v)=\sum_{h=0}^n c_hu^hv^h$, with $$\begin{aligned}
\forall\,h\in\N:0\leq h\leq n\quad c_h&=&\sum_{i=0}^{n-h}(-1)^i{n \choose n-i}\psi_\aa(n-h-i)\\
&=&\sum_{j=0}^{n-h}(-1)^{n-h-j}{n\choose h+j}\psi_\aa(j)
\end{aligned}$$ In particular, if $\phi:\widehat{\XX}_\aa\longrightarrow\XX_\aa$ is a crepant resolution then $$h^{p,q}(\widehat{\XX}_\aa)=\hst^{p,q}(\XX_\aa)=\left\{\begin{array}{cc}
c_p & \text{when $p=q$} \\
0 & \text{otherwise}
\end{array}
\right.$$
Assume now $\aa=\aa_0:=(\1_{n},d-n)$ with $d=|\aa_0|\geq n+2$. By Proposition \[prop:n+1\], $\XX_{\aa_0}$ has not Gorenstein singularities, but $\Q$-Gorenstein ones. This is clear also by recalling the anti-canonical polytope of $\XX_{\aa_0}$, given by $$\D_{-K_{\XX_{\aa_0}}}=\left(
\begin{array}{cccc}
\e_1&\cdots&\e_n&-1/(d-n)\1_n
\end{array}
\right)$$ This means that $\vf_K$ assumes integer values on $M\cap|\Si_{\aa_0}\setminus\langle\L_{\aa_0}^{\{n+1\}}\rangle|$, while it may assume rational, non integer, values on $$M\cap\langle\L_{\aa_0}^{\{n+1\}}\rangle=M\cap\left\langle
\begin{array}{c}
d-1 \\
-1 \\
\vdots \\
-1 \\
\end{array}\,
\begin{array}{c}
\\
\ldots\\
\\
\\
\end{array}\,
\begin{array}{c}
-1 \\
\vdots \\
-1 \\
d-1\\
\end{array}\right\rangle$$
\[prop:risoluzione\] Consider the $n\times (2n+1)$ integer matrix $$\L'_{\aa_0}:=\left(
\begin{array}{ccccc}
\L_{\aa_0} & | & \e_1 & \cdots & \e_n \\
\end{array}
\right)$$ where $\L_{\aa_0}$ is like in (\[Deltaa\]), and let $\Si'_{\aa_0}\in\SF(\L'_{\aa_0})$ be a subdivision of the fan $\Si_{\aa_0}$ and consider the associated toric variety $\XX'_{\aa_0}(\Si'_{\aa_0})$. Then:
1. $\XX_{\aa_0}'$ has at worst Gorenstein singularities,
2. there exists a birational morphism $f:\XX'_{\aa_0}\longrightarrow\XX_{\aa_0}$ which is a partial resolution of $\XX_{\aa_0}$,
3. there exists a crepant resolution $\phi':\widehat{\XX}_{\aa_0}\longrightarrow\XX'_{\aa_0}$ such that $$\phi=f\circ\phi':\widehat{\XX}_{\aa_0}\longrightarrow\XX_{\aa_0}$$ is a resolution of singularities of $\XX_{\aa_0}$.
In particular, (1) shows that the canonical support function of $\XX'_{\aa_0}$ is a well defined continuous function $\vf_K:M_\R\longrightarrow\R$ such that
- $\vf_K(\ll'_i)=1$ for every column $\ll'_i$ of the fan matrix $\L'_{\aa_0}$ of $\XX_{\aa_0}'$,
- $\vf_K$ is linear on each cone $\s'\in\Si'_{\aa_0}$,
- $\vf_K(\m)\in\Z$ for every $\m\in M$.
Then, by setting $$\begin{aligned}
\forall\,h\in\N\quad\vf_{\aa_0}(h)&:=&\left|\{\m\in M\,|\,\vf_K(\m)=-h\}\right|\\
\forall\,p\in\N:0\leq p\leq n\quad\quad\quad c'_p&:=&\sum_{h=0}^{n-p}(-1)^h{n \choose n-h}\vf_{\aa_0}(n-p-h)\\
&=&\sum_{h=0}^{n-p}(-1)^{n-p-h}{n\choose p+h}\vf_{\aa_0}(h)
\end{aligned}$$ one has $$h^{p,q}(\widehat{\XX}_{\aa_0})=\hst^{p,q}(\XX'_{\aa_0})=\left\{\begin{array}{cc}
c'_p & \text{when $p=q$} \\
0 & \text{otherwise}
\end{array}
\right.$$
Keeping in mind the expression of the anti-canonical polytope $\D_{-K_{\XX_{\aa_0}}}$ and the definition, given in (\[KsuppFunct\]), of the canonical support function $\vf_K$, lattice points $\e_1,\ldots,\e_n\in N$ turn out defining $\vf_K$ on every maximal cone in $\Si'_{\aa_0}(n)$ which is not contained in the $n$-cone $\langle\L_{\aa_0}^{\{n+1\}}\rangle\in\Si_{\aa_0}$. Moreover, the introduction of new rays $\langle\e_1\rangle,\ldots,\langle\e_n\rangle\subset\langle\L_{\aa_0}^{\{n+1\}}\rangle$ determines a subdivision of the $n$-cone $\langle\L_{\aa_0}^{\{n+1\}}\rangle\in\Si(n)$ in $n+1$ simplicial maximal cones of $\Si'_{\aa_0}(n)$. The definition of $\vf_K$ on these $n+1$ cones is then, respectively, assigned by the following $n+1$ vectors $$\left(\begin{array}{c}
-d+n \\
-\1_{n-1}\end{array}\right)
\,,\cdots,\,
\left(\begin{array}{c}
-\1_{n-1} \\
-d+n
\end{array}\right)\,,\,-\1_n \in N$$ Then $\vf_K:M_\R\longrightarrow\R$ is well defined and satisfying properties $(i),(ii),(iii)$ in the statement. This suffices to guarantee that $\XX'_{\aa_0}$ admits at worst Gorenstein singularities, so proving (1). For (2), notice that $\Si'_{\aa_0}$ is a subdivision of $\Si_{\aa_0}$. Then the identity map $\operatorname{id}_{M_\R}$ induces a map of fans $f_\#:\Si'_{\aa_0}\longrightarrow\Si_{\aa_0}$ and then a well defined birational morphism $f:\XX'_{\aa_0}\longrightarrow\XX_{\aa_0}$.
Finally, the crepant resolution $\widehat{\XX}_{\aa_0}$ is obtained by further subdividing $\Si'_{\aa_0}$ by adding all the new rays associated with the $A$-triangulation of $\D_{\aa_0}$, in the sense of [@Batyrev94 Def. 2.2.15], obtained by setting $$\label{A}
A=\{\m\in M\,|\,\vf_K(\m)\leq 1\}$$ Then $\phi'$ is constructed like $f$ and it is a crepant resolution of $\XX'_{\aa_0}$ [@Batyrev94 Thm. 2.2.24], so proving (3).
The last part of the statement, about the computation of Hodge numbers of $\widehat{\XX}_{\aa_0}$, follows immediately by Definition \[def:stringy\] and Propositions \[prop:smooth-E\] and \[prop:E-crepant\].
### [Kähler ]{}moduli of the generic $f$-dual hypersurface $\widehat{Y}^\vee\subset\widehat{\XX}_{\aa_0}$
Consider the calibrated $f$-process $(\P^n,D_{\aa_0})\leftrightsquigarrow(\XX_{\aa_0},D'_{\bb_0})$, with $$\aa_0=(\1_{n},d-n)\quad\text{and}\quad\bb_0=((d-n)\1_{n},1)$$ Let $Y^\vee$ be the generic hypersurface of $\XX_{\aa_0}$ in the linear system $|D'_{\bb_0}|$ whose defining polynomial is $f^\vee\in\operatorname{Cox}(\XX_{\aa_0})$. Define the transformed hypersurface $$\label{trasformata}
\widehat{Y}^\vee:=\phi^{-1}(Y^\vee)$$ as the zero-locus of $\phi^*(f)\in\operatorname{Cox}(\widehat{\XX}_{\aa_0})$ under the resolution $\phi:\widehat{\XX}_{\aa_0}\longrightarrow\XX_{\aa_0}$, constructed in Proposition \[prop:risoluzione\]. Assuming $n\geq 4$, we now, subdivide the computation of $h^{1,1}(\widehat{Y}^\vee)$ in three different cases: $d=n+1$, $d=n+2$ and $d\geq n+3$.
$d=n+1$.Let $Y\subset\P^n$ be a generic hypersurface of degree $d=n+1$. It is a [Calabi-Yau ]{}variety and what follows is a very particular case of the Batyrev duality for anti-canonical hypersurfaces of Fano toric varieties described in [@Batyrev94].
Since $d=n+1$ then $\aa_0=\1_n$ and also $\bb_0=\1_n$.
\[thm:B-mirror0\] $\widehat{Y}^\vee$ is a smooth [Calabi-Yau ]{}hypersurface of $\widehat{\XX}_\1$ with $$h^{1,1}(\widehat{Y}^\vee)=-(n+1){n\choose n-1}+\sum_{i=1}^n (-1)^{n-i}{n+1\choose i+1}{in+i-1\choose n}$$ In particular, $$h^{n-2,1}(Y)= m^n_{n+1}={2n+1\choose n+1}-(n+1)^2=h^{1,1}(\widehat{Y}^\vee)$$ so giving the B-side mirror symmetry between the generic anti-canonical hypersurface $Y\subset \P^n$ and $\widehat{Y}^\vee\subset\XX_\1$.
Recalling Theorem \[thm:m\*=k\], this means that $(Y,\widehat{Y})$ is a pair of topological and Hodge mirror symmetric partners.
Consider the generic hypersurface $Y^\vee\in|D'_\1|$. It is the zero-locus of $$f=\sum_{i=1}^{n+1} c_i\, x_i^{n+1} +c_{n+2}\,\prod_{j=1}^{n+1} x_j$$ as can be immediately deduced by setting $d=n+1$ in (\[f\]). Then $Y^\vee$ is quasi-smooth, as $\widetilde{Y}^\vee=\pi^{-1}(Y^\vee)$ is a smooth hypersurface of $\P^n$, where $\pi:\P^n\twoheadrightarrow\XX_\1$ is the canonical projection of the $G_{\1}$-action described in Lemma \[lem:Ga\]. Therefore $\widehat{Y}^\vee$ is a smooth hypersurface of the resolution $\widehat{\XX}_\1$ of $\XX_\1$. In particular, it is a [Calabi-Yau ]{}hypersurface as $\phi:\widehat{\XX}_\1\longrightarrow\XX_\1$ is a crepant resolution, by Corollary \[cor:hdgnbr\_n+1\]. The same Corollary gives $$\begin{aligned}
h^{1,1}(\widehat{\XX}_\1)=\hst^{1,1}(\XX_\1)=c_1&=&\sum_{j=0}^{n-1}(-1)^{n-1-j}{n\choose j+1}\psi_\1(j)\\
&=&\sum_{j=0}^{n-1}(-1)^{n-1-j}{n \choose j+1} \left[l(j\D_\1)-l((j-1)\D_\1)\right]\\
&=&\sum_{j=0}^{n-1}(-1)^{n-1-j}{n+1\choose j+2}l(j\D_\1)\\
&=&\sum_{i=1}^{n}(-1)^{n-i}{n+1\choose i+1}{in+i-1\choose n}\end{aligned}$$ where, the first passage is the definition of $\psi_\1$ given in Proposition \[prop:n+1\], the second one is obtained by recalling that ${n\choose j+1}+{n\choose j+2}={n+1\choose j+2}$ and the last one follows by setting $i=j+1$ and observing that $$l((i-1)\D_\1)=h^0\left(\cO_{\P^n}((i-1)(n+1))\right)={in+i-1\choose n}$$ Computing $h^{1,1}(\widehat{Y}^\vee)$ is now a consequence of Poincaré Duality [@DK §1.4.(f)] and Lefschetz Theorem [@DK Thm. 3.7] by observing that, being smooth and generic, $\widehat{Y}^\vee$ is a non-degenerate hypersurface of $\widehat{\XX}_\1$, in the sense of Definition \[def:Si-regolare\]: then, in the resolution of $\XX_\1$, a blow up whose exceptional divisor is the closure of the torus orbit of a ray generated by an interior point of a facet of $\D_\1$, does not give any contribution to $h^{1,1}(\widehat{Y}^\vee)$: in fact $\widehat{Y}^\vee$ can be assumed not passing through the center of such a blow up, which is just a point. Consequently $$\begin{aligned}
h^{1,1}(\widehat{Y}^\vee)&=&-\sum_{\Theta<^1\D_\1}l^*(\Theta)+h^{1,1}(\widehat{\XX}_\1)\\
&=&-(n+1){n\choose n-1}+\sum_{i=1}^n (-1)^{n-i}{n+1\choose i+1}{in+i-1\choose n}\end{aligned}$$ Finally, the second part of the statement is just the evaluation for $d=n+1$ of the combinatorial Lemma \[lem:combinatorica\].
$d=n+2$.Let $Y\subset\P^n$ be a generic hypersurface of degree $d=n+2$. Then $Y$ is the lowest degree case of a projective hypersurface of general type. In particular, $\aa_0=(\1_{n},2)$ and $\bb_0=(2\cdot\1_{n},1)$ and let $\phi':\widehat{\XX}_{(\1,2)}\longrightarrow\XX'_{\aa_0}$ be the crepant resolution, constructed in Proposition \[prop:risoluzione\], such that $\phi=f\circ\phi'$ is a resolution of $\XX_{\aa_0}$.
\[thm:B-mirror1\] There exists a partial crepant resolution $\vf:\XX''_{\aa_0}\longrightarrow\XX'_{\aa_0}$, factorizing the crepant resolution $\phi'$ and such that the transformed hypersurface $$Y''=(f\circ\vf)^{-1}(Y^\vee)\subset\XX''_{(\1,2)}$$ is quasi-smooth and $$h^{1,1}(Y'')=-{n+1\choose n-1}+c'_1=-{n+1\choose n-1}+\sum_{h=0}^{n-1}
(-1)^{n-1-h}{n\choose h+1}\vf_{(\1,2)}(h)$$ where $\vf_{(\1,2)}(h)$ admits the following recursive expression $$\begin{aligned}
\vf_{(\1,2)}(0)&=&1\\
\forall\,h\in\N\setminus\{0\}\quad\vf_{(\1,2)}(h)&=&{h(n+2)+n\choose n}\\
&&-\sum_{j=0}^{h-1}\left[{(j+1)(n+1)+h-1\choose n-1}+\vf_{(\1,2)}(j)\right]
\end{aligned}$$ In particular, $$\label{B per n+2}
m^n_{n+2}={2n+2\choose n}-(n+1)^2=h^{1,1}(Y'')$$ that is, recalling Definition \[def:A,B-mirror\], the generic $Y^\vee\subset\XX_{(\1,2)}$ is a $B$-mirror partner of the generic hypersurface $Y\subset \P^n$ of degree $n+2$. Recalling Theorem \[thm:m\*=k\], this means that $(Y,Y^\vee)$ is a pair of topologically mirror partners.
Consider the generic hypersurface $Y^\vee\in|D'_{(2\cdot\1,1)}|$ and its transformed hypersurface $\widehat{Y}^\vee=\phi^{-1}(Y^\vee)\subset\widehat{\XX}_{(\1,2)}$ as defined in (\[trasformata\]). Local analysis explained in the next §\[ssez:locale\] suffices to show that $\widehat{Y}^\vee$ is smooth. To compute $h^{1,1}(\widehat{Y}^\vee)$ we still use Poincaré Duality [@DK §1.4.(f)] and Lefschetz Theorem [@DK Thm. 3.7] as in the previous case $d=n+1$, but we need to be more careful in determining which blow up in the resolution $\phi:\widehat{\XX}_{(\1,2)}\longrightarrow\XX_{(\1,2)}$ does not contribute to $h^{1,1}(\widehat{Y}^\vee)$. In fact, now, $Y^\vee$ is a degenerate hypersurface of $\XX_{(\1,2)}$. Nevertheless, singularities of $Y^\vee$ are all contained in the prime toric divisor $D'_{n+1}\subset\XX_{(\1,2)}$, that is, $Y':=Y^\vee\setminus D'_{n+1}$ is a smooth hypersurface of the Zariski open subset $\XX':=\XX_{(\1,2)}\setminus D'_{n+1}$ of $\XX_{(\1,2)}$. Notice that $\XX'$ is the toric variety associated with the sub-fan $\Si'\subset\Si_{(\1,2)}$ defined by the maximal cone $\s_{n+1}=\langle\ll_1,\ldots,\ll_n\rangle$ and all its faces. $\s_{n+1}$ is the cone of $\Si_{(\1,2)}$ generated by the facet $$\Theta_{n+1}=\operatorname{Conv}(\ll_1,\ldots,\ll_n)$$ of the polytope $\D_{(\1,2)}$, opposite to the vertex $\ll_{n+1}$. In particular, recalling Definition \[def:Si-regolare\], $Y'$ turns out to be a $\Si'$-regular hypersurface of $\XX'$, meaning that it has empty intersection with the torus orbit of the whole cone $\s_{n+1}$. That is, the following stratum is empty $$Y^\vee_{\s_{n+1}}=Y'_{\s_{n+1}}=Y'\cap\T\cdot x_{\s_{n+1}}=Y'\cap\{[0:\cdots:0:1]\}=\emptyset$$ as can also be immediately deduced from the polynomial $f'$ of the generic $Y^\vee$. In other word, recalling the definition of $\phi$ by means of the $A$-triangulation defined in (\[A\]), $Y'$ can be assumed not passing through the points of $\XX'$ which are centers of those blow ups determined by lattice points in $A\cap\s_{n+1}^\circ$, where $\s_{n+1}^\circ$ denotes the relative interior of $\s_{n+1}$. The same assumption holds for $Y^\vee$, too. On the contrary, recalling (\[f\]) and setting $\s_i=\langle \ll_j\,|\,j\neq i\rangle\in\Si_{(\1,2)}(n)$, $$\forall\,1\leq i\leq n\quad [0:\cdots:\underset{i}{1}:\cdots:0]\in Y^\vee\ \Longrightarrow\ Y^\vee_{\s_i}=\T\cdot x_{\s_i}$$ meaning that $Y^\vee$ passes through every center of those blow ups composing $\phi$ and determined by interior lattice points of the facet $\Theta_i=\operatorname{Conv}(\ll_j\,|\,j\neq i)$, for $1\leq i \leq n$. Notice that the desingularization process described in §\[ssez:locale\] requires the choice of $n$ lattice points among the $$(n-1)(d-1)=n^2-1$$ lattice points in $\bigcup_{i=1}^{n-1}\operatorname{Relint}(\operatorname{Conv}(\ll_i,\ll_{n+1}))$. Then one can stop the resolution process of $\XX'_{\aa_0}$ avoiding the blowups associated with $\binom{d-1}{n-1}-n=(n^2-n)/2$ of lattice points in $\bigcup_{i=1}^{n-1}\operatorname{Relint}(\operatorname{Conv}(\ll_i,\ll_{n+1}))$, so getting the following factorization of $\phi'$: $$\xymatrix{\widehat{\XX}_{\aa_0}\ar[rd]_-{(n^2-n)/2\ \text{blowups}}^-{\phi''}\ar[rr]^-{\phi'}&&\XX'_{\aa_0}\ar[r]^-f&\XX_{\aa_0}\\
&\XX''_{\aa_0}\ar[ur]^-{\vf}&&}$$ Then $Y''=(f\circ\vf)^{-1}(Y^\vee)$ is quasi-smooth as $$n^2-1- n \ge {n^2-n\over 2}$$ by the dimensional assumption $n\ge 4>2$. Moreover, recalling Proposition \[prop:risoluzione\], $$h^{1,1}(\widehat{Y}^\vee)= h^{1,1}(\widehat{\XX}_{(\1,2)})-|A\cap\s_{n+1}^\circ| - {n^2-n\over 2}= c'_1 - \binom{n+1}{n-1}$$ which is given by (\[B per n+2\]) as a consequence of the following combinatorial Lemma \[lem:h11st\].
*$d=n+\d$ with $\d\geq 3$*.Let $Y\subset\P^n$ be a generic hypersurface of degree $d=n+\d\geq n+3$. Then $\aa_0=(\1_{n},\d)$ and $\bb_0=(\d\cdot\1_{n},1)$ and let $\widehat{\XX}_{(\1,\d)}$ be the resolution of $\XX_{(\1,\d)}$ constructed in Proposition \[prop:risoluzione\].
\[thm:B-mirror2\] There exists a smooth hypersurface $\widehat{Y}^\vee\subset\widehat{\XX}_{(\1,\d)}$ giving a resolution of $Y^\vee\subset\XX_{(\1,\d)}$ such that $$h^{1,1}\left(\widehat{Y}^\vee\right)=\binom{d+n}{n}-\binom{d}{n}-n<m_d^n$$ Moreover, there exists a birational morphism $f':\widehat{\XX}'_{(\1,\d)}\longrightarrow\widehat{\XX}_{(\1,\d)}$, which is a composition of divisorial blowups, such that, calling $\vf:=\phi\circ f'$, the associated transformed hypersurface $(\widehat{Y}^\vee)':=\vf^{-1}(Y^\vee)=(f')^{-1}(\widehat{Y}^\vee)$ of $Y^\vee$ is smooth and $$h^{1,1}\left((\widehat{Y}^\vee)'\right)= h^{1,1}(\widehat{\XX}'_{(\1,\d)})-n= m^n_{d}$$ that is, recalling Definition \[def:A,B-mirror\], the generic $Y^\vee\subset\XX_{(\1,\d)}$ is a $B$-mirror partner of the generic hypersurface $Y\subset \P^n$ of degree $d=n+\d$. Recalling Theorem \[thm:m\*=k\], this means that $(Y,Y^\vee)$ is a pair of topologically mirror partners.
Consider the generic hypersurface $Y^\vee\in|D'_{(\d\cdot\1,1)}|$ and its transformed hypersurface $\widehat{Y}^\vee=\phi^{-1}_*(Y^\vee)\subset\widehat{\XX}_{(\1,\d)}$, as defined in (\[trasformata\]). Local analysis explained in the next §\[ssez:locale\] suffices to show that $\widehat{Y}^\vee$ is smooth. The computation of $h^{1,1}(\widehat{Y}^\vee)$ proceeds exactly as in the proof of the previous Theorem \[thm:B-mirror1\], recalling the definition of the $A$-triangulation of $\D_{\aa_0}$ given in (\[A\]). Then, the following Lemma \[lem:h11st\] gives $$h^{1,1}\left(\widehat{Y}^\vee\right)= \hst^{1,1}\left(\XX'_{(\1,\d)}\right)- |A\cap\s_{n+1}^\circ|= \binom{d+n}{n}-\binom{d}{n}-n$$ Since $\d\ge 3$, one has $$\binom{d}{n}+n=\binom{n+\d}{n}+n>(n+1)^2$$ so giving that $$h^{1,1}\left(\widehat{Y}^\vee\right)<\binom{d+n}{n} - (n+1)^2=m^n_d$$ Proceed, now, to blowing up $\widehat{\XX}_{(\1,\d)}$ in $s:=\binom{d}{n}+n-(n+1)^2$ points belonging to $\widehat{Y}$, to getting the birational morphism $$\xymatrix{f':\widehat{\XX}'_{(\1,\d)}:=B_s\widehat{\XX}_{\1,\d}\ar[r]&\widehat{\XX}_{\1,\d}}$$ Then, finally $$\begin{aligned}
h^{1,1}\left((\widehat{Y}^\vee)'\right)&=& h^{1,1}\left(\widehat{\XX}'_{(\1,\d)}\right) - n= s+\hst^{1,1}\left(\XX'_{(\1,\d)})\right) - n\\
&=& \binom{d+n}{n} - (n+1)^2 =m^n_d
\end{aligned}$$
Local analysis of singularities and resolutions {#ssez:locale}
-----------------------------------------------
The present paragraph is devoted to giving a proof of the existence of a smooth resolution $\widehat{Y}^\vee\subset\widehat{\XX}_{\aa_0}$ of a generic $Y^\vee\in|D'_{\bb_0}|$, as claimed in previous Theorems \[thm:B-mirror1\] and \[thm:B-mirror2\], being $\phi:\widehat{\XX}_{\aa_0}\longrightarrow \XX_{\aa_0}$ the resolution given in Proposition \[prop:risoluzione\](3).
Let $Y$ be a generic hypersurface in $\P^n$ of degree $d=n+\d$, with $\d\geq 2$. Then, after acting an automorphism of $\XX_{\aa_0}$, as in (\[automorfismo\]), the generic hypersurface $Y^\vee\in|D'_{\aa_0}|$ is defined by the following polynomial in $\operatorname{Cox}(\XX_{\aa_0})\cong\C[\x]$ $$f=\prod_{i=1}^n x_i^{\d-1}\left(\sum_{i=1}^n x_i^d +\psi\,\prod_{j=1}^{n+1} x_j\right)+x_{n+1}^{n+1}$$ where $\psi$ is the unique complex modulus of the family (recall Theorem \[thm:m\*=k\]).
First of all recall that $\operatorname{Sing}(Y^\vee)\subset D'_{n+1}$. Then, from here on, we will restrict to consider singularities in the affine open subset $U_n:=\{x_n\neq 0\}\subset\XX_{\aa_0}$: this suffices as the treatment of singularities in the remaining affine subsets $U_i$, $1\leq i\leq n-1$, is completely analogous and $Y^\vee\cap U_{n+1}$ is smooth. Then $Z_n:=Y^\vee\cap U_n$ is the hypersurface defined in $\C^n$ by the polynomial $$\begin{aligned}
\label{fn}
\nonumber
f_n&:=&\prod_{i=1}^{n-1}z_i^{\d-1}\left(\sum_{i=1}^{n-1} z_i^{n+\d}+1+\psi z_n\prod_{i=1}^{n-1}z_i\right)+z^{n+1}_{n}\\
&=&\zeta\prod_{i=1}^{n-1}z_i^{\d-1}+z^{n+1}_{n}\end{aligned}$$ where: $$\forall\,i=1,\ldots,n-1 \quad z_i=x_{i}/x_n\,,\quad z_n=x_{n+1}/x_n^\d\,,\quad\zeta:=\sum_{i=1}^{n-1} z_i^{n+\d}+1+\psi z_n\prod_{i=1}^{n-1}z_i$$ This means that we are studying the open subset determined by the cone $\s_n:=\langle\L_{\aa_0}^{\{n\}}\rangle=\langle\ll_1,\ldots,\ll_{n-1},\ll_{n+1}\rangle$, opposite to the ray $\rho_{n}=\langle\ll_{n}\rangle$. The singular locus $\operatorname{Sing}(Z_n)$ is contained in the hyperplane $\{z_n=0\}$, which is determined by the torus orbit of the distinguished point of the ray $\rho_{n+1}=\langle\ll_{n+1}\rangle$. Then, a resolution of $Z_n$ has to be obtained as a birational transform induced by a suitable subdivision of $\s_n$, admitting a sub-cone $\s\subset\s_n$ such that:
- $\s$ is a cone of a subdivision of $\s_n$ constructed by adding a suitable number of new rays, associated with exceptional divisors of successive blowups,
- $\rho_{n+1}$ is still a ray of $\s$.
More precisely, notice that every 2-dimensional sub-cone $\langle\ll_i,\ll_{n+1}\rangle$, for $1\leq i\leq n-1$, is the cone spanned by the 1-dimensional face $\operatorname{Conv}(\ll_i,\ll_{n+1})<\D_{\aa_0}$. Every such 1-dimensional face contains exactly $d-1$ lattice points in its relative interior. Recalling definition (\[A\]) of the $A$-triangulation of $\D_{\aa_0}$, giving rise to the resolution $\phi':\widehat{\XX}_{\aa_0}\longrightarrow \XX'_{\aa_0}$ in Proposition \[prop:risoluzione\], all these lattice points determine new rays, hence exceptional divisors of the resolution $\phi'$. The choice of an interior lattice point $\ll^{(i)}\in\operatorname{Int}(\operatorname{Conv}(\ll_i,\ll_{n+1}))$ induces a crepant blowup of $\XX'_{\aa_0}$, whose exceptional divisor is the (closure of the) torus orbit of the distinguished point of $\langle\ll^{(i)}\rangle$.
Let us start the resolution process by considering the birational transform of $Z_n$ in the blowup determined by $\ll^{(1)}\in\operatorname{Int}(\operatorname{Conv}(\ll_1,\ll_{n+1}))$. This is induced by the blowup of the codimension 2 linear subvariety $\{z_1=z_n=0\}\subset\operatorname{Spec}(A_{\s_n})$, under notation introduced in §\[ssez:TV\], giving the following total transform of $Z_n$ $$\left\{\begin{array}{c}
z_1= t_{1}z_n \\
f_n(\z)=0
\end{array}
\right\}\ \Longrightarrow\ \zeta^{(1)}t_1^{\d-1}z_n^{\d-1}\prod_{i=2}^{n-1}z_{i}^{\d-1}+z_n^{n+1}=0$$ being $\zeta^{(1)}$ the obvious transform of $\zeta$. Then, the associated birational transform $B_1Z_n\subset\operatorname{Spec}(A_{\s_n^{(1)}})$, with $\s_n^{(1)}:=\langle\ll^{(1)},\ll_2,\ldots,\ll_{n-1},\ll_{n+1}\rangle$, is given by the zero locus of $$f^{(1)}_n:=\left\{\begin{array}{cc}
\zeta^{(1)}t_1^{\d-1}\prod_{i=2}^{n-1}z_{i}^{\d-1}+z_n^{n-\d+2}\in\C[t_{1},z_2,\ldots,z_n] & \text{if $\d\leq n+2$} \\
\zeta^{(1)}t_1^{\d-1}z_n^{\d-n-2}\prod_{i=2}^{n-1}z_{i}^{\d-1}+1\in\C[t_{1},z_2,\ldots,z_n] & \text{otherwise}
\end{array}\right.$$ Then, singularities are not yet resolved if $n\geq \d$. In this case, we go on by considering the blowup determined by $\ll^{(2)}\in\operatorname{Int}(\operatorname{Conv}(\ll_2,\ll_{n+1}))$ and the associated birational transform $B_2Z_n\subset\operatorname{Spec}(A_{\s_n^{(2)}})$, with $\s_n^{(2)}:=\langle\ll^{(1)},\ll^{(2)},\ll_3,\ldots,\ll_{n-1},\ll_{n+1}\rangle$.
Assuming $\d\geq3$, this process terminates after $s=\left[{n\over\d-1}\right]+1\leq n-1$ blowups, producing a smooth resolution $\widehat{Z}_n=B_sZ_n\longrightarrow Z_n$.
On the other hand, if $\d=2$ then $B_{n-1}Z_n\subset\operatorname{Spec}(A_{\s_n^{(n-1)}})$ is the zero-locus of $$f_n^{(n-1)}=\zeta^{(n-1)}\prod_{i=1}^{n-1}t_i^{\d-1}+z_n^2\in\C[t_1,\ldots,t_{n-1},z_n]$$ with $\s_n^{(n-1}:=\langle\ll^{(1)},\ldots,\ll^{(n-1)},\ll_{n+1}\rangle$. Then, a final resolution can be obtained by choosing a further lattice point $\ll^{(n)}\in\operatorname{Int}(\operatorname{Conv}(\ll^{(n-1)},\ll_{n+1}))$: this is possible up to possibly change the previous choice of $\ll^{(n-1)}\in\operatorname{Int}(\operatorname{Conv}(\ll_{n-1},\ll_{n+1}))$, as there are $d-1=n+1\geq 5>2$ possible choices of useful interior lattice points.
Two combinatorial Lemmas {#ssez:combinatorica}
------------------------
This section is devoted to prove some combinatorial formulas needed to compute Hodge numbers in previous Theorems \[thm:B-mirror0\], \[thm:B-mirror1\] and \[thm:B-mirror2\].
\[lem:combinatorica\] For every positive integers $n\geq 1$ and $d\geq n+1$ the following is an identity $$\label{identita}
\binom{2d-1}{n}-(n+1){d-1\choose n}=\sum_{i=1}^n (-1)^{n-i}{n+1\choose i+1}{id-d+n\choose n}$$
Set $P(k):={(k-1)d-d+n\choose n}$ and think it as a polynomial of degree $\leq n$ in $k$. Then, the $(n+1)^{\text{}st}$ finite differences of $P(k)$ vanish, that is, $$\sum_{k=0}^{n+1} (-1)^k{n+1\choose k}P(k)=0\ \Longrightarrow\ \sum_{k=2}^{n+1} (-1)^k{n+1\choose k}P(k)=-P(0)+(n+1)P(1)$$ Multiplying by $(-1)^{n-1}$ gives $$(-1)^nP(0)-(n+1)(-1)^nP(1)=\sum_{k=2}^{n+1} (-1)^{n-1+k}{n+1\choose k}P(k)$$ Notice that $$(-1)^nP(0)=\binom{2d-1}{n}\ ,\quad (-1)^nP(1)=\binom{d-1}{n}\ ,\quad \forall\,k\quad (-1)^{n-1+k}=(-1)^{n-1-k}$$ Then setting $k=i+1$ in the summation gives (\[identita\]).
\[lem:h11st\] Consider the Gorenstein partial resolution $\XX_{\aa_0}'\longrightarrow\XX_{\aa_0}$ constructed in Proposition \[prop:risoluzione\]. Then, for every positive integers $n\geq 4$ and $d\geq n+2$, the following is an identity $$\label{identita2}
\hst^{1,1}(\XX'_{\aa_0})=c'_1=\binom{d+n}{n}-\binom{d}{n}$$
By Proposition \[prop:risoluzione\] $$\label{c1'}
c'_1=\sum_{h=0}^{n-1}(-1)^{n-1-h}{n\choose 1+h}\vf_{\aa_0}(h)$$ where, by definition, $\vf_{\aa_0}(h)=\left|\{\m\in M\,|\,\vf_K(\m)=-h\}\right|$, being $\vf_K$ the canonical support function of $\XX'_{\aa_0}$. Clearly $\vf_{\aa_0}(0)=1$. Moreover $$\forall\,h\in\N\setminus\{0\}\quad\vf_{\aa_0}(h)=\binom{n+hd}{n}-\sum_{j=1}^h\left(\sum_{i=1}^{d-n-1}
\binom{jd-j+h-i}{n-1}\right)-\sum_{l=0}^{h-1}\vf_{\aa_0}(l)$$ giving rise to a recursive equation admitting the following unique solution $$\begin{aligned}
\vf_{\aa_0}(1) &=& \binom{d+n}{n}-\binom{d}{n}+n\\
\forall\,h\geq 2\quad \vf_{\aa_0}(h) &=& \sum_{i=1}^{n-1}\binom{hd+i}{n-1}+\sum_{l=0}^{h-1}\binom{ld+n+h-l-1}{n-1}- \sum_{l=1}^{h-1}\binom{ld+h-l-1}{n-1}
\end{aligned}$$ Define $$\label{Pk}
\forall\,k\in\Z\quad P(k):= \sum_{i=1}^{n-1}\binom{(k-1)d+i}{n-1}+\sum_{l=0}^{k-2}\binom{ld+n+k-l-2}{n-1}- \sum_{l=1}^{k-2}\binom{ld+k-l-2}{n-1}$$ with the following usual conventions: $$\begin{aligned}
\label{convenzioni}
\binom{-a}{b} &=& (-1)^{b}\binom{a+b-1}{b} \\
\nonumber
\forall\,a,b,c\in\Z\quad\sum_{i=a}^{b-1}f(i)+ \sum_{j=b}^cf(j)&=&\sum_{l=a}^cf(l)
\end{aligned}$$ In particular, the second convention implies that $$\sum_{i=1}^0 f(i)=\sum_{i=0}^{-1}f(i)=0,\ \sum_{i=0}^{-2}f(i)=-f(-1),\ \sum_{i=1}^{-2}f(i)=-f(0)-f(-1),\ \sum_{i=1}^{-1}f(i)=-f(0)$$ Therefore, one has $$\begin{aligned}
P(0)&=&\sum_{i=1}^{n-1}\binom{-d+i}{n-1}-\binom{-d+n-1}{n-1}+\binom{-2}{n-1}+\binom{-d-1}{n-1}\\
&=&\binom{-d+n}{n}-\binom{-d}{n}+(-1)^n\binom{d-1}{n-1}+(-1)^{n-1}n+(-1)^{n-1}\binom{d+n-1}{n-1}\\
&=&(-1)^n\left[\binom{d-1}{n}+\binom{d-1}{n-1}\right]+(-1)^{n-1}\left[\binom{d+n-1}{n}+\binom{d+n-1}{n-1}+n\right]\\
&=&(-1)^{n-1}\left[\binom{d+n}{n}-\binom{d}{n}+n\right]=(-1)^{n-1}\vf_{\aa_0}(1)\\
P(1)&=& \sum_{i=1}^{n-1}\binom{i}{n-1}+\binom{-1}{n-1} = 1+(-1)^{n-1}\\
P(2)&=& \sum_{i=1}^{n-1}\binom{d+i}{n-1}+n=\binom{d+n}{n}-\binom{d}{n}+n=\vf_{\aa_0}(1)\\
P(k)&=&\vf_{\aa_0}(k-1)\quad\text{for $k\geq 3$}
\end{aligned}$$ Definition (\[Pk\]), taking into account conventions (\[convenzioni\]), allows one to check that $n$-th finite differences have to vanish for $P(k)$, that is, $$\label{DF}
\sum_{k=0}^n (-1)^k\binom{n}{k}P(k)=0$$ Then, (\[c1’\]) gives $$\begin{aligned}
c'_1&=&(-1)^n\sum_{k=1}^n(-1)^k\binom{n}{k}\vf_{\aa_0}(k-1)\\
&=&(-1)^n\left[-n+\binom{n}{2}\vf_{\aa_0}(1)+
\sum_{k=3}^n(-1)^k\binom{n}{k}P(k)\right]
\end{aligned}$$ and vanishing (\[DF\]) allows us to conclude that $$\begin{aligned}
c_1' &=& (-1)^n\left[-n+\binom{n}{2}\vf_{\aa_0}(1))-
\sum_{k=0}^2(-1)^k\binom{n}{k}P(k)\right] \\
&=& \vf_{\aa_0}(1)-n =\binom{d+n}{n}-\binom{d}{n}\end{aligned}$$
Extending the duality to complete intersections in toric varieties {#sez:CI}
==================================================================
The present section is devoted to extending $f$-duality to families of complete intersection varieties in a fixed toric variety $X$, keeping in mind §\[sez:dualita-hyp\] and how Borisov generalized the Batyrev duality [@Borisov], [@BB96].
Let $(X,D_\aa=\sum_{j=1}^m a_jD_j)$ be a ftv and $V=\left(
\begin{array}{ccc}
\v_1 & \cdots & \v_m \\
\end{array}
\right)
$ be a fan matrix of $X$, where $m=n+r$, recalling notation §\[sssez:notazione\]. A *partition* of the framing $D_\aa$ is the datum of a partition $$\exists\,l\in \N:\quad I_1\cup\cdots\cup I_l=\{1,\ldots,m\}\ ,\quad\forall\,i\neq j\quad I_i\cap I_j =\emptyset$$ and divisors $D_{\aa_1},\ldots,D_{\aa_l}$ such that $$\forall\,k=1,\ldots,l\quad D_{\aa_k}:=\sum_{i\in I_k}a_iD_i$$ Clearly $D_\aa=\sum_{k=1}^l D_{\aa_k}$, that is, $\aa=\sum_{k=1}^l \aa_k$.\
The ftv $(X,D_\aa)$ with a framing partition $\aa=\sum_{k=1}^l\aa_k$ is called a *partitioned ftv* and denoted by $(X,\aa=\sum_{k=1}^l\aa_k)$.
$f$-process for complete intersections
--------------------------------------
Given a partitioned ftv $$(X,D_\aa=\sum_{k=1}^l D_{\aa_k})$$ consider the following algorithm.
### The partitioned $f$-process algorithm {#algoritmoDnef}
1. Let $\D_\aa$ and $\D_{\aa_1},\ldots,\D_{\aa_l}$ be the polytopes associated with divisors $D_\aa$ and $D_{\aa_1},\ldots,D_{\aa_l}$, respectively, that is $$\begin{aligned}
\D_\aa&=&\{\m\in M_\R\,|\,V^T\cdot\m \geq -\aa\}\\
\forall\,k=1,\ldots,l\quad\D_{\aa_k}&=&\{\m\in M_\R\,|\,V^T\cdot\m \geq -\aa_k\}\end{aligned}$$ In particular, it turns out that $$\label{sommaintersezione}
\bigcap_{k=1}^l\D_{\aa_k}=\{\0\}\quad \text{and}\quad \D_\aa=\sum_{k=1}^l\D_{\aa_k}$$ where the sum denotes the Minkowski sum of polytopes.
2. Define $$\cv{\D}_\aa:=\operatorname{Conv}(\D_{\aa_1},\ldots,\D_{\aa_l})\subset M_\R$$ Clearly $\cv{\D}_\aa\subseteq\D_\aa$ and relations (\[sommaintersezione\]) ensure that $\0\in\operatorname{Int}(\cv{\D}_\aa)$. Recalling Definition \[def:Deltapolytope\], relations (\[sommaintersezione\]) still hold for multiple polytopes $k_0\D_\aa$ and $k_0\D_{\aa_1},\ldots,k_0\D_{\aa_l}$, so giving that $$\bigcap_{k=1}^l[k_0\D_{\aa_k}]=\{\0\}\quad \text{and}\quad \0\in\operatorname{Int}(\D(X,\aa))$$ since $\D(X,\aa)=[\sum_{k=1}^lk_0\D_{\aa_k}]$. Then $\0\in\operatorname{Int}(\cv{\D}(X,\aa))$, being $\cv{\D}(X,\aa):=[k_0\cv{\D}_\aa]$.
3. Set $$\cv{\XX}_\aa:=\XX_{\cv{\Si}_\aa}\quad\text{where}\quad\cv{\Si}_\aa:=\Si_{\cv{\D}(X,\aa)}$$ and let $\cv{\L}_\aa\in\mathbf{M}(n\times \cv{m};\Z)$ be a fan matrix of $\cv{\XX}_\aa$, where $\cv{m}=|\cv{\Si}(1)|$. Notice that $\cv{\XX}_\aa$ is a complete toric variety, by Proposition \[prop:Fmatricedipolitopo\].
4. For every $k=1,\ldots,l$, set $m_k:=|I_k|$ and consider the matrix $$\cv{M}_{\aa_k}:= V_{I_k}^T\cdot\cv{\L}_\aa\in\mathbf{M}(m_k\times \cv{m};\Z)$$ and let $\bb_k=(b_{jk})_{j=1}^{\cv{m}}$ be *the minimum non-negative column vector* such that $$\cv{M}_{\aa_k}^T+B_k\geq \0\quad\text{where}\quad B_k:=\underbrace{\left(\,\bb_k\ \cdots\ \bb_k\,\right)}_{m_k\ \text{times}}\,\in \mathbf{M}(\cv{m}\times m_k,\N)$$ Then, define $\cv{\bb}:=\sum_{k=1}^l\bb_k$. Calling $\cv{D}_1,\ldots,\cv{D}_{\cv{m}}$ the torus invariant generators of $\Weil(\cv{\XX}_{\aa})$, there is a unique induced partition $$J_1\cup\cdots\cup J_l=\{1,\ldots,\cv{m}\}$$ such that $\left(\cv{\XX}_\aa,\cv{D}_{\cv{\bb}}=\sum_{k=1}^l\cv{D}_{\bb_k}\right)$, with $\cv{D}_{\bb_k}:=\sum_{j\in J_k}b_{jk}\cv{D}_j$, is a partitioned ftv.
5. Analogously to step (1), let ${\D}_{\cv{\bb}}$ and $\cv{\D}_{\bb_1},\ldots,\cv{\D}_{\bb_l}$ be the polytopes associated with divisors $\cv{D}_{\cv{\bb}}$ and $\cv{D}_{\bb_1},\ldots,\cv{D}_{\bb_l}$, respectively, that is $$\begin{aligned}
{\D}_{\cv{\bb}}&=&\{\n\in N_\R\,|\,\cv{\L}_\aa^T\cdot\n \geq -\cv{\bb}\}\\
\forall\,k=1,\ldots,l\quad\cv{\D}_{\bb_k}&=&\{\n\in N_\R\,|\,\cv{\L}_\aa^T\cdot\n \geq -\bb_k\}\end{aligned}$$ Then $$\label{sommaintersezione-b}
\bigcap_{k=1}^l\cv{\D}_{\bb_k}=\{\0\}\quad \text{and}\quad {\D}_{\cv{\bb}}=\sum_{k=1}^l\cv{\D}_{\bb_k}$$
6. Analogously to step (2), define $$\cv{\D}_{\cv{\bb}}:=\operatorname{Conv}(\cv{\D}_{\bb_1},\ldots,\cv{\D}_{\bb_l})\subset N_\R$$ Clearly $\cv{\D}_{\bb}\subseteq{\D}_{\cv{\bb}}$ and relations (\[sommaintersezione-b\]) ensure that $\0\in\operatorname{Int}(\cv{\D}_{\cv{\bb}})$. Then, (\[sommaintersezione-b\]) still holds for multiple polytopes $k_1{\D}_{\cv{\bb}}$ and $k_1\cv{\D}_{\bb_1},\ldots,k_1\cv{\D}_{\bb_l}$, so giving that $$\bigcap_{k=1}^l[k_1\cv{\D}_{\bb_k}]=\{\0\}\quad \text{and}\quad \0\in\operatorname{Int}(\D(\cv{\XX}_{\aa},\cv{\bb}))$$ since $\D(\cv{\XX}_{\aa},\cv{\bb})=[\sum_{k=1}^lk_1\cv{\D}_{\bb_k}]$, when $k_1$ is defined as the minimum positive integer such that $\0\in\operatorname{Int}([k_1\D_{\cv{\bb}}])$. Then $\0\in\operatorname{Int}(\cv{\D}(\cv{\XX}_{\aa},\cv{\bb}))$, being $$\cv{\D}(\cv{\XX}_{\aa},\cv{\bb}):=[k_1\cv{\D}_{\cv{\bb}}]$$
7. Analogously to step (3), set $$\cv{\XX}_{\cv{\bb}}:=\XX_{\cv{\Si}_{\cv{\bb}}}\quad\text{where}\quad\cv{\Si}_{\cv{\bb}}:=
\Si_{\cv{\D}(\cv{\XX}_{\aa},\cv{\bb})}$$ and let $\cv{\L}_{\cv{\bb}}\in\mathbf{M}(n\times \widetilde{m};\Z)$ be a fan matrix of $\cv{\XX}_{\cv{\bb}}$, for some $\widetilde{m}\in\N$. As above, $\cv{\XX}_{\cv{\bb}}$ is a complete toric variety, by Proposition \[prop:Fmatricedipolitopo\].
8. Analogously to step (4), for every $k=1,\ldots,l$, set $\cv{m}_k:=|J_k|$ and consider the matrix $$\cv{M}_{\aa_k,\cv{\bb}}:=(\cv{\L}_\aa)_{J_k}^T\cdot\cv{\L}_{\cv{\bb}}\in\mathbf{M}(\cv{m}_k\times \widetilde{m};\Z)$$ and let $\cc_k=(c_{j,k})_{j=1}^{\widetilde{m}}$ be *the minimum non-negative column vector* such that $$\cv{M}_{\aa_k,\cv{\bb}}^T+C_k\geq \0\quad\text{where}\quad C_k:=\underbrace{\left(\,\cc_k\ \cdots\ \cc_k\,\right)}_{\cv{m}_k\ \text{times}}\,\in \mathbf{M}(\widetilde{m}\times \cv{m}_k,\N)$$ Then, $(\cv{\XX}_{\cv{\bb}}, \cv{\cc}:=\sum_{k=1}^l\cc_k)$ is a partitioned ftv, whose partitioned framing is given by $\widetilde{D}_{\cv{\cc}}= \sum_{j=1}^{\widetilde{m}}c_{jk}\widetilde{D}_j$, calling $\widetilde{D}_1,\ldots,\widetilde{D}_{\widetilde{m}}$ the torus invariant generators of $\Weil(\cv{\XX}_{\cv{\bb}})$.
Following the previous algorithm \[algoritmoDnef\], the partitioned ftv $(\cv{\XX}_\aa,\cv{\bb}=\sum_{k=1}^l\bb_k)$, is called a *partitioned $f$-dual* of $(X,\aa=\sum_{k=1}^l\aa_k)$.
A double application of partitioned $f$-duality defines a *partitioned $f$-process* $$\label{nefDprocess}
\left(X,\aa=\sum_{k=1}^l\aa_k\right)\ {\rightsquigarrow}\ \left(\cv{\XX}_\aa,\cv{\bb}=\sum_{k=1}^l\bb_k\right)\ {\rightsquigarrow}\ \left(\cv{\XX}_{\cv{\bb}},\cv{\cc}=\sum_{k=1}^l\cc_k\right)$$ which is called *calibrated* if there exist $\Xi\in\SF(V)$ and $\Xi'\in\SF(\cv{\L}_{\cv{\bb}})$, refining $\Si$ and $\cv{\Si}_{\cv{\bb}}$, respectively, such that $$\left(\widehat{X},\vf^*D_\aa\right)\ {\cong}\ \left(\widehat{X}',(\vf')^*\widetilde{D}_{\cv{\cc}}\right)$$ are isomorphic framed toric varieties, where $$\vf:\widehat{X}(\Xi)\longrightarrow X(\Si)\quad\text{and}\quad\vf':\widehat{X}'(\Xi')\longrightarrow \cv{\XX}_{\cv{\bb}}(\cv{\Si}_{\cv{\bb}})$$ are the small resolutions associated with the choice of $\Xi$ and $\Xi'$, respectively.
The following characterization of a calibrated partitioned $f$-process is a direct consequence of Theorem \[thm:Deltatriviale\].
In the above notation, up to identifying lattices $M$ (hence $N$) of $X$ and $\cv{\XX}_{\cv{\bb}}$, the partitioned $f$-process (\[nefDprocess\]) is calibrated if and only if $$\begin{aligned}
\label{nefDcalibrato}
\nonumber
\cv{\L}_{\cv{\bb}} &=& V\quad\text{up to a permutation of columns} \\
\forall\,k=1,\ldots,l\quad \cc_k&=& \aa_k\end{aligned}$$
\[def:nefpmirror\] Given the partitioned ftv $(X,\aa=\sum_{k=1}^{l}\aa_k)$, assume that the associated partitioned $f$-process (\[nefDprocess\]) is calibrated. Consider the complete intersection subvariety $$Y:=\bigcap_{k=1}^lY_k\subset X\quad\text{with}\quad Y_k\in|D_{\aa_k}|$$ The generic complete intersection subvariety $$Y^\vee:=\bigcap_{k=1}^l Y^\vee_k\subset \cv{\XX}_\aa\quad\text{with}\quad Y_k\in|\cv{D}_{\bb_k}|$$ is called a *$f$-mirror partner of* $Y$.
If $l=1$, that is, the partition is trivial, the $f$-mirror duality defined by the previous Definition \[def:nefpmirror\] reduces to give the $f$-mirror duality between hypersurfaces in toric varieties defined in Definition \[def:mirror\].
\[rem:CIequazioni\] This is the analogue of what described by Remark \[rem:famiglie\] when $l=1$.\
One can explicitly describe the defining polynomials of both $Y$ and $Y^\vee$ in the Cox rings of $X$ and $\cv{\XX}_\aa$, respectively. Namely:
- for every $k=1,\ldots,l$, the lattice polytope $[\D_{\aa_k}]$ is the Newton polytope of $Y_k\in|D_{\aa_k}|$; call $\overline{\L}_{\aa_k}$ a matrix whose columns are given by all the lattice points in $[\D_{\aa_k}]$: it is well defined up to a permutation of columns; setting $l_k:=|\D_{\aa_k}\cap M|$, then $\overline{\L}_{\aa_k}$ is a $n\times l_k$ integer matrix; define $$\overline{M}_{\aa_k}:= V^T\cdot \overline{\L}_{\aa_k}\quad\text{and}\quad \overline{A}_k:=\underbrace{\left(\,\aa_k\ \cdots\ \aa_k\,\right)}_{l_k\ \text{times}}\,\in \mathbf{M}(m\times l_k;\N)\,;$$ then the polynomial of $Y_k$ is given by $$f_k=\sum_{j=1}^{l_k} c_j\x^{\m_j} \in \operatorname{Cox}(X)\cong\C[x_1,\ldots,x_m]$$ where $\m_j=(m_{i,j})$ is the $j$-th column of $\overline{M}_{\aa_k}+\overline{A}_k$ and $\x^{\m_j}:=\prod_{i=1}^m x_i^{m_{i,j}}$;
- recalling step (5) in the algorithm \[algoritmoDnef\], the lattice polytope $[\D_{\bb_k}]$ is the Newton polytope of $Y_k^\vee\in|\cv{D}_{\bb_k}|$; call $\overline{\L}_{\bb_k}$ a matrix whose columns are given by all the lattice points in $[\D_{\bb_k}]$; setting $l'_k:=|\D_{\bb_k}\cap N|$, then $\overline{\L}_\bb$ is a $n\times l'_k$ integer matrix; define $$\overline{M}_{\aa,\bb_k}:= \cv{\L}_\aa^T\cdot \overline{\L}_{\bb_k}\quad\text{and}\quad \overline{B}_k:=\underbrace{\left(\,\bb_k\ \cdots\ \bb_k\,\right)}_{l'_k\ \text{times}}\,\in \mathbf{M}(\cv{m}\times l'_k;\N)\,;$$ then the polynomial of $Y^\vee_k$ is given by $$f^\vee_k=\sum_{j=1}^l c_j\x^{\n_j} \in \operatorname{Cox}(\cv{\XX}_\aa)\cong\C[x_1,\ldots,x_{\cv{m}}]$$ where $\n_j=(n_{i,j})$ is the $j$-th column of $\overline{M}_{\aa,\bb_k}+\overline{B}_k$ and $\x^{\n_j}:=\prod_{i=1}^{\cv{m}} x_i^{n_{i,j}}$.
Notice that, for every $k$, both $f_k$ and $f_k^\vee$ are homogeneous polynomials, with respect to degrees induced by class groups. In fact, columns of both $\overline{M}_{\aa_k}$ and $\overline{M}_{\aa,\bb_k}$ determine trivial divisors, up to linear equivalence. Then $$\deg(f_k)=[D_{\aa_k}]\in\operatorname{Cl}(X)\quad\text{and}\quad\deg(f_k^\vee)=[\cv{D}_{\bb_k}]\in\operatorname{Cl}(\cv{\XX}_\aa)$$
Generalizing Batyrev-Borisov duality
------------------------------------
Definition \[def:nefpmirror\] is clearly motivated by the case when $X$ is a Fano toric variety and $\aa=\1$, that is $D_\aa=-K_X$. In fact, in this case a framing partition $\aa=\sum_{k=1}^{l} \aa_k$ such that $D_{\aa_k}$ is a nef divisor, for every $k=1,\ldots,l$, is precisely a Borisov nef partition of the anti-canonical divisor [@Borisov Def. 2.5, Rem. 2.6], [@BB96 Def. 4.6]. In this case, $f$-duality reduces to give the well known Batyrev-Borisov mirror symmetry between [Calabi-Yau ]{}complete intersections in Fano toric varieties.
\[ex:nefD\] To fixing ideas, an easy example is here presented. Consider the partitioned ftv $$\left(X,\aa=\sum_{k=1}^l\aa_k\right)=\left(\P^2,(1,1,2)=(1,0,0)+(0,1,2)\right)$$ where weights of the partition are referred to primitive generators of the 1-skeleton of the fan defining $\P^2$ and given by columns of the fan matrix $$V=\left(
\begin{array}{ccc}
1 & 0 & -1 \\
0 & 1 & -1 \\
\end{array}
\right)$$ Notice that the framing partition $(1,1,2)=(1,0,0)+(0,1,2)$ is actually a *nef* partition, as both of the summands give back nef divisors. We are then considering a generic complete intersection $Y\subset\P^2$ of a line and a cubic (hence 3 points) whose equations are given by Newton polytopes (step (1) in algorithm \[algoritmoDnef\]) $$\begin{aligned}
\D_{\aa_1} &:=& \operatorname{Conv}(\L_{\aa_1})\ ,\quad \L_{\aa_1}:=\left(
\begin{array}{ccc}
0 & -1 & -1 \\
0 & 1 & 0 \\
\end{array}
\right)
\\
\D_{\aa_2} &:=& \operatorname{Conv}(\L_{\aa_2})\ ,\quad \L_{\aa_2}:=\left(
\begin{array}{ccc}
3 & 0 & 0 \\
-1 & 2 & -1 \\
\end{array}
\right)
\end{aligned}$$ Notice that $\D_\aa=\D_{\aa_1}+\D_{\aa_2}$ (see Fig. \[Fig3\]).
![\[Fig3\] The calibrated partitioned $f$-process of Example \[ex:nefD\].](nefD.jpg){width="12truecm"}
By part (a) in Remark \[rem:CIequazioni\], polynomials defining $Y$ are then given by $$\begin{aligned}
\overline{M}_{\aa_1}+\overline{A}_1 &=& V^T\cdot\L_{\aa_1}+\left(
\begin{array}{ccc}
\aa_1 & \aa_1 & \aa_1 \\
\end{array}
\right) = \left(
\begin{array}{ccc}
1 & 0 & 0 \\
0 & 1 & 0 \\
0 & 0 & 1 \\
\end{array}
\right)
\\
&\Longrightarrow& f_1=a_1x_1+a_2 x_2+a_3 x_3\\
\overline{M}_{\aa_2}+\overline{A}_2 &=& V^T\cdot\overline{\L}_{\aa_2}+\underbrace{\left(
\begin{array}{ccc}
\aa_2 & \cdots & \aa_2 \\
\end{array}
\right) }_{\text{10 times}}
\\
&=& \left(
\begin {array}{cccccccccc} 0&1&0&2&1&0&3&2&1&0\\ 3&2&2&1&1&1&0&0&0&0\\ 0&0&1&0&1&2&0&1&2&3\end {array}
\right)
\\
&\Longrightarrow& f_2=b_1x_2^3+b_2x_1x_2^2+b_3x_2^2x_3+b_4x_1^2x_2+b_5x_1x_2x_3\\
&&\quad\quad\ +b_6x_2x_3^2+b_7x_1^3+b_8x_1^2x_3+b_9x_1x_3^2+b_{10}x_3^3\end{aligned}$$ Then, generically, $Y$ is given by 3 distinct aligned points.
A mirror partner $Y^\vee$ of $Y$ is determined by part (b) in Remark \[rem:CIequazioni\]. Namely, by step (2) in algorithm \[algoritmoDnef\], one has $$\cv{\D}_\aa:=\operatorname{Conv}\left(\D_{\aa_1},\D_{\aa_2}\right)=\operatorname{Conv}\left(
\begin{array}{ccccc}
-1 & -1 & 3 & 0 & 0 \\
1 & 0 & -1 & 2 & -1 \\
\end{array}
\right)$$ Then, passing to step (3), one has $$\cv{\L}_\aa=\left(
\begin{array}{ccccc}
-1 & -1 & 3 & 0 & 0 \\
1 & 0 & -1 & 1 & -1 \\
\end{array}
\right)$$ and this is enough to determine the fan $\cv{\Si}_\aa$ of $\cv{\XX}_\aa$. In particular, an easy check gives that $\cv{\XX}_\aa$ is the blow up of $\P(1,2,1)$ in two distinct points.
Step (4) in algorithm \[algoritmoDnef\] allows us to compute the partitioned framing $\cv{\bb}=\bb_1+\bb_2$ over $\cv{\XX}_\aa$. Namely $$\begin{aligned}
\cv{M}_{\aa_1}^T &=& \cv{\L}_\aa^T\cdot\left(
\begin{array}{c}
1 \\
0 \\
\end{array}
\right)= \left(
\begin{array}{c}
-1 \\
-1 \\
3 \\
0 \\
0 \\
\end{array}
\right)\ \Longrightarrow\ \bb_1=(1,1,0,0,0)
\\
\cv{M}_{\aa_2}^T &=& \cv{\L}_\aa^T\cdot\left(
\begin{array}{cc}
0 & -1 \\
1 & -1 \\
\end{array}
\right)= \left(
\begin{array}{cc}
1 & 0 \\
0 & 1 \\
-1 & -2 \\
1 & -1 \\
-1 & 1 \\
\end{array}
\right)\ \Longrightarrow\ \bb_2=(0,0,2,1,1)\end{aligned}$$ Then, step (5) gives polytopes associated with divisors $\cv{D}_{\cv{\bb}}, \cv{D}_{\bb_1}, \cv{D}_{\bb_2}$, namely $$\begin{aligned}
\cv{\D}_{\bb_1} &=& \operatorname{Conv}\left(
\begin{array}{cc}
1 & 0 \\
0 & 0 \\
\end{array}
\right)
\\
\cv{\D}_{\bb_2} &=& \operatorname{Conv}\left(
\begin{array}{cccc}
0 & 0 & -1 & -1/3 \\
0 & 1 & -1 & 1 \\
\end{array}
\right)
\\
\D_{\cv{\bb}} &=& \operatorname{Conv}\left(
\begin{array}{ccccc}
1 & 1 & 0 & -1/3 & -1 \\
0 & 1 & -1 & 1 & -1 \\
\end{array}
\right)\end{aligned}$$ Notice that $\D_{\cv{\bb}}=\cv{\D}_{\bb_1}+\cv{\D}_{\bb_2}$. By a direct check (use e.g. [@RT-Ample Thm. 3]), divisors $\cv{D}_{\cv{\bb}}$, $\cv{D}_{\bb_1}$ and $\cv{D}_{\bb_2}$ turn out to be semi-ample and line bundles $\cO_{\cv{\XX}_{\aa}}(3\cv{D}_{\cv{\bb}})$, $\cO_{\cv{\XX}_{\aa}}(\cv{D}_{\bb_1})$ and $\cO_{\cv{\XX}_{\aa}}(3\cv{D}_{\bb_2})$ be globally generated.
Passing to step (6) one gets $$\cv{\D}_{\cv{\bb}}:=\operatorname{Conv}(\cv{\D}_{\bb_1},\cv{\D}_{\bb_2})=\operatorname{Conv}\left(
\begin{array}{cccc}
1 & 0 & -1 & -1/3 \\
0 & 1 & -1 & 1 \\
\end{array}
\right)$$ Therefore $[\cv{\D}_{\cv{\bb}}]=\operatorname{Conv}(V)$, so giving $\cv{\L}_{\cv{\bb}}=V$, up to a permutation on columns, which is the first condition in (\[nefDcalibrato\]). To check the second one, by step (8) one gets $$\begin{aligned}
\cv{M}_{\aa_1,\cv{\bb}}^T &=& V^T\cdot\L_{\aa_1}=\left(
\begin{array}{ccc}
0 & -1 & -1 \\
0 & 1 & 0 \\
0 & 0 & 1 \\
\end{array}
\right)\ \Longrightarrow\ \cc_1=(1,0,0)=\aa_1
\\
\cv{M}_{\aa_2,\cv{\bb}}^T &=& V^T\cdot\L_{\aa_2}=\left(
\begin{array}{ccc}
3 & 0 & 0 \\
-1 & 2 & -1 \\
-2 & -2 & 1 \\
\end{array}
\right)\ \Longrightarrow\ \cc_2=(0,1,2)=\aa_2\end{aligned}$$ and the partitioned $f$-process associated with $(\P^3,(1,0,0)+(0,1,2))$ turns out to be calibrated. Then, recalling part (b) of Remark \[rem:CIequazioni\], polynomials defining the mirror partner $Y^\vee$ are given, in the Cox ring $\operatorname{Cox}(\cv{\XX}_\aa)\cong\C[x_1,\ldots,x_5]$, by $$\begin{aligned}
\overline{M}_{\aa,\bb_1}+\overline{B}_1 &=& \cv{\L}_\aa^T\cdot\left(
\begin{array}{cc}
1 & 0 \\
0 & 0 \\
\end{array}
\right)
+\left(
\begin{array}{cc}
\bb_1 & \bb_1 \\
\end{array}
\right) = \left(
\begin{array}{cc}
0 & 1 \\
0 & 1 \\
3 & 0 \\
0 & 0 \\
0 & 0 \\
\end{array}
\right)
\\
&\Longrightarrow& f^\vee_1=a_1x_3^3+a_2 x_1x_2\\
\overline{M}_{\aa,\bb_2}+\overline{B}_2 &=& \cv{\L}_\aa^T\cdot\left(
\begin{array}{ccc}
0 & -1 & 0 \\
1 & -1 & 0 \\
\end{array}
\right)
+\left(
\begin{array}{ccc}
\bb_1 & \bb_1 &\bb_1 \\
\end{array}
\right) =
\\
&=& \left(
\begin {array}{ccc} 1&0&0\\ 0&1&0\\
1&0&2\\
2&0&1\\
0&2&1\end {array}
\right)
\\
&\Longrightarrow& f^\vee_2=b_1x_1x_3x_4^2+b_2x_2x_5^2+b_3x_3^2x_4x_5\end{aligned}$$ Therefore $Y^\vee=Y^\vee_1\cap Y^\vee_2\subset\cv{\XX}_\aa$, where $Y^\vee_1$ is an hypersurface of degree $(3,3,0)\in \operatorname{Cl}(\cv{\XX}_\aa)$ and $Y^\vee_2$ is an hypersurface of degree $(2,3,2)\in \operatorname{Cl}(\cv{\XX}_\aa)$.
Mirroring projective complete intersections
-------------------------------------------
After Theorem \[thm:dualita\] and Corollary \[cor:ipersuperfici\], one may expect that analogous statements could still hold for suitable partitioned framed projective spaces and projective complete intersections. This is actually the case, holding the following
\[thm:CI\] Let $Y_d=\bigcap_kY_{d_k}\subseteq\P^n$ be a complete intersection of $l$ generic projective hypersurface, of degree $d=\prod_kd_k\geq n+1$. Then there always exists a partitioned framing ${\aa}=\sum_k\aa_k$ of $\P^n$ such that $Y_{d_k}\sim D_{\aa_k}$, for every $k$, and the associated partitioned $f$-process is calibrated.
For the proof and any further detail, the interested reader is referred to the incoming paper [@R-fCI].
Further examples, remarks and open problems {#sez:open}
===========================================
This final section is devoted to collect a series of suggestions and perspectives coming from the previous treatment of $f$-duality and the induced mirror web, which will be the main objects of incoming works. Let us first of all recall, in order of appearance, main problems earlier arisen.
1. Understanding the generalized Krawitz duality and LG/Hypersurfaces correspondence as sketched in §\[sssez:LG/Hyp\] and in particular what is concerning Remark \[rem:K-dualità\].
2. Recalling §\[ssez:KKP-compct\], understanding relations between $f$-duality, log geometry and Intrinsic Mirror Symmetry in the sense of the Gross-Siebert program.
3. Conjecture \[conj:LGmirror\] and, more in general, the HMS implications of $f$-duality, taking into account the previous items, as observed in Remark \[rem:LGmirrors\].
4. Check several MS instances, among those listed in §\[ssez:mirrortest\], for some further examples of hypersurfaces and complete intersections in toric varieties.
In the following, we present some further interesting remark and related problems.
$f$-mirrors of Kodaira negative hypersurfaces {#ssez:NegKod}
---------------------------------------------
In Definition \[def:ftv\] we asked for a *framing* to be a strictly effective divisor. This is motivated by the willing of giving rise to an involutive duality between pairs of framed toric varieties (recall Remark \[rem:negKod\]). On the other hand, Givental’s approach [@Givental-ICM] produces mirror partners of complete varieties admitting non-negative first Chern class, by means of by LG models, so introducing a strong asymmetry in the mirror correspondence, the [Calabi-Yau ]{}case. Actually he proved a Mirror Theorem in the case of toric complete intersections [@Givental96]. Consequently, we are led to relax the definition of a framing, just requiring it is no more than an *effective divisor*.
\[def:wftv\] A *weak framed toric variety* is a couple $(X,D_\aa)$ (also denoted $(X,\aa)$) where:
- $X$ is a complete toric variety, with $\dim(X)=n$ and $\operatorname{rk}(\operatorname{Pic}(X))=r$,
- $D=\sum_{\rho\in\Si(1)}a_\rho D_\rho=\sum_{i=1}^m a_i D_i\in\Weil(X)$, with $m=n+r$, is an effective torus invariant Weil divisor, called a *weak framing* of $X$.
By Proposition \[prop:gg\], the associated polyhedron $\D_\aa$ is still a polytope, but in general $\0\in M$ is no more a relative interior point of $\D_\aa$, but just a lattice point of $\D_\aa$. Recalling Definition \[def:Deltapolytope\], define the $f$-polytope associated with a wftv $(X,\aa)$ to be the following $$\D(X,\aa):=[k_0\D_\aa]\quad,\quad k_0:=\min\{k\in\N\,|\,\0\in[k\D_\aa]\}$$ Then, the toric variety over the polytope $\D(X,\aa)$, that is, $\XX_\aa:=\XX_{\D(X,\aa)}$ is no more complete, in general, and we cannot hope to reconstructing a mirror wftv of $(X,\aa)$. By the way, we can adopt the Givental’s asymmetry and thinking of $\XX_\aa$, endowed with a sort of framing we are going to define in a moment, in terms of LG mirror model. More precisely, calling $\L_\aa$ the fan matrix of $\XX_\aa$ obtained by the $f$-polytope $\D(X,\aa)$ as in Proposition \[prop:Fmatricedipolitopo\], and recalling (\[b\]), let us define the *mirror framing* as the *minimum non-negative column vector* $\bb=(b_j)_{j=1}^{m'}$ (i.e. effective divisor $D'_\bb\in \Weil(\XX_\aa)$) such that $$\label{wb}
M_\aa^T+B\geq \0\quad\text{where}\quad B:=\underbrace{\left(\,\bb\ \cdots\ \bb\,\right)}_{m\ \text{times}}\,\in \mathbf{M}(m'\times m;\N)$$ being $V$ a fan matrix of $X$ and $M_\aa:=V^T\cdot\L_\aa\in\mathbf{M}(m\times m';\Z)$, as usual. One can now go on as in §\[sssez:LG/Hyp\], by setting:
- $\T_\aa\cong(\C^*)^n$ be the maximal acting torus on $\XX_\aa$,
- $f^\vee_\bb:=f^\vee/\x^\bb\in\C[\x,\x^{-1}]$, where $f^\vee$ is the generic polynomial given in (\[fdual\]), generated by the columns of the matrix $\overline{M}_{\aa,\bb}+\overline{B}$, with $M_{\aa,\bb}=\L_\aa^T\cdot V$: then $f^\vee_\bb$ is the generic polynomial generated by the columns of the matrix $\overline{M}_{\aa,\bb}$.
\[def:LGmirror\] Given a wftv $(X,\aa)$, let $f$ be the generic polynomial constructed as in (\[f-WT\]) and generated by the columns of $\overline{M}_\aa+\overline{A}$, and let $Y\in|D_\aa|$ be the hypersurface defined by $f$. Then:
1. the hypersurface $Y^\vee\subset\XX_\aa$ defined by the generic polynomial $f^\vee$ is called an *$f$-mirror partner* of $Y$,
2. the LG model given by $(\T_\aa,f^\vee_\bb)$ is called a *$f$-mirror LG model* of $Y$.
### According with Givental’s LG mirror model
The passing from a framing to a weak framing, that is the dropping of the word “strictly”, explains why one cannot expect a complete mirror model for toric hypersurfaces (and complete intersections, adapting to a weak framing what described in §\[sez:CI\]) associated with a weak framing. The LG mirror model given in Definition \[def:LGmirror\], turns out to share a certain similarity with the Givental’s LG mirror model (see e.g. the following Remark \[rem:LGparagone\]). Actually, showing that these LG models are equivalent from the HMS point of view is a completely open task, at least at the best of author’s knowledge.
Here we consider the case of hypersurfaces $Y_d\subset\P^n$ of degree $d\leq n$, for sake of completeness what analyzed in §\[sez:ipersuperfici\].
\[prop:d<n+1\] For every $d=1,\ldots,n$, set $\aa_d:=(\1_d,\0_{n+1-d})$ and consider the wftv $(\P^n,\aa_d)$. Then $$\label{wf-politopo}
\D(\P^n,\aa_d)=[\D_{\aa_d}]=\D_{\aa_d}$$ turns out to be the Newton polytope of the generic degree $d$ homogeneous polynomial in $\C[\x]$, whose zero-locus defines the generic hypersurface $Y_d\subset\P^n$. Moreover, $\XX_{\aa_d}$ is a non-complete toric variety and
1. an $f$-mirror partner of $Y_d$ is given by the hypersurface $Y^\vee_d$ of $\XX_{\aa_d}$ defined as the zero-locus of $$\begin{aligned}
f^\vee_1 &=& \prod_{i=1}^nx_i\cdot\left(1+\sum_{i=1}^n x_i\right)+1\in\operatorname{Cox}(\XX_{\aa_1})\quad\text{if $d=1$} \\
f^\vee_d &=& \prod_{k=1}^{n+1}x_k\cdot\left(1+\sum_{j=d+1}^{n+1}x_j^{d}\right)+\sum_{i=1}^d x_i^d \in\operatorname{Cox}(\XX_{\aa_d}) \quad\text{if $2\le d\le n$}
\end{aligned}$$
2. an $f$-mirror LG model of $Y_d$ is given by $(\T_{\aa_d},f^\vee_{\bb_d})$, where $\T_{\aa_d}\cong(\C^*)^n$ is the maximal acting torus on $\XX_{\aa_d}$ and $$f^\vee_{\bb_d}=\left\{\begin{array}{cc}
1+\sum_{i=1}^n x_i+1/\prod_{i=1}^nx_i & \text{if $d=1$} \\
1+\sum_{j=d+1}^{n+1}x_j^{d}+\left(\sum_{i=1}^d x_i^d\right)/ \prod_{k=1}^{n+1}x_k & \text{if $2\le d\le n$}
\end{array}
\right.$$
Equalities (\[wf-politopo\]) are immediately obtained by definitions.
Assume now $d=1$. Then $$\L_{\aa_1}=\left(
\begin{array}{ccccc}
-1 & -1& \cdots&-1 & -1 \\
1 &0 &\cdots& 0&0 \\
0&1&0&\cdots&0\\
\vdots&\ddots&\ddots&\ddots&\vdots\\
0&\cdots&0&1&0
\end{array}
\right)=\left(
\begin{array}{ccc}
&-\1_n& \\
\hline
I_{n-1}&\vline &\0_{n-1}^T \\
\end{array}
\right)\in\M(n,n;\Z)$$ Then $\XX_{\aa_1}\cong\C^n$ is affine and $\operatorname{Cox}(\XX_{\aa_1})\cong\C[x_1,\ldots,x_n]$. Moreover, (\[wb\]) gives $$M_{\aa_1}^T=\L_{\aa_1}^T\cdot V= \left(
\begin{array}{cc}
-\1_n^T & I_n \\
\end{array}
\right)\ \Longrightarrow\ \bb_1=\1_n\ \Longrightarrow\ \x^{\bb_1}=\prod_{i=1}^nx_i$$ $$\overline{M}_{\aa_1,\bb_1}+\overline{B}=\L_{\aa_1}^T\cdot \overline{V}+\overline{B}=\left(
\begin{array}{cccccc}
1 & 0 & 2 & 1 & \cdots & 1 \\
\vdots& \vdots & 1 & 2 & \cdots & 1 \\
\vdots& \vdots & \vdots & \ddots & \ddots & \vdots \\
1& 0 & 1 & \cdots & 1 & 2 \\
\end{array}
\right)\in\M(n,n+2;\Z)$$ $$\begin{aligned}
&\Longrightarrow&f_1^\vee=\prod_{i=1}^nx_i\cdot\left(1+\sum_{i=1}^n x_i\right)+1\\
&\Longrightarrow& f^\vee_{\bb_1}=1+\sum_{i=1}^n x_i+{1\over\prod_{i=1}^nx_i}\end{aligned}$$ Assume now $2\le d\le n$. Then $$\L_{\aa_d}=\left(
\begin{array}{ccccc}
&-\1_{d,n+1}&+&\left(dI_d\,\vline\,\0_{d,n+1-d}\right)& \\
\hline
\0_{n-d,d}&\vline &dI_{n-d}&\vline&\0_{n-d}^T \\
\end{array}
\right)\in\M(n,n+1;\Z)$$ and $\operatorname{Cox}(\XX_{\aa_d})\cong\C[x_1,\ldots,x_{n+1}]$. Moreover, (\[wb\]) gives $$M_{\aa_d}^T=\L_{\aa_d}^T\cdot V= \left(
\begin{array}{c|c}
-\1_{d,d}+dI_d & \0_{d,n+1-d} \\ \hline
-\1_{n+1-d,d}& dI_{n+1-d}
\end{array}
\right)\ \Longrightarrow\ \bb_d=\1_{n+1}\ \Longrightarrow\ \x^{\bb_d}=\prod_{k=1}^{n+1}x_k$$ $$\overline{M}_{\aa_d,\bb_d}+\overline{B}=\L_{\aa_d}^T\cdot \overline{V}+\overline{B}=\left(
\begin{array}{c|c|c}
\1_{d}^T&dI_{d}&\1_{d,n+1-d}\\
\hline \1_{n+1-d}^T & \0_{n+1-d,d} & (d+1)I_{n+1-d}
\end{array}
\right)$$ $$\begin{aligned}
&\Longrightarrow&f_d^\vee=\prod_{k=1}^{n+1}x_k\cdot\left(1+\sum_{j=d+1}^{n+1}x_j^{d}\right)+\sum_{i=1}^d x_i^d\\
&\Longrightarrow& f^\vee_{\bb_d}=1+\sum_{j=d+1}^{n+1}x_j^{d}+{\sum_{i=1}^d x_i^d\over \prod_{k=1}^{n+1}x_k}\end{aligned}$$
Notice that for $d\geq 2$, $\XX_{\aa_d}$ is covered by at least two open affine subsets. In particular, $f^\vee_d$ restricted to one such open affine subset becomes of the same shape as $f^\vee_1$, that is, setting e.g. $x_1=1$, one gets $$f^\vee_d|_{\{x_1=1\}}= \prod_{k=2}^{n+1}x_k\cdot\left(1+\sum_{j=d+1}^{n+1}x_j^{d}\right)+1+\sum_{i=2}^d x_i^d= \prod_{i=1}^{n}y_i\cdot\left(1+\sum_{i=1}^{n}y_i^d\right)+1+ \sum_{i=1}^{d-1}y_i^d$$ by setting $y_i=x_{i+1}$. In particular, imposing $d=1$, the right hand side gives $f_1^\vee(\y)$.
Moreover, for $d\geq n+1$, the construction above is precisely the one already analyzed in §\[sez:ipersuperfici\].
\[rem:LGparagone\] If $d=1$ then $Y_d\cong\P^{n-1}$ embedded in $\P^n$ by setting $x_1=0$. One can then check the relation between the LG mirror model $((\C^*)^n, f_{\bb_1}^\vee)$ given in Proposition \[prop:d<n+1\] and the Givental’s LG mirror model as given, e.g., in the Introduction of [@GKR] and in Ex. 2.2 of [@KKP]. In particular, the LG model here presented turns out to be the section $x_{n+1}=1$ of the LG model presented in [@GKR], after the dimensional correction needed to comparing the two constructions.
What happens when the $f$-process is not calibrated? {#ssez:noncalibrato}
----------------------------------------------------
Recalling Definition \[def:Deltaproc banale\], assume that the $f$-process $$(X,\aa)\stackrel{f-dual}{\rightsquigarrow}(\XX_\aa,\bb)\stackrel{f-dual}{\rightsquigarrow}(\XX_\bb,\cc)$$ is not calibrated. This fact means that $f$-duality cannot be involutive or, in other words, that it is asymmetric: this is not a new situation, as for instance the case of the Givental’s Fano/LG model correspondence and, more in general, as for $f$-duality on a weak framed toric variety just considered in the previous §\[ssez:NegKod\].
Le us then assume, by definition as done in Definition \[def:dual-ftv\], that:
- $(\XX_\aa,\bb)$ *is the $f$-dual ftv of $(X,\aa)$ and $(\XX_\bb,\cc)$ is the $f$-dual ftv of $(\XX_\aa,\bb)$*.
Calling $Y,Y',Y''$ the generic hypersurfaces in $|D_\aa|,|D'_\bb|,|D''_\cc|$, respectively, many questions are naturally arising.
1. There is a relation between $(X,\aa)$ and $(\XX_\bb,\cc)$? For instance, is there a birational map $f:\XX_\bb\dashrightarrow X$ such that $D''_\cc=f^{-1}(D_\aa)$? If not, may a similar birational transformation relate $(X,\aa)$ with the final ftv obtained after a finite and even number of $f$-dual passages?
2. Recalling §\[ssez:mirrortest\], which mirror symmetric aspect is relating hypersurfaces in the ordered pairs $(Y,Y')$ and $(Y',Y'')$?
3. Is there a relation linking $Y$ and $Y''$? For instance, is it true that $h^{p,q}(\widehat{Y})=h^{p,q}(\widehat{Y}'')$ for suitable resolutions $\widehat{Y}\longrightarrow Y$ and $\widehat{Y}''\longrightarrow Y''$? Are they equivalent from the HMS point of view?
The present paper is already too long, to starting analyzing these and related problems, but they are interesting questions to be settled in future works.
Generalized complete intersections, BH-transpolarity and $f$-duality {#ssez:Hubsch}
--------------------------------------------------------------------
Recently, physicists Anderson *et al.* [@AAGGL] described a method to produce examples of new [Calabi-Yau ]{}varieties which are not compete intersections. The basic idea is taking an hypersurface (or complete intersection) $Y$ in an ambient variety $P$ and then considering hypersurfaces (or complete intersections) $X$ in $Y$ for which there need not exist sections of two (or $r+s$, resp.) line bundles on $P$ whose common zero locus is $X$. The [Calabi-Yau ]{}condition is resumed by a constraint on involved degrees of $Y$ and $X$: hence it is not an essential tool of the geometric construction of these varieties, called *generalized* complete intersections (gCI). This method has been further studied by Berglund and Hübsch [@BH-CYgCI] and rigorously (and nicely) explained in cohomological terms, in the basic case $r=s=1$, by mathematicians Garbagnati and Van Geemen [@G-vG_gCI], who presented $X$ as the zero locus of a global section $\xi$ of a suitable negatively twisted line bundle on $P$, restricted to $Y$.\
In their preprint , Berglund and Hübsch conjecturally described a method to extending Batyrev-Borisov mirror duality on [Calabi-Yau ]{}complete intersections to that kind of generalized [Calabi-Yau ]{}complete intersections, by means of a, so called, *trans-polarity between VEX polytopes*, that is, a sort of a finite patching of Batyrev-Borisov dualities on convex pieces composing a not necessarily convex polytope, arising as the Newton polytope associated with the global section $\xi$ (). Very recently, T. Hübsch pointed me out (private communication) that, dropping [Calabi-Yau ]{}condition in the above mentioned transpolarity may correspond to replacing BB-duality on the convex pieces by $f$-duality. This observation opens interesting, although possibly intricate, perspectives to extending $f$-duality to generalized complete intersections in a toric ambient variety $P$.
Toric degeneration: extending $f$-duality via geometric transitions {#ssez:Tdegenarazione}
-------------------------------------------------------------------
Following Batyrev’s ideas given in [@Batyrev02] (see also [@Rossi-JGP §6.3]), since $f$-mirror partners come in families, one can easily extend the $f$-mirror definition to complete algebraic varieties admitting a toric degeneration.
\[def:degenerazione\] Let $Y$ be a smooth and complete algebraic variety isomorphic to the generic fiber of a flat family $y:\cY\longrightarrow B$, endowed with a special point $0\in B$ such that $Y_0:=y^{-1}(0)$ is isomorphic to a complete intersection subvariety of a complete toric variety $X(\Si)$, determined by a nef-partitioned framing $D_\aa=\sum_kD_{\aa_k}$ of $X$: $Y_0$ is called a *toric degeneration* of $Y$. Assume that the nef-partitioned process associated to $(X,\aa)$ is calibrated. Then the generic complete intersection $Y_0^\vee$, giving a $f$-mirror partner of $Y_0$, is also an *$f$-mirror partner of* $Y$.
\[conj:topmirror\] In the same notation of the previous Definition \[def:degenerazione\], there exists a partitioned ftv $(X,\aa=\sum_k\aa_k)$ and a suitable resolution $\widehat{Y}_0^\vee\longrightarrow Y_0^\vee$ such that the $f$-mirror partner $Y_0^\vee$ of $Y$ is a topological mirror partner of $Y$, that is, $$k_{\widehat{Y}_0^\vee}=m_Y\quad\text{and}\quad k_Y=m_{Y_0^\vee}$$
Notice that, calling $\widehat{Y}_0\longrightarrow Y_0$ a resolution of singularities, the process $$\label{g.t.}
\xymatrix{\widehat{Y}_0\ar[r]&Y_0\ar@{<~>}[r]&Y}$$ is a *geometric transition* (see [@Rossi-JGP Def. 1.4] for a definition, here considered in a broader sense, beyond the [Calabi-Yau ]{}setup). Recalling Morrison’s argumentation given in [@Morrison] (see also [@Rossi-JGP §6.2]), the extension of Batyrev’s mirror duality, given by $f$-mirror duality, allows one to formulate, beyond the [Calabi-Yau ]{}setup, the following
\[conj:reverset.\] Under notation given in Definition \[def:degenerazione\] and Conjecture \[conj:topmirror\], and given the geometric transition (\[g.t.\]), there should exist a *reverse* geometric transition $$\xymatrix{\widehat{Y}_0^\vee\ar[r]&Y^\vee_0\ar@{<~>}[r]& Y^\vee}$$ such that $Y_0$ is a topological mirror partner of $Y^\vee$, that is, $$k_{\widehat{Y}_0}=m_{Y^\vee}\quad\text{and}\quad k_{Y^\vee}=m_{Y_0}$$ In particular, $Y_0^\vee$ is a toric degeneration of $Y^\vee$, meaning $Y^\vee$ is isomorphic to the generic fiber of a flat family $y^\vee:\mathcal{Y}^\vee\longrightarrow B^\vee$, endowed with a special point $0^\vee\in B^\vee$ such that $(y^{\vee})^{-1}(0^\vee)\cong Y_0^\vee$.
Following the lines given in [@Batyrev02], in the [Calabi-Yau ]{}setup, examples satisfying Conjecture \[conj:reverset.\] are constructed by means of the monomial-divisor correspondence [@AGM]. More or less, the same argumentation may be extended beyond the [Calabi-Yau ]{}setup. In fact, the meaning of the monomial-divisor correspondence is that of the differential of the mirror map. Assume that there exist well defined isomorphisms (actually differentials of mirror maps) $$\xymatrix{\mu_A:K_Y\ar[r]^\cong &M_{Y_0^\vee}}\ ,\quad \xymatrix{\mu'_B:K_{\widehat{Y}_0}\ar[r]^\cong &M_{Y^\vee}}$$ where $K_Y,K_{\widehat{Y}_0}$ are the tangent spaces to the [Kähler ]{}moduli spaces of $Y$ and $\widehat{Y}_0$, respectively, and analogously $M_{Y_0^\vee},M_{Y^\vee}$ are the tangent spaces to the complex moduli speces of $Y_0^\vee$ and $Y^\vee$, respectively: assume all of them are well defined! The isomorphism $\mu_A$ comes from the $A$-side topological mirror symmetry of $(Y,Y_0^\vee)$ and the isomorphism $\mu'_B$ comes from the $B$-side topological mirror symmetry of $(Y^\vee,Y_0)$ (recall Definition \[def:A,B-mirror\]). The geometric transition (\[g.t.\]) induces an inclusion $K_Y\hookrightarrow K_{\widehat{Y}_0}$, via the inclusion of the associated Picard groups. Therefore, the subspace $\mu'_B(K_Y)\subset M_{Y^\vee}$ defines a first-order deformation of $Y^\vee$ which should give rise to the toric degeneration to $Y_0^\vee$.
General hyperelliptic curve {#ssez:iperellittica}
---------------------------
The only examples of toric hypersurfaces considered throughout the present paper, were hypersurfaces and complete intersections in some projective space. The present subsection is devoted to consider a general hyperelliptic curve of genus $g\geq 2$ as presented in [@KKOY §4.1], that is, a divisor $Y$, of bi-degree $(2,g+1)$, in the Hirzebruch surface $\FF_0=\P(\cO_{\P^1}\oplus\cO_{\P^1})$. The latter is a toric variety of Picard number $r=2$, hence a substantially different example from the case of $\P^n$. The reader is warmly invited to comparing the $f$-mirror (complete) model here proposed with Landau-Ginzburg mirror models proposed in [@KKOY] and, for $g=2$, in [@Seidel], and, moreover, for the case of a general curve of genus $g\geq 2$ in [@Efimov], generalizing Seidel’s approach.
A fan matrix of $\FF_0$ is given by $$V=\left(
\begin{array}{cccc}
1 & -1 & 0 & 0 \\
0 & 0 & 1 & -1 \\
\end{array}
\right)$$ and a framing $D_\aa$ of bi-degree $(2,g+1)$ is given, e.g., by $\aa=(1,1,1,g)$. Then $$\L_\aa=\left(
\begin{array}{cccc}
1 & 1 & -1 & -1 \\
g & -1 & g & -1 \\
\end{array}
\right)\ \Longrightarrow\ \D_\aa=\operatorname{Conv}(\L_\aa)$$ In particular $D_\aa$ is an ample divisor of $\FF_0$. Recalling (\[b\]), $$M_\aa^T=\L_\aa^T\cdot V= \left(
\begin{array}{cccc}
1&-1&g&-g\\
1&-1&-1&1\\
-1&1&g&-g\\
-1&1&-1&1
\end{array}
\right)\ \Longrightarrow\ \bb=(g,1,g,1)$$ Then $$\D_{\bb}=\operatorname{Conv}\left(
\begin{array}{cccc}
2g\over g+1 & 0 & 0 & -{2g\over g+1} \\
-{g-1\over g+1} & 1 & -1 & g-1\over g+1 \\
\end{array}
\right)\ \Longrightarrow\ \left[\D_\bb\right]=\NN=\operatorname{Conv}(V)$$ so giving $\L_\bb=V$, up to a permutation of columns. Moreover, (\[c\]) gives $$M_{\aa,\bb}^T=V^T\cdot\L_\aa=M_\aa\ \Longrightarrow\ \cc=(1,1,1,g)=\aa$$ implying that the $f$-process is calibrated, by Theorem \[thm:Deltatriviale\].
Part (b) of Remark \[rem:famiglie\] gives the polynomial $f^\vee\in\operatorname{Cox}(\XX_\aa)\cong\C[x_1,\ldots,x_4]$, defining the generic element $Y^\vee$ of the mirror family, $$f^\vee=c_1x_1^{2g}x_3^{2g}+c_2x_1^{g+1}x_2^2x_3^{g-1}+c_3x_1^{g}x_2x_3^{g}x_4+c_4x_1^{g-1}x_3^{g+1}
x_4^2+c_5x_2^2x_4^2$$ Recalling (\[aut\]), one gets $\dim(\operatorname{Aut}(\XX_\aa))=2$ and the generic $f^\vee$ can be reduced to the following shape $$\label{dual f}
f^\vee=x_1^{2g}x_3^{2g}+x_1^{g+1}x_2^2x_3^{g-1}+x_1^{g}x_2x_3^{g}x_4+\psi\, x_1^{g-1}x_3^{g+1}x_4^2+\vf\, x_2^2x_4^2$$ As a Cox quotient, $\XX_\aa\cong \left(\C^4\setminus Z\right)/H_\aa$, where the irrelevant locus $Z$ is the union of two plains meeting in the origin of $\C^4$, namely $Z=\{x_1=x_2=0\}\cup\{x_3=x_4=0\}$, and $$H_\aa\cong\left\{\begin{array}{cc}
(\C^*)^2 & \text{if $g=2h$ is even} \\
(\C^*)^2\times\boldsymbol{\mu}_2 & \text{if $g=2h+1$ is odd}
\end{array}\right.\quad h\in\N\setminus\{0\}$$ In particular the weight matrix defining the action of $H_\aa$ over $\C^4\setminus Z$ is given by $$\begin{aligned}
Q_g&=&Q_{2h}=\left(
\begin{array}{cccc}
1 & g & 1 & g \\
0 & g+1 & 2 & g-1 \\
\end{array}
\right)=\left(
\begin{array}{cccc}
1 & 2h & 1 & 2h \\
0 & 2h+1 & 2 & 2h-1 \\
\end{array}
\right)\ \text{if $g=2h$}\\
Q_g&=&Q_{2h+1}\times \boldsymbol{\tau}=\left(
\begin{array}{cccc}
1 & h & 0 & h+1 \\
0 & h+1 & 1 & h \\
\end{array}
\right)\times\left(
\begin{array}{cccc}
1 & 1 & 1 & 1 \\
\end{array}
\right)\ \text{if $g=2h+1$}\end{aligned}$$ meaning that the action is given by $$\begin{aligned}
\label{azione_g}
&\xymatrix{((\l,\mu),\x)\ar@{|->}[r]&(\l x_1,\l^g\mu^{g+1}x_2,\l\mu^2x_3,\l^g\mu^{g-1}x_4)}& \quad\text{if $g=2h$} \\
\nonumber
&\xymatrix{((\l,\mu,\pm 1),\x)\ar@{|->}[r]&(\pm\l x_1,\pm\l^h\mu^{h+1}x_2,\pm\mu x_3,\pm\l^{h+1}\mu^hx_4)} &\quad\text{if $g=2h+1$}\end{aligned}$$ Notice that $f^\vee$ is equivariant to both these actions. In particular, as an element of $\operatorname{Cox}(\XX_\aa)$, which is graded on $$\operatorname{Cl}(\XX_\aa)\cong\left\{\begin{array}{cc}
\Z^2 & \text{if $g=2h$ is even} \\
\Z^2\oplus\Z/2\Z & \text{if $g=2h+1$ is odd}
\end{array}\right.\quad h\in\N\setminus\{0\}$$ $f^\vee$ turns out to be homogeneous of degree either $(4g,4g)$ or $(2g,2g,\overline{0})$, respectively. In particular, it turns out that, if $g$ is even then $(g+1)D_\bb$ is ample and, analogously, if $g=2h+1$ is odd then $(h+1)D_\bb$ is ample (apply e.g. [@RT-Ample Thm. 3]).
### Hori-Vafa type LG mirror models {#sssez:HViperellittica}
In [@KKOY §4.1] a LG mirror model of the general hyperelliptic curve of genus $g$ is proposed, by adopting the Hori-Vafa recipe [@Hori-Vafa] for an hypersurface of bi-degree $(2,g+1)$ in $\FF_0$. By a different approach, Seidel proposed a further LG mirror model for the case $g=2$ [@Seidel], obtained as an unramified quotient of the Hori-Vafa LG mirror model of a plane quintic curve. Seidel’s methods have been generalized by Efimov [@Efimov] for every $g\geq 2$. In particular we get a double proposals of LG mirror models for the generic hyperelliptic curve of genus $g\geq 2$. In all these cases, authors checked one direction of HMS.
Recalling what observed in §\[ssez:HoriVafa\], and in particular in §\[sssez:HVinvariante\], we can obtain a further proposal of LG mirror model for the general hyperelliptic curve of genus $g\geq 2$, by considering the LG model $(\L_{g,\vf,\psi},w_{g,\vf,\psi})$ so defined:
- $\L_{g,\vf,\psi}\cong(\C^*)^4$ is an irreducible component of the reducible torus complete intersection $$\L_{g,\vf,\psi}:=\left\{\tau_1\,y_1^2=
x_1^{2g}x_2^2x_3^{2g}x_4^2=
\tau_2\,y_2^2\right\}\subset (\C^*)^4\times (\C^*)^2$$ where $4\vf^{2}=\tau_1=e^{t_1}$ and $4\psi^{2}=\tau_2=e^{t_2}$, being $t_1,t_2\in\C^*$ [Kähler ]{}parameters related with volumes of the two rulings on $\FF_0$;
- $w_{g,\vf,\psi}$ is the restriction to $\L_{g,\vf,\psi}$ of the regular function $\widetilde{w}_{g,\vf,\psi}:\C^6\longrightarrow\C$ defined by $$\widetilde{w}_{g,\vf,\psi}(\x,\y):=x_1^{2g}x_3^{2g}+x_1^{g+1}x_2^2x_3^{g-1}+\psi (x_1^{g-1}x_3^{g+1}x_4^2+y_2)+\vf (x_2^2x_4^2+y_1)$$
When restricted to $\L_{g,\vf,\psi}$, the superpotential $\widetilde{w}_{g,\vf,\psi}$ can then be rewritten as $f^\vee$ in (\[dual f\]), that is, $$w_{g,\vf,\psi}(\x)= x_1^{2g}x_3^{2g}+x_1^{g+1}x_2^2x_3^{g-1}+x_1^{g}x_2x_3^{g}x_4+\psi\, x_1^{g-1}x_3^{g+1}x_4^2+\vf\, x_2^2x_4^2 =f^\vee$$ This gives the following global picture, analogous to (\[LGquoziente\]), $$\xymatrix{ \{\0\}\ar@{^(->}[r]&\C&\L_{g,\vf,\psi}\cong(\C^*)^{4}\ar@{^(->}[r]\ar[l]_-{w_{g,\vf,\psi}}
\ar@{->>}[d]_-{/H_\aa}&\C^4\setminus Z\ar@{->>}[d]_-{/H_\aa}\\
w_{g,\vf,\psi}^{-1}(0)/H_\aa\ar[u]\ar@{^(->}[rr]&&\T_\aa\ar@{^(->}[r]&\XX_{\aa}}$$ where $\T_\aa$ is the acting torus on $\XX_{\aa}$. Then the $f$-mirror $Y^\vee$ of $Y$, as proposed in Definition \[def:mirror\] and defined by $f^\vee\in\operatorname{Cox}(\XX_\aa)$, is precisely the closure $$Y^\vee=\overline{w_{g,\vf,\psi}^{-1}(0)/H_\aa}\subset\overline{\T}=\XX_{\aa}$$ induced by the open embedding $\T_\aa\hookrightarrow \XX_{\aa}$. Evidences seem enough to motivating the following
A LG mirror model of the general hyperelliptic curve of genus $g\geq2$ of [Kähler ]{}parameters $t_1,t_2$, is given by $((\C^*)^4,w_{g,\vf,\psi})$, with $2(2+\ln(\vf))=t_1$ and $2(2+\ln(\psi))=t_2$.
Accordingly with Hori-Vafa terminology [@Hori-Vafa], a similar LG model admits an associated gauged linear sigma model whose gauge action is given by the $(\C^*)^2$-action described in (\[azione\_g\]). Quotienting by such a gauge action gives back, up to a possibly further quotient by $\Z/2\Z$ depending on the parity of the genus $g$, the LG model $(\T_\aa,f^\vee_\bb)$ described in §\[sssez:LG/Hyp\], and admitting the (generalized) KKP-compactification $(\XX_\aa,\overline{f}^\vee_\bb)$, described in §\[ssez:KKP-compct\]. In particular: $$(\overline{f}^\vee_\bb)^{-1}(0)=Y^\vee\ ,\quad(\overline{f}^\vee_\bb)^{-1}(\infty)=gD'_1+D'_2+gD'_3+D'_4=D'_\bb$$ Also in the present case, the ftv $(\XX_\aa,D'_\bb)$ can be thought of a log (no Calabi-Yau) pair, opening the door to an intrinsic mirror symmetric interpretation, in the sense of Gross-Siebert [@GS-IMS].
Checking if the LG mirror model here proposed and those proposed in [@KKOY], [@Seidel] and [@Efimov] are actually each other equivalent from the HMS point of view, is a completely open task!
[^1]: Moreover, after [@Hu-Keel Prop. 1.11], $\operatorname{Nef}(X)\cong\vf^*(\operatorname{Nef}(X))$ lives on the boundary of $\operatorname{Nef}(Y)$: actually Hu-Keel proved this fact when $X$ is projective, but this hypothesis is unnecessary (see e.g. [@R-wMDS Thm. 3.7] and references therein).
[^2]: Actually in [@CoxKatz Prop. 3.6.2] authors assume $X$ to be Gorenstein. Under this assumption $\D_{-K_X}$ is a lattice polytope, making easier to understand relation (\[aut\]). Anyway, by the previous result stated in [@CoxKatz Prop. 3.6.1], the Gorenstein assumption may be dropped in getting (\[aut\]).
[^3]: Artebani, Comparin and Guilbot asked for $\Q$-Fano toric varieties with *torsion free class group*, when presenting their generalization. Actually this hypothesis is unnecessary, as it was confirmed to me by Artebani (private communication).
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'literatur.bib'
---
[ **A critical evaluation of network and pathway based classifiers for outcome prediction in breast cancer**]{}
**[C Staiger$^{1, 2, *}$, S Cadot$^2$, R Kooter$^{3}$, M Dittrich$^{4}$, T Müller$^{4}$, GW Klau$^{1, 5, +, *}$, LFA Wessels$^{2, 3, 6, +, *}$ ]{}**
$^1$ Centrum Wiskunde & Informatica, Life Sciences Group, Science Park 123, 1098 XG Amsterdam, Netherlands\
$^2$ Bioinformatics and Statistics, The Netherlands Cancer Institute, Plesmanlaan 121, 1066 CX Amsterdam, Netherlands\
$^3$ Delft Bioinformatics Lab, Faculty of Electrical Engineering, Mathematics and Computer Science, 2600 GA Delft, Netherlands\
$^4$ Department of Bioinformatics, Biocenter, Am Hubland, 97074 University of Würzburg, Germany\
$^5$ Netherlands Institute for Systems Biology, Amsterdam, Netherlands\
$^6$ Cancer Systems Biology Center, The Netherlands Cancer Institute, Plesmanlaan 121, 1066 CX Amsterdam, Netherlands\
$^\ast$ E-mail: [email protected], [email protected], [email protected]\
$^+$ shared last authorship
Abstract {#abstract .unnumbered}
========
Recently, several classifiers that combine primary tumor data, like gene expression data, and secondary data sources, such as protein-protein interaction networks, have been proposed for predicting outcome in breast cancer. In these approaches, new composite features are typically constructed by aggregating the expression levels of several genes. The secondary data sources are employed to guide this aggregation. Although many studies claim that these approaches improve classification performance over single gene classifiers, the gain in performance is difficult to assess. This stems mainly from the fact that different breast cancer data sets and validation procedures are employed to assess the performance. Here we address these issues by employing a large cohort of six breast cancer data sets as benchmark set and by performing an unbiased evaluation of the classification accuracies of the different approaches. Contrary to previous claims, we find that composite feature classifiers do not outperform simple single gene classifiers. We investigate the effect of (1) the number of selected features; (2) the specific gene set from which features are selected; (3) the size of the training set and (4) the heterogeneity of the data set on the performance of composite feature and single gene classifiers. Strikingly, we find that randomization of secondary data sources, which destroys all biological information in these sources, does not result in a deterioration in performance of composite feature classifiers. Finally, we show that when a proper correction for gene set size is performed, the stability of single gene sets is similar to the stability of composite feature sets. Based on these results there is currently no reason to prefer prognostic classifiers based on composite features over single gene classifiers for predicting outcome in breast cancer. Supplementary data can be downloaded from <http://homepages.cwi.nl/~staiger/supplement.pdf> .
Introduction {#introduction .unnumbered}
============
Modern high-throughput methods provide the means to observe genome wide changes in gene expression patterns in breast cancer samples. Gene expression signatures have been proposed [@Veer2002; @Wang2005] to predict prognosis in breast cancer patients, but were shown to vary substantially between data sets. One possible explanation for this effect is that the data sets on which the predictors are trained are typically poorly dimensioned, consisting of many more genes than samples. Integrating secondary data sources like, for example, protein-protein interaction (PPI) networks, co-expression networks or pathways from databases such as KEGG, has recently been proposed to overcome variability of prognostic signatures and to increase their prognostic performance [@Chuang2007; @Lee2008; @Taylor2009; @Ma2010; @Abraham2010]. Many of these studies claim that combining gene expression data with secondary data sources to construct composite features results in higher accuracy in outcome prediction and higher stability of the obtained signatures. In addition, inclusion of the secondary sources raises the hope that the obtained signatures will be more interpretable and thus provide more insight into the molecular mechanisms governing survival in breast cancer.
The underlying idea of these methods is that genes do not act in isolation, and that complex diseases such as cancer are actually caused by the deregulation of complete processes or pathways, representing ‘hallmarks of cancer’ [@Hanahan2011]. This is unlikely to happen due to an aberration in a single gene, and often multiple genes need to be perturbed to disable a process. This leads to the notion that aggregating gene expression of functionally linked genes smooths out noise and provides more power to detect deregulation of complete functional units and hence to obtain a clearer picture of the biological process underlying tumorigenesis and disease outcome.
The observed improvement in classification accuracy achieved by the approaches employing secondary data is hard to assess since it is dependent on many factors such as the specific data sets and evaluation protocol employed. To shed more light on this issue we performed an extensive comparison of a simple, single gene based classifier with three of the most popular approaches that include secondary data sources in the construction of the classifier. More specifically, we included the approaches proposed by Chuang [*et al.*]{}[@Chuang2007], Lee [*et al.*]{}[@Lee2008] and Taylor [*et al.*]{}[@Taylor2009]. We investigated how these methods perform with respect to classification accuracy and stability of the set of features included in the classifiers. We will now briefly outline how the approaches work and point out some of the claims made by the authors. Detailed descriptions are provided in the Methods section.
Chuang [*et al.*]{}[@Chuang2007] describe a greedy search algorithm on PPI networks. For each defined subnetwork, a composite feature is defined as a variant of the average of the expression values of the genes included in the subnetwork. The score that guides the search is the association of the composite feature with patient outcome. Significance testing and a feature selection step are employed to select the set of composite features employed in the final classifier. The authors claim that classification based on subnetwork markers improves prediction performance on two breast cancer data sets. Moreover, they state that subnetwork markers are more reproducible across different breast cancer studies than single gene markers.
Lee [*et al.*]{}[@Lee2008] employ gene sets from the MsigDB database [@Subramanian2005] as secondary data source. The association of the composite feature with patient outcome is used as performance criterion, and a greedy search is employed to select a subset of genes from a gene set to constitute the composite feature. The value of the composite feature is derived from the expression values of the subset of genes as defined in Chuang [*et al.*]{} [@Chuang2007]. In contrast to Chuang [*et al.*]{}, Lee [*et al.*]{}do not exploit the connectivity of the pathway in the construction of the composite features. Lee [*et al.*]{}claim that by using these pathway activities a higher classification performance can be achieved on different cancer types, most notably leukemia, lung and breast cancer. They also report a higher overlap between genes in the top scoring composite features compared to the top scoring single genes.
The underlying assumption in the study by Taylor [*et al.*]{}[@Taylor2009] is that disease-causing perturbations disturb the organization of the interactome, which then has an effect on outcome. They concentrate on highly connected proteins, so-called hubs, as these proteins act as organizers in the molecular network. In contrast to Lee [*et al.*]{}and Chuang [*et al.*]{}, Taylor [*et al.*]{}detect aberrations in the correlation structure between a hub and its immediate interactors. As correlation between two genes cannot be assessed for a single sample, Taylor [*et al.*]{}employ the pairwise expression difference between the hub and each of its interactors as features for the classifier. While no claims are made regarding performance improvements, we included this approach in the comparison as it is a recently proposed, novel approach for exploiting secondary data sources to predict outcome in breast cancer.
Table \[tab:methods\_overview\] provides a summary of the characteristics of all methods included in the comparison. It lists a description of each approach, the secondary data sources employed, the types of (composite) features and how the value of a (composite) feature is computed for a single tumor.
All three studies listed above use their own specific cross-validation (CV) protocol and evaluate their method on different (combinations of) data sets. This makes it hard to assess the improvement over other methods. In this work, we therefore employ an unbiased training and validation protocol and present a comprehensive evaluation of cross data set classification performance and stability on six publicly available breast cancer data sets. Given that these classifiers are intended to predict the unknown outcome of a patient, we suggest a cross-validation procedure that does not assume any knowledge about the samples used for testing. Thus, we strictly separate the training data set from the test data set, [*i.e.*]{}composite feature construction, the selection of the optimized number of features for classification and the training of the final classifier are all performed on the training data set, while the testing of this trained classifier is performed on a completely separate test set without calibrating the classifier on the test data. See Figure \[fig:cv\] and Algorithm \[algo:cv\] for details. In other words, in contrast to previous studies, we strictly distinguish between training and test data.
To prevent biases associated with a specific secondary data source, we tested the algorithms on different types of secondary data sources. (See the Materials and Methods section for detailed descriptions of all these data sources.) We also used two different classifier types, the nearest mean classifier (NMC) and logistic regression (LOG) to evaluate the influence of the classifier on prediction performance. We chose these classifiers since Popovici [*et al.*]{}[@Popovici2010] confirmed earlier findings that these classifiers performed best on various breast cancer related classification tasks. Similarly, different feature extraction strategies were employed. While the included set of feature extraction approaches is by no means exhaustive, we employed approaches that were shown to perform well on gene expression based diagnostic problems [@Wessels2005]. All evaluations were performed on a curated collection consisting of six breast cancer cohorts [@Reyal2008] including the cohort from the Netherlands Cancer Institute[@Vijver2002].
In contrast to earlier findings we find that when we apply a proper correction for the number of genes appearing in the composite features employed by the composite feature classifier, the stability of single gene feature sets is comparable to the stability of composite feature sets. Much to our surprise, and in contrast to other studies, we also find that integrating secondary data, as done in the evaluated methods, does *not* lead to increased classification accuracy when compared to simple single gene based methods. Our findings are partly consistent with the findings of Abraham [*et al.*]{}[@Abraham2010], where the authors show that averaging over gene sets does not increase the prediction performance over a single gene classifier.
We investigated several possible factors that may explain the disappointing performance of approaches incorporating secondary data. First, we looked into the effect of the way the number of features is selected. Second, we looked into the effect of the exact size and composition of the starting gene set. This factor could play a role since not all genes are included in secondary data sources, hence classifiers employing secondary data sources may be at a disadvantage compared to single gene classifiers that select the gene set from all genes on the chip. Third, we investigated the effect of sample size. Finally, we looked into the effect of heterogeneity of the data sets on classifier performances. We find that none of these factors change our general findings.
In addition to all these technical factors, we also investigated whether the biological information captured in the secondary data contributes to the classification performance of the composite feature classifiers. To our astonishment we found that composite classifiers constructed from 25 *randomized* versions of the secondary data sources performed comparably to composite classifiers trained on the original, non-randomized data.
We conclude that further research has to be done on finding effective ways to integrate secondary data sources in predictors of outcome in breast cancer. In order to facilitate this research, and to ensure a standardized and objective way of establishing improvements over state-of-the-art approaches, we make all the code, data sets and results employed in this comparison available for download and use upon request.
[**Method**]{} [**Symbol**]{} [**Description**]{} [**Secondary**]{} [**data**]{} [**Feature**]{} [**Feature value**]{}
---------------------------------- ----------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------- --------------------------------------- -------------------------------------------------------------------------
Single genes SG Calculates t-statistic between mRNA expression distributions of the two patient groups None Single gene mRNA expression of the gene
Lee [*et al.*]{}[@Lee2008] L,[*Lee*]{} Calculates for each pathway a set of genes with high t-statistic between the averaged gene expression and the two patient groups MsigDB, KEGG Subset of genes in a pathway Averaged mRNA expression of genes in set
Chuang [*et al.*]{}[@Chuang2007] C, [*Chuang*]{} Calculates subnetworks with high mutual information between the averaged gene expression of the genes in the subnetworks and the class labels of the two patient groups KEGG, HPRD, I2D, NetC Genes in a subnetwork Averaged mRNA expression of the genes in the network
Taylor [*et al.*]{}[@Taylor2009] T, [*Taylor*]{} Finds hub proteins that, given the two patient groups, show different Pearson correlation of the mRNA expression between the hub proteins and all of their direct interactors KEGG, HPRD, I2D, NetC Edge between a hub and its interactor Difference of mRNA expression of hub and its interactors (edge weights)
: Overview of evaluated feature extraction methods.
As secondary data sources we used the KEGG database [@Kanehisa2010] and the C2 data set of the MsigDB [@Subramanian2005], as PPI data we used the information from KEGG, HPRD [@Prasad2009] and the OPHID/I2D databases [@Brown2005]. In addition we used the PPI network published by Chuang [*et al.*]{}[@Chuang2007] (NetC). \[tab:methods\_overview\]
Results {#results .unnumbered}
=======
Current composite feature classifiers do not outperform single gene classifiers on six breast cancer data sets {#sec:SGbetter .unnumbered}
--------------------------------------------------------------------------------------------------------------
We compared the performance of a nearest-mean classifier (NMC) using single genes with a NMC employing feature extraction methods based on pathway and PPI data. The results are depicted in Figure \[fig:auc\_boxplot\]. For each combination of secondary data source and feature extraction approach, Figure \[fig:auc\_boxplot\]A shows the box plots of the area under the receiver-operator characteristics curve (AUC) values obtained for each pair of data sets – using one data set of the pair as training set and the other data set in the pair as test set. The feature extraction approaches are ranked in descending order based on the median AUC values. The box plots suggest that no composite classifier performs better than the single gene classifier. Indeed, testing whether the mean performance of the single gene classifier is different from the mean performance of any composite classifier reveals that there is no difference (null hypothesis can not be rejected) except for [*Taylor*]{}and [*Chuang*]{}-I2D, where the single gene classifier is clearly superior. See Table S1 for details. This fact is confirmed by the pairwise comparisons between all classifiers, see Figure \[fig:auc\_boxplot\]B. A green square means that the combination in the row won more frequently over the combination in the column across the data set pairs. The good performance of the single gene classifier is reflected by the fact that the bottom row does not contain a single red box. Also, the generally poor performance of *Taylor* is clearly reflected in the dark red rows associated with this approach.
We also provide the classification results for the LOG classifiers in Figure S1 and Table S2. In general, the performances are lower than for the NMC, with the best combination not even reaching an AUC of 0.7 while several NMC classifiers clearly exceed 0.7. Apart from [*Lee*]{}all composite LOG classifiers perform equally or even worse than the single gene LOG classifier. However, it should be noted that the performance of the LOG classifier is highly variable as a function of the number of included features—see Figures S2-S4. In addition, the training procedure does not converge for all feature values as is evident in the AUC vs. number of features curves that end abruptly. The high sensitivity to the number of features is most evident for the [*Taylor*]{}composite NMCs. Clearly, the LOG classifier as employed here (and as employed by Chuang [*et al.*]{}) requires additional regularization to ensure convergence across the whole range of feature values. Also in combination with this classifier, [*Taylor*]{}performed poorly. This together with the high computational burden associated with this method, prompted us to omit [*Taylor*]{}from the remaining analyses.
Based on the results presented in Figure \[fig:auc\_boxplot\], we conclude that on the six breast cancer data sets employed in this comparison, composite classifiers employing secondary data sources do not outperform single gene classifiers on the task of predicting outcome in breast cancer, provided that a robust single gene classifier is employed.
**A**\
![ For each combination of feature extraction method and secondary data source and each pair of data sets we obtained one AUC value resulting in 30 AUC values per combination. The number of features for each classifier was determined in the cross-validation procedure (CV-optimized). **A:** Each box plot shows the median, the $25\%$ and $75\%$ percentiles and the standard deviation of the 30 AUC values. Outliers are depicted by crosses. The boxes are sorted in descending order according to the median. **B:** This panel shows the result of pairwise comparisons between all combinations of feature extraction methods and secondary data sources. If, for a given combination of training and test data set, the AUC value of classifier $i$ is higher (lower) than the AUC value of classifier $j$ on the same test data set, it is counted as a win (loss) for classifier $i$. Element $(i,j)$ in the matrix represents the $\log_2$ ratio of wins to losses of method $i$ compared to method $j$. Green indicates an overall win, red an overall loss and white represents draws. The rows and columns are sorted as in Panel A. [**Abbreviations:**]{} SG: Single genes; C: [*Chuang*]{}; L: [*Lee*]{}and T: [*Taylor*]{}. []{data-label="fig:auc_boxplot"}](Plots/Boxplot_permuted_Experiment_01_NMCv1.pdf "fig:"){width="3.5in"}
**B**\
![ For each combination of feature extraction method and secondary data source and each pair of data sets we obtained one AUC value resulting in 30 AUC values per combination. The number of features for each classifier was determined in the cross-validation procedure (CV-optimized). **A:** Each box plot shows the median, the $25\%$ and $75\%$ percentiles and the standard deviation of the 30 AUC values. Outliers are depicted by crosses. The boxes are sorted in descending order according to the median. **B:** This panel shows the result of pairwise comparisons between all combinations of feature extraction methods and secondary data sources. If, for a given combination of training and test data set, the AUC value of classifier $i$ is higher (lower) than the AUC value of classifier $j$ on the same test data set, it is counted as a win (loss) for classifier $i$. Element $(i,j)$ in the matrix represents the $\log_2$ ratio of wins to losses of method $i$ compared to method $j$. Green indicates an overall win, red an overall loss and white represents draws. The rows and columns are sorted as in Panel A. [**Abbreviations:**]{} SG: Single genes; C: [*Chuang*]{}; L: [*Lee*]{}and T: [*Taylor*]{}. []{data-label="fig:auc_boxplot"}](Plots/Winchart_permuted_Experiment_01_NMCv1.pdf "fig:"){width="3.5in"}
Four hypotheses regarding the lack of observed performance differences {#four-hypotheses-regarding-the-lack-of-observed-performance-differences .unnumbered}
----------------------------------------------------------------------
Next we formulated a number of hypotheses that could explain why classifiers employing secondary data sources do not outperform single gene classifiers. These hypotheses relate to (1) the feature selection approach employed; (2) the starting set of genes employed in each of the approaches; (3) the effect of the training set size on performance and (4) the homogeneity of the data set employed. In the following sections we will investigate these hypotheses one by one.
### The number of selected features does not effect relative performances {#the-number-of-selected-features-does-not-effect-relative-performances .unnumbered}
In the cross-validation protocol that we proposed for unbiased performance evaluation and also employed in the comparison, we employ individual feature filtering to select an optimized number of features to employ in the classifier. While this approach is sub-optimal, we (Wessels [*et al.*]{}[@Wessels2005]) and others have shown that these simple approaches perform the best in predicting phenotypes based on gene expression data. However, we observed in the curves showing the AUC values as a function of the number of ranked features included in the classifier (Figures S2-S4) that the AUC values for the NMC are very stable across a large range of features for most approaches, and that the absolute maximal AUC value chosen during the feature selection routine might only marginally differ from the performance obtained with other feature values. For this reason, and since the selection of the optimized number of features introduces additional variability between the approaches, we decided to fix the number of features to 50, 100 and 150 for most approaches. We chose these values as they covered the feature ranges across which the performance remained stable in all approaches. The results for fixing the number of features to 50 are depicted in Figure \[fig:boxplot50\] while results for 100 and 150 features are presented in Figures S5 and S6. When accounting for multiple testing, no classifier using a fixed number of features performs significantly different from its counterpart using the number of features determined by cross-validation. See p-values of the pairwise Wilcoxon rank test in Tables S4, S6 and S8. As expected, there are only minor differences between the performance of classifiers when the number of features is restricted to 50, 100 and 150 (Tables S3, S5 and S7) with any significant differences favoring single genes. This confirms that the number of features is not a critical parameter. Based on these results, we can conclude that the number of selected features does not explain the observed differences between composite feature classifiers and single gene classifiers.
![ Comparison of the performance of the classifiers when the number of features is trained in the CV procedure (denoted as ‘CV-opt number of features’, same values as in Figure \[fig:auc\_boxplot\]) and when the 50 best scoring features (denoted as ‘50 best features’) are selected for classification. We cannot show the values for [*Chuang*]{}-KEGG, [*Taylor*]{}-KEGG or [*Taylor*]{}-HPRD since for some data sets, the number of significant composite features was lower than 50. Abbreviations of methods as in Figure \[fig:auc\_boxplot\]. []{data-label="fig:boxplot50"}](Plots/Boxplot_Experiment_01_50_features_NMCv1.pdf){width="5in"}
### Restricted gene sets are not detrimental to composite feature classifiers {#restricted-gene-sets-are-not-detrimental-to-composite-feature-classifiers .unnumbered}
We next hypothesized that the lack of difference in the performance between composite classifiers and single gene classifiers could be caused by the fact that the composite features are bound to the genes annotated in the secondary data while single gene classifiers can employ all genes on the microarray. To test this hypothesis we retrained the single gene classifier, but restricted the set of genes from which features for the final classifier could be selected to the genes that are present in the respective secondary data sources. The resulting classifiers are denoted by the secondary data source from which the gene set is derived, while the single gene classifier employing features from the whole microarray is denoted by *unrestr*. The results of this analysis are depicted in Figure \[fig:restricted\_genes\]A. There is significant difference in the performance of the classifiers employing genes annotated in the I2D, KEGG and MsigDB (Table S9). However, when accounting for multiple testing only the difference between *unrestr* and I2D remains significant. Moreover, as indicated earlier, the optimization of the number of features by cross-validation introduces significant variation in the number of features without resulting in large performance changes. To eliminate this source of variation from the comparison, we fixed, as before, the number of features to 50, 100 and 150 and repeated the comparisons. The results are depicted in Figures \[fig:restricted\_genes\]B and S7, and Tables S10-S12. We can only find significant differences between the unrestricted single gene classifier and KEGG when the 50 best features are selected and the I2D when employing the 150 best features. However, both of these differences disappear when multiple testing correction is performed. We therefore conclude that the starting gene set has a minor influence on the single gene classifiers. Hence we can reject the hypothesis that feature extraction approaches employing secondary data sources are put at a disadvantage since they can not exploit the full set of genes present on the array.
A\
![ We compared the performance of the single gene classifier trained on all genes present on the microarray ([*unrestr.*]{}) with the performance of single gene classifiers that only employ genes present in the secondary data sources. The resulting classifiers are indicated by the secondary data source whose gene set was employed to train the classifier. [**A:**]{} The number of single genes was determined during the cross-validation procedure; [**B:**]{} the 50 best scoring single genes were employed.[]{data-label="fig:restricted_genes"}](Plots/Boxplot_Experiment_04_restrictedSG.pdf "fig:"){width="\linewidth"}
B\
![ We compared the performance of the single gene classifier trained on all genes present on the microarray ([*unrestr.*]{}) with the performance of single gene classifiers that only employ genes present in the secondary data sources. The resulting classifiers are indicated by the secondary data source whose gene set was employed to train the classifier. [**A:**]{} The number of single genes was determined during the cross-validation procedure; [**B:**]{} the 50 best scoring single genes were employed.[]{data-label="fig:restricted_genes"}](Plots/Boxplot_Experiment_04_restrictedSG_100_NMCv1.pdf "fig:"){width="\linewidth"}
### Training set size has no significant effect on performance differences {#training-set-size-has-no-significant-effect-on-performance-differences .unnumbered}
A third possible factor that could explain the lack of performance difference between the composite feature classifiers and the single gene classifier is the size of the training set. Recall that in the cross-validation procedure we train on one data set and then test on another data set. We repeat this procedure for all possible pairs of data sets; excluding, of course, training and testing on the same data set (*paired setting*). We can, however, also follow an alternative scheme where we train on *all* data sets except the test set, the so-called *merged setting*. More specifically, in this setting four of the five Affymetrix data sets were merged to form a single training data set and the fifth data set was used as test set. Thus, we receive for each feature selection method five AUC values. This increases the size of the training set, and by comparing the results obtained in this setting with the results from the paired setting, we can investigate the effect of the training set size on the classifier performances.
Figure \[fig:boxplot\_pw\_vs\_merged\] depicts the results for the merged setting and the pairwise setting for the CV-optimized feature sets and when only the top 50 features are selected. (Note that, in contrast to the results in Figure \[fig:auc\_boxplot\], this pairwise setting only employs the Affymetrix datasets). The results for the top 100 and 150 features are similar, see Figure S8. Statistical testing shows that in the paired setting (Tables S13-S16) when the number of features is set to 150, [*Lee*]{}employing the MsigDB performs better than the single gene classifier. However, this difference disappears when correcting for multiple testing. More importantly, there are no significant differences between the performances of the single gene and composite feature classifiers in the merged setting (Tables S17-S20).
Hence, we can also reject the hypothesis that the lack of performance difference is due to the sizes of the employed training sets.
[c|c|c]{} & CV-opt features & 50 best features\
merged
& ![**Classification results for merged and paired setting**. In the merged setting one Affymetrix data set is set aside as test and the remaining four Affymetrix data sets are merged into a single data set. This is repeated until every one of the five data sets acted as a test set. [**Top row:**]{} Results for the merged setting. The red lines indicate the median. [**Bottom row:**]{} Only the five Affymetrix data sets were used in the paired setting.[]{data-label="fig:boxplot_pw_vs_merged"}](Plots/Dotplot_Experiment_08_AFFYmerged.pdf "fig:"){width="3in"} & ![**Classification results for merged and paired setting**. In the merged setting one Affymetrix data set is set aside as test and the remaining four Affymetrix data sets are merged into a single data set. This is repeated until every one of the five data sets acted as a test set. [**Top row:**]{} Results for the merged setting. The red lines indicate the median. [**Bottom row:**]{} Only the five Affymetrix data sets were used in the paired setting.[]{data-label="fig:boxplot_pw_vs_merged"}](Plots/Doxplot_Experiment_08_AFFYmerged_50_features.pdf "fig:"){width="3in"}\
pairwise
& ![**Classification results for merged and paired setting**. In the merged setting one Affymetrix data set is set aside as test and the remaining four Affymetrix data sets are merged into a single data set. This is repeated until every one of the five data sets acted as a test set. [**Top row:**]{} Results for the merged setting. The red lines indicate the median. [**Bottom row:**]{} Only the five Affymetrix data sets were used in the paired setting.[]{data-label="fig:boxplot_pw_vs_merged"}](Plots/Boxplot_Experiment_07_NMCv1.pdf "fig:"){width="3in"} & ![**Classification results for merged and paired setting**. In the merged setting one Affymetrix data set is set aside as test and the remaining four Affymetrix data sets are merged into a single data set. This is repeated until every one of the five data sets acted as a test set. [**Top row:**]{} Results for the merged setting. The red lines indicate the median. [**Bottom row:**]{} Only the five Affymetrix data sets were used in the paired setting.[]{data-label="fig:boxplot_pw_vs_merged"}](Plots/Boxplot_Experiment_07_50_features_NMCv1.pdf "fig:"){width="3in"}\
### Dataset homogeneity affects single genes and composite classifiers similarly {#dataset-homogeneity-affects-single-genes-and-composite-classifiers-similarly .unnumbered}
Breast cancer is a collection of several heterogeneous diseases that show very different gene expression patterns [@Gatza2010]. Expression patterns predictive of outcome might vary between subtypes, which typically leads to problems when training classifiers on gene expression data derived from breast tumors. If this is not explicitly taken into account during classifier training it could result in poor performance and unstable classification, as the selected genes may depend on the composition of the training set. In this section we control the heterogeneity in both the training and test sets by only selecting the relatively homogeneous ER positive breast cancer sub-population. Since the training sets become too small in the paired setting if we only select the ER positive cases, we followed the merged setting outlined above. More specifically, we created a test set consisting of all ER positive cases of a single data set and a training set by pooling all ER positive cases from the remaining data sets. We repeated this procedure across the six data sets and thus obtained six AUC values per classifier. Figure \[fig:ER\_pos\] depicts the results for the CV-optimized feature sets and the top 50 features. As before, the classifiers employing a fixed number of features perform similar to classifiers based on a feature set optimized in the CV procedure. See Figure \[fig:ER\_pos\]B and Figure S9. In general, and in accordance with earlier observations as made, [*e.g.*]{}, by Popovici [*et al.*]{}[@Popovici2010], the performance of all classifiers is substantially better on the ER positive cases compared to the unstratified case. More importantly, also in this setting there are no significant performance differences between the single gene classifiers and composite feature classifiers (Tables S21-S24).
$$$$$$$$
**A**\
![ The ER positive cases from a single data set were set aside as test set while ER positive cases from the remaining five data sets were merged into a single training set. This was repeated until each data set was employed as left-out test set, resulting in six AUC values. The red lines indicate the median. [**A**]{}: CV-optimized number of features; [**B**]{}: 50 best features.[]{data-label="fig:ER_pos"}](Plots/Dotplot_Experiment_09_ERpos.pdf "fig:"){width="3.5in"}
**B**\
![ The ER positive cases from a single data set were set aside as test set while ER positive cases from the remaining five data sets were merged into a single training set. This was repeated until each data set was employed as left-out test set, resulting in six AUC values. The red lines indicate the median. [**A**]{}: CV-optimized number of features; [**B**]{}: 50 best features.[]{data-label="fig:ER_pos"}](Plots/Dotplot_Experiment_09_ERpos_50_features.pdf "fig:"){width="3.5in"}
Equal classification using real or randomized networks and pathways {#equal-classification-using-real-or-randomized-networks-and-pathways .unnumbered}
-------------------------------------------------------------------
In the previous sections we showed that composite classifiers do not perform significantly better than classifiers employing single genes as features. We investigated several factors that could influence the performances of these two approaches, but failed to find any factor that induces significant performance differences on the data sets we employ in this study. This lead us to question whether prior knowledge sources really contain information that is of value in constructing features for classifiers predicting outcome in breast cancer. Chuang [*et al.*]{}[@Chuang2007] compared their PPI-based classifier to classifiers derived from randomized PPI networks. They concluded that their classifier performed significantly better than random classifiers. We decided to repeat this analysis for a subset of the classifiers in our comparison to determine whether prior knowledge sources really contain information relevant for predicting outcome in breast cancer. To this end, we generated, for each prior knowledge source, 25 random instances. More specifically, we maintained the structure of the pathways, networks and gene sets, and randomly permuted the identities of the genes. In doing so, the original topology of the secondary data is preserved while the biological information is destroyed. We then repeated the whole validation procedure on all 25 random instances for the feature extraction methods [*Lee*]{}and [*Chuang*]{}. The results of this analysis are presented in Figure \[fig:rand\_vs\_pws\]. Strikingly, classifiers derived from secondary data sources suffer no significant performance degradation when employing randomized secondary data sources. The performance of *Chuang* on randomized PPI data clearly has a large variance, and there are instances of classifiers derived from random networks that perform much worse and much better than classifiers derived from the non-randomized networks. Furthermore, we found that most classifiers based on randomized secondary data show performances similar to the classifiers derived from the real secondary sources. To formalize this observation, we performed a statistical test. We have reason to believe that the results derived from the real data should be better than the results derived from random data. Hence we performed one-sided paired Wilcoxon rank tests to determine whether the null hypothesis that the mean ‘real’ AUC-value is larger than each of the the 25 ‘randomized’ mean AUC-values can be rejected. We performed a Bonferroni correction to account for multiple testing. The results in Figures S10-S12 and Tables S25-S30 show that in the vast majority of the cases the null hypothesis can not be rejected. Conversely, it is very simple to generate a randomized secondary data source that performs equally well as the real data source. This result shows that further research must be done on the utility of secondary data sources in predicting breast cancer outcome.
![. [**Left:**]{} AUC values obtained with the feature extraction method *Lee* on real and randomized MsigDB pathways. [**Right:**]{} AUC values obtained with the feature extraction method *Chuang* on real and randomized PPI networks.[]{data-label="fig:rand_vs_pws"}](Plots/Boxplot_25rand_Experiment_05_RandomisedNetworks_L_MsigDB.pdf "fig:"){width="2in"} ![. [**Left:**]{} AUC values obtained with the feature extraction method *Lee* on real and randomized MsigDB pathways. [**Right:**]{} AUC values obtained with the feature extraction method *Chuang* on real and randomized PPI networks.[]{data-label="fig:rand_vs_pws"}](Plots/Boxplot_25rand_Experiment_05_RandomisedNetworks_C_NetC.pdf "fig:"){width="2in"}
Current composite feature classifiers do not increase the stability of gene markers {#current-composite-feature-classifiers-do-not-increase-the-stability-of-gene-markers .unnumbered}
-----------------------------------------------------------------------------------
Apart from performance improvements, it is also frequently claimed that features derived from classifiers employing secondary data sources are far more stable than single gene classifiers. In other words, whereas single gene signatures extracted from different data sets show very little overlap, features extracted by approaches that employ secondary data sources are claimed to show a large degree of overlap, even though the features were derived from separate data sets.
In this section we determined whether feature sets extracted from secondary data sources do, in fact, show a larger degree of stability than single gene feature sets. As a measure of stability we calculated the pairwise Jaccard index between the features derived from different data sets for a given feature extraction method. Since the number of features determined in the CV procedure varies, we performed this comparison for the cases where the top 50, 100 and 150 features are selected. The Jaccard indices for the best 50 features are depicted in Figures \[fig:overlap\] while Figure S13 depicts the results for the best 100 and 150 features. It is clear that the overlap of feature sets consisting of single genes is relatively low, albeit slightly higher than the overlap of [*Lee*]{}-MsigDB. The highest consistent stability is obtained by [*Chuang*]{}-HPRD with [*Chuang*]{}-KEGG showing high variance in stability. One can therefore clearly conclude that, when compared in terms of *signature genes* overlap, single genes are generally less stable than feature sets extracted by including secondary data sources. However, strictly speaking, such a comparison compares the proverbial apples and oranges, since a single feature constructed based on secondary data sources can contain many genes. In order to ensure that the low overlap of single genes is not only due to the fact that the best single gene feature sets contain fewer genes than the other feature sets, we controlled the single gene feature sets for size. More specifically, for each data set and each feature selection approach employing secondary data sources, we obtain a single best feature set consisting of $n^*$ features (networks, gene sets or pathways) that, in turn, consists of $m$ genes. We then determine a size-matched single gene set by choosing the best $m$ single genes on that same expression data set. We also required these single genes to be annotated in the respective secondary data source. For these size-matched single gene feature sets, we computed the Jaccard index. The results, depicted in Figure \[fig:overlap\_featuresNMC\_cfs\] and S14, show that when this size correction is applied, the stability of single gene feature sets are as high as features extracted by employing secondary data sources. See also Tables S31-S33.
![ For each method the Jaccard index was calculated between the gene sets extracted from two different data sets. This was repeated for all pairwise combinations of data sets. Thus, 15 values were obtained.[]{data-label="fig:overlap"}](Plots/Boxplot_Experiment_11_Overlap_50.pdf){width="3.5in"}
![. Box plots of the Jaccard indices computed for all pairs of gene sets derived from two different data sets. The green box plots represent the Jaccard indices for genes constituting composite features, while the blue box plots (denoted as ‘Control for size SG’) represent the gene-size-corrected Jaccard indices for single gene classifiers. The white stars represent the mean of the distributions.[]{data-label="fig:overlap_featuresNMC_cfs"}](Plots/Boxplot_Experiment_11_Overlap_contrlSize_restricted_50.pdf){width="5in"}
Discussion {#discussion .unnumbered}
==========
In this study we evaluated the prediction performance of network and pathway-based features on six breast cancer data sets. In contrast to previous studies we found that none of the classifiers employing composite features derived from secondary data sources can outperform a simple single gene classifier. Moreover, we did not find any evidence that composite features show a higher stability across the six breast cancer cohorts. Our findings suggest that with the feature extraction methods tested in this study, we cannot extract more knowledge from secondary data sources than we find in the expression of single genes.
There are several issues that could potentially contribute to that situation. First, secondary data sources are, to a large degree, generated by high-throughput biological experiments and could thus contain a level of noise that deems them inappropriate for outcome prediction in breast cancer. On the other hand, the search algorithms could be unsuitable to detect biologically meaningful networks. All three feature extraction methods only extract local information without taking into account the full structure of the network or pathway data. This local information is then combined in the classifiers in a rather crude way, namely by simply averaging the expression of the genes associated with a feature that was found to be associated with outcome, [*i.e.*]{}treating each single sub-network or sub-pathway as a single dimension in the classification space. Possible dependencies between features are not taken into account. Also, exploring the subnetwork search space in a heuristic manner may decrease classification performance. The recent method by Dittrich [*et al.*]{}[@Dittrich2008] computes provably maximally deregulated connected subnetworks based on a sound statistical score definition. This method has not yet been used for classification.
Other recent algorithms as presented by Ulitsky [*et al.*]{}[@Ulitsky2010], Chowdhury [*et al.*]{}[@Chowdhury2011] and Dao [*et al.*]{}[@Dao:2010kt] search for deregulated subnetworks in subsets of samples. These subnetworks are sets of genes that are deregulated in most, but not necessarily all, patients with poor disease outcome. The heuristic method by Chowdhury [*et al.*]{}[@Chowdhury2011] has been shown to perform well on cross-platform classification of colorectal cancer outcome. Dao [*et al.*]{}[@Dao:2010kt] improved on these results by exact enumeration of all dense subnetworks with the above-mentioned property. Looking at subsets of samples in a class, [*i.e.*]{}a subset of the poor outcome samples, is an interesting aspect for further evaluation, in particular for breast cancer outcome prediction as it may quite accurately capture the large phenotypic variety of this rather inhomogeneous disease.
Lee [*et al.*]{}[@Lee2008] and Chuang [*et al.*]{}[@Chuang2007] average the expression values of single genes comprising a subnetwork, to determine the ‘activity’ of the subnetwork. This is, however, a very simplistic view of the dynamics in a subnetwork itself. In contrast to these two algorithms, the method by Taylor [*et al.*]{}[@Taylor2009] predicts outcome by trying to capture the disruption of the regulation of a hub protein over its interactors in poor outcome patients. This is implemented by using every edge leading to a hub as a feature in classification space. Yet, in this way, the classification space becomes too large to allow for good classification results. This method is thus not appropriate for solving the classification problem and this is clearly demonstrated in the poor performance of this algorithm in the comparison.
To find a subnetwork scoring function remains one of the biggest problems when including promising gene sets into a classification framework. Abraham [*et al.*]{}[@Abraham2010] tested the classification performance of gene sets provided by the MsigDB. The authors employed several set statistics like mean, median and first principle component to score the gene sets. They found that none of the classifiers employing gene sets and scoring them with the set statistics performed better than a single gene classifier.
In our experiments where we shuffle the genes in the secondary data we showed that features determined on this nonsensical biological data perform equally well in classification than features determined on the real secondary data. This again could possibly be caused by the low quality of the network and pathway data. However, the nature of gene expression patterns in breast cancer, and specifically its association with outcome can also explain these findings. Since many genes are involved in breast cancer and are differentially expressed and associated with outcome, as shown, for example by Ein-Dor [*et al.*]{}[@Ein-Dor2005], the chance that those genes span decently sized subnetworks in the randomized secondary data is high. Both algorithms, [*Chuang*]{}and [*Lee*]{}, look for highly differentially expressed subnetwork or pathway markers, and these can also be found in the randomized data. Furthermore, overlaying networks or pathways that contain protein level information with mRNA expression data might result in erroneous results. These data sources reflect events on very different molecular levels. While gene expression and protein expression is undoubtedly linked, there are many processes that prevent this from being a trivial one-to-one mapping. Thus, we may measure, for a set of genes, an effect on the mRNA level that leads to differential expression between the two patient classes but this may have little bearing on the relationships between these genes as captured in the PPI graph. In conclusion, our results show that it may not be sufficient to search for sub-networks or sub-pathways that are differentially expressed *on average* but that complex interactions between entities as well as the more complicated relationships between mRNA levels and the topology of the PPI graph need to be taken into account. Our different classification results are partially owed to the fact that we used a different cross-validation procedure, which, in our opinion, fits the clinical situation, better. The studies by Chuang [*et al.*]{}[@Chuang2007] and Lee [*et al.*]{}[@Lee2008] also determined possible features on one data set. However, in contrast to our work they reranked the features on the second (test) data set. Furthermore, they determined the number of features and the classification performance on this second data set using five-fold cross-validation. In our opinion, this does not represent a *bona fide* independent validation of the classifier.
In summary, we introduced a framework to test the use of feature extraction methods with respect to the prediction of their determined features. We used this framework to specifically test the superiority of feature extraction methods based on network and pathway data over classifiers employing single genes. Across six breast cancer cohorts, we showed that the three tested methods do not outperform the single gene classifier nor do they provide more stable gene signatures for breast cancer.
An important aspect that hampers progress in the field of network and pathway based classification is the lack of proper evaluation of proposed algorithms. In our opinion this is caused by (1) lack of reproducibility of the results; (2) lack of large and standardized benchmark sets to test proposed algorithms and (3) lack of a standardized, unbiased protocol to assess the performances of proposed methods on the benchmark sets. To overcome these issues, we have created a software pipeline that implements all the classifiers as faithfully as possible and also runs our validation protocol. We have also established a large collection of breast cancer datasets (and this is currently being expanded) on which the algorithms can be tested. Both the datasets and the pipeline are freely available. In the long term, we envision a web service where a classifier can be submitted as a software package. The server will then autonomously evaluate the performance of the classifier using the standardized pipeline on the benchmark set.
Materials and Methods {#materials-and-methods .unnumbered}
=====================
Microarray data {#microarray-data .unnumbered}
---------------
The microarray data sets used in this work is listed in Table \[tab:data\_sets\]. To combine the five Affymetrix arrays with the Agilent arrays we first matched the probes on the arrays to Entrez GeneIDs. Only those genes were included in the feature sets that appeared on both platforms, resulting in 11601 genes in total. In case that several probes on one chip matched to the same gene the expression values of the probe with the highest variance was taken. The final expression matrices were then z-normalized such that the expression distribution of each gene has a mean of zero and a standard deviation of one. Samples in the data sets were labeled ‘good’ outcome if no event, that is, a distant metastasis or death, occurred within five years. Otherwise samples were labeled ‘poor’ outcome.
Dataset Outcome poor/good platform
------------------------ --------------------------- ------------ ------------
Chin [@Chin2006] Metastasis within 5 years 68/29 Affymetrix
Desmedt [@Desmedt2007] Metastasis within 5 years 91/29 Affymetrix
Loi [@Loi2007] Metastasis within 5 years 92/28 Affymetrix
Miller [@Miller2005] Death within 5 years 156/37 Affymetrix
Pawitan [@Pawitan2005] Death within 5 years 120/22 Affymetrix
Vijver [@Vijver2002] Metastasis within 5 years 178/70 Agilent
: **[Microarray expression data]{}**
All data sets were processed as described in [@Vliet2008] and contained 11601 genes with z-normalized expression values afterwards. Column ‘poor/good’ contains the number of samples with poor or good outcome, respectively.
\[tab:data\_sets\]
ER status of patients was predicted from the expression values of the gene ESR1. For more detail of the processing of the data see van Vliet [*et al.*]{}[@Vliet2008].
Network and pathway data {#network-and-pathway-data .unnumbered}
------------------------
All feature extraction methods were run on the databases KEGG [@Kanehisa2010] and HPRD [@Mishra2006]. The algorithm by Lee [*et al.*]{}[@Lee2008] was also run on the MsigDB C2 database [@Subramanian2005].
### KEGG {#kegg .unnumbered}
We collected all pathway information available for *Homo sapiens* (hsa) from the KEGG database, version December 2010. The entries contained information on metabolic pathways, pathways involved in genetic information processing, signal transduction in environmental information processing, cellular processes and pathways active in human disease and drug development. We obtained 215 pathways. The nodes contained in the pathways were matched with the KEGG gene database such that each node carries an Entrez GeneID. In this way we obtained a network composed of 200 pathways containing 4066 nodes and 29972 interactions of which 3110 nodes are also contained in the expression sets.
### MsigDB {#msigdb .unnumbered}
As second pathway database we used the C2 collection of the MsigDB (version 3) [@Subramanian2005], which was also used in Lee [*et al.*]{}[@Lee2008] (version 1.0). It contains gene sets from online pathway databases such as KEGG, gene sets made available in scientific publications and expert knowledge. We obtained 3272 gene sets of which 2714 could be entirely or partially mapped the six data sets. The MsigDB does not contain any edges, thus this database was only usable for the algorithm by Lee [*et al.*]{}[@Lee2008].
### HPRD {#hprd .unnumbered}
The HPRD (version 9) provides information on protein-protein interactions (PPI) derived from the literature. The HPRD contains 9231 proteins and 35853 interactions. The proteins were mapped to their genes carrying Entrez GeneIDs. There are 7390 genes contained in both the HPRD and the expression sets.
### OPHID/I2D {#ophidi2d .unnumbered}
The OPHID/I2D database, downloaded in April 2011, contains protein-protein interactions derived from BIND, HPRD and MINT as well as predicted interactions from yeast, mouse and *C. elegans*. The database contains 12643 nodes and 142309 edges. 9453 of the nodes are also present in the six breast cancer studies examined here.
### Protein-protein interaction network by Chuang [*et al.*]{}[@Chuang2007] (NetC) {#protein-protein-interaction-network-by-chuang-et-al.-netc .unnumbered}
The network curated by Chuang [*et al.*]{}[@Chuang2007] consists of 57228 interactions and 11203 nodes of which 8572 are contained in the six breast cancer studies. The network is curated from yeast two hybrid experiments and interactions predicted from co-citation.
Algorithms {#algorithms .unnumbered}
----------
### Notation {#notation .unnumbered}
Let $E_{k\times n}$ be a gene expression matrix, as we obtain it from a microarray study, with $k$ samples and $n$ genes. Each entry $e_{i, j}$ contains the expression value of gene $j$ in sample $i$. All samples carry a binary class label $l_i \in \{0, 1\}$ denoting outcome, where 1 denotes ‘poor outcome’ and 0 denotes good outcome’. The label vector of all samples is denoted by $L = (l_1, \dots l_k)^T$. We denote a network by $N =(G, I)$ where $G$ is the set of genes in the network and $I$ is the set of interactions between these genes, also called edges in the following. We define a subnetwork as the connected graph $N^\prime = (G^\prime, I^\prime)$ with $G^\prime \subseteq G$ and $I^\prime \subseteq I$, and a pathway as a gene set $G^\prime \subseteq G$. Let $G^\prime$ be such a pathway or the set of genes in a subnetwork then according to [@Chuang2007; @Lee2008] the activity of the pathway or subnetwork in sample $i$ is given as $$\label{activity}
a_{G^\prime, i} = \sum_{j \in G^\prime} \frac{e_{i, j}}{\sqrt{|G^\prime|}}$$
### Feature extraction method by Chuang [*et al.*]{}[@Chuang2007] {#feature-extraction-method-by-chuang-et-al. .unnumbered}
Given a network $N = (G, I)$, the algorithm by Chuang [*et al.*]{}[@Chuang2007] carries out a greedy search starting from a seed—a single gene in the network. It then iteratively extends the network by adding neighboring genes to find subnetworks with high mutual information (MI) of the activity of the pathway and sample labels. Each node in $N$ is used once as seed. In each step, an additional gene is picked that leads to a maximal MI improvement. If no improvement is possible, the search stops.
More precisely, the association between the subnetwork activity and the class labels is computed as follows: Given a subnetwork $N^\prime = (G^\prime, I^\prime)$ the activity vector $a_{G^\prime} = (a_{G^\prime, 1}, \dots, a_{G^\prime, k})^T$ is calculated using Equation (\[activity\]). To compute the MI, vector $a_{G^\prime}$ is discretized. Given a dissection of the interval $[\min_{i \in [1, \dots, k]} a_{G^\prime, i}, \max_{i \in [1, \dots, k]}
a_{G^\prime, i}]$ into $\lceil\log_2 k\rceil$ bins let be a function that assigns a network activity to a sample with one of these bins, where $\delta_i$, $i = 1, \dots, \lceil\log_2 k\rceil$, denote the bins. We define the mutual information $s_{\mathit{MI}}$ between the probability density of the bins $p(\delta(\cdot))$ and the probability distribution of the class labels $p(l)$ as
$$s_{MI}(\delta, l) = \sum_{i = 1}^{\lceil\log k\rceil} \sum_{l = 0}^1 p(\delta_i, l)\log\frac{p(\delta_i, l)}{p(\delta_i)p(l)}$$
where $p(\delta_i, l)$ is the joint distribution of $p(\delta(\cdot))$ and $p(l)$. The algorithm also performs three statistical tests to extract only networks that show significantly high mutual information. The ranking of the networks is given by ordering the networks according to their mutual information $s_{\mathit{MI}}$.
In our study we use PinnacleZ, an implementation of the algorithm provided by the authors. As feature values for classification the subnetwork activity as given in Equation (\[activity\]) of the determined subnetworks was used.
Before determining the subnetworks, PinnacleZ performs a z-normalization of the given data set. This is undesirable when looking at subsets of data sets as we do in the five-fold cross-validation. In order to skip the normalization step, we implemented a patch in the PinnacleZ source code. This patch adds a “-z” option that instructs PinnacleZ to *omit* its usual gene-wise z-normalization step.
One problem when mapping the expression data to the network data is that for some nodes there is no expression data. Chuang [*et al.*]{}[@Chuang2007] do not state in their paper how they handled this problem although their identified subnetworks contain such nodes. We therefore filtered out proteins for which no expression data is available before running PinnacleZ. For further issues we encountered when working with PinnacleZ see the supplementary information.
### Algorithm by Lee [*et al.*]{}[@Lee2008] {#algorithm-by-lee-et-al. .unnumbered}
The algorithm by Lee [*et al.*]{}[@Lee2008] uses the t-statistic to rank pathways according to their overall differential expression. For this it first defines sets of genes, the so called condition responsive genes (CORGs), which contain the most differentially expressed genes of a pathway. These genes are found by applying a greedy search. For each pathway the genes are ordered according to their t-statistics. Given the two sample groups let $E^1_j$ be the expression values of gene $j$ for all samples with class label $1$ and $E^0_j$ the expression values of gene $j$ for all samples with class label $0$, respectively. Let $k^0$ and $k^1$ denote the number of samples in each group; $\mu_{E^0_j}$ and $\mu_{E^1_j}$ denote the means of the two groups and $\sigma_{E^0_j}$ and $\sigma_{E^1_j}$ the standard deviation in the two groups. The t-statistics between $E^1_j$ and $E^0_j$ is given by
$$\label{t-test_eq}
t(j) = \frac{\mu_{E^1_j}-\mu_{E^0_j}}{\sqrt{\sigma_{E^1_j}/k^1+\sigma_{E^0_j}/k^0}}$$
The genes in a pathway are sorted either in ascending order, if the highest absolute $t$ value is negative, or in descending order, if the highest absolute $t$ value is positive. Given this order the greedy search iteratively combines genes and calculates their average expression, or *pathway activity*, across the samples as it is given in Equation (\[activity\]), [*i.e.*]{}$a_{G^\prime} = (a_{G^\prime, 1}, \dots, a_{G^\prime, k})^T$ is calculated where $G^\prime$ contains the best $m$ genes according to the ranking. To evaluate the combined discriminative power of the genes that have been averaged, the t-statistic of the averaged expression is computed as follows: $$\label{t-test_activity}
s_{G^\prime} =
t(G^\prime) =
\frac{\mu_{a^1_{G^\prime}}-\mu_{a^0_{G^\prime}}}{\sqrt{\sigma_{a^1_{G^\prime}}/k^1+\sigma_{a^0_{G^\prime}}/k^0}}$$ where $\mu_{a^0_{G^\prime}}$ and $\mu_{a^1_{G^\prime}}$ represent the means and $\sigma_{a^0_{G^\prime}}$ and $\sigma_{a^1_{G^\prime}}$ represent the standard deviations of the averaged activities. If the resulting value $t(G^\prime)$ is higher than the previous value of the t-statistics then the search continues adding the gene to the already determined CORGs, otherwise the search stops. The score $s_{G^\prime}$ of the final CORGs $G'$ is then used to rank the pathway.
As mentioned beforehand, [*Lee*]{}can only be executed on predefined gene sets. Those gene sets are normally not provided in a PPI database. Thus, we used the KEGG and MsigDB databases to evaluate this algorithm. In order to decrease the running time the authors executed a pathway ranking by employing the algorithm by Tian [*et al.*]{} [@Tian2005] and just taking the top 10% of pathways into account prior to executing their algorithm. In our setting we ranked all of the pathways according to the algorithm by Lee [*et al.*]{}[@Lee2008] and considered for determining the optimized number of features in the final classifier all pathways in KEGG and the top 400 pathways in MsigDB. As feature values for classification the pathway activity, as computed according to Equation (\[activity\]) for all condition-responsive genes (CORGs), is employed. Here again we excluded proteins in the pathways for which no expression data is available.
### Algorithm by Taylor [*et al.*]{}[@Taylor2009] {#algorithm-by-taylor-et-al. .unnumbered}
In contrast to the algorithms by Chuang [*et al.*]{}[@Chuang2007] and Lee [*et al.*]{}[@Lee2008], the algorithm by Taylor et al. [@Taylor2009] first identifies organizer proteins in the network, the so-called hubs, and then attaches a weight to each edge between the hubs and their direct neighbors in the network. These weights are later used to train a classifier.
Candidate hubs are the 15% most densely connected proteins in the network data, independent from their expression status. For the following calculations proteins without expression data are excluded. To identify hubs that significantly change their interaction behaviour between the two classes the authors introduce the hub difference and the average hub difference which are based on the Pearson correlation. The Pearson correlation between a hub $h$ and an interactor $n$ of this hub is defined as the Pearson correlation between their expression profiles across the $k$ samples $$P(h, n) = \frac{\sum^{k}_{i = 1}(e_{n, i} - \mu_{E_n})(e_{i, h} - \mu_{E_h})}{(k - 1)\sigma_{E_n}\sigma_{E_h}}\enspace.$$
$E_n$ and $E_h$ denote the distribution of expression values across the $k$ samples and $\mu$ and $\sigma$ are their means and standard deviations. The hub difference is defined as the difference of the Pearson correlation $P(h, n)$ given the two sample classes, indicated by the superscript $0$ and $1$,$$\label{hubdiff}
\mathit{d}(h, n) = P^0(h, n)-P^1(h, n)\enspace .$$
Let $\delta(h)$ denote the set of neighbors of a hub $h$ then the average hub difference is$$\mathit{\overline{d}}(h) = \frac{\sum_{n\in \delta(h)} |\mathit{d}(h,n)|}{|\delta(h)|}\enspace .$$ To extract only those hubs that show a significant average hub difference the value is compared to a distribution of the average hub difference for a permuted dataset, using a p-value cut off of $0.05$. This distribution is calculated by 1) randomly permuting the class labels and 2) recalculating the average hub difference and repeating these two steps 1000 times. The significant hubs are ranked by their average hub difference.
As feature values in the classifier differences of the expression of the hub and each of its interactors were used. For example, suppose $h_1, \dots, h_p$ were found significant and suppose ${(h_1, i_{1, 1}), \dots, (h_1, i_{1, m_1}), \dots, (h_p, i_{p, 1}), \dots, (h_p, i_{p, m_p})}$ are the edges to their interactors. Then for one sample the vector $(e_{h_1} - e_{i_{1, 1}}, \dots, e_{h_1} - e_{i_{1, m_1}}, \dots, e_{h_p} - e_{i_{p, 1}}, \dots, e_{h_p} - e_{i_{p, m_p}})$ contains the feature values for the classifier.
Since the edges attached to a hub are not ranked, all those edges were included in the classifier, given that the hub shows a significant average hub difference. For the cross-validation procedure this means that we can not train the number of features but only a number of feature sets.
### Classifiers {#classifiers .unnumbered}
In our study we employed a nearest mean classifier (NMC) and logistic regression (LOG). As distance metric for the NMC we employed the Euclidean distance. More specifically, a sample is projected on the line connecting the two class means, and the Euclidean distance of he projected sample to each class mean is computed. The sample is assigned to the closest class. In addition to the NMC we executed all features extraction methods in combination with the LOG. We found that simple LOG without any regularization parameters cannot be trained properly since for higher numbers of features (approximately 50 features and more) the training step does not converge on the breast cancer data. Moreover, we found that for many features different implementations of LOG return different weight vectors. Thus, we used three different implementations (the R GLM, R NNET and Python SciKits implementation) and only accepted the classification result of the R GLM implementation when all three versions converged to the same weight vector.
### Cross-validation and classification {#cross-validation-and-classification .unnumbered}
In the cross-validation procedure we employed, we rigorously separate the training and test data sets. For details, see Figure \[fig:cv\] and Algorithm \[algo:cv\]. The training phase consists of determining the best performing number of features and training the final classifier with this number of features. The trained classifier is then tested in the test phase. The data sets used in these two phases are completely independent, [*i.e.*]{}the test set is never used in training the classifier.
To determine the optimized number of features in our classifiers we employed five fold cross-validation. In this cross-validation, we first determined all required composite features (if necessary) and their ranking on four splits of the training data set. Then a series of classifiers is trained on the same four splits by gradually adding features according to the ranking. These classifiers are then evaluated on the fifth split of the data set. Since this is done in a five fold cross-validation we obtain for each of the classifiers five different AUC values. The optimized number of features extracted corresponds to the number of features yielding the highest mean performance. Once the best performing number of features is determined, the features are calculated using the whole training data set and the final classifier is trained also using the complete training data set. The classifier is then tested on the test data set. For each method we used all possible pairs of data sets as training and test set respectively. Since we have six data sets available this resulted in 30 AUC values for each method.
![[]{data-label="fig:cv"}](pipeline1_crop){width="9cm"}
\
$(d_1, d_2, d_3, d_4, d_5) \leftarrow D_1$ $\mathit{Val} \leftarrow d_s$ $\mathit{Train} \leftarrow D_1 \setminus d_s$ $\mathit{rankedF} \leftarrow m(\mathit{Train}, N)$ $n \leftarrow $ number of $\mathit{rankedF}$ $\mathit{AUC}_s$ = empty list $C_{s, i} \leftarrow \mathit{trainCl}(C,
\mathit{rankedF}[1, \dots, i],
\mathit{Train})$ $\hat p \leftarrow C_{s, i}(e_{\mathit{Val}})$ $\mathit{AUC}_{s, i} \leftarrow \mathit{calcAUC}(\hat p,
l_\mathit{Val})$\
$\mathit{meanAUC}[j] \leftarrow mean(\mathit{AUC}_1[j], \cdots, \mathit{AUC}_5[j])$ $n^* \leftarrow {\operatornamewithlimits{argmax}}_j \mathit{meanAUC}[j]$ $\mathit{networks} \leftarrow m(D_1, N)$ $C_{n^*} \leftarrow \mathit{trainCl}(C, \mathit{rankedNWs}[1, \ldots, n^*], \mathit{D_1})$ $\hat p \leftarrow C_{n^*}(E_{2})$ $\mathit{AUC} \leftarrow \mathit{calcAUC}(\hat p, L_2)$ $\mathit{AUC}$
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.